Understanding Feedback Expectations on Facebook

Nir Grinberg1,2, Shankar Kalyanaraman2, Lada A. Adamic2 and Mor Naaman1

nir@cs.cornell.edu, kshankar@fb.com, ladamic@fb.com, mor@jacobs.cornell.edu1Jacobs Institute, Cornell Tech 2Facebook Inc.

ABSTRACTWhen people share updates with their friends on Facebookthey have varying expectations for the feedback they will re-ceive. In this study, we quantitatively examine the factorscontributing to feedback expectations and the potential out-comes of expectation fulfillment. We conducted two sets ofsurveys: one asking people about their feedback expectationsimmediately after posting on Facebook and the other askinghow the amount of feedback received on a post matched theparticipant’s expectations. Participants were more likely toexpect feedback on content they evaluated as more impor-tant, and to a lesser extent more personal. Expectations alsodepended on participants’ age, gender, and level of activityon Facebook. When asked about feedback expectations fromspecific friends, participants were more likely to expect feed-back from closer friends, but expectations varied consider-ably based on recency of communication, geographical prox-imity, and the type of relationship (e.g. family, co-worker).Finally, receiving more feedback relative to expectations cor-related with a greater feeling of connectedness to one’s Face-book friends. The findings suggest implications for the theoryand the design of social network sites.

Author KeywordsFeedback expectations; Computer-mediated communication;Social Media; Facebook; Information Sharing

ACM Classification KeywordsH.5.m. Information Interfaces and Presentation (e.g. HCI):Miscellaneous

INTRODUCTIONPosting and receiving feedback shapes the experience of peo-ple on social media. Feedback, whether it is expressed vialightweight one-click communication or more, carries socialvalue that motivates people to post, provides social and emo-tional support, and shapes relationships over time [12,13,19].While a considerable body of work has studied the role offeedback in social network sites [12, 15, 19, 26, 37, 46], littleresearch examined the expectations for feedback people havewhen sharing content to their social network. In this paper

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full cita-tion on the first page. Copyrights for components of this work owned by others thanACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re-publish, to post on servers or to redistribute to lists, requires prior specific permissionand/or a fee. Request permissions from permissions@acm.org.

CSCW ’17, February 25-March 01, 2017, Portland, OR, USA

© 2017 ACM. ISBN 978-1-4503-4335-0/17/03. . . $15.00

DOI: http://dx.doi.org/10.1145/2998181.2998320

we focus on the feedback expectations associated with post-ing content on Facebook, and the way that expectations varyfrom one person to another, are dependent on the propertiesof the post, and are impacted by the relationship to other in-dividuals.

Expectations are an important measure that guides social be-havior and attitude, which can inform the design of socialnetwork sites. Expectations motivate us to take action andhelp us choose among alternatives. For example, expectationof feedback is a key motivating factor for participation in on-line forums, contribution to Wikipedia, and posting on socialmedia [17, 28, 31, 39, 41]. In a recent study, we showed thatpeople visit Facebook more often after posting a status up-date, potentially in expectation of feedback, even when therewas no evidence of actual feedback received [27].

Despite the importance of feedback expectations, previous re-search did not directly model people’s expectations for feed-back from their online social networks, nor did it examinethe implications of expectation fulfillment. Existing theoriesof interpersonal communication such as Expectancy Viola-tion Theory [5, 8] do not immediately translate to expecta-tions from online social networks, where feedback is oftenaggregated and knowledge of viewership is lacking or incom-plete. Previous studies on social media touched on several as-pects related to expectations such as the “imagined audience”,perceived audience size, feedback preferences, norms evolu-tion and violation [3, 13, 32, 33, 35, 44], but did not directlymodel feedback expectations. Without a clear understandingof feedback expectations it is unclear how actual feedbackis perceived and how that feedback (or lack thereof) affectspeople’s experience on social media.

This paper examines people’s expectations for receivingLikes and Comments on Facebook, and the relation betweenfulfillment of expectations to feeling connected to one’s Face-book friends. We build on Expectancy Violation Theory(EVT) [5,8] as inspiration for the conceptual framework usedin this work. We conduct a comprehensive, in-context exam-ination of the factors associated with feedback expectationsimmediately after posting on social media. Our investigationborrows from EVT the key elements of the model that con-tribute to expectation: properties of the communicated con-tent properties, characteristics of the individual who posted it,and individual’s relationships to others on the platform. Notonly do we look at factors that contribute to expectations, wealso study the fulfillment of feedback expectations and its re-lation to feeling of connectedness to one’s Facebook friends,an important outcome for individual well being [15, 29].

To this end, we use two large-scale surveys to ask peopleabout feedback expectations, fulfillment of expectations, andconnectedness to friends. First, we surveyed people imme-diately after posting on Facebook and asked them about theirfeedback expectations on that particular post, both in terms oftotal feedback and from specific Facebook friends. We com-plemented survey responses with de-identified, aggregatedlog data to understand how the characteristics of the individu-als, posts, and interpersonal relationships are associated withfeedback expectations. Using this dataset, we built predictivemodels of feedback expectations. In addition, we conducted aseparate survey, asking participants about an earlier post theymade, and how the amount of feedback received compared totheir expectations. We also asked participants in this laggedsurvey how connected they feel to their Facebook friends inorder to establish a link between fulfillment of expectationsand connectedness.

This study offers a general framework for thinking aboutfeedback, behavior and attitude on social media in the contextof expectations. We identify the significant factors associatedwith higher than usual feedback expectations on a post andthe important properties of relationships linked with expecta-tions from specific friends. In addition, we show an associa-tion between the congruency of feedback and expectations ona post to an important outcome – individuals’ connectednessto their friends. Last, our predictive models can be used inpractice to evaluate how well people’s expectations are met,and to explore ways of potentially improving the experienceof people when posting on social media platforms.

RELATED WORKPrevious work identified several benefits of posting on socialmedia, among them are self-expression, relational develop-ment, social validation, and approval [1]. Social media usehad been shown to impact both social capital and well be-ing [14–16, 19, 20, 30]. Many of the benefits of social me-dia use come through feedback mechanisms such as Likesand Comments on Facebook. Support and help via feedbackare important for alleviating loneliness [15, 29], getting emo-tional support after losing a job or when sharing emotionalcontent [11,12], enabling information seeking [26,37], main-taining relationships [19,46,47], and more. While feedback isa necessary component for all of these benefits, as Bazarovaet al. point out, it is one’s subjective satisfaction from thefeedback received that determines its value for the communi-cating individual [2].

Other work studied people’s perceptions around audienceand feedback in computer-mediated communications. Peoplehave an imagined audience in mind when posting to friendson social media [32, 33], but as Bernstein et al. showed, peo-ple’s mental model of audience underestimates the number ofpeople who actually see their posts and overestimate the rateat which friends give Likes and Comments [3]. Wang et al.found that posters and outsiders evaluate Facebook updatesdifferently, particularly around topics of self-presentation andrelationships [51]. Perceptions about feedback are also highlysubjective – different people may have different interpreta-tions of social interactions online. A recent study by Scis-

Figure 1. The conceptual framework used in this work, inspired by Ex-pectancy Violation Theory by Burgoon [6].

sors et al. examined the perceptions around lightweight com-munications on Facebook and found that most people do notconsider receiving “enough” Likes as important and assign-ing importance to getting enough Likes is positively corre-lated with high levels of self-monitoring and negatively cor-related with self-esteem [44]. Previous work did not directlytie the diverse perspectives people have about activities on so-cial media to expectations, which offer a more general viewof social behavior as we describe next.

The notion of expectations is central to many theories abouthuman behavior as it proposes a general framework for un-derstanding behavior and attitude. The conceptual frameworkused in the current work was inspired by Expectancy Viola-tion Theory (EVT) in communication [5, 8]. EVT was orig-inally developed based on studies of proxemic behavior inface-to-face communication and was later extended to a vari-ety of behaviors [5, 7, 9, 10, 49]. Burgoon defines expectan-cies as “enduring pattern of anticipated behavior”, which de-rive from three classes of factors: communicator (e.g. de-mographics, personality, appearance), relationship (e.g. fa-miliarity, similarity, status difference), and context (e.g. pri-vate/public environment, the message communicated) [6].According to EVT, expectancies “serve as framing devicesthat define and shape interpersonal interactions . . . [and] sig-nificantly influence how social information is processed”.The congruency between enacted behavior and expectationsdetermines how behavior is perceived, the impression peoplehave of each other, and the outcomes of the interaction. Pre-vious work applied EVT to study norm evolution and viola-tion on Facebook [4,21,35], but focused more on incidents ofnorm violations by individual friends, well-aligned with theoriginal theory. However, as EVT focuses on single individu-als’ behavior, it does not directly apply to studying aggregateexpectations as we do here. Instead, we use the overall frame-work of EVT as inspiration for our research model.

