Time Series Clustering for Enhanced Dynamic Allocation in A/B Testing

Abstract

An A/B-Test is a method for evaluating online experiments on target items and observing which A/B/C/... variations are better through log reports and statistical analysis of the rewards earned by each variation. Recent advancements in A/B-Tests through reinforcement learning encompass dynamic allocation employing multiarmed bandits (MAB). MABs provides A/B-Tests with fast identification of the best variation (A or B) and helps limit the loss of the test i.e. the cost of exploring low-reward variation. When partial information is available before assigning variations, dynamic allocation is extended to the contextual multiarmed bandit problem (CMAB). Current state-of-the-art approaches for empirically estimating the context-dependent reward function for each variation demonstrate strong performance in limiting test loss and personalized tests. However, few studies have addressed this problem in the context of variable-sized time series. This paper presents a new reinforcement learning methodology to handle A/B-Tests with variable-sized time series as context information. We provide two new methods that obtain a minimization of the cumulative regret with a soft computational cost. This paper also provides numerical results on real A/B-Test datasets, in addition to public data, to demonstrate an improvement over traditional methods.