Recommender Systems, Consumer Preferences, and Anchoring Effects Gediminas Adomavicius Jesse Bockstedt Shawn Curley Jingjing Zhang University of Minnesota George Mason University University of Minnesota University of Minnesota Minneapolis, MN Fairfax, VA Minneapolis, MN Minneapolis, MN gedas@umn.edu jbockste@gmu.edu curley@umn.edu jingjing@umn.edu ABSTRACT recommender system’s use and value, as illustrated in Figure 1. Recommender systems are becoming a salient part of many e-commerce The figure also illustrates how consumer ratings are commonly websites. Much research has focused on advancing recommendation used to evaluate the recommender system’s performance in terms technologies to improve the accuracy of predictions, while behavioral of accuracy by comparing how closely the system-predicted aspects of using recommender systems are often overlooked. In this ratings match the later submitted actual ratings by the users. In study, we explore how consumer preferences at the time of consumption our studies, we focus on the feed-forward influence of the are impacted by predictions generated by recommender systems. We recommender system upon the consumer ratings that, in turn, conducted three controlled laboratory experiments to explore the effects of serve as inputs to these same systems. We believe that providing system recommendations on preferences. Studies 1 and 2 investigated user preferences for television programs, which were surveyed consumers with a prior rating generated by the recommender immediately following program viewing. Study 3 broadened to an system can introduce anchoring biases and significantly influence additional context—preferences for jokes. Results provide strong consumer preferences and, thus, their subsequent rating of an evidence viewers’ preferences are malleable and can be significantly item. As noted by [7], biases in the ratings provided by users can influenced by ratings provided by recommender systems. Additionally, lead to three potential problems: (i) biases can contaminate the the effects of pure number-based anchoring can be separated from the inputs of the recommender system, reducing its effectiveness; (ii) effects of the perceived reliability of a recommender system. Finally, the biases can artificially improve the resulting accuracy, providing a effect of anchoring is roughly continuous, operating over a range of distorted view of the system’s performance; (iii) biases might perturbations of the system. allow agents to manipulate the system so that it operates in their 1. INTRODUCTION favor. Recommender systems have become important decision aids in Predicted Ratings (expressing recommendations for unknown items) the electronic marketplace and an integral part of the business models of many firms. Such systems provide suggestions to consumers of products in which they may be interested and allow firms to leverage the power of collaborative filtering and feature- Recommender System (Consumer preference Accuracy Consumer based recommendations to better serve their customers and (Item consumption) estimation) increase sales. In practice, recommendations significantly impact the decision-making process of many online consumers; for example, it has been reported that a recommender system could account for 10-30% of an online retailer’s sales [25] and that Actual Ratings (expressing preferences for consumed items) roughly two-thirds of the movies rented on Netflix were ones that users may never have considered if they had not been Figure 1. Ratings as part of a feedback loop in consumer- recommended to users by the recommender system [10]. recommender interactions. Research in the area of recommender systems has focused almost For algorithm developers, the issue of biased ratings has been exclusively on the development and improvement of the largely ignored. A common underlying assumption in the vast algorithms that allow these systems to make accurate majority of recommender systems literature is that consumers recommendations and predictions. Less well-studied are the have preferences for products and services that are developed behavioral aspects of using recommender systems in the independently of the recommendation system. However, electronic marketplace. researchers in behavioral decision making, behavioral economics, and applied psychology have found that people’s preferences are Many recommender systems ask consumers to rate an item that often influenced by elements in the environment in which they have previously experienced or consumed. These ratings are preferences are constructed [5,6,18,20,30]. This suggests that the then used as inputs by recommender systems, which employ common assumption that consumers have true, non-malleable various computational techniques (based on methodologies from preferences for items is questionable, which raises the following statistics, data mining, or machine learning) to estimate consumer question: Whether and to what extent is the performance of preferences for other items (i.e., items that have not yet been recommender systems reflective of the process by which consumed by a particular individual). These estimated preferences are elicited? In this study, our main objective is to preferences are often presented to the consumers in the form of answer the above question and understand the influence of a “system ratings,” which indicate an expectation of how much the recommender system’s predicted ratings on consumers’ consumer will like the item based on the recommender system preferences. In particular, we explore four issues related to the algorithm and, essentially, serve as recommendations. The impact of recommender systems: (1) The anchoring issue— subsequent consumer ratings serve as additional inputs to the understanding any potential anchoring effect, particularly at the system, completing a feedback loop that is central to a point of consumption, is the principal goal of this study: Are Copyright is held by the author/owner(s). people’s preference ratings for items they just consumed drawn Decisions@RecSys’11, October 27, 2011, Chicago, IL, USA. toward predictions that are given to them? (2) The timing issue— 35 does it matter whether the system’s prediction is presented before that users showed high test-retest consistency when being asked to or after user’s consumption of the item? This issue relates to one re-rate a movie with no prediction provided. However, when possible explanation for an anchoring effect. Showing the users were asked to re-rate a movie while being shown a prediction prior to consumption could provide a prime that “predicted” rating that was altered upward/downward from their influences the user’s consumption experience and his/her original rating for the movie by a single fixed amount (1 rating subsequent rating of the consumed item. If this explanation is point), they tended to give higher/lower ratings, respectively. operative, an anchoring effect would be expected to be lessened when the recommendation is provided after consumption. (3) The Although [7] did involve recommender systems and preferences, system reliability issue—does it matter whether the system is our study differs from theirs in important ways. First, we address characterized as more or less reliable? Like the timing issue, this a fuller range of possible perturbations of the predicted ratings. issue is directed at illuminating the nature of the anchoring effect, This allows us to more fully explore the anchoring issue as to if obtained. If the system’s reliability impacts anchoring, then this whether any effect is obtained in a discrete fashion or more would provide evidence against the thesis that anchoring in continuously over the range of possible perturbations. More recommender systems is a purely numeric effect of users applying fundamentally, the focus of [7] was on the effects of anchors on a numbers to their experience. (4) The generalizability issue—does recall task, i.e., users had already “consumed” (or experienced) the anchoring effect extend beyond a single context? We the movies they were asked to re-rate in the study, had done so investigate two different contexts in the paper. Studies 1 and 2 prior to entering the study, and were asked to remember how well observe ratings of TV shows in a between-subjects design. Study they liked these movies from their past experiences. Thus, 3 addresses anchoring for ratings of jokes using a within-subjects- anchoring effects were moderated by potential recall-related design. Consistency of our findings supports a more general phenomenon, and preferences were being remembered instead of phenomenon that affects preference ratings immediately following constructed. In contrast, our work focuses on anchoring effects consumption, when recommendations are provided. that occur in the construction of preferences at the time of actual consumption. In our study, no recall is involved in the task 2. BACKGROUND AND HYPOTHESES impacted by anchors, participants consume the good for the first Behavioral research has indicated that judgments can be time in our controlled environment, and we measure the constructed upon request and, consequently, are often influenced immediate effects of anchoring. by elements of the environment in which this construction occurs. Still, [7] provide a useful model for the design of our studies, with One such influence arises from the use of an anchoring-and- two motivations in mind. First, their design provides an excellent adjustment heuristic [6,30], the focus of the current study. Using methodology for exploring the effects of recommender systems on this heuristic, the decision maker begins with an initial value and preferences. Second, we build upon their findings to determine if adjusts it as needed to arrive at the final judgment. A systematic anchoring effects of recommender systems extend beyond recall- bias has been observed with this process in that decision makers related tasks and impact actual preference construction at the time tend to arrive at a