关键词: 关系预测, 深度学习, 网络结构, 节点属性

Abstract: In reality, social systems from various domains can be effectively characterized through network models, often exhibiting structural properties distinct from random networks, such as small-world and scale-free characteristics. The formation of these non-trivial structural properties is closely associated with the establishment of relationships (i.e., links) among individuals (i.e., nodes) in the network. Consequently, accurately predicting potential relationships in the network not only helps deepen our understanding of the underlying mechanisms driving network formation but also further elucidates the relationship between network topology and system function. Thus, the prediction of links between nodes has become an important research problem in the field of network science. For link prediction, a commonly used method is heuristic algorithms based on similarity. However, in more complex network scenarios, such methods struggle to effectively address high-dimensional non-linear problems resulting from network scale expansion or node feature growth. In recent years, the emergence of deep learning-based approaches has provided new opportunities by transforming complex network information into low-dimensional representation vectors. However, most existing deep learning-based approaches primarily achieve link prediction through the similarity of embedding representation vectors of network structures. Many empirical studies indicate that the formation of links in the network is influenced by node attributes, and similarity alone is not the sole criterion for link formation. Therefore, the link prediction approach based on deep learning is worth further exploration.
In this paper, we propose a deep walk-deep neural network for link prediction (DDLP) model, which integrates network structure and node attribute information for link prediction. This model consists of two stages, i.e., the stage of node feature embedding and the stage of link prediction. In the first stage, network structure information is embedded using deep walks. Then, to obtain node feature vectors, the embedded structure feature vectors are merged with standardized node attribute feature vectors through early fusion. In the second stage, a deep learning model is constructed to capture the link patterns between node feature vectors through supervised learning, thereby achieving relationship prediction. We select real network data from three different domains, including open-source software development, patent research and development, and scientific collaboration, to examine the effectiveness of the model. Additionally, in the experimental sample networks, we compare the predictive performance of the proposed model with traditional methods such as common neighbors (CN) and resource allocation (RA), deep learning methods that only consider node structural information like deep walk and node2vec, as well as models that can incorporate node attributes like variational graph auto encoders (VGAE) and graph convolutional networks (GCN).
The results show that the DDLP model, based on node feature embedding, effectively captures the distribution patterns of links in the network. Its performance (precision, recall, and F1 score) significantly surpasses that of traditional models based on vector similarity (such as CN and RA) and deep learning models such as node2vec and VGAE. Furthermore, compared to predictive methods that only incorporate network structural information, the integration of node attributes has significantly enhanced the predictive capabilities of both the DDLP model and comparative models such as VGAE and GCN. Particularly, the DDLP model has the highest performance metrics, indicating that the incorporation of node attributes allows it to learn a richer set of rules for link formation, thereby offering superior performance. This also further reveals that it is not enough to predict the link formation only by the similarity of node vectors, and there is a need for more refined processing to enable the model to better learn the rules of link formation within networks.
This study not only proposes a deep learning framework that integrates network structure and node attribute information for link prediction but also lays the methodological foundation for related applications, such as system recommendations. In future work, we will explore the portability of the framework to other network analysis tasks such as mechanism analysis of the network formation, link prediction in heterogeneous networks, etc. through more extensive experiments. In addition, we intend to optimize the DDLP model to reduce computational complexity, making it more suitable for link prediction in ultra-large-scale networks.

Key words: link prediction, deep learning, network structure, node attribute