Research map / Label-efficient and robust learning / Recognition and long-tail

ScoreNet: Consistency-driven Framework with Multi-side Information Fusion for Session-based Recommendation

AAAI 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ScoreNet, a session-based recommendation framework that fuses multi-side information into candidate scores via a multi-path re-engagement network (MPRE) and a position-enhanced consistent scoring network (PECS), evaluated on Diginetica, Tmall, and Retailrocket.

Problem

Side information fusion in session-based recommendation introduces semantic heterogeneity between item IDs and side attributes, causing inconsistent attention weights and inaccurate candidate scores, and existing methods miss re-interaction patterns over side information.

Approach

Decomposes decoding scores into a re-engagement mode (per side-information type, using position/difference embeddings of re-interaction indices, semantic-agnostic) and a generalization mode (transformer session encoder with item-only, consistent L2-normalized scoring plus position-aware attention), then combines them into a unified softmax scoring.

Key results

• ScoreNet reaches P@20 of 56.27 on Diginetica, 52.24 on Tmall, and 59.37 on Retailrocket, outperforming the best baseline (CORE/DIF-SR) on all three datasets.
• Removing MPRE lowers P@20 by about 9.61% on average across datasets (e.g., Tmall 52.24 to 39.84).
• Improvement over best baseline on Tmall MRR@20 is +21.35%.

Measured effects reported in the paper

Surprising findings

• Replacing MPRE with RepeatNet-style GRU re-interaction modeling drops performance on Tmall P@20 from 52.24 to 52.03 and the no-MPRE variant drops it to 39.84.

Limitations

• Reliance on hyperparameters (alpha2 = 0.3, tau = 12) tuned by grid search on validation data
• Retailrocket has most side information missing beyond categories, limiting multi-side fusion evaluation
• Evaluated only on three e-commerce datasets

Experimental setup

Datasets: Diginetica Tmall Retailrocket

Benchmarks: Diginetica Tmall Retailrocket

Metrics: P@10 P@20 MRR@10 MRR@20

Baselines: GRU4Rec NARM STAMP SR-GNN RepeatNet DSAN CORE Atten-Mixer ICAI-SR NOVA DIF-SR

Builds on: DIF-SR NOVA CORE RepeatNet

Closest papers by meaning

• Dual Sparse Attention Network For Session-based Recommendation (AAAI 2021 · similarity 0.87)
• Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience (AAAI 2025 · similarity 0.86)
• Asymmetrical Hierarchical Networks with Attentive Interactions for Interpretable Review-Based Recommendation (AAAI 2020 · similarity 0.86)
• ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.84)
• LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.84)
• Top-one Recommendation with Anonymous User Behaviors (Student Abstract) (AAAI 2025 · similarity 0.84)