Forecasting user behaviors in call detail records using LSTM models

Download
2021-2-13
Kocaman, Hasan
City planners, governors and mobile phone operators utilize quite from Call DetailRecords (CDR) data in the fields like optimizations of traffic congestion, event de-tection, billing and advertisement policies. Both individual and crowd analyses helpimproving the quality of services.In this thesis, we present regression and classification analysis for three main prob-lems. In regression analysis, we experiment on call counts and call time-sums ofspecified numbers of users. We propose cluster based and outlier separating modelsin these two tasks for the purpose of improving the results of individual user-basedmodels comprised of various Long Short Term Memory(LSTM) layers. In the clas-sification analysis, on the other hand, we present models that predict next locationson the trajectories of the users. We improve the results of base LSTM model withtwo-predictions-at-once approach. The analyses show that recurrent neural networkswork well with sequential data and optimizations on top of the models yield promis-ing results.
Citation Formats
H. Kocaman, “Forecasting user behaviors in call detail records using LSTM models,” M.S. - Master of Science, Middle East Technical University, 2021.