Learning to rank videos personally using multiple clues
Songhua Xu, Hao Jiang, Francis C. M. Lau · 2009
In this paper, we introduce a new learning based video content similarity model. The model leverages on multiple clues on the contents of a video and can be used to rank videos in a personalized way. The key to produce a personalized video ranking is to have a good estimate of pairwise video content similarity, which is realized through meta-learning using a radial-basis function network. Four aspects of a video are considered in deriving the video content similarity in our method. The training data to our model are acquired in the form of user judged preference relationships regarding video content similarities. With the optimized video content similarity estimation obtained by our algorithm, we can produce a personalized video ranking that matches more closely an individual user's watching interest over a collection of videos. The video ranking results generated by our prototype system are compared with the groundtruth rankings supplied by the individual users as well as rankings by the commercial video website YouTube. The results confirm the advantages of our method in generating personalized video rankings.