Predictive Sequential Pattern Mining via Interpretable Convolutional Neural Networks
Abstract: We present an algorithm using explainable convolutional neural networks for mining sequential patterns from event log data. The key to our approach is utilizing structured regularization to achieve sparse parameter values that closely resemble the results of typical pattern mining algorithms, and allows the learned convolution filters to be interpreted easily. Our method can handle both sequences of individual, unique elements and concurrent multiple-element sequences, which represents most situations where sequences may occur in logs of student actions. We applied our structured regularization method to a self-supervised problem predicting future actions from past actions in two different educational datasets as example applications. Furthermore, we generated features from the learned patterns to evaluate the utility of patterns and trained a supervised model with these features to predict academic outcomes via transfer learning. Our algorithm improves the correlation of sequences with outcomes by an average of r = .131 on one dataset and r = .101 on the other dataset versus a traditional sequential pattern mining algorithm. Finally, we visualize the extracted patterns and demonstrate that they can be interpreted as a sequence of actions.