Research map / Label-efficient and robust learning / Recognition and long-tail

ACCon: Angle-Compensated Contrastive Regularizer for Deep Regression

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ACCon, an angle-compensated supervised contrastive regularizer that maps label distances to target angles on a semi-hypersphere so feature similarity decreases linearly with label distance, improving deep regression on age estimation and STS-B.

Problem

Deep regression needs representations that reflect the relationships (order and distance) among continuous labels; existing order-aware and distance-weighted contrastive methods do not capture this well, especially for imbalanced data and small sample sizes.

Approach

Hypothesizes a linear negative correlation between label distance and representation similarity, and derives an angle-compensated negative cosine term so that negatives are pushed to label-dependent angles on a semi-hypersphere; the loss is plug-and-play into SupCon-style frameworks.

Key results

• On AgeDB-Natural, ACCon reaches MAE 6.724 vs 7.044 for vanilla (4.54% improvement) and 1.02% better than the best prior method.
• On STS-B-Natural, ACCon achieves Pearson 0.791 vs 0.756 for vanilla, and MAE improves 9.27% over vanilla.
• On AgeDB-DIR, ACCon reduces MAE from 7.77 (vanilla) to 6.81, and ACCon+ConR reaches 6.82; the learned cosine similarity shows Pearson -0.735 with normalized label distance.

Measured effects reported in the paper

Surprising findings

• ACCon underperforms most baselines on STS-B-DIR relative to RankSim-based variants, which the authors attribute to limited text data and constrained augmentation.
• Learned cosine similarities mostly fall in [0.3, 1] rather than the ideal [-1, 1] range, attributed to label imbalance.

Limitations

• Label-imbalance causes cosine similarities to concentrate in a narrow range rather than spanning the full [-1, 1] interval.
• Underperforms on STS-B-DIR, attributed to limited data and restricted augmentation for text.
• Few-shot region shows increased variance in cosine similarity as label distance grows.

Experimental setup

Datasets: AgeDB IMDB-WIKI STS-B

Benchmarks: AgeDB-Natural STS-B-Natural IMDB-WIKI-Natural AgeDB-DIR IMDB-WIKI-DIR STS-B-DIR

Metrics: MAE G-means R2 Pearson MSE many/medium/few-shot MAE

Baselines: Vanilla NaiveSupCon AdaSupCon BMSE RankSim RNC ConR LDS Focal-R SqrtINV

Builds on: supervised contrastive learning ConR RankSim Balanced MSE Label Distribution Smoothing

Closest papers by meaning

• ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.91)
• Rank-N-Contrast: Learning Continuous Representations for Regression (NeurIPS 2023 · similarity 0.86)
• Semi-Supervised Contrastive Learning for Deep Regression with Ordinal Rankings from Spectral Seriation (NeurIPS 2023 · similarity 0.86)
• Gradient Aligned Regression via Pairwise Losses (ICML 2025 · similarity 0.84)
• Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.84)
• Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression (AAAI 2025 · similarity 0.83)