Multi-Dimensional Causality Fairness Learning

Abstract

Causal learning is a recent and widely adopted paradigm to handle algorithmic discrimination. Contemporary causality-based studies on fairness only capture the unfair causal effect of a single-dimensional sensitive attribute (i.e., individual-dimension, like gender) on the decision. They neglect the socially constructed nature of individual attributes, such as macro-dimensional factors. However, social science research shows that discrimination against an individual may be related to disadvantaged treatments, which operate at the macro-dimension (e.g., neighborhood economic level). This multi-dimensional conceptualization is pertinent to matters of fairness, and it is crucial to be fair for individuals across multiple dimensions. The hidden confounder is another bottleneck for addressing fairness concerns based on causal techniques. To tackle these issues, we present an approach, called MultiCFL, which accounts for multi-dimensional sources of discrimination and unifies them via causal tools. To handle hidden confounders, MultiCFL first trains a causal effect variational autoencoder as the causal estimator to learn the causal mechanisms behind observational data. Subsequently, it makes selective use of estimated causal relationships to construct a predictive model with multi-dimensional fairness. Experimental results confirm the effectiveness of MultiCFL, and prove the necessity of considering multiple dimensional properties to mitigate unfairness.