Visualized Episode Mining with Feature Granularity Selection

Abstract

There has been much effort in the last decade to provide of the shelf software for episode mining, but still it remains a challenge. This is especially true for sequences that contain hierarchical events, e.g. diagnosis codes. Then, the user must not only decide which features should be used, but also in which granularity. This is even complicated due to the fact that, visualizations of the results of episode mining are rare. Therefore, we first introduce an extension to LifeLines2 that is able to mine and visualize sequential data and second, propose a feature granularity selection method to handle hierarchical event sequences. The granularity selection of features applies a greedy bottom up search that also incorporates support and run time constraints. A significant run time improvement is achieved by a parallelization approach. Finally, we show the applicability on real world medical data sets.