Unsupervised multi-graph propagation for ranking based on order change tendency

Graph ranking has attracted remarkable attention in ranking field since it exploited the cluster and connectivity assumption. However, itpsilas difficult to combine inhomogeneous features into one graph, such as the link relation and content similarity in text retrieval. To address the above problems, a novel multi-graph propagation algorithm for ranking named MGP is proposed, which can be applied to both directed and undirected graphs. It is implemented by constructing multiple graphs in inhomogeneous views and combining results to maximize the agreement of multiple graphs. Compared to existing multi-view learning approaches using labeled data as agreement, MGP introduces order change tendency as agreement substituting for label information, which also ensures the combination being uniform. The theoretical analysis on the convergence of MGP is given. Experimental results on Cora and CiteSeer data sets indicate that MGP can make use of inhomogeneous features sufficiently to enhance the ranking performance.