RAA#2: A Semantic Content-Based Recommender System Integrating Folksonomies for Personalized Access

1. APA Citation: Lops, P., Gemmis, M., Semeraro, G., Musto, C., Narducci, F., & Bux, M. (2009). A Semantic Content-Based Recommender System Integrating Folksonomies for Personalized Access. In G. Castellano, L. C. Jain, & A. M. Fanelli (Eds.), Web Personalization in Intelligent Environments (Vol. 229, pp. 27-47). Berlin, Heidelberg: Springer Berlin Heidelberg. Retrieved from http://www.springerlink.com.login.ezproxy.lib.purdue.edu/content/e6603207661850m7/

2. Problem Statement: Traditional content-based personalization (recommender system) is usually driven by string (keywords) matching operations, i.e., to match attributes of user profiles with the attributes of content objects. This method is unable to capture the semantics of the user interests behind the keywords.

3. Purpose of the Research: This paper aims at improving recommender system by incorporating user ratings and tags, and also utilizing semantic analysis techniques derived from research in Information Filtering, Machine Learning, and Natural Language Processing.

4. Research Question: Does the integration of tags cause an increase of the prediction accuracy in the process of filtering relevant items for users?

5. Some Definitions: Static Content, SocialTags(I), PersonalTags(U,I)

Static Content: compared with the dynamic tags provided by the users, the title, author, original descriptions of the items are static content.

SocialTags and PersonalTags: Given an item I, the set of tags provided by all the users who rated I is denoted as SocialTags(I), while the set of tags provided by a specific user U on I is denoted by PersonalTags(U, I).

6. Methods:

(1) A system called FIRSt (Folksonomy-Based Item Recommender syStem) is designed integrating semantics analysis algorithms and User Generated Content (interest ratings and tags). FIRSt allows users to give an interest rating (1-5, 1=strongly dislike, 5=strongly like ) and free tags to museum paintings.

(2) A user experiment is conducted. 30 non-experts and 10 experts are recruited to rate and tag 45 paintings chosen from the collection of the Vatican picture-gallery. The participants are selected according to availability sampling strategy. The 30 non-experts are from young people having a master degree in Computer Science or Humanities, while experts are teachers in Arts and Humanities disciplines.  They did statistics evaluation using classification accuracy measures (precision and recall) to investigate whether using only the personal tags or the whole set of social tags are more accurate to recommendation, and whether expertise will influence the accuracy of recommendation.

7. Main findings:

(1) The highest overall accuracy is reached when user preference learned from both static content and personal tags (not social tags) are exploited in the recommendation process.
(2) The expertise of users contributing to the folksonomy does NOT actually affect the accuracy of recommendations.

8. Analysis: This paper is very related to our research team’s current and future work in building the personalization and collaboration system for engineering education community.

(1) The conclusion here that expertise doesn’t matter is interesting. This recommendation system is mainly for recommendation of artworks, movie, music and books. If the items being expanded to others such as academic papers, I am afraid that expertise will matter.

(2) My current work is designing the UI that allow the users to specify interest ratings, write tags and comments to papers. For one things, this will driven the personalization in My Library and My Gallery (the two personalized components in our system). For another, we also thinking about ask the users to specify quality ratings of the paper which could contribute to the whole community in making sense of academic papers in this field.

(3) I have a deep concern about the meaning of personalization and recommendation, that is, the filter bubble, or social segmentation, or the echo chamber. We only see things that’s recommended to us by the machine, and being isolated from the world outside of our world.