Addressing Label Noise in Causation Classification via Kernel Embeddings

A basic task of causal inference is to infer whether there exists a causeeffect relationship between two sets of vectors of interest, akin to a binary classification problem.With a sequence of independent and identically distributed paired vectors, one may employ the kernel mean embedding of probability distribution to map the empirical distribution to a feature space, and then train a classifier in the feature space to infer the causation for a future pair of vectors.This strategy, however, is susceptible to mislabeling, a common challenge in causation studies.In this paper, we explore this issue and quantify mislabeling effects.We develop valid learning methods with the mislabeling effects accounted for and theoretically justify the validity of the proposed methods.