Reviews: Learning Supervised PageRank with Gradient-Based and Gradient-Free Optimization Methods

This paper analyzes the convergence of a non-convex loss-minimization problem for learning the parameters of a general graph-based ranking model, that is defined by a random walk conducted by weights of nodes and edges, which are in turn defined by random walks defined by nodes' and edge's features. The optimization problem can not be solved by existing optimization methods which require exact values of the objective function. The proposed approach hence operates in two level. At the first level, a linearly convergent method is used to estimate an approximation to the stationary distribution of Markov random walk. This approach is validated among others and the authors show the value of the loss function can be approximated with any given precision.