Big Data Preprocessing for Multivariate Time Series Forecast
Mikael Kylänpää · UTUPub (University of Turku) · 2020
Big data platforms alleviate collecting and organizing large datasets of varying content. A downside of this is the heavy preprocessing required to analyze their data by conventional analysis techniques. Especially time series data is found challenging to transform from platform-provided raw format into tables of feature and target values, required by supervised machine learning models. This thesis presents an experiment of preprocessing a data-platform-extracted collection of multivariate time series and forecasting it by machine learning models such as neural networks and support vector machines. Reviewed techniques of data preprocessing and time series analysis literature are utilized, but also custom solutions such as log level-based target variable, and valuedistribution-based feature elimination are developed. No significant forecasting accuracies are achieved, which indicates the difficulty of modelling big data. The expected reason for this is the inadequate validation of model parameters and preprocessing decisions, which would require excessive testing to improve.