SessionRec: Next Session Prediction Paradigm For Generative Sequential Recommendation

📝 Paper Summary

Generative Sequential Recommendation Session-based Recommendation

SessionRec shifts generative recommendation from predicting the next single item to predicting the entire set of positive items in the next session using session-level representations and a joint retrieval-ranking objective.

Core Problem

Traditional Next Item Prediction Paradigm (NIPP) models train by predicting one item at a time autoregressively, which misaligns with real-world scenarios where users interact with multiple items per session/request.

Why it matters:

Real-world recommendations occur at the session level (e.g., a homepage refresh), requiring simultaneous generation of multiple relevant items rather than a single next item
Standard autoregressive models struggle to model diverse user interests efficiently because they focus on a single positive label per step
Item-level sequences with massive negative interactions create high computational complexity (O(N^2)) for Transformer backbones

Concrete Example: In a food delivery app, a user opens the app (a session) and sees a list of restaurants. NIPP models try to predict just the *one* next restaurant they will click. However, the user might be interested in *any* of 5 different restaurants shown. NIPP fails to capture this set of potential interests simultaneously and ignores the negative feedback (ignored items) effectively within the session context.

Key Novelty

Next Session Prediction Paradigm (NSPP)

Hierarchical Encoding: Aggregates item-level interactions (both positive and negative) into compact session-level embeddings before sequence modeling, significantly reducing sequence length
Session-Level Prediction: Instead of predicting the next token, the model predicts a set of positive items for the next session simultaneously
Joint Retrieval and Ranking: Incorporates a rank loss within the generative framework to prioritize positive items over hard negatives within the same predicted session

Architecture

The SessionRec framework, detailing the flow from Item Sequence to Session Representation to User Interest to Prediction.

Evaluation Highlights

Achieves up to +27% performance gain on average over best baselines (like HSTU and BERT4Rec) on public datasets
Online A/B testing in Meituan App showed +1.4% gain in Gross Merchandise Value (GMV) and +1.1% in Order quantity
Reduces computational complexity of the sequence encoder by a factor of roughly M^2 (where M is average session length) due to session aggregation

Breakthrough Assessment

8/10

Proposes a fundamental paradigm shift from item-level to session-level generative recommendation that better aligns with industrial reality. Significant efficiency and performance gains, validated in a large-scale real-world system.

⚙️ Technical Details

Problem Definition

Setting: Sequential recommendation predicting the set of items a user will interact positively with in the next session

Inputs: Historical interaction sequence S_u containing items, side information, and session identifiers (including positive and negative interactions)

Outputs: A set of items (top-K) that the user is likely to interact with in the next session

Pipeline Flow

Embedding Module: Items → Embeddings
Item-based Session Encoder (ISE): Item Embeddings → Session Representations
Session-based Sequence Encoder (SSE): Session Representations → User Interest Representation
Session Prediction Module: User Representation → Top-K Items

System Modules

Embedding Module

Converts item IDs and side information into dense vectors

Model or implementation: Lookup tables + MLP

Item-based Session Encoder (ISE) (Encoding)

Aggregates intra-session item embeddings into a single token

Model or implementation: Pooling (Max/Mean) or Transformer (Pooling is best practice)

Session-based Sequence Encoder (SSE) (Encoding)

Models dependencies between sessions to generate user representation

Model or implementation: HSTU or Transformer (Decoder-only)

Session Prediction Module

Predicts next-session positive items using retrieval and rank objectives

Model or implementation: Dot Product Similarity + Loss Calculation

Novel Architectural Elements

Two-stage hierarchical encoding (ISE + SSE) transforming item sequences to session sequences before main modeling
Integration of Rank Loss specifically on intra-session negatives (hard negatives) alongside standard retrieval loss

Modeling

Base Model: HSTU (Hierarchical Sequential Transduction Unit) as the preferred backbone for SSE

Training Method: Supervised learning with joint loss

Objective Functions:

Purpose: Retrieve relevant items from the massive candidate set.

