Interactive Sequence Mining

Abstract

Understanding how people use a program is vital for proper development and support of any application. Numerous ways to gain this insight exist, but those often require user feedback or further adjustments to the program. In this thesis we run multiple pattern mining techniques on existing program traces to reveal patterns in application usage. This process is illustrated and verified by taking existing log files of the application TeleDIA which is currently used in production in various healthcare institutions. To make numerous algorithms available to users new to the field of pattern mining we created an interactive tool. The tool supports the user by providing sensible default parameters and inferring others from the dataset. To further support the user the program provides an interactive graph-view that visualises patterns. The view also allows the user to interactively grow the patterns to specific interesting sequences he is interested in, which are then used as a highly-effective filter on all pattern suggestions. With this method we tackle part of the ever-present problem of pattern explosion in pattern mining. Experiments show we are able to mine almost all patterns present in the trace files. Next to quantitative results we interviewed key users to confirm that the mined patterns are interesting, and in some cases novel. This is supported by the fact that changes were made to TeleDIA, changes based on insight gained using our interactive sequence mining prototype.