The research, led by Vu-Linh Nguyen of the Université de technologie de Compiègne in France and Xuan-Truong Hoang of the Japan Advanced Institute of Science and Technology, together with Sébastien Destercke, Cassio de Campos and Van-Nam Huynh, recasts multi-label classification as a problem of learning order structures. Instead of treating each label as an independent yes-or-no question, the authors model the relationships between labels as preferences: one label may be strictly preferred to another, the two may be considered equivalent, or they may be incomparable. The result is a richer representation that, the authors argue, can be exploited to make multi-label learning both more accurate and more robust.

The core idea is a formal transformation. For any pair of labels, the possible joint outcomes form four categories: the first label is relevant and the second is not, the reverse, both are relevant, or neither is. A standard multi-label classifier predicts a binary vector, but the new framework maps that vector onto a relation between the two labels. A label vector in which one label is active and the other inactive becomes a strict preference between them; two active labels become an equivalence; and two inactive labels become another form of indifference or incomparability, depending on the chosen structure. Learning to classify instances then becomes learning to predict the most probable order structure, a problem the authors call Bayes-optimal order structure prediction.

This reframing is not merely cosmetic. Because order structures obey transitivity, they carry constraints that ordinary independent binary predictions lack. The authors show that finding the most probable preorder, an order structure that allows both strict preferences and equivalences, can be formulated as an integer linear programming problem. The transitivity conditions translate naturally into linear constraints, and the logarithm of the probability of an order structure decomposes into a sum over label pairs, making the optimization tractable with standard solvers. The framework also accommodates partial orders, which are more parsimonious but, as the authors demonstrate through careful examples, can lose crucial information by lumping together the cases where both labels are relevant and where neither is.

The distinction matters in high-stakes applications. The authors illustrate the point with a medical scenario involving two diseases encoded as two labels. A patient whose predicted joint outcome places both diseases as almost certainly present should clearly be sent for further diagnosis, while a patient in whom both are almost certainly absent should not. A framework based on partial orders treats these two situations as the same relation and would issue identical recommendations for both patients. The preorder representation, by contrast, distinguishes them, preserving exactly the kind of information that separates a useful clinical decision support system from a dangerous one.

The new approach also exposes weaknesses in an established preference-learning method for multi-label classification known as calibrated label ranking. That earlier technique ranks labels by pairwise comparison and uses a threshold, estimated from the expected number of irrelevant labels, to separate relevant from irrelevant ones. Through worked examples, the authors show that calibrated label ranking can misclassify the most probable preference relation in cases where the two labels are most likely equivalent, because its scoring scheme collapses the four joint outcome probabilities into a single pairwise score. The richer representation, which keeps all four probabilities, avoids these failures by construction.

On the theoretical side, the paper establishes which order structures are Bayes-optimal under different evaluation metrics. Under subset accuracy, the optimal prediction is simply the most probable order structure. Under Hamming-style accuracies, which score each pairwise relation independently, the optimal structure can be obtained pair by pair, a considerable computational simplification. The authors further extend their results to cost-sensitive variants, where different errors carry different penalties, and sketch how confusion-matrix-derived metrics such as the F-measure and the Jaccard index can be generalized to the order-structure setting, while noting that efficient optimization for those generalized metrics remains an open problem.

The empirical study covers ten publicly available multi-label datasets, with random forests and LightGBM as base learners to estimate the pairwise probabilities. The experiments focus on two regimes where order structures should shine. The first is robustness: the researchers artificially corrupted training labels at varying noise levels and with imbalanced classes, then measured how well each method predicted the true binary vectors. The second is partial abstention, where a classifier is allowed to withhold predictions on individual labels it is unsure about rather than guessing. In both regimes, the preorder-based classifiers consistently outperformed established baselines including binary relevance, classifier chains, calibrated label ranking and ensembles of classifier chains, particularly on the F-measure, a balanced metric that rewards both precision and recall.

The abstention results are especially striking. Allowing predictions with abstention led to impressive improvements in F-measure, and the preorder-based classifiers achieved competitive gains with significantly lower abstention rates than their partial-order counterparts. In other words, the richer structure not only made better predictions when it did predict, but also needed to hedge less often. Runtime experiments on the Enron email dataset, which has 53 labels, showed that the integer linear programming approach exhibits polynomial-like scaling in the number of labels, with per-instance prediction times of at most a few seconds when implemented with open-source solvers, a practical footprint attributable in part to the sparse constraint matrices of the underlying optimization problems.

