I was invited to give a talk about the CHIP project for the IR group at the University of Glasgow as part of the K-Space Network of Excellence exchange. The group is head by Keith van Rijsbergen. Joemon Jose also leads the group quite a bit, and is in charge of its involvement with K-Space.
This group has long been research media retrieval. They have an increasing involvement with user-centered design, user studies and the Semantic Web. They are also interested in recommenders and the Semantic Web, which is why they asked me to talk.
K-Space participant Jana Urban is wrapping up her PhD thesis with the group. She will likely stay with the group for at least a year or so afterward. Her thesis is about EGO, a system she developed for organizing photographs. She performed user studies on EGO as part of her thesis work.
EGO takes in many image files. It handles three types of image metadata: groups, tags and features. A user can drag and drop the images into groups, which he or she can then label. EGO lets its users assign tags to images as well. Finally, EGO extracts a set of features from all of its images.
Given images and metadata, EGO offers content-based recommendation. To do so, EGO takes images the user has selected and recommends other images that share groups, tags and features with them.
Jana mentioned that their group is currently discussing what policy and practice to have for user-centered design and user studies in its projects. She mentioned Hyowen Lee, whom she described as DCU's in-house usability expert, helping make sure all its project follow effective user-oriented practice. Hyowen is now wrapping up a two-week visit at CWI-INS2, also as part of a K-Space exchange.
Another K-Space participant, Frank Hopfgartner, is now starting his PhD work based on his Masters work on a video retrieval system, which he applies to TRECvid. This system lets users type in keywords for search. It then displays the returns as thumbnails of keyframes from video scene segments speech-recognized text contains these keywords. His system's focus is the subsequent search refinement, which finds other keywords that frequently occur in the search returns. The user can then choose to apply these keywords in refinements of the search query.
Joemon discussed his desire to use ontologies for video retreival, applying it, for example, to Frank's system. I pointed him to the CHOICE project as a starting point for semantic processing of video and for finding relevant ontologies for their work.
Ali Azimi Bolourian is a PhD student researching event detection in blogs. That is, he processes tens of thousands of blog entries to detect events as they start being discussed, as well as central bloggers in this discussion. He is also discussing starting a project with museums in Glasgow where their whole collections are put on-line. Motivation includes making the items in storage also accessible.
Questions during the talk including some about the algorithms we use and plan to use for processing the recommendations and for the "Why?" button certainty displays. People were also interested in the size measurements for our metadata. The large image for each artwork is an important feature for them.
They also asked about the precise nature of the stored user model. I clarified that we currently store just explicit ratings (and some logging and status) because that is our research focus. We then discussed having general user data, such as age, country, gender and profession, serve to help the recommendation, particularly at the very beginning. This relates to minimizing user interaction. Would knowing only these four basic facts about the user generate more accurate predictions than ratings for the four most strategically selected paintings? We could check this with Henriette's data.