Back to search

University of Ioannina

Advanced machine learning techniques for data analysis and processing

Abstract

dc:description

Time series classification is a fundamental machine learning task with applications spanning healthcare monitoring, human activity recognition, industrial fault detection, and financial analysis. Among the many approaches developed for this task, symbolic representations occupy a distinctive position: by discretizing continuous signals into sequences of discrete tokens, they offer a combination of computational efficiency, noise robustness, and interpretability that purely numerical or deep-learning approaches rarely match. The central technique in this family is Symbolic Aggregate approXimation (SAX), which reduces a time series to a compact symbolic string via Piecewise Aggregate Approximation (PAA) and Gaussian breakpoint discretization. Despite its two-decade prevalence, SAX and the Bag-of-Patterns (BOP) classifier built on top of it carry several well-documented weaknesses: parameter selection is typically heuristic, trend information is discarded by the piecewise averaging step, cross-channel dependencies are not natively captured in multivariate settings, sequential order between symbolic words is lost in BOP’s frequency histogram, and the symbolic vocabulary has not previously been exploited as a first-class input to modern sequence models. This thesis develops a coherent programme of extensions that address these weaknesses systematically, while preserving the interpretability that motivates the symbolic approach in the first place. The thesis begins by conducting a systematic review of the SAX literature and organizing the identified extensions into a taxonomy structured around the specific weakness of each address, establishing the conceptual map from which the original contributions depart. The first original contribution introduces two independent extensions targeting separate weaknesses: Multichannel Intelligent Icons (MII), which captures cross-channel dependencies in multivariate time series, and Slopewise Aggregate Approximation (SAA), which recovers trend information through a parallel angle-based discretization branch. Both are evaluated on a human activity recognition benchmark. The second contribution addresses the parameter selection problem by applying the DIRECT global optimization algorithm to the joint tuning of SAX parameters across four feature-extraction strategies, evaluated on twenty-four UEA multivariate benchmark datasets. The proposed approach achieves competitive classification accuracy without gradient information or exhaustive search, while maintaining full symbolic interpretability. The third contribution bridges symbolic time series representations and Large Language Models (LLMs). Observing that SAX already produces ordered discrete token sequences – the native input format of LLMs – the thesis develops SAX_HAR-LLM, a dual-stream framework that pairs a trend-aware case-modulated symbolic encoding with lightweight kinematics-informed descriptors and fine-tunes a causal LLM via QLoRA on structured prompts derived from inertial sensor data. The framework is evaluated on two human activity recognition benchmarks and includes an attention-based interpretability analysis that links predictions to specific symbolic motifs on specific sensor axes. The fourth contribution addresses BOP’s loss of sequential order by constructing directed graphs over SAX-word transitions and learning graph embeddings via Graph Neural Networks. Two graph construction schemes are proposed and evaluated on twelve UCR univariate benchmark datasets, establishing that graph-based representations complement BOP’s frequency information most effectively on series compressed into longer symbolic strings. Taken together, these contributions constitute a systematic and empirically grounded exploration of symbolic time series representation, advancing the state of the art along four distinct dimensions while maintaining a coherent commitment to interpretability throughout.

Degree

thesis:*
Grantor dc:publisher
University of Ioannina
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Pappa, Lamprini
  • Παππά, Λαμπρινή

Subjects

dc:subject × 26

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62510

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

Pappa, Lamprini; Παππά, Λαμπρινή. Advanced machine learning techniques for data analysis and processing. University of Ioannina, 2026. http://hdl.handle.net/10442/hedi/62510