The work, supported by funding from the French National Research Agency, the Office of Naval Research Global, the European Defence Fund and the Dutch Research Council, points toward a broader lesson for machine learning: sometimes the path to robustness lies not in bigger models or more data, but in choosing a representation whose structure matches the problem. By treating labels not as isolated switches but as elements of an ordered system, the researchers have shown that the rich mathematics of preferences and orders can absorb noise, tolerate imbalance and signal uncertainty in ways that conventional multi-label pipelines cannot. As datasets grow noisier and applications demand more honest predictions, that shift in perspective may prove increasingly hard to ignore.

Subject of Research: Transforming multi-label classification into order structure learning via preference learning to improve robustness against noisy and imbalanced labels and to enable partial abstention

Article Title: Robust Multi-label Classification via Preference Learning

Article References: Nguyen, V.-L., Hoang, X.-T., Destercke, S., de Campos, C., & Huynh, V.-N. (2026). Robust Multi-label Classification via Preference Learning. Machine Learning, 115(9), Article 211. https://doi.org/10.1007/s10994-026-07147-2

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07147-2

Keywords: multi-label classification, preference learning, order structures, robustness, noisy labels, imbalanced labels, partial abstention, Bayes-optimal prediction, integer linear programming, calibrated label ranking, machine learning, label ranking

The technical core of the new system rests on transfer learning from molecular pretraining. The authors adopted the two-dimensional graph encoder from GraphMVP, a framework whose parameters were previously learned through three-dimensional-informed pretraining, in which a model is trained to align what it sees in a molecule’s 2D graph with the geometric realities of its 3D conformations. Importantly, once this knowledge is distilled into the encoder, the downstream flavor-prediction task requires no explicit 3D coordinates at all. No conformers are generated during fine-tuning or inference. The researchers are careful to frame this as 3D-supervised pretraining rather than direct 2D/3D geometric integration at prediction time, a distinction that matters for both computational cost and reproducibility, since generating reliable conformers for diverse molecules is itself a difficult and error-prone step. The practical upshot is a model that carries the benefit of geometric understanding while remaining fast enough to screen large chemical libraries.

The second major contribution is an algorithmic solution to a problem that plagues nearly every flavor dataset: class imbalance. Sensory databases are lopsided by nature. Thousands of molecules have been cataloged as sweet or bitter, but far fewer carry rare labels, and a model trained naively on such data becomes excellent at the common flavors and nearly blind to the rare ones. The team proposes ML-ROS-improved, an adaptive oversampling algorithm built from three interacting components. First, dynamic thresholding identifies which labels qualify as minority classes at each stage of training. Second, weighted minority-label sampling deliberately boosts the frequency with which those rare-labeled molecules appear during learning. Third, constrained graph augmentation creates new training variants of minority molecules, but the augmentation is not arbitrary. The team systematically compared several graph augmentation strategies, and the most successful one constrains modifications using the Molecular Connectivity Index, a topological descriptor that quantifies branching patterns in a molecular skeleton. By enforcing MCI constraints, the algorithm ensures that synthetic augmentations remain chemically plausible and topologically faithful to the parent structure, avoiding the distribution drift that can poison oversampling methods.

The benchmark results are striking under the reported experimental settings. Across the FlavorMiner and FART benchmarks, FlavorGPN achieved the highest observed Macro-F1 and Macro AUC-ROC point estimates among all evaluated baselines. On the FART benchmark in particular, the model recorded a Macro-F1 score of 0.8542 and a Macro AUC-ROC of 0.9796. These are meaningful metrics for multi-label problems. Macro-F1 averages the harmonic mean of precision and recall across all flavor labels, giving rare categories the same weight as abundant ones, which means the oversampling strategy is being tested where it matters. Macro AUC-ROC, meanwhile, measures how well the model ranks molecules above non-molecules for each label regardless of the decision threshold. A value approaching 0.98 indicates near-perfect separation between positive and negative examples for the average flavor category. The authors also report literature-based contextual comparisons suggesting the unified model performs competitively on key flavor categories including Sweet, Bitter, and Sour, although they are explicit that such comparisons do not constitute controlled head-to-head evaluations.

