Multi-Behavior Sequential Transformer Recommender (SIGIR)
Metadata
| Field | Content | |------|------| | Title | Multi-Behavior Sequential Transformer Recommender | | Venue | SIGIR | | Source link | https://dl.acm.org/doi/abs/10.1145/3477495.3532023 | | Resource type | Academic paper | | Category | Sequential Recommendation |
Overview
This entry indexes a SIGIR paper on multi-behavior sequential recommendation. In large-scale recommender systems, users do not interact with items in a single way: a typical e-commerce session mixes clicks, adding items to carts, favoriting, and purchases. These heterogeneous behavior types carry different signals about user intent and have different levels of commitment. The paper proposes a Transformer-based sequential recommendation approach that explicitly takes multiple behavior types into account when modeling user interaction sequences, aiming to improve next-item prediction over methods that flatten all behaviors into a single stream.
Problem Setting and Motivation
Sequential recommendation has long faced challenges around user-intent understanding, efficiency, and scalability. Classic pipelines treat retrieval, ranking, and generation separately, and standard sequential models (e.g., session-based Transformers) often ignore *which type* of behavior occurred, focusing only on the ordered list of item IDs. Multi-behavior modeling tries to recover the richer intent structure behind the sequence:
- Different behaviors imply different interest strengths (a purchase is stronger evidence than a click);
- Behavior order and type transitions encode short-term intent;
- Single-behavior data (purchase only) is typically sparse, while clicks are abundant but noisy.
- Datasets: public recommendation datasets containing multiple behavior types;
- Metrics: Recall@k, Hit@k, nDCG@k;
- Baselines: single-behavior sequential recommenders and other multi-behavior methods;
- Ablations: contribution of behavior-type modeling versus sequence modeling alone.
- BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
- EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
- Efficient On-Device Session-Based Recommendation
- How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
- LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models
- Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models
- Original paper: *Multi-Behavior Sequential Transformer Recommender*, SIGIR. DOI link: <https://dl.acm.org/doi/abs/10.1145/3477495.3532023>
Method Outline
Following the standard pipeline for this line of work:
1. Input and representation: encode user interaction sequences of items together with their behavior-type labels; 2. Core module: a Transformer backbone processes the combined item–behavior sequence; 3. Learning strategy: training on behavior-aware next-item prediction over logged user sequences; 4. Inference: score candidate items from the final sequence representation.
For the precise architecture details, loss design, and hyperparameters, refer to the original PDF at the ACM link above — this post is an index entry, and quantitative claims should be verified against the paper itself.
Evaluation Practice
Papers in this area are typically evaluated with:
Key Takeaways for Search / Rec / Personalization
1. Architecture: behavior-aware sequence encoders remain a strong baseline family; the trend is toward generative and LLM-based recommenders, but heterogeneous behavior signals are still crucial supervision; 2. Data: interaction logs of different behavior types are the key asset; filtering and type alignment matter; 3. Evaluation: offline ranking metrics should be cross-checked against online satisfaction metrics; 4. Deployment: latency, cost, and interpretability constraints remain hard requirements in production.
Limitations and Future Directions
Likely limitations include benchmark–reality distribution mismatch, sparse behavior labels, and unknown cross-domain generalization. Future work can explore deeper integration with knowledge graphs, causal and fairness constraints, and generative recommendation over multi-behavior logs.