Data Imputation with Iterative Graph Reconstruction

Effective data imputation demands rich latent ``structure" discovery capabilities from ``plain" tabular data. Recent advances in graph neural networks-based solutions show their structure learning potentials by translating as bipartite graphs. However, due to a lack of relations between samples, they treat all samples equally which is against one important observation: ``similar sample should give more information about missing values." This paper presents novel Iterative Generation and Reconstruction framework for Missing imputation(IGRM). Instead treating equally, we introduce the concept: ``friend networks" represent different among samples. To generate an accurate friend network with data, end-to-end reconstruction solution designed allow continuous optimization during learning. The representation optimized network, turn, used further optimize process differentiated message passing. Experiment results on eight benchmark datasets that IGRM yields 39.13% lower mean absolute error compared nine baselines 9.04% than second-best. Our code available at https://github.com/G-AILab/IGRM.