The study, conducted by Haotian Zhang, Chen Wang, and Jun Moon of the Department of Electrical Engineering at Hanyang University in Seoul, introduces TACCO, short for TD3+BC with Actor-Critic Contrastive Optimization. The method builds on TD3+BC, a minimalist and widely used offline reinforcement learning algorithm, and augments it with supervised contrastive learning branches that explicitly model out-of-distribution actions. The authors define out-of-distribution actions as those that lack local data support in a given state, do not match the current state, or exhibit low reliability. Rather than treating these problematic actions as an afterthought, TACCO makes them a first-class target of representation learning, reshaping the internal embeddings of the agent so that unreliable actions become visibly distinguishable from trustworthy ones.

To understand why this matters, it helps to see how offline reinforcement learning goes wrong. When an agent learns only from a fixed dataset, its value function, the critic that estimates the long-term reward of taking an action in a state, is only trustworthy for state-action pairs that actually appear in the data. During policy improvement, the actor proposes new actions that may drift outside the support of the dataset. The critic, having never seen these pairs, tends to overestimate their value, a phenomenon known as value overestimation. The actor then chases these inflated estimates, proposing ever more exotic actions, and the whole learning process spirals into degradation. This feedback loop between an overconfident critic and an adventurous actor is the central failure mode of offline reinforcement learning.

Existing approaches generally respond in one of two ways. Policy-constraint methods, such as TD3+BC itself, tether the learned policy to the behavior policy recorded in the data, penalizing any action that strays too far from what the dataset contains. Conservative value regularization methods, exemplified by CQL, pessimistically lower the estimated values of out-of-distribution actions so the critic never overhypes them. Both families work, but as the Hanyang team points out, few methods explicitly model the problematic actions from a representation learning perspective. That is the gap TACCO fills: instead of only constraining behavior or deflating values, it teaches the network’s internal representations to separate reliable actions from unreliable ones, providing a structural safeguard that complements the existing toolkit.

TACCO comes in two variants, each targeting a different side of the actor-critic architecture and a different reward regime. The first, TACCO-A, is designed for dense reward environments, where feedback signal is abundant. It attaches a supervised contrastive learning branch to the actor. The actor consists of a shared state-encoding backbone, a standard policy output head, and an additional embedding head that concatenates the hidden state representation with the proposed action and projects the result into a 64-dimensional embedding. Contrastive learning then operates on three kinds of actions: data-driven actions drawn from the offline dataset, random actions, and cross-state policy actions. The training objective pulls embeddings of dataset-supported actions together while pushing embeddings of random and mismatched actions apart, so the actor’s representation space develops a clear geometry in which out-of-distribution actions stand out and can be suppressed.

The second variant, TACCO-B, is tailored to sparse reward environments, where rewards arrive rarely and unreliable value estimates are especially corrosive. Here the contrastive branch moves to the critic. TACCO-B filters training samples on the critic side based on two criteria: high Q-values and low uncertainty. Samples that pass both filters are treated as reliable, while the remainder are flagged as suspect. The method then applies contrastive constraints to both reliable and remaining samples simultaneously, encouraging the critic’s shared state-action encoder, a small network ending in a LayerNorm-normalized 64-dimensional embedding, to learn discriminative representations that reflect this reliability structure. By shaping the critic’s representation space rather than merely clipping its outputs, TACCO-B aims to stabilize value learning even when the reward signal is thin and misleading bootstrapped targets are easy to come by.

The architectural details are deliberately modest, which is part of the method’s appeal. Both variants retain the dual Q-network structure of TD3, with two independent critics that help mitigate function approximation error, and both keep the behavioral cloning term of TD3+BC that anchors the policy to the dataset. The contrastive modules are small multilayer perceptrons, and the authors report that performance is relatively robust to the weighting of the contrastive loss. Sensitivity analyses reveal a more interesting story for the other knobs: in TACCO-A, relaxing the behavioral cloning weight makes the policy more adventurous but destabilizes training, and increasing the contrastive weight steadily compensates for that instability. In TACCO-B, the fraction of high-value samples used for filtering must be tuned carefully, since too few samples starve the model of coverage while too many reintroduce the very low-quality actions the filter was meant to exclude.

The empirical evaluation is broad. The team compared TACCO against six state-of-the-art offline reinforcement learning algorithms, including CQL, TD3+BC, the Decision Transformer, IQL, DTQL, and O-DICE, across 27 tasks spanning the MuJoCo locomotion suite and the Adroit dexterous manipulation suite within the D4RL benchmark. All methods were trained under an identical protocol: one million gradient steps, evaluation every five thousand steps, a batch size of 256, and results averaged over five random seeds with 95 percent confidence intervals, using the standard D4RL score normalization. According to the authors, TACCO achieved superior performance on most of these tasks, which they interpret as validation that mitigating out-of-distribution action interference through contrastive optimization enhances policy stability, in both dense and sparse reward settings.

The broader significance of the work lies in its reframing of a classic problem. Distribution shift has usually been handled as a constraint or a penalty, a blunt instrument applied to actions or values. TACCO instead treats it as a representation problem: if the network’s embeddings make the boundary between supported and unsupported actions explicit, then both the actor and the critic can exploit that structure during training. This connects offline reinforcement learning to the wider wave of contrastive learning that has transformed vision and speech processing, and to a growing line of research that uses contrastive objectives for goal-conditioned control and planning. It also suggests a practical path forward, since the contrastive branches add little architectural complexity on top of a minimalist baseline that many practitioners already deploy.

Challenges remain before such methods reach real-world robotics and industrial control, where offline data is often suboptimal, heterogeneous, and collected under shifting conditions. The authors note that code will be made available on request, and the work was supported by Korean government research programs through KETEP, MOTIE, and IITP. As offline reinforcement learning matures into a serious option for settings where online exploration is prohibitive, techniques like TACCO point toward a hybrid future: conservative enough to trust the data, expressive enough to improve on it, and, crucially, aware of exactly where its own knowledge ends.

Subject of Research: Contrastive actor-critic optimization for mitigating out-of-distribution actions in offline reinforcement learning

Article Title: TACCO: Actor–Critic Contrastive Optimization for OOD Action Mitigation in Offline Reinforcement Learning

Article References: Zhang, H., Wang, C., & Moon, J. (2026). TACCO: Actor–Critic Contrastive Optimization for OOD Action Mitigation in Offline Reinforcement Learning. Machine Learning, 115(10), Article 241. https://doi.org/10.1007/s10994-026-07170-3

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07170-3

Keywords: offline reinforcement learning, contrastive learning, out-of-distribution actions, TD3+BC, actor-critic, D4RL benchmark, value overestimation, distribution shift, representation learning, MuJoCo, Adroit, machine learning

To understand why ALTAI matters, it helps to grasp the peculiar challenge that Chinese poses to machines. In English, a named entity such as a company or a person is usually delimited by spaces and signaled by an initial capital letter, so a model can often spot boundaries with relative ease. Chinese text offers no such signposts. The same character sequence can be segmented into words in several plausible ways, and the meaning of a character often depends on which words it belongs to. Consider that a single character might stand alone as a word, join with a neighbor to form a two-character word, or sit inside a longer multi-character phrase. A model that only looks at characters misses the lexical knowledge embedded in dictionaries, while a model that only looks at words risks losing sight of the fine-grained compositional structure of the language. The best-performing systems therefore try to use both levels of information simultaneously.

For years, the dominant approach has been the so-called word-character lattice framework. The idea is to build a lattice structure over the sentence: characters form the backbone, and every dictionary word that matches a span of characters is attached as an additional node. This allows the model to consult word-level information without committing to a single segmentation. A well-known implementation of this idea is the Flat-Lattice Transformer, which folds lattice nodes into a transformer architecture so that characters and words attend to one another. Lattice-based models have delivered solid gains, and they remain widely used in Chinese NER research. Yet, according to the authors of the new study, these methods share a common blind spot.

That blind spot concerns how lexical information is actually integrated. In existing lattice systems, word-level knowledge is typically injected through a dedicated encoder architecture, and the interactions between character representations and word representations are handled implicitly, as a byproduct of attention mechanisms. The authors argue that this design overlooks the complex interactions across different granularities between character-level and word-level representations. In other words, the model may receive word information, but it does not explicitly learn how that information should reshape its understanding of the individual characters that compose the words. Semantic dependencies, too, are underexploited, particularly in their interactions with character representations. The result is that lexical knowledge, one of the most valuable resources available for Chinese NER, is not fully exploited.

