Research map / Label-efficient and robust learning / Recognition and long-tail
Multiple Purchase Chains with Negative Transfer Elimination for Multi-Behavior 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 MPC, which builds multiple purchase chains (e.g., view→cart→purchase and view→purchase) and removes negative transfer by subtracting embeddings learned from a negative interaction graph, for multi-behavior recommendation.
Problem
Multi-behavior recommendation typically uses only a single purchase chain from auxiliary to target behaviors and assumes all auxiliary interactions help, ignoring negative transfer from items viewed/carted but not purchased.
Approach
Construct all valid purchase chains from view to purchase, initialize each behavior's embeddings from the previous behavior on the chain, and build a negative graph (items interacted with the previous behavior but not the next) whose learned embeddings are subtracted to eliminate negative transfer.
Key results
- MPC improves over the best baseline by 5.04% Recall@10 and 5.41% NDCG@10 on Beibei, and by 76.90% Recall@10 and 89.08% NDCG@10 on Taobao.
- Ablation on Beibei: removing negative transfer elimination drops Recall@10 from 0.2501 to 0.2299; using only view→cart→purchase drops it to 0.1324.
- Taobao gains are much smaller at Recall@50 (1.96%) than at Recall@10, attributed to low cart usage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MPC rather than the strongest baseline HPMR | multi-behavior recommendation · Beibei | Recall@10: 0.2375 -> 0.2501 | Helped |
| use MPC rather than the strongest baseline HPMR | multi-behavior recommendation · Taobao | NDCG@10: 0.0586 -> 0.1108 | Helped |
| add multiple purchase chains and negative transfer elimination (ablation) | MPC · Beibei | Recall@10: 0.1922 -> 0.2501 | Helped |
| add multiple purchase chains and negative transfer elimination (ablation) | MPC · Taobao | NDCG@10: 0.0442 -> 0.1108 | Helped |
| use multiple purchase chains instead of only view→purchase (ablation) | MPC · Beibei | Recall@10: 0.1671 -> 0.2501 | Helped |
| use multiple purchase chains instead of only view→purchase (ablation) | MPC · Taobao | Recall@10: 0.1598 -> 0.1677 | Helped |
| use multiple purchase chains instead of only view→cart→purchase (ablation) | MPC · Beibei | NDCG@10: 0.0637 -> 0.1425 | Helped |
| use multiple purchase chains instead of only view→cart→purchase (ablation) | MPC · Taobao | NDCG@10: 0.1103 -> 0.1108 | No clear effect |
| add negative transfer elimination (ablation) | MPC · Beibei | Recall@10: 0.2299 -> 0.2501 | Helped |
| add negative transfer elimination (ablation) | MPC · Taobao | Recall@10: 0.1606 -> 0.1677 | Helped |
| increase negative-information filtering ratio α from 0 to 0.9 (ablation) | MPC · Beibei | NDCG@10: 0.2299 -> 0.2500 | Helped |
| increase negative-information filtering ratio α from 0 to 1.0 (ablation) | MPC · Taobao | NDCG@10: 0.1672 -> 0.1644 | Hurt |
Surprising findings
- On Taobao, the best negative-filter ratio alpha is 0.1 while Beibei prefers 0.9, which the authors tie to differing behavior conversion rates.
- Removing the single chain view→cart→purchase hurts Beibei more than removing view→purchase (Recall@10 0.1324 vs 0.1671), but on Taobao the opposite holds.
Limitations
- The optimal negative-filter ratio varies across datasets and must be tuned manually; learning it is left to future work.
- Gains on Taobao are much larger at Recall@10 than at Recall@50, indicating less consistent improvement at longer lists.
- Negative items are defined only by absence of an interaction, which may include unobserved positives.
Experimental setup
Datasets: Beibei Taobao
Benchmarks: Beibei Taobao
Metrics: Recall@10 NDCG@10 Recall@50 NDCG@50
Baselines: NCF LightGCN MC-BPR NMTR EHCF GHCF HPMR
Builds on: LightGCN EHCF NMTR Context-aware behavior encoder (GHCF)
Closest papers by meaning
- BLADE: A Behavior-Level Data Augmentation Framework with Dual Fusion Modeling for Multi-Behavior Sequential Recommendation (AAAI 2026 · similarity 0.86)
- On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · similarity 0.83)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.83)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.83)
- Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.83)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.83)
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.