PAWS Meeting April 1st, 2014
Today Chirayu was presenting the progress of his thesis, "Using External Sources to Improve Research Talk Recommendation in Small Communities". The keyword-based Content-based and Content-boosted Collaborative Filtering models work pretty well. In contrast, The SVD-based in (25,75, and 100 latent topics) ones underperform below the keyword-based ones in the baseline and other three external source models.
While both techniques perform decently, clustering the keyword-based and SVD-based (100 latent topics) performed very poorly. The group has suggested to explored on more latent topics (25 - 1000 topic) and tried on the spatial clustering technique, in which Julio and Sherry used it in the other related work and performed very well in the very spare dataset, comparing with other techniques. Sherry also suggested to try to use only certain portions of the external sources, for example, the first page of homepage, the latest k bibliographies, or the latest k external academic bookmarked articles.