Behavioral Sequence Classification with Constraints

The algorithms for sequence classification usually include three steps: first, discovering the supporting features from the sequences; next, using these features to build the classifier; and last, inputting the new sequences into the classifier to get predictable results. The interesting Behavioral Constraint Miner (iBCM), as one of the constraint-based algorithms, can deliver predefined informative templates that indicate item positions, binary relations, and absence information for different labels of the sequences. Furthermore, when combined with window- based approaches, iBCM can shorten the long sequences used to capture concept drift. Based on this, we propose a novel behavioral analysis algorithm named BSClass to firstly analyze the individual performances of the old templates and make a few modifications to their original designs. Using a flexible way to search for the best templates and window combinations, we improved the calculation for obtaining frequent itemsets in the test. As evidenced by 10-fold cross-validation, these improvements can significantly enhance the accuracy of the challenging classification task and also reduce the size of the features.