judgment that is skewed toward the initial of consumption. Grounded in the explanations for anchoring, as anchor. Prior research on anchoring effects spans three decades discussed above, our research goes beyond their findings to see if and represents a very important aspect of decision making, recommender system anchoring effects are strong enough to behavioral economics, and marketing literatures. Epley and manipulate a consumer’s perceptions of a consumption experience Gilovich [9] identified three waves of research on anchoring: (1) as it is happening. establishes anchoring and adjustment as leading to biases in judgment [5,9,21,29,30], (2) develops psychological explanations Since anchoring has been observed in other settings, though different than the current preference setting, we begin with the for anchoring effects [5,9,13,21,23], and (3) unbinds anchoring conjecture that the rating provided by a recommender system from its typical experimental setting and “considers anchoring in serves as an anchor. Insufficient adjustment away from the all of its everyday variety and examines its various moderators in anchor is expected to lead to a subsequent consumer preference these diverse contexts” ([9], p.21) [14,17]. Our study is primarily rating that is shifted toward the system’s predicted rating. This is located within the latter wave while informing the second wave— captured in the following primary hypothesis of the studies: testing explanations—as well; specifically, our paper provides a contribution both (a) to the study of anchoring in a preference Anchoring Hypothesis: Users receiving a recommendation situation at the time of consumption and (b) to the context of biased to be higher will provide higher ratings than users recommender systems. receiving a recommendation biased to be lower. Regarding the former of these contextual features, the effect of One mechanism that may underlie an anchoring effect with anchoring on preference construction is an important open issue. recommendations is that of priming, whereby the anchor can serve Past studies have largely been performed using tasks for which a as a prime or prompt that activates information similar to the verifiable outcome is being judged, leading to a bias measured anchor, particularly when uncertainty is present [6]. If this against an objective performance standard (also see review by [6]. dynamic operates in the current setting, then receiving the In the recommendation setting, the judgment is a subjective recommendation prior to consumption, when uncertainty is higher preference and is not verifiable against an objective standard. The and priming can more easily operate, should lead to greater application of previous studies to the preference context is not a anchoring effects than receiving the recommendation after straightforward generalization. consumption. Manipulating the timing of the recommendation provides evidence for tying any effects to priming as an Regarding our studies’ second contextual feature, very little underlying mechanism. research has explored how the cues provided by recommender systems influence online consumer behavior. The work that Timing Hypothesis: Users receiving a recommendation prior comes closest to ours is [7], which explored the effects of system- to consumption will provide ratings that are closer to the generated recommendations on user re-ratings of movies. It found recommendation (i.e., will be more affected by the anchor) than users receiving a recommendation after viewing. 36 Another explanation proposed for the anchoring effect is a Hypothesis) using artificial anchors; (2) to perform the content-based explanation, in which the user perceives the anchor exploratory analyses of whether participants behave differently as providing evidence as to a correct answer in situations where with high vs. low anchors; (3) to test the Timing Hypothesis for an objective standard exists. When applied to the use of anchoring effects with system recommendations (i.e., concerning recommender systems and preferences, the explanation might differential effects of receiving the recommendation either before surface as an issue of the consumer’s trust in the system. Prior or after consuming the item to be subsequently rated) ; (4) to test study found that increasing cognitive trust and emotional trust the Perceived System Reliability Hypothesis for anchoring effects improved consumer’s intentions to accept the recommendations with system recommendations (i.e., concerning the relationship [15]. Research also has highlighted the potential role of human- between the perceived reliability of the recommender system and computer interaction and system interface design in achieving anchoring effects of its recommendations); and (5) to build a high consumer trust and acceptance of recommendations database of user preferences for television shows, which would be [7,19,22,28]. However, the focus of these studies differs from used in computing personalized recommendations for Study 