Figure 1 shows the conceptual framework used in this work.At the top are three categories of factors that mirror the orig-inal EVT model [6], which we also expect to affect expecta-tions of feedback: individual, relationship, and context prop-erties. The characteristics of the individual can include de-

mographics, personality traits, and more. Relationship prop-erties may consist of tie strength between two individuals,differences in status, shared interests and other factors thataffect interactions between people. Context includes the ad-ditional factors needed to describe a particular situation. Inthe current study, context primarily consists of properties ofthe posted content and past interactions on previous posts.The three categories of factors at the top of the figure jointlycontribute to people’s expectations, which are then comparedagainst the actual feedback received. If, for example, theamount of feedback an individual received from friends ex-ceeded their expectations, they may be more inclined to postin the future, and may feel more connected to their friends.In contrast, unsatisfying feedback experiences may be oneof the mechanisms behind departure from online social plat-forms [55]. As proposed by EVT, and suggested here, thecongruency between the observed feedback and expectationsdetermines people’s attitude and subsequent behavior.

Much like the early studies of proxemic behavior in face-to-face communication [5, 9], we seek to understand the impor-tant factors behind feedback expectations on social networksites (top of Figure 1). Our first two research questions fo-cus on characteristics of the individual (left) and the postedstatus update, i.e. the context (right). It is important to dis-tinguish expectations across people because individuals expe-rience social media very differently: seeing different sets ofstories and a variety of interactions with them (see [53] foran example). Similarly, distinguishing between expectationsfor different posts is important because posts vary in contentand importance to the individual [38, 50]. Therefore, our firsttwo research questions are:

RQ1: What are the characteristics of individuals (e.g. age,gender, etc.) that affect feedback expectations?

RQ2: What are the properties of posts (e.g. length, topic,etc.) that affect feedback expectations?

Next, we study the third category of factors in Figure 1 thataffects feedback expectations – relationship properties. Pre-vious studies showed that people have different preferencesfor feedback from strong and weak ties on Facebook [13, 35]and from different social groups (e.g. close friends, family,co-workers, etc.) [44]. However, prior work did not directlyexamine expectations from specific friends (rather than ab-stract social groups), on a specific post, for different formsof feedback (e.g Likes and Comments), and at the time ofposting (rather than retroactively). The complexity of socialrelations calls for a joint examination of all the above aspectsof relationships in order to gain a more comprehensive un-derstanding of relationship expectations. Therefore, our thirdresearch question is:

RQ3: What are the relationship properties (e.g. based onrelationship type, tie strength, age difference, etc.) that affectfeedback expectations?

Finally, motivated by work on social capital and well be-ing [14–16, 18–20, 30], we investigate one potential outcomeof fulfilling feedback expectations. Self-Determination The-ory describes a basic human need for relatedness – to belong

and feel connected to the people, group or culture sharing theindividual’s goals [43]. In offline settings, greater related-ness was shown to contribute to daily well being [42]. Sincegetting feedback is one of the key motivating factors for par-ticipation online [17, 28, 31, 39, 41] it is likely that the ac-tual feedback received would affect people’s satisfaction withtheir relationships. In fact, two recent studies point in that di-rection, showing that when people share emotional content onFacebook, friends respond with more emotional and support-ive Comments, which is associated with greater satisfactionwith communication goals [2, 11]. Previous findings, how-ever, did not extend beyond emotional content or tie receivedfeedback to relationship satisfaction. The conceptual frame-work in Figure 1 borrows from EVT to suggest that both ex-pectations and observed behavior (whether they are met) willaffect outcomes. Specifically, our question is:

RQ4: How does the fulfillment of feedback expectations re-late to feeling connected to one’s Facebook friends?

With these questions in mind, we performed a quantitativemixed-methods study of feedback expectations on Facebook,as we describe next.

METHODSIn this section we describe the mixed-methods approach weused in order to address our research questions about feed-back expectations. Following previous work, we use surveymechanisms to ask people about their subjective perceptionsof social media activities [3,44]. We then complement surveyresponses with observational log data to gain better under-standing of the contextual factors that explain expectations.

SurveysWe conducted two online surveys by recruiting participantson Facebook’s web interface (Facebook.com)1. The first sur-vey asked participants about their expectations for feedbackfor a specific post. The survey was offered to people immedi-ately after they posted a status update, as a pop-up dialog. Werefer to it here as the Immediate survey. The second surveywas offered to people 23-27 hours after they posted a statusupdate, as a banner on their Facebook page. This Lagged sur-vey asked participants about fulfillment of expectations forthat day-old post. The surveys were limited to English speak-ers in the U.S. with 20 friends or more that did not participatein any survey conducted by Facebook in the six months priorto ours. Participation in both surveys was voluntary and didnot involve compensation in any form. The surveys ran for 20days, between July 27, 2015 to August 16, 2015. The samplesfor both surveys were drawn from the same sampling frame,but were non-overlapping (people were invited to participatein either one of the surveys, but not both2). The response ratevaried between the two surveys, with a lower response rate of

1Due to the length and complexity of surveys we left the develop-ment of mobile versions to future work.2we choose this design over repeated surveys of the same peoplefor two reasons: 1) to eliminate the potential bias that answeringquestions in the Immediate survey influences answers to the Laggedsurvey, and 2) lower the burden on people of answering repeatedsurveys day after day.

Average difference Immediate Laggedfrom a random sample survey survey

# Log in days out of 28 3.4∗ 4.4∗# Log in days out of 7 1.0∗ 1.2∗Friend count 151∗ 6Gender (% Female) 0.0% 5%∗

Age -0.7 1.9∗

N 2788 4032Response rate 33% 78%∗p<0.001 using 2-sample t-test comparing each surveyseparately to the random sample.

Table 1. Usage and demographic statistics of survey participants rela-tive to a random sample of English speakers in the US who logged-in toFacebook on the web at least once in the month prior to both surveys.

33% for the Immediate survey versus 78% for the Lagged sur-vey. The difference was probably due to the Immediate sur-vey’s disruptive nature as a pop-up immediately after postingcompared to the more organic banner of the Lagged survey3.In total, 2788 people completed the Immediate survey and4032 completed the Lagged survey.

Table 1 summarizes key differences between demographicand usage characteristics of survey participants and other peo-ple on Facebook. We compare the participants in the surveysto a random sample of US, English speaking individuals whoaccessed Facebook’s web interface at least once in the monthprior to our surveys. As highlighted in Table 1, survey partic-ipants are more active Facebook users compared to the ran-dom sample, logging-in to Facebook about 4 additional daysover the course of 28 days. Participants in the Immediate sur-vey have more friends on average, while more females andolder individuals participated in the Lagged survey. Overall,we conclude that the participants in the surveys are slightlymore active on Facebook than a random sample, but no othermajor differences are evident.

MeasuresWe now turn to describe the measures included in our surveysabout feedback expectations and their fulfillment. Certaincommon elements are likely to affect both the expectationson a particular post, measured in the Immediate survey, andfulfillment of expectations, measured in the Lagged survey.Therefore, we include in both surveys the following 5-pointLikert scale questions:

• Connectedness: how connected do you feel to your Face-book friends? (1=very disconnected, 5=very connected).

• Importance: how important is this post to you comparedto your average post? (1=much less than usual, 5=muchmore than usual).

• Personal: how personal is this post? (1=not at all, 5=verypersonal).

3A pop-up was necessary in the Immediate survey to capture re-sponses before any other action is taken on the site. We opted fora banner in the second survey because a pop-up in this case wouldhave been out-of-context and hence much more disruptive to the userexperience.

Figure 2. The Friends Grid question that was populated with a stratifiedsample of the participant’s friends (in random order). Participants wereasked to specify for each friend whether they expect a Like and/or aComment on their latest post, shortly after posting it. Profile picturesand names are blurred in this figure, drawn for demonstration from thefirst author’s account, in order to preserve individuals’ privacy.

In addition, the Immediate survey asked participants aboutfeedback expectations for the post:

• Post-level expectations: how many Likes and Commentsdo you expect to get on your latest post? (1=far fewer thanusual, 5=far more than usual).

