An adaptive user interface based on spatiotemporal structure learning

We developed a user interface prototype for the Android smartphone, which recommends a number of applications to best match the user's context. To consider the user's context of use, we utilized 5 prototypical variables; time, location, weather, emotion, and activities. The developed system derives the best three recommended applications based on the results of supervised machine learning from such data sets. To consider the history of past context information, in addition to the current one, we developed a novel and effective probabilistic learning and inference algorithm named "Spatiotemporal Structure Learning." By extending Naive Bayesian Classifier, the spatiotemporal structure learning can create a probability model which represents relationship between time-series contextual variables. We implemented a prototype system which shows the current context and the inferred recommendation of applications. For the prototype system, we developed an Android widget application for the user interface and a Java-based server application which learns structure from training data and provides inference results in real time. To gather training data and evaluate the proposed system, we conducted a pilot study which showed 69 percent accuracy in predicting the user's application usage. The prototype demonstrated the feasibility of an adaptive user interface applied to a state of the art smartphone. We also believe that the suggested spatiotemporal structure learning can be applied to number of application areas including healthcare or energy problems.