METHOD FOR FORMING A/B TEST GROUPS WITH MINIMAL INTERFERENCE USING THE STOER–WAGNER ALGORITHM

Abstract

This article examines contemporary methods for constructing experimental groups in A/B testing for digital products, with a specific focus on minimizing mutual interference between participants and reducing spillover effects arising from social or behavioral interactions. The challenge of uncontrolled user-to-user influence has grown increasingly important in highly con- nected online environments, where actions, recommendations, and shared content can propagate across a network and distort the measured impact of an experimental treatment. Such interference undermines the validity of causal inference and often leads to biased or unstable conclusions about product performance. The paper provides an overview of established approaches, including cluster-based randomization, geographic exper- iments, and hierarchical assignment strategies, as well as practical implementations widely adopted by leading technology companies such as Meta and Google. These organizations regularly face large-scale network effects and have developed robust methodologies to ensure reliable experimentation even when user interactions cannot be ignored. A central contribution of the article is the introduction of a graph-based method for constructing experimental groups. In this approach, users are represented as nodes within an interaction graph, and the strength of their connections reflects the probability or intensity of potential interference. Group formation is achieved through identifying a minimum cut in this graph, computed using the Stoer–Wagner algorithm. This algorithm enables efficient and scalable partitioning of large networks while guaranteeing an optimally minimal sum of cross-group connections. As a result, the proposed method effectively reduces spillo- ver effects and significantly improves the internal validity of A/B tests. The advantages of minimum-cut partitioning include natural alignment with real-world interaction patterns, suitability for graphs containing millions of users, reduced risk of contamination between treatment and control groups, and consist- ency with experimentation frameworks observed in industry practice. These benefits underscore the relevance and applicability of graph-based experimental design for modern digital ecosystems.