Learning Classifiers From a Relational Database of Tutor Logs

A bottleneck in mining tutor data is mapping heterogeneous event streams to feature vectors with which to train and test classifiers. To bypass the labor-intensive process of feature engineering, AutoCord learns classifiers directly from a relational database of events logged by a tutor. It searches through a space of classifiers represented as database queries, using a small set of heuristic operators. We show how AutoCord learns a classifier to predict whether a child will finish reading a story in Project LISTEN’s Reading Tutor. We compare it to a previously reported classifier that uses hand-engineered features. AutoCord has the potential to learn classifiers with less effort and greater accuracy. ________________________________________________________________________