In the era of widespread use of smartphones, researchers have been exploring new detection methods to cope with the threat of constantly upgrading Android malware. With the emergence of graphical neural networks, people’s interest in using this technology to detect Android malware has surged. However, external connections between Android malware are often overlooked, and the graph structure is filled with redundant information. To address these challenges, this article introduces AnGSL, a graph based method designed specifically for Android malware detection. This method first uses the MTF-IDF algorithm to extract feature entities, and then calculates the Euclidean distance between the feature vectors of two Android applications to determine whether there are edges between them. Subsequently, two subgraphs were constructed using the K-nearest neighbor and diffusion matrix algorithms, and optimized and merged to obtain the final graph structure. Finally, by minimizing the cross entropy loss between the subgraph and the optimized graph, a contrastive learning method is used to train the model. Experimental evaluations conducted on publicly available Android datasets showed a detection accuracy of 98.24%. Research has shown that compared with previous model methods, the method proposed in this paper has a certain degree of performance improvement, indicating that graph based Android malware detection has broad potential and prospects in the research field.