Improving Players' Profiles Clustering from Game Data Through Feature Extraction

The study of players profiles is an emerging area which commonly uses demographic characteristics as determinant features. However, there is the need for a deeper understanding that these characteristics alone do not provide. Another problem is the use of predefined profiles that might be oversimplified or too embracing. This paper investigates players profiles based on their gameplay data, besides demographics, using an educational math game as a testbed. Thus, problems such as noise, mixed data types and high dimensionality must be tackled. To this end, we investigated two feature extraction methods to mitigate these difficulties, Principal Component Analysis (PCA) and Features Agglomeration (FA). Then, two unsupervised learning algorithms were used to find the profiles in our experiments, showing that both PCA and FA improved clustering performance, wherein the best results indicated four profiles: advanced, skilled, beginners and intermediated. Our findings provide game designers with insights about playing styles, can be used to adapt the game in real-time and to assess how distinct players profiles perform in the educational subject, as well as their playing performance.