A New Approach to Contextual Suggestions Based on Word2Vec

We report our participation in the contextual suggestion track of TREC 2014 for which we submitted two runs using a novel approach to complete the competition. The goal of the track is to generate suggestions that users might fond of given the history of users’ preference where he or she used to live in when they travel to a new city. We tested our new approach in the dataset of ClueWeb12-CatB which has been pre-indexed by Luence. Our system represents all attractions and user contexts in the continuous vector space learnt by neural network language models, and then we learn the user-dependent profile model to predict the user’s ratings for the attraction’s websites using Softmax. Finally, we rank all the venues by using the generated model according the users’ personal preference.