Recommendations for the long tail by term-query graph

We define a new approach to the query recommendation problem. In particular, our main goal is to design a model enabling the generation of query suggestions also for rare and previously unseen queries. In other words we are targeting queries in the long tail. The model is based on a graph having two sets of nodes: Term nodes, and Query nodes. The graph induces a Markov chain on which a generic random walker starts from a subset of Term nodes, moves along Query nodes, and restarts (with a given probability) only from the same initial subset of Term nodes. Computing the stationary distribution of such a Markov chain is equivalent to extracting the so-called Center-piece Subgraph from the graph associated with the Markov chain itself. Given a query, we extract its terms and we set the restart subset to this term set. Therefore, we do not require a query to have been previously observed for the recommending model to be able to generate suggestions.