Efficient temporal graph analytics : Using large scale telecommunication data for mobility modeling and infrastructure maintenance

Abstract

Behavioral changes in society or technology can be represented as a graph with dimensions in space and time. Such graphs represent the link between events in the real world and their abstract representation. By analyzing such data, insights are derived, impacting decisions taken in the real world. The datasets collected at a telecommunication company commonly contain these dimensions; for example, the usage of mobile phones or the telemetry of a cable modems in a network. The former can be helpful to determine the change of characteristics of society and its behavior at the scale of whole countries and the latter for predictive maintenance of the network. The scalability of particularly costly operations such as geospatial or graph algorithms is essential when handling such data sets. We develop distributed scalable primitives here for geospatial operations or perform smart aggregations. These primitives are applied to analyze the impact of non-pharmaceutical interventions (e.g. lockdowns) on society. Systemic risk is the possibility that an event at the company level could trigger severe instability or collapse an entire industry or economy. The Systemic risk contribution of companies was hitherto not quantifiable since supply networks on the company-level did not exist except for very few countries. Here we use telecommunication data to reconstruct nationwide company-level supply networks. The resulting networks allow us to quantify the systemic risk of individual companies reliably and thus estimate a country's economic resilience. The method can be used for objectively monitoring change in production processes which might become essential in the green transition. We could achieve impact in the corporate domain for the predictive maintenance of the cable network. For hybrid fiber-coaxial (HFC) networks, searching for upstream high noise in the past was cumbersome and time-consuming. Even with machine learning, the task remains challenging due to the heterogeneity of the network and its topological structure and noisy data. We solve the task by sessionizing the data per-incident and reformulating the classification into a ranking job. We present the automation of a simple business rule (largest change of a specific value), compare its performance with state-of-the-art machine-learning methods, and conclude that the precision@1 can be improved by 2.3 times using the developed machine learning approach.