2. that underlying our research questions. In particular, the aforementioned prior studies focused on interface design 3.1. Methods (including presentation of items, explanation facilities, and rating 216 people completed the study. Ten respondents indicated scale definitions) rather than the anchoring effect of having seen some portion of the show that was used in the study recommendations on the construction of consumer preferences. (all subjects saw the same TV show episode in Study 1). Our work was motivated in part by these studies to specifically Excluding these, to obtain a more homogeneous sample of highlight the role of anchoring on users’ preference ratings. subjects all seeing the show for the first time, left 206 subjects for analysis. Participants were solicited from a paid subject pool and In their initial studies, Tversky and Kahneman [30] used anchors paid a fixed fee at the end of the study. that were, explicitly to the subjects, determined by spinning a wheel of fortune. They still observed an effect of the magnitude In Study 1 subjects received artificial anchors, i.e., system ratings of the value from this random spin upon the judgments made (for were not produced by a recommender system. All subjects were various almanac-type quantities, e.g., the number of African shown the same TV show episode during the study and were countries in the United Nations). [27] also demonstrated asked to provide their rating of the show after viewing. anchoring effects even with extreme values (e.g., anchors of 1215 Participants were randomly assigned to one of seven experimental or 1992 in estimating the year that Einstein first visited the United groups. Before providing their rating, those in the treatment States). These studies suggest that the anchoring effect may be groups received an artificial system rating for the TV show used purely a numerical priming phenomenon, and that the quality of in this study. Three factors were manipulated in the rating the anchor may be less important. provision. First, the system rating was set to have either a low (1.5, on a scale of 1 through 5) or high value (4.5). Since [29] In contrast, [20] found that the anchoring effect was mediated by found an asymmetry of the anchoring effect such that high the plausibility of the anchor. The research cited earlier anchors produced a larger effect than did low anchors in their connecting cognitive trust in recommendation agents to users’ study of job performance ratings, we used anchors at both ends of intentions to adopt them [15] also suggests a connection between the scale. reliability and use. To the extent that the phenomenon is purely numerically driven, weakening of the recommendation should The second factor in Study 1 was the timing of the have little or no effect. To the extent that issues of trust and recommendation. The artificial system rating was given either quality are of concern, a weakening of the anchoring should be before or after the show was watched (but always before the observed with a weakening of the perceived quality of the viewer was asked to rate the show). This factor provides a test of recommending system. the Timing Hypothesis. Together, the first two factors form a 2 x 2 (High/Low anchor x Before/After viewing) between-subjects Perceived System Reliability Hypothesis: Users receiving a design (the top four cells of the design in Table 1). recommendation from a system that is perceived as more reliable will provide ratings closer to the recommendation Intersecting with this design is the use of a third factor: the (i.e., will be more affected by the anchor) than users perceived reliability of the system (strong or weak) making the receiving a recommendation from a less reliable system. recommendation. In the Strong conditions for this factor, subjects were told (wording is for the Before viewing/Low anchor To explore our hypotheses, we conducted three controlled condition): “Our recommender system thinks that you would rate laboratory experiments, in which system predictions presented to the show you are about to see as 1.5 out of 5.” Participants in the participants are biased upward and downward so our hypotheses corresponding Weak conditions for the perceived reliability factor can be tested in realistic settings. The first study explores our saw: “We are testing a recommender system that is in its early hypotheses by presenting participants with randomly assigned stages of development. Tentatively, this system thinks that you artificial system recommendations. The second study extends the would rate the show you are about to see as 1.5 out of 5.” This first and uses a live, real-time recommender system to produce factor provides a test of the Perceived System Reliability predicted recommendations for our participants, which are then Hypothesis. At issue is whether any effect of anchoring upon a biased upward or downward. The final study generalizes to recommendation is merely a numerical phenomenon or is tied to preferences among