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

Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification · 4 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.

Decomposes the deep imbalanced regression objective into group classification plus per-group regression, and proposes an ordinal group-aware contrastive loss, a multi-experts regressor, and symmetric descending soft labels for group classification.

Problem

Deep imbalanced regression (DIR), where continuous targets have a highly skewed training distribution, leading to poor generalization on balanced test sets; prior classification-based regularizers ignore label continuity and similarity across data.

Approach

Bayesian decomposition of the DIR objective into group classification and within-group regression (divide-and-conquer), with an ordinal group-aware contrastive loss, a multi-expert regressor routed by predicted group, and symmetric descending soft labels that encode similarity between neighboring groups.

Key results

• On AgeDB-DIR, the method reaches MAE 6.87 (All), the best overall MAE among compared methods, and GM 4.30.
• On IMDB-WIKI-DIR, MAE 7.22 overall with best Median (11.42) and Few-shot (20.25) MAE, while VIR is slightly better on overall MAE (7.19).
• On STS-B-DIR, MSE 0.887 and Pearson correlation 77.4, with best Median and Few-shot results, while VIR has higher Pearson on Many-shot and All splits.

Measured effects reported in the paper

Surprising findings

• Standard cross-entropy and logit adjustment classifiers over-estimate or fail on minority groups when used for group prediction in DIR, motivating the soft labels.
• Group prediction accuracy drops as the number of groups increases, attributed to data dependence across neighboring groups.

Limitations

• Group prediction accuracy degrades as the number of groups grows due to data dependence across neighboring groups.
• Method underperforms VIR on several Many-shot and overall metrics (e.g., IMDB-WIKI-DIR overall MAE, AgeDB-DIR Many-shot MAE), and the Few-shot AgeDB-DIR MAE trails VIR.

Experimental setup

Datasets: IMDB-WIKI-DIR AgeDB-DIR STS-B-DIR

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

Metrics: MAE GM b-MAE MSE Pearson correlation group classification accuracy

Baselines: SMOTER SMOGN RRT FOCAL-R SQINV LDS+FDS VAE DER Con-R RankSim VIR BMC

Builds on: Balanced MSE Rank-based similarity regularization (RankSim) VIR (variational imbalanced regression) Logit adjustment Label distribution smoothing

Closest papers by meaning

• ConR: Contrastive Regularizer for Deep Imbalanced Regression (ICLR 2024 · similarity 0.89)
• Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data (ICML 2025 · similarity 0.86)
• Improve Representation for Imbalanced Regression through Geometric Constraints (CVPR 2025 · similarity 0.86)
• PRIME: Deep Imbalanced Regression with Proxies (ICML 2025 · similarity 0.85)
• Deep Imbalanced Regression via Hierarchical Classification Adjustment (CVPR 2024 · similarity 0.85)
• The Group Robustness is in the Details: Revisiting Finetuning under Spurious Correlations (NeurIPS 2024 · similarity 0.84)