A Non-Isomorphic-Discriminated Inductive Graph-based Matrix Completion Model for High-Dimensional and Incomplete Data

High-dimensional and incomplete (HDI) data are generally used to analyze the potential links between real entities in the big data era. To perform this key task, an inductive graph-based matrix completion (IGMC) model is proposed to predict the link ratings w/o using the side information and pre-training. However, the isomorphic representation in the IGMC model only promotes the representation accuracy of the links with the adjacent ratings while does not consider the link representation accuracy of non-adjacent ratings. To address this issue, this paper proposes a local Non-Isomorphic-Discriminated IGMC (NID-IGMC) model to represent the links with distinct rating values more exactly. The main idea of the NID-IGMC model is two-fold. First, the NID-IGMC model proposes the definition of the non-isomorphic representation, constructs its representative functions and divides into three sub-categories. Second, the NID-IGMC model incorporates the non-isomorphic representation into the isomorphic representation and re-construct the model. Two kinds of experiments on industrial HDI matrices are performed, which indicate the high accuracy of the proposed NID-IGMC model.