Network-based information filtering algorithms: Ranking and recommendation

Abstract

This chapter gives an overview of applications of random walks to information filtering, focusing on the tasks of ranking and recommendation in particular. Despite the amount of work done in these two directions, multiple important research challenges still remain open. Due to the massive amounts of available data, scalability of algorithms is of critical importance. Even when full computation is possible, one can think of potential approaches to update the output gradually when new data arrives. To achieve that, one can use or learn from perturbation theory which is a well-known tool in physics. It has been seen that results based on random walks often correlate strongly with mere popularity (represented by degree) of nodes in the network. Such bias toward popularity may be beneficial for an algorithm’s accuracy but it may also narrow one’s view of the given system and perhaps create a self reinforcing loop further boosting popularity of already popular nodes. Thus it is needed that information filtering algorithms converge less to the center of the given network. Random walks biased by node centrality or time information about nodes and links could provide a solution to this problem. As a beneficial side effect, this line of research could yield algorithms pointing us to fresh and promising content instead of highlighting old victors over and over again.