SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation (2502.10157v2)

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.

• The paper introduces SessionRec, replacing next item prediction with a next session prediction paradigm to mirror real user interaction patterns.
• It employs hierarchical sequence aggregation and dual loss functions, achieving a 27% performance gain and enhancing item ranking significantly.
• The model-agnostic framework is scalable and has been successfully deployed in industrial settings, improving both click-through and conversion rates.

SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation

The paper "SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation" introduces SessionRec, a generative sequential recommendation framework. It aims to address limitations in existing models by aligning the predictive framework with how users naturally interact in real-world scenarios.

Introduction

The paper identifies the misalignment between the widely used Next Item Prediction Paradigm (NIPP) and real-world recommendation needs. While NIPP focuses on predicting the next item in a sequence, it does not correspond to users' actual session-based interaction patterns. SessionRec introduces the Next Session Prediction Paradigm (NSPP) for generative sequential recommendations. NSPP focuses on session-aware representation learning through hierarchical sequence aggregation, aiming to reduce computational complexity while modeling complex sequences of user interactions. Moreover, it incorporates both positive and negative interactions within sessions for broader context capture, significantly enhancing ranking capabilities.

Figure 1: Next Item Prediction Paradigm (NIPP) vs. our proposed Next Session Prediction Paradigm (NSPP).

SessionRec establishes a new foundation for developing industrial-scale generative recommendation systems, illustrated by strong empirical results and power-law scaling laws akin to those found in LLMs.

The Architectural Framework of SessionRec

SessionRec employs a hierarchical session-based encoding mechanism involving intra-session and inter-session aggregation. The framework is model-agnostic, enabling integration with sequence encoding architectures like RNN, Transformer, or more recent models like HSTU.

Figure 2: Overall architecture of our proposed SessionRec.

Embedding Module

The model ingests historical interaction sequences, transforming discrete item-ID and side attributes into embeddings via lookup tables and concatenation into item representations.

Item-based Session Encoder (ISE)

ISE aggregates these item representations into session-level embeddings via methods like pooling, simplifying the representation while maintaining essential context.

Session-based Sequence Encoder (SSE)

SSE further processes these session embeddings to derive user interest representations. It grants the flexibility to utilize any sequence encoding backbone and enhances computational efficiency by reducing input sequence complexity, especially with extensive negative interactions.

Optimization via Loss Functions

SessionRec uses dual loss functions:

1. Retrieval Loss: A sampled cross-entropy loss, designed to prioritize positive item retrieval per session.
2. Rank Loss:
   • A novel contribution improving intra-session ranking and discriminating negative to positive item interactions.
   • Balances retrieval and ranking tasks, weighted by a coefficient $\alpha$, which influences the focus on retrieval accuracy versus ranking quality.

     Figure 3: Impact of ranking loss weight on model performance improvement.

Analysis within experimental settings reveals that using a combination of retrieval and rank loss (optimal $\alpha$ around 0.2) enhances the model's performance significantly.

\section{Empirical Evaluation and Results}

SessionRec was evaluated against established baselines like GRU4Rec, SASRec, SASRec+, BERT4Rec, HSTU, and HSTU+ using public datasets with session labels and real-world interactions, allowing the model benefits to be tested in both controlled and realistic environments.

Figure 2: Overall architecture of our proposed SessionRec.

Figure 4: Performance and Training Time per Epoch

Figure 5: Scaling laws of SessionRec with data volume.

SessionRec demonstrated excellent performance, achieving an average performance gain of 27% over the best existing baseline on public datasets like KuaiSAR and RecFlow across both NDCG and Recall metrics (Table \ref{tab:model_performance}). This gain reflects the paradigm's ability to comprehend diverse interactions in user sessions, notably improving item ranking capabilities with an average performance boost additionally exceeding 39% (Table \ref{tab:model_impovement}).

Figure 3: Impact of ranking loss weight on model performance improvement.

The introduction of a rank loss, which distinguishes negative from positive interactions within sessions, has proven to boost performance significantly, with a precise tuning evidenced in Figure \ref{fig:performance_rankloss}.

Furthermore, SessionRec demonstrates notable scalability (Figure \ref{fig:scaling-law}), enabling improved performance with an increasing volume of training data, closely aligned with the scaling laws observed in LLMs.

To assess online performance, SessionRec was deployed in Meituan's recommendation system, leading to significant improvements in both payment and click-through conversion rates (Figure \ref{fig:online_deploy}).

Conclusion

The paper introduces SessionRec, an innovative shift in the generative sequential recommendation paradigm towards next-session predictions, which are more in line with actual user behavior patterns. Through a model-agnostic approach, the framework provides significant improvements in ranking performance and computational efficiency. This research opens avenues for future exploration into single-model architectures that integrate both retrieval and ranking in recommendation systems, showing potential for transformative impact in industrial applications. SessionRec not only enhances model accuracy but also offers scaling potential, pointing to its robustness and adaptability across different data volumes and contexts. A/B test results demonstrate its practical value in improving key business metrics in industry settings such as the Meituan App's recommendation system. Further optimizations in real-time capabilities hold additional promise for performance enhancement.