ALTAI addresses this weakness with two tightly coupled components. The first is a Cross Transformer, a module designed to explicitly model interactions among three kinds of representations: the character representations that form the basic units of the text, the word-level lexical representations drawn from matching dictionary entries, and semantic representations that capture higher-level meaning. Rather than letting these views of the sentence mix only incidentally, the Cross Transformer forces a direct exchange of information across the different granularities. Each character can consult the words it participates in, and the semantic layer can modulate how both characters and words are interpreted. This explicit cross-granularity modeling is what distinguishes ALTAI from earlier lattice encoders, where the dialogue between levels of representation was left largely to chance.

The second component is Cross-Granularity Contrastive Learning, a training strategy borrowed from the recent wave of self-supervised learning research. Contrastive learning works by teaching a model to pull related representations together and push unrelated ones apart in the embedding space. In ALTAI, the technique is applied across granularities: character representations are aligned with their corresponding lexical and semantic views, so that the model is encouraged to produce consistent representations across the character, word, and semantic spaces. A character that participates in a meaningful word should end up with an embedding that reflects that word’s presence, and the semantic view of the sentence should agree with what the character-level view suggests. This alignment acts as an additional training signal, shaping the representation space so that entities become more discriminative, meaning that the vectors for genuine entities stand apart more clearly from those of ordinary text.

The combination of these two ideas produces what the authors describe as discriminative entity representations by modeling cross-granularity interactions. The intuition is straightforward: when a character-level representation has been explicitly refined by lexical and semantic information, and when the training process has actively enforced consistency among those views, the final representation of an entity span carries far richer evidence than a representation built from characters alone. Downstream, the model’s sequence labeling decisions, deciding where an entity starts, where it ends, and what type it is, rest on this enriched foundation. The contrastive objective also serves as a form of regularization, discouraging the model from latching onto spurious patterns that appear in one granularity but not in others.

The empirical case for ALTAI rests on experiments across four benchmark datasets for Chinese named entity recognition. The authors report that ALTAI consistently outperforms strong baseline models, including the lattice-based architectures that have defined the state of the art in this area. Consistency across multiple datasets is a meaningful claim in NER research, where a method that excels on one corpus, say one dominated by news text, may falter on another drawn from social media or specialized domains. By holding its advantage across all four benchmarks, ALTAI suggests that its benefits stem from the general principle of explicit cross-granularity interaction rather than from quirks of any particular dataset. The work was supported by the National Natural Science Foundation of China and several Guangdong provincial research programs, reflecting sustained institutional investment in Chinese-language AI research.

The significance of this study extends beyond Chinese NER itself. Named entity recognition is a foundational task that feeds countless downstream applications: information extraction, question answering, knowledge graph construction, search, and recommendation systems all depend on reliably identifying entities in raw text. Improvements at this level propagate through entire pipelines. Moreover, the core insight of ALTAI, that representations at different granularities should be explicitly interacted and aligned rather than merely concatenated or implicitly mixed, is a design principle that could transfer to other languages and tasks where multiple levels of linguistic structure coexist. Languages without clear word boundaries, and tasks such as relation extraction or event detection that similarly straddle character, word, and sentence levels, are natural candidates.

The study also highlights a broader trend in machine learning: the migration of contrastive learning from computer vision into structured language tasks. Originally popularized for learning image representations without labels, contrastive objectives are now being adapted to align representations across views, modalities, and, as here, linguistic granularities. For Chinese NER specifically, the arrival of ALTAI signals a shift in how researchers think about lexical knowledge, not as a static feature to be injected into an encoder, but as a set of interacting perspectives that a model must learn to reconcile. As the field moves forward, the open-access publication of this work, published on 1 September 2026 with a permanent DOI, means that researchers worldwide can examine, reproduce, and build upon the approach, potentially accelerating progress on a problem that sits at the heart of making machines truly literate in one of the world’s most widely spoken languages.

Subject of Research: Cross-granularity contrastive learning for Chinese named entity recognition

Article Title: Semantic-aware cross-granularity contrastive learning for Chinese named entity recognition

Article References: Mao, S., Feng, Z., Li, J., Sun, C., Xiang, D., & Jiang, Y. (2026). Semantic-aware cross-granularity contrastive learning for Chinese named entity recognition. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02483-1

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02483-1

Keywords: Chinese named entity recognition, natural language processing, contrastive learning, Cross Transformer, lattice framework, lexical knowledge, semantic representations, representation learning, benchmark datasets, Complex & Intelligent Systems, deep learning, information extraction

The core problem is deceptively simple. In supervised learning, a model learns from labeled examples, so the supervision signal is handed to it directly. In clustering, no labels exist. The network must simultaneously learn discriminative features and assign instances to clusters, with each task supposed to help the other. Without any ground truth, the algorithm needs some source of guidance to construct its own supervision. That guidance, the survey contends, is always a prior: an assumption, whether explicit or implicit, that constrains the space of acceptable solutions. From the earliest deep clustering methods to today’s state-of-the-art systems, the history of the field is really the history of priors evolving.

The authors organize the landscape into six categories of prior knowledge. The first and oldest is the structure prior, inherited directly from classical clustering algorithms such as K-means, DBSCAN, spectral clustering, and agglomerative clustering. K-means assumes instances form spherical structures around centroids; DBSCAN assumes clusters are contiguous high-density regions; spectral methods assume data lie on a locally linear manifold whose neighborhood relations should be preserved. Early deep clustering methods such as the Deep Embedding Network, SpectralNet, and PARTY simply transplanted these mature assumptions into neural network objectives, computing graph Laplacians or enforcing self-representation properties in the latent space learned by autoencoders. These methods were interpretable and theoretically grounded, and they already outperformed classic K-means on raw features thanks to the neural network’s superior feature extraction.

The second category, the distribution prior, assumes that instances from different semantic classes follow distinct probability distributions. This assumption gave rise to generative deep clustering, built on variational autoencoders and generative adversarial networks. The landmark method VaDE fit a Gaussian mixture model in the latent space, sampling a cluster distribution, then a latent vector conditioned on that cluster, and finally reconstructing the input image, with all components jointly optimized by maximizing a variational evidence lower bound. ClusterGAN later replaced explicit Gaussian components with an adversarially learned latent space, introducing a discrete one-hot variable alongside the continuous noise vector to capture cluster identity, and penalizing deviations between encoded and sampled variables to keep clusters distinct. The survey notes that Gaussian components proved redundant and could blur discriminability, which motivated this shift toward implicit distribution learning.

The third and arguably most transformative prior is augmentation invariance, the idea that different transformed views of the same image, such as crops, color distortions, and rotations, preserve its semantic content. Rather than mining structure already present in the data, researchers began constructing new supervisory signals by augmentation. Mutual-information-based methods like IMSAT and IIC maximize the information shared between the cluster assignments of an image and its augmented counterpart, while regularized information maximization pushes assignments to be both unambiguous, by minimizing conditional entropy, and balanced across clusters, by maximizing marginal entropy. Contrastive methods take a related route, pulling positive pairs together in the representation space while pushing negative pairs apart, with theoretical work showing that instance-level contrastive learning is equivalent to maximizing mutual information. Methods such as PICA, Contrastive Clustering, and DRC extended this logic to the cluster level, treating entire cluster assignment vectors as instances to be contrasted, and Twin Contrastive Learning fused instance and cluster representations into a unified embedding.

The fourth prior, neighborhood consistency, exploits a striking empirical observation: features learned through self-supervised pretext tasks map semantically similar instances to nearby points in latent space. SCAN capitalized on this by training a cluster head to make consistent predictions for each instance and its k-nearest neighbors, with an entropy term preventing collapse onto a single cluster. NNM and GCC went further, folding neighborhood information directly into contrastive objectives. GCC in particular builds a normalized symmetric graph Laplacian from a k-nearest-neighbor graph and uses it to reweight the contrastive loss, attracting neighbors rather than only augmented views of the same image. This elegantly mitigates the false-negative problem, the situation where two instances of the same class are wrongly treated as negatives and pushed apart, a known failure mode of vanilla contrastive learning.