• Friend-level expectations: we presented a personalizedFriends Grid, a two column grid populated with a sampleof up to 22 of the participant’s friends (described below),asking participants to indicate whether they expect a Likeand/or a Comment from each individual friend.

The Lagged survey, on the other hand, first showed partici-pants their post along with its feedback (as it appears in NewsFeed). Then, it asked participants about:

• Fulfillment of feedback expectations: how did the Likesand Comments received so far match your expectations?(1=far fewer than expected, 5=far more than expected).

Figure 2 shows the layout of the Friends Grid used in the Im-mediate survey. In order to get a more balanced sample offriends with and without feedback expectations we includedin the Friends Grid a stratified sample of the participant’sfriends. We chose a stratified sample of friends over a ran-dom sample because a random sample is mostly dominatedby weak ties that may not be associated with any feedbackexpectations. Our stratification randomly picked friends ofthe participant from Facebook lists the person may maintain(close friends, acquaintances), friends with overlapping pro-file information (same workplace, college, high-school, hometown or current city), self-reported family ties (parent, child,sibling, spouse), and most recent interactions (last Like orComment, given or received). In addition to sampling a ran-dom friend from each of these groups we included the friendfrom each group that the person communicated with most fre-quently (without introducing duplication).

Checkboxes in the Friends Grid may remain unchecked be-cause the participant had no feedback expectations from thatfriend or because it is the default option. To address this bias,we use an assumption that people scan items visually from

top to bottom, and from left to right4. This linear scanningassumption has been studied extensively in the analysis ofsearch results and was shown to improve results relevance[45]. In our case, a friend is associated with “no expectation”only if the participant made a selection in a lower position inthe grid or to the right. After processing the raw responses,our dataset consisted of 568 participants who labeled 5,256of their friends with expectations for only a Like (30.1%),only a Comment (3.8%), Like and a Comment (11.4%), or nofeedback (54.7%).

Log dataWe complemented the survey responses with Facebook’sserver logs in the 12 weeks prior to the survey in order tobetter understand the context of reported expectations. Alllog data were observational – no experiment was performedand no individual’s experience on the site was altered. Thelog data includes the posts that participants were asked about,profile information such as education or work history, andfriendship information. We took significant steps to ensurepeople’s privacy: all data were de-identified and analyzed inaggregate such that no individual’s text could be viewed byresearchers.

POST-LEVEL EXPECTATIONSThis section addresses our first two research questions aboutfeedback expectations of different people (RQ1) on differentposts (RQ2), and then focuses on estimating how accuratelythese expectations can be predicted.

We address our first two research questions by fitting a logis-tic regression to the reported expectations on a post (from theImmediate survey) with covariates that describe the individ-ual, the feedback received on the individual’s previous posts,and the newly posted content. We include both individualand post-level covariates in the regression model in order tounderstand how each group of factors varies while the oth-ers are held constant. For example, in addressing RQ1 weexamine the characteristics of individuals that associate withhigher than usual expectations while holding the properties ofthe post constant at their mean value. Our dependent variablein the regression is positive whenever a person reported ex-pecting more than usual feedback on her post (4 or 5 on theLikert scale, where 3 was labeled “about the same as usual”)and negative otherwise. We focus in particular on cases withhigher than usual expectations since these are most likely toresult in an unsatisfying experience when unmet.

Individual differences: Different people are likely to havedifferent expectations. Therefore, we include in the re-gression information about age, gender, tenure on Facebook(years since creating the Facebook account), friend count,and the number of days in the past week that the participantlogged in to Facebook (L7). Age, tenure and L7 were cen-tered; friend count was log-transformed (base 2) to account

4The left to right assumption is reasonable since all of our partici-pants are English speakers.

for skew prior to standardization and the rest were centeredand scaled using two standard deviations5.

Past feedback: Feedback on previous posts is also likely toaffect expectations. Therefore, we compute the median andinterquartile range (IQR)6 for the following measures of feed-back based on the individual’s posts in the prior 12 weeks:number of Likes per post, Comments per post, Likes perview, and Comments per view. We also include the numberof Likes and Comments on the most recent post of the indi-vidual since these might have greater impact on expectations.All variables were standardized as described earlier, exceptfor the number of Likes/Comments per view which were log-transformed prior to standardization.

Content properties: Posts interest people to various degreesand therefore result in different expectations. Our datasetcontained only few posts with photos and therefore we ex-cluded those and focused only on textual posts. Our contentproperties include a variety of features: basic (word count,does the post contain a URL?), subjective assessments (howpersonal/important is this post?), topics, and emotional di-mensions. The text was preprocessed and converted to low-ercase, tokenized, and punctuation, stopwords and terms ap-pearing in less than 5 posts were removed. All of the tex-tual features were extracted using standard scripts over de-identified content such that no member of the research groupexamined any individual post.

We used Supervised Latent Dirichlet Allocation (sLDA) tomodel the topics that appear in posts [34]. The benefit ofsLDA over “standard” unsupervised LDA is that topics are fitto better separate class labels. In our case, we used higherthan usual feedback expectations as binary class labels. Weexperimented with different numbers of topics ranging from10 to 60 (in increments of 10) and found no significant im-provement in log-likelihood beyond using 20 topics. Usingthe trained sLDA model (using 10-fold cross validation) weget a single probability that represents the likelihood that apost is associated with higher than usual feedback expecta-tions. We include the sLDA prediction in our final model.

Emotional dimensions were extracted using the 2007 versionof Linguistic Inquiry and Word Count (LIWC) [40]. Mostfine-grained LIWC categories (e.g. filler words) had no orvery little support in our dataset and therefore we only in-cluded high-level categories such as function words, positiveand negative emotions, social terms, achievement terms, andtime orientation information (references to past, present or fu-ture). All of the LIWC features were included in the form ofproportion of the total number of words in the post.

Without limiting the number of content properties in our re-gression, we run the risk of overfitting the data and findingspurious correlations as statistically significant. We addressthis concern in two different ways. First, we keep the numberof covariates small relative to the number of survey responses5Unless specified otherwise, all continuous covariates were centeredand scaled by two standard deviation in order to put them on roughlythe same scale as untransformed binary variables [23].6a robust measure of dispersion, defined as the difference betweenthe upper and lower quartiles.

Higher than usual feedback exp. ∼ Coef. SEIntercept −1.56∗∗∗ .15Individual differencesAge (years) 0.0098∗∗ .0037Is male 0.13 .13Tenure (years) −0.14∗∗∗ .03Log2(Friend count) 0.93∗∗∗ .14Connectedness 0.58∗∗∗ .13L7 −0.041 .055Posts per day −0.002 .024

Past feedback:Likes(last post) −0.26 .20Comments(last post) 0.02 .16IQR(Likes per post) 0.01 .21Median(Likes per post) 0.19 .22IQR(Comments per post) 0.48∗ .19Median(Comments per post) −0.38 .21IQR(Likes per view) 0.00 .19Log2(Median(Likes per view)) 0.09 .15IQR(Comments per view) −0.19 .22Log2(Median(Comments per view)) 0.33∗ .16

Post:Importance 1.22∗∗∗ .15Personal 0.36∗ .14Has link −0.22 .20Log2(Word count) −0.08 .17sLDA prediction 0.26∗ .13LIWC:funct 0.18 .16posemo −0.27 .16negemo 0.19 .12social 0.04 .14percept 0.22 .12bio −0.17 .13achieve 0.32∗∗ .12past −0.19 .15present −0.14 .15future −0.10 .15

P(Y |X) 20.47%Log Likelihood -813.4Akaike Inf. Crit. 1886.8

N = 2, 788; ∗p<0.5; ∗∗p<0.01; ∗∗∗p<0.001Table 2. Coefficients of Bayesian logistic regression for having higherthan usual feedback expectations on a post.

(2,788) by limiting the number of topics and emotional di-mensions we include. Second, we use Bayesian logistic re-gression with a non-informative Cauchy prior (0 median and2.5 scale) to pull regression coefficients slightly towards zeroapriori, but allow for large coefficients when the data doessupport it [24]. The coefficients in a Bayesian logistic re-gression have the same interpretation as those of a “standard”logistic regression.

FindingsTable 2 shows the resulting coefficients of the Bayesian lo-gistic regression with the binary dependent variable of havinghigher than usual feedback expectations on a post. All vari-ance inflation factors (VIF) were less than two, indicating thatmulticollinearity is not an issue in our independent variables.The logistic regression assigns a probability of 20.47% forhaving high expectations to the average person (designatedby P(Y |X) in the table), closely matching the empirical pro-portion in the dataset with less than 0.01% in difference. Sig-nificant coefficients appear in all three categories of features,as we describe next.

