8. CONCLUSIONS
In this paper, we focused on the task of related video
suggestion, commonly found in many online applications.
The majority of the current approaches to related video
suggestion are based on the collaborative filtering analysis.
Most generally, the collaborative filtering approaches suggest
videos that are likely to be co-viewed with the currently
watched video.
While such suggestions work well for popular videos with
many views, it is less applicable in situations where little or
no view data is available. This includes fresh videos, or tail
videos with few views. To address this challenge, we propose
a hybrid approach to video suggestion, which combines the
video content with the co-view information to improve the
related video suggestions.
To this end, we represent the video using a set of topics
mined from various sources including the video metadata,
frequent uploader keywords, common search queries, playlist
names and Freebase entities. In order to achieve an effective
topical video representation, we assign weights to each topic,
and propose two approaches for topic weighting.
The first approach is based on the standard information
retrieval heuristics such as topic frequency and inverse document
frequency weighting and stopword removal. The downside
of this approach is that it does not directly take into
the account the user behavior.