AI-enabled Multi-modal Network Anomaly Association: A Deep Self/Semi-Supervised Learning Approach

In nowadays large-scale networks, it is challenging for network operation and maintenance systems to analyze the reported massive network anomaly information. To handle this problem, we proposed a deep multi-modal learning approach called multi-modal anomaly root cause analysis, which enables network operation and maintenance systems to automatically and effectively associate the related network anomalies that appear from different modalities or aspects, and then locate the root causes. As a self/semi-supervised approach, our proposal is capable of realizing self-learning, self-adapting, and does not rely on a large number of manual annotations. According to the experimental results in a real large-scale network, without any annotations, our approach achieves up to 14% accuracy improvement in terms of multi-modal network anomaly association and root cause locating compared to the classical association rule mining algorithm Apriori, while its performance turns even much better when a few of labeled training samples are provided. The experiment also well proves the versatility and self-adaptability of our approach, which means our learning-based approach is able to not only achieve fast convergence but also automatically adapt itself to network changes.