The new method, called PMFS-NRDLC, was developed by Zhenzhen Sun, Shuhao Shan, Jinghua Liu, Hongbo Zhang and Jin Gou of Huaqiao University, together with Yuanlong Yu of Fuzhou University. Writing in the journal Data Mining and Knowledge Discovery, they describe a framework that unifies label noise removal and dimensionality reduction inside a single optimization process. The motivation is straightforward: multi-label datasets routinely contain thousands of feature dimensions, and existing partial multi-label methods, which focus almost exclusively on disentangling true labels from false ones, leave themselves exposed to the curse of dimensionality. Redundant and irrelevant features dilute the signal the model needs, inflate computational costs, and ultimately cap the accuracy any disambiguation technique can achieve.

The core insight behind the approach lies in what the authors call the exclusive prior of partial multi-label noise. In this learning paradigm, each example comes with a candidate label set that contains all of its true labels plus some spurious ones, but crucially, any label outside that candidate set is guaranteed to be false. PMFS-NRDLC exploits this guarantee by building a position-constrained regression model that imposes strict zero constraints on the noise terms corresponding to those non-candidate labels. In practical terms, the model is forbidden from ever assigning confidence to a label that is known to be wrong. This structural constraint, the researchers argue, fundamentally prevents the spurious confidence assignment that plagues softer disambiguation schemes, allowing the algorithm to strip away label noise with far greater precision than methods that must guess which candidate labels are genuine.

Denoising alone, however, is not enough, because the estimated confidence of each label must remain faithful to the geometry of the data itself. To enforce this, the method introduces feature-space manifold topology under a smoothness assumption: examples that sit close together in the original feature space should carry similar label confidence profiles. By regularizing the confidence-learning process with this manifold structure, the algorithm ensures that the local structural characteristics of the raw features are preserved in the disambiguated labels. This coupling matters because a denoising step that ignores the underlying data distribution can easily produce confidences that are internally tidy but disconnected from what the features actually support, quietly degrading any downstream classifier.

Label relationships add a further layer of complexity. In real multi-label problems, categories are not independent: a photo tagged as beach is more likely to also show sand or ocean, and in biology, genes sharing a molecular function often share related functions too. Yet these correlations are latent and change as the model learns, which is why the team designed an iterative dynamic label graph update strategy. Rather than fixing a static graph of label similarities at the outset, the algorithm revises the graph during optimization, capturing time-varying latent correlations as the confidence estimates evolve. The graph, in turn, regularizes the feature weights: labels that are correlated are constrained to share similar feature importance, so evidence relevant to one category is naturally propagated to its close relatives.

The final piece of the framework is the feature selection mechanism itself. PMFS-NRDLC adopts the l2,1-norm, a sparsity-inducing mathematical device well known in embedded feature selection, to constrain the matrix of feature weights. Unlike filters that score features independently of any model, or wrappers that repeatedly train classifiers on candidate subsets, embedded selection learns the weights and the predictor simultaneously. The l2,1-norm encourages entire rows of the weight matrix to shrink to zero, effectively eliminating uninformative features while keeping those that matter across the label set. The result is a compact feature subset selected in full awareness of both the noise structure of the labels and the correlations among them.

To test whether this architectural ambition translates into measurable gains, the researchers ran extensive experiments across a range of benchmark datasets, comparing PMFS-NRDLC against state-of-the-art competitors in partial multi-label feature selection. According to the paper, the method consistently outperformed the alternatives. The breadth of the comparison was considerable: the cited literature spans probabilistic graph matching, adversarial disambiguation, subspace optimization, dual-graph regularization, label distribution learning, and anchor-graph-based fast adaptive schemes, reflecting how rapidly this subfield has grown. One competitor, G3WI, failed to produce results within an acceptable timeframe on the YeastMF dataset, and its performance there was treated as the worst for ranking purposes — a reminder that scalability remains a genuine differentiator in high-dimensional settings.

The experimental design also attended to statistical rigor and robustness. The authors employed established procedures for comparing classifiers over multiple datasets, and parameter sensitivity analyses indicated that the method’s behavior was stable across its tuning knobs, with sensitivity experiments on additional evaluation metrics yielding similar results. The work was supported by the Natural Science Foundation of Fujian Province, the Foundation for Cultivated Young Talents of Fujian Province, the National Natural Science Foundation of China, and the Fundamental Research Funds for the Central Universities of Huaqiao University. The authors state that all data included in the study will be made available on request from the first author, and they declare no conflicts of interest.