The fifth category, pseudo-labeling, rests on a self-referential assumption: predictions in which the model is highly confident are probably correct, and can therefore serve as surrogate labels. DEC, a pioneering method, computed soft assignments with a Student’s t-distribution over distances to learnable centroids, then sharpened those assignments by squaring the probabilities and training the network to match the sharpened targets via KL divergence. DeepCluster iterated between K-means on learned features and supervised training on the resulting pseudo-labels, though its performance was limited by weak initial representations. ProPos later showed that running the same expectation-maximization loop on features from a state-of-the-art self-supervised paradigm like BYOL dramatically improves results, demonstrating that pseudo-label quality is only as good as the semantics of the representation behind it. Newer methods such as TCL and SPICE refine the selection of confident samples, using cluster-wise top-K selection or prototype-based re-assignment to keep pseudo-labels balanced and accurate, and borrow semi-supervised techniques like FixMatch to squeeze more value from them.

The sixth and most recent prior breaks with the others entirely: instead of extracting knowledge from the data itself, it imports external knowledge, most notably textual semantics. SIC builds a semantic space from meaningful texts resembling category names, generates image pseudo-labels by matching image embeddings to text centers in a CLIP-pretrained space, and then trains the cluster head with cross-entropy plus neighborhood consistency. TAC retrieves a text counterpart for each image among representative nouns, improving K-means without extra training, and introduces a mutual distillation paradigm in which image and text modalities teach each other through cluster-level contrastive losses applied to each modality’s assignments and their cross-modal nearest neighbors. The survey’s benchmark experiments confirm the payoff: external-knowledge methods achieve state-of-the-art results across five widely used image benchmarks, including CIFAR-10, CIFAR-100, STL-10, ImageNet-10, and the fine-grained ImageNet-Dogs.

Stepping back, the authors identify two grand trends in how priors have evolved. The first is a shift from mining to constructing: early methods passively extracted assumptions already latent in the data, such as manifold structure or density, whereas augmentation-based methods actively manufacture supervision by transforming inputs. The second is a shift from internal to external: the field is moving from priors derived within the dataset toward knowledge imported from open-world resources, including pretrained vision-language models. The performance gains from different priors also turn out to be largely independent and composable, which explains why methods combining augmentation invariance, neighborhood consistency, and pseudo-labeling, such as ProPos, outperform approaches relying on any single prior.

The survey also maps the road ahead. Fine-grained clustering, such as distinguishing biological subspecies that differ only in subtle markings, defeats coarse priors because color and shape augmentations may erase exactly the features that matter. Non-parametric clustering, where the number of clusters is unknown, remains computationally expensive, though DeepDPM’s Dirichlet Process mixture framework with Metropolis-Hastings-guided split-and-merge operations offers a promising template. Fair clustering must confront biases in sensitive attributes like gender and race that can distort partitions in high-stakes domains such as healthcare and employment, with recent information-theoretic metrics beginning to quantify both quality and fairness. Multi-view clustering, which fuses complementary and consistent information from different sensors or modalities, continues to expand. The authors close with a provocative suggestion: as large pre-trained models like ChatGPT and GPT-4V mature, the next generation of clustering systems may draw supervision from external knowledge sources far richer than anything the data alone can provide, completing the field’s journey from mining priors to importing them.

Subject of Research: Prior knowledge in deep clustering methods for unsupervised data grouping

Article Title: A survey on deep clustering: from the prior perspective

Article References: Lu, Y., Li, H., Li, Y., Lin, Y., & Peng, X. (2024). A survey on deep clustering: from the prior perspective. Vicinagearth, 1(1), Article 4. https://doi.org/10.1007/s44336-024-00001-w

Image Credits: AI Generated

DOI: 10.1007/s44336-024-00001-w

Keywords: deep clustering, unsupervised learning, prior knowledge, contrastive learning, pseudo-labeling, neural networks, data augmentation, representation learning, generative models, external knowledge, machine learning survey, clustering benchmarks

The answer proposed by the researchers is a framework called SCLS, short for shapelet-based soft contrastive learning for multivariate time series clustering. Published online in the journal Knowledge and Information Systems, the method combines two powerful ideas from machine learning, discriminative subsequences known as shapelets and a gentler form of contrastive learning, into a four-stage pipeline that avoids the noise and uncertainty that plague earlier approaches. In experiments spanning dozens of real-world datasets, the authors report that SCLS achieved the highest performance ranking compared with state-of-the-art clustering methods.

To appreciate why this matters, it helps to understand what shapelets are and why they have become a cornerstone of time series analysis. A shapelet is a short, distinctive subsequence that captures the essence of a pattern, such as the characteristic spike in an electrocardiogram that signals a particular cardiac event or the telltale wobble in a sensor reading that precedes a machine failure. Since shapelets were introduced as a primitive for data mining in 2009, researchers have used them for classification and clustering because they are inherently interpretable: a human can look at a shapelet and see exactly which fragment of the signal drove the decision. That transparency is precious in domains like medicine and industrial monitoring, where a bare accuracy number is not enough.

Contrastive learning, meanwhile, has become one of the dominant paradigms for learning representations without labels. The idea is to teach a neural network to pull similar examples together in an embedding space while pushing dissimilar ones apart. In time series work, this is usually done through data augmentation: the same series is distorted in two different ways, for example by jittering, scaling, or cropping, and the network is trained to recognize that both distorted views came from the same source. The problem, as Wang and Yang point out, is that augmentation introduces noise and uncertainty of its own. A badly chosen distortion can destroy the very pattern that defines a class, and the augmented views may no longer resemble anything in the real data. Worse, in shapelet-based methods that lean on augmentation, the extracted shapelets can drift away from true subsequences of the original signals, eroding both interpretability and clustering quality.

SCLS sidesteps this difficulty by working directly with the original sequences rather than their distorted copies. The framework unfolds in four carefully sequenced stages. In the first, an attention mechanism fuses multiple feature components drawn from the original series. Attention, the same mechanism that powers modern language models, allows the network to weigh different parts of the signal according to their relevance, which reduces uncertainty and helps extract more representative shapelets. Rather than treating every time point and every channel as equally informative, the model learns where the discriminative structure lives.

The second stage mines multi-scale shapelets using a perceptually important points selection method based on information gain. Perceptually important points are the salient peaks, valleys, and inflection points of a curve, the skeleton of a signal that survives even when the details are smoothed away. By selecting these anchor points and measuring how much each candidate subsequence contributes to distinguishing between groups, quantified through information gain, the method can discover shapelets at multiple lengths and granularities. This multi-scale view matters because the discriminative fingerprint of a time series may be a brief transient in one dataset and a long, slow trend in another.

Third, the framework employs a shapelet encoder enhanced with a pseudo-label clustering mechanism. Pseudo-labeling is a well-known strategy in semi-supervised learning in which a model’s own provisional cluster assignments are treated as soft training signals. Here, the encoder embeds the mined shapelets into structure-aware representations, meaning the embedding preserves the relationships among shapelets rather than treating them as isolated vectors. These structure-aware embeddings then serve as the substrate for the final stage of learning.

That final stage is the soft contrastive learning strategy that gives the method its name and its distinctive character. Instead of the hard binary logic of conventional contrastive learning, where two samples are either a positive pair to be attracted or a negative pair to be repelled, soft contrastive learning assigns graded, probabilistic relationships between samples. The hierarchical strategy introduced in SCLS jointly optimizes the shapelet representation space at two levels: the temporal level, where fragments within a series are compared, and the sample level, where whole series are compared with one another. The result is a representation space whose geometry is explicitly shaped for clustering analysis, with coherent groups emerging naturally rather than being forced by rigid pairwise constraints.

The authors evaluated the framework through extensive experiments on dozens of real-world datasets, measuring clustering quality against a battery of state-of-the-art competitors. According to the published results, SCLS attained the highest performance ranking among the methods tested, suggesting that the combination of augmentation-free shapelet extraction and hierarchical soft contrastive optimization translates into a genuine advantage rather than a theoretical curiosity. The study draws on a rich lineage of prior work, from the original shapelet primitive through shapelet transformers presented at the 2024 KDD conference, multiview unsupervised shapelet learning, and recent soft contrastive approaches for time series presented at ICLR, positioning SCLS as a synthesis that resolves the specific weaknesses of each ingredient when used alone.

