Improving the quality of recommendation using semi-structured user feedback
Résumé rédigé par
Directeur de thèse:
Unité de recherche INRIA 0 Institut National de Recherches en Informatique et en Automatique
Recommendation systems mainly rely on two forms of explicit user feedback: free form text or stars rating. Star rating does not reveal why the user likes or doesn’t like the rated item; this information is indeed mandatory to perform a reliable recommendation. On the other extreme, making recommendation from free form text (what we call unstructured reviews) is extremely complex and requires supervised techniques such as sentiment analysis. Semi-structured reviews (i.e. short sequences of words describing user’s experience), or “tags”, would give more detailed information than star ratings and would eliminate the complex and error-prone filtering of the free-form reviews. The objective of this thesis is to (1) build a system (android app) that will enable users to enter tags about anything they want to review in the most possible intuitive way, (2) distribute the app in order to collect large amount of tags, (3) analyze the data collected to evaluate different recommendation techniques and compare to state of the art, (4) improve recommender system to provide more accurate recommendations based on our semi-structured reviews, and (5) compare to state-of-the-art recommender systems. One major improvement consists in recommending users what tags they are most likely to enter based on their other reviews and based on what other people who reviewed the same item said. Note that because these semi-structured reviews do not exist today, we will have to design an algorithm to summarize reviews in a small number of tags. This is by itself a difficult challenge.