Why should a general audience care about an optimization framework buried in a data mining journal? Because weakly supervised learning with noisy labels is quietly becoming the norm rather than the exception. Web-scale image collections are tagged by algorithms and crowds, medical records are coded by busy clinicians, and scientific databases accumulate annotations of uneven quality. Any system that must learn from such data faces the twin burdens PMFS-NRDLC addresses: figuring out which annotations to trust, and doing so efficiently despite feature spaces that have exploded in size. A method that handles both simultaneously, rather than treating them as separate preprocessing chores, points toward learning systems that are more robust by construction.

There are also practical dividends. Feature selection reduces the dimensionality that downstream models must process, which can translate into faster training, lower memory footprints, and models that are easier to interpret — a valuable property when practitioners need to understand which features drive predictions. The dynamic label graph, meanwhile, offers a template for how latent structure can be revised during learning rather than frozen in advance, an idea likely to resonate beyond partial multi-label learning. As the volume of imperfectly labeled, high-dimensional data continues to grow, frameworks that fuse noise removal, structural regularization, and embedded dimensionality reduction into one coherent objective are likely to shape how the next generation of learning algorithms is built. The study, published as volume 40, article 114 of Data Mining and Knowledge Discovery on 12 October 2026, is available via its DOI, 10.1007/s10618-026-01279-8.

Subject of Research: Feature selection for partial multi-label learning combining label noise removal and dynamic label correlation

Article Title: Feature selection for partial multi-label learning via noise removal and dynamic label correlation

Article References: Feature selection for partial multi-label learning via noise removal and dynamic label correlation. (n.d.). https://doi.org/10.1007/s10618-026-01279-8

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01279-8

Keywords: partial multi-label learning, feature selection, label noise removal, label confidence, dynamic label correlation, dimensionality reduction, weakly supervised learning, manifold regularization, l2,1-norm sparsity, machine learning, data mining, optimization

The problem the authors tackle is one of the thorniest in applied cryptography and computer vision at once. Content-based image retrieval, or CBIR, is the technology behind searching a collection by visual content rather than by file name or tags. It works by extracting a feature vector from each image, a long list of numbers that summarizes what the picture shows, and then comparing the query’s vector against those stored on the server. But if the server holds the feature vectors in plain form, it effectively holds a compressed description of every image, which is precisely what a privacy-conscious owner wants to hide. Encrypting the vectors, meanwhile, makes them meaningless to the very algorithms that need to measure similarity between them.

Researchers have spent the past decade building privacy-preserving content-based image retrieval, known as PPCBIR, to thread this needle. The field has produced schemes based on order-preserving encryption, randomized locality-sensitive hashing, additive secret sharing, and encrypted vision transformers, among other approaches. Yet the authors of the new study argue that existing designs suffer from a recurring set of weaknesses: an inherent trade-off between security and efficiency, suboptimal retrieval speed, and compromised retrieval accuracy. Protecting features more strongly tends to make searches slower; making searches faster tends to leak information or degrade the quality of the results. EMG-IR is presented as an attempt to break that triangle rather than merely shift along its edges.

The first pillar of the system is a feature extraction stage built on an improved ResNet50, a widely used deep convolutional neural network architecture. The team modified the network by adding a channel attention module, a mechanism that lets the model learn which feature channels, out of the hundreds produced in its deeper layers, matter most for distinguishing the content of an image. Channel attention effectively reweights the network’s internal representations, amplifying informative channels and suppressing noisy ones. The authors acknowledge that this addition increases the number of parameters compared with the original ResNet50, but they report that it delivers a significant improvement in accuracy, a trade they consider worthwhile for a retrieval system whose usefulness depends on finding the right images.

From that backbone, EMG-IR extracts multi-granularity features, meaning descriptors that capture visual information at several levels of detail, from fine local textures to coarse global composition. Multi-granularity representation matters because a single image often contains multiple objects and concepts at different scales; a street scene might include a car, a pedestrian, and an overall urban context, each relevant to a different query. The system then combines these features with a dual-channel hash encoding method to achieve efficient multi-label generation. Hashing, in this context, compresses high-dimensional feature vectors into short binary codes that can be compared with fast bitwise operations, and the dual-channel design lets the model encode complementary semantic information in parallel streams. The result is a compact, multi-label description of each image that supports both accurate matching and rapid search.