The practical implications reach well beyond the benchmark suite. Clustering multivariate time series underpins customer segmentation in dynamic markets, electricity consumption analysis, fault diagnosis in industrial equipment, and the discovery of patterns in online popularity dynamics, all areas represented in the study’s bibliography. Because the shapelets at the heart of SCLS remain anchored to true subsequences of the original data, practitioners gain not only clusters but also an explanation of what distinguishes them, a feature that matters whenever algorithmic decisions must be audited or acted upon by domain experts. The code supporting the findings is available from the corresponding author upon reasonable request, and the research was supported by the Guangdong Basic and Applied Basic Research Foundation and the National Natural Science Foundation of China. As sensors multiply and the volume of unlabeled temporal data grows, methods like SCLS point toward a future in which machines can find the hidden rhythms of the world without needing humans to label every beat first.

Subject of Research: Shapelet-based soft contrastive learning for multivariate time series clustering

Article Title: SCLS: Shapelet-based soft contrastive learning for multivariate time series clustering

Article References: SCLS: Shapelet-based soft contrastive learning for multivariate time series clustering. (n.d.). https://doi.org/10.1007/s10115-026-02902-2

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02902-2

Keywords: multivariate time series, clustering, shapelets, contrastive learning, attention mechanism, pseudo-labeling, information gain, representation learning, data mining, unsupervised learning, feature fusion, Knowledge and Information Systems

The core problem the team set out to solve is a familiar one in multimodal machine learning. Most existing pipelines process each modality independently, using a vision network for frames, an audio network for sound and a language model for captions or titles, and only combine the outputs in a late fusion stage. That design is convenient, but it comes at a cost. Information that only makes sense across modalities, such as the way a piece of music reinforces the mood of a scene, or the way a caption disambiguates a visually ambiguous object, can be lost before fusion ever happens. The separate encoders may also learn feature distributions that are poorly aligned with one another, so that the fusion layer receives streams that are difficult to reconcile. The result, the authors argue, is inefficient representation learning and a ceiling on classification accuracy for tasks such as microvideo category recognition.

MTA departs from that recipe by building extraction and fusion into a shared Transformer structure. Transformers, which rely on self-attention to weigh relationships between elements of a sequence, have become the backbone of modern artificial intelligence, powering everything from large language models to video vision transformers such as ViViT and multimodal self-supervised systems such as VATT. What distinguishes MTA from a conventional Transformer is what happens in its fusion layer. Instead of applying a static attention mechanism that treats all inputs uniformly, MTA incorporates a dynamic modality-aware attention mechanism. In practice, this means the model learns to adaptively attend to modality-specific information, shifting its focus depending on what a particular clip demands. For a clip in which the soundtrack carries most of the semantic signal, the audio stream can receive greater weight; for a clip driven by on-screen text or visual content, the mechanism can redirect attention accordingly. The dynamic weighting also helps align the heterogeneous feature distributions produced by different encoders, reducing the mismatch that plagues late-fusion designs.

A second conceptual innovation in the paper is the notion of categorical modality. In many multimodal classifiers, the fused representation is compared against class labels only at the very end, through a final classification layer that treats labels as inert symbols. MTA instead facilitates direct interaction between the fused multimodal features and label semantics during prediction. By giving the label side a learned semantic representation that can exchange information with the fused features, the model can exploit the relationships between categories themselves, for example when two microvideo categories share visual or textual vocabulary. The authors report that this end-to-end interaction improves learning effectiveness, because the prediction stage is no longer a passive readout but an active participant in shaping the shared representation.

The third pillar of the study is an investigation of parameter-sharing strategies. Sharing parameters across the branches of a multimodal network is a double-edged sword: it can regularize the model, reduce its size and encourage features that generalize across modalities, but excessive sharing can blur modality-specific distinctions. The researchers examined two sharing strategies within the MTA framework and found that both further boosted model performance, suggesting that a carefully chosen degree of weight sharing complements the dynamic attention mechanism rather than competing with it. This finding is practically significant for deployment, because parameter sharing reduces memory footprint and training cost, which matters when models must process the enormous volumes of short video that platforms handle daily.

Evaluation was carried out on a category-balanced benchmark dataset, a deliberate choice that removes the distorting effect of imbalanced class distributions, in which a model can appear accurate simply by favoring frequent categories. Across this benchmark, MTA consistently outperformed state-of-the-art multimodal fusion methods in classification accuracy. The authors present this as evidence of the robustness and scalability of the approach for real-world microvideo understanding tasks, and they have released their code publicly on GitHub, allowing other researchers to reproduce the results and build on the architecture.

The significance of the work becomes clearer when set against the broader landscape of multimodal research. Recent years have produced spectacular large multimodal models, from Flamingo and BLIP-3 to Llama 3 and the Qwen vision-language family, which achieve broad capabilities through massive pretraining. Yet specialized tasks such as microvideo category recognition, scene recognition and venue classification still benefit from architectures tailored to the specific structure of the problem, where clips are short, modalities are tightly interleaved and label sets are well defined. Prior work on microvideos has explored joint sequential-sparse modeling, neural multimodal cooperative learning, gated fully convolutional sequence models and attention-enhanced joint learning networks, each attacking pieces of the fusion problem. MTA’s contribution is to unify extraction and fusion within a single Transformer backbone while making the fusion itself adaptive, rather than fixed.

The applications that motivated the research are far-reaching. Content recommendation systems depend on accurate representations to surface clips that match user interests, and better multimodal understanding translates directly into more relevant feeds. Regulatory compliance is an equally pressing driver, since platforms must detect content that violates policies even when the violation is conveyed by the combination of audio and imagery rather than by either alone. Multimedia search, meanwhile, requires representations that capture what a clip is about, not merely what it looks like. A model that can weigh modalities dynamically and reason over label semantics offers a stronger foundation for all three use cases than pipelines that lose cross-modal information before fusion.

Technically, the design reflects several lessons from the Transformer literature. Self-attention allows every element of a sequence to attend to every other element, which makes it a natural fit for aligning features across modalities that unfold at different rates and granularities. Prior multimodal Transformers, such as the Multimodal Transformer for unaligned language sequences, demonstrated that cross-modal attention can bridge streams that are not temporally aligned. MTA extends this line by making the attention itself modality-aware and dynamic, so the alignment process is conditioned on the content being fused. Combined with the categorical modality mechanism, the architecture treats classification not as an afterthought but as an integrated stage of representation learning, echoing the end-to-end philosophy that has driven progress across deep learning.

The study, received in April 2025 and accepted in July 2026, was supported by the 111 Project, the Natural Science Foundation of Liaoning Province and the National Natural Science Foundation of China. Its limitations are those inherent to the field: the benchmark is category-balanced, while real-world feeds are anything but, and the reported data are available only on request due to privacy and ethical restrictions. Nevertheless, the consistent gains over existing fusion methods, the public code release and the architectural clarity of the approach position MTA as a meaningful step toward short-video AI systems that genuinely watch, listen and read at the same time. As platforms grapple with billions of daily microvideo views, architectures that fuse modalities adaptively within a shared Transformer structure are likely to shape how the next generation of recommendation, moderation and search systems makes sense of the world’s shortest, densest media.

Subject of Research: Multimodal microvideo understanding using a shared Transformer architecture with dynamic modality-aware attention

Article Title: Multimodal microvideo understanding based on the shared transformer structure

Article References: Zhong, Z., Zhang, W., Yu, H., & Zhu, Z. (2026). Multimodal microvideo understanding based on the shared transformer structure. Neural Computing and Applications, 38(19), Article 758. https://doi.org/10.1007/s00521-026-12409-0

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12409-0

Keywords: microvideo understanding, multimodal fusion, transformer, attention mechanism, deep learning, video classification, multimedia search, content recommendation, neural networks, representation learning, Neural Computing and Applications, Multimodal

The core limitation the researchers identify is conceptual rather than computational. Existing expert-finding methods are built to score the suitability of one expert for one question, treating each user as an isolated candidate. That design works reasonably well when a question falls squarely within a single specialty, but it breaks down when a question spans multiple domains or requires complementary skills. The team formation methods that do exist, meanwhile, tend to rely on heuristic search procedures that are computationally expensive and disconnected from the learning of the representations that describe experts in the first place. The new approach bridges this gap by jointly learning how to represent experts, questions and their interactions within one unified architecture.