jokes, studied using a within-subjects design the perceived reliability and quality of the recommendation. and varying levels of rating bias. The next three sections provide details about our experiments and findings. Since there was no basis for hypothesizing an interaction between timing of the recommendation and strength of the system, the 3. STUDY 1: IMPACT OF ARTIFICIAL complete factorial design of the three factors was not employed. RECOMMENDATIONS For parsimony of design, the third factor was manipulated only The goals of Study 1 were fivefold: (1) to perform a test of the within the Before conditions, for which the system primary conjecture of anchoring effects (i.e., Anchoring recommendation preceded the viewing of the TV show. Thus, 37 within the Before conditions of the Timing factor, the factors of collected by survey, including both demographic data (e.g., Anchoring (High/Low) and Reliability of the anchor gender, age, occupation) and questionnaire responses (e.g., hours (Strong/Weak) form a 2x2 between-subjects design (the bottom watching TV per week, general attitude towards recommender four cells of the design in Table 1). systems), as covariates and random factors. However, none of these variables or their interaction terms turned out to be In addition to the six treatment groups, a control condition, in significant, and hence we focus on the three fixed factors. which no system recommendation was provided, was also included. The resulting seven experimental groups, and the We begin with analysis of the 2x2 between-subjects design sample sizes for each group, are shown in Table 1. involving the factors of direction of anchor (High/Low) and its timing (Before/After viewing). As is apparent from Table 2 (rows Subjects participated in the study using a web-based interface in a marked as Design 1), and applying a general linear model, there is behavioral lab, which provided privacy for individuals no effect of Timing (F(1,113) = 0.021, p = .885). The interaction participating together. Following a welcome screen, subjects of Timing and High/Low anchor was also not significant (F(1, were shown a list of 105 popular, recent TV shows. TV shows 113) = 0.228, p = .634). There is a significant observed anchoring were listed alphabetically within five genre categories: Comedy, effect of the provided artificial recommendation (F(1, 113) = Drama, Mystery/Suspense, Reality, and Sci Fi/Fantasy. For each 14.30, p = .0003). The difference between the High and Low show they indicated if they had ever seen the show (multiple conditions was in the expected direction, showing a substantial episodes, one episode, just a part of an episode, or never), and effect between groups (one-tailed t(58) = 2.788, p = .0035, then rated their familiarity with the show on a 7-point Likert scale assuming equal variances). Using Cohen’s (1988) d, which is an ranging from “Not at all familiar” to “Very familiar.” Based on effect size measure used to indicate the standardized difference these responses, the next screen first listed all those shows that the between two means (as computed by dividing the difference subject indicated having seen and, below that, shows they had not between two means by a standard deviation for the data), the seen but for which there was some familiarity (rating of 2 or effect size is 0.71, in the medium-to-large range. above). Subjects rated each of these shows using a 5-star scale that used verbal labels parallel to those in use by Netflix.com. Table 2. Mean (SD) Ratings of the Viewed TV Show by Half-star ratings were also allowed, so that subjects had a 9-point Experimental Condition in Study 1. scale for expressing preference. In addition, for each show, an Design Design Group (timing-anchor- N Mean (SD) option of “Not able to rate” was provided. Note that these ratings 1 2 reliability) were not used to produce the artificial system recommendations in * * Before-High-Strong 31 3.48 (1.04) Study 1; instead, they were collected to create a database for the * After-High-Strong 28 3.43b (0.81) recommender system used in Study 2 (to be described later). Control 29 3.22 (0.98) * Before-High-Weak 31 3.08 (1.07) Table 1 Experimental Design and Sample Sizes in Study 1. * Before-Low-Weak 29 2.83 (0.75) Control: 29 * After-Low-Strong 29 2.88 (0.79) Reliability condition Timing condition Low High * * Before-Low-Strong 29 2.78 (0.92) (anchor) (anchor) Strong (reliability) After (timing) 29 28 Using only data within the Before conditions, we continue by Strong (reliability) Before (timing) 29 31 analyzing the second 2 x 2 between-subjects design in the study Weak (reliability) Before (timing) 29 31 (Table 2, rows marked as Design 2), involving the factors of direction of anchor (High/Low) and perceived system reliability Following the rating task, subjects watched a TV episode. All (Strong/Weak). The anticipated effect of weakening the subjects saw the same episode of a situation comedy. A less well- recommender system is opposite for the two recommendation known TV