Researchers Karim Hallal, Adam Dandan, Sireen Hammoud and Seifedine Kadry of the Lebanese American University have now tackled this selection problem head-on with a framework they describe as meta-learning for pseudo-label utility prediction. Their central insight is deceptively simple: rather than running every candidate clustering algorithm on every new dataset, one can learn from past experience which algorithm is likely to serve a given dataset best. The team introduces a metric called Label Substitution Efficiency, or LSE, which quantifies how well pseudo-labels can stand in for human annotation. LSE is computed by training two classifiers on the same dataset, one on pseudo-labels produced by a clustering algorithm and one on the ground-truth labels, and then comparing their balanced accuracy. If the pseudo-label-trained classifier nearly matches the ground-truth-trained one, the clustering algorithm has done its job of substituting for the annotator.

The technical machinery behind the study is worth unpacking. The authors benchmark six pseudo-label generators across 94 datasets drawn from OpenML, the open repository of machine learning datasets that has become a standard testbed for meta-learning research. For each dataset and each clustering algorithm, they compute a six-dimensional LSE vector, one entry per algorithm, capturing the full profile of how each candidate performs as a label substitute. This vector becomes the prediction target for the meta-learners. The idea of using downstream classification performance as the yardstick for clustering quality is itself a departure from tradition. Classical clustering validation indices, such as silhouette scores or Davies-Bouldin measures, judge clusters by their internal geometry. LSE instead asks a more practical question: do these clusters, converted into labels, actually teach a classifier something useful?

With the LSE tables in hand, the researchers trained two complementary families of meta-learners. The first is a classification meta-learner that treats algorithm selection as a discrete recommendation problem: given a description of a new dataset, it directly outputs the clustering method it predicts will yield the highest LSE. The second is a regression meta-learner that predicts the entire six-dimensional LSE vector, offering far more flexibility. Because it estimates the utility of every candidate rather than committing to a single winner, the regression model supports confidence thresholding, ranking and shortlisting strategies. A practitioner could, for example, ask for the top three algorithms and run only those, or reject a recommendation outright if the predicted utility falls below a threshold. This graded output acknowledges a truth that discrete recommenders often gloss over: sometimes no clustering algorithm will substitute well for labels, and it is better to know that before investing compute.

A crucial design question in any meta-learning system is how to describe a dataset to the learner. These descriptions, called meta-features, traditionally include simple statistical summaries such as the number of instances, the number of attributes, class entropy proxies, correlation statistics and measures of skewness or dimensionality. The study evaluates five different meta-feature representations, including a novel one the authors call Option C2, a multi-scale dictionary learning representation. Dictionary learning, a technique rooted in sparse coding, represents each dataset as a sparse combination of learned basis elements across multiple scales, potentially capturing structural signatures that hand-crafted statistics miss. The approach draws on earlier work on coupled dictionary learning for unsupervised feature selection and on online matrix factorization methods developed by Mairal and colleagues.

The headline result is striking. The best meta-classifier, a logistic regression model operating on the C2 dictionary-learning features, achieves 43.6 percent top-1 accuracy under leave-one-out validation, meaning it picks the single best clustering algorithm for a previously unseen dataset nearly half the time. That figure may sound modest until it is compared against the baseline: the classic de Souto ranking approach from 2008, a foundational meta-learning method for clustering algorithm selection, manages only 22.3 percent on the same task. In other words, the new framework nearly doubles the performance of the established default. In a domain where the best algorithm varies substantially across datasets and where running all candidates defeats the purpose of annotation-free learning, that improvement translates into real savings of time and computation.

Statistical rigor underpins the claims. Paired significance testing shows that the meta-learner significantly outperforms a weak heuristic baseline with a p-value of 0.003, a result that would clear conventional thresholds for statistical significance with room to spare. Perhaps more interesting is what the tests did not find: no single meta-feature representation significantly outperformed another, with all pairwise comparisons yielding p-values above 0.08. The authors interpret this as evidence that the framework’s benefit is robust to the specific feature-engineering choice rather than dependent on any one representation. For practitioners, that is good news. It suggests the framework does not hinge on a fragile, bespoke encoding of datasets; the meta-learning signal is strong enough to survive whatever reasonable description of the data one feeds it.

The study situates itself within a rapidly growing literature on automated algorithm selection. Earlier systems such as cSmartML combined meta-learning with hyperparameter tuning for clustering, while more recent efforts like CLAMS have pursued zero-shot model selection, and deep learning approaches such as ClustRecNet have attempted end-to-end recommendation of clustering pipelines. Work on Dataset2Vec has even explored learning meta-features directly from raw data rather than computing them by hand. What distinguishes the new framework is its target: instead of predicting abstract clustering quality, it optimizes for downstream classification efficiency, tying the recommendation directly to the end goal of building a working classifier from pseudo-labels. The authors also employ SHAP, the game-theoretic attribution method introduced by Lundberg and Lee, to interpret which meta-features drive the recommendations, adding a layer of explainability to what could otherwise be an opaque selection process.

