User Modeling for Contextual Suggestion

This paper describes our work on the Contextual Suggestion Track of the Twenty-Third Text REtrieval Conference (TREC 2014). The key to our approach is user interest modeling. By building explicit models of user interests and information needs, we are able to make suggestions relevant to the user. We extended our Reinforcement and Aging Modeling Algorithm (RAMA) to create user interest models using the rated examples in a user profile as explicit relevance feedback. Two models, one for specific interests and the other for general interests, are built for each user profile. To ensure that the recommendations are contextually appropriate, we have also built a simple model to capture contextual relevance of a recommendation. Candidate suggestions are retrieved from the Yelp® website using its application programming interface. For each candidate, we calculate three component scores based on the specific interest model, the general interest model, and the context model, respectively. Final scoring and ranking are computed as a weighted linear combination of the component scores. We hypothesize that the relative weighting of the components may affect the performance of our system. To test the hypothesis, we have submitted two runs with different weighting schemes. In particular, RUN1 has a specific interest priority whereas RAMARUN2 has a general interest priority. TREC evaluation reveals that both runs performed significantly better than the median of all submitted runs (i.e., the Track Median) on three performance metrics. In addition, RAMARUN2 has a slight performance edge over RUN1. The effectiveness of our approach is evidenced by the TREC evaluation result that RAMARUN2 and RUN1 ranked #2 and #6 out of the 31 runs submitted by the 17 participating teams from around the world.