A Linear Regression Model for Decentralized Search

Decentralized search is a classic problem that arises in message-passing between people or P2P networks and in information searching by following hyperlinks or paper citations. Previous approaches use node degree and similarity to perform decentralized search. However, there is an abundance of other locally-available information, such as path history and clustering coefficient, that has not previously been used in decentralized search algorithms. By combining various heuristics (including the recently-developed metric node2vec) in a linear regression model, we increased the success rate and decreased the average path length in both synthetic graphs and real-world networks, compared to previous work. We found that node2vec can enable similarity-based approaches to work in graphs that otherwise do not have a natural notion of similarity between nodes. Furthermore, in graphs (e.g., citation networks) that do have a natural notion of similarity, node2vec often provides a better similarity metric than the inherent similarity derived from the real-world meaning of the graph. Finally, we analyze the weights produced by the linear regression model to reveal insights about the decentralized search problem as a whole.