The practical implications extend well beyond the benchmark. Pseudo-labeling has become a staple technique in domains where labels are scarce, from remote sensing applications that delineate snow cover in satellite imagery to self-training pipelines for tabular data in medicine and finance. In each of these settings, someone must currently choose a clustering algorithm, often by trial and error, and each trial consumes computational resources that scale with dataset size. A meta-learner that recommends the right algorithm from dataset properties alone, before any clustering is run, collapses that search to a single prediction. The regression variant’s shortlisting capability offers a middle path for risk-averse users: run only the top-ranked candidates and stop early if one achieves a predicted utility above a confidence threshold.

The authors have made their work reproducible and accessible. All 94 OpenML dataset identifiers and the computed LSE tables are publicly available, along with the complete codebase on GitHub, inviting the community to extend the benchmark, test additional meta-feature representations and plug in new clustering algorithms. As annotation costs continue to rise with the growing scale of tabular data in industry and science, frameworks like this one point toward a future in which machines not only learn from data but also learn which learning strategy to use, quietly and automatically, before a single human label is ever requested.

Subject of Research: Meta-learning for unsupervised clustering algorithm selection via pseudo-label utility prediction

Article Title: Meta-learning for pseudo-label utility prediction: unsupervised algorithm selection via downstream classification efficiency

Article References: Hallal, K., Dandan, A., Hammoud, S., & Kadry, S. (2026). Meta-learning for pseudo-label utility prediction: unsupervised algorithm selection via downstream classification efficiency. International Journal of Data Science and Analytics, 22(1), Article 331. https://doi.org/10.1007/s41060-026-01296-2

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01296-2

Keywords: meta-learning, algorithm selection, pseudo-labeling, clustering, Label Substitution Efficiency, OpenML, dictionary learning, tabular data, unsupervised learning, classification, SHAP, automated machine learning

The research team, led by Jinhyeok Jang and corresponding author Daeseon Choi, frames the problem with unusual clarity. Label flipping is a form of data poisoning in which the observed class of a training example is changed while the underlying features remain intact. Because the features themselves look perfectly normal, simple feature-level screening catches nothing. Worse, verifying whether a label is truly correct in domains like finance or healthcare often requires scarce expert knowledge, so corrupted labels can survive unnoticed all the way into production. As organizations increasingly rely on crowdsourced annotation, outsourced labeling, and third-party data integration to cut costs, the attack surface only grows.

What makes the new work particularly timely is its focus on how modern label-flipping attacks have evolved. Early attacks flipped labels at random, scattering noise across the dataset. Recent strategies are far more surgical. Decision-boundary attacks use surrogate models to identify low-margin samples sitting perilously close to the classifier’s decision frontier, while optimization-based methods such as ALFA and its ALFA-Tilt variant solve constrained optimization problems to select the flip set that maximally degrades the learner. The authors’ visualizations show that these targeted attacks produce localized poisoning patterns, with flipped samples clustering together in feature space rather than appearing as isolated outliers. That clustering defeats density-based and global outlier detectors, because the poisoned points masquerade as a legitimate local group.

The team’s answer is a defense built on a counterintuitive idea: learn what the data looks like without trusting any labels at all. Their method, called BYOL-Union, first trains a self-supervised encoder using BYOL, or Bootstrap Your Own Latent, a technique originally developed for images that learns representations by predicting augmented views of the same instance without any negative pairs or labels. Applied to tabular data, the encoder absorbs the intrinsic geometric structure of the feature space, untouched by whatever corruption may lurk in the observed labels. Only after this label-free representation learning is complete does the system consult the labels, and then solely to measure how much each sample disagrees with its neighbors.

The disagreement scoring is where the method earns its name. For each sample, the system finds its k nearest neighbors in the embedding space and counts how many of them carry a different observed label. A genuinely flipped sample tends to sit among neighbors whose true labels differ from its own forged one, producing a high disagreement score. Crucially, the method does not rely on a single fixed view of the neighborhood. Just as image-based self-supervised learning uses crops and flips to see one object from multiple angles, the framework generates a second view by adding Gaussian noise, with a standard deviation of 0.1, to continuous features only, leaving categorical features untouched to avoid inventing invalid categories. Each sample is then probed in both the original and the perturbed embedding spaces, and suspicious sets from the two views are merged by a union rule.

