We empirically evaluate these two topical representation
methods in a large scale live experiment on YouTube traffic.
We find that both of the representations have a significantly
positive effect on the quality of video suggestions, compared
to the standard collaborative filtering approach. This effect
is especially visible for videos with rich topical representation
such as videos in the News and Science and Technology
categories (more than 3.5% increase in watch time), and for
fresh videos with sparse co-view data (more than 3% increase
in watch time).