Addressing Information Asymmetry: Deep Temporal Causality Discovery for Mixed Time Series

Area of Science:

• Computational neuroscience
• Machine learning
• Causal inference

Background:

• Causal discovery methods primarily address continuous time series.
• Mixed time series (continuous and discrete variables) present unique challenges due to nonlinearity and high dimensionality.
• Discrete variables often arise from underlying continuous processes, losing information through discretization.

Purpose of the Study:

• To develop a novel framework for causal discovery in mixed time series data.
• To address the challenges posed by nonlinearity, high dimensionality, and the mixed nature of variables.
• To recover the intrinsic continuity of discrete variables by leveraging their relationship with continuous variables.

Main Methods:

• Proposed a generic deep mixed time series temporal causal discovery framework.
• Developed a contextual adaptive Gaussian kernel embedding technique for latent continuity recovery.
• Employed a two-stage training process for self-supervised latent continuity recovery and sparsity-induced causal structure learning.

Main Results:

• The proposed framework successfully recovers latent continuous variables from discrete data.
• Causal discovery is performed in a unified continuous-valued space, integrating information from both continuous and discrete variables.
• Experimental evaluations demonstrate the superior performance of the framework compared to existing methods.

Conclusions:

• The framework effectively handles mixed time series data for causal discovery.
• Recovering latent continuous variables is crucial for accurate causal inference in mixed data.
• This approach offers a significant advancement for understanding complex systems with heterogeneous data types.