CAGAIN: Column Attention Generative Adversarial Imputation Networks

Jun Kawagoshi, Yuyang Dong, Takuma Nozawa, Chuan Xiao

Abstract: Imputation for missing values is a key operation in building data analysis models. In this paper, we target numerical and categorical values in tabular data. While previous studies have demonstrated the effectiveness of state-of-the-art methods, a major limitation is that these methods lack robustness and their performance significantly varies across datasets and the missing rate of values, hence posing considerable overhead of selecting and tuning models in a real-world scenario. To tackle this problem, we propose a Column Attention Generative Adversarial Imputation Network (CAGAIN), an imputation model which employs a generative adversarial network (GAN) and the attention mechanism. The generator of CAGAIN mimics the distribution of original data and generates imputed samples similar to real ones. The discriminator of CAGAIN distinguishes real and generated samples, so as to improve the quality of the imputed data. At the same time, the attention mechanism captures the correlation between attributes and focuses on the most significant attributes that determine the values of the missing positions. By inheriting the advantages of GAN and the attention mechanism, our model is endowed with robustness to shifting datasets and missing rates, which is demonstrated by experiments using 9 real datasets.