Back to search

Aristotle University Of Thessaloniki (AUTH)

Machine learning approaches for time series problems

Abstract

dc:description

This thesis presents original research in the intersection of machine learning and time series. Time series data is everywhere in nature or our daily life. Every sequence of data where the order in terms of time matters consists of a time series. This type of data has unique properties. Time series data emerge from the dynamics of complex systems and are high dimensional. These aspects bring out many new challenges when trying to extract useful information from time series data using machine learning methods. We focus on both practical and theoretical aspects of machine learning methods when dealing with time series data. The three main pillars of our contributions include novel machine learning architectures, time series embedding representations and designing deep learning models based on information theoretic principles. Our research goals are driven by three common time series problems such as blind source separation, classification and regression. The domains of the data that are used in this work span energy, speech and health. Regarding the first domain we selected the problem of non-intrusive load monitoring (NILM). From the domain of speech the problems that are tackled are speech commands classification and speaker recognition. Finally, in the domain of health we conducted experiments related to mining electrocardiograms (ECGs) and bone age assessment. One of our first contributions is a comprehensive investigation of the aforementioned real-world problems, where we identified research opportunities. The first pillar of our scientific contributions comprises novel machine learning solutions for the problem of NILM. We propose a new stacking machine learning system, a transfer learning mechanism that utilises imaging techniques for time series and three novel neural architectures. Our methods are compared against existing ones and demonstrate state-of-the-art results. Furthermore, we address several evaluation issues by introducing new evaluation measures/approaches and developing new benchmark frameworks. The second major contribution of this research regards time series representations. Time series representation has been an active research direction the last few decades. With the rise of the internet of things (IoT) and the explosion of time series data from various sensors, it is apparent that efficient representations without losing useful information are imperative. We propose a novel time series embedding representation, named Signal2Vec. Signal2Vec is an unsupervised approach to transform a given time series into vectors. The proposed algorithm is compared against existing time series representation methods from the domain of signal processing and chaos theory. The evaluation tasks are a multiclass classification of energy data and the problem of NILM seen as a multilabel classification approach. Signal2Vec outperforms the other methods and demonstrates state-of-the-art results on the above-mentioned tasks. Our third scientific addition touches two research disciplines, deep learning and information theory. Information theory has been a powerful tool for machine learning since the first steps of the field. Nowadays, information theoretic principles are utilised in order to bridge the gap between theory and practice in deep learning. Despite the success of deep learning in many domains, designing such models relies on the process of trial and error. A rigorous mathematical theory of overparameterized models is still missing. Our efforts focus on a popular component of modern neural architectures, called pooling and includes both theoretical and empirical outcomes. We study existing pooling operations via the lens of information theory and try to answer the question of which function should be chosen when designing a new neural network given a task and some data. Then, we develop two novel pooling operations based on the principle of maximum entropy and the principle of information bottleneck. Our theoretical outcomes are verified via experimentation. The proposed methods are compared against others and show superior performance in classification and regression tasks, when solving problems in the areas of speech and health. This scientific work is a stepping stone towards more reliable and efficient machine learning methods for time series problems. We hope that our findings will motivate future researchers and will serve as tools for engineers in high impact industrial applications.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2022

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Nalmpantis, Christoforos
  • Ναλμπάντης, Χριστόφορος

Subjects

dc:subject × 8

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/52028

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Nalmpantis, Christoforos; Ναλμπάντης, Χριστόφορος. Machine learning approaches for time series problems. Aristotle University Of Thessaloniki (AUTH), 2022. http://hdl.handle.net/10442/hedi/52028