pairwise ranking

This paper investigates whether binary comparisons can outperform ordinal comparisons in ranking recovery. For tie-free ordinal data, we propose a generalized additive parametric model that jointly characterizes the link function for preference differences and the mode function for the response distribution. Theoretically, we show that when the signal-to-noise ratio is dominated by the mode function, binarizing ordinal data significantly improves ranking accuracy; moreover, ranking error under binary feedback exhibits faster exponential convergence—challenging the conventional belief that ordinal information is inherently richer. These findings are empirically validated on both synthetic and real-world MovieLens data. We further identify sufficient conditions under which binarization maximizes ranking gains, thereby establishing a new paradigm for feedback mechanism design in preference learning.

This work addresses the limitations of existing contrastive learning methods, which overlook ordinal relationships among labels, and conventional ordinal learning approaches, which struggle to capture global ordinal structures. To bridge this gap, the authors propose ConOrd, a novel contrastive ordinal learning framework that seamlessly integrates contrastive learning with ordinal regression. ConOrd introduces rank-difference-based soft affinities and disparity weights to enable fine-grained modeling of ordinal relationships across all sample pairs within a batch, thereby effectively capturing global ordinal structure. The method achieves state-of-the-art performance on diverse tasks including facial age estimation and blind image and video quality assessment, demonstrating strong generalization capability.

This work reframes the evaluation of explanation quality as a learning-to-rank problem, moving beyond conventional approaches that rely on generating a single optimal explanation or pointwise regression, which struggle to distinguish among explanations of varying quality levels. The study introduces listwise ranking methods—specifically ListNet, LambdaRank, and RankNet—to train a reward model capable of performing relative quality assessment over multiple candidate explanations while preserving their ordinal structure. Experimental results demonstrate that ranking-based losses consistently outperform regression-based counterparts across all domains. Furthermore, policy optimization using ranking-derived rewards achieves stable convergence, whereas regression-based rewards fail entirely. The findings also highlight that data quality exerts a more decisive influence than model scale, enabling smaller models to match the performance of significantly larger ones.

Existing ordinal regression methods typically employ standard cross-entropy loss, disregarding the inherent ordinal structure among class labels; moreover, prevailing unimodality modeling strategies lack theoretical foundations and rely on heuristic designs. Method: This paper systematically characterizes the geometric structure of unimodal distributions within the probability simplex for the first time, proposes a theory-driven loss term based on set projection, and designs an end-to-end unimodal neural network architecture. The method strictly constrains predicted distributions to the unimodal set via differentiable projection, enabling efficient optimization. Contribution/Results: On multiple benchmark datasets, the proposed architecture achieves top-2 performance. The new loss significantly outperforms state-of-the-art baselines while maintaining high unimodality—establishing the first theoretically grounded unimodal modeling paradigm for ordinal regression.

Existing Bradley–Terry–based alignment methods (e.g., DPO, RLHF) rely on pairwise comparisons, which inadequately capture the full ranking structure among multiple responses and thus limit alignment with diverse human preferences. This work introduces, for the first time, the information retrieval metric Normalized Discounted Cumulative Gain (NDCG) into LLM alignment, proposing a differentiable ordinal preference optimization framework. We construct a surrogate loss via a differentiable NDCG approximation to enable end-to-end multi-response ranking optimization. Furthermore, we integrate ordinal preference modeling with a negative-sample pool expansion strategy to mitigate interference from trivial negatives. Extensive evaluation on benchmarks—including AlpacaEval—demonstrates significant improvements over DPO and RLHF, validating the method’s effectiveness in enhancing both response quality and fidelity to human ordinal preferences.

This work addresses the problem of robust rank estimation under structured ordinal label noise by proposing a stochastic ranking learning framework that explicitly models the uncertainty of ordinal labels as probabilistic ranking relations, thereby moving beyond the conventional assumption of deterministic labels. The method jointly optimizes a discriminative loss and a stochastic ranking loss, integrating discriminative instance-to-centroid interactions with probabilistic ranking constraints in the embedding space to effectively capture the distribution over multiple plausible rankings. Experimental results demonstrate that the proposed framework significantly enhances both the robustness and accuracy of rank estimation across diverse datasets and noise configurations.

Existing representation similarity measures often suffer from poor interpretability, sensitivity to outliers, and limited scalability to large datasets. This work proposes a representation alignment framework grounded in ordinal similarity, which quantifies the consistency of local ordinal relationships between representations using triplet (TSI) and quadruplet (QSI) similarity indices. The study establishes, for the first time, a theoretical equivalence between ordinal similarity and local alignment criteria such as mutual nearest neighbors. Combining interpretability, robustness, and scalability, the proposed method leverages efficient ranking algorithms and demonstrates significant performance gains over existing metrics in large-scale experiments, offering a reliable tool for representation understanding and design.

This work addresses the challenges in ordinal regression posed by ambiguous label boundaries and noise arising from subjective annotations, as well as the bias amplification inherent in conventional methods that rely on fixed supervision targets. To overcome these limitations, the authors propose the D3O framework, which replaces static supervision with dynamic distribution distillation. D3O leverages a self-distillation mechanism to evolve ordinal label distributions and incorporates a contrastive ordinal-aware label augmentation module alongside a cumulative distribution function (CDF)-based cross-layer interactive distillation strategy. This approach effectively models inter-class ambiguity and instance-level uncertainty while integrating vision-language alignment to enrich label representation. Evaluated across four ordinal regression benchmarks, D3O consistently outperforms state-of-the-art methods, demonstrating exceptional robustness under severe class imbalance and highly noisy supervision.

This work addresses the performance degradation in ordinal classification caused by existing methods' neglect of the natural order among classes. To this end, we propose ADABORD, a novel framework that, for the first time, integrates both an ordinal splitting criterion and an error function accounting for inter-class distances within AdaBoost. Specifically, ADABORD employs decision stumps based on an ordinal Gini impurity measure as base learners and introduces an absolute ranking probability score to more appropriately update sample and model weights. Experimental results on the TOC-UCO benchmark—the largest evaluation suite for ordinal classification—demonstrate that ADABORD significantly outperforms seven state-of-the-art methods, with particularly pronounced gains on datasets containing five or more ordinal classes.

This work addresses the limitation of conventional loss functions—such as cross-entropy—in neglecting structural relationships among classes, which hinders their ability to handle structured label noise or incorporate prior knowledge. The authors propose Conveyance, a novel framework that introduces, for the first time, a unified loss function capable of jointly addressing tasks with structured label spaces, including hierarchical classification, ordinal regression, and multiple-instance learning. By modeling class relationships through a graph structure, the method avoids the need for complex joint distributions or manually designed utility matrices. It further incorporates a double-margin maximization mechanism to optimize decision margins across varying class partitions. The proposed loss enjoys favorable theoretical properties, such as monotonicity and partial convexity, and achieves performance on par with or superior to task-specific methods across multiple benchmark datasets, demonstrating its generality and effectiveness.