Self-Optimizing Feature Transformation

Abstract

Feature transformation aims to extract a good representation (feature) space by mathematically transforming existing features. It is crucial to address the curse of dimensionality, enhance model generalization, overcome data sparsity, and expand the availability of classic models. Current research focuses on domain knowledge-based feature engineering or learning latent representations; nevertheless, these methods are not entirely automated and cannot produce a traceable and optimal representation space. When rebuilding a feature space for a machine learning task, can these limitations be addressed concurrently? In the preliminary work, we present a self-optimizing framework for feature transformation that employs three cascading reinforced agents to learn effective policies for selecting candidate features and operations for feature crossing. While it has achieved good performance, the effectiveness of the reconstructed feature space can be improved by: (1) obtaining an advanced state representation for enabling reinforced agents to comprehend the current feature set better; and (2) resolving Q-value overestimation in reinforced agents for learning unbiased and effective policies. Thus, in the journal version, to fulfill goal 1, it is assumed that an effective state representation should not only contain the entire feature set knowledge, but also capture feature-feature correlations. First, we build a complete graph based on feature-feature correlations. Then, using message-passing, we aggregate feature knowledge based on the learned graph structure to obtain a graph embedding. The graph embedding can recover the original graph, which is the state representation. To fulfill goal 2, we undertake Q-learning using two Q-networks: one for action selection and the other for Q-value estimation. We divide the learning process in each Q-Network into two streams: state value estimation and action advantage value estimation. The Q-value is the combination of the two streams' estimations. Finally, to make experiments more convincing than the preliminary work, we conclude by adding the outlier detection task with five datasets, evaluating various state representation approaches, and comparing different training strategies. Extensive experiments and case studies show that our work is more effective and superior.