Unsupervised Pattern Discovery for Mobile Network Tuning Optimization

To ensure reliable service, mobile network operators continuously monitor and manage their networks. Radio network engineers analyze the performance of mobile cells and adjust configuration parameters to optimize operation, through a process commonly known as tuning. However, tuning is time-consuming and repetitive, which makes it inefficient at scale. This thesis addresses the problem of detecting recurring performance patterns across mobile cells to support semi-automated network tuning. Instead of relying on predefined rules or known issues, the proposed approach uses unsupervised machine learning techniques to group mobile cells based on their performance similarities. This allows engineers to analyze entire clusters, identify shared issues, and apply common solutions, ultimately reducing workload and improving the response times. The thesis explores various data representations, including image-based and graph-based representations, and compares several embedding learning techniques, such as convolutional neural networks, autoencoders, and graph-based models. It evaluates multiple clustering algorithms, including k-means, DBSCAN, and Spectral clustering. Given the unsupervised nature of the problem, it proposes an evaluation pipeline using multiple visualizations (e.g., PCA, t-SNE, UMAP, custom heatmaps) and clustering metrics (e.g., clustering coefficients, Rand index, mutual information).