Swayed by Friends or by the Crowd?
Zeinab Abbassi
Dept. of Computer ScienceColumbia Universityzeinab@cs.columbia.edu
Christina Aperjis
Social Computing GroupHP Labschristina.aperjis@hp.com
Bernardo A. Huberman
Social Computing GroupHP Labsbernardo.huberman@hp.com
ABSTRACT
We have conducted three empirical studies of the effects of friend recommendations and general ratings on how onlineusers make choices. These two components of social influ-ence were investigated through user studies on MechanicalTurk. We find that for a user deciding between two choicesan additional rating star has a much larger effect than an ad-ditional friend’s recommendation on the probability of se-lecting an item. Equally important, negative opinions fromfriends are more influential than positive opinions, and peo-ple exhibit more random behavior in their choices when thedecision involves less cost and risk. Our results can be gen-eralized across different demographics, implying that indi-viduals trade off recommendations from friends and ratingsin a similar fashion.
Author Keywords
SocialInfluence, Friends, Ratings, Choice, DynamicsofMar-ket Share
ACM Classification Keywords
H.5.m Information Interfaces and Presentation (e.g., HCI):Miscellaneous; K.4.4 Computers and Society: [ElectronicCommerce]
General Terms
Human Factors, Experimentation.
INTRODUCTION
When making choices, people use information from a num-ber of sources including friends, family, experts, media, andthe general public. Two sources that are particularly relevantin an online setting are the opinions of friends and ratingsfrom the general public. Friends influence choices becauseoften their interests overlap with their friends’ interests. Inmany cases, however, recommendations from one’s friendsare in stark contrast to opinions of individuals in the generalpublic who are not one’s friends.
Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.
CHI 2012
, May 5–10, 2012, Austin, TX, USA.Copyright 2012 ACM xxx-x-xxxx-xxxx-x/xx/xx...
$
10.00.
An interesting question, which we study in this paper, is howthe decision of an online user is influenced by friend recom-mendations and ratings from the general public, in partic-ular when these two sources of information are in conflictwith each other. This question is interesting for two reasons:First, understanding how people trade off friends’ opinionswith ratings from the general public helps to determine theweight assigned by consumers to these two sources whenthey are uncertain about choosing one of two possible op-tions. Second, from a normative standpoint, this informationcan be used when designing algorithms that display thesetwo sources of information in order to increase the probabil-ityofauserselectingoneoftheoptions. Forexample, anon-line social network platform that has information about howa user’s friends and the general public have rated two differ-ent items can display to the user the item that she is morelikely to select. On the other hand, if users tend to disregardsome source of information, this source need not be shownto the user. Finally, an advertiser that wishes to make theuser choose a certain item may strategically choose whichpieces of information to show.Specifically, focusing on friends and the general public astwocomponentsofsocialinfluenceisimportantbecausethesesources of social information are already used in a varietyof algorithms and applications online. Social recommendersystems take into account the actions of a user’s friends andmake recommendations accordingly. Social search is alsogaining more attention. Google recently launched its +1 but-ton for search results and ads in order to improve its searchalgorithm. If a user thinks that a search result or an ad is use-ful she can click on the +1 button. The +1 will be displayedalong with the user’s name in the search results to all herfriends who subsequently search a similar query. For userswho are not friends, only the number of +1’s will be dis-played. Facebook uses a similar approach for business pageswith the intention of getting higher click-through rates. Themodel that we suggest can be leveraged to design better al-gorithms for these and other similar applications.In addition to quantifying how an individual makes choicesin the presence of social influence, we also consider how agroup of users makes decisions over time when exposed toinformation from friends as well as ratings from the gen-eral public. While previous work has studied how groupsmake decisions in the presence of social influence, it has notconsidered the different sources of this influence. For in-stance, the model of Bendor et al. that considers individualsmaking choices takes into account one source of informa-1
tion, namely the number of others that have chosen one itemout of two options[7]. By contrast we consider how mar-ket share of the two options evolves in a setting where thereare two sources of social influence, namely opinions fromfriends and ratings from the general public.In particular, in this paper we asked the following questions:
•
R1:
How much are one’s choices influenced by the opin-ions of her friends compared to ratings from the generalpublic? What mathematical model explains this?
