126 6. MULTIMODAL SEQUENTIAL LEARNING
Interest: Cooking Baby Dance Basketball
Click Not Click
Like
Follow
Interactions:
t
1
t
2
Figure 6.1: Illustration of a user’s historical interactions with micro-videos, which reflects the
user’s diverse, dynamic, and multi-level interest.
spired by this, some researches consider users’ interest as dynamic when designing recommenda-
tion systems and have achieved better performance [28, 52, 132, 152, 157, 203]. ey, however,
overlook the diverse and multi-level characteristics of users’ interest. Moreover, all the afore-
mentioned methods commonly assume that items not been clicked by users are negative and
utilize them [60, 127, 137] or sample part of them as negative samples [61, 66] to represent
users’ uninterested items. However, these presumed negative samples may be not truly negative,
and they hence may confuse the recommendation system. As we can see, the existing studies
neither consider the diverse and multi-level interest nor exploit users’ true uninterested items
to model the recommendation system. erefore, they cannot be directly applied to the micro-
video recommendation.
6.3 RELATED WORK
Recommender systems are vital in video communities, such as Youtube, Vimeo,
1
and Veoh.
2
e
exiting methods can be roughly categorized as collaborative filtering-based methods [7, 25, 69],
content-based methods [34, 113, 129, 206, 209], and hybrid methods [48, 201]. In terms of
collaborative filtering, Baluja et al. [7] utilized the random walk through a co-view graph to rec-
ommend YouTube videos. Chen et al. [25] integrated an attention mechanism into collaborative
filtering with implicit feedback and evaluated its effectiveness in multimedia recommendation.
However, collaborative filtering based methods cannot well solve the cold start problem. By con-
trast, the content-based methods recommend videos by calculating the similarity between new
videos and users’ historical accessed videos. For example, Mei et al. [113] proposed a contex-
tual recommendation system based on multimodal fusion and relevance feedback. With respect
1
https://vimeo.com/
2
https://www.veoh.com/
..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset