Interpretable Outlier and Anomaly Detection for Mobile Networks from Small Tabular Data

Abstract

Outliers and anomalies in mobile networks refer to significant deviations of the Key Performance Indicator (KPI) from expected values, often degrading user experience. Therefore, detecting and understanding these atypical events is crucial for troubleshooting. To monitor network performance, operators continuously collect data using different testing strategies. One such strategy involves drive-tests that capture datasets with many parameters but limited sample sizes, rendering them unsuitable for deep learning approaches, which require large datasets for effective learning. This paper proposes ROAD (Interpretable Outlier and Anomaly Detection), an unsupervised machine learning methodology designed to detect and understand atypical operational scenarios in mobile networks from drive-tests. This methodology comprises an unsupervised detection stage and introduces an interpretability module that is applied separately to outliers and anomalies. The interpretability module identifies variables and samples associated with atypical events, quantifies the degree of similarity between each variable and the anomaly pattern, and builds a decision tree to reveal the ranges of variables describing atypical scenarios. We implemented the methodology in software and evaluated its performance using real drive-test data. Our method provides high accuracy in detecting outliers and anomalies separately, while reducing the identification of false positives (recall) between 39\% and 63\% compared to an existing explainable detection method.