Research map / Continual learning / Class-incremental
EmoGrowth: Incremental Multi-label Emotion Decoding with Augmented Emotional Relation Graph
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 AESL, a multi-label class-incremental emotion decoding framework using an augmented emotional relation graph with graph-based label disambiguation, GIN-based emotion embeddings, and relation-based distillation from affective dimensions (arousal/valence).
Problem
Multi-label class-incremental learning for fine-grained emotion decoding, where new emotion categories arrive over time and emotions co-occur, causing past-missing and future-missing partial label problems and catastrophic forgetting.
Approach
Augments the emotional relation graph incrementally using graph-based label disambiguation to produce soft labels for old classes, learns label embeddings with a graph autoencoder to guide semantic-specific feature decoupling, and uses relation-based knowledge distillation aligning features with an affective dimension space (arousal/valence) to address future-missing labels.
Key results
- On Brain27, AESL gives a 9.6% relative mAP improvement (4 protocols averaged) over the second-best method.
- On Audio28 B0-I7, AESL reaches 49.0 average mAP versus 47.3 for AGCN (the strongest baseline).
- Removing the relation-based distillation reduces last mAP by 0.6 on Audio28 B0-I7 (42.7 to 42.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AESL instead of KRT-R | incremental emotion decoding · Brain27 B0-I9 | Avg. Acc mAP: 42.5 -> 44.2 | Helped |
| Use AESL instead of KRT-R | incremental emotion decoding · Brain27 B0-I3 | Avg. Acc mAP: 44.3 -> 43.8 | Hurt |
| Use AESL instead of OCDM | incremental emotion decoding · Video27 B0-I9 | Avg. Acc mAP: 43.1 -> 44.6 | Helped |
| Use AESL instead of AGCN | incremental emotion decoding · Audio28 B16-I3 | Avg. Acc mAP: 44.5 -> 47.8 | Helped |
| Add emotional semantics learning and label disambiguation (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 47.6 -> 49.0 | Helped |
| Replace learned emotion embeddings with sentence embeddings from LLaMA 3.1-8B (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 49.0 -> 45.3 | Hurt |
| Replace learned emotion embeddings with learnable embeddings (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 49.0 -> 46.8 | Hurt |
| Directly integrate sample-wise affective dimensions into category-wise emotion embeddings (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 49.0 -> 48.4 | Hurt |
| Add label disambiguation when constructing the adjacency matrix (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 48.1 -> 49.0 | Helped |
| Add relation-based knowledge distillation (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 48.3 -> 49.0 | Helped |
| Replace relation-based knowledge distillation with a linear regressor for affective dimensions (ablation) | AESL on Audio28 B0-I7 · Audio28 B0-I7 | Avg. Acc mAP: 49.0 -> 47.0 | Hurt |
Surprising findings
- Rehearsal-based methods (ER, RS, PRS, OCDM, KRT-R) perform poorly on maF1 and miF1, which the authors attribute to replayed labels aggravating partial-label problems.
Limitations
- Relies on affective dimension ratings (arousal/valence) whose noise can degrade performance when the distillation weight is large.
- Rehearsal-free design; no storage of historical instances, so performance gaps on some metrics remain versus oracle upper bound.
Experimental setup
Datasets: Brain27 Video27 Audio28
Benchmarks: Brain27 Video27 Audio28
Metrics: mAP maF1 miF1 average accuracy last accuracy
Baselines: Finetune EWC LwF ER RS AGCN PRS OCDM KRT-R
Builds on: AGCN Graph Isomorphism Network Relation-based knowledge distillation Graph-based label propagation KRT-R
Closest papers by meaning
- Beyond Single Emotion: Multi-label Approach to Conversational Emotion Recognition (AAAI 2025 · similarity 0.85)
- EEmo-Logic: A Unified Dataset and Multi-Stage Framework for Comprehensive Image-Evoked Emotion Assessment (ICML 2026 · similarity 0.85)
- Emotion-LLaMA: Multimodal Emotion Recognition and Reasoning with Instruction Tuning (NeurIPS 2024 · similarity 0.84)
- OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition (ICML 2025 · similarity 0.84)
- A Brain-Inspired Way of Reducing the Network Complexity via Concept-Regularized Coding for Emotion Recognition (AAAI 2024 · similarity 0.84)
- Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis (AAAI 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.