Modelling Psychological Needs for User-dependent Contextual Suggestion

This paper presents our approach for the Contextual Suggestion track of 2014 Text REtrieval Conference (TREC). The task aims to provide recommendations on points of interests (POI) for various kinds of users under different contexts. This becomes challenging due to the limited amount of training data provided by TREC and the demanding constraints for a suggestion to be judged as relevant. Our approach does not deviate from existing Machine Learning based methods in principle, but sticks closely to the defined relevance judgement criteria, by focusing primarily on modelling users’ preferences on POI categories, and investigating upon their psychological expectations on the textual descriptions of the POIs. The latter is considered as our novelty in this work. Support Vector Regression was used for suggestion ranking, an ad-hoc web information extractor was used to collect POI descriptions, and a description evaluation mechanism was engaged to select proper POI descriptions subject to the nature of the POIs. Our results suggest that our methods are effective in obtaining satisfying user-specific POI rankings and generating descriptions that meet users’ psychological expectations.