Interpreting Computational Models of Interactive Software Usage

Evaluation of how users actually interact with interactive software is challenging because users’ behaviours can be highly heterogeneous and even unexpected. Probabilistic, computational models inferred from low-level logged events offer a higher-level representation from which we can gain insight. Automatic inference of such models is key when dealing with large sets of log data, however interpreting these models requires significant human effort. We propose new temporal analytics to model and analyse logged interactions, based on learning admixture Markov models and interpreting them using probabilistic temporal logic properties and model checking. Our purpose is to discover, interpret, and communicate meaningful patterns of usage in the context of redesign. We illustrate by application to logged data from a deployed personal productivity iOS application