Formally: Sampled Cross-Entropy Loss L_retrieval using random negatives
Purpose: Rank positive items higher than exposed-but-ignored items (hard negatives) in the next session.

Formally: Rank Loss L_rank = sum(-log(sigmoid(score_pos - score_neg))) over positive/negative pairs in the target session
Purpose: Balance retrieval and ranking.

Formally: L = L_retrieval + alpha * L_rank

Key Hyperparameters:

embedding_dimension: d
rank_loss_weight_alpha: Typically small (e.g., 0.1 or 0.2 based on ablation sensitivity)
average_session_length: M (approx 23 in Meituan scenario)

Compute: Complexity scales with O( (N/M)^2 ) where N is item sequence length and M is session length, significantly faster than O(N^2)

Comparison to Prior Work

vs. HSTU: SessionRec uses hierarchical session aggregation and session-level prediction targets, whereas HSTU uses item-level autoregression
vs. SASRec/BERT4Rec: SessionRec incorporates negative interactions and ranking objectives explicitly, while these typically use only positive sequences and Next Item Prediction
vs. DIN/DIEN [not cited in paper]: SessionRec is a generative sequential model, whereas DIN/DIEN are attention-based CTR models that typically don't generate sequences directly

Limitations

Relies on the definition of 'sessions', which may be ambiguous or hard to define in some domains
Requires handling of negative feedback (exposure data), which significantly increases data volume compared to positive-only sequences
Performance gain depends on the tuning of the rank loss weight alpha; too high damages retrieval

Reproducibility

No replication artifacts mentioned in the paper. Code URL is not provided. Dataset details (public) are mentioned but specific preprocessing scripts are not. Meituan proprietary dataset is not available.

📊 Experiments & Results

Evaluation Setup

Next-session item prediction on public and industrial datasets

Benchmarks:

Public Datasets (Sequential Recommendation)
Meituan Industrial Dataset (Large-scale industrial recommendation) [New]

Metrics:

HitRate@K
NDCG@K
MRR
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Public dataset experiments demonstrate significant improvements over standard baselines.
Public Datasets (Avg)	Performance Gain	Not reported in the paper	Not reported in the paper	+27% (Relative)
Online A/B testing results from Meituan App deployment.
Meituan App	GMV	Not reported in the paper	Not reported in the paper	+1.4%
Meituan App	Order Quantity	Not reported in the paper	Not reported in the paper	+1.1%

Main Takeaways

SessionRec consistently outperforms item-level baselines (SASRec, BERT4Rec, HSTU) by modeling session coherence and multiple interests.
The rank loss is critical; it forces the model to distinguish between clicked items and exposed-but-ignored items within the predicted session, improving ranking accuracy.
Pooling (Max/Mean) is sufficient for the Item-based Session Encoder (ISE), performing comparably to complex encoders like Transformers while being more efficient.
The model exhibits power-law scaling properties similar to LLMs: performance improves log-linearly with increased data volume.

📚 Prerequisite Knowledge

Prerequisites

Sequential Recommendation (Transformer-based models like SASRec, BERT4Rec)
Generative Recommendation paradigms
Loss functions (Cross-Entropy, Rank Loss)
Scaling laws in LLMs

Key Terms

NSPP: Next Session Prediction Paradigm—a framework that treats the user session (a group of interactions) as the basic unit for encoding and prediction, rather than individual items

NIPP: Next Item Prediction Paradigm—the traditional approach where models predict the single next item in a sequence autoregressively

HSTU: Hierarchical Sequential Transduction Unit—a high-performance generative recommendation backbone used as a baseline and component in this paper

ISE: Item-based Session Encoder—a module that aggregates item embeddings within a single session into a single session representation

SSE: Session-based Sequence Encoder—a module that models the temporal dependencies between session representations

Hard negative: Items that a user was exposed to but did not interact with (e.g., did not click), which are harder to distinguish than random negatives

GMV: Gross Merchandise Value—total value of merchandise sold, a key business metric in e-commerce

Scaling Law: The observation that model performance improves predictably as a power-law function of model size, data size, or compute