Technically, the framework rests on two complementary encoders. The first is a structure encoder built on graph convolutional networks, the now-standard family of neural architectures for learning over graph-structured data. The community question answering platform is modeled as a graph in which users, questions and answers form interconnected nodes, and the graph convolutional network aggregates information from a node’s neighbors to produce an embedding that captures higher-order topological proximity, meaning an expert’s position in the wider web of past collaborations and answered questions. The authors use a two-layer graph convolutional network, a depth that prior work has shown is sufficient to capture these higher-order relationships without over-smoothing the representations.

The second component is a content encoder that reads the actual text of questions and answers. Rather than relying on exact keyword overlap, which is brittle in the face of vocabulary variation, the model employs a kernel pooling mechanism built on radial basis function kernels. This technique, originally developed for neural ad-hoc ranking in web search, transforms soft word-matching similarities between a new question and an expert’s past answers into a vector of soft term-frequency statistics. Eleven kernels are used: one narrow kernel tuned to capture exact matches, and ten evenly spaced kernels that act as soft bins for graded degrees of similarity. The result is a rich, differentiable signal of how well the language of a new question aligns with the demonstrated expertise of a candidate.

What makes the architecture genuinely novel is how these signals are fused and trained. The outputs of the structure encoder and the content encoder are passed into a multi-layer perceptron ranker that produces final relevance scores, and the entire pipeline, from word embeddings through graph convolutions to the ranking layer, is optimized jointly with gradient descent using the Adam optimizer. Gradients flow backward through the kernel pooling layer into the translation matrix of word similarities and ultimately into the word embeddings themselves, meaning the model learns not just how to rank experts but how to represent language in a way that serves expert discovery. This end-to-end discipline stands in contrast to heuristic team formation methods, which involve iterative combinatorial search and cannot benefit from representation learning at all.

The empirical case for the method rests on experiments across four real-world datasets drawn from Stack Exchange communities: android, history, dba and physics. For individual expert finding, the model outperformed the strongest baselines by at least 4.4 percent in NDCG and 6.7 percent in MAP, two standard ranking metrics that capture, respectively, how closely a predicted ranking matches the ideal ordering of experts and how robustly relevant experts appear across the ranking list. In help-hurt analyses across individual test questions, the proposed method beat the best baseline on 36.4 percent of questions for NDCG and improved MAP over the best baseline by 32.5 percent, indicating that the gains are broad rather than concentrated in a few easy cases.

For team formation, the researchers evaluated three complementary metrics. Skill coverage measures how completely the tags of questions a discovered team has previously answered overlap with the tags of a new question, capturing whether the team collectively possesses the background knowledge required. Collaboration level counts how often team members have answered the same questions in the past, normalized across all pairs, and the gold standard match score measures the fraction of a discovered team that overlaps with the experts who actually answered the question in reality. On skill coverage, the new model achieved at least 4.7 percent broader coverage than collaborative expert finding baselines, and it beat the strongest team formation baselines on 15.2 to 18.4 percent of test questions while losing on only 3.5 to 7.4 percent.

Perhaps the most revealing findings concern trade-offs. The strongest baseline on collaboration level, a method called EnC, tends to retrieve high-degree nodes in the network, users who have answered many questions and therefore share many neighbors, which inflates collaboration scores. But the analysis shows that high collaboration does not guarantee skill coverage: teams formed by EnC had on average 5.91 percent lower skill coverage than those discovered by the new model, while the new model’s teams carried on average 1.7 times higher collaboration than the best team formation baseline. On the gold standard match metric, the new method retrieved actual answerers roughly twice as often as the best baseline and achieved 4.18 times the score of the second-best team formation method, suggesting its teams resemble the real collaborative groups that form organically on these platforms.

Scalability, often the Achilles heel of graph neural approaches, receives careful treatment. The authors show theoretically that the graph convolutional layers scale linearly in the number of edges times the embedding dimension, while the ranker scales with the number of nodes times the squared embedding dimension, a complexity profile comparable to existing neural baselines and free of the iterative search costs of heuristic methods. Empirically, throughput declines as datasets grow, as expected, but even on the largest physics dataset the model processed more than half a sample per second during training and more than one sample per second during inference, speeds that the authors argue are compatible with batch processing and even near real-time expert recommendation in latency-sensitive settings.

Beyond the immediate application to question answering platforms, the work signals a broader shift in how machine learning treats collective expertise. Ablation results show that combining topological and textual signals yields on average 4.4 percent and 8.61 percent better NDCG and MAP than topology alone, although structural data proved the more powerful single signal, improving performance by 16.44 and 19.7 percent over text-only variants. The lesson is that expertise lives simultaneously in what people write and in where they sit within a network of past collaborations, and that models capable of learning both jointly, rather than stitching them together after the fact, are better positioned to find not just who knows the answer, but who can answer it together.

Subject of Research: Joint representation learning with graph neural networks and kernel pooling for discovering individual experts and collaborative expert teams in community question answering

Article Title: Joint Representation Learning for Expert Discovery

Article References: Etemadi, R., Zihayat, M., Feng, K., Adelman, J., Zarrinkalam, F., & Bagheri, E. (2026). Joint Representation Learning for Expert Discovery. Machine Learning, 115(10), Article 234. https://doi.org/10.1007/s10994-026-07166-z

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07166-z

Keywords: expert finding, community question answering, graph neural networks, kernel pooling, team formation, representation learning, question routing, skill coverage, Stack Exchange, machine learning, neural ranking, collaborative expertise

The motivation comes from a well-known blind spot in graph neural networks. These architectures, along with classical spectral embedding methods, build their representations by aggregating local neighborhood information, layer by layer or eigenvector by eigenvector. That strategy works well when the properties that matter are themselves local, such as the content similarity of connected documents. But it systematically misses global metric properties: how quickly the volume of a space grows with distance, whether the space is hyperbolic in the sense of Gromov, and what the boundary at infinity looks like. Those coarse-geometric signatures are precisely what distinguish a grid from a tree, or a flat lattice from a negatively curved hierarchical network, even when the two graphs are indistinguishable at the level of small neighborhoods.

Enakoutsa’s framework formalizes this intuition with three mathematical guarantees. First, the learned representations are stable under quasi-isometries: if two metric spaces are coarsely equivalent, meaning one can be mapped onto the other with bounded distortion, their distance-based feature representations remain correspondingly close. This is the property that makes the features genuinely geometric rather than artifacts of a particular graph labeling. Second, the framework satisfies a universal approximation property for the distance-generated algebra on compact metric spaces, meaning that with enough anchors and appropriate nonlinearities, the model can approximate any continuous function in that algebra arbitrarily well. Third, the method extends naturally to the boundary at infinity of Gromov-hyperbolic spaces, the idealized boundary points that encode how geodesics diverge, which is where the hierarchical structure of many real networks actually lives.

The parametric model works as follows. A set of anchor points is chosen in the metric space, and each data point is described not by its adjacency relations but by the vector of its distances to those anchors. These distances are then passed through nonlinear activations with learnable scales, so that the entire feature map is differentiable and the anchors themselves can be trained by gradient descent. The result is what the author calls a nonlinear feature algebra: a family of functions generated by distance measurements, closed under the nonlinear transformations the network applies. Because distance to an anchor is a global quantity, every feature already carries information about the position of a point relative to the whole space, not just its immediate surroundings.

The experimental test was designed to isolate exactly the capability that local methods lack. Enakoutsa constructed pairs of graphs that share low-order local structure but differ in their coarse geometry, such as grid graphs and tree graphs. Locally, both look like ordinary networks; globally, a grid is flat and grows quadratically while a tree is hyperbolic and grows exponentially. A tanh-activated distance representation achieved a mean accuracy of 77.8 percent, with a standard deviation of 5.8 percent, in separating grid from tree topology across ten runs. The local and spectral baselines managed only 48.7 plus or minus 5.9 percent and 45.5 plus or minus 5.1 percent respectively, essentially chance performance, and the difference was highly significant with a paired t-test yielding p less than ten to the minus five.

The activation function turned out to matter enormously. When the saturating tanh nonlinearity was replaced with a non-saturating ReLU, performance on the same grid-versus-tree task jumped to 99.9 plus or minus 0.4 percent, near-perfect separation. The same ReLU-activated representation also achieved 96.2 plus or minus 2.4 percent accuracy in distinguishing a grid from a Barabási-Albert scale-free graph, another pair whose members differ dramatically in growth rate and curvature. The contrast between the two activations suggests that the saturating behavior of tanh compresses exactly the long-range distance information the method depends on, while the unbounded ReLU preserves the dynamic range needed to encode global geometry.

