Predicting User-Cell Association in Cellular Networks from Tracked Data

We consider the problem of predicting user location in the form of user-cell association in a cellular wireless network. This is motivated by resource optimization, for example switching base transceiver stations on or off to save on network energy consumption. We use GSM traces obtained from an operator, and compare several prediction methods. First, we find that, on our trace data, user cell sector association can be correctly predicted in ca. 80% of the cases. Second, we propose a new method, called "MARPL", which uses Market Basket Analysis to separate patterns where prediction by partial match (PPM) works well from those where repetition of the last known location (LAST) is best. Third, we propose that for network resource optimization, predicting the aggregate location of a user ensemble may be of more interest than separate predictions for all users; this motivates us to develop soft prediction methods, where the prediction is a spatial probability distribution rather than the most likely location. Last, we compare soft predictions methods to a classical time and space analysis (ISTAR). In terms of relative mean square error, MARPL with soft prediction and ISTAR perform better than all other methods, with a slight advantage to MARPL (but the numerical complexity of MARPL is much less than ISTAR).