Abstract
Game log data have great potential to provide actionable information about the in-game behavior of players. However, these low-level behavioral data are notoriously difficult to analyze due to the challenges associated with extracting meaning from sparse data stored at such a small grain size. This paper describes a three-step solution that uses cluster analysis to determine which strategies players use to solve levels in the game, sequence mining to identify changes in strategy across multiple attempts at the same level, and state transition diagrams to visualize the strategy sequences identified by the sequence mining. In the educational video game used in this case study, cluster analysis successfully identified 15 different in-game strategies. The sequence mining found an average of 40 different sequences of strategy use per level, which the state transition diagrams successfully displayed in an interpretable way.
Highlights
Game log data provide an opportunity to collect data about players’ problem-solving strategies and mistakes [1,2,3] and can do so in an unobtrusive manner [4, 5] that is both feasible and cost-effective [6].Because log data from games record the exact behavior of players [7], they can provide useful information about players’ thought processes, intentions, and abilities
The data are at such a small grain size (e.g., “selected tool A,” “moved game character to the left”) that there is often no known theory to help identify precisely which actions are important to the construct being measured [18]
The cluster analysis results in this study come from an earlier analysis of the same Save Patch data [37], wherein 15 different strategies for solving levels were identified
Summary
Game log data provide an opportunity to collect data about players’ problem-solving strategies and mistakes [1,2,3] and can do so in an unobtrusive manner [4, 5] that is both feasible and cost-effective [6].Because log data from games record the exact behavior of players [7], they can provide useful information about players’ thought processes, intentions, and abilities. Log data can be used to provide detailed measures of the extent to which players have mastered specific goals [8] or to support diagnostic claims about players’ processes [9]. Examining fine-grained behaviors in such highly unstructured environments creates some inherent problems [19] This is because log data are sparse (in that any given player produces a lot of actions, but any given action may only be produced by a few players), noisy (in that irrelevant actions can vastly outnumber relevant ones, and relevant actions are not usually identifiable a priori), and so large that they are prohibitively costly to examine by hand
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have