Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval
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 GPR4DUR, which represents each user as a Gaussian process posterior over item embeddings (density-based user representation) and retrieves multi-interest candidates via Thompson sampling or UCB, evaluated on three offline datasets and a simulated online setting.
Problem
Modeling users' diverse and dynamic multiple interests for candidate retrieval in recommender systems, where single-point and multi-point user representations struggle with adaptability, uncertainty, and efficiency.
Approach
Represent each user by a personalized Gaussian process regressor fit on fixed pretrained item embeddings with implicit feedback, so the number of interests is determined non-parametrically and posterior variance gives uncertainty for bandit-style exploration during retrieval.
Key results
- On Amazon, GPR4DUR achieves IC@20 of 0.739 vs. 0.717 for the best baseline (REMI), and IR@100 of 0.643 vs. 0.627 for REMI.
- On Taobao, GPR4DUR leads all six interest metrics, e.g., IR@50 0.624 and IC@100 0.891 with Recall@50 of 0.906 and nDCG@50 of 0.936 on the ranking task.
- In an online simulation with 1000 users and 3000 items, UCB and Thompson sampling policies achieve higher cumulative interest coverage than greedy/random at later iterations (Thompson reaches 0.98 at t=10 vs. 0.92 for greedy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · Amazon | IC@50: 0.888 -> 0.895 | Helped |
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · Amazon | IR@50: 0.550 -> 0.560 | Helped |
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · MovieLens | IC@50: 0.973 -> 0.974 | Helped |
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · MovieLens | IR@50: 0.847 -> 0.862 | Helped |
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · Taobao | IC@50: 0.583 -> 0.601 | Helped |
| Use GPR4DUR instead of the strongest listed baseline | Multi-interest retrieval · Taobao | IR@50: 0.573 -> 0.624 | Helped |
| Use an RBF kernel rather than a cosine similarity kernel (ablation) | GPR4DUR · Amazon | IC@50: 0.809 -> 0.895 | Helped |
| Use an RBF kernel rather than a cosine similarity kernel (ablation) | GPR4DUR · MovieLens | IC@50: 0.940 -> 0.974 | Helped |
| Use an RBF kernel rather than a cosine similarity kernel (ablation) | GPR4DUR · Taobao | IC@50: 0.518 -> 0.601 | Helped |
| Increase the number of Ward clusters from 10 to 20 (ablation) | GPR4DUR · Taobao | IR@50: 0.602 ± 0.012 -> 0.624 ± 0.011 | Helped |
| Increase the number of Ward clusters from 20 to 50 (ablation) | GPR4DUR · Taobao | IR@50: 0.624 ± 0.011 -> 0.625 ± 0.011 | No clear effect |
| Use categorical auxiliary loss when pretraining item embeddings (ablation) | GPR4DUR · MovieLens | category consistency | Helped |
Surprising findings
- All methods, including GPR4DUR, have negative TEI@k, so none improve tail-category exposure relative to the holdout sets, which the authors attribute to popularity bias.
- GPR4DUR remains effective in interest coverage at very low embedding dimensions (d=8 and d=4), whereas SUR and MUR baselines degrade.
Limitations
- GPR training/inference cost is cubic in history length, so long histories require truncation to recent interactions.
- Tail-category exposure (TEI) is negative for all evaluated methods, so popularity bias is not addressed.
- User interest collections are assumed not to be highly dynamic, so the model is used as a periodically updated candidate generator rather than a real-time model.
Experimental setup
Datasets: Amazon MovieLens Taobao
Benchmarks: Amazon MovieLens Taobao
Metrics: Interest Coverage (IC@k) Interest Relevance (IR@k) Exposure Deviation (ED@k) Tail Exposure Improvement (TEI@k) Recall@50 nDCG@50 cumulative interest coverage (online simulation)
Baselines: Random MostPop YoutubeDNN GRU4Rec BERT4Rec gSASRec MIND ComiRec CAMI PIMI REMI
Builds on: ComiRec MaxMF YoutubeDNN extreme multi-class classification
Closest papers by meaning
- Learning Multiple User Distributions for Recommendation via Guided Conditional Diffusion (AAAI 2025 · similarity 0.86)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.85)
- Auto Encoding Neural Process for Multi-interest Recommendation (AAAI 2025 · similarity 0.84)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.84)
- Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach (AAAI 2025 · similarity 0.84)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · 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.