As our retrieval methods are topic based, we expect them
to be most beneficial for videos with rich topical content
and videos with low co-view signal. The results in Table 1
confirm this hypothesis.
For categories with richer topical representations like News
and Science and Technology, the improvements obtained by
our method are approximately three times higher than the
average improvement in Figure 6. The highest improvements
are obtained for the Pets and Animals category, which
has many tail videos with little co-view information. On the
other hand, for Music and Gaming categories, which have
more popular videos with co-view data, the topic retrieval
has smaller positive benefits (or even a slight negative impact
for the IRTopics method in the Music category).
Similarly, for fresh videos (e.g., less than a month old),
our improvements are more significant (especially for the
TransTopics method) than for the older videos. This is due
to the fact that the fresher videos have a weaker co-view
signal.
The results in Table 1 demonstrate that our topic retrieval
methods improve the related video suggestions by augmenting
the co-view results with fresher and more topically relevant
videos. In agreement with the general results in Figure
6, the TransTopics method is always significantly more
effective than the IRTopics method across different video
categories and age groups.