Adequacy of Data for Characterizing Caller Behavior

The increase of advanced service offered by cellular networks draws lots of interest from researchers to study the networks and phone user behavior. From the phone user’s point of view, we are interested in learning caller behavior. In this paper, we characterize caller behavior using probabilistic models based on caller’s call arrival, inter-arrival, and talk time from the call logs. A probabilistic model is generally used to predict or estimate the future observation which is conditioned by a knowledge of the historical data. The question is how much historical data is adequate? We answer this question by presenting a technique to detect and compute the adequate amount of historical data to capture the caller behavior. In fact, this adequate amount of historical data is proved to be more relevant to the future caller behavior than considering the entire historical data and hence useful for constructing a predictive model for caller behavior. In addition, we show the improvement in the performance of a Call Predictor [16] when applying adequacy of data. For our analysis, we use the real-life call logs of 94 mobile users collected at MIT over nine months by the Reality Mining Project group. This paper extends our understanding of caller behavior. We believe that the results are useful in constructing a predictive model of a time series.