Activity Monitoring Using Topic Models

Activity monitoring is the task of continual observation, which in many cases, needs to verify a window of data that is captured from a data stream. For some types of categorical data, such as zip codes and phone numbers, hundreds of unique attribute values exist and the frequency vector of a window will be high-dimensional and sparse. This vector is then hard to be compared to the frequency vector of the training set that was collected from a much longer period of time. In this paper, using topic models, we present a method for dimensionality reduction which can detect anomalous windows with low false positive and false negative rates. We also address the problem of the variable nature of normal data by updating the model parameters along with the gradual changes of the data. Our experiments on several real life datasets show that our model outperforms state-of-the-art methods for categorical data with large domains of unique attribute values.