A Context Based Generalized User Modeling Framework for Desktop Applications

While current desktop applications provide input for an application’s information processing by monitoring peripheral activity (keyboard, mouse, internal clock), these applications do not access external information from a user’s environment to better adapt their behavior towards individual users. To address this lack of personalization, I design Sycophant, my context-aware user modeling framework. Sycophant leverages user-related information from a desktop’s environment, usercontext, to learn a user’s preferences for different application actions. Results from four real-world user studies show that Sycophant can successfully predict a userpreferred action for a media player and a calendar. In the first pilot study, Sycophant establishes the feasibility of a context based approach to predict user-preferred calendar alarm types. The second study with Google Calendar further extends this alarm type personalization for multiple participants thereby validating my framework’s generalizability across participants. Consistent with the second study’s results, findings from the third study confirm that Sycophant successfully predicts a participant’s preference for Winamp, a media player. Sycophant’s Winamp personalization validates the generalizability of my framework across multiple applications. The fourth study examines and verifies Sycophant’s user-preferred calendar alarm type prediction accuracy over long-term application use. In addition to highlighting Sycophant’s generalizability across participants and applications, these user studies also show that removing user-context features significantly degrades the action prediction accuracy of various machine learning algorithms. This performance degradation emphasizes the need for context based approaches to personalize desktop applications and improve human-computer interaction.