as “one of the most inuential recent theories of public opinion
formation” [8].
We highlight the unique characteristics of our empirical study. (1)
We study the behavior of giving ratings, while previous empirical
studies focus on the biases in ratings [
11
–
13
,
33
]. (2) We study the
dynamic aspect which has never been addressed by existing MNAR
models [
1
–
6
]. For example, as time passes by, the domination of
majority opinion grows, which results in the increased willingness
to rate for majority opinion holders and the decayed willingness
for the minority opinion holders.
Two challenges arise in the empirical study. Firstly, survey stud-
ies in a lab environment [
11
–
14
] is problematic, because the nd-
ings are based on hypothetical willingness instead of actual will-
ingness [
15
]. A recent survey [
16
] shows that the hypothetical
willingness is a poor indicator of actual willingness. To address
this challenge, we form our empirical study on the basis of actual
willingness. Two scenarios are included: (1) Scenario I considers
that a user’s willingness to display or hide his/her ratings is oered
(Sec. 3); (2) Scenario II considers that a user’s inclination to rate is
not available (Sec. 4).
Secondly, there are counter cases according to the Spiral of Si-
lence Theory, i.e. the minority opinion that remains at the end of the
spiral is called the
hardcores
. Mixing hardcore and non-hardcore
users can hurt the performance of a recommendation model as the
two user groups behave dierently. To tackle this challenge, we
present formal denition to distinguish hardcore users and study
the characteristics of hardcore users in Sec. 5.
In order to demonstrate the impact of our empirical study, we
present a straightforward application of the two major ndings,
i.e. the existence of spiral and the existence of hardcore users. We
develop a
M
issing
C
onditional on
P
ersona (MCP) model which
mimics the generation process of ratings and responses. The proba-
bility of giving response is related to the perceived opinion climate,
individual rating, and the persona of the user (i.e. hardcore user or
non-hardcore user). Experiments show that the MCP model outper-
forms state-of-the-art recommendation models, including models
with and without MNAR assumptions.
Our contributions are three-fold.
Practical contributions.
We verify the existence of a spiral of
silence by large-scale empirical studies on 8 real recommendation
data sets. We validate the group of hardcore users and provide
detailed insights into the personal traits of hardcore users. These
ndings are particularly useful in niche marketing.
Methodological contributions.
We design solutions to con-
duct large-scale eld research of the spiral of silence theory on
recommendation systems. For example, we give formal denitions
to the core concepts including majority opinion holders and hard-
core users. In hypothesis testing, we conduct a series of trend studies
to capture the dynamic aspect of the theory which has hardly been
addressed by previous studies.
Model contributions.
We present a MCP model based on the
major ndings of our empirical study. We experimentally show
that the MCP model outperforms state-of-the-art recommendation
models. The signicant improved performance reveals the potential
impact of our empirical ndings.
The paper is organized as follows. Sec. 2 summarizes the related
works. In Sec. 3 and Sec. 4, we testify the existence of a spiral of
silence over dierent scenarios. Sec. 3 is devoted to empirical study
in which a user’s willingness to rate is available. Sec. 4 corresponds
to conducting empirical study on recommender systems in which
a user’s willingness to rate is not available. In Sec. 5, we study
the existence of hardcore users in the formation of the spiral of
silence, and we reveal the characteristics of hardcore users. Sec. 6
presents the MCP model which is developed by embedding the
empirical ndings. Sec. 7 presents and analyzes the experimental
results. Finally, Sec. 8 concludes our contributions and insights into
the future work.
2 RELATED WORK
Biases in Recommender System
is related to our research. Some
researchers observed a trend of increasing average ratings [
17
–
19
];
others found that later ratings are on average lower [
20
]. Hu et.al
observed a J-shaped distribution [
21
]. A recent work [
22
] showed
the existence and strength of conformity. These works focused on
the biases in ratings, while we study the biases in responses, i.e. mi-
norities are less likely to give ratings. In addition, the explanations
in previous studies are not appropriate for all recommender systems,
i.e. the choice-supportive bias [
18
] only applies to recommender
systems with reviews.
The
rich gets richer
(Matthew eect) cliche is another line of
research we want to distinguish our work with. Though the “rich
gets richer” assumption generates a similar phenomena to the spi-
ral process, it does not explain the formation of public opinion as
the spiral theory does. The Matthew eect suggests group dynam-
ics. It is not suitable to derive a recommendation model because
personalization is not retained.
MNAR Models
in RS are aware that ratings are missing not at
random. Probabilistic models were presented to relate a missing
to various factors, e.g. the value of a hidden rating [
1
–
4
] or to the
item to be rated [
1
,
5
,
6
]. As we mentioned above, they were based
on heuristics that are neither empirically veried nor theoretically
proven. Furthermore they are unable to explain the evolution of
ecology and several phenomena in the recommender systems, e.g.
a high rated item gets more praises. Our work aims to reveal these
hidden patterns from a social science perspective, and thus serves
as a guiding light for future MNAR models.
Empirical Study on Spiral of Silence Theory
has a long his-
tory. They adopted a “train test” type of experiments, i.e. the subjects
are questioned about their willingness to discuss with a stranger
on a train about any topic. Most works [
11
–
14
] observe a positive
correlation between perceived opinion climate and willingness to
rate, both of which are collected during the survey of train test.
However the result is based on hypothetical willingness. We believe
that our work is the rst to verify the spiral model in large scale real
life recommender systems. Moreover, they only proved the “social
conformity hypothesis" [
9
]. Emphasis on time in the formation
of the spiral has not been reected on the methodologies. On the
contrary, we acknowledge the dynamic nature of the spiral model.
3 EXISTENCE OF SPIRAL: SCENARIO I
In this section we testify the fundamental assumption of the theory
in recommender systems: the spiraling process, in which two key
activities repeatedly occur. (1) A user is prompted to show his
Session 4: FATE & Privacy
WSDM ’19, February 11–15, 2019, Melbourne, VIC, Australia