Label noise-oriented graph node classification method

• Main Authors: DONG SHOUBIN, HU JINLONG, CHEN LANG
• Format: Patent
• Language: Chinese; English
• Published: 20.11.2020
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention discloses a label noise-oriented graph node classification method. The method comprises the following steps of: 1) constructing a plurality of graph convolutional network (GCN) models; 2) estimating joint distribution on the graph data, pruning noise data to obtain primarily cleaned data, and training a graph convolutional network model; 3) training a graph convolution network modelon the graph data, predicting unmarked nodes together with the graph convolution network model in the step 2) to obtain a pseudo label and a prediction probability matrix, and jointly estimating full-data joint distribution and pruning noise data in combination with the label with the marked nodes and the prediction probability matrix; and 4) training a graph convolutional network model by using the graph data subjected to secondary cleaning obtained in the step 3), and predicting a node category. According to the invention, on the basis of a graph convolution network model, a noise joint distribution estimation and p