Herb Community Detection from TCM Prescription Based on Graph Embedding

Taking advantage of machine learning models, many researchers are exploring to dig out valuable information in large Traditional TCM(TCM) data. In view of the problems of existing TCM herb community detection methods, including poor flexibility, poor extensibility, poor performance of the herb network map, the difficulty in handling the network with small granularity, and the poor balance of detection results, this paper innovatively proposes a new herb community detection idea based on Graph Embedding. This idea mainly has three steps. The first step is constructing the TCM prescription network. The second step is mapping every herb node in the network to herb vector. The third step is using common vector clustering algorithm to get herb communities by dividing the network. In this paper, the second step is the core step. In order to reflect one-to-one and one-to-many relationship of herb nodes, this paper proposes two herb vector construction methods based on two Graph Embedding methods, respectively are matrix decomposition method and improved random walk method. In order to evaluate the experiment results, this paper proposes a comprehensive evaluation metrics which combining modularity, balance, and manual analysis and conducts experiments on relevant outpatient prescription record data. Experimental results show that the new herb community detection methods proposed in this paper has great performance in evaluation metrics than the traditional community detection algorithm, at the same time, the proposed vector construction method can also find potential new herb communities and help innovation of constructing prescription.