Usando redes neurais para reconstruir traços de sessões de usuários de sistemas de larga escala

Abstract

The process of monitoring online entities (e.g., users, IoT devices, hosts) in large-scale distributed systems is fundamental for understanding their behaviors, analyzing their dynamics, properties, limitations and identifying opportunities for improvement. This process can be relevant in a variety of situations, such as monitoring the availability of internet connections in the last mile and assessing the impact of extreme natural events such as weather, earthquakes, and fires, in the region’s infrastructure. In many systems, the online presence of entities can be monitored by sampling – at regular intervals – the entities currently online. Examples include online users in distributed applications, or active stations on the internet. Monitoring can be flawed, and some entities may not appear as online in one or more lists, thus compromising the accuracy of the data collected. Previous investigations have applied statistical methods to identify the occurrence of such failures, and used thresholds to correct them. In the present work, we investigate the potential of machine learning methods to regenerate monitoring data collected via sampling. In particular, we study the potential of correcting data using deep learning techniques, which can be substantially improve results in comparison with existing statistical methods. Key-words: Machine Learning. Distributed Systems. Monitoring.