Résumé: The contextual suggestion task is how to generate better recommendations to users with prior knowledge including user background, preference history, city location, time of the year and many other contextual information. In this talk, we will introduce our approach to the task at TREC 2016. We have two major models. The first model is a global interest regressor trained to model popular interests loved by all users (E.g. Museums and National Parks). The second model introduces word embeddings to captures individual user preference. Both user profiles and candidate places are represented as word vectors in a same Euclidean space. So that a similarity score between user and attraction can be measured by their vector distance. The combined model of the two gained the highest Precision at 5 among all other track results.
Note: Cette présentation sera donnée en anglais.