More about HKUST
Multivariate Time Series Analysis with Deep Learning
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Multivariate Time Series Analysis with Deep Learning"
By
Mr. Shuhan ZHONG
Abstract:
Multivariate time series (MTS) serves as a fundamental data modality across
numerous science domains and real-world applications, enabling multiple
analysis tasks. Based on sampling regularity, MTS can be categorized into
regular (RMTS) and irregular (IMTS) types. Each task and data type presents
distinct challenges for deep learning models, which must capture complex
temporal and channel-wise dependencies. This thesis aims to design such deep
learning models for different MTS analysis scenarios that achieve both
effectiveness and efficiency.
For RMTS analysis, the core challenge is modeling the complex multi-scale
temporal compositions. Addressing this challenge, we propose MSD-Mixer, a
multi-scale decomposition MLP-Mixer, which learns to explicitly decompose the
input RMTS into multi-scale temporal patches for modeling. Each layer employs
specialized multi-layer perceptrons to mix intra-patch, inter-patch, and
cross-channel features, while a novel residual loss ensures decomposition
completeness. MSD-Mixer beats state-of-the-art performance across forecasting,
imputation, anomaly detection, and classification tasks.
For IMTS classification, the primary obstacle is the severe channel-wise
asynchrony, which cripples standard attention mechanism. Overcoming this
limitation, we propose MTM, a multi-scale token mixing Transformer, which
mitigates the asynchrony via multi- scale down-sampling and introduces a
proactive token mixing mechanism that identifies and exchanges pivotal tokens
across unsynchronized channels, enabling superior IMTS classification accuracy
over existing methods.
For IMTS forecasting, where existing methods trade off between effectiveness
and efficiency, we propose EMIRA. It extends the efficient Mamba model for
IMTS forecasting with two key innovations: a context-aware selectivity that
integrates global irregularity context into the scan process, and a query
prompting mechanism that uses forecasting queries to proactively retrieve
relevant observations. EMIRA outperforms existing methods in both forecasting
accuracy and efficiency.
Date: Wednesday, 11 March 2026
Time: 3:00pm - 5:00pm
Venue: Room 2612A
Lifts 31-32
Chairman: Prof. David COOK (ECON)
Committee Members: Prof. Gary CHAN (Supervisor)
Prof. Andrew HORNER
Prof. Xiaofang ZHOU
Dr. Sisi JIAN (CIVL)
Dr. Wentao CHENG (BNBU)