Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
409826 | Neurocomputing | 2012 | 11 Pages |
Many web applications today thrive on offering services for large-scale multimedia data, e.g., Flickr for photos and YouTube for videos. However, these data, while rich in content, are usually sparse in textual descriptive information. For example, a video clip is often associated with only a few tags. Moreover, the textual descriptions are often overly specific to the video content. Such characteristics make it very challenging to discover topics at a satisfactory granularity on this kind of data. In this paper, we propose a generative probabilistic model named Preference-Topic Model (PTM) to introduce the dimension of user preferences to enhance the insufficient textual information. PTM is a unified framework to combine the tasks of user preference discovery and document topic mining together. Through modeling user-document interactions, PTM cannot only discover topics and preferences simultaneously, but also enable them to inform and benefit each other in a unified framework. As a result, PTM can extract better topics and preferences from sparse data. The experimental results on real-life video application data show that PTM is superior to LDA in discovering informative topics and preferences in terms of clustering-based evaluations. Furthermore, the experimental results on DBLP data demonstrate that PTM is a general model which can be applied to other kinds of user–document interactions.