In order to achieve an effective topical video representation,
we assign weights to each topic associated with a video.
We propose two techniques for topical video representation.
The first technique is based on the well known information
retrieval heuristics such as computing topic frequencies
and topic inverse document frequencies. The second technique
leverages the implicit user feedback (such as video
co-watches) available in the online setting. This feedback is
used for a supervised learning of the optimal topic weights.
We empirically evaluate these two topical representation
methods. We find that both of these representations have
a significantly positive effect on the quality of video suggestions.
Furthermore, our evaluation demonstrates that
learning topic weights from user feedback can increase the
user engagement (compared to the collaborative filtering approach)
by more than 80% over the standard information
retrieval representation.
To evaluate our approaches, we conduct a large scale live
experiment on millions of YouTube video. The live experiment
demonstrates that a hybrid video suggestion system
that incorporates topic-based retrieval significantly outperforms
a purely co-view based suggestion system. These improvements
are especially high for fresh videos and videos
with rich topical representations.
There are several important contributions in this paper.
First, we formulate the video suggestion task as an information
retrieval problem and demonstrate that this formulation
enables effective and efficient deployment of video suggestion
on a very large scale. Previous work on video retrieval was
mainly performed offline on relatively small collections such
as TRECVID collections [20], or small pre-selected samples
of online videos [31]. To the best of our knowledge, this is
the first published study to address the challenges of realtime
video retrieval using topical representation in a large
scale web collection.
Second, we demonstrate that the richness of the implicit
user feedback available in the online setting can be leveraged
to improve the effectiveness of topical video representations.
To this end we employ a novel learning technique that derives
the optimal topic weights from co-view information.
Third, we describe the architecture of the hybrid video
suggestion system that combines the collaborative filtering
information with the topic-based video information.
Finally, we thoroughly evaluate the performance of our
system using both user simulation and a large scale live experiment.
This evaluation demonstrates the superiority of