Machine learning for performance forecast based on CPU/Memory heap data flow and backend usage for anomaly detection

This thesis work tried to realize a Machine Learning algorithm capable of autonomously identify and detect the application status monitored by the APM organizational unit. To this end, two different approaches were adopted: \textbf{Supervised} and \textbf{Unsupervised Learning}. Both can be used in real use case scenarios, depending on the type of monitoring required. A mathematical and theoretical explanation of the various Machine Learning types was given, highlighting the necessary characteristics to know in order to understand the work carried out. The Supervised Learning consisted in the realization of a classifier (created in two different libraries) able to distinguish the application state after being trained on a manually exported dataset. Then the various performance metrics were measured and compared, in order to tune and improve the network. The export, filtering and choice of the dataset was not a trivial problem: in fact, this study highlighted how the number of features and their correlation influenced the results. The initial versions of the datasets brought to misleading results, when tried on classifiers or other algorithms that performed well on other reference dataset. The datasets quality was also reflected in the unsupervised learning phase, that, once the right features and dataset settings were chosen, returned good results. In particular three different types of clustering were tested: k-means, hierarchical and spectral. Also in this case, the performance metrics were described from a mathematical and theoretical point view, and then tested in a practical way. This system, after building a stand-alone, deployable version, is capable to read in real time the data delivered by the APM agents.