Machine Learning Analysis of Player Reviews and Video Game Development: What Information Is Useful for Games Developers?

This paper investigates how video game developers can use machine learning to extract useful information from large volumes of online player reviews, particularly on the Steam platform. It positions the work against prior research on software development, which has identified useful reviews as those supporting maintenance, revealing user preferences, and suggesting new features. The authors argue that games present distinct challenges because they must engage players emotionally and because judgements of whether a game is "fun" are highly subjective, complicating the processing of player preference data.

The methodology combines two strands. First, BERTopic, a neural topic modelling technique, was applied to anonymised Steam reviews of games released by the participating companies to identify recurring themes. Second, semi-structured interviews were conducted with developers from five companies ranging from independent studios to global publishers, exploring what information they seek from reviews. Preliminary interview findings indicate that developers value reviews that describe bugs in detail, sometimes supplemented by gameplay videos, and that binary likes and dislikes or numeric aspect ratings are useful for summarising satisfaction. Developers regarded complaints about core mechanics or a game's fundamental design as unconstructive, since such feedback falls outside realistic development scope.

Feature requests were widely seen as important but were weighed against business strategy, team capacity, and the stage of development. Interviewees stressed the time and resource costs of responding to player suggestions and noted the risk of player attrition when feature delivery is slow. The authors conclude that their approach can filter out less useful content and refine future directions for game review analysis, though the findings are described as preliminary and drawn from a small interview sample.