Handling Missing Data in Graph Networks with Adaptive Graph Convolutional Network

With recent advancements in social networks and knowledge graphs, graph data has become more popular and the use of Graph Convolutional Networks (GCN) in graph analysis tasks has been a great success. GCN model assumes all features of all nodes are available and it works by smoothening node features. But, due to privacy-protection policies, copyright restrictions, and sensor malfunctioning during data collection, real-world graphs suffer from incompleteness and missing features. Conventional methods could only produce inconsistent and unstable results, the main reason for which is, that feature imputation and graph learning are unrelated. To address this gap there are network models that are capable of imputing node feature information and learning the graph within the same architecture. GCN with missing features (GCN
MF
) is a modified GCN model that outperformed conventional methods for node classification. The proposed method utilizes the capabilities of GCN
MF
while producing a better estimate of missing features. This is achieved by performing an unsupervised clustering of the graph data using K-means and applying GCN
MF
for each cluster. Feature matrix from each cluster is then combined and passed through a GCN model. Experiments conducted over the model with two popular graph datasets for node classification tasks produced a better accuracy score