Interruptibility prediction using contextual information on smartphones

Abstract

We are frequently interrupted by smartphones for a variety of reasons in our lives. However, such interruption at inappropriate moment can be a huge psychological burden. Owing to its practicality and importance, interruptibility prediction has gained popularity in both academia and industry. In this dissertation, among the various interruption purposes, we predict interruptibility for notifications and phone calls. Because smartphones are widely spread and provide us with rich contextual data, they are currently one of the most popular data sources to understand user behavior, and thus, we use smartphones in our research. The difference in perspective on the present makes our research stand out from previous studies. Most existing studies extracted the feature values from the current moment (i.e., the predicting point), and relatively subsequent studies additionally considered the immediate past (e.g., 5 minutes). However, the events that occurred several hours ago can also affect the current interruptibility owing to a variety of factors such as mood. Motivated by this, we reflect on the whole current day to understand the current moment in-depth. Depending on how the long past is exploited for the prediction, we suggest two methodologies. The first methodology is based on conventional machine learning (ML). To consider long periods in the past systematically, the method divides it into several timeslots, and then, for each timeslot, extracts relevant features (e.g., game playing time in the morning). These new features, along with those derived from the immediate past and the present, are used in a conventional way. The second methodology is based on a recurrent neural network (RNN). We propose the hierarchical LSTM network by leveraging a hierarchical characteristic that the time system has: the timestamps form the timeslots. The method samples context vectors at a certain interval (e.g., 15 minutes) and concatenates them into a sequence. Then, the hierarchical LSTM network understands the context vector sequence and predicts the current interruptibility. Further, we have two directions to improve the models. The former one is with respect to the missing values. Actually, interpolated values based on very far real values are not reliable and can act as noise values. We evaluate reliability of each interpolation and incorporate it into our data set. The latter one is with respect to the generalization of the model. We try to create a generalized model using the data sets of the heavy users. For the evaluation, we used two data sets, KAIST data set and device analyzer data set. The KAIST data set has been obtained from 25 participants during four weeks and is used for notification interruptibility. The other is a large-scale public data set constructed from 907 users during approximately nine months and is used for call interruptibility. The ML model is applied to both notification interruptibility and call interruptibility, whereas the RNN model is applied to only the call interruptibility due to the small scale of the KAIST data set. The experimental results show that the ML model improves the prediction accuracy by up to 16% and 7% compared with the baseline and state-of-the-art methods, respectively. The RNN model achieves a prediction accuracy of 76.6%, which outperforms both a conventional ML-based method and a simple LSTM model. Moreover, when missing values are imputed properly, the accuracy is further increased to 81.3%. The generalized model relieved the cold start problem by achieving a prediction accuracy of about 70%, which is reasonable, for 200 real new users.