Grier: graph repairing based on iterative embedding and rules

Identifying and repairing the erroneous and missing information in the graphs is crucial. Rule-based studies rely on graph quality rules to detect the inconsistencies among different entities and make a one-time repair to resolve these inconsistencies. The missing data can be predicted and imputed by graph embedding methods by preserving the inherent structure of a reliable graph. However, both lines of way need more evidence to improve. In this paper, we propose Grier, a novel repairing method to enrich the evidence by taking advantage of rule-based repairing and graph embedding. Specifically, Grier iteratively learns the graph embeddings with guidance from logical rules, which has significant power in knowledge acquisition and inference. The erroneous triples are detected and repaired by evaluating the correctness of the triples in the embedding space. The newly repaired triples are used as training data to update the embedding module for better learning. Extensive experiments on three graphs demonstrate the effectiveness of Grier even with very few rules.