That union rule is deliberately recall-oriented. A sample flagged as suspicious in either view enters the final suspicious set, maximizing the chance of catching poisoned labels that are exposed in at least one neighborhood perspective, at the cost of some additional false positives. The authors also designed for realistic deployment, where the true poisoning ratio is unknown: instead of assuming an oracle budget, the practical version uses z-score thresholding at 1.0 with any-view exceedance, which their ablations show performs close to the oracle reference while remaining entirely oracle-free.

The evaluation is unusually broad. Six public tabular benchmarks spanning network intrusion detection, finance, and healthcare, including NSL-KDD, UNSW-NB15, Bank Marketing, Credit Card Fraud, BRFSS 2015 Diabetes Health Indicators, and the Diabetes 130-US Hospitals dataset, were poisoned at ratios of 10, 20, and 30 percent under random, decision-boundary, and optimization-based attacks. Four heterogeneous target models, spanning support vector machines, deep neural networks, FT-Transformers, and XGBoost, were then trained on sanitized data. The results show a consistent pattern: under decision-boundary flipping, the original feature-space kNN detector achieved a recall of only 0.601, while the BYOL-based and SCARF-based variants reached 0.773 and 0.792 respectively, and paired Wilcoxon tests confirmed that the SSL-based detectors significantly improved both recall and F1 over Curie, LS-SVM, and original-space kNN baselines.

The study is equally candid about limits. Under ALFA-Tilt, the strongest optimization-based attack tested in a small-scale setting, detection became markedly harder for every detector, and the robust-training defense FLORAL achieved the smallest downstream accuracy gap, though it cannot identify which specific samples are poisoned and is tied to particular model architectures. The authors stress that improved detection does not always translate directly into recovered test accuracy, because removing suspicious samples changes the training set itself and can thin out supervision for minority classes. Detection quality, they argue, should be judged on its own terms, with downstream recovery treated as a complementary, setting-dependent benefit rather than the sole criterion.

Perhaps the most striking result concerns adaptive adversaries. The team constructed defense-aware attacks in which the attacker knows the sanitization mechanism and selects flips that remain locally plausible in the detector’s own neighborhood space. Against a raw feature-space detector, such an adaptive attack cut recall from 0.668 to 0.435. Yet the SSL-based union detectors held firm: BYOL-Union maintained a recall of 0.760 even when the attacker targeted its own representation space, and SCARF-Union reached 0.800 under the corresponding adaptive attack. The layered evidence from learned representation spaces, it appears, is not fully dismantled by an adversary who optimizes against any single neighborhood view.

The broader lesson resonates beyond this one defense. As machine learning systems are deployed in domains where a single misclassified transaction or missed intrusion carries real consequences, the integrity of training labels deserves the same security attention as model architecture. The Soongsil team’s framework, which will see code released on request, offers a practical, model-agnostic preprocessing step: it flags suspicious samples before any downstream classifier is fit, works alongside tree ensembles and neural networks alike, and leaves room for future extensions toward label correction, sample reweighting, and robust retraining with uncertainty. In a field where attackers increasingly aim at the data rather than the model, defenses that learn to see the data on its own terms may prove essential.

Subject of Research: Label-flipping attack detection and data sanitization in tabular machine learning using multi-view self-supervised representations

Article Title: Multi-view self-supervised learning for label-flipping robustness in tabular data: a comparative study

Article References: Multi-view self-supervised learning for label-flipping robustness in tabular data: a comparative study. (n.d.). https://doi.org/10.1007/s10489-026-07452-2

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07452-2

Keywords: label flipping, data poisoning, self-supervised learning, BYOL, tabular data, data sanitization, k-nearest neighbors, adversarial machine learning, SCARF, network intrusion detection, machine learning security, Applied Intelligence

The new method, called CausalCACTUS, is described in an open-access paper published in the journal Machine Learning on 28 September 2026. It extends an earlier framework known as CACTUS, which generated counterfactual explanations in a compressed latent space learned by a variational autoencoder, while allowing users to specify contextual constraints such as age or family status that the explanation must respect. CausalCACTUS adds what CACTUS lacked: an explicit model of cause and effect among the input features, learned automatically from data rather than supplied by human experts.

The technical core of the approach rests on three pillars. First, the researchers use a causal discovery algorithm called ReX, which trains predictive models for each feature and applies bootstrapped Shapley value attributions to identify which features depend on which. The output is a directed acyclic graph, a map of parent-child relationships such as the chain linking revolving credit utilization to the number of credit lines and then to a person’s debt ratio. Second, this graph is converted into dynamic feasibility constraints during the counterfactual search: a child feature is only allowed to change if at least one of its parents changes as well. Third, a phase-based optimization strategy first drives the search toward any valid counterfactual and then refines it for contextual alignment and minimal distance, with a weighting parameter that gradually shifts emphasis from prediction validity to context preservation.