In terms of individual differences (RQ1), four properties ofthe person posting the content are statistically significant:age, tenure on Facebook, number of friends and Connect-edness. Each additional year of age is associated with aexp(0.0098) = 1.009 = +0.9% increase in the odds of havinghigher than usual expectations. More significantly, doublingthe number of friends on Facebook and feeling more con-nected to friends increases the odds by 38.6% and 29.6%, re-spectively. In contrast, each additional year of having a Face-book profile is associated with a -13.3% decrease in the oddsof having higher than usual expectations.

Past feedback also contributes to higher feedback expecta-tions, but only through Comments. There is an increase of11.4% in the odds of high expectations for every additionalComment in the individual’s interquartile range, and 23.9%increase in odds when doubling the median rate of Commentsper view. None of the measures based on past Likes or feed-back on the last post were significant. These findings sug-gest that greater variability in past Comments (but not Likes)contributes to higher feedback expectations, and that peoplelearn over time the rate at which their friends comment ontheir posts even without explicit knowledge about views.

Several aspects of the post’s content affect feedback expec-tations (RQ2). First, increases in the Importance and Per-sonal scales translate into higher expectations (78% and 16%increase in odds, respectively)7. The fact that personal andimportant posts are associated with considerably higher feed-back expectations highlights the value in better understandingthese subjective assessments by people, which is beyond thescope of the current work. Second, sLDA predictions basedon broad topics found in the post increase the odds ratio mod-estly (+5.5%). The only emotional aspect that is significantis the occurrence of achievement terms. None of the otherLIWC dimensions such as positive or negative emotions sig-nificantly affect expectations.

In summary, our findings show that individual differences,past feedback and posts’ content are linked to higher thanusual feedback expectations on a post. Age, number offriends and Connectedness are positively associated high ex-pectations, while tenure on Facebook is negatively associated.Past Comments (and not Likes) affect expectations, and poststhat are important, personal, and refer to achievements havehigher expectations. Overall, we see that different peoplehave different expectations for different posts, and that pastbehavior of friends (mostly in commenting) affects future ex-pectations.

Next, we evaluate how well feedback expectations for a postcan be predicted in order to assess the practicality of integrat-ing feedback expectations into the design of social networkssites.

Predicting post-level expectationsIn this part we examine how well different subsets of the fea-tures from Table 2 predict the feedback expectations that anauthor has for her Facebook post. We test the following three7recall that all of surveyed content had the same privacy settings:posts shared by participants with all of their Facebook friends.

Higher than usual exp. ∼ AUC P@R5 P@R50 P@R95

baseline:last post percentile

49.0

53.057.0

22.6

29.536.5

19.6

21.322.9

19.3

20.622.0

individual differencesAge + Gender +# Friends

60.2

62.665.0

33.8

55.978.0

25.0

29.534.0

18.8

21.424.0

+ past feedback61.0

63.465.7

39.4

51.663.8

26.5

30.233.8

21.0

22.524.0

+ content:67.4

70.773.9

57.9

69.681.3

38.7

43.949.1

21.9

23.324.7

+ self-reports:Connectedness +Post importance + Personal

75.4

77.780.0

68.6

81.794.7

46.6

50.454.2

23.0

25.127.2

Table 3. The predictive power of different feature sets obtained usingeither glmnet or gbm. P@R stands for precision at different recall levelsof 5, 50 or 95 percent. Numbers above/below in each cell represent 95%confidence intervals.

predictive models as implemented in R: Elastic-Net Regular-ized Generalized Linear Models (glmnet [22]), GeneralizedBoosted Regression Modeling (gbm [54]), and Support Vec-tor Machine (from the e1071 package [36]).

Table 3 summarizes the results of 10-fold cross-validation ofthe best-performing model for each feature-set. Our base-line, which uses a personalized percentile of feedback re-ceived on the last post relative to the individual’s posts inthe prior 12 weeks, only performs marginally better than ran-dom. The predictive performance improves significantly overthe baseline when including user information (62.6% AUC),adding past feedback (63.4% AUC) and finally content fea-tures, reaching 70% AUC. Using log data alone, the modelidentifies posts with higher than usual feedback expectationswith a precision of about 70% when retrieving only 5% ofposts with high expectations. In other words, at the level of5% recall, the model will return one out of 20 posts with highexpectations and would correctly identify the expectations forseven out of every 10 posts returned. As shown in Table 3,precision naturally deteriorates when increasing the recall to50% or 95%. Last, including participants’ answers to surveyquestions improves performance even further, showing thatsubjective information is important and not fully captured byother variables. In particular, feeling connected to friends,knowing the post’s importance and how personal it is, are allimportant predictors of high feedback expectations.

Next, we address our third research question about the charac-teristics of relationships that affect expectations for feedbackfrom one friend and not another.

FRIEND-LEVEL EXPECTATIONSIn this section we use the responses from the Friends Gridin the Immediate survey to address RQ3: which properties ofrelationships affect expectations for feedback from differentsocial ties? We examine how similarities and differences be-tween two people as well as long-term and short-term com-munication patterns relate to expectations of feedback fromthat person. We fit two separate logistic regression models,

one for Like expectations, the other for Comment expecta-tions, on the same set of responses and relationship features.

Before we describe in greater detail the relationship prop-erties used for addressing RQ3 we first specify the controlsincluded in our models. We control for the order in whichfriends appeared in the Friends Grid, the characteristics ofparticipants and the properties of posts. Despite the randomorder of friends in the Friends Grid, certain positions in thegrid may receive more attention. Therefore, we include theposition information in our model relative to the top (1-top to11-bottom) and relative to the left (0-left, 1-right). In addi-tion, as we saw in the previous section, individuals have dif-ferent expectations for different posts. Since our focus here ison dyadic properties that affect expectations from a specificfriend we control for non-dyadic features that were identifiedas significant in addressing RQ1 and RQ2. The complete listof control variables can be found in Table 4.

We describe below the three families of features we consid-ered in our model for friend-level expectations: dyadic differ-ences, topical similarity, and relationship properties.

Dyadic differences: the relative differences between partic-ipants and their friends may affect expectations. We includein our model demographic and activity information about theparticipant’s friends in relative form, e.g. the difference be-tween the friend’s age and the participant’s age. We also con-sidered interactions between covariates, since different sub-populations may have different expectations. For example,age difference may be linked to higher expectations in gen-eral, but the gap can matter differently for younger and olderadults.

Dyadic topics similarity: the perceived interest of a friendin a topic is likely to affect expectations for a response whenthe topic is discussed. Here, we develop a set of featuresaimed at capturing the overlap in topical interests of partici-pants and their friends, and quantifying how a particular postfits into this overlap. For every individual, we aggregatedthe topics of all posts that they interacted with into a vectorof high-level topical interests. Then, we computed interestssimilarity using cosine similarity between the participant (de-noted as u) and their friend (denoted as f ), and between thepost (denoted as p) and the friend’s interests. Due to the largeamount of posts involved, we used a TagSpace model to ex-tract the proportions of topics in posts [52], and reduced thetopic space to 20 high-level most frequent topics (e.g. music,entertainment, education).

We also compute friend specificity to a topic in two differ-ent ways. First, we calculated the percent of the participant’sfriends that are highly interested in each of the post’s topics8

(denoted as AUD(topic)). As a second measure of specificity,we compute weighted friend share (WFS (topic)), based onthe relative frequency the friend of interacts with a topic. Weuse tie strength (described below) as weights in WFS in or-

8We define “high interest” as exceeding a topic-specific thresholdthat is set to the upper quartile in the population. For example, aperson who interacts with more than 8% of political posts would beconsidered as having high interest in politics.

der to give strong ties greater influence on the final measureof specificity than weaker ties. For both measures of topicalspecificity we took the maximal specificity score among thepost’s topics.

Dyadic relationship properties: this set of features focuseson social structure and communications between the partici-pant and their friends. To represent social structure we createa set of indicator variables that designate whether the friend isa close family member (parent, child, sibling, spouse), mem-ber of a Facebook list that the participant maintains (closefriends, acquaintances), or has overlapping profile informa-tion (same workplace, college, high-school, home town, orcurrent city). In order to understand how expectations devi-ate for members of the same social structure, we also includea binary variable (designated by best) to indicate the strongesttie in that social circle.

