Research map / Continual learning / Online and replay-based
Influential Exemplar Replay for Incremental Learning in Recommender Systems
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 INFER, a model-agnostic incremental recommender framework that replays influence-selected exemplars, with a one-step influence approximation (INFERONCE) that avoids inverse-Hessian computation.
Problem
Recommender models degrade when trained once on static data and fail to adapt to evolving user interactions; full retraining is costly, and incremental learning for recommendation (unlike vision) must handle user-item interaction streams with catastrophic forgetting.
Approach
Select replay exemplars by their influence on model-level recommendation loss (via influence functions), rather than data-level properties like frequency; the one-step approximation drops the inverse Hessian, reducing the score to a gradient dot product between candidate and training-data gradients.
Key results
- INFERONCE achieves the best AUC on all four datasets for DeepFM/TwoTower CTR prediction in most settings, e.g., 0.8486 on Gowalla with DeepFM versus 0.8403 for Fine-Tune.
- On Top-N recommendation with LightGCN, INFERONCE reaches Recall@20 of 0.1249 on Foursquare, above LWC-KD at 0.1194.
- Applying INFERONCE replay to GraphSAIL, SGCT, and LWC-KD improves their Recall@20 on all four datasets (e.g., SGCT on Foursquare 0.1157 to 0.1284).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use INFERONCE rather than Fine-Tune for CTR prediction | DeepFM and TwoTower · Lastfm-2k, Taobao2014, Gowalla, Foursquare | AUC | Helped |
| use INFERONCE rather than other incremental methods for CTR prediction | DeepFM and TwoTower · Lastfm-2k, Taobao2014, Gowalla, Foursquare | AUC | Mixed |
| use INFERONCE rather than competing methods for Top-N recommendation | BPR-MF and LightGCN · Lastfm-2k, Taobao2014, Gowalla, Foursquare | Recall@20 | Helped |
| replace vanilla influence-based replay with one-step influence estimation (ablation) | INFER · Gowalla and Foursquare | Recall@20: 0.0591 -> 0.0621; 0.0847 -> 0.0876; 0.0913 -> 0.0944; 0.1163 -> 0.1249 | Helped |
| replace vanilla influence-based replay with one-step influence estimation (ablation) | INFER · Lastfm-2k and Taobao2014 | Recall@20: 0.0513 -> 0.0508; 0.0695 -> 0.0704; 0.0074 -> 0.0072; 0.0083 -> 0.0086 | Mixed |
| replace vanilla influence-based replay with one-step influence estimation (ablation) | INFER · CTR prediction, all four datasets | AUC: 0.6863 -> 0.6884; 0.7099 -> 0.7134; 0.6338 -> 0.6370; 0.6528 -> 0.6551; 0.8432 -> 0.8486; 0.8052 -> 0.8237; 0.9547 -> 0. | Mixed |
| increase the replay ratio (ablation) | INFERONCE with TwoTower · Gowalla | AUC: 0.8247 to 0.8127 | Mixed |
| increase the replay ratio from 0.05 toward 0.25 (ablation) | INFERONCE with TwoTower · Gowalla | AUC | Helped |
| increase the replay ratio from 0.05 toward 0.30 (ablation) | INFERONCE with LightGCN · Gowalla | Recall@20 | Helped |
| add INFERONCE replay to GraphSAIL | GraphSAIL · Lastfm-2k, Taobao2014, Gowalla, Foursquare | Recall@20: 0.0553 -> 0.0708; 0.0078 -> 0.0082; 0.0725 -> 0.0815; 0.1042 -> 0.1195 | Helped |
| add INFERONCE replay to SGCT | SGCT · Lastfm-2k, Taobao2014, Gowalla, Foursquare | Recall@20: 0.0620 -> 0.0722; 0.0080 -> 0.0088; 0.0783 -> 0.0881; 0.1157 -> 0.1284 | Helped |
| add INFERONCE replay to LWC-KD | LWC-KD · Lastfm-2k, Taobao2014, Gowalla, Foursquare | Recall@20: 0.0674 -> 0.0721; 0.0081 -> 0.0088; 0.0792 -> 0.0902; 0.1194 -> 0.1292 | Helped |
Surprising findings
- The vanilla influence-based method (INFER-Vanilla) underperforms on larger datasets (Gowalla, Foursquare) for Top-N recommendation, attributed to sensitivity of the inverse-Hessian influence to model scale.
- Performance drops when the replay ratio is pushed beyond about 0.25 (TwoTower) or 0.30 (LightGCN), attributed to historical data introducing bias under preference drift.
Limitations
- Full-data retraining still beats INFERONCE on Top-N Recall@20, leaving a gap.
- Replay performance is sensitive to the reservoir size; over-large reservoirs degrade results.
- Approximates historical data loss with the reservoir loss, presuming close correlation between them.
Experimental setup
Datasets: Lastfm-2k Taobao2014 Gowalla Foursquare
Benchmarks: Lastfm-2k Taobao2014 Gowalla Foursquare
Metrics: AUC Recall@20 runtime
Baselines: Fine-Tune FD-Retrain Uniform ER-MIR InvDeg GraphSAIL SGCT LWC-KD INFER-Vanilla
Builds on: influence functions (Koh and Liang 2017) experience replay GraphSAIL ER-MIR GDumb
Closest papers by meaning
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.87)
- Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.87)
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.86)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.86)
- On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · similarity 0.86)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.86)
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.