Using Data Mining to Identify Common Behavior Patterns in Computer-Assisted Math Learning

• In Event: Self-Regulated Learning Analytics: Aligning Data and Their Treatment to the Assumptions of Theory

Abstract

Objectives & Theoretical Framework. Computer-based learning environments provide extensive data on student learning behaviors, which can be examined using data mining techniques. Previous sequence mining research has investigated self-regulated learning (Kinnebrew et al., 2014; Nesbit et al., 2007) and behaviors of high-achieving and low-achieving learners (Martinez et al., 2011), among other topics. This study uses a sequence mining tool, LASAT (Kinnebrew, Loretz, & Biswas, 2013), to identify learning behavior patterns that are common to high and low achievers in math, and to students with a particular degree of efficacy or belief about intelligence. Specifically, we examined logs of student behavior in an intelligent tutor system to better understand how behavior patterns differ between groups of students who 1) view their intelligence as a fixed asset compared to those who believe they can increase their intelligence, 2) have high or low self-efficacy for math, and 3) ultimately achieve high and low grades in math (i.e., retrospective analysis of behaviors contributing to achievement). We examine the frequency of these behaviors in the group, and the differential frequency across groups, and also incorporate data to explore how students’ behavior patterns changed over time.
Methods. Participants (n=155) were from advanced and middle track algebra classes in a rural high school. Students used Cognitive Tutor Algebra (CTA) twice per week in class and for homework during one academic year. Surveys administered prior to ITS use determined students’ self-efficacy and mindset, and logs of students’ transactions with CTA software during math problem solving were captured for each problem step. Students had multiple opportunities at a particular algebra skill and these were ordered by opportunity number to create a temporal sequence. Individual transactions for each problem attempt within an opportunity were transformed into a sequence and these were grouped based on student characteristics (i.e. fixed/growth mindset, high/low self-efficacy, high/low achievement). A differential sequence mining analysis using LASAT was run on the grouped sequences to identify common behavior patterns and those that were differentially frequent patterns between groups.
Results. Students with a fixed view of intelligence showed greater tendency to make repeated errors and showed a greater reliance on hints to solve problems than peers with a growth mindset (Table 1). Highly efficacious students made greater use of learning resources (e.g., worked examples, glossary) than less efficacious students (Table 2), who tended instead to make repeated errors and request hints. Students who ultimately struggled in Algebra (i.e., earn D & F final grades) frequently make repeated errors and are also more reliant upon hints to solve problems as compared to students who earned an A grade in math (Table 3).
Significance. This data-driven approach revealed new knowledge that further informs our understanding of the ways beliefs affect learning behavior. The study demonstrates the power of data-driven analysis for investigating complex learning processes, and the findings can be used by instructors who use ITS to consider how students’ beliefs might lead them to use technology in a fashion that ultimately associates with a desirable and undesirable learning outcomes.

Authors

• Erica Marti

• Matthew L. Bernacki, University of Nevada - Las Vegas

• Kira Albers, University of Nevada - Las Vegas

• John Kinnebrew, Vanderbilt University

• Vincent Aleven, Carnegie Mellon University

• Timothy James Nokes-Malach, University of Pittsburgh