SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation

Lei Huang, Hao Guo, Linzhi Peng, Long Zhang, Xiaoteng Wang, Daoyuan Wang, Shichao Wang, Jinpeng Wang, Lei Wang, Sheng ChenarXiv 2025 · DOI · arXiv

Abstract

Modern sequential recommendation systems typically rely on the Next Item Prediction Paradigm (NIPP), treating user interactions as a sequence of individual item tokens. However, this approach often suffers from structural misalignment with real-world request-based systems, high computational complexity, and a narrow focus on single-item interests. We present SessionRec, a framework that shifts the focus to a Next Session Prediction Paradigm (NSPP). By aggregating interactions into session-level tokens, SessionRec achieves a quadratic reduction in computational complexity while better capturing diverse user interests. Experimental results on large-scale datasets and online A/B tests demonstrate that SessionRec significantly outperforms state-of-the-art baselines in both accuracy and efficiency.

Introduction

Sequential recommendation has seen significant progress with models like SASRec and BERT4Rec, which adapt the "next token prediction" logic of Large Language Models (LLMs) to item sequences. This Next Item Prediction Paradigm (NIPP) treats every click or view as a discrete time step. Despite its success, NIPP faces three critical challenges:

  1. Structural Misalignment: Industrial recommendation systems usually serve items in batches (e.g., pages or sessions). NIPP's item-by-item approach ignores the grouped nature of these interactions.
  2. Computational Inefficiency: The $O(N^2)$ complexity of self-attention mechanisms becomes prohibitive as user history length $N$ increases.
  3. Interest Narrowness: Optimizing for the single next positive item fails to account for the multi-faceted interests a user might exhibit within a single future session.

To address these, we propose the Next Session Prediction Paradigm (NSPP), implemented via the SessionRec framework.

Comparison between Next Item Prediction Paradigm (NIPP) and Next Session Prediction Paradigm (NSPP).

Methodology: The SessionRec Framework

SessionRec introduces a hierarchical, model-agnostic architecture designed to process sequences at the session level.

Item-based Session Encoder (ISE)

The ISE module aggregates all items within a single session—including both positive (clicked) and negative (exposed but not clicked) interactions—into a single "session token." While various architectures like RNNs or Transformers can be used, empirical evidence suggests that simple mean pooling is highly effective. It preserves the core representation without introducing the noise often associated with more complex intra-session modeling.

Session-based Sequence Encoder (SSE)

The SSE serves as the backbone, processing the sequence of session tokens. By compressing a sequence of $N$ items into $N/M$ sessions (where $M$ is the average session length), the attention complexity is reduced by a factor of $M^2$. We recommend the Hierarchical Sequential Transduction Unit (HSTU) as the SSE backbone for optimal performance.

Dual Loss Function

SessionRec utilizes a composite loss function to refine both retrieval and ranking:

  • Retrieval Loss: A standard sampled cross-entropy loss used to identify potential positive items for the upcoming session.
  • Rank Loss: A novel component that encourages the model to distinguish between positive and negative interactions within the predicted session, effectively integrating fine-grained ranking into the retrieval phase.

Experimental Results

We evaluated SessionRec using datasets that include session identifiers and negative feedback: KuaiSAR-S (search logs), RecFlow (industrial-scale exposure data), and Meituan production data.

Performance and Efficiency

SessionRec-HSTU demonstrated a substantial performance leap, outperforming state-of-the-art baselines like standard HSTU and SASRec by an average of 27%. Crucially, this gain in accuracy does not come at the cost of speed. As shown in Figure 3, SessionRec-HSTU achieves significantly higher Recall@500 while maintaining a lower training time per epoch compared to item-level counterparts.

Model performance (Recall@500) vs. training time per epoch, highlighting the efficiency and accuracy of SessionRec variants.

Online Impact and Scaling Laws

In a 7-day A/B test on the Meituan App involving over 100 million daily active users, SessionRec led to a 0.603% increase in Page Views for Payment (Pay PV) and a 0.564% increase in conversion rates (PVCTCVR). Furthermore, the framework exhibits clear scaling laws: performance increases linearly on a log-scale as training data volume grows, mirroring the behavior of large-scale language models.

Conclusion

SessionRec represents a fundamental shift in how sequential recommendation models are structured. By treating the session as the atomic unit of recommendation, the framework aligns more closely with industrial deployment realities and achieves superior computational efficiency. Looking forward, the integration of rank loss within the NSPP framework suggests a potential "collapse of the recommendation funnel," where a single generative model could replace the traditional multi-stage pipeline of retrieval, pre-ranking, and ranking.

Summary updated

Abstract

We introduce SessionRec, a novel next-session prediction paradigm (NSPP) for generative sequential recommendation, addressing the fundamental misalignment between conventional next-item prediction paradigm (NIPP) and real-world recommendation scenarios. Unlike NIPP's item-level autoregressive generation that contradicts actual session-based user interactions, our framework introduces a session-aware representation learning through hierarchical sequence aggregation (intra/inter-session), reducing attention computation complexity while enabling implicit modeling of massive negative interactions, and a session-based prediction objective that better captures users' diverse interests through multi-item recommendation in next sessions. Moreover, we found that incorporating a rank loss for items within the session under the next session prediction paradigm can significantly improve the ranking effectiveness of generative sequence recommendation models. We also verified that SessionRec exhibits clear power-law scaling laws similar to those observed in LLMs. Extensive experiments conducted on public datasets and online A/B test in Meituan App demonstrate the effectiveness of SessionRec. The proposed paradigm establishes new foundations for developing industrial-scale generative recommendation systems through its model-agnostic architecture and computational efficiency.
BibTeX
@misc{huang2025sessionrec,
  author = {Lei Huang and Hao Guo and Linzhi Peng and Long Zhang and Xiaoteng Wang and Daoyuan Wang and Shichao Wang and Jinpeng Wang and Lei Wang and Sheng Chen},
  title = {{SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation}},
  year = {2025},
  eprint = {2502.10157},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.10157},
}
Figure