Time Series Domain Adaptation via Latent Invariant Causal Mechanism

Area of Science:

• Machine Learning
• Causal Inference
• Time Series Analysis

Background:

• Time series domain adaptation seeks to transfer temporal dependencies from labeled to unlabeled domains.
• Existing methods use causal mechanisms of observed variables but struggle with high-dimensional data and indirect causal links.
• Challenges include modeling precise causal structures in complex, high-dimensional datasets like videos.

Purpose of the Study:

• To address limitations in current time series domain adaptation methods for high-dimensional data.
• To propose a framework for modeling causal mechanisms in latent variables for improved domain adaptation.
• To develop a method that guarantees the uniqueness and identifiability of reconstructed latent causal structures.

Main Methods:

• Proposed a latent causal mechanism identification framework by modeling causal mechanisms of temporal latent processes.
• Identified latent variables using historical information changes and enforced sparsity for identifiable latent causal structures.
• Developed the Latent Causality Alignment (LCA) model using variational inference with intra- and inter-domain latent sparsity constraints.

Main Results:

• The Latent Causality Alignment (LCA) model demonstrated improved performance on domain-adaptive time series classification and forecasting tasks.
• Achieved general improvement across eight benchmark datasets, validating the method's effectiveness.
• Successfully addressed challenges in modeling precise causal structures in high-dimensional time series data.

Conclusions:

• The proposed latent causal framework effectively captures domain-invariant temporal dependencies by modeling underlying latent causal structures.
• LCA offers a robust solution for real-world time series domain adaptation problems, particularly with high-dimensional data.
• The method enhances the applicability of causal inference techniques to complex time series datasets.