Attribute-based feature learning for network-structured data

Networks are used widely in various systems because of their capacity to capture the interactions (edges) among data (nodes) in various real-world settings. This ability has led to the widespread adoption of networks. In the field of machine learning, constructed networks may be used to produce different predictive-based analyses. However, current machine learning methods are based on the assumption that data instances are independent of each other and prevent network information from being included in machine learning due to the irregular and variable-size nature of network structured data. Recently, feature learning methods have been proposed that aim to learn a low-dimensional vector representation for each node while preserving the structural information in the network. These techniques only use the network's topology information and do not consider the node attributes. However, real-world networks often have large numbers of nodes and high-dimensional node attributes. This study demonstrates how node2vec, which adopts the random walk principle, can be extended to take advantage of node attributes as well as topology information. With node2vec, a random walk is represented as an array of node identifiers. To utilize attributes, it is expressed as an array of feature vectors. The node2vec-attribute model is proposed for networks with a single categorical attribute and the node2vec-attribute+ model is for multi-attribute networks. Various experiments were performed on real-world datasets to evaluate the effectiveness of the proposed approaches on the similarity of structural roles and node classification. All the experimental studies performed prove the effectiveness of each algorithm. Overall, the presented results demonstrate the effectiveness and robustness of the proposed models in capturing and encoding network-related node content information.