show was chosen to maximize the likelihood that the directions. A High-Weak recommendation is expected to be less majority of subjects were not familiar with it. The episode was pulled in the positive direction compared to a High-Strong streamed from Hulu.com and was 23 minutes 36 seconds in recommendation; and, a Low-Weak recommendation is expected duration. The display screen containing the episode player had a to be less pulled in the negative direction as compared to Low- visible time counter moving down from 20 minutes, forcing the Strong. So, we explore these conjectures by turning to the direct respondents to watch the video for at least this time before the tests of the contrasts of interest. There is no significant difference button to proceed to the next screen was enabled. between the High and Low conditions with Weak Either immediately preceding (in the Before conditions) or recommendations (t(58) = 1.053, p = .15), unlike with Strong immediately following (in the After conditions) the viewing recommendations (as noted above, p = .0035). Also, the overall display, subjects saw a screen providing the system effect was reduced for the Weak setting, compared to the Strong recommendation with the wording appropriate to their condition recommendation setting, and was measured as a Cohen’s d = 0.16, (Strong/Weak, Low/High anchor). This screen was omitted in the less than even the small effect size range. Thus, the subjects were Control condition. Following, subjects rated the episode just sensitive to the perceived reliability of the recommender system. viewed. The same 5-star (9-point) rating scale used earlier was Weak recommendations did not operate as a significant anchor provided for the preference rating, except that the “Not able to when the perceived reliability of the system was lowered. rate” option was omitted. Finally, subjects completed a short Finally, we check for asymmetry of the anchoring effect using the survey that included questions on demographic information and control group in comparison to the Before-High and Before-Low TV viewing patterns. groups. (Similar results were obtained using the After-High and After-Low conditions as comparison, or using the combined High 3.2. Results and Low groups.) In other words, we already showed that the All statistical analyses were performed using SPSS 17.0. Table 2 High and Low groups were significantly different from each shows the mean ratings for the viewed episode for the seven other, but we also want to determine if each group differs from the experimental groups. Our preliminary analyses included data Control (i.e., when no recommendation was provided to the users) 38 in the same manner. When an artificial High recommendation system’s predicted rating). High and Low conditions were was provided (4.5), ratings were greater than those of the Control included to learn more about the asymmetry effect observed in group, but not significantly so (t(58) = 0.997, p = .162). But when Study 1. In addition to the three treatment groups, a control group an artificial Low recommendation was provided (1.5), ratings was included for which no system recommendation was provided. were significantly lower than those of the Control group (t(56) = The numbers of participants in the four conditions of the study are 1.796, p = .039). There was an asymmetry of the effect; however, shown in Table 4 (Section 4.2). the direction was opposite to that found by Thorsteinson et al. (2008). To study the effect further, Study 2 was designed to Based on the TV show rating data collected in Study 1, an online provide further evidence. So, we will return to the discussion of system was built for making TV show recommendations in real the effect later in the paper. time. We compared seven popular recommendation techniques to find the best-performing technique for our dataset. The In summary, analyses indicate a moderate-to-strong effect, techniques included simple user- and item-based rating average supporting the Anchoring Hypothesis. When the recommender methods, user- and item-based collaborative filtering approaches system was presented as less reliable, being described as in test and their extensions [2,4,24], as well as a model-based matrix phase and providing only tentative recommendations, the effect factorization algorithm [11,16] popularized by the recent Netflix size was reduced to a minimal or no effect, in support of the prize competition [3]. Each technique was evaluated using 10- Perceived System Reliability Hypothesis. Finally, the Timing fold cross validation based on the standard mean absolute error Hypothesis was not supported – the magnitude of the anchoring (MAE) and coverage metrics. Although the performances are effect was not different whether the system recommendation was comparable, the item-based CF performed slightly better than received before or after the viewing experience. This suggests other techniques (measured in predictive accuracy and coverage). that the effect is not attributable to a