Spiral of Silence in Recommender Systems
Dugang Liu, Chen Lin
Department of Computer Science, Xiamen University
Xiamen, China
Zhilin Zhang
School of Computing Science, Simon Fraser University
Burnaby, Canada
Yanghua Xiao
School of Computer Science, Fudan University
Shanghai, China
Alibaba Group
Hangzhou, China
Hanghang Tong
School of Computing, Informatics and Decision Systems
Engineering, Arizona State University
Tempe, U.S.A.
ABSTRACT
It has been established that, ratings are missing not at random in
recommender systems. However, little research has been done to
reveal how the ratings are missing. In this paper we present one
possible explanation of the missing not at random phenomenon.
We verify that, using a variety of dierent real-life datasets, there
is a spiral process for a silent minority in recommender systems
where (1) people whose opinions fall into the minority are less likely
to give ratings than majority opinion holders; (2) as the majority
opinion becomes more dominant, the rating possibility of a majority
opinion holder is intensifying but the rating possibility of a minority
opinion holder is shrinking; (3) only hardcore users remain to rate
for minority opinions when the spiral achieves its steady state.
Our empirical ndings are benecial for future recommendation
models. To demonstrate the impact of our empirical ndings, we
present a probabilistic model that mimics the generation process of
spiral of silence. We experimentally show that, the presented model
oers more accurate recommendations, compared with state-of-
the-art recommendation models.
KEYWORDS
Spiral of Silence, Recommender System, Missing not at Random,
Hardcore
ACM Reference Format:
Dugang Liu, Chen Lin, Zhilin Zhang, Yanghua Xiao, and Hanghang Tong.
2019. Spiral of Silence in Recommender Systems. In The Twelfth ACM Inter-
national Conference on Web Search and Data Mining (WSDM ’19), February
11–15, 2019, Melbourne, VIC, Australia. ACM, New York, NY, USA, 9 pages.
https://doi.org/10.1145/3289600.3291003
1 INTRODUCTION
R
ecommender
S
ystems (RS) have received extensive attentions
from both research communities and industries. The power of an
RS is highly dependent on the assumption that the collection of
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specic permission and/or a
fee. Request permissions from permissions@acm.org.
WSDM ’19, February 11–15, 2019, Melbourne, VIC, Australia
© 2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-5940-5/19/02... $15.00
https://doi.org/10.1145/3289600.3291003
Table 1: A toy example of 5 users’ ratings on 5 movies. Alice’s
ratings on Aliens and Eskiya are hidden.
Aliens Ben-Hur Casino Dangal Eskiya
Alice (2) 3 3 (5)
Bob 5 3 2 2
Clare 5 4 5 1 2
Diane 2 2 2 5
Elle 5 2 4 2
ratings correctly reects the users’ opinions. Most recommender
systems, however, suer from extremely sparse rating data. More
challengingly, it is rare that users tell “the truth and the whole truth”
at all times. When a missing rating occurs due to the user’s choice
of non-response, the representativeness of the ratings is degraded
and the inference of a recommendation model is distorted. Consider
a conventional collaborative ltering RS running on a toy example
illustrated in Table 1. Suppose, for some reason, Alice is not willing
to give her ratings on the movie Aliens and Eskiya. The RS will
make a wrong judgement that Alice’s nearest neighbor is Bob, based
on the two common items Alice and Bob have, while in fact Diane
shares the most similar taste with Alice.
In the literature of RS, models [
1
–
6
] which assume ratings are
M
issing
N
ot
A
t
R
andom (MNAR models) are recognized to have a
superior ranking performance. Existing MNAR models mimic the
generation of responses under dierent heuristics, i.e. the possibility
of a response is related to the exact value of the rating [
2
–
4
] or to
an unknown feature of the item [
5
,
6
]. Unfortunately, none of the
heuristics is empirically veried on real datasets, or supported by
theoretical social studies.
In real scenarios, there could be various factors that lead to
missing responses. Our goal in this paper is to provide a possible
explanation for missing ratings and identify the key factors under-
lying users’ decision process of whether or not to rate an item in
recommender systems.
Towards this goal, we rst empirically examine the response
patterns in recommender systems. We verify the existence of a
spiral process
in which users are more and more likely to rate
if they perceive that they are supported by the opinion climate
(i.e. the dominant opinion), while the minority opinion holders are
more and more reticent. Such a spiral process can be explained by
the Spiral of Silence Theory
1
[
7
], which has been acknowledged
1
In the remaining of this paper, the Spiral of Silence theory will be referred to as “the
theory”.
Session 4: FATE & Privacy
WSDM ’19, February 11–15, 2019, Melbourne, VIC, Australia