An Improved KNNG Driven Graph Transfer Diagnosis Method Via Edge Predictor

By establishing relationships between cross-domain samples, graph data-driven transfer diagnosis methods have achieved a great success. However, there are still some limitations: 1) Cross-domain relationships cannot be correctly established due to data distribution discrepancy, thereby degenerating graph quality; 2) Calculation of distance matrix consumes much computational cost. To overcome above limitations, an improved KNNG driven graph transfer diagnosis method via edge predictor is proposed. Firstly, few source-domain samples are used to construct initial K-nearest neighbor graph (IKNNG) for edge predictor training. Then, a pre-trained edge predictor, learning the graph construction strategy and restoring the graph structure of the input graph, directly outputs distance matrix without calculation of distance matrix. All the source-domain and target-domain samples are input to the pre-trained edge predictor, obtaining corresponding distance matrix. Based on the generated distance matrix, source-domain and target-domain samples are converted into source-domain and target-domain graphs, respectively. Both two kinds of graphs are used to train the Chebyshev graph convolutional network, achieving transfer diagnosis tasks. Comparative experimental results show that the proposed method achieves a competitive result.