Activity Recognition from Call Detail Record: Relation Between Mobile Behavior Pattern and Social Attribute Using Hierarchical Conditional Random Fields

Mobile phone, as a kind of most commonly used vehicle of communication, keep records of every movements of each person. For each cell phone user, the different social attribute, leading to various mobility behaviors and social cliques, reflect on dissimilarity of their call behavior patterns. How to deduce the social attribute from the calling behavior is discussed in this paper, by estimated the time he spent on his business, his family or his friends. The data contains 749 users 3 months call detail records (CDR) with the 5 different jobs, which is selected randomly from database of a telecommunication operator who refuse to apprize its name. In this paper, the daily behavior of one user is divided into 48 parts with every half an hour as a basic element which is labeled with one activity-mode. There are eight activity-modes, inferred using hierarchical conditional random fields (HCRF), including four work-purpose states, two chat-purpose states and two other states as 3 basic elements of calling behavior. The cluster result is shown and the analyses of relation between the cluster and the job are made.