Detect Exam Cheating Pattern by Data Mining

This aim of this project is to apply a series of pattern detection Data Mining algorithms to accurately identify cheating by one or more students during classroom test exams. JMP software was utilized to analyze correlation among exam scores for 75 students (sitting at 25 different tables due to space constraints) who took a multiple-choice assessment exam. During the exam, three students were seated per table, each given an exam with the same questions but arranged in different order to prevent cheating. Each of the three students, therefore, was given Versions “A,” “B,” and “C” of the exam, respectively, per table. Nonetheless, the possibility of cheating by students existed since they could still synchronize the question sequence prior to, and during the exam. To detect if a pattern could be identified on the answer keys between students not attributable to chance alone (and therefore attributed to cheating), multivariate statistics tools were used to determine whether there was any association pattern among the students from the same exam table. Hierarchical Clustering and Dendrogram Tree were used to identify the grouping affinity behavior related to exam cheating pattern. The clustering analysis would group students with similar answering patterns among the 75 students who took the exam. The cheating pattern could be identified among the first few groupings if students were seated at the same test table during the exam. Authors also used JMP Graph Builder and Graphical Heat Map to identify and recognize patterns in exam scores among students using visual analysis. To further improve the prediction confidence of these combined tools, the authors also selected the top 20% of questions considered the most difficult ones (as identified by the instructor), in order to increase the detection signal-noise ratio. The probability of picking the same wrong answers on the difficult questions are even more unlikely by chance alone as compared to picking the right answers for the easy questions. It is statistically even more improbable that students would unintentionally select the same wrong answers on difficult questions, and therefore provides very evidence of cheating. Principle component analysis was also used to identify pairs of students who cheated, with separation of pairs based on the top two principle components. From the analysis presented using a unique set of Data mining tools, three tables were summarized in this paper and all supported evidence of cheating in the same student pairs. The predictive model approach using Data Mining tools was very powerful for analysis of the complex exam cheating patterns. This case study has been included in standard curriculum in graduate school course to discourage students cheating on exams. The approach herein can also be used to study patterns in students' multiple-choice answers across subject matter, and to help instructors design their future curriculums based on pattern recognition tools derived from these Data mining algorithms.