Research map / Graph learning / Knowledge and temporal graphs
Dynamic Multi-Interest Graph Neural Network for Session-Based Recommendation
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.
Introduces DMI-GNN, which brings multi-interest learning to session-based recommendation using multiple positional pattern learning and a dynamic multi-interest regularizer that adapts interest diversity to session length.
Problem
Session-based recommendation typically learns a single session representation, missing multiple user interests, while existing multi-interest methods generate excessive interests on short session data.
Approach
Builds on GCE-GNN context learning with a multi-head self-attention multi-interest extractor that adds learnable positional embeddings (multiple positional patterns), and a dynamic regularizer (sigmoid of session-length-scaled interest similarity) that lets the model adjust how many distinct interests are used per session; prediction takes the max score over interests.
Key results
- DMI-GNN reports HR@20 of 40.26 on Tmall versus 37.74 for the best baseline A-Mixer.
- On RetailRocket, DMI-GNN reaches HR@20 57.53 and MRR@20 31.27, above MGIR (56.52, 29.91).
- Cov@20 rises to 94.00 on Tmall versus 71.16 for ComiRec, indicating greater recommendation diversity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DMI regularization (ablation) | DMI-GNN · Tmall | HR@20: 38.27 -> 40.26 | Helped |
| Add DMI regularization (ablation) | DMI-GNN · Tmall | MRR@20: 18.32 -> 18.58 | Helped |
| Add DMI regularization (ablation) | DMI-GNN · LastFM | HR@20: 25.62 -> 25.66 | Helped |
| Add DMI regularization (ablation) | DMI-GNN · LastFM | MRR@20: 9.34 -> 9.39 | Helped |
| Add multiple positional patterns learning (ablation) | DMI-GNN · Tmall | HR@20: 36.93 -> 40.26 | Helped |
| Add multiple positional patterns learning (ablation) | DMI-GNN · Tmall | MRR@20: 17.84 -> 18.58 | Helped |
| Add multiple positional patterns learning (ablation) | DMI-GNN · LastFM | HR@20: 25.63 -> 25.66 | Helped |
| Add multiple positional patterns learning (ablation) | DMI-GNN · LastFM | MRR@20: 9.34 -> 9.39 | Helped |
| Use Multi-Interest rather than single-interest modeling (ablation) | DMI-GNN · Tmall | HR@20: 33.33 -> 40.26 | Helped |
| Use Multi-Interest rather than single-interest modeling (ablation) | DMI-GNN · Tmall | MRR@20: 16.74 -> 18.58 | Helped |
| Use Multi-Interest rather than single-interest modeling (ablation) | DMI-GNN · LastFM | HR@20: 24.42 -> 25.66 | Helped |
| Use Multi-Interest rather than single-interest modeling (ablation) | DMI-GNN · LastFM | MRR@20: 9.02 -> 9.39 | Helped |
Surprising findings
- Plain MIND performs poorly on session data, attributed to generating excessive interests.
- Replacing max-over-interests with summed scores (voting) degrades performance to roughly single-interest levels.
Limitations
- Number of interests U and the regularization hyperparameters (beta, eta) must be tuned per dataset.
- Evaluated only on three benchmark datasets with preprocessing following SR-GNN.
Experimental setup
Datasets: Tmall RetailRocket LastFM
Benchmarks: Tmall RetailRocket LastFM
Metrics: HR@20 MRR@20 Cov@20
Baselines: FPMC GRU4Rec NARM SR-GNN GCE-GNN MIND ComiRec MGIR A-Mixer
Builds on: GCE-GNN ComiRec SR-GNN MIND
Closest papers by meaning
- LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.91)
- Auto Encoding Neural Process for Multi-interest Recommendation (AAAI 2025 · similarity 0.87)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.85)
- Behavior Importance-Aware Graph Neural Architecture Search for Cross-Domain Recommendation (AAAI 2025 · similarity 0.84)
- Active Large Language Model-Based Knowledge Distillation for Session-Based Recommendation (AAAI 2025 · similarity 0.84)
- ScoreNet: Consistency-driven Framework with Multi-side Information Fusion for Session-based Recommendation (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.