The second pillar is the encrypted index. The team designed a multi-label guided hierarchical searchable encrypted index structure, which organizes the encoded images in a tree-like hierarchy whose branches correspond to groups of labels. Instead of scanning every encrypted descriptor for every query, the server can descend the hierarchy, following the labels relevant to the query and pruning entire branches that cannot contain matches. Crucially, this structure is combined with a secure k-nearest neighbor algorithm built on the learning with errors problem, or LWE, a lattice-based hardness assumption widely regarded as resistant to quantum attacks and a cornerstone of modern post-quantum cryptography proposals. The LWE-based mechanism protects the confidentiality of the feature vectors, the index itself, and the query process, so the cloud server learns neither what the images depict nor what the user is looking for.

A third design choice addresses a practical pain point that static encrypted databases have long suffered from: the inability to add new data without rebuilding everything. EMG-IR adopts hierarchical indexing technology to enable dynamic updates of image data, allowing new images to be inserted and existing categories to expand without disrupting the overall structure. In real deployments, image collections grow constantly, and a retrieval scheme that requires re-encrypting and re-indexing the entire corpus after every batch of uploads would be unusable at scale. The hierarchical design gives the system flexibility to adapt to new images and evolving label sets, which the authors present as a key advantage for cloud environments where data is anything but static.

The team evaluated the scheme on two standard public benchmarks: NUS-WIDE, a real-world web image database from the National University of Singapore containing hundreds of thousands of images with multiple labels each, and MS-COCO 2017, Microsoft’s large-scale dataset of common objects in context, which is a standard proving ground for multi-label recognition. Both datasets are publicly available for scientific research, which means other groups can reproduce and stress-test the reported results. The experiments compared EMG-IR against existing comparison schemes on retrieval accuracy, retrieval efficiency, storage cost, and security properties.

The reported numbers are striking. According to the security analysis and experimental evaluation, EMG-IR improves retrieval accuracy by at least 3 percent compared with the comparison scheme, enhances retrieval efficiency by 34 percent, and reduces storage costs by at least 12 percent, while offering higher security. Each of those gains addresses one leg of the security-efficiency-accuracy trade-off that has constrained the field, and the authors argue that the combination, rather than any single metric, is what distinguishes the work. The accuracy gain traces to the attention-enhanced multi-granularity features and the multi-label hashing; the efficiency gain comes from the hierarchical index and compact binary codes; the storage reduction follows from the same compression; and the security improvement stems from the LWE-based encryption covering vectors, index, and queries alike.

The implications reach well beyond personal photo albums. Encrypted image search is a building block for privacy-preserving medical imaging, where hospitals outsource archives to cloud platforms but are bound by confidentiality obligations to patients; for biometric systems, where face or iris templates must never be exposed; for financial and smart healthcare infrastructures in the Internet of Things, where cameras generate sensitive imagery continuously; and for multi-user enterprise settings where different owners share a single cloud repository. Prior work in the field has explored verifiable retrieval, multi-key settings, search-pattern suppression, and traceable search, and EMG-IR adds to that growing toolkit a design that emphasizes dynamic updates and multi-label semantics. The research was supported by the National Natural Science Foundation of China, and the authors declare no competing interests.

Challenges remain before such systems become routine infrastructure. The channel attention module adds parameters and therefore computational cost at feature extraction time, which runs on the data owner’s side before anything is encrypted. Lattice-based secure computation, while theoretically robust, still carries overheads that scale with the dimensionality of the protected vectors, and the balance between code length and accuracy will always require tuning for a given application. Nonetheless, the study offers a concrete demonstration that the long-standing assumption of a rigid trade-off between confidentiality, speed, and accuracy in encrypted image search can be loosened. As cloud storage continues to absorb the world’s visual data, schemes like EMG-IR point toward a future where outsourcing images no longer means surrendering them, and where the server that stores your pictures can find the right one without ever knowing what it is looking at.

