Abstract: Aiming at the limitation of the simplifying the processing of homophily relational classifiers based on first-order Markov assumption, when constructing the class vector and reference vector in the class-distribution relational neighbor classifier, we introduced the activation spreading algorithm of local graph ranking, combined with the relaxation labeling collective inference method. By appropriately expanding the range of neighboring nodes during classification, we increased the homophily of nodes to be classified in network data, thereby reducing the error rate of classification. The comparative experimental results show that this method expands the  neighborhood of nodes to be classified, and has good classification accuracy  on network data.

Key words: artificial intelligence, network data classification, activation spreading, class-distribution relational neighbor classifier, collective inference