Homogeneous graph neural networks for third-party library recommendation

During mobile application development, developers often use various third-party libraries to expedite the development process and enhance application functionality. Real datasets often show significant long-tailed distribution characteristics, where a few third-party libraries are widely adopted, while most are seldom used, leading to extreme data sparsity. This distribution phenomenon challenges recommendation algorithms, which typically recommend widely used third-party libraries for basic functionality, failing to meet developers' specific feature needs. To address these limitations, we propose HGNRec, a third-party library recommendation model based on a homogeneous graph neural network. First, to overcome the limitations of fusing heterogeneous node information, we decompose the heterogeneous graph network into two homogeneous graph networks using a statistical method. Second, the two constructed GNN models use separate aggregation and nonlinear transformation network structures for adaptive aggregation, along with edge-level and feature-level constraint methods to optimize model performance. In homogeneous graph networks, low-order and high-order neighbor information of nodes are propagated and aggregated in the same knowledge space, capturing the complex interactions among homogeneous nodes. Furthermore, we validate the superiority of HGNRec compared to several state-of-the-art methods using real datasets. Source code will be available at https://github.com/dacilab/HGNRec.