Deconfounded classification by an intervention approach

For an automatic system, image classification is much challenging, partly due to noises that obscure or reduce the clarity of data. Thus, noise suppression has become one of the core tasks of classification. There is often a key assumption in classification that any noise source only affects one side of images and labels. However, this assumption overlooks the confounding scenario that the same noise source affects both images and labels. In this paper, we propose an intervention approach to learn a deconfounded classification model. Classification problem is firstly formulated as causal inference, where intervention is used to untangle the causal from the correlatives, and derive a causal effect formula for deconfounded classification. The WAE (Wasserstein Auto-Encoder) objective is then expanded for classification, with a new regularizer defined for learning unobserved confounders. To build a robust network architecture, a probability factorization is performed in conjunction with d-separation rule to find useful dependency patterns in data. The deconfounded classification model is finally obtained by rearranging the components of the learnt decoder according to the causal effect formula. The experimental results demonstrate our approach outperforms significantly the existing state-of-the-art classification models, particularly on imbalanced data.