Date of Award

8-19-2025

Degree Type

Thesis

Degree Name

Doctor of Philosophy (PhD)

Discipline

Business

First Advisor

Prof. SEE-TO Wing Kuen Eric

Second Advisor

Prof. YANG Chen

Abstract

Time series generated by complex systems, such as industrial IoT and user behavior systems, confront two core challenges: structured missingness (e.g., continuous or periodic gaps) that disrupt temporal dependencies, and the difficulty in effectively modeling dynamic long- and short-term temporal dependencies inherent in evolving patterns (e.g., user interests). Traditional approaches struggle to balance the preservation of local dependency continuity and the rational association of global long-range dependencies in structured missing scenarios, often incurring high computational costs. In temporal pattern modeling, the lack of adaptive mechanisms to fuse evolving long- and recent behavior trends (e.g., stable interest inertia vs. short-term preference shifts) leads to suboptimal modeling of nonlinear associations in dynamic systems.

To address these issues, this thesis conducts research from two independent dimensions—data quality enhancement and efficient temporal modeling—and constructs a dual-module methodological framework:

Structured Missingness Imputation for Time Series: A local-global dynamic fusion framework is proposed. It employs near-neighbor temporal averaging for preprocessing to retain local dependencies at missing segment boundaries, combined with a lightweight dynamic weight adjustment mechanism to adaptively integrate local interpolation and global dependency modeling. This resolves the failure of dependency reconstruction in structured missing scenarios while reducing computational overhead. Multidimensional temporal feature encoding is introduced to enhance hierarchical modeling of high-frequency fluctuations (e.g., equipment transient anomalies) and low-frequency trends (e.g., long-term operational patterns).

Dynamic Temporal Dependency-Driven Interest Transfer Modeling: An adaptive decoupling framework for long- and recent behavior trends is designed. It utilizes attentionenhanced recurrent neural networks to capture stable historical behavior trends, incorporates bidirectional temporal encoding to model forward-backward dependencies in short-term behaviors, and employs a semantic similarity gating mechanism to allocate weights between long- and recent behavior trends dynamically. This overcomes the bottleneck of traditional models in modeling nonlinear interest evolution. A domain-knowledge-driven interest quantification method is integrated to transform prior information, such as user behavior frequency and duration, into temporal interest representations, improving the model’s responsiveness to dynamic scenarios.

The proposed dual-module framework focuses on the essential characteristics of time series in complex systems, addressing issues encountered in real-world scenarios through methods such as dynamic weight adjustment and attention mechanism optimization. The research outcomes provide reusable modular solutions for industrial data governance, personalized recommendation, and other domains, advancing the analysis of complex system time series from application-specific modeling toward a methodology-driven universal technical system. Future work may explore intelligent collaboration mechanisms between the two modules and extensions to cross-modal complex systems.

Language

English

Recommended Citation

Fan, Z. (2025). Machine learning research on time series data (Doctoral thesis, Lingnan University, Hong Kong). Retrieved from https://commons.ln.edu.hk/otd/246/

Share

COinS