The counterfactual search itself takes place not in the original feature space but in a four-dimensional latent space learned by a beta-variational autoencoder. This composite model includes an auxiliary classification head that predicts the user-defined context from the latent representation, allowing the optimizer to penalize candidates that drift outside the specified context, for example an applicant older than fifty. A causal distance loss, computed with an adaptive binary mask that is updated at every iteration, penalizes changes to features whose causal parents have not yet been modified. Feature values are also clipped to ranges observed in the training data after each decoding step, keeping every candidate within plausible bounds.

The evaluation covered five publicly available tabular datasets spanning credit risk, income prediction and legal education, ranging from 462 to 16,668 instances, with binary context variables derived from demographic and financial attributes. Against four baseline methods, including LatentCF++, two variants of PrototypeCF and the original CACTUS, CausalCACTUS ranked among the top two on validity in eight of ten comparisons and improved validity over the original CACTUS from roughly 0.7 to around 0.9. More strikingly, it achieved the best context-alignment scores, measured by a local outlier factor computed within the target context, in nine of ten scenarios, and the best causal-compact consistency, a metric combining causal constraint satisfaction with sparsity of changes, in nine of ten configurations.

Statistical analysis reinforced these findings. A Friedman test followed by a Wilcoxon-Holm post-hoc analysis across ten benchmark tasks showed that CausalCACTUS achieved the best average ranks, between 1.0 and 1.4, on proximity, context alignment, compactness and causal consistency. Ablation studies revealed a division of labor between the two new components: the phase-based optimization primarily boosted validity and compactness, while the causal feasibility constraints mainly improved context alignment and causal consistency. The researchers also mapped how the loss-weighting parameter alpha trades validity against context alignment, identifying a balanced setting near 0.7.

A concrete example from the credit dataset illustrates why the framework matters. For a loan applicant over fifty initially classified as high risk, competing methods proposed income changes ranging from a decrease of more than 15,000 dollars to increases exceeding 5,000 dollars, figures the authors describe as implausible for someone at that life stage. CausalCACTUS instead suggested a moderate adjustment of 1,605 dollars, required fewer overall feature modifications, and respected the learned dependency chain between credit utilization, credit lines and debt ratio, which the baseline methods violated. Correlation analysis of feature changes across test instances showed that CausalCACTUS produced strong correlations concentrated on causally related feature pairs, closely matching the learned graph, while baselines either showed weak correlations or correlations misaligned with causal structure.

The authors are candid about the limitations of their approach. Because it relies on a causal graph estimated by a discovery algorithm rather than a fully specified structural causal model, missing edges in the estimated graph could introduce inaccurate dependencies into the search; expert-defined graphs can substitute where reliable domain knowledge exists. The generative model is also a bottleneck: reconstruction errors or misalignment between the latent representation and the underlying causal mechanisms may cause information loss or unintended feature changes. Future work, the team suggests, could replace the beta-variational autoencoder with diffusion models, normalizing flows or causally informed generative variants, and could jointly learn latent representations and causal structures with uncertainty estimates attached to the learned graph.

The implications reach well beyond credit scoring. Counterfactual explanations have been linked to the right to explanation under the General Data Protection Regulation, and as algorithmic decisions spread through finance, hiring and education, the gap between mathematically valid explanations and practically actionable ones becomes a matter of public trust. By showing that causal consistency and user-defined context can be enforced simultaneously in a latent-space search, without demanding a fully specified causal model as a prerequisite, CausalCACTUS offers a template for explanations that a rejected applicant, a loan officer and a regulator could all accept. The source code has been released publicly on GitHub, and the authors point toward extensions to time-series classification and sequential prediction as the next frontier for causally grounded, context-aware explainability.

Subject of Research: Causally consistent and context-aware counterfactual explanations for machine learning classifiers

Article Title: CausalCACTUS: A Causally Consistent and Context-Aware Framework for Counterfactual Explanations

Article References: Wang, Z., García, D., Enguita, J. M., Chatterjee, S., & Jansson, M. (2026). CausalCACTUS: A Causally Consistent and Context-Aware Framework for Counterfactual Explanations. Machine Learning, 115(10), Article 233. https://doi.org/10.1007/s10994-026-07155-2

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07155-2

Keywords: counterfactual explanations, explainable AI, causal discovery, machine learning, variational autoencoder, structural causal models, latent space optimization, algorithmic recourse, model interpretability, context-aware learning, tabular data, GDPR

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