BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260624T153736EDT-6697EuBlHb@132.216.98.100
DTSTAMP:20260624T193736Z
DESCRIPTION:Abstract\n\nTime-series analysis and sequential machine learnin
 g have emerged as fundamental pillars of modern data science\, with applic
 ations spanning quantitative finance\, weather prediction\, electricity ma
 nagement\, healthcare monitoring\, and industrial process control. The inc
 reasing complexity and scale of temporal data necessitate sophisticated me
 thodologies that can capture intricate patterns\, long-range dependencies\
 , and nonlinear dynamics. This thesis addresses fundamental challenges in 
 sequential machine learning by proposing novel architectures and methodolo
 gies that advance the state-of-the-art in time-series modeling\, forecasti
 ng\, and representation learning. The main contributions of this thesis ar
 e organized into three primary categories.\n\nFirst\, we propose a novel M
 ulti-resolution Time-Series Transformer (MTST) architecture for multivaria
 te time series forecasting. This framework employs a multi-branch architec
 ture that simultaneously models diverse temporal patterns at different res
 olutions by adjusting patch-level tokenization\, enabling the capture of b
 oth short-term fluctuations and long-term seasonal trends. Unlike previous
  works that rely on subsampling\, MTST constructs multi-resolution represe
 ntations through different patch sizes\, with each branch processing tempo
 ral patterns at distinct frequencies. The architecture employs relative po
 sitional encoding\, which is naturally aligned with capturing periodic tem
 poral patterns. Extensive experimental evaluation demonstrates that MTST a
 chieves state-of-the-art performance across seven benchmark datasets and f
 our prediction horizons\, outperforming previous patch-based transformers 
 with statistical significance in the majority of cases.\n\nSecond\, we est
 ablish SKOLR\, a novel approach that connects Koopman operator theory with
  linear Recurrent Neural Networks. By leveraging an extended state space o
 f lagged observations\, we demonstrate an equivalence between structured K
 oopman operators and linear RNN updates\, enabling the development of fore
 casting architectures that combine theoretical rigor with computational ef
 ficiency. SKOLR implements a structured Koopman operator through a highly 
 parallel linear RNN stack\, where learnable spectral decomposition of the 
 input signal allows different RNN chains to attend to different dynamical 
 patterns from different representation subspaces. The resulting architectu
 re achieves exceptional performance on various forecasting benchmarks and 
 dynamical systems\, demonstrating superior capabilities in handling both s
 hort-term and long-term forecasting tasks across diverse temporal patterns
 .\n\nThird\, we introduce GraphTNC\, a framework for learning joint repres
 entations of graph-structured time series through contrastive learning. Th
 e framework addresses the challenge of unsupervised representation learnin
 g for multivariate time-series data\, particularly in settings where the d
 ata exhibits graph-structured relationships that evolve over time. GraphTN
 C incorporates both temporal smoothness and graph-structured relationships
  into the contrastive learning objective\, assuming piecewise smooth dynam
 ics in both time-series and graph evolution. This enables joint learning o
 f graph and temporal representations that can be effectively utilized for 
 downstream tasks such as classification. Experimental results demonstrate 
 that GraphTNC learns meaningful representations that improve performance o
 n various graph-structured time-series tasks.\n\nCollectively\, these cont
 ributions advance both the theoretical understanding and practical capabil
 ities of time-series modeling\, with demonstrated improvements in forecast
 ing accuracy\, computational efficiency\, and representation quality acros
 s diverse benchmark datasets and application domains.\n
DTSTART:20260604T170000Z
DTEND:20260604T190000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Yitian Zhang – Advanced Sequential Machine Learning 
 Models for Time-Series Signals
URL:https://www.mcgill.ca/ece/channels/event/phd-defence-yitian-zhang-advan
 ced-sequential-machine-learning-models-time-series-signals-373046
END:VEVENT
END:VCALENDAR