Design and Applications of Machine Learning Algorithms for Time-Series Forecasting

Abstract

In the last decade, big data have gained extensive attention from both academia and industry. Machine learning has been introduced to learn from data, which plays a significant role in providing predictive analytical solutions for large-scale datasets. Therefore, the use of machine learning has increased dramatically during the last decade and it is continuously evolving and developing. In recent years, with the development of machine learning an extensive research has been carried out for solving real-world problems, such as self-driving cars, online recommendations, speech and face recognition systems, human activity recognition systems, and fraud detection in a banking system. One of the important subjects in machine learning is concerned with analyzing and forecasting time series datasets. Time series forecasting models help to predict future values of events having significant impact on our life, such as timely healthcare services and prediction of health conditions of a patient, forecasting the natural hazards and stock pries. Time series forecasting is an important avenue of research in machine learning due to many prediction problems that include a time variable. Nevertheless, when considering the time component as an additional information, it already makes time series problems more challenging to handle compared to other prediction tasks. In literature, most of the data-driven forecasting models are based on supervised learning algorithms. Nevertheless, those models are not time dependent and time variable t is not taken into account during the learning process. In practice, features of a time series dataset are time-variant and its corresponding output (target) value depends on time t. In predicting future time series data value, a historical data points should be used along with its corresponding time instances. Hence, the output of the forecasting model will then be a predicted value y(t0) at a time instance t0. In this regard, reliable and robust forecasting models have to be designed, developed and evaluated. In this thesis, we present the state-of-the-art regarding the machine learning models for analysis and forecasting of time series datasets. Then, we propose a novel neural network-based model for time series forecasting that takes into account a time component. In particular, there two goals are aimed, which are 1) designing efficient neural network-based forecasting algorithm and 2) application and validation of the proposed prediction models on real-world datasets. The main contribution of the thesis is designing a new neural network algorithm that uses a periodic function for forecasting periodic and seasonal time series datasets. We demonstrate that our proposed model outperforms existing other prediction models. We also show that the proposed model can efficiently handle forecasting of time series with missing data. The usefulness of the proposed model is validated on other application domain. We proposed a novel neural network-based parameter computation method to simulate a deterministic channel model, such as Gaussian channel and ocean waves. We demonstrate that the proposed parameter computation method outperforms the existing parameter computation methods. Finally, we present a new key stroke recognition system which enables to recognize the password entered on point-of-sale (POS) terminal. The results show that the proposed model can identify the password consisting of 6 digits by 73 % of accuracy. We also propose a novel kernel function for support vector machine (SVM) used for recognition of keystrokes entered on POS terminal.