Abstract:

With the development of the progress of information technology, the deficiency of traditional psychological testing is becoming more obvious, such as test anxiety and test exposure. Some researchers have begun to test individuals using game-based assessment, which has many advantages, such as increasing the motivation and input level of the participants, and providing the possibility for the implementation of log-file technology. However, the current data analysis and scoring logic ignore substantial information of process, and thus cannot accurately assess individual characteristics and abilities. The advantages of machine learning in data analysis provide a new direction. The machine learning algorithm can analyze the log-file data by building a complex model.

The present study attempted to use game-based assessment combining game log-file and machine learning techniques to predict participants’ ability: reasoning ability and mathematical achievement. Participants were 360 first and second grade students from a middle school in Beijing; predictive variables were a series of features extracted from the game log-file, outcome variables were dichotomous variables calculated from Raven test and mathematics achievement, which took 25th and 75th percentile as the cutoff line. In the model training, the random forest algorithm was selected, 70% samples were randomly selected for cross validation and hyper parametric search, and then the prediction was carried out on the other 30% of samples.

Results showed that the logarithm of the ratio of the first step time to the average execution time was the highest features of average importance ratio, and the number of steps that are different from the optimal solution, thinking time ratio, execution between fluctuation, proportion of repeat steps all contributed to the mathematical achievement prediction model; reasoning ability prediction model was similar. With these important features, it could be found that the reasoning ability prediction model had 76.11% precision, 65.72% accuracy, 63.10% recall and 65.01% F1 scores; the mathematical achievement prediction model had 83.07% precision, 73.70% accuracy, 73.33% recall and 75.57% F1 score.

The finding of the present study showed that the random forest model had acceptable predictive effect when predicting reasoning ability and mathematics achievement classification based on the game log-file, with 75% precision of reasoning and 80% precision of math. In conclusion, the research provides a new method to predict the cognitive ability and academic achievement of the students; the game log-file combined with machine learning can establish an effective discrimination model. This result can provide some reference and direction for the development of educational psychological assessment.

Key words: video game, Sokoban, machine learning, reasoning ability, mathematical achievement.