Communication between people provides an additional di-mension to social structure. We calculate a rough approxi-mation of the tie strength between two people using the long-term frequency at which they communicate in any form (e.g.liking, commenting, tagging, direct messages) as recorded inour logs over 12 weeks. Gilbert and Karahalios showed thatfrequency of communication is one of strongest predictors oftie strength [25]. In order to evaluate the effect of recent com-munications, we include indicator variables for the most re-cent friend who gave a Like or a Comment to the participantor received one from her.

FindingsTable 4 and Figure 3 describe the separate logistic regressionmodels we fitted (on the same feature set) to Like and Com-ment expectations from individual friends. Both models con-verged and produced comparable estimates (P(Y |X)) to theempirical percentages of expectations reported in the survey(41.68% for Likes, 15.2% for Comments). Significant coeffi-cients appear in every category of features as we discuss next.

Many of the findings for post-level expectations also hold forexpectations from specific friends, but a more nuanced pic-ture emerges. For example, participants with relatively fewerfriends (in the lower quartile Q f rnds

1 with 20 − 146 friends)are more likely to expect a Comment from a friend whilethose with more friends (in the upper quartile Q f rnds

4 with632 − 5000 friends) are more likely to expect a Like. Thisshift in expectations is in line with previous findings show-ing a preference for composed communications from strongties and lightweight communications from weak ties [13]. Inaddition, we see that male participants expect more feedbackthan females, but as we will see next this effect is mitigatedby the friend’s gender.

Only some dyadic differences have a significant effect onfeedback expectations from specific friends. The differencein activity levels between the listed friend and the participant(∆L7) is not significant on its own unless the participant her-self is more active than average. Gender differences in gen-eral do not show a significant effect, but the interaction term(“Is male-female rel.”) shows that males have lower expecta-tions for Likes from their female friends. Gaps in the number

Likes Comments

Friend expectation ∼ Coef. SE Coef. SE

Controls:Intercept −0.82∗∗∗ .22 −2.95∗∗∗ .27Position top −0.02 .03 0.081∗ .036Position right 0.130 .078 0.125 .099Age 0.0071∗∗ .0024 0.0072∗ .0030Is male 0.27∗∗ .10 0.48∗∗∗ .13Q f rnds

1 −0.18 .11 0.40∗∗ .13Q f rnds

3 0.039 .093 0.03 .12Q f rnds

4 0.23∗ .10 0.06 .13Tenure −0.099∗∗∗ .020 −0.141∗∗∗ .025L7 −0.137∗∗ .044 −0.009 .061Connectedness −0.070 .066 −0.291∗∗∗ .080Importance 0.293∗∗∗ .073 0.43∗∗∗ .10Personal 0.222∗∗ .073 0.22∗ .10

Individual diff.:∆Age 0.0041 .0044 0.0048 .0057Age × (∆Age) 0.07 .15 0.17 .21Is diff gender 0.146 .086 0.01 .11Is male-female rel. −0.28∗ .14 −0.32 .18∆ Friends −0.18 .22 −0.25 .30Q f rnds

1 × (∆ Friends) 0.13 .30 −0.44 .40Q f rnds

3 × (∆ Friends) 0.24 .26 0.05 .36Q f rnds

4 × (∆ Friends) 0.22 .24 0.46 .32∆Tenure −0.023 .033 −0.133∗∗ .041Tenure ×(∆Tenure) −0.07 .14 0.33 .18∆ L7 0.032 .067 0.114 .091L7 × (∆L7) 0.63∗ .28 −0.19 .37

Topical interests:cos(u, f ) 0.228∗∗∗ .068 0.25∗ .10cos(p, f ) −0.073 .069 0.08 .13maxtopicAUD(topic) −0.071 .068 0.108 .079maxtopicWFS (topic) 0.152∗ .078 0.198∗∗ .071

Relationship:Relationship type (See Figure 3)Tie strength (TS) 2.91∗∗∗ .19 1.79∗∗∗ .24TS × Importance −0.65∗ .30 −0.35 .40TS × Personal 0.12 .30 −0.19 .39

P(Y |X) 41.27% 13.3%Log Likelihood -2840.2 -1910.5Akaike Inf. Crit. 5796.3 3937.1

N = 5, 256; ∗p<0.5; ∗∗p<0.01; ∗∗∗p<0.001Table 4. Coefficients of Bayesian logistic regressions for expecting a Likeand a Comment on a post from a particular friend.

of friends and age between participants and their friends arenot significant.

Shared topical interests and specificity of close ties are asso-ciated with higher feedback expectations. The fact that ourweighted measure (WFS) is statistically significant while thenon-weighted measure (AUD(topic)) is not suggests that top-ical specificity matters more for close ties than weaker ties.These elevated levels of expectations can be explained bysimilarity to close ties or by the fact that one is more likely toknow the topical interests of their close ties.

Other relationship properties are strongly correlated withfeedback expectations. Doubling the frequency of commu-nication with a friend increases the odds of a Like or a Com-ment expectation tremendously – by 5-17 times. The only ex-ception to this general trend is for Likes on important posts,which can be seen in the negative coefficient of the interac-tion term of tie strength and the post’s importance in the Likes

HighschoolCollege

AcquaintancesAcquaintancesbest

Current citySpouse

HometownWorkChild

SiblingCollegebest

Close friendsHighschoolbest

WorkbestClose friendsbestLast given Like

ParentHometownbest

Current citybestLast received Like

Last given CommentLast received Comment

0% 20% 40% 60% 80%

Expectation Comment Like

P(exp. | rel. type, ..., controls)

Figure 3. Probability of expecting a Like or a Comment from differentsocial ties (95% CIs). Xbest indicates the friend in social circle X that theparticipant most frequently communicates with.

model (TS × Importance). An important post is associatedwith slightly lower expectations from close ties, which showsthat content moderates expectation differently for Likes andComments, and from different social ties.

Figure 3 shows the impact of social structure on feedback ex-pectations. The figure shows how the probability of expecta-tions (x-axis) changes for different types of relationships (y-axis), when all other variables from Table 4 are held constantat their mean value. Comment expectations are presentedwith green points and Like expectations with blue triangles.For example, the second row of the figure shows that partic-ipants are about 80% likely to expect a Like from the lastperson whom they gave a Comment to.

The results in Figure 3 highlight four important aspects of so-cial structures: recency, geographical proximity, family ties,and close friendships. The appearance of recently commu-nicated friends at the top of the figure highlights the strongassociation between recency and feedback expectations, evenafter controlling for longer-term tie strength and a variety ofother measures. In fact, the level of expectations for recentlycommunicated friends is above and beyond close-family tiesand most frequently communicated friends in every other so-cial circle. We also see that the best friends from the currentlocation (city or hometown) are associated with higher expec-tations than many other social ties. Parents and siblings areexpected to comment more than friends with similar Likesexpectations, while spouses are expected to like more thanfriends with similar Comments expectations9. It is possible

9The small number of spouses in our sample, 26, increases its con-fidence interval, but the gap between Likes and Comments’ expec-tations is statistically significant with p < 0.001.

Friend exp. ∼ AUC P@R5 P@R50 P@R95

Baseline:grid position

59.4

60.962.4

62.3

70.979.5

50.7

53.556.2

44.9

47.149.3

Demographics &activity info.:

62.9

64.065.0

59.6

66.272.8

55.1

57.159.1

47.5

49.050.6

+ Tie strength74.0

75.877.6

88.4

92.396.3

71.4

73.776.0

49.6

51.353.1

+ Topical similarityand specificity

74.4

76.278.0

85.3

89.794.1

70.6

73.776.8

50.2

52.254.1

+ Social structure79.7

80.781.8

92.3

95.899.4

77.7

80.182.5

51.3

53.656.0

+ self-reports:Connectedness +Post importance + Personal

79.8

81.382.7

91.4

96.3100.0

78.0

81.184.2

51.9

54.857.7

Table 5. Predicting friend-level expectations for a Like or a Comment us-ing different feature sets obtained using gbm. P@R stands for precisionat different recall levels of 5, 50 or 95 percent. Numbers above/below ineach cell represent 95% CIs.

that lower expectations for Comments from spouses reveal apreferences for face-to-face feedback over online communi-cations in this case. Last, we note that the best friend indi-cator variables were found significant in every social groupexcept Acquaintances, for whom people have lower expec-tations in general. The significance of best friend variablesdemonstrates that the closest-ties in most social circles havemuch higher feedback expectations, beyond what is expectedbased on the frequency of communication with them or anyother factor in our models. Therefore, we conclude that re-cent communication, geographical proximity, family ties, andclose friendships increase feedback expectations consider-ably, even after controlling for individual differences, postedcontent, and other relationship properties.