•
R2:
Dofriends’negativeopinionshaveastrongerorweakereffect than friends’ positive opinions about an item?
•
R3:
Do friends’ opinions have the same effect on one’sdecision in higher risk situations versus lower risk situa-tions?
•
R4:
How does market share evolve over time in the pres-ence of different aspects of social influence?To answer the first three questions, we performed user stud-ies on Mechanical Turk using positive and negative opinionsfrom friends, as well as ratings from the general public; thelatter was represented by the average number of stars. Wefind that the choice between two options fits a logit model.Our major findings are that (1) an additional recommenda-tion star has a much larger effect than an additional friend’srecommendation on the probability of selecting an item, (2)negative opinions from friends are more influential than pos-itive opinions, and (3) people exhibit more random behav-ior in their choices when the decision involves less cost andrisk. Our results are quite general in the sense that peo-ple across different demographics trade off recommenda-tions from friends and ratings in a similar fashion. Withrespect to R4, we performed simulations and observed thatthe variability of outcomes increases when the variability of ratings increases, and inequality increases when individualsbecome more selective in their recommendations.
RELATED WORKEmpirical Studies on Social Influence
A number of empirical studies have considered the effectof social influence in various contexts, including prescrip-tion drug adoption and use[16], viral and word of mouthmarketing [4], health plans [23], crime rates[10], and in-
vestment in the stock market [1, 15]. Tucker et al. focus onhow quality and popularity influence decisions on a weddingwebsite [25].Guo et al. study the role of social networks in online shop-ping[12]. Their study is performed on data from a leadingChinese e-commerce website, which deploys a messagingsystem among its users. The data contains users’ purchasebehavior as well as message exchanges among them; how-ever, the authors did not have access to the content of themessages. The main finding of the paper that is related toour work is that, a message sent from a user that has pur-chased an item increases the probability of purchase by theuser who got the message by 1%. The causal relation is notdefinite, since the content of the messages is not known.Salganik et al. study the effects of social influence over timeon the popularity of songs on an artificial online music mar-ket [22]. They find that social influence may lead to unpre-dictable and unequal outcomes. This work is related to ourstudy of the dynamics of market share in the presence of so-cial influence; however, we consider two sources of socialinfluence. As we discuss below, the dynamics of social in-fluence have also been analyzed theoretically.
Theoretical Models on Dynamics under Social Influence
Theoretical models focus mainly on the dynamics of mar-ket share in the presence of social influence: before makinga decision, an individual observes what choices others havemade in the past[5, 8,26]. More related to our study are [7,
9]. Bendor et al. presented a model that considers individu-als making choices for which the objective evidence of theirrelative value is weak [7]. In that case, individuals tend torely heavily on the prior choices of people in similar roles.They then showed that the dynamics of market shares leadto outcomes that appear to be deterministic in spite of be-ing governed by a stochastic process. In other words, whenthe objective evidence for the value of the choices is weak,any sample path of this process quickly settles down to afraction of adopters that is not predetermined by the initialconditions: ex ante, every outcome is just as (un)likely asevery other. In[9], the authors define the decision functionas a stochastic function of the social influence (in terms of market share) and the inherent quality of the product. Theirmain result is that market shares converge to an equilibrium.These processes only takes into account one source of socialinformation, namely the number of others which have cho-sen one item over the other. By contrast we consider howmarket share evolves in a setting with two sources of socialinfluence, namely opinions from friends and ratings from thegeneral public.
Sources of Social Influence
Mostpreviousworkmodelchoosingbetweenitemsasafunc-tion of two factors: the individual’s own judgment and socialinfluence. However, different groups of people have differ-ent levels of influence on one’s decision. For example, whenone wants to choose a movie to watch, some people relymore on what the general ratings for the movie are and someothers more on what their friends who have already watchedthe movie say about it. Currently, with the availability of so-cial network data, in many online settings, recommendationsfrom friends are available in addition to ratings and reviewsfrom other people. Due to this fact, in this work we focus onthese two sources of social influence.
