Research map / Probabilistic and scientific modeling / Molecules and proteins
Counterfactual Task-augmented Meta-learning for Cold-start Sequential 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 CTM, a meta-learning method for cold-start sequential recommendation that augments user histories with counterfactual sequences and aggregates meta-path neighbors, reporting gains over Wide&Deep, TransRec, SASRec, MeLU and Meta TL on Electronics, MovieLens 100k and KuaiRec.
Problem
Cold-start sequential recommendation with sparse user histories; existing meta-learning methods build meta-tasks from regular users, introducing preference bias for cold-start users.
Approach
Intervene on real interaction sequences to create counterfactual sequences that form counterfactual tasks within a MAML-style meta-learning loop (local updates on support, global updates on combined actual and counterfactual query losses), plus self-attention aggregation over meta-path item neighbors.
Key results
- On KuaiRec, CTM reaches MRR 0.520 and Hits@5 0.846 versus 0.371 and 0.603 for the best baseline Meta TL.
- On MovieLens 100k, CTM reaches MRR 0.347 and Hits@5 0.539, versus 0.293 and 0.458 for Meta TL.
- Ablations on KuaiRec show removing counterfactual sequence augmentation drops MRR from 0.512 to 0.476; removing meta-path aggregation drops it to 0.437.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Counterfactual Sequence Augmentation (ablation) | CTM · KuaiRec | MRR: 0.476 -> 0.512 | Helped |
| add Meta-path Neighbor Aggregation (ablation) | CTM · KuaiRec | MRR: 0.437 -> 0.512 | Helped |
| add Counterfactual Sequence Augmentation (ablation) | CTM · MovieLens 100k | MRR: 0.304 -> 0.347 | Helped |
| add Meta-path Neighbor Aggregation (ablation) | CTM · MovieLens 100k | MRR: 0.312 -> 0.347 | Helped |
| set initial interactions k to 3 rather than 2 or 4 (ablation) | CTM · Electronics, MovieLens 100k, and KuaiRec | performance | Helped |
| vary counterfactual intervention degree σ (ablation) | CTM · MovieLens 100k and KuaiRec | performance | Mixed |
Surprising findings
- Ablation on MovieLens 100k shows removing meta-path neighbor aggregation lowers MRR but does not uniformly lower every metric, and the MRR values reported in the main table differ from the ablation's complete-model values for KuaiRec (0.520 vs 0.512).
Limitations
- Counterfactual sequences are generated by random item replacement, so sampling randomness is not studied (acknowledged as future work).
- Evaluation uses only three datasets and five baselines, with no comparison to recent cold-start sequential recommenders.
- Reported numbers are inconsistent across tables (KuaiRec complete model MRR 0.520 in Table 2 vs 0.512 in Table 4).
Experimental setup
Datasets: Electronics (Amazon Review subset) MovieLens 100k KuaiRec
Benchmarks: Electronics MovieLens 100k KuaiRec
Metrics: MRR NDCG@5 Hits@5 Hits@1 NDCG@1
Baselines: Wide&Deep TransRec SASRec MeLU Meta TL
Builds on: MAML MeLU Meta TL SASRec
Closest papers by meaning
- CoDeR: Counterfactual Demand Reasoning for Sequential Recommendation (AAAI 2025 · similarity 0.88)
- Temporally and Distributionally Robust Optimization for Cold-Start Recommendation (AAAI 2024 · similarity 0.87)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.86)
- Customizing Language Models with Instance-wise LoRA for Sequential Recommendation (NeurIPS 2024 · similarity 0.85)
- Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation (AAAI 2024 · similarity 0.85)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · 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.