Research map / Learning theory and algorithms / Prediction and decision losses

Ready for You When You Are Back: Content-Driven Session-Based Recommendation for Continuity of Experience

AAAI 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Ranking models from comparisons · 0 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 content-driven session boundaries (split when the dominating item category changes, using BERT-embedding clusters as category labels) instead of fixed time-threshold sessions, and plugs them into six session-based recommenders, improving recall and MRR/NDCG on four datasets.

Problem

Time-driven session-based recommenders split sessions by a fixed idle threshold θ, so when users return after leaving the platform, recommendations differ from what a continued session would have produced, breaking continuity of experience; the θ value is also hard to choose.

Approach

Create sessions from content homogeneity: items get category labels via k-means clustering of BERT embeddings of item descriptions, and a session continues while a single category dominates, splitting when a different category begins to dominate. Timestamps are used only for ordering.

Key results

• Content-driven sessions improve Recall@20 over time-driven sessions for all baselines on MovieLens, Goodreads, LastFM and Amazon; e.g., GRU4Rec on Amazon rises from 0.4034 to 0.5466 Recall@20.
• Tr4Rec on MovieLens NDCG@20 improves from 0.0942 to 0.1371 with content-driven sessions.
• Authors report no change in the testing procedure; only session IDs differ between the two settings, and the CD method removes the need to tune the time threshold θ.

Measured effects reported in the paper

Surprising findings

• CD-HRNN, which uses item descriptions, still gains from content-driven sessions, which the authors take as evidence the gain is not due to content information alone.
• STAMP performed worse when user IDs were removed, and evaluation changes for single-item test sessions lowered performance overall for both session types.

Limitations

• Requires item descriptions or categories; LLM or web scraping is needed when they are missing.
• Category labels come from k-means clustering with the elbow method, and the number of clusters is a heuristic choice.
• Gains were measured with retrained baselines on sessions only; no live user study of continuity of experience is reported.

Experimental setup

Datasets: MovieLens Goodreads Book LastFM Amazon e-commerce behavior

Benchmarks: MovieLens Goodreads Book LastFM Amazon e-commerce behavior

Metrics: Recall@10 Recall@20 MRR NDCG

Baselines: STAMP NARM GRU4Rec HRNN Cd-HRNN Tr4Rec

Builds on: GRU4Rec HRNN NARM STAMP Tr4Rec

Closest papers by meaning

• ScoreNet: Consistency-driven Framework with Multi-side Information Fusion for Session-based Recommendation (AAAI 2025 · similarity 0.86)
• Top-one Recommendation with Anonymous User Behaviors (Student Abstract) (AAAI 2025 · similarity 0.85)
• Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.85)
• ORBIT - Open Recommendation Benchmark for Reproducible Research with Hidden Tests (NeurIPS 2025 · similarity 0.85)
• Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · similarity 0.85)
• Influential Exemplar Replay for Incremental Learning in Recommender Systems (AAAI 2024 · similarity 0.85)