NSPP: Next Session Prediction Paradigm

• NSPP is a paradigm that models and predicts entire future sessions, aligning recommendations with real-world multi-intent user behaviors.
• It employs a hierarchical approach with intra-session and inter-session aggregation to reduce computational complexity while capturing diverse user intents.
• Empirical results show significant gains, including 14–50% improvements in ranking metrics and predictable scaling with increased data volumes.

The Next Session Prediction Paradigm (NSPP) defines a shift in the evaluation and modeling of recommender systems by moving from predicting single next-item interactions to directly modeling and predicting entire future sessions as coherent units. Unlike traditional item-level autoregressive approaches—which generate interactions one at a time—NSPP aligns more closely with real-world session-based behaviors found in practical recommendation scenarios, such as e-commerce transaction blocks or web browsing navigation. NSPP emphasizes session-aware hierarchical representation learning, computational scalability, robust negative interaction modeling, and session-level predictive objectives, resulting in enhanced accuracy, improved computational efficiency, and better support for multi-intent user behavior.

1. Conceptual Foundations and Distinctiveness

NSPP was introduced to address the misalignment between the classical Next-Item Prediction Paradigm (NIPP) and real user interaction patterns in modern recommender systems (Huang et al., 14 Feb 2025). In NIPP, the model predicts the next item given the history, operating in a strictly autoregressive, token-by-token regime: $v_{n+1} = \mathrm{NIPP}(S_u; \mathrm{SeqEncoder})$ where $S_u$ is the user’s interaction history.

NSPP generalizes this by predicting a set of items representing the user’s future session: $$[v_{n+1}, v_{n+2}, \ldots, v_{n+k}] = \mathrm{NSPP}(S_u; \mathrm{SeqEncoder})$$ This establishes a session-aware prediction task, enabling the model to recommend all relevant items a user may interact with in their next session, rather than just the immediate next item. NSPP supports multi-item recommendations, better captures users’ diverse intents within a session, and is fundamentally more scalable and efficient for real-world settings, especially when sessions involve multiple simultaneous or short-term interests (Huang et al., 14 Feb 2025).

2. Session-Aware Hierarchical Sequence Aggregation

A defining methodological element of NSPP is hierarchical representation learning over user behavior data:

• Intra-session aggregation: Each session is encoded as a block-level representation via aggregation of individual item embeddings. This can be implemented using an Item-based Session Encoder (ISE), which pools item records within a session to produce a session token.
• Inter-session aggregation: These session-level tokens are then processed with a Session Sequence Encoder (SSE), such as a Transformer or GRU, capturing cross-session dependencies and long-term user intent.

Hierarchical modeling achieves two key outcomes:

• Reduced computational complexity: Attention-based models move from $O(N^2)$ time/space (N = total events) to $O(S^2)$, where S is the number of sessions—a potentially large reduction if sessions are longer.
• Improved expressivity: Intra-session diversity (multiple intents per session) and inter-session evolution (long-term interest dynamics) are both captured in the hierarchical embedding space.

3. Session-Based Prediction Objective and Loss Functions

NSPP’s prediction objective is to generate all positive (e.g., clicked, purchased) items in the next session. Rather than only predicting the most likely next item, the model predicts a set $$\mathcal{V}_{\mathrm{next}} = \{v_{n+1}, \ldots, v_{n+k}\}$$.

Training objectives consist of two principal components:

• Session-based retrieval loss: A sampled cross-entropy loss encouraging the model to retrieve all true positive items from a large candidate pool:

$$\mathcal{L}_{\mathrm{retrieval}} = -\sum_{i} \sum_{v_j \in \mathcal{V}_{s_i}^{+}} \log \frac{\exp(P_{s_i}^{(v_j)})}{\exp(P_{s_i}^{(v_j)}) + \sum_{v_k \in \mathcal{V}_C^-} \exp(P_{s_i}^{(v_k)})}$$

where $P_{s_i}^{(v)} = \mathrm{DotProduct}(E_u, E_v)$ for the session/user representation $E_u$ and item embedding $E_v$.

• Rank loss within session: To improve intra-session item ranking, a second term penalizes poor ordering among positive and negative items within the predicted session:

$$\mathcal{L}_{\mathrm{rank}} = -\sum_{i} \sum_{v_j \in \mathcal{V}_{s_i}^{+}} \log \frac{\exp(P_{s_i}^{(v_j)})}{\exp(P_{s_i}^{(v_j)}) + \sum_{v_k \in \mathcal{V}_{s_i}^{-}} \exp(P_{s_i}^{(v_k)})}$$

The two are combined: $$\mathcal{L}_{\mathrm{NSPP}} = \mathcal{L}_{\mathrm{retrieval}} + \alpha \cdot \mathcal{L}_{\mathrm{rank}}$$ where $\alpha$ controls the ranking-retrieval trade-off (Huang et al., 14 Feb 2025).

