Unsupervised Diagnosis in Large Scale Complex Networks
Zhaobin Liu, Xiaoxu Li, Tong Zhu And Yiyang Zhao
Diagnosis in modern large scale network is really frustrating, as the interactions among network components become more and more complex, as well as the domain knowledge is limited. In this work, we propose a novel approach U-diagnosis for troubleshooting in a large complex network. This approach does not need any supervision from the domain knowledge. Instead, U-diagnosis applies an inference model that encodes inner dependencies among network components and an online fault inference scheme. Compared with prior methods, our model is richer in representing complex dependencies among components which does not rely on domain knowledge. We propose to learn the dependencies from historical monitored data and construct state and dependency library to facilitate the fault inference. By borrowing knowledge from other components, our solution can deal with the lack of historical data and adapt to large network size and high dynamics. We implement and evaluate our approach in a large scale campus network. The experimental results verify the efficiency of our solution.
Keywords: Diagnosis; network management