Subject of Research: Privacy-preserving content-based image retrieval over encrypted data in cloud computing

Article Title: Efficient multi-label generation-based image retrieval with privacy-preserving in cloud computing

Article References: Chen, Z., Zhang, Q., & Xu, R. (2026). Efficient multi-label generation-based image retrieval with privacy-preserving in cloud computing. Cluster Computing, 29(15), Article 835. https://doi.org/10.1007/s10586-026-06662-0

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06662-0

Keywords: image retrieval, privacy-preserving, cloud computing, searchable encryption, deep hashing, multi-label learning, ResNet50, LWE, k-nearest neighbor, hierarchical index, data confidentiality, hash coding

The core idea behind hashing-based image retrieval is to translate each image into a short string of binary digits, its hash code, so that similar images end up with similar codes. Once images are represented this way, comparing a query against a massive database reduces to counting differing bits, an operation that modern processors execute at extraordinary speed and that requires only a few dozen bits of storage per image instead of thousands of floating-point numbers. For single-label images, where each picture has one clear category, this translation is relatively straightforward. But most real-world photographs are messier: a single street scene might simultaneously depict a person, a car, a dog, a building, and an overcast sky. Multi-label retrieval demands that a compact code capture all of these meanings at once, a far harder compression problem.

The authors identify two persistent weaknesses in existing approaches. First, most supervised hashing methods rely on a semantic similarity matrix that simply declares whether two images are similar or dissimilar, based on how many labels they share. That binary or pairwise judgment, the researchers argue, throws away the richer guidance that semantic information could provide throughout the learning process. The semantic signal is reduced to a static target rather than an active participant in shaping how image features flow through the network. Second, conventional architectures are symmetric: they push both the incoming query images and the entire database of stored images through the same full deep network every time codes must be generated. Because the database side vastly outnumbers the query side, this symmetry wastes enormous computational effort and inflates the time cost of training and updating the retrieval index.

TCADH tackles both problems with a three-stream design. Two asymmetric image streams process, respectively, the original images and augmented versions of those same images, while a third stream carries the semantic label information itself. All three streams are mapped into a common discrete Hamming space, the mathematical arena where binary hash codes live and where distance is measured by counting bit differences. The word collaborative in the method’s name refers to this alignment: instead of letting labels sit passively in a similarity matrix, the framework actively pulls original images, their augmented counterparts, and their semantic labels into mutual agreement within the same binary space. The result is a code that reflects not just what an image looks like but what it means, from multiple complementary perspectives simultaneously.

The image-processing backbone of the framework is built on ResNet34, a widely used convolutional neural network architecture known for its balance of depth and efficiency. The authors construct a Siamese network from this backbone, meaning that two structurally identical copies of the network process the original and augmented images in parallel with shared parameters. A key innovation sits inside this backbone: a sign-consistency-guided feature enhancement block. Deep hashing requires converting continuous, real-valued features into binary codes, typically by taking the sign of each feature dimension, and this dimensionality reduction inevitably discards information. The enhancement block uses the sign of the features as a guide to reinforce the information that survives the compression step, mitigating the loss that would otherwise degrade the quality of the final hash codes.

Supervision in TCADH comes from a weighted pairwise loss applied across the three streams. The loss function is designed to maximize semantic consistency among different data streams that originate from the same image source. In practice, this means the hash code of an original image, the code of its augmented twin, and the representation derived from its semantic labels are all pushed toward agreement, with weights that let the model calibrate how strongly each pair of streams should align. Because the label stream provides fine-grained semantic supervision rather than a coarse similar-or-not verdict, the network receives a much denser training signal. Every label attached to an image contributes to shaping the code, which matters enormously in multi-label settings where a single image may carry half a dozen or more categories.

The asymmetric learning strategy is the framework’s answer to the computational bottleneck of symmetric hashing. Rather than forcing both queries and database images through identical encoding pipelines, TCADH learns to capture similarities directly between the discrete hash codes and the real-valued image features. This decoupling means the expensive deep feature extraction can be handled differently on each side of the retrieval task, improving retrieval precision while simultaneously accelerating the model’s convergence during training. The practical implication is significant for anyone operating large-scale image services: the database side of the system, which may contain millions or billions of entries, no longer pays the full price of symmetric deep encoding every time the index is refreshed.

