Realization of Causal Representation Learning to Adjust Confounding Bias in Latent Space

Causal graphs are usually considered in a 2D plane, but it has rarely been noticed that within multiple relatively independent timelines, which is comparatively common in causality machine learning, the individual-level differences may lead to Causal Representation Bias (CRB). More importantly, such a blind spot has brought obstacles to interdisciplinary applications. Deep Learning (DL) methods overlooking CRBs confront the trouble of models' generalizability, while statistical analytics face difficulties in modeling individual-level features without a geometric global view. In this paper, we initially discuss the Geometric Meaning of causal graphs regarding multi-dimensional timelines; and, accordingly, analyze the scheme of CRB and explicitly define causal model generalization and individualization from a geometric perspective. We also spearhead a novel framework, Causal Representation Learning (CRL), to construct a valid learning plane (in latent space) for causal graphs, propose a particular autoencoder architecture to realize it, and experimentally prove the feasibility. Involved causal data includes Electronic Healthcare Records (EHR) to estimate medical effects and a hydrology dataset to forecast the environmentally influenced streamflow.