Research map / Bandits and online learning / Contextual bandits and Thompson sampling
Neural Combinatorial Clustered Bandits for Recommendation Systems
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 NeUClust, a neural contextual combinatorial bandit that clusters base arms by context and uses two neural networks (base-arm reward and monotonic super-arm reward) to pick super arms without an oracle, with an Õ(d̃√T) regret bound and experiments on MovieLens and Yelp.
Problem
Contextual combinatorial bandits for recommendation where the agent selects a size-K subset of arms, learns unknown base-arm and super-arm reward functions online, and needs to avoid restrictive reward models and NP-hard optimization oracles.
Approach
Cluster arms by their contexts (k-means), use a monotonic neural network to estimate super-arm reward and a neural UCB base-arm network, then select the cluster with highest average UCB and play its top-K arms, removing the need for an exact or approximate optimization oracle.
Key results
- NeUClust achieves final cumulative regret of 152 on MovieLens versus 168 for CN-UCB, 207 for K-LinUCB, 227 for Neural-MAB and 403 for CC-MAB.
- On Yelp, NeUClust achieves regret 157 versus 182 for CN-UCB, 185 for Neural-MAB and 282 for K-LinUCB.
- Theory proves cumulative regret Õ(d̃√T) where d̃ is the effective dimension of the NTK matrix.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use NeUClust instead of CN-UCB | MovieLens 25M recommendation · MovieLens 25M | cumulative regret: 168 -> 152 | Helped |
| use NeUClust instead of CN-UCB | Yelp recommendation · Yelp | cumulative regret: 182 -> 157 | Helped |
| use NeUClust instead of the strongest other baseline | MovieLens 25M recommendation · MovieLens 25M | cumulative regret: 168 -> 152 | Helped |
| use NeUClust instead of the strongest other baseline | Yelp recommendation · Yelp | cumulative regret: 182 -> 157 | Helped |
Surprising findings
- Theoretical bound scales linearly in super arm size K for the UCB term, because no optimization oracle is used.
- CCMAB was excluded from Yelp experiments because its performance was considerably worse in runtime and regret.
Limitations
- Regret bound grows linearly in K and becomes linear in T if clusters are not identifiable or delta_t does not decay.
- Restricting super arm selection to a single cluster may lead to suboptimal reward.
- Experiments are on simulated binary rewards derived from MovieLens/Yelp genre contexts with synthetic mean functions.
Experimental setup
Datasets: MovieLens 25M Yelp Open Dataset
Benchmarks: MovieLens 25M Yelp Open Dataset
Metrics: cumulative super arm regret runtime (s)
Baselines: CN-UCB Neural-MAB CC-MAB K-LinUCB
Builds on: CN-UCB Neural-UCB MonoNet CC-MAB
Closest papers by meaning
- Online Clustering of Dueling Bandits (ICML 2025 · similarity 0.89)
- Neural Dueling Bandits: Preference-Based Optimization with Human Feedback (ICLR 2025 · similarity 0.88)
- From Contextual Combinatorial Semi-Bandits to Bandit List Classification: Improved Sample Complexity with Sparse Rewards (NeurIPS 2025 · similarity 0.87)
- Optimal and Efficient Contextual Combinatorial Semi-bandits with General Function Approximation (NeurIPS 2026 · similarity 0.87)
- Demystifying Online Clustering of Bandits: Enhanced Exploration Under Stochastic and Smoothed Adversarial Contexts (ICLR 2025 · similarity 0.87)
- Bandits Meet Mechanism Design to Combat Clickbait in Online Recommendation (ICLR 2024 · similarity 0.87)
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.