Foxtrot fact-file

What's the Problem?

The Foxtrot recommender system addresses the problem of recommending on-line research papers to over 200 computer science staff and students at the University of Southampton for a full academic year. Researchers need to be able to search the system for specific research papers and have interesting papers autonomously recommended. Unobtrusive monitoring methods are preferred because researchers have their normal work to perform and would not welcome interruptions from a new system. Very high accuracy on individual recommendations is not critical, however, since recommendations are made in sets, and poor recommendations can be ignored in favour of better ones.

Towards a Solution

Foxtrot is an evolution of a previous project Quickstep. Web browsing is unobtrusively monitored using a Squid web proxy and all browsed research papers shared between users within a central paper database. Foxtrot uses pearson-r correlation to recommend and boosted kNN classification to profile user interests. An ontological approach is taken to represent user profiles which allows inference of interests, bootstrapping from an external ontology and profile visualization.

Ontological inference is applied to profile topics, using a is-a taxonomy described within the ontology, to infer interest in superclass topics. This is applied recusrively by the profiling algorithm, thus adding to user profiles those topics of potential interest not observed directly. Two experiments conducted with the Quickstep recommender system have shown inference of profile interests to be advantagous.

Since profiles are represented using an ontology, external ontiologies can be mapped to the Foxtrot ontology, allowing external knowledge to be used to bootstrap the recommender system. This has been shown via experimentation to be effective in reducing the cold-start problem recommender systems often face.

Profile visualization is made possible because the profiles are represented in ontological terms, terms which users can understand. Visualization of a interest profile thus helps users build a conceptual model of the system and acts as a useful tool to facilite direct feedback via the graphical interface. Users can literally draw their own profile when they think the system has got things wrong.

A year long trial of Foxtrot was conducted with over 200 staff and students from the computer science departement at the University of Southampton. The results are published in the papers referenced below. Two smaller experiments were also conducted with the Quickstep system, also detailed in the papers below.

Further Reading

Middleton, S.E. Alani, H. Shadbolt, N.R. De Roure, D.C. (2002) Exploiting Synergy between Ontologies and Recommender Systems. In Proceedings Semantic Web Workshop 2002, Hawaii, USA.

Middleton, S.E. De Roure, D.C. Shadbolt, N.R. (2002) Foxtrot Recommender System: User profiling, Ontologies and the World Wide Web, Poster, The Eleventh International World Wide Web Conference (WWW2002), Hawaii, USA

Middleton, S.E. De Roure, D.C. Shadbolt, N.R. (2001) Capturing knowledge of user preferences: ontologies in recommender systems. In Proceedings First International Conference on Knowledge Capture, pages 100-107, Victoria, British Columbia, Canada.

Who's Who in Recommender Systems - Non-Attendees, 2001 ACM SIGIR Workshop on Recommender Systems

Semantic representation

View in the AKT Triplestore Browser or as RDF.

Also available in DOAP RDF (Description Of A Project)