Graph Neural Networks for Missing Value Classification in a Task-Driven Metric Space

Incomplete instances with various missing values in real-world scenes have brought challenges to the classification tasks. Many existing methods impute the incomplete instances based on their neighbouring instances before classification. However, the construction of such neighbourhood relationships in the original data space may become unreliable since missing values could disturb the traditional distance measurement. Moreover, these methods decouple the imputation from the classification, which makes it difficult for the former to learn from the supervised information, resulting in sub-optimal performance. To this end, this paper proposes graph neural networks for missing values classification (GNN4MV), which directly classify the incomplete instances based on their neighbourhood relationships constructed in a novel task-driven metric space. Specifically, the supervised information is taken as additional guidance in a task-driven metric space to reduce the impact of the missing values for neighbourhood relationship construction. Furthermore, a novel neighbourhood graph convolutional network is proposed in GNN4MV, which enables direct classification of the incomplete instances without imputation by utilizing the graph topology in the constructed neighbourhood relationships. Experiments on real-world datasets demonstrate the robustness and effectiveness of the proposed algorithm.