The significance of this work becomes clearer when placed in the context of how flavor has historically been studied. Human taste perception involves an interplay of taste receptors, olfactory signals, and cognitive processing, and mapping that perceptual space onto chemical structures has resisted simple rules for decades. The human bitter taste receptor family TAS2R, for example, exhibits broad and overlapping receptive ranges that defy straightforward structure-activity mappings. Earlier computational efforts, from fingerprint-based classifiers for bitterants and sweeteners to BERT-style models for bitter peptides and language models for general taste prediction, have each advanced the field but typically handled a narrow slice of the problem, such as a single flavor category or a binary sweet-versus-bitter decision. FlavorGPN’s contribution is unification. A single model, trained once, outputs probabilities for the full multi-label flavor profile of any query molecule, which is exactly the format a food chemist or flavorist needs when deciding whether a candidate compound belongs in a formulation.

Potential applications extend well beyond academic curiosity. The food and beverage industry spends enormous resources on sensory evaluation, employing trained human panels whose work is slow, expensive, and inherently variable across individuals and sessions. A reliable computational pre-screening step could prioritize which candidates from a virtual library deserve actual sensory testing, dramatically shrinking development timelines for sweeteners, bitterness modifiers, flavor enhancers, and aroma compounds. Similar logic applies in agriculture and plant breeding, where molecular markers of desirable flavor have guided crop improvement, and in safety contexts, where predicting that a food-grade additive carries a bitter or off-flavor label can flag problems before they reach consumers. The authors also note the model’s applicability domain analysis, included in the released code, which helps users understand when predictions can be trusted and when a query molecule strays too far from the training distribution, a critical safeguard in any deployed machine learning system.

The broader scientific lineage of this approach traces back to the message-passing paradigm in neural networks, where information flows along molecular graph edges through successive layers, allowing each atom’s representation to incorporate its chemical neighborhood. Graph neural networks have already proven themselves in drug discovery, predicting drug-target binding affinities and molecular properties with hierarchical and attention-equipped variants, and the transfer of those tools to chemosensory prediction is a natural and overdue extension. What FlavorGPN adds is the combination of pretrained geometric awareness with a carefully engineered answer to multi-label imbalance, two ingredients that previous flavor-prediction efforts had addressed only separately. The MCI-constrained augmentation result in particular suggests a general lesson for molecular machine learning: when augmenting graph data, respecting topological invariants can be the difference between helpful synthetic data and harmful noise.

The open science posture of the study strengthens its potential impact. The datasets used are publicly available through the MoleculeNet repository, a widely adopted benchmark collection for molecular machine learning, and the complete source code, including model implementation, training scripts, evaluation scripts, data partitioning, and applicability domain analysis, has been released on GitHub. This means other research groups can immediately benchmark alternative architectures against FlavorGPN, adapt the ML-ROS-improved oversampling algorithm to other imbalanced multi-label domains such as drug side-effect prediction or protein localization, or extend the framework with additional pretraining schemes. Reproducibility has been a persistent weakness in applied deep learning, and the full availability of code and data here lowers the barrier considerably.

There remain, as the authors themselves acknowledge, important caveats. Benchmark performance measured under specific experimental settings does not guarantee superiority in every deployment scenario, and the literature-based comparisons with specialized single-category models were not controlled experiments. Sensory labels themselves are human constructs, subject to panel composition, cultural context, and concentration effects, and any model trained on such labels inherits those limitations. VirtualTaste and other web servers in this space have made similar predictions available for years, and the community will need prospective validation studies, in which the model’s predictions on newly synthesized or newly tested compounds are checked against fresh sensory data, before FlavorGPN’s numbers translate into industrial trust. Time-split validation, which mimics the realistic scenario of predicting compounds discovered after model training, remains a best practice in this field, and future work will likely need to demonstrate robustness under such conditions.

Even with those caveats, the arrival of a unified, high-performing, imbalanced-aware graph neural network for flavor prediction marks a notable step in the computational understanding of chemosensation. It demonstrates that representations learned with three-dimensional supervision can be transferred efficiently to a perceptual task without ever touching a conformer, and that principled, structure-aware data augmentation can unlock rare classes that conventional pipelines miss. As the food industry confronts demands for healthier formulations, sustainable ingredients, and plant-derived flavor systems, tools that can predict how a molecule will taste before anyone synthesizes or tastes it will only grow in value. FlavorGPN offers a template for how deep learning, pretraining science, and careful data engineering can converge on one of the senses we most take for granted.