Scholarlink

This paper develops an ML-based random forest multi-classification model for early-phase prediction of board game market success using characteristics scraped from BoardGameGeek (BGG), analyzing through the domains of game mechanics, language dependence, and social interaction. The variable number of plays was selected as a proxy label to define market success, then logarithmically transformed and binned by quantiles. A random forest classifier was separately trained for social interaction variables, yielding a 0.706 accuracy with precision and recall highs of 0.91 and 0.89, respectively, for extreme classes. The cumulative model using all variables achieved only marginal improvements at 0.71 accuracy. Feature analysis emphasized the predictive dominance of social interaction variables, especially organic community-driven engagement. Novel insights included the unexpected importance of language dependence (ranked fifth overall), suggesting latent influence from accessibility, while game mechanics had minimal impact. The models offer actionable insights for practical application, and future work can optimize models by exploring the observed asymmetrical misclassification patterns.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Random Forest Multi-classification and Gini-Based Feature Analysis of Board Game Success via BGG Characteristic Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tianle Liang,
  • Kieran Rosenfeld,
  • Nathan Lu

摘要

This paper develops an ML-based random forest multi-classification model for early-phase prediction of board game market success using characteristics scraped from BoardGameGeek (BGG), analyzing through the domains of game mechanics, language dependence, and social interaction. The variable number of plays was selected as a proxy label to define market success, then logarithmically transformed and binned by quantiles. A random forest classifier was separately trained for social interaction variables, yielding a 0.706 accuracy with precision and recall highs of 0.91 and 0.89, respectively, for extreme classes. The cumulative model using all variables achieved only marginal improvements at 0.71 accuracy. Feature analysis emphasized the predictive dominance of social interaction variables, especially organic community-driven engagement. Novel insights included the unexpected importance of language dependence (ranked fifth overall), suggesting latent influence from accessibility, while game mechanics had minimal impact. The models offer actionable insights for practical application, and future work can optimize models by exploring the observed asymmetrical misclassification patterns.