BayesCNS: A Unified Bayesian Approach to Address Cold Start and Non-Stationarity in Search Systems at Scale

TL;DR

BayesCNS tackles cold-start and non-stationarity in large-scale information retrieval by learning an expressive empirical Bayesian prior over user-item interactions via a neural network that outputs Gamma-Poisson parameters. This prior, parameterized as $p(x|z)$ with $x$ following a Gamma-Poisson mixture (equivalently a Negative Binomial with $r=oldsymbol{\\alpha}$ and $\rho=\boldsymbol{\beta}/(1+\boldsymbol{\beta})=\boldsymbol{\sigma}(\boldsymbol{\zeta})$), is updated online using Thompson sampling to balance exploration and exploitation in ranking. The prior network outputs $\boldsymbol{\alpha}$ and $\boldsymbol{\zeta}$ (via $\log\boldsymbol{\alpha}$) from contextual features $\boldsymbol{z}$, enabling an efficient, ranker-guided online learning loop that updates posterior parameters $\boldsymbol{\alpha}_t(\boldsymbol{z})$ and $\boldsymbol{\beta}_t(\boldsymbol{z})$ under non-stationarity with a tunable rate parameter $\gamma$. Experiments across stationary and non-stationary simulations, benchmark datasets, and a large-scale online A/B test demonstrate improved new-item interactions (10.60% increase) and modest but meaningful gains in overall success rate (1.05%), validating BayesCNS as a practical approach for scalable, dynamic IR systems.

Abstract

Information Retrieval (IR) systems used in search and recommendation platforms frequently employ Learning-to-Rank (LTR) models to rank items in response to user queries. These models heavily rely on features derived from user interactions, such as clicks and engagement data. This dependence introduces cold start issues for items lacking user engagement and poses challenges in adapting to non-stationary shifts in user behavior over time. We address both challenges holistically as an online learning problem and propose BayesCNS, a Bayesian approach designed to handle cold start and non-stationary distribution shifts in search systems at scale. BayesCNS achieves this by estimating prior distributions for user-item interactions, which are continuously updated with new user interactions gathered online. This online learning procedure is guided by a ranker model, enabling efficient exploration of relevant items using contextual information provided by the ranker. We successfully deployed BayesCNS in a large-scale search system and demonstrated its efficacy through comprehensive offline and online experiments. Notably, an online A/B experiment showed a 10.60% increase in new item interactions and a 1.05% improvement in overall success metrics over the existing production baseline.

Figures (3)

• Figure 1: (a) An Illustration of our proposed approach. The contextual query-item features $\boldsymbol{z}$ is used to construct a learned prior model $f_{\boldsymbol{\theta}}(\boldsymbol{z})$. This model is used to construct prior estimates of the user interaction features $p(\boldsymbol{x}|\boldsymbol{z})$. The model then performs online learning through Thompson sampling. By providing user interaction feature estimates $\hat{\boldsymbol{a}}$ to a ranking model $p_{\boldsymbol{\psi}}(\cdot)$, the model can explore the space of relevant rankings $p(\boldsymbol{c} | \boldsymbol{x}, \boldsymbol{z})$ while being guided by relevant contextual information $\boldsymbol{z}$. The model receives feedback rewards in the form of user interactions $\boldsymbol{x}_t$ which are used to subsequently update the posterior distribution $p(\boldsymbol{x}_t | \boldsymbol{x}_{1:t-1}, \boldsymbol{z})$ in the next timestep $t$. (b) The neural network architecture used for $f_{\boldsymbol{\theta}}(\boldsymbol{z})$ trained to output prior parameters $\alpha$ and $\zeta$ by optimizing a Gamma-Poisson loss function $\mathcal{L}(\boldsymbol{\theta})$.
• Figure 2: Click-through Rate (CTR (%)) comparison of different approaches in simulated environments with varying predictive power ($w$).
• Figure 3: Click-through Rate (CTR (%)) rewards per time step of different methods measured in a simulated non-stationary environment. We ablate our method (Bayesian Nonstationary), removing the non-stationary assumption (Bayesian Stationary), and removing the use of user interaction features (Non-behavioral).