Research map / Probabilistic and scientific modeling / Molecules and proteins
BayesCNS: A Unified Bayesian Approach to Address Cold Start and Non-Stationarity in Search Systems at Scale
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 BayesCNS, an empirical-Bayes Gamma-Poisson prior network over user-interaction features with non-stationary Thompson-sampling updates guided by a ranker, addressing item cold start and non-stationarity in search.
Problem
Learning-to-rank models in search and recommendation depend on user-interaction features, which are sparse for new items (cold start) and shift non-stationarily over time, and existing approaches handle these separately or via costly periodic retraining.
Approach
Treat cold start and non-stationarity jointly as online Bayesian learning: a neural network predicts Gamma-Poisson (negative binomial) priors over interaction counts from contextual features, which are updated with decay (gamma) as new clicks arrive and sampled via Thompson sampling to feed a ranker.
Key results
- Online A/B test over one month: overall new item interactions (impression) +10.60% and success metric +1.05% versus production baseline (both significant at p<0.05).
- On CiteULike Recall@100, BayesCNS reaches .7001 versus .6852 for Heater, the strongest baseline.
- In simulation, BayesCNS outperforms both non-behavioral and behavioral rankers across low and high predictive power w, and adapts across non-stationary episodes where the stationary variant fails.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BayesCNS instead of the production baseline | Search system, Overall A/B test cohort · Online A/B test, Overall cohort | Succ. (+%): 1.76 -> 2.81; Difference 1.05* | Helped |
| Use BayesCNS instead of the production baseline | Search system, Overall A/B test cohort · Online A/B test, Overall cohort | Impression (+%): 7.62 -> 18.22; Difference 10.60* | Helped |
| Use BayesCNS rather than the non-behavioral approach (ablation) | Cold-start simulation · Stationary simulated environment, across predictive power settings | CTR | Helped |
| Use BayesCNS rather than the behavioral approach without the learned prior (ablation) | Cold-start simulation · Stationary simulated environment, across predictive power settings | CTR | Helped |
| Remove the non-stationary assumption (Bayesian Stationary) (ablation) | BayesCNS · Non-stationary simulation, successive episodes | CTR | Hurt |
| Remove user interaction features (Non-behavioral) (ablation) | BayesCNS · Non-stationary simulation, successive episodes | CTR | Hurt |
Surprising findings
- The behavioral (interaction-feature) ranker performs worse than the non-behavioral ranker at high predictive power due to selection bias on previously interacted items.
Limitations
- Interaction features are simplified to a single scalar click count in simulation.
- Benchmark results are described as only competitive, not uniformly best, with BayesCNS trailing Heater on some precision metrics.
- Online A/B test is from a single proprietary search system and results are anonymized.
Experimental setup
Datasets: CiteULike LastFM XING simulated search environment
Benchmarks: CiteULike LastFM XING simulated stationary environment simulated non-stationary environment
Metrics: Recall@k Precision@k NDCG@k CTR success rate new item impression rate
Baselines: KNN LinMap NLinMap DropoutNet Heater Non-behavioral Bayesian Stationary
Builds on: Thompson sampling Gamma-Poisson / negative binomial DropoutNet Heater
Closest papers by meaning
- Ada-Retrieval: An Adaptive Multi-Round Retrieval Paradigm for Sequential Recommendations (AAAI 2024 · similarity 0.84)
- Evidential Stochastic Differential Equations for Time-Aware Sequential Recommendation (NeurIPS 2024 · similarity 0.84)
- AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking (NeurIPS 2025 · similarity 0.83)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.83)
- Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation (AAAI 2025 · similarity 0.83)
- BayesRAG: Probabilistic Mutual Evidence Corroboration for Multimodal Retrieval-Augmented Generation (NeurIPS 2026 · similarity 0.83)
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.