Conditional Independence-based Causal Discovery Algorithm for Few-Shot Feature Selection
Ying Yang, Lianglun Cheng, Guoheng Huang · International Conference on Frontiers of Electronics, Information and Computation Technologies · 2021
Feature selection can select the optimal subset of features by eliminating the irrelevant or redundant features, which has been an important challenge in machine learning tasks. Existing neural network methods for data processing often requires large number of features to be filtered, extracted, and transformed. However, the characteristics of few-shot are usually accompanied by high dimensionality and sparse sample size, which making it arduous for neural network methods to locate the truly relevant features. Consequently, a causal feature selection layer based on conditional independence is proposed to embed the traditional neural network. In particular, effective features can be screened out by mining the causal relationships through conditional independence, and it makes the data modeling with certain interpretability. Experimental results show the feasibility of our theory on both simulated and real dataset.