Biclustering of Log Data: Insights from a Computer-Based Complex Problem Solving Assessment.

Area of Science:

• Educational Measurement and Assessment
• Data Mining and Machine Learning in Education
• Cognitive Science and Problem-Solving

Background:

• Computer-based assessments generate valuable behavioral data through log files.
• Understanding student problem-solving requires analyzing these complex process data.
• Traditional clustering methods analyze either students or features, but not simultaneously.

Purpose of the Study:

• To apply biclustering algorithms for simultaneous classification of students and assessment features.
• To evaluate the effectiveness of biclustering in identifying homogeneous subgroups within process data.
• To explore the utility of biclustering for analyzing action sequence and timing data in assessments.

Main Methods:

• Utilized biclustering algorithms on log file data from the PISA 2012 Computer-Based Assessment (CBA)
• Applied biclustering to the 'Ticket' task, analyzing both action sequences and timing data.
• Compared biclustering results with traditional one-mode clustering approaches.

Main Results:

• Biclustering successfully identified homogeneous biclusters, grouping students with similar behavioral patterns on specific features.
• Specific feature subsets were found to be critical for effective bicluster identification.
• Incorporating time-based features significantly improved the understanding of student actions and outcomes.

Conclusions:

• Biclustering offers a powerful approach for uncovering fine-grained insights into student problem-solving behaviors from log data.
• This method provides a more nuanced understanding than one-mode clustering by analyzing students and features concurrently.
• The integration of temporal data enhances the interpretability and depth of behavioral analysis in educational assessments.