The empirical evaluation spans four of the most demanding multi-label benchmarks in computer vision research: NUS-WIDE, MIRFLICKR-25K, MS-COCO, and VOC2012. These datasets collectively contain hundreds of thousands of images annotated with dozens of possible labels, from everyday objects to abstract scene attributes, and they have long served as the proving grounds for retrieval algorithms. Across all four, the authors report that TCADH outperformed state-of-the-art deep supervised hashing methods on multi-label retrieval tasks. While the abstract does not enumerate specific numerical margins, the consistency of the advantage across datasets with very different characteristics, from the social-media imagery of Flickr to the densely annotated everyday scenes of MS-COCO, suggests that the framework’s benefits are robust rather than tailored to one particular data distribution.

The significance of this work extends beyond a leaderboard improvement. As visual data continues to grow at a staggering pace, driven by smartphones, surveillance networks, e-commerce catalogs, and scientific imaging, the energy and latency costs of exhaustive similarity search become prohibitive. Hashing remains one of the few techniques that scales gracefully to web-scale corpora, and improvements in how faithfully binary codes preserve multi-label semantics translate directly into better user experiences and lower infrastructure costs. The asymmetric design philosophy championed by TCADH, in which the query side and the database side are treated as fundamentally different computational problems, is likely to influence how future retrieval systems are architected, particularly as researchers explore combinations of hashing with large vision-language models that could supply even richer semantic streams.

The study, received in April 2026 and accepted that September, arrives at a moment when the field of deep hashing is maturing from proof-of-concept demonstrations into engineering-grade systems. By weaving semantic supervision directly into the multimodal information flow, compressing images with sign-guided feature enhancement, and breaking the symmetry that has long constrained training efficiency, Wang and Cui have outlined a template for retrieval systems that are simultaneously more accurate and more economical. The article is available open access under a Creative Commons Attribution 4.0 license, allowing researchers and practitioners anywhere to examine the method in full technical detail and to build upon its three-stream, collaboratively aligned vision of what image search can become.

Subject of Research: Deep hashing for multi-label image retrieval

Article Title: Triple collaborative asymmetric deep hashing for multi-label image retrieval

Article References: Wang, W., & Cui, Z. (2026). Triple collaborative asymmetric deep hashing for multi-label image retrieval. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02533-8

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02533-8

Keywords: image retrieval, deep hashing, multi-label learning, computer vision, asymmetric learning, Hamming space, ResNet34, semantic supervision, NUS-WIDE, MS-COCO, binary codes, machine learning

The blind spot, which the authors call direction-blindness, arises from a property most people never think about when they compute a distance. Standard Euclidean distance is symmetric: it depends only on the absolute magnitude of the difference between two feature values, not on whether one value sits above or below the other. That means a feature whose informative signal for a label lies entirely on one side of its average value — say, a gene whose under-expression flags a disease, or a word whose absence marks a document topic — is, from the perspective of pairwise distances, indistinguishable from a feature whose signal is symmetrically distributed around the mean. The classifier sees proximity, but not direction, and can collapse two semantically very different patterns into a single similarity score.

Adaptive metric-learning techniques such as LMNN and NCA can address this problem, but they do so inside the classifier, downstream of any feature selection step. The Sassari team, Filippo Casu, Andrea Lagorio and Giuseppe A. Trunfio, took a different route: expose directional evidence at the representation level, before the classifier ever runs. Their approach is an embedded feature-selection method, meaning it learns which features matter while simultaneously fitting a predictive model, rather than scoring features one at a time with a filter statistic. Embedded methods are attractive because they account for interactions among features and labels, and because they return a ranking that can be plugged into any downstream algorithm.

The core idea of PSDFS is a mathematical trick the authors call a signed-deviation lift. Every original feature is split into two nonnegative channels: one measuring how far a value sits above the training-fold mean, and one measuring how far it sits below. This two-sided hinge-style expansion, which echoes classical constructions in multivariate adaptive regression splines and in convex nonnegative matrix factorization, has a powerful consequence. In a conventional nonnegative reconstruction model, the contribution of a feature is a monotonically increasing function of its value, so low feature values can never act as explicit evidence for a label. The lift breaks that monotonicity barrier while keeping every weight nonnegative — a constraint that preserves the stability of multiplicative updates and keeps feature contributions easy to interpret. The researchers prove formally that the class of functions representable after lifting strictly contains the class representable without it.