priming of one’s attitude Also because the similarities between items could be pre- prior to viewing. Instead, anchoring is likely to be operating at computed, the item-based technique performed much faster than the time the subject is formulating a response. other techniques. Therefore the standard item-based collaborative filtering approach was selected for our recommender system. Overall, viewers, without a system recommendation, liked the episode (mean = 3.22, where 3 = “Like it”), as is generally found During the experiments, the system took as input subject’s ratings with product ratings. However, asymmetry of the anchoring of shows that had been seen before or for which the participant effect was observed at the low end: Providing an artificial low had indicated familiarity. In real time, the system predicted recommendation reduced this preference more so than providing a ratings for all unseen shows and recommended one of the unseen high recommendation increased the preference. This effect is shows for viewing. To avoid possible show effects (e.g., to avoid explored further in Study 2. selecting shows that receive universally bad or good predictions) as well as to assure that the manipulated ratings (1.5 points 4. STUDY 2: IMPACT OF ACTUAL above/below the predicted rating) could still fit into the 5-point RECOMMENDATIONS rating scale, only shows with predicted rating scores between 2.5 Study 2 follows up Study 1 by replacing the artificially fixed and 3.5 were recommended. When making recommendations, the anchors with actual personalized recommendations provided by a system examined each genre in alphabetical order (i.e., comedy well-known and commonly used recommendation algorithm. first, followed by drama, mystery, reality, and sci-fi) and went Using the user preferences for TV shows collected in Study 1, a through all unseen shows within each genre alphabetically until recommender system was designed to estimate preferences of one show with a predicted rating between 2.5 and 3.5 was found. subjects in Study 2 for unrated shows. Because participants This show was then recommended to the subject. When no show provide input ratings before being shown any recommendations or was eligible for recommendation, subjects were automatically re- other potential anchors, the ratings were unbiased inputs for our assigned to one of the treatment groups in Study 1. own recommendation system. Using a parallel design to Study 1, Our TV show recommender system made suggestions from a list we examine the Anchoring Hypothesis with a recommender of the 105 most popular TV shows that have aired in the recent system comparable to the ones employed in practice online. decade according to a ranking posted on TV.com. Among the 105 shows, 31 were available for online streaming on Hulu.com at the 4.1. Methods time of the study and were used as the pool of shows 197 people completed the study. They were solicited from the recommended to subjects for viewing. Since our respondents same paid subject pool as used for Study 1 with no overlap rated shows, but viewed only a single episode of a show, we between the subjects in the two studies. Participants received a needed a procedure to select the specific episode of a show for fixed fee upon completion of the study. viewing. For each available show, we manually compared all In Study 2, the anchors received by subjects were based on the available episodes and selected the episode that received a median recommendations of a true recommender system (discussed aggregated rating by Hulu.com users to include in the study. This below). Each subject watched a show that he/she had indicated procedure maximized the representativeness of the episode for not having seen before – that was recommended by an actual real- each show, avoiding the selection of outlying best or worst time system based on the subject’s individual ratings. Since there episodes that might bias the participant’s rating. Table 3 shows was no significant difference observed between subjects receiving the distributions of rated and viewing-available shows by genre. system recommendations before or after viewing a show in Study The procedure was largely identical to the Before and Control 1, all subjects in the treatment groups for Study 2 saw the system- conditions used for Study 1. However, in Study 2, as indicated provided rating before viewing. earlier, subjects did not all view the same show. TV episodes Three levels were used for the recommender system’s rating were again streamed from Hulu.com. The episode watched was provided to subjects in Study 2: Low (i.e., adjusted to be 1.5 either approximately 22 or 45 minutes in duration. For all points below the system’s predicted rating), Accurate (the subjects, the viewing timer was set at 20 minutes, as in Study 1. system’s actual predicted rating), and High (1.5 points above the Subjects were instructed that they would not be able to proceed 39 until the timer reached zero; at which time they could choose to the overall