Transform-Based Multiresolution Decomposition for Unsupervised Learning and Data Clustering of Cellular Network Behavior

Abstract

The growing complexity of cellular networks makes it harder for network operators to control and manage the system. To ease the management and automatically detect network problems, unsupervised techniques have been put to use. This work proposes a novel method that combines Multi-Resolution Analysis (MRA) by wavelet transforms and unsupervised clustering with pre-initialized Gaussian Mixture Models (GMMs) for the totally unsupervised grouping of cellular network behaviors using different metrics. The application of multi-resolution decomposition, allows the much simpler clustering technique to take into account temporal information that would require of a much complex method otherwise, being useful for cluster analysis by experts as different duration issues are now segregated and automatically labeled. The generated labels are indicative of the intensity and duration of the anomalies, such labeling can be linguistic or visual, providing faster issue identification. The proposed approach has been tested with real network data, successfully separating different behaviors analyzed in the evaluation section of the manuscript.

Keywords

• Anomaly detection
• cellular networks
• clustering
• multiresolution analysis (MRA)
• Gaussian mixture model (GMM)