Towards learning segmented temporal sequences: A decision tree approach

Time has been included in learning procedure in domains where observations are recorded on a time basis. However, most recent works take input as a stream of events from the same entity. In those scenarios where a cohort of entities have a short event history, applying existing methods will lead to an isolated or inaccurate (if not both) prediction model for each entity. To address this problem, our work learns the segmented sequences merged from consecutive temporal events of all entities. Instead of using static entropy as splitting metric in the training process, we employ feedback-directed bottom-up approach to build the decision tree. This work adopts a probabilistic model based on Bayes' theorem to enhance prediction accuracy. Moreover, it supports automatic parallelization to reduce overhead. Experimental results demonstrate that the proposed approach not only improves the accuracy of prediction, but also facilitates fast and adaptive parameter tuning, which is essential for its diversified use cases.