On real-world networks, however, the picture became more nuanced, and the author is careful to report this scope dependence rather than overclaim. The distance-based representation outperformed spectral embedding on the Zachary karate club network and on the Les Misérables co-occurrence network, both small graphs whose structure is strongly shaped by community and hierarchy. But on the Cora and PubMed citation networks, which reach up to roughly 20,000 nodes, spectral embedding won. The reason, according to the paper, is that the labels in those citation datasets track content homophily, the tendency of papers to cite similar papers, rather than coarse geometry. When the signal in the data is local and content-driven, local aggregation remains the right tool; when the signal is global and metric, the distance-based approach dominates.

This honest boundary of applicability may prove as informative as the headline results. The work does not claim that distance features should replace message passing everywhere. Instead, it identifies a class of problems, those governed by growth rate, hyperbolicity, and boundary structure, where local methods are provably and empirically inadequate, and supplies a principled alternative with mathematical guarantees. The connection to coarse geometry also gives the field a vocabulary it has lacked: quasi-isometry invariance is a much weaker and more robust equivalence than graph isomorphism, and methods that respect it should generalize across graphs that look different locally but share the same large-scale shape.

The implications reach into several active areas of machine learning. Hyperbolic embeddings have become popular for hierarchical data such as knowledge graphs, taxonomies, and biological phylogenies, precisely because negatively curved spaces can embed tree-like structure with low distortion. The new framework offers a way to detect and exploit that hyperbolicity directly from distances, rather than committing to a particular embedding space in advance. It also connects to geometric deep learning, the broad research program aiming to build neural architectures whose inductive biases respect the symmetries and structure of the underlying data domain, and to the theory of neural and nonlocal operators, both listed among the paper’s keywords.

Published open access under a Creative Commons Attribution 4.0 license and released as a citable version of record with a permanent DOI, the paper arrives from an unexpected direction: a mathematics department rather than a computer science lab, and without external funding. That provenance is fitting for a contribution whose main strength is conceptual. By grounding representation learning in the mathematics of large-scale metric structure, the work suggests that the next advance in how machines understand networks may come not from deeper stacks of local message passing, but from teaching them to measure distances and read the shape of the whole space at once. Whether the approach can be scaled efficiently to graphs with millions of nodes, and whether hybrid architectures can combine local content signals with global geometric ones, are the natural next questions for this emerging line of research.

Subject of Research: Distance-based representation learning for graphs using nonlinear functions of distances to learnable anchor points, grounded in coarse geometry

Article Title: Distance-Based Representation Learning with Nonlinear Feature Algebras

Article References: Enakoutsa, K. (2026). Distance-Based Representation Learning with Nonlinear Feature Algebras. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11882-x

Image Credits: AI Generated

DOI: 10.1007/s11063-026-11882-x

Keywords: geometric deep learning, coarse geometry, representation learning, graph neural networks, hyperbolic geometry, universal approximation, spectral embedding, metric spaces, quasi-isometry, networks, machine learning, Gromov-hyperbolic spaces

In a paper published in Knowledge and Information Systems, a team led by Tianyi Ma and Yanfang Ye of the University of Notre Dame, together with collaborators at the University of Connecticut and Amazon GenAI, introduces BHyGNN+, a self-supervised framework for learning representations of heterophilic hypergraphs, structures in which a single relationship, called a hyperedge, can link any number of entities that may have little or nothing in common. The work extends the group’s earlier supervised model, BHyGNN, and removes its most restrictive requirement: the need for labeled training data.

The heterophily problem is easy to illustrate. Consider a shopping dataset in which each hyperedge is a shopping cart. A single cart routinely contains products from wildly different categories, such as diapers, a phone charger, and a novel, so the items joined by that hyperedge are semantically dissimilar by design. Political datasets behave similarly: a bill co-sponsored by legislators spans party lines, so nodes connected through the same hyperedge often belong to different classes. When a neural network averages information across such dissimilar neighbors, it blurs the very distinctions that downstream tasks, like classifying a product or predicting a legislator’s party, depend on. The researchers found that on heterophilic benchmarks, even a simple multilayer perceptron that ignores network structure entirely can outperform several sophisticated hypergraph neural networks, a striking sign that conventional message passing can actively hurt in these settings.

The earlier BHyGNN model tackled this by making propagation selective. Rather than forcing every node to exchange information with every hyperedge it touches, the model learns, for each node-hyperedge pair, whether the node should broadcast its features into the hyperedge or receive information back from it. This decision is made by a Variational Broadcast Autoencoder Network that samples discrete broadcast and receive actions through a Gumbel-Softmax relaxation, and the resulting representations are refined by an incorporation transformer with multi-head attention. Trained end to end, the model worked well, but its propagation actions were guided by downstream class labels, which are precisely the resource that is scarce, expensive, or entirely absent in many practical applications, from emerging illicit drug trafficking networks to newly formed e-commerce catalogs.

BHyGNN+ replaces that label dependence with a contrastive self-supervised objective built on an elegant structural idea: hypergraph duality. Every hypergraph has a dual in which the roles of nodes and hyperedges are swapped. Each original hyperedge becomes a node in the dual, and each original node becomes a hyperedge connecting the dual nodes it belonged to. Mathematically, this amounts to transposing the incidence matrix, and it is a lossless operation. The authors prove that the dual preserves all task-relevant information, since the transform is its own inverse, and that the normalized walk operators of a hypergraph and its dual share the same nonzero eigenvalues. In spectral terms, the two views are isospectral: they encode identical nontrivial structure while exposing complementary, node-centric versus hyperedge-centric, pathways through the data.

The training procedure capitalizes on this complementarity. The framework applies four stochastic augmentations, namely node attribute masking, hyperedge perturbation that only removes members to avoid injecting noisy connections, hyperedge dropping, and node dropping, to generate corrupted views of both the original hypergraph and its dual. Each augmented view is passed through the shared hypergraph encoder, and a readout function pools node and hyperedge embeddings into a single global representation. The contrastive loss then maximizes the cosine similarity between the representation of an augmented view and that of its corresponding augmented dual. Crucially, this formulation needs no negative samples at all. Most existing hypergraph contrastive methods, such as TriCL and HyperGCL, rely on negative pairs to prevent representation collapse, but constructing high-quality negatives in large hypergraphs is difficult and computationally costly. By contrasting a structure against its own dual, BHyGNN+ sidesteps the problem entirely, using only positive pairs.

The authors also supply a theoretical argument for why duality is especially well suited to heterophilic data. They show that two hyperedges sharing even a single node, which become neighbors in the dual, have provably higher expected similarity in their class composition than random hyperedge pairs, with a strictly positive contribution arising from the shared structure alone. In a heterophilic hypergraph, where members of one hyperedge span many classes, this means the dual view reorganizes apparently incoherent node-level structure into locally coherent hyperedge-level neighborhoods. Because the global readout aggregates over all of these dual neighbors, the coherence propagates into the contrastive signal, giving the encoder a stable learning signal even when node-level homophily is absent. The bound is tightest precisely when shared nodes carry maximally diverse labels, which is exactly the heterophilic regime where other methods struggle most.

The empirical evaluation is unusually broad, spanning eleven datasets and thirteen baselines. The heterophilic benchmarks include Senate, Congress, and House, where nodes are legislators labeled by party affiliation, and Walmart, where nodes are co-purchased products; the homophilic set covers co-authorship networks such as Cora-CA and DBLP, co-citation networks including Cora, PubMed, and Citeseer, and a Twitter hypergraph for classifying roles in illicit drug trafficking networks, plus a synthetic four-class heterophilic hypergraph generated from a contextual stochastic block model. Across these benchmarks, BHyGNN+ consistently outperformed both supervised and self-supervised state-of-the-art methods, including ED-HNN, SheafHGNN, TriCL, and HypeBoy, despite never seeing a label during pretraining. Paired t-tests with Bonferroni correction confirmed that the improvements over most baselines are statistically significant, with adjusted p-values near zero on the heterophilic datasets, and the advantage was most pronounced against structure-ignoring MLPs and the generative method HypeBoy.