This loss structure encourages the model to retrieve relevant items while optimizing their relative order within the session, accounting for both positive and negative feedback (e.g., clicks and skips).

4. Computational Efficiency and Scalability

The hierarchical aggregation of events into sessions yields a radical reduction in computational demands for sequence modeling, particularly in O(N²)-complexity models. For example, in transformer-based architectures, moving from event-level to session-level sequences reduces both memory and computation by a factor of the average session length squared:

• Input event sequence $N$ → number of sessions $S = N / M$. Relative attention complexity improves by $M^2$ (Huang et al., 14 Feb 2025).

This paradigm enables training and inference on industrial-scale data volumes, as empirically demonstrated in large-scale deployments such as the Meituan App, where models must process hundreds of millions of candidate items or daily events.

5. Empirical Performance and Scaling Laws

Extensive experiments on public datasets (e.g., KuaiSAR, RecFlow) and online A/B tests support NSPP’s practical advantages (Huang et al., 14 Feb 2025):

• SessionRec, the reference NSPP implementation, produces 14–50% improvements in ranking metrics (Recall@K, NDCG@K) versus traditional sequential models like GRU4Rec, SASRec, and BERT4Rec.
• Session-level objectives paired with the rank loss further sharpen accuracy in top-K recommendations, aligning better with real user interaction distributions.
• NSPP models exhibit power-law scaling: retrieval accuracy (e.g., Recall@500) shows linear gains with exponential increases in data, paralleling scaling behaviors observed in LLMs. This property confirms that NSPP can exploit larger and richer behavioral logs with predictable quality improvements.

6. Industrial Deployment and Model-Agnostic Integration

The session prediction framework is model-agnostic: the hierarchical aggregation layer and session-level objectives can be combined with a wide array of sequence encoders, such as GRUs, vanilla transformers, or specialized architectures (e.g., HSTU). Production deployment leverages approximate nearest neighbor (ANN) search for candidate retrieval (e.g., Faiss) and further optimizations through dual-stream processing for historical and recent events.

Advantages in practice:

• Plug-and-play integration with minimal alteration to existing architectures.
• Scalable to real-world session logs with millions of unique items and users.
• Enables retrieval and ranking in a unified generative recommendation model, potentially replacing traditional multi-stage (retrieval, pre-ranking, ranking) cascades.

7. Limitations and Future Directions

NSPP represents a significant paradigm shift but introduces new avenues and challenges for exploration:

• Fine-tuning the ranking-retrieval loss weighting ($\alpha$) to balance between retrieval volume and intra-session ranking is dataset- and product-specific.
• Extending NSPP objectives to incorporate richer cross-statistical features, additional negative feedback, and user context (demographics, temporal cues) could yield further gains.
• Exploring single unified generative models for both retrieval and ranking, as suggested by observed scaling laws, could simplify production pipelines.
• Further paper of session definition granularity (session boundaries, session merging) and the capture of multi-intent sub-sequences remains an open area.

A plausible implication is that the continued development of session- and session-transition–aware architectures—possibly leveraging foundation models and richer item-user co-embeddings—will sustain gains observed under the NSPP as data and application scale increase.

Summary Table: NSPP Key Features

NSPP Characteristic	Description	Impact
Prediction unit	Predicts all items in next session	Multi-intent/realistic modeling
Hierarchical aggregation	Intra-/Inter-session embedding	Lower complexity, richer representation
Loss formulation	Retrieval + session-based rank loss	Improves both recall and in-session ranking
Negative interaction use	Implicitly models negatives in session objective	Better robustness, addresses exposure bias
Scaling property	Demonstrates LLM-like power laws with data/model size	Predictable improvements with more data
Deployment	Model-agnostic, computationally efficient architecture	Industrial-scale applicability, easy adoption

Conclusion

The Next Session Prediction Paradigm fundamentally revises the objective and evaluation protocol for sequential recommender systems by predicting user engagement at the session—rather than item—level. Through hierarchical representation, session-level objective functions, and efficient computation, NSPP achieves demonstrable gains in both academic benchmarks and production environments, and it provides a strong foundation for future development of unified, industrial-scale generative recommendation architectures (Huang et al., 14 Feb 2025).