Predicting friend-level expectationsModels that identify friend-level feedback expectations canhelp social network sites evaluate how often people’s expec-tations from their friends are met, identify possible reasonsfor unmet expectations, and ultimately guide the design ofplatforms to do better targeting and deliver more satisfyingexperiences to people. Therefore, our goal in this section isto assess how well a predictive model can identify friend-levelexpectations in practice.

We test different subsets of features from Table 4 in predictingthe feedback expectations from a particular friend, withoutdistinguishing between Likes and Comments for simplicity.Again, we use three different machine learning models forpredicting expectations (glmnet, gbm and SVM), this timefor feedback from a friend (Like or Comment) as reportedin the Friends Grid. Table 5 summarizes the results of 10-fold cross-validation of the best-performing model for eachfeature-set.

The baseline model obtains 60% AUC using informationabout the friend’s position in the grid alone. However, simple

demographic and activity information about people’s Face-book activity surpasses the baseline with 64% AUC. Then,including tie strength improves the predictive ability consid-erably, from 64% to 75.8% AUC. The topical features aloneachieve 66% AUC (not in the table), but when added tothe rest of the features improve the predictive accuracy onlymarginally. A second considerable increase in performanceis obtained using information about the relationship type –the AUC increases from 76% to closer to 81%, demonstratingthat social structure carries valuable information about expec-tations that is not captured by other variables. Last, subjectiveinformation about Connectedness and the post’s importanceand intimacy only adds little to the predictive ability of themodel. Most likely, the self-reported variables encode infor-mation already captured by other variables.

Our predictive model outperforms the baseline and identifiesfriend-level expectations with good accuracy (80%) using logdata alone (no self-reported measures). When the model is setto retrieve only half of the cases with expectations (recall of50%) it will correctly identify feedback expectations of heldout individuals for four out of five of their friends (80%). Atthis level of performance, social media platforms can begin toestimate how well people’s expectations are met and exploredesigns that improve people’s satisfaction from their onlineinteractions.

FULFILLMENT OF EXPECTATIONS & CONNECTEDNESSFinally, we analyze participants’ responses from the Laggedsurvey to understand the relationship between fulfillment ofexpectations and feeling of Connectedness (RQ4). First, weexamine the relationship between two of the measures fromthe Lagged survey: Fulfillment of feedback expectation andConnectedness. Then, we establish that feedback expecta-tions carry valuable information about connectedness that isnot captured by the raw amount of feedback received or othermeasures.

Figure 4 shows a positive correlation between Connected-ness and fulfillment of expectations, as reported by partici-pants in the Lagged survey. The measure of Connectedness(y-axis) is presented using numerical values (1=very discon-nected, 3=neither connected nor disconnected, and 5=veryconnected) and the measure of Fulfillment of expectations ispresented on the x-axis. For example, people who reported re-ceiving about the same amount of feedback as they expectedaveraged 3.86 on the 1-5 Connectedness scale.

The results in Figure 4 highlight an important relation be-tween fulfillment of feedback expectations and connected-ness. The more feedback received relative to expectationsthe more connected people feel to their Facebook friends:each unit increase on the fulfillment of feedback expectationsscale translates into an addition 0.26 of connectedness. Over-all, people move from 3 (neither connected nor disconnected)when their expectations are far from being met closer to 5(very connected) when their expectations are exceeded con-siderably.

We ran an additional control survey to rule out the possibil-ity that the response on the Connectedness question, which

3.5

4.0

4.5

1-Far fewer than expected

3-About whatI expected

5-Far morethan expected

How did the feedback received so far match your expectations?

How connected do you feel to your Facebook friends?

Conn ~ Estimate Std. err

Intercept 3.86*** 0.06 Exp. match 0.26*** 0.01

*** p < 2e-16, R^2 = 0.05

Figure 4. Responses from the 24h Lagged survey about fulfillment ofexpectations (x-axis) and feeling connected to Facebook friends (y-axis)with 95% CIs.

AUC

Connectedness ∼ Linear Regression SVM

Random answer 50.0% 50.0%

log(1 + WF) × log(#Friends) 55.0% 55.5%

WFglbl × log(#Friends) 55.0% 55.8%

WFinvd × log(#Friends) 54.4% 55.3%

Fulfillment of expectations 58.8% 58.8%

All of the above 61.0% 62.7%

Table 6. The predictive power of feedback and expectation using linearregression and SVM. Fulfillment of expectations is the single strongestpredictor for Connectedness with 58.8% AUC, only second to the modelthat uses feedback and expectations jointly.

appeared first, may affected the response on fulfillment ofexpectations question. In the control survey we omitted theConnectedness question, and found no significant differencein the distribution of responses to the fulfillment of expecta-tions question (χ2 = 5.56, df = 4, p-value = 0.23). Therefore,we conclude that there is no evidence of a priming effect be-tween the two measures.

Next, we establish that knowledge of feedback expectationsprovides valuable information that is not captured otherwise.In particular, we show that neither the feedback on the par-ticular post nor the feedback on previous posts can better ex-plain connectedness than knowledge about people’s expecta-tions. We do so by using both linear and non-linear regres-sion models to predict the Connectedness people reported inthe Lagged survey. Our goal is not to perfectly explain Con-nectedness, which is a complex social and psychological con-struct, but rather show that expectations carry important ad-ditional information that is not captured by past (or present)feedback.

We calculate three different measures of feedback and com-pare them to the single measure of Fulfillment of expectationsfrom the Lagged survey. First, for simplicity, we combine theLikes and Comments into a single measure of Weighted Feed-

back: WF = Likes + 5 × Comments that gives more weightto Comments since they are more rare10. WF is based on theLikes and Comments received on a single post in the 24 hoursafter posting, similar to timing of our Lagged survey. We thencompute the percentile of WF relative to the distribution of allposts in our log data (denoted as WFglbl) or the individual’sprevious posts (denoted as WF indv). Since our measure of ful-fillment of expectations implicitly incorporates knowledge ofthe friends network, we include the friend count of the indi-vidual both as a separate predictor and as an interaction termwith the WF measures.

Table 6 shows separately the predictive power of our mea-sures of received feedback, fulfillment of expectations, andthe combination of actual feedback and expectations. Forexample, the model that uses received WF on the post andfriend count to predict Connectedness achieves 55% area un-der the curve (AUC) using linear regression and 55.5% us-ing SVM. The measure of fulfillment of feedback expec-tations is the single strongest predictor for Connectedness(with 58.8% AUC), outperforming models using the actualfeedback (WF), feedback relative to the global distribution(WFglbl), and a personalized measure of feedback (WF indv).These results demonstrate that knowledge of expectations isvaluable for important concepts like Connectedness, and can-not be simply substituted by feedback information.

DISCUSSIONIn this study, we complemented survey responses with logdata to better understand people’s expectations for feedbackon Facebook. We have shown that when feedback expecta-tions are met people feel more connected to their friends onFacebook. We also presented a nuanced view of how thoseexpectations are shaped. We showed that whether a partici-pant expected a post to receive more feedback than usual de-pends on the importance, intimacy, and content of the post.This expectation also depended on the characteristics of theindividuals themselves: their age and gender, as well as howlong they had been active on Facebook. Furthermore, wedemonstrated how the expectation for feedback from a partic-ular friend varies depending on tie strength, recency of com-munication, geographical proximity, relationship type, andthe relative strength of relationship within the social groupit is embedded in.

In addressing our first two research questions we found sup-porting evidence that links some characteristics of individ-uals and posts to higher than usual feedback expectations.The subjective importance and intimacy of posts were the twostrongest content properties associated with higher than usualfeedback expectations. The result about intimacy of contentis in line with self-disclosure literature [1,48], which showedthat people seek more social validation when broadcasting tomany friends. Greater desire for social validation could alsolead to increased feedback expectations. The fact that boththe importance and intimacy were significant for post-levelexpectations highlights the need to better understand these

10other weighting of Comments did not significantly change the re-sults.

important concepts. Future work could investigate what prop-erties of the content (e.g. linguistic features, style, topics,etc.) makes a post subjectively more important for an individ-ual. Similarly, future research can investigate the mechanismsbehind some of the individual differences we found in feed-back expectations on a post. For example, higher feedbackexpectations of older adults can be due to building strongerties over time or because expectations are less calibrated.