The framework’s advantages widen as labels become scarcer. Under a split in which only ten percent of nodes are used for training, BHyGNN+ showed a substantial improvement over its supervised predecessor BHyGNN, which struggles to learn sensible propagation actions from limited supervision. In controlled experiments on synthetic hypergraphs with homophily scores ranging from 0.4 to 0.8, all methods improved as homophily increased, but BHyGNN+ achieved the best results in the most heterophilic settings, beating the strongest baseline, TriCL, by roughly four percentage points under both moderate and high feature noise. Sensitivity analyses showed stable performance across hidden dimensions from 64 to 1024, with an augmentation ratio of 0.2 proving optimal, since heavier perturbation injects too much noise. Diagnostic plots of the embedding space ruled out representation collapse: training loss never reached zero, embedding norms remained stable, and the effective rank of the embedding matrix stayed high, notably higher for heterophilic datasets, suggesting the model automatically allocates richer representational capacity to harder structures.

The implications reach well beyond benchmark tables. Hypergraphs arise wherever group interactions matter, including biochemical networks, transaction records, legislative activity, and online communities, and in most of these domains, ground-truth labels are rare or lag reality. A method that extracts usable representations from raw structure alone, without negatives and without annotations, lowers the barrier to applying higher-order network learning in exactly those settings. The authors note open questions that remain, including extending the framework to hyperedge-level and graph-level tasks, designing augmentations tailored specifically to heterophilic structure, and deepening the theory of why duality works so well. For now, though, the message is clear: when a network looks incoherent from one angle, its mirror image may hold the pattern.

Subject of Research: Self-supervised representation learning on heterophilic hypergraphs using hypergraph duality

Article Title: Bhygnn+: unsupervised representation learning for heterophilic hypergraphs

Article References: Ma, T., Qian, Y., Wang, Z., Zhang, Z., Zhang, C., & Ye, Y. (2026). Bhygnn+: unsupervised representation learning for heterophilic hypergraphs. Knowledge and Information Systems, 68(1), Article 265. https://doi.org/10.1007/s10115-026-02834-x

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02834-x

Keywords: hypergraph neural networks, self-supervised learning, contrastive learning, heterophily, hypergraph duality, representation learning, graph neural networks, unsupervised learning, node classification, higher-order networks, machine learning, Knowledge and Information Systems

The research, led by Ronghua Tian and Hong Yu of the Chongqing Key Laboratory of Computational Intelligence at Chongqing University of Posts and Telecommunications, together with Yaogang Geng and Xiaoling Wang of East China Normal University, addresses multimodal knowledge graph completion, abbreviated MMKGC. Unlike classical knowledge graph completion, which relies purely on symbolic triples, multimodal completion also exploits the rich information attached to entities in other forms: images, textual descriptions, and numerical attributes. A knowledge graph describing a famous painting might carry not only the fact that it was created by a particular artist but also a digital image of the canvas and a descriptive paragraph about its style. The premise of multimodal completion is that these additional signals can help a model infer missing links that the structure alone cannot reveal.

There is, however, a catch that the authors identify as a fundamental weakness in the field. Real-world multimodal knowledge graphs are profoundly uneven in how much information they provide for each entity. Some entities are lavishly documented, with high-resolution images and lengthy descriptions, while others may have only a name and a single relation. Existing methods, the paper argues, tend to overlook this imbalance of modal information among entities. The consequence is that models develop an overreliance on the dominant modality, typically the visual one when images are plentiful or the textual one when descriptions dominate. When a model leans too heavily on one channel, the weaker channels contribute little, and for entities whose strong modality is missing or uninformative, completion performance degrades sharply. The imbalance is not a marginal nuisance; it is baked into the structure of the data itself.

To confront this problem, the researchers propose a method called MI-MKGC, which attacks modality imbalance from two complementary directions. The first component is a relation knowledge-guided dynamic modal fusion module. The key insight here is that the type of relation connecting two entities carries prior knowledge about which modality is likely to matter most. Consider a relation such as has image or looks similar to: for these, visual features are clearly decisive. For a relation like was born in or is the author of, textual and structural information carries far more weight. Rather than fusing modalities with fixed weights that apply uniformly across all triples, MI-MKGC dynamically adjusts the weight assigned to each modality depending on the relational type between the entities involved. The relation acts as a guide, telling the fusion mechanism where to look before it decides how to blend the available signals.

This dynamic fusion stands in contrast to a long line of earlier approaches. Foundational embedding models such as TransE, which translated entities through relation-specific vectors in a continuous space, and ComplEx, which extended the idea into complex-valued spaces, established the template of learning purely from graph structure. Later systems incorporated literals and images, and more recent transformer-based architectures such as MEAformer and optimal-transport-based methods like OTKGE pushed multimodal fusion further. But in most of these designs, the fusion weights are either learned globally or computed without reference to the specific relation at hand. MI-MKGC’s use of relational types as priors for modality weighting represents a targeted response to the observation that different relations draw on different kinds of evidence, and that a one-size-fits-all fusion strategy systematically shortchanges entities whose informative modality happens to be the weaker one.

The second component of the method is a joint modality adversarial training module, which addresses the imbalance from the data side rather than the fusion side. Inspired by generative adversarial networks, a framework introduced by Goodfellow and colleagues in 2014, this module generates negative samples specifically designed to enhance weak modality information. In adversarial training, a generator produces challenging examples while a discriminator learns to distinguish them from genuine data, and the two improve in tandem. Here, the generator crafts negative triples that are hard to reject, with a particular focus on strengthening the representations of modalities that are underrepresented in the training signal. By forcing the model to grapple with adversarially constructed examples that probe its weak channels, the training process gradually amplifies the contribution of those channels, reducing the dominance of the strong modality and producing more balanced entity representations.

The choice of adversarial training is grounded in a substantial body of prior work. Wasserstein generative adversarial networks and their improved variants provided more stable training dynamics for generative models, and adversarial techniques have already been applied to knowledge graph settings, for instance in KBGAN, which used adversarially generated negative samples to sharpen knowledge graph embeddings, and in GAN-based approaches to text classification and information retrieval. MI-MKGC adapts this lineage to the specific problem of modality imbalance, using the adversarial game not merely to improve discrimination between true and false triples but to deliberately rebalance the informational diet of the model across modalities.

The empirical evaluation is notably thorough. The authors tested MI-MKGC on three public multimodal knowledge graph datasets against sixteen state-of-the-art baseline methods, spanning the range from purely structural embedding models to the latest multimodal fusion architectures. The results showed consistent advantages, with the most striking gain appearing in the Hit@1 metric, which measures the proportion of cases where the model’s top-ranked prediction is exactly correct. On this metric, MI-MKGC improved by up to 5.2 percent over the best existing compared method. In a field where incremental gains of fractions of a percent are common, an improvement of that magnitude on the strictest ranking metric is a meaningful advance, and it suggests that directly modeling modality imbalance, rather than treating it as background noise, unlocks performance that fusion tweaks alone cannot reach.

The practical implications extend well beyond benchmark datasets. Multimodal knowledge graphs underpin question answering systems that combine language models with structured knowledge, recommendation engines that blend product images with user behavior, and video captioning systems that must reconcile visual and textual evidence. In all of these applications, the entities involved vary enormously in how well documented they are, and a system that silently favors well-illustrated or well-described entities will perform unevenly across the long tail of less prominent items. By dynamically reweighting modalities according to relational context and actively training up weak channels through adversarial sampling, MI-MKGC offers a blueprint for more equitable and robust multimodal reasoning, particularly for the rare entities where completion is hardest and most valuable.

The team has released its implementation code publicly on GitHub, lowering the barrier for other researchers to build on the approach. The work was supported in part by the National Natural Science Foundation of China, and the authors report no conflicts of interest. As knowledge graphs continue to expand into multimodal territory, absorbing images, text, and structured facts at scale, the problem of imbalance will only grow more pronounced, since no curation process can guarantee uniform coverage across billions of entities. What this study demonstrates is that the solution lies not in richer data alone but in algorithms that recognize, at the level of individual relations and individual entities, which signals to trust and which to strengthen. In teaching models to listen to their quietest channels, the researchers may have pointed the way toward knowledge systems that are not only more accurate but fundamentally fairer in how they represent the world.

