Learning about clients from mobile call graph

Based on a dataset provided by a telecommunications operator with fully anonymized information about mobile phone communications, the objective is to construct a mobile call graph and evaluate to which extent are network metrics informative for the prediction of age and gender of users from the call graph. A mobile call graph is constructed to represent the social network of mobile phone users and facilitate the understanding of their communication behaviors. Extensive exploratory data analysis is performed to assess the quality of the data and investigate social strategies associated to the network structure. Since the state of the art is not reproducible, a preliminary approach with classical classification methods is proposed for gender and age inference in order to assess the added value of network metrics as features. This thesis shows in detail that several social behaviors adopted by mobile phone users can be identified in call graphs, mostly in accordance to state of the art based on other datasets, thus proving the correlation between demographics and mobile phone communication behaviors within the network. Classification tree-based methods are tested with and without network features as predictors, with negligible performance difference. Therefore, even though the state of the art and the social strategies confirmed in this dataset show the merit of network structure in inference, the tree-based classifiers do not benefit much from it. This verification opens the door for exploring more complex graph-based models, such as probabilistic graphical models, which can better leverage the network structure for improved prediction