Pattern profiling of users' behaviour.

Owing to growth in the popularity of mobile phones, solutions for more efficient mobile network resource management have been increasingly demanded by network operators. Predicting the future state of the network and allocating the network resources based on the predicted state has been proposed as an effective method for efficient management of the network resources by the research community. One of the major factors that changes the future state of network is changes in the behavior of users. As the result, to forecast the future state of network, a major task is to predict the future behaviors of users. This task is accomplished by User Behavior Prediction Models (UBPrMs). In order to maintain the quality of the service, such methods are expected to provide sufficiently accurate prediction. However, the existing methods often are not able to meet this performance requirement. The accuracy of a predictive model is affected by two distinct sources of error, namely Modeling Error (ME) and Sampling Error (SaE). As the result, one ought to consider both sources of error while improving the performance of a model. To do this, this thesis aims to study and alleviate the impact of the mentioned sources of error on the performance of a UBPrM. To treat the ME, we propose a novel group-level user behaviors prediction framework as a more accurate alternative for population-level user behaviors prediction models and a more computationally efficient alternative for individual-level user behaviors prediction. The novel framework is called Event Profiling Method (EPM). To diminish the impact of ME, the proposed event-based method takes advantage of similarities amongst users' behavior and the existing underlying patterns that repetitively occur in the network. To evaluate the proposed framework, EPM method needs to be implemented in real-world scenarios. Video popularity prediction is considered as a suitable use case for EPM. For this purpose, this thesis utilizes the ideas of EPM framework to propose a novel approach for enhancing the video popularity prediction models. Using the proposed approach, we enhance three popularity prediction techniques that outperform the accuracy of the prior state-of-the-art solutions. The major components of the proposed approach are three novel mechanisms for "user grouping", "content classification" and "dominant-follower users identification". The user grouping method is an unsupervised clustering approach that divides the users into an adequate number of user groups with similar interests. The content classification approach identifies the classes of videos with similar early popularity trends. The dominant-follower identification technique divides the users in each group into two distinct subgroups based on their reaction time to the released videos. To predict the popularity of the newly-released videos, our proposed popularity prediction model trains its parameters in each user group and its associated video popularity classes and user subgroups. Evaluations are performed through a 5-fold cross validation and on a dataset containing one month video request records of 26,706 number of BBC iPlayer users. Our analysis shows that the accuracy of the proposed solution outperforms the state-of-the-art including S-H, ML, MRBF models on average by 59%, 27% and 21%, respectively. Afterwards, this thesis proposes a novel combination technique for multi-dimensional user profiles that is able to treat the SaE. In doing so, the proposed technique considers the samples of other users' behavior (or in general, other items) as a biased approximation of each user (or an item). The method utilizes two conditions on the magnitude and sign of the estimated bias between two users to decide on combining their profiles or not. The proposed technique is evaluated against synthesized and real-world datasets. Our results show that the proposed method provides better estimations of the statistics of the synthesized datasets than the standard method. To test the proposed method in the real world, we utilize one-month content request records of 787 users of BBC iPlayer. We show that our proposed method performs on average 15% better than the standard method.