Projection-Wise Disentangling for Fair and Interpretable Representation Learning: Application to 3D Facial Shape Analysis

Confounding bias is a crucial problem when applying machine learning to practice, especially in clinical practice. We consider the of representations independent multiple biases. In literature, this mostly solved by purging information from learned representations. however expect strategy harm diversity representation, and thus limiting its prospective usage (e.g., interpretation). Therefore, we propose mitigate while keeping almost all latent representations, which enables us observe interpret them as well. To achieve this, project features onto vector direction, enforce independence between biases projected rather than features. mapping input data, projection-wise disentangling: sampling reconstruction along direction. The proposed method was evaluated on analysis 3D facial shape patient characteristics (N = 5011). Experiments showed that conceptually simple achieved state-of-the-art fair prediction performance interpretability, showing great potential for applications.