Node Similarity-based Search Method for Medical Insurance Heterogeneous Information Network

Abstract

Among the current medical insurance fraud detection methods, fraudulent nodes that can be detected mainly depend on empirical rules to determine medical insurance fraud patterns or a large amount of fraud data for machine learning. However, since the increasing concealment of fraud, rules and models designed based on empirical rules cannot cope with the rapidly changing fraud patterns. At the same time, the labels of fraudulent samples in medical insurance data are small in magnitude and unevenly distributed, making it challenging to support active mining. In addition, if the classical meta-path or meta-structure is used to model the medical insurance heterogeneous information network, patient information will be lost when connecting multi-order paths such as patient-physician-disease and bring ambiguous information that does not match the actual medical insurance data. In order to solve the above problems, this paper proposes a node similarity-based search method for medical insurance heterogeneous information network. On the one hand, the method uses GraphSAGE to learn the global low-dimensional representation of patient nodes and recalls nodes relevant to the query node as search candidate sets. On the other hand, by defining weighted meta-path and weighted meta-structure, the method solves the problem of ambiguity in the representation of heterogeneous information network. Based on weighted meta-path and weighted meta-structure, new algorithms W-Pathsim and W-Strucsim are proposed to calculate the similarity of nodes in heterogeneous information network. Finally, our method uses a multi-layer perceptron to return the nodes list that is highly similar to the query node in the candidate set to assist in medical insurance review. Experiments show that our method is better than the compared baseline methods.