Multi-View Denoised Graph Collaborative Filtering for Third-Party Library Recommendation

International Joint Conference on Neural Networks (IJCNN), pp. 1–8

Abstract

Third-party libraries are crucial for accelerating software development, yet identifying suitable candidates within the expanding open-source ecosystem remains challenging. Graph neural network (GNN)-based collaborative filtering emerges as a promising solution due to its ability to effectively model high-order interaction relationships. However, data sparsity leads to biased representations, hindering effective node representation learning. Moreover, standard multi-layer convolutions blend irrelevant node features, causing noise accumulation and suboptimal embeddings. To address these challenges, a Multi-View Denoised Graph Collaborative Filtering (MDGCF) approach is proposed for TPL recommendation. Specifically, MDGCF constructs complementary graph views to alleviate sparsity and employs an implicit supervision module to guide representation learning with auxiliary signals. To mitigate noise, we propose a denoised graph convolution that calculates normalized weights for even-order nodes to suppress irrelevant propagation, alleviating noise accumulation. Furthermore, a sampled softmax loss with graph topology-based negative sampling is carefully designed to replace the conventional BPR loss, strengthening the differentiation between positive and negative samples and enhancing recommendation accuracy. Experiments on a real-world dataset demonstrate that MDGCF outperforms state-of-the-art baselines, validating the model’s effectiveness in improving the performance of TPL recommendation.

Authors 7

1. Yuxiang Kuang

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST

2. Xiaoxiang Liao

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST

3. Guosheng Kang

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST

4. Ye Cao

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST

5. Xiao Cui Aachen

   Leiden University

   Affiliation as printed

   Leiden University

6. Rong Hu

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST

7. Jiayan Xiang

   Hunan University of Science and Technology

   Affiliation as printed

   HNUST