Splitting each feature into two channels, however, creates a new hazard. If a sparsity penalty acts independently on the lifted components, the optimizer may select only one half of a pair — an XOR-like artifact the authors quantify with a diagnostic called the SplitRatio. Under an unpaired penalty, this ratio often approaches one, meaning the reported selection fragments the identity of the original feature and produces rankings that are sensitive to an arbitrary decomposition. PSDFS solves this with a paired group-sparsity penalty that treats the above- and below-center channels of each feature as a single selection unit, ranking the original features directly by the combined norm of their two channels. The result is a single, coherent ranking in the original feature space, with no split artifacts by construction.

The optimization machinery is deliberately conservative. For fixed training-fold statistics, the objective — a Frobenius reconstruction loss plus the paired group penalty — is convex in the weight matrix, and the multiplicative updates descend monotonically with a sublinear O(1/T) bound on average suboptimality, properties the authors establish in a series of propositions with full proofs. Empirically, most of the objective decrease happens within roughly ten iterations, and a budget of forty updates proved safe across every benchmark tested. Additional propositions show that the optimum carries genuine directional semantics: an active above-center weight requires positive correlation between the above-center deviation and the label residual, and vice versa, so the relative magnitudes of the two channels reveal which side of the mean carries the signal.

The experimental evaluation is unusually thorough. The team compared PSDFS against seven recent embedded feature-selection methods — LRMFS, LSMFS, LRDG, GRRO, RFSFS, SRFS and SCNMF — on sixteen benchmark datasets drawn from the Mulan repository and the Yahoo topic collection, spanning text, image and bioinformatics domains with anywhere from 194 to nearly 14,000 instances and up to 374 labels. All methods ran on identical stratified folds with no per-dataset hyperparameter tuning, a discipline that guards against information leakage. Evaluated with ML-kNN as the downstream classifier, PSDFS achieved the best average rank on all seven reported metrics, including Micro-F1, Macro-F1, Hamming Loss, precision-recall areas and label-ranking measures. After Holm–Bonferroni correction, its improvements were statistically significant against every baseline on Micro-F1 and Macro-F1, and against all but one baseline — GRRO — on Hamming Loss.

Targeted ablations confirm that each design element earns its place. Removing the lift collapses performance back toward the monotonicity-limited baseline, with the lifted variant winning on twelve to fourteen of sixteen datasets across the core metrics. Replacing the nonnegative constraint with an unconstrained solver degrades results on nearly every metric, suggesting that nonnegativity regularizes the solution and prevents the two channels from canceling each other algebraically. A robustness check with a second fixed-metric classifier, binary relevance with an RBF-kernel logistic regression, preserved the ranking of methods, reducing the risk that the advantage is an artifact of ML-kNN’s neighborhood bias. The method is also fast: at an average of 1.37 seconds per fold, it outruns most of its competitors, only trailing the two baselines that solve simpler linear systems.

Perhaps the most practically valuable analysis concerns low-resource regimes. When the training split is subsampled to as little as five percent, PSDFS degrades gracefully and is never overtaken by any baseline, though its separation from the strongest reconstruction-based competitors narrows to a statistical tie. The authors identify the mechanism precisely: the lift estimates twice as many channel weights from disjoint subsets of the data, so its contribution is the first casualty of scarce labels. They offer concrete deployment guidance — use the method as-is when the sample size exceeds roughly twice the lifted dimensionality, and treat fine-grained directional annotations with caution below that threshold. The authors also note two honest limitations: the model remains additive and does not capture higher-order feature interactions, and it has not been validated inside end-to-end deep visual pipelines. A reference implementation is publicly available on GitHub, and given how widely fixed-metric distance classifiers are deployed, the idea of making feature selection direction-aware seems likely to travel far beyond the sixteen benchmarks where it was born.

Subject of Research: Direction-aware feature selection for distance-based multi-label classification

Article Title: Direction-aware multi-label feature selection via paired signed-deviation lifting

Article References: Direction-aware multi-label feature selection via paired signed-deviation lifting. (n.d.). https://doi.org/10.1007/s10618-026-01265-0

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01265-0

Keywords: multi-label learning, feature selection, machine learning, distance-based classifiers, nonnegative reconstruction, group sparsity, ML-kNN, signed-deviation lift, metric learning, data mining, benchmark evaluation, interpretable AI