ABCPRec: Adaptively Bridging Consumer and Producer Roles for User-Generated Content Recommendation (SIGIR 2019)
ABCPRec: Adaptively Bridging Consumer and Producer Roles
for User-Generated Content Recommendation
What is the characteristic of UGC?
How to recommend UGC to a user?
How to learn consumer/producer vectors?
What are the experimental results?
C
Non UGC: User = Consumer UGC: User = Consumer + Producer
C
P
𝝂𝝂𝑢𝑢
𝑝𝑝
A user has two vectors
corresponding to two roles:
consumer and producer.
P
C 𝝂𝝂𝑢𝑢
𝑐𝑐
𝑢𝑢 ?
𝑖𝑖
C 𝝂𝝂𝑢𝑢
𝑐𝑐
= �𝑥𝑥𝑢𝑢𝑢𝑢 = 𝛼𝛼 + 𝛽𝛽𝑢𝑢 + 𝛽𝛽𝑖𝑖 + 𝝂𝝂𝑢𝑢
𝑐𝑐
, 𝜸𝜸𝑖𝑖 + 𝝂𝝂𝑢𝑢
𝑐𝑐
, 𝝂𝝂𝑝𝑝𝑖𝑖
𝑝𝑝
?
where 𝝂𝝂𝑢𝑢
𝑐𝑐
, 𝜸𝜸𝑖𝑖
𝑝𝑝
is the affinity between user 𝑢𝑢 and item 𝑖𝑖
where 𝝂𝝂𝑢𝑢
𝑐𝑐
, 𝝂𝝂𝑝𝑝𝑖𝑖
𝑝𝑝
is the affinity between user 𝑢𝑢 and 𝑖𝑖’s producer 𝑝𝑝𝑖𝑖
𝜸𝜸𝒊𝒊 𝝂𝝂𝑢𝑢
𝑝𝑝
𝑝𝑝𝑖𝑖
P
𝒖𝒖’s consumed items
𝝂𝝂𝑢𝑢
𝑐𝑐
𝝂𝝂𝑢𝑢
𝑝𝑝
𝒖𝒖’s produced items
Consume Consume
Target user 𝑢𝑢
Consume Consume
Target user 𝑢𝑢
𝝂𝝂𝑢𝑢
𝑐𝑐
𝝂𝝂𝑢𝑢
𝑝𝑝
If 𝒖𝒖’s nature as a consumer is similar to that as a producer, 𝝂𝝂𝑢𝑢
𝑐𝑐
and 𝝂𝝂𝑢𝑢
𝑝𝑝
should be close.
Compute the similarity based on the overlap between
the users who consumed 𝒖𝒖’s consumed items and those who consumed 𝒖𝒖’s produced items.
If the overlap is big, 𝝂𝝂𝑢𝑢
𝑐𝑐
and 𝝂𝝂𝑢𝑢
𝑝𝑝
should be close. If the overlap is small, 𝝂𝝂𝑢𝑢
𝑐𝑐
and 𝝂𝝂𝑢𝑢
𝑝𝑝
should not be close.
𝑲𝑲 PopRec BPR Vista FMs CPRec NBCPRec ABCPRecH1 ABCPRecH2
Flickr
20 0.6737 0.8698 0.8436 0.8764 0.8563 0.8839 0.8861 0.8900
50 0.6737 0.8772 0.8435 0.8822 0.8664 0.8937 0.8949 0.8992
80 0.6737 0.8777 0.8394 0.8810 0.8712 0.8955 0.8988 0.9028
Reddit
20 0.6392 0.8713 0.8829 0.8960 0.9138 0.9209 0.9296 0.9340
50 0.6392 0.8721 0.8918 0.8999 0.9201 0.9302 0.9346 0.9391
80 0.6392 0.8709 0.8946 0.9001 0.9211 0.9322 0.9376 0.9408
CPRec is the state-of-the-art method for UGC recommendations proposed at RecSys’18.
ABCPRecH1/ABCPRecH2 focus on the users who consumed 𝒖𝒖’s consumed/produced items.
ABCPRec statistically outperformed CPRec in terms of AUC.
User-item matrix
𝒖𝒖’s consumed items 𝒖𝒖’s produced items
UGC is content produced
by ordinary people rather
than by professionals and
distributed on the Web.
National Institute of Advanced Industrial Science and Technology (AIST)Kosetsu Tsukuda, Satoru Fukayama, Masataka Goto
photo by Emma
photo by CetusPhotography
photo by cuatrok77
photo by Bryan Ochalla
photo by Laura D'Alessandro
photo by panli54
photo by K-nekoTR
photo by Tambako The Jaguar
photo by ClieistD
photo by ClieistD
photo by localpups