Machine Learning Based Mobile Network Data Analysis and Prediction in Wireless Communication Network

With the rapid development of wireless networks, more and more online services significantly raise mobile data traffic demands, which causes a massive challenge for wireless network operators. In addition to deploying more communication facilities to improve the whole wireless network capacity, real-time observation and prediction of the mobile data traffic to achieve a dynamic balance of network load can further improve the efficiency of the services from those operators and reduce energy waste. In general, the accuracy of mobile data traffic prediction directly impacts the entire network system's Quality of Service (QoS) and Operating Expenditure (OPEX). Therefore, mobile network traffic prediction is the main research direction of this thesis. Firstly, the user's Point of Interest (POI) exploration is chosen as a key point for analysis. This kind of user mobility modelling represents an essential branch of mobile traffic analysis. By applying machine learning algorithms, clear summaries of the mobility pattern characteristics of typical wireless users are obtained. Through the analysis of these regular characteristics, the value of mobility information related to the user's POI is initially demonstrated. Subsequently, this thesis introduces a data prediction model based on the Long Short-Term Memory (LSTM) model, a typical neural network for sequence prediction. Through verification of predictions using real sampled user data, it further demonstrates that the user's POIs tend to be relatively fixed and exhibit periodicity. Additionally, by comparing the prediction results with those of other models, the advantages of neural networks, particularly LSTM, in sequence prediction are evident. Subsequently, this thesis aims to enhance the accuracy of wireless data traffic prediction by exploring location information. Although former researches have indicated that the distance relationship may affect the similarity of mobile communication traffic across different base stations, there is a lack of studies regarding the selection of the training dataset scope for urban mobile traffic. Building upon the previous research on the user's POI characteristics, this thesis verifies that the mobile network data trends of base stations in distant regions could also exhibit high similarity with applying real-world data, thereby expanding the training sample range. After that, a multi-task learning framework called MTL-STPN is designed to incorporate these highly correlated mobile traffic data as auxiliary content for predicting target region mobile traffic data. The results demonstrate that the designed model achieves nearly a 10\% improvement in mobile traffic prediction compared to the state-of-the-art traffic prediction models with Root Mean Square Error (RMSE) measurement as prediction metrics. This outcome substantiates that reasonable correlations between mobile network traffic samples can be applied to enhance the performance of appropriate algorithms. Finally, to address the more complex and bursty but highly valuable application-level mobile network traffic prediction, specifically Instant Messaging (IM), this thesis further improves upon the characteristics of the sub-models extracted from the multi-task framework and proposes a novel deep stacked learning architecture called SLIM-TP. After operating the sub-models for extracting the spatiotemporal dependencies of traffic as well as the mobile users’ equipment (UEs) behavioural information, the meta-learner is employed to make optimal decisions regarding these features and effectively retain the factors that can enhance prediction accuracy. Experimental results based on a large dataset collected from a real cellular network demonstrate that the proposed model achieves over 40\% improvements in WeChat traffic prediction performance compared to the state-of-the-art traffic prediction models through RMSE measuring. It shows the effectiveness of incorporating high-dimensional data such as user location and related traffic as auxiliary features in complex mobile network traffic prediction scenarios.