Anomaly detection by means of consecutive pattern discovery

Time series analysis is a research field belonging to the Machine Learning and Data Science research cluster. It covers branches ranging from similarity search to anomaly detection, with Pattern Discovery Algorithms being the key to countless problems. State-of-the-art Pattern Discovery Algorithms struggle to extract meaningful information from long time series without incurring in speed and scalability issues, condition that portrays anomaly detection as an unattainable task for large datasets. Adversities in such matters find their roots in the methods carried out for solving specific problems, which try to break away from the standard brute-force algorithms but inevitably mantain their same nature, causing the resulting algorithms to carry the burden of quadratic complexities. This thesis work presents an approach that aims to change the perspective from what problems are looked upon, presenting a novel General Purpose Pattern Detection Algorithm, which claims better scalability and provides meaningful results for anomaly detection problems. The General Purpose Pattern Detection Algorithm adopts a different, linear approach, which paves the way for research that goes beyond the anomaly detection field: the algorithm is applicable to contexts that include financial data analysis, presenting itself as a possible starting point for more ambitious projects such as pattern classification and prediction.