As for expectations from specific friends (RQ3), we provideda nuanced view that integrates tie strength and social struc-ture. Recent communicators are associated with the highestfeedback expectations. Gilbert and Karahalios showed theimportance of recent communications in the prediction of tiestrength [25], but the considerable effect of recent communi-cations on feedback expectation was not shown before. Wealso found a more nuanced preference for Comments overLikes that depends not only on tie strength but also on socialstructure. For example, spouses and best workplace friendshave relatively low commenting expectations despite beingstrong ties, which perhaps reveals an expectation of face-to-face communications from these friends. These results addto the findings of Burke and Kraut about different commu-nication preferences for strong and weak ties [13]. Finally,the best friends in each social group have higher expecta-tions associated with them, even after controlling for their tiestrength, social structure, and all other properties included inour models. Taken together, these results provide a glimpseinto the complex and inter-connected nature of expectationsfrom different social ties.

Our findings also suggest that the fulfillment of expectationson a single post has a sizable effect on how connected peo-ple feel to their friends (RQ4), and that this effect is not fullycaptured by information about feedback alone. This resultshows that for one important outcome, people’s sense of con-nectedness to their friends, the general framework of expecta-tions does indeed help in understanding people’s attitude bet-ter than any other measure in our models. Moreover, this re-sult highlights one potential mechanism, fulfillment of expec-tations, through which social media use contributes to one’ssubjective well being (as found in other studies [15, 19]). Weemphasize, however, that the correlation we found betweenfulfillment of expectation and Connectedness does not war-rant a causal relation (despite our additional control survey).The experimental evidence in the literature (e.g. [42]) leadsus to believe that fulfillment of expectations does indeed af-fect Connectedness, but further work is needed to establish acausal link.

Feedback expectations are not only important, but also pre-dictable, which paves the way for studying how expectationscan be incorporated in social systems. Our models identifiedfeedback expectations for posts with good accuracy. The pre-diction of expectations for a representative sample of friends(rather than the stratified sample used in our study) is likelyto attain even higher accuracy due to the higher proportionof weak ties that are likely to have no feedback expectationsassociated with them. However, it remains an open questionwhether systems should adapt to posters’ expectations, and if

so, to what extent. How can the prediction of feedback expec-tations be used to improve the experience for both the personcreating the content and their friends? Is content associatedwith higher-than-usual feedback expectations more likely tobe interesting to a wider range of a person’s friends? Howdoes the fulfillment of expectations on one post affect expec-tations on subsequent posts? And if having one’s feedbackexpectations met correlates with a greater feeling of connect-edness, would understanding one’s audience [3] be helpful?We leave these and other questions for future work.

Limitations and future workDespite our attempts to capture feedback expectations as ac-curately as possible, our study design has several limitations.People may have different interpretations for expectations,which may vary from the bare minimum of feedback thatwould be “enough” to desires and hopes. Moreover, directlyasking people about expectations may elicit expectations thatdid not exist before taking the survey and may not alwaysbe well-calibrated. In addition, the length and complexityof surveys led us to rely on single-item measures, which aregenerally less reliable than multiple-item measures. As notedearlier, our findings are based on observational analysis thatdoes not allow us to infer causality.

There are several remaining gaps that would be fruitful av-enues for future research. First, our work identified many im-portant factors for feedback expectations that are worthy offurther investigation. For example, why does age contributeto feedback expectations? why do males have higher expec-tations from specific friends? does the impact of recent com-munications on expectations stem from memory mechanismsor reciprocity? Second, our surveys focused on a single postat a certain point in time. Therefore, it is not yet clear howexpectations evolve over time, and whether people’s expecta-tions and feedback received on prior posts influence expecta-tions on any subsequent post. In addition, our surveys merelyasked people about the existence of Like and Comment ex-pectations, which do not capture the full range of possible re-sponses11. For example, people may expect supportive Com-ments from some friends, sarcastic replies from others, andmore informative responses from acquaintances. Our workalso identified potential value in developing language mod-els that would better capture the subjective importance andintimacy of posts.

Our results may not accurately represent expectations in otherpopulations, forms of media, and platforms. While the gen-eral framework of expectations was shown to be relativelyuniversal [6], the concrete expectations people have may bespecific to a certain culture, language, or community. Evenwithin the population of people on Facebook, the stratifiedsample of friends we used may not accurately represent theentire friends network, especially in cases where people’sprofile information is incomplete. In addition, sharing otherforms of media, such as photos, videos or mixed-media con-tent may be associated with other factors that affect expec-tations. Finally, different social networks bring about differ-

11Our study was conducted before the introduction of Facebook Re-actions, which are an extension of Likes.

ent social dynamics and with it varied feedback expectations.For example, if people share more public content on Twit-ter then expectations for may perhaps shift towards Retweetsrather than Likes. While some variables in our models areFacebook-specific, we believe that the categories we devel-oped in this work and their relation to feedback expectationswill generalize to other social network sites. That being said,the current work only looked at one social media site, at onepoint in time, and surveyed a tiny fraction of the people onFacebook. We look forward to other works that would utilizeour survey design and conceptual framework to contrast ourfindings with those from other social media platforms.

ConclusionThis work demonstrated that feedback expectations vary con-siderably across people, posts, and interpersonal relation-ships. Higher than usual feedback expectations on a post arelinked to the characteristics of the post (importance, intimacy,and content), individual (age, gender, activity on Facebook),and past Comments. Neither the length of posts nor thesentiment of posts were found significantly correlated withfeedback expectations. People have higher expectations fromcloser ties in general, but these are moderated by recency ofinteractions, geographical proximity, relationship type, andclose friendships. Moreover, we found that the fulfillmentof expectations is associated with feeling more strongly con-nected to friends, thus potentially contributing to individuals’well being. Last, our predictive models can estimate people’sexpectations with good accuracy, which paves the way for fu-ture research into the benefits and limitations of integratingexpectations in social systems.

REFERENCES1. Natalya N. Bazarova and Yoon H. Choi. 2014.

Self-Disclosure in Social Media: Extending theFunctional Approach to Disclosure Motivations andCharacteristics on Social Network Sites. Journal ofCommunication 64, 4 (2014), 635–657.

2. Natalya N. Bazarova, Yoon H. Choi, VictoriaSchwanda Sosik, Dan Cosley, and Janis Whitlock. 2015.Social sharing of emotions on Facebook: Channeldifferences, satisfaction, and replies. In Proceedings ofCSCW 2015.

3. Michael S Bernstein, Eytan Bakshy, Moira Burke, andBrian Karrer. 2013. Quantifying the invisible audiencein social networks. In Proceedings of CHI 2013.

4. Jennifer L. Bevan, Pei-Chern Ang, and James B. Fearns.2014. Being unfriended on Facebook: An application ofexpectancy violation theory. Computers in HumanBehavior 33 (2014), 171–178.

5. Judee K. Burgoon. 1978. A communication model ofpersonal space violations: Explication and an initial test.Human Communication Research 4, 2 (1978), 129–142.

6. Judee K. Burgoon. 1993. Interpersonal expectations,expectancy violations, and emotional communication.Journal of Language and Social Psychology 12, 1-2(1993), 30–48.

7. Judee K. Burgoon and Stephen B. Jones. 1976. Toward atheory of personal space expectations and theirviolations. Human Communication Research 2, 2(1976), 131–146.

8. Judee K Burgoon and Beth A. Le Poire. 1993. Effects ofcommunication expectancies, actual communication,and expectancy disconfirmation on evaluations ofcommunicators and their communication behavior.Human communication research 20, 1 (1993), 67–96.

9. Judee K. Burgoon, Valerie Manusov, Paul Mineo, andJerold L. Hale. 1985. Effects of gaze on hiring,credibility, attraction and relational messageinterpretation. Journal of Nonverbal Behavior 9, 3(1985), 133–146.

10. Judee K. Burgoon, Joseph B. Walther, and E. JamesBaesler. 1992. Interpretations, evaluations, andconsequences of interpersonal touch. HumanCommunication Research 19, 2 (1992), 237–263.

11. Moira Burke and Mike Develin. 2016. Once More withFeeling: Supportive Responses to Social Sharing onFacebook. In Proceedings of CSCW 2016.

12. Moira Burke and Robert Kraut. 2013. Using Facebookafter losing a job: Differential benefits of strong andweak ties. In Proceedings of CSCW 2013.

