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

CoDeR: Counterfactual Demand Reasoning for Sequential Recommendation

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 3 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 CoDeR, a graph-based sequential recommender that models user demand shifts with self-attention and demand graphs and applies counterfactual backdoor adjustment to separate true demand shifts from noisy ones.

Problem

Sequential recommendation struggles to detect subtle demand shifts in short sequences, which are often confounded by noise such as accidental clicks, clickbait, or position bias.

Approach

Combines a self-attention demand extractor with a Louvain-modularity demand graph, and uses causal total/natural direct effect analysis with backdoor adjustment (KL-divergence stability and modularity-gain deviation) to treat demand drift as a confounder and debias recommendations.

Key results

• On Diginetica, CoDeR reaches R@20 54.180 and N@20 29.207, the best among all listed baselines (DiffuRec R@20 45.825; CORE R@20 52.894).
• On Tmall, CoDeR reaches R@20 27.269 vs. 26.755 for the second-best baseline (CORE 26.755 is listed as the runner-up by the table ordering, DiffuRec 22.180 at R@10).
• On TaFeng, CoDeR reaches R@20 11.548 and N@20 5.525, with the CDR module giving minimal gain (ablation R@20 11.336 without CDR).

Measured effects reported in the paper

Surprising findings

• Removing the counterfactual reasoning module has little effect on TaFeng, which the authors attribute to stable grocery demand.
• DiffuRec and CORE underperform several simpler baselines on TaFeng, and the gains on Diginetica and Tmall NDCG are smaller than on TaFeng.

Limitations

• Gains on Diginetica and Tmall NDCG are modest, attributed to promotional noise in the data.
• The counterfactual module's benefit is small on stable-demand datasets such as TaFeng.

Experimental setup

Datasets: Diginetica Tmall Tafeng

Benchmarks: Diginetica Tmall Tafeng

Metrics: Recall@10 Recall@20 Recall@40 NDCG@10 NDCG@20 NDCG@40

Baselines: POP item-KNN FPMC GRU4Rec NARM STAMP SR-GNN LightGCN CORE DiffuRec

Builds on: DAGNN SR-GNN Louvain algorithm backdoor adjustment

Closest papers by meaning

• Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation (AAAI 2025 · similarity 0.88)
• Alleviating Dual Biases in Recommendation (Student Abstract) (AAAI 2025 · similarity 0.87)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.85)
• Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.85)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.84)
• SIGMA: Selective Gated Mamba for Sequential Recommendation (AAAI 2025 · similarity 0.84)