125
C H A P T E R 6
Multimodal Sequential
Learning for Micro-Video
Recommendation
6.1 BACKGROUND
As the micro-videos surge, it becomes increasingly difficult and expensive for users to locate their
desired micro-videos from the vast candidates. In light of this, it is crucial to build a personalized
recommendation system to intelligently route micro-videos to the target users.
6.2 RESEARCH PROBLEMS
Building a personalized recommendation system for micro-video services is non-trivial, due
to the following reasons. (1) Diverse and dynamic interest. On the one hand, users’ interest
evolves over time, and it is hence a sequential expression. For example, as shown in Figure 6.1,
a user likes cooking videos at time t
1
, but may prefer dance videos at t
2
. On the other hand,
users’ interest is diverse, namely a user may be fond of multiple topics at the same time. In a
sense, personalized recommendation requires to simultaneously model users’ dynamic and di-
verse interest information. (2) Multi-level interest. Users may have different interaction types
on micro-videos, including “click,” “like,” and “follow,” which signal different degrees of in-
terest. For example, “click” means the user is attracted to the micro-video, “like” is one much
enjoys and appreciates the micro-video, and “follow” refers to the user likes the micro-video very
much and wishes to see it again in future. Heretofore, how to integrate the various degrees of
interest into personalized recommendation is largely untapped. (3) True negative samples. As
we know, prior methods commonly assume that nonpositive items are negative samples, which
is hardly reliable to infer which item a user did not like. Different from these models, we are
able to obtain true negative samples, i.e., micro-videos that users preview the thumbnails yet no
“click” occurs. erefore, how to utilize these true negative samples to explicitly model users’
uninterested information becomes a crucial problem.
For the past few years, several studies have been conducted on the personalized rec-
ommendation, such as collaborative filtering based models [7, 25, 69], content-based sys-
tems [34, 113, 129, 206, 209], and hybrid methods [48, 201]. Although these methods produce
promising performance on recommendation, most of them suppose users’ interest as static. In-
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.