METHOD
In this section we describe the experimental design and re-port some statistics about the data we collected. Our goal isto study how people trade off information from friends andthe general public when choosing between two items. More-over, our experiments allow us to compare a setting wherethe information from friends consists of positive recommen-dations to a setting where the information from friends con-sists of negative opinions.2
Furthermore, we compare people’s choices with respect totwo types of decisions: one that involves a monetary cost(booking a hotel) and a low risk decision that involves nomonetary cost (watching a movie trailer). We chose bookinghotels because the user cannot go and check it out before de-ciding and should rely on the information she gets from oth-ers. Similarly, a user may not have any information about amovie trailer before she watches it. We can think of the set-ting with the movie trailers as a less serious decision, sinceit involves less cost (just a couple of minutes of one’s time)and risk. Users often make choices of this type online, e.g.,when watching Youtube videos, clicking on a link or ad, etc.In total, we conducted three user studies: booking a ho-tel with positive recommendations from friends (Study 1),booking a hotel with negative opinions from friends (Study2) and watching a movie trailer with positive recommenda-tions from friends (Study 3).
Online survey
To collect the data we conducted the three studies in theform of surveys on Amazon’s Mechanical Turk during Julyand August 2011. Amazon’s Mechanical Turk (MTurk) isa crowdsourcing online marketplace where
requesters
usehuman intelligence of
workers
to perform certain tasks, alsoknownasHITs(HumanIntelligenceTasks). Workersbrowseamong existing tasks and complete them for a monetary pay-ment set by the requester [18].Once a worker completes the task, the requester can decidewhether to approve it. In particular, if the requester believesthat the worker completed the task randomly, he can rejecther work. In that case, the worker does not get paid for theparticular task and her approval rate is decreased. For ourstudies, we only hired workers that had approval rates of over 95%, that is, workers who had performed well in thepast.We asked each worker to put herself in the following hypo-thetical situation: she is about to book a hotel (resp. watcha movie trailer) an on e-commerce site (resp. online), andamongtheoptions, shehascomedowntotwobetweenwhichshe is indifferent. The website has an underlying social net-work of friends (or it runs on top of an online social net-work). For each of the two options, we provide the followinginformation:(i) the overall rating (in terms of stars on the scale of 1 to5) based on ratings from a large number of previous cus-tomers (resp. users) in the case of selecting which hotel tobook (resp. which movie trailer to watch)(ii) the number of friends who recommend (resp. have neg-ative opinions about) the option in the case of positive(resp. negative) recommendationsFor each question, the option that has more stars is the onethat is less recommended by friends; that is, we did not usea pair of options where one clearly dominated the other.
Figure 1. Sample questions for Studies 1, 2, and 3 (top to bottom).
“Which one would you choose?”
and 2)
“Which one do you think others would choose?”
The answer to the first question provides informa-tion on how a particular worker trades off information fromfriends and the general public when making her choice. Weonly use the data from this question for our analysis.Even though we do not use the answers to the second ques-tion for our analysis, we included it in the survey for two rea-sons. First, by asking the second question the subjects feltthat the survey’s goal was to study how people think theirdecisions are similar to the others, weakening the reactivityeffect[14]. Second, we offered three bonus payments ($5 or$10 each, which is 50 or 100 times the amount we paid foreach HIT) to the three workers whose answers to the secondquestion was closest to the others’ answers to the first ques-tion in order to deter workers from answering randomly.
Validity check.
Apart from using the bonus to incentivizeworkers to put some thought when answering the survey, weincorporated two “validity check” questions in the survey. If a worker did not answer these two questions correctly, wedid not include any of her responses in our analysis (and re- jected her work). In the first such question one option clearlydominated the other in terms of both the number of starsand friends’ recommendations (one option had only one starand 10 negative recommendations from friends, whereas theother had more stars and 10 positive recommendations fromfriends). The second test question was a repeated questionwith the order of choices reversed and also with no graphics.Responses with inconsistencies in the answers to these two3
Search History:
Result 00 of 00
00 results for result for
p.
More From This User
Notes
Load more