13. Moira Burke and Robert E. Kraut. 2014. GrowingCloser on Facebook: Changes in Tie Strength ThroughSocial Network Site Use. In Proceedings of CHI 2014.

14. Moira Burke and Robert E. Kraut. 2016. TheRelationship between Facebook Use and Well-Beingdepends on Communication Type and Tie Strength.Journal of Computer-Mediated Communication 21, 4(2016), 265–281.

15. Moira Burke, Cameron Marlow, and Thomas Lento.2010. Social network activity and social well-being. InProceedings of CHI 2010.

16. Mina Choi and Catalina L. Toma. 2014. Social sharingthrough interpersonal media: Patterns and effects onemotional well-being. Computers in Human Behavior36 (2014), 530–541.

17. Utpal M. Dholakia, Richard P. Bagozzi, and Lisa K.Pearo. 2004. A social influence model of consumerparticipation in network-and small-group-based virtualcommunities. International journal of research inmarketing 21, 3 (2004), 241–263.

18. Nicole B. Ellison, Rebecca Gray, Cliff Lampe, andAndrew T. Fiore. 2014. Social capital and resourcerequests on Facebook. New Media & Society 16, 7(2014), 1104–1121.

19. Nicole B. Ellison, Charles Steinfield, and Cliff Lampe.2007. The benefits of Facebook “friends”: Social capitaland college students’ use of online social network sites.Journal of Computer-Mediated Communication 12, 4(2007), 1143–1168.

20. Nicole B. Ellison, Jessica Vitak, Rebecca Gray, andCliff Lampe. 2014. Cultivating social resources onsocial network sites: Facebook relationship maintenancebehaviors and their role in social capital processes.Journal of Computer-Mediated Communication 19, 4(2014), 855–870.

21. Eric M. Fife, C. Leigh Nelson, and Kristine Bayles.2009. When You Stalk Me, Please Don’t Tell Me AboutIt: Facebook and Expectancy Violation Theory.Kentucky Journal of Communication 28, 1 (2009).

22. Jerome Friedman, Trevor Hastie, and Robert Tibshirani.2010. Regularization Paths for Generalized LinearModels via Coordinate Descent. Journal of StatisticalSoftware 33, 1 (2010), 1–22.

23. Andrew Gelman. 2008. Scaling regression inputs bydividing by two standard deviations. Statistics inmedicine 27, 15 (2008), 2865–2873.

24. Andrew Gelman, Aleks Jakulin, Maria G. Pittau, andYu-Sung Su. 2008. A weakly informative default priordistribution for logistic and other regression models. TheAnnals of Applied Statistics (2008), 1360–1383.

25. Eric Gilbert and Karrie Karahalios. 2009. Predicting tiestrength with social media. In Proceedings of CHI 2009.

26. Rebecca Gray, Nicole B. Ellison, Jessica Vitak, andCliff Lampe. 2013. Who wants to know?:question-asking and answering practices amongfacebook users. In Proceedings of CSCW 2013.

27. Nir Grinberg, P Alex Dow, Lada A Adamic, and MorNaaman. 2016. Changes in Engagement Before andAfter Posting to Facebook. In Proceedings of CHI 2016.

28. Adam N. Joinson. 2008. Looking at, looking up orkeeping up with people?: motives and use of facebook.In Proceedings of CHI 2008.

29. Funda Kivran-Swaine, Jeremy Ting, Jed R. Brubaker,Rannie Teodoro, and Mor Naaman. 2014. UnderstandingLoneliness in Social Awareness Streams: Expressionsand Responses.. In Proceedings of ICWSM 2014.

30. Cliff Lampe, Rebecca Gray, Andrew T. Fiore, andNicole B. Ellison. 2014. Help is on the way: Patterns ofresponses to resource requests on Facebook. InProceedings of CSCW 2014.

31. Cliff Lampe, Rick Wash, Alcides Velasquez, and ElifOzkaya. 2010. Motivations to participate in onlinecommunities. In Proceedings of CHI 2010.

32. Eden Litt. 2012. Knock, knock. Who’s there? Theimagined audience. Journal of Broadcasting &Electronic Media 56, 3 (2012), 330–345.

33. Alice E. Marwick and danah boyd. 2011. I tweethonestly, I tweet passionately: Twitter users, contextcollapse, and the imagined audience. New media &society 13, 1 (2011), 114–133.

34. Jon D. Mcauliffe and David M. Blei. 2008. Supervisedtopic models. In Advances in neural informationprocessing systems.

35. Caitlin McLaughlin and Jessica Vitak. 2012. Normevolution and violation on Facebook. New Media &Society 14, 2 (2012), 299–315.

36. David Meyer, Evgenia Dimitriadou, Kurt Hornik,Andreas Weingessel, and Friedrich Leisch. 2015. e1071:Misc Functions of the Department of Statistics. Rpackage version 1.6.7.

37. Meredith R. Morris, Jaime Teevan, and KatrinaPanovich. 2010. What do people ask their socialnetworks, and why?: A survey study of status messageQ&A behavior. In Proceedings of CHI 2010.

38. Mor Naaman, Jeffrey Boase, and Chih-Hui Lai. 2010. Isit really about me?: message content in social awarenessstreams. In Proceedings of CSCW 2010.

39. Oded Nov. 2007. What motivates wikipedians?Commun. ACM 50, 11 (2007), 60–64.

40. James W. Pennebaker, Martha E. Francis, and Roger J.Booth. 2001. Linguistic inquiry and word count: LIWC2001. Word Journal Of The International LinguisticAssociation (2001).

41. Sheizaf Rafaeli, Tsahi Hayat, and Yaron Ariel. 2009.Knowledge building and motivations in Wikipedia:Participation as “Ba”. In Cyberculture and new media,Francisco J. Ricardo (Ed.). Rodopi Press,Amsterdam/New-York, 51–68.

42. Harry T. Reis, Kennon M. Sheldon, Shelly L. Gable,Joseph Roscoe, and Richard M. Ryan. 2000. Dailywell-being: The role of autonomy, competence, andrelatedness. Personality and social psychology bulletin26, 4 (2000), 419–435.

43. Richard M. Ryan and Edward L. Deci. 2000.Self-determination theory and the facilitation of intrinsicmotivation, social development, and well-being.American psychologist 55, 1 (2000), 68.

44. Lauren Scissors, Moira Burke, and Steven Wengrovitz.2016. What’s in a Like?: Attitudes and behaviors aroundreceiving Likes on Facebook. In Proceedings of CSCW2016.

45. C. J. Van Rijsbergen. 1979. Information Retrieval (2nded.). Butterworth-Heinemann.

46. Jessica Vitak. 2012. Keeping connected in the Facebookage: The relationship between Facebook use,relationship maintenance strategies, and relationaloutcomes. Ph.D. Dissertation. Michigan StateUniversity.

47. Jessica Vitak. 2014. Facebook makes the heart growfonder: relationship maintenance strategies amonggeographically dispersed and communication-restrictedconnections. In Proceedings of CSCW 2014.

48. Jessica Vitak and Jinyoung Kim. 2014. You can’t blockpeople offline: Examining how Facebook’s affordancesshape the disclosure process. In Proceedings of CSCW2014.

49. Joseph B. Walther and Judee K. Burgoon. 1992.Relational communication in computer-mediatedinteraction. Human communication research 19, 1(1992), 50–88.

50. Yi-Chia Wang, Moira Burke, and Robert E. Kraut.2016a. Modeling Self-Disclosure in Social NetworkingSites. In Proceedings of CSCW 2016.

51. Yi-Chia Wang, Hayley Hinsberger, and Robert E. Kraut.2016b. Does Saying This Make Me Look Good?: HowPosters and Outsiders Evaluate Facebook Updates. InProceedings of CHI 2016.

52. Jason Weston, Sumit Chopra, and Keith Adams. 2014.#TagSpace: Semantic Embeddings from Hashtags. InProceedings of EMNLP 2014.

53. David S White and Alison Le Cornu. 2011. Visitors andResidents: A new typology for online engagement. FirstMonday 16, 9 (2011).

54. Greg Ridgeway with contributions from others. 2015.gbm: Generalized Boosted Regression Models. Rpackage version 2.1.1.

55. Shaomei Wu, Atish Das Sarma, Alex Fabrikant, SilvioLattanzi, and Andrew Tomkins. 2013. Arrival anddeparture dynamics in social networks. In Proceedingsof WSDM 2013.