This document describes the research vision and goals for my contribution to the CHIP project. It is an active and continually changing document, stating what I'm doing know, where I'm going with it, and recording what I've done and how the research direction has evolved.
My role in CHIP is studying the impact of the Semantic Web on recommendation processing. This is in part about how we can apply Semantic Web technologies to improve the generation of recommendations from user ratings and artifact annotations, thus applying new background processing to improve the performance of an established interface paradigm. I will also explore the new interface paradigms the combination of recommendation processing and the Semantic Web facilitate, such as the generation of certain aspects of the personalized tour.
In the long run, I wish to address the emerging Semantic Web User Interaction community. My long-term focus is semantic to presentation structure transform. This relates strongly with user profiles and their relation to particular semantic domains. Together, user profiles about particular semantic repositories guides the transformation of a repository to personalized presentation.
See CHIP Member Recommender Page
The context of the Semantic Web is new for content-based recommendation. Its novel contributions include:
We can guide our research plan with a "map" of the problem space that CHIP could solve. CHIP can then plot a course through it. The space would be personalization on the (museum) Semantic Web. Components includes user profiles based on conceptual domains and what personalizable variables exist in interfaces to them, such as content selection, ordering and group, in terms of local and various global access means.
Recommendation represents one type of user profile over a domain: that of expressed interest in individual concepts. Other forms of relavance may be similarly calculated, such as familiarity and task vs goal vs permanent relevance. For CHIP, relavance makes a weighted graph from semantics, which guides transform to presentation. For the demo, we see now the importance of structuring lists in various parts of the interface. Such transform could start giving tangible results in the demo soon, as well as scientific results.
We can handle predicted and explicit interest of concepts as relevance in various aspects of interface generation. The best established aspect for applying relevance to is search. Here, the user enters a search query whose processing generates relevance measures for matching repository components. Typically, this relevance measure determines the order the system presents the matching concepts in. Most simply, we could consider either or both of explicit and predicted interest as the relevance measure for sorting the presentation of a set of concepts. More complexly, interest can be a factor in relevance, combined with query match and perhaps other kinds of relevance measure.
Certain predicates can correlate with interest, either positive, negative or both. If the different values of a predicate annotate concepts for which the user has explicit interest value with low distribution, this this predicate correlates strongly with user interest. For example, if the user consistently likes artwork created by Rembrandt, but consistently dislikes artwork created by Van Gogh, and is consistently neutral about artwork by Rietveld, then the "created by" predicate tends to determine the user's interest. Simply put, the user's interest or taste in an artwork is often determined by its creator.
The system can exploit predicates that correlate with interest by grouping presentations based on different values of that predicate more often. For example, if a user's interest is often determined by artist, then generated presentations will group their components based on artist, and perhaps artist styles and groups, for that user. Another potential exploitation is that the dialog strategy can detect when such correlating predicates appear and apply them to selecting artworks for the user to rate. Here, hypotheses of interest will more often be based on different values for that predicate, which in turn selects artworks for rating that have these values.