Research map / Label-efficient and robust learning / Recognition and long-tail
CoreRec: A Counterfactual Correlation Inference for Next Set Recommendation
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 CoreRec, a counterfactual correlation inference framework for next set recommendation that contrasts regular and intervened (inter-set-suppressed) item graphs and uses a post-trained switch to pick item representations.
Problem
Next set recommendation predicts items in a user's next purchase set, but implicit cross-set correlations introduce spurious links and high-frequency items distort correlation modeling, hurting low-frequency items.
Approach
Build a regular graph (intra- and inter-set edges, weighted by co-occurrence and time-interval) and an intervened graph (intra-set edges only); estimate the causal effect of inter-set information as the difference between their predictions, and train a post-trained MLP switch that selects which item representation to use.
Key results
- On JingDong at K=20, CoreRec reaches 61.46 PHR@20 and 47.28 Recall@20 versus ETGNN's 56.52 and 38.12 (reported gains of 8.74% and 22.77%).
- On TaoBao at K=20, CoreRec achieves 36.05 PHR@20 vs ETGNN's 32.52, and the best NDCG/Recall at most K values.
- On TaFeng at K=20, CoreRec is slightly below ETGNN on PHR (62.38 vs 62.45) but better on NDCG (16.91 vs 16.47) and Recall (22.27 vs 21.06).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoreRec instead of the strongest baseline | next-set recommendation · TaFeng | PHR@20: 60.01 -> 62.38 | Helped |
| Use CoreRec instead of the strongest baseline | next-set recommendation · TaoBao | PHR@20: 32.52 -> 36.05 | Helped |
| Use CoreRec instead of the strongest baseline | next-set recommendation · JingDong | PHR@20: 56.52 -> 61.46 | Helped |
| Add a short-term sequence (ablation) | CoreRec · TaFeng | PHR@50: 72.67 -> 73.07 | Helped |
| Add time interval matrices (ablation) | CoreRec · TaFeng | PHR@50: 72.08 -> 73.07 | Helped |
| Add the regular graph (ablation) | CoreRec · TaFeng | PHR@50: 71.78 -> 73.07 | Helped |
| Add the intervened graph (ablation) | CoreRec · TaFeng | PHR@50: 72.18 -> 73.07 | Helped |
| Add the post-trained switch module (ablation) | CoreRec · TaFeng | PHR@50: 72.06 -> 73.07 | Helped |
Surprising findings
- On TaFeng at K=20, CoreRec does not improve PHR over ETGNN (62.38 vs 62.45), so gains are not uniform across metrics and datasets.
Limitations
- Gains are not uniform across datasets and metrics; e.g., PHR on TaFeng at K=20 is slightly below ETGNN.
- Frequency-group analysis shows CoreRec performs comparably to DNNTSP for items with more than 500 purchases.
- The causal framing is heuristic: the intervention is implemented by deleting inter-set graph edges rather than a formal do-operator on a learned variable.
Experimental setup
Datasets: TaFeng TaoBao JingDong
Benchmarks: TaFeng TaoBao JingDong
Metrics: PHR@K NDCG@K Recall@K
Baselines: TOP DeepFM Sets2Sets DSNTSP DNNTSP ETGNN
Builds on: ETGNN DNNTSP Sets2Sets CauseRec
Closest papers by meaning
- Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.84)
- Causal Representation Learning for Generalisable Recommendation (NeurIPS 2026 · similarity 0.83)
- CoDeR: Counterfactual Demand Reasoning for Sequential Recommendation (AAAI 2025 · similarity 0.83)
- CURE4Rec: A Benchmark for Recommendation Unlearning with Deeper Influence (NeurIPS 2024 · similarity 0.82)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.81)
- Social Recommendation via Graph-Level Counterfactual Augmentation (AAAI 2025 · similarity 0.81)
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.