Significant Words Language Models for Contextual Suggestion

In this paper, we present the participation of the University of Amsterdam’s ExPoSe team in the TREC 2016 Contextual Suggestion Track. The main goal of contextual suggestion track is to evaluate methods for providing suggestions for activities or points of interest to users in a specific location, at a specific time, taking their personal preferences into consideration. One of the key steps of contextual suggestion methods is estimating a proper model for representing different objects in the data like users and attractions. Here, we describe our approach which is employing Significant Words Language Models (SWLM) [2] as an effective method for estimating models representing significant features of sets of attractions as user profiles and sets of users as group profile. We observe that using SWLM, we are able to better estimate a model representing the set of preferences positively rated by users as their profile, compared to the case we use standard language model as the profiling approach. We also find that using negatively rated attractions as negative samples along with positively rated attractions as positive samples, we may loose the performance when we use standard language model as the profiling approach. While, using SWLM, taking negatively rated attractions into consideration may help to improve the quality of suggestions. In addition, we investigate groups of users sharing a property (e.g. of a similar age) and study the effect of taking groupbased profiles on the performance of suggestions provided for individual users. We noticed that group-based suggestion helps more when users have a tendency to rate attraction in a neutral way, compared to the case users are more subjective in their rating behavior.