A Novel Community Detection Algorithm Based on Deep Learning Algorithm

With the extremely rapid growth of data scale, feature reduction in communities is gradually gaining attention. Community detection methods based on deep learning can discover deep network information and complex relationships, and better handle high-dimensional data. In this paper, we apply deep learning to community detection by using a deep sparse autoencoder to reduce the dimension of the input matrix of the network graph and extract data features from it, and continuously reduce the reconstruction error until the optimal solution is found to achieve the classification of communities. The community results obtained from K-means clustering using adjacency matrix directly, K-means clustering using the similarity matrix, K-means clustering using the feature matrix after dimensionality reduction by the autoencoder, and K-means clustering using the feature matrix after dimensionality reduction by the sparse autoencoder are compared. Experiments on Strike, Football and Karate network datasets were conducted to test the community detection results under the above four methods, and the detection results were analyzed using three evaluation metrics, such as NMI. In addition to hop count threshold, there are parameters such as decay factors that have an impact on the community detection results, so control variate method is used to explore effect on experiments when the parameters take various values.