Recommendation Systems Based On Behaviour Logs Of Mobile Users

Along with our mobile life,a large number of behaviour-related logs(i.e.the Cellular data)are passively recorded.Cellular data(e.g.,voice calls,SMS,traffic data)contain information about the time,location and type of service for every record,but do not include the specific content.With the help of these records,service providers are able to character customers' tastes and interests,so as to make personalized service recommendation.As a consequence,recommending based on logs from operators has become one of the most popular aspects in academia.Compared to the dataset processed by traditional recommenders,the cellular data consist of implicit feedback about customers' actual behaviours.It requires additional side information when building personalized recommendation based on the implicit feedback.To this end,the dissertation has achieved the following innovative results in three aspects,that is the introduction of the semantic location-based network,the behaviour similarity network and the user-item bipartite graph to make personalized recommendation based on users' implicit feedback from the cellular data.1.A hybrid recommender that introduces users' semantic location-based network is proposed.In this method,we assume that users with neighbouring semantic locations are akin to each other,and build a novel implicit network for users' preference similarities based on seman-tic locations.The algorithm first determines their semantic locations such as residences or workplaces based on their historical movements over different time periods;Then,a seman-tic location-based network is constructed,which approximates preference similarities among users by their semantic distances;Finally,by combining semantic location-based network and the item popularity-based weighted matrix factorization,a novel hybrid recommender is proposed for implicit feedback of mobile users.We evaluate this model in a practical s-cenario:the smart-phone recommendation based on operator records.The empirical results conclude that the hybrid recommender combining with the semantic location-based network outperforms the state-of-the-art methods.2.A hybrid recommender that introduces the users' behaviour similarity network is proposed.In this method,we assume that users who behaved similarly tend to have similar preference.The algorithm first obtains users' call-log based network according to users' historical in-teractions;Then,an implicit network is constructed based on historical usage,which is then combined with the call-log based network;Finally,this composited network is merged with the matrix factorization technology to obtain a personalized recommender for application usage of mobile users.The empirical results illustrate that our hybrid model combining the users' behaviour similarity network enhances the performance of models with the call-log based network and other traditional methods.3.An improved metric-based recommender which considers side information from user-item bipartite graph is proposed.In this part,we take the users' relative preferences among their interacted items into consideration for a fine-grained recommendation.To alleviate it,we propose an improved recommender in this paper.The algorithm first builds the user-item bipartite graph based on their historical interactions,and learns their topological similarities from this graph.These similarities are applied in a cascade manner to measure the dis-tance between interacted user-item pairs in the metric space.Finally,under the assumption pulling interacted items according to their topological similarities and pushing missing items away than interacted items,we propose the "Graph Embedded Metric Factorization based model".We evaluate this model through two public real-world datasets and practical us-age records from the mobile service provider.Empirical results conclude that our improved recommender outperforms the state-of-the-art methods when making personalized recom-mendation based on users' implicit feedback.