Causal Correlation Detection in Telecommunication Networks Failures via Graph Analysis

The work arises upon a long-lasting relationship between Tim S.p.A. and the Politecnico di Torino, aimed at developing an intelligent alarm management system capable to handle all network failures rising along the national territory and cluster them based on causal correlation. Indeed, the company is currently detecting all failures at a single-device granularity, lacking therefore an high-level partitioning linking causally-correlated alarms which would help TIM S.p.A. in quickly identifying the problem and operating accordingly. As of now, clustering is performed following ad hoc heuristics which, while presenting comforting results, are still only slightly validated by TIM S.p.A. domain experts due to the time-consuming process required. This thesis is intended as a parallel approach to the one currently adopted – based on a sounder mathematical foundations – with the goal of not only providing a promising new method, but also a possible validation of the heuristics in use. Whereas the current approach is based on a priori chosen temporal windows which group together all alarms falling in them (hence yielding a strong prior), the proposed method is based on graph modeling, following the intuition that causal correlation can be modeled as an arc between two nodes representing correlated failures. More in detail, starting from an high-level geographical clustering, all failures are modeled as nodes of a graph, building directed edges between each alarm and all those arising later on, weighting also those arcs on such temporal distance. It follows that this method can be seen as an extension of the heuristic one, where the goal coincides with the one of detecting connected components, if one provides a temporal upper threshold for arcs existence. A more elegant process is instead adopted in this work, where the Leiden algorithm with a Constant Potts Model optimization (that is, the state of the art for community detection) is implemented with the goal of determining a meaningful partitioning under causal correlation. A simple post-processing based on a weighted label propagation algorithm is finally applied in order to reduce the sometimes highly-granular community structure detected by the algorithm. While still requiring a validation from TIM S.p.A. domain experts, the results are in line with those obtained following the heuristics determined a priori with the company, hence supporting with more confidence such direction. Moreover, the Leiden-based approach presents itself as a more general parallel technique which can be exploited and further improved in the years ahead, eventually taking advantage from non-spatio-temporal information as well.