Subject of Research: Multimodal knowledge graph completion using relation-guided fusion and adversarial training to address imbalanced modality information

Article Title: Multimodal knowledge graph completion method for addressing imbalance of modal information

Article References: Tian, R., Yu, H., Geng, Y., & Wang, X. (2026). Multimodal knowledge graph completion method for addressing imbalance of modal information. Data Mining and Knowledge Discovery, 40(6), Article 106. https://doi.org/10.1007/s10618-026-01272-1

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01272-1

Keywords: knowledge graph completion, multimodal learning, modality imbalance, adversarial training, knowledge graph embeddings, generative adversarial networks, link prediction, dynamic modal fusion, negative sampling, data mining, machine learning, representation learning

The study, published open access in the journal Machine Learning by Sintija Stevanoska, Christian L. Camacho Villalón, Sašo Džeroski and Katharina Dost, tackles a well-known pain point in applied machine learning. Labels are expensive: healthcare systems accumulate millions of patient records, yet expert-annotated outcomes remain scarce. Self-supervised learning promises to bridge that gap by defining pretext tasks from the intrinsic structure of data — masking features and reconstructing them, corrupting values and asking a model to spot the damage, or learning invariant representations across contrastive views. For tables, however, these methods inherit an awkward problem. Unlike images or sentences, tabular rows lack spatial or sequential structure, mix numerical and categorical features, and often contain missing values and tangled inter-dependencies. And pretraining on very large unlabeled pools is computationally expensive while potentially importing noisy or unrepresentative samples that actively degrade learning.

The researchers asked a deceptively simple question: is a small, well-chosen subset of unlabeled data sufficient for effective pretraining? To answer it, they ran what they describe as the first large-scale evaluation of unlabeled data subsampling for tabular self-supervised learning, spanning 25 benchmark datasets drawn from the OpenML-CC18 suite and covering domains from biology and finance to medicine and climate science. Four representative self-supervised models were tested — TabNet, VIME, SCARF and SAINT — deliberately chosen to span the field’s three main paradigms. TabNet and VIME learn by masking or corrupting features and reconstructing them, SCARF relies on contrastive views generated by resampling features from their empirical distributions, and SAINT combines contrastive and denoising objectives inside a transformer that attends over both rows and columns. This diversity of pretext tasks and encoder architectures ensured the conclusions would not hinge on one particular model family.

At the heart of the study lies a family of selection strategies borrowed and adapted from active learning, the branch of machine learning concerned with identifying the most informative examples to label. The team compared random sampling against three principled approaches. Disagreement sampling, based on the classic query-by-committee method, trains a committee of 100 decision trees on bootstrap samples of the labeled data, restricting each tree to a random subset of features in the manner of random forests, and scores every unlabeled point by the normalized entropy of the committee’s votes — a value from 0, indicating complete agreement, to 1, indicating uniform disagreement. Diversity sampling applies k-means clustering to the unlabeled pool and picks one representative per cluster, the point closest to each centroid, enforcing a minimum cluster size to avoid crowning representatives of tiny outlier groups. Optimal transport sampling treats the labeled and unlabeled sets as empirical distributions, computes an entropically regularized transport plan between them with the Sinkhorn algorithm, and assigns each unlabeled point its expected transport cost to the labeled distribution — a low score signaling strong alignment with the labeled data, a high score signaling underrepresented regions of the feature space.

Crucially, all of these strategies are deliberately model-agnostic. They are computed before any self-supervised pretraining begins, using no gradients, losses, embeddings or augmentations from the downstream model. Score-based methods partition the unlabeled pool into percentile brackets — low, medium and high — and samples are then drawn from a chosen bracket, either uniformly at random or via the diversity-based cluster representatives. This design allowed the authors to evaluate the sampling approaches consistently across all four pretext objectives, isolating the effect of data selection from the idiosyncrasies of any particular architecture.

The experimental scale was formidable. Labeled subsets of 100, 500 and 1000 examples were drawn under both unbiased, stratified conditions and a deliberately biased regime in which labeled examples were preferentially selected from one branch of a decision stump, simulating real-world selection bias. Unlabeled subsets of 100 to 1000 examples were drawn under every combination of strategy, subset size, bracket and five random seeds. Performance was measured through three lenses: downstream classification quality using macro-averaged area under the precision-recall curve, total wall-clock time including both selection and training, and representation quality assessed by linear probing — training a simple logistic regression classifier on frozen pretrained embeddings to test whether class information is linearly encoded in the learned latent space.

The headline finding is striking. Subsampling only 100 to 1000 unlabeled examples often matched or exceeded full-data pretraining across both labeled-data regimes. In the low-label, unbiased setting, the best sampling strategies reached up to 119 percent of the full-data baseline and 114 percent of a purely supervised baseline that skipped pretraining entirely. Under biased labeled selection the gains grew even more pronounced, climbing to 123 percent of the full-data baseline and 115 percent of supervised learning. The effect was strongest exactly where it matters most: when labels were scarcest, and the choice of pretraining data had the greatest opportunity to shape what the model learned. Optimal transport distance and disagreement-based sampling performed most consistently, but the humble random sampler was surprisingly competitive, hinting that for small labeled sets, simply reducing redundancy and avoiding overfitting to the unlabeled distribution may itself act as a form of regularization.

The computational story is equally compelling. Pretraining on the full unlabeled pool consistently demanded 60 to 70 percent more time than the sampling strategies, and the selection step itself added a negligible share of the total runtime. Random sampling typically finished in 29 to 42 percent of full-data time, with optimal transport close behind; diversity-based selection was the most expensive to compute, consuming up to 70 percent of the full-data baseline’s runtime without delivering consistent accuracy dividends. In other words, competitive or better predictive performance was achieved at roughly 30 to 40 percent of the computational cost — a margin that alone would justify subsampling in resource-constrained environments even without accuracy improvements. Linear probing reinforced the picture: targeted subsampling consistently matched or exceeded the representation quality of full-data pretraining, and in many configurations full-data self-supervision actually fell below the supervised baseline, an adverse effect the selective strategies appeared to mitigate, particularly under label bias.

To understand when and why subsampling pays off, the team conducted a meta-analysis linking dataset characteristics to observed gains. They extracted statistical and complexity-related metafeatures for each dataset — among them the performance of a nearest-neighbor classifier built on the single most important feature, the stability of local neighborhood structure across cross-validation folds, the minimum support of decision-tree leaves, the median absolute deviation of the features, and trimmed means capturing distributional asymmetry. The patterns that emerged were model-specific rather than universal. TabNet and SCARF gained most from subsampling on datasets where a single dominant feature carries the predictive signal, while SAINT and VIME tended to benefit more on structurally complex tasks; datasets containing a feature with strongly asymmetric central mass favored using all the data. No single metafeature predicted gains across all architectures, pointing to a decisive interplay between dataset geometry and each model’s inductive bias.

The authors are candid about limitations and open questions. The number and size of unlabeled subsets and the granularity of the score brackets were fixed by design, so the study does not determine an optimal data budget for a given dataset, and the meta-analysis, based on 25 datasets, is exploratory rather than confirmatory. Future directions include adaptive budget selection guided by validation performance or meta-learned dataset properties, curriculum-style ordering of the unlabeled pool, incorporating self-supervised model feedback into the selection loop, and training a meta-learner to recommend strategy and subset size automatically. Even so, the message for practitioners is already actionable and, in its own way, subversive: in tabular self-supervised learning, the brute-force assumption that more unlabeled data is always better does not survive scrutiny. A few hundred well-chosen — or even randomly chosen — examples can deliver richer representations and better predictions at a fraction of the cost, a result that could reshape how resource-conscious teams approach pretraining on the structured data that quietly runs the world.

Subject of Research: Unlabeled data selection for efficient self-supervised learning on tabular data

Article Title: Less is More: Unlabeled Data Selection for Efficient Tabular Self-supervised Learning

Article References: Stevanoska, S., Camacho Villalón, C. L., Džeroski, S., & Dost, K. (2026). Less is More: Unlabeled Data Selection for Efficient Tabular Self-supervised Learning. Machine Learning, 115(9), Article 208. https://doi.org/10.1007/s10994-026-07133-8

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07133-8

Keywords: self-supervised learning, tabular data, data selection, unlabeled data, active learning, optimal transport, query-by-committee, computational efficiency, representation learning, machine learning benchmarks, TabNet, VIME