LSA as Ground Truth for Recommending “Flickr-Aware” Representative Tags

Abstract

Most item recommendation systems nowadays are implemented by applying machine learning algorithms with user surveys as ground truth. In order to get satisfactory results from machine learning, massive amounts of user surveys are required. But in reality obtaining a large number of user surveys is not easy. Additionally, in many cases the opinions are subjective and personal. Hence user surveys cannot tell all the aspects of the truth. However, in this paper, we try to generate ground truth automatically instead of doing user surveys. To prove that our approach is useful, we build our experiment using Flickr to recommend tags that can represent the users’ interested topics. First, when we build training and testing models by user surveys, we note that the extracted tags are inclined to be too ordinary to be recommended as “Flickr-aware” terms that are more photographic or Flickr-friendly. To capture real representative tags for users, we apply LSA in a novel way to build ground truth for our training model. In order to verify our scheme, we define Flickr-aware terms to measure the extracted representative tags. Our experiments show that our proposed scheme with the automatically generated ground truth and measurements visibly improve the recommendation results.

This work (No. 2011-0029729) was supported by Mid-career Researcher Program through NRF grant funded by the MEST.