TCB: A feature transformation method based central behavior for user interest prediction on mobile big data

Although traditional spatial-temporal features, such as gyration, probability, and the intervals between consecutive records, have contributed to model human dynamics, the importance of these basic spatial-temporal features in predicting mobile user interest is not fully investigated. Moreover, these typical features ignore the fact that human behaviors are highly predictable and centralized. Specifically, human mobility is constrained in a small area depicted by several hotspots, and users tend to access mobile Internet intensively on several particular timeslots, which are defined as hot-times in this article. Thus, this article proposes a feature transformation method based central behavior to construct informative feature sets. Transformation method based central behavior only requires small amount of records to extract hotspots/hot-times information for every user, and projects original records into a relative vector space, of which coordinates represent the effects suffered from corresponding centralities (hotspots/hot-times). Then, the new space is further enriched by statistical summaries related to hotspots/hot-times. Based on the state-of-the-art classification algorithms, the proposed transformation method based central behavior is validated on a large Usage Detail Records dataset generated in real physical world. Results show that features generated by transformation method based central behavior surpass traditional spatial-temporal features and preference in the terms of precision, recall, and f1-score.