Time-Series Analysis Framework
Pedro Moreira Costa · Portuguese National Funding Agency for Science, Research and Technology (RCAAP Project by FCT) · 2019
A time series consists of a data structure that relates an event observation with a time instance.It is the type of data that occurs in a variety of fields and whose analysis (such as Forecasting and Anomaly Detection) provides a more sensitive understanding regarding events' behavior.Time series analysis is usually carried out through the use of plots and classical models.However, Machine Learning (ML) approaches have seen a rise in the state of art for Forecasting and Anomaly Detection because they provide comparable results at appropriate time and data constraints.However, one of their major drawbacks is how costly it is to obtain the best data preparation, model selection and parametrization.With this in mind, this work's main goal is overcoming the complexity involved with data processing, model selection and model tuning, through the support of an autonomous approach selector, based on Case-based Reasoning (CBR) and Bayesian Optimization, present in both forecasting and anomaly detection frameworks.The framework handles Univariate Time Series (UTS) for Forecasting and Multivariate Time Series (MTS) for Forecasting and Anomaly Detection of different categories and bearing different attributes to deliver an optimized approach.The drawn results for each type of analysis, for each type of time series, suffer from insufficient examples, especially MTS, but, for Forecasting UTS, the model selection results show an averaged Macro weighted F1-score of 0.38.Regarding the effects of Bayesian Optimization, all models produced results within an acceptable error (Symmetric Mean Absolute Percentage Error (sMAPE) lower than 16% and F1-score higher than 0.60), proving the efficacy of this component.