analysis of Study 2 at the High and Low ends. stop and proceed to the next part of the study or to watch the remainder of the episode before proceeding. To pursue the results further, we recognize that one source of variation in Study 2 as compared to Study 1 is that different shows Table 3. Distribution of Shows. were observed by the subjects. As it turns out, 102 of the 198 subjects in Study 2 (52%) ended up watching the same Comedy Genre Number of Shows Available for Viewing show. As a result, we are able to perform post-hoc analyses, Comedy 22 7 Drama 26 8 paralleling the main analyses, limited to this subset of viewers. Mystery/Suspense 25 4 The mean (standard deviation) values across the four conditions Reality 15 4 of these subjects for the main response variable are shown in Sci Fi and Fantasy 17 8 Table 5. Using the same response variable of rating drift, the Total 105 31 overall effect across the experimental conditions was marginally maintained (F(2, 77) = 2.70, p = .07. Providing an accurate recommendation still did not significantly affect preferences for 4.2. Results the show, as compared to the Control condition (two-tailed t(47) = Since the subjects did not all see the same show, the preference 0.671, p = .506). Consistent with Study 1 and the overall ratings for the viewed show were adjusted for the predicted analyses, the High recommendation condition led to inflated ratings of the system, in order to obtain a response variable on a ratings compared to the Low condition (one-tailed t(51) = 2.213, p comparable scale across subjects. Thus, the main response = .016). The effect size was also comparable to the overall effect variable is the rating drift, which we define as: magnitude with Cohen’s d = 0.61, a medium effect size. Rating Drift = Actual Rating – Predicted Rating. However, for the limited sample of subjects who watched the same episode, the effects at the High and Low end were not Predicted Rating represents the rating of the TV show watched by symmetric. Compared to receiving an Accurate recommendation, the user during the study as predicted by the recommendation there was a significant effect of the recommendation being raised algorithm (before any perturbations to the rating are applied), and (t(52) = 1.847, p = .035, Cohen’s d = .50), but not of being Actual Rating is the user’s rating value for this TV show after lowered (t(51) = 0.286, p = .388). watching the episode. Therefore, positive/negative Rating Drift values represent situations where the user’s submitted rating was Table 5. Mean(SD) Rating Drift for Subjects Who Watched higher/lower than the system’s rating, as possibly affected by the Same Comedy Show in Study 2. positive/ negative perturbations (i.e., high/low anchors). Group N Mean (SD) Similarly to Study 1, our preliminary analyses using general linear High 27 0.81 (0.82) models indicated that none of the variables collected in the survey Control 22 0.53 (0.76) (such as demographics, etc.) demonstrated significance in Accurate 27 0.37 (0.93) explaining the response variable. The mean (standard deviation) Low 26 0.30 (0.86) values across the four conditions of the study for this variable are Thus, the indicated asymmetry of the anchoring effect is different shown in Table 4. Using a one-way ANOVA, overall the three from the asymmetry present in Study 1, being at the High end experimental groups (i.e., High, Low, and Accurate) significantly rather than the Low end. Also, the asymmetry is not robust across differed (F(2, 147) = 3.43, p = .035). the overall data. Indicated is that the underlying cause of Table 4. Mean (SD) Rating Drift of the Viewed TV Show by asymmetries is situational, in this case depending upon specific Experimental Condition, Study 2. TV show effects. When looking at effects across different TV shows (Table 4), the show effects average out and symmetry is Study 2 observed overall. When looking at effects for a particular show Group N Mean (SD) (Tables 2 and 5), idiosyncratic asymmetries can arise. High 51 0.40 (1.00) Control 48 0.14 (0.94) 5. STUDY 3: ACTUAL Accurate 51 0.13 (0.96) Low 47 -0.12 (0.94) RECOMMENDATIONS WITH JOKES Study 3 provides a generalization of Study 2 within a different Providing an accurate recommendation did not significantly affect content domain, applying a recommender system to joke preferences for the show, as compared to the Control condition preferences rather than TV show preferences. As in Study 2, the (two-tailed t(97) = 0.023, p = .982). Consistent with Study 1, the procedure uses actual recommendations provided by a commonly High recommendation condition led to inflated ratings compared used recommendation algorithm. A within-subjects design also to the Low condition (one-tailed t(96) = 2.629, p = .005). The allows us to investigate behavior at an individual level of analysis, effect size was of slightly less magnitude with Cohen’s d = 0.53, a rather than in the aggregate. We apply a wider variety of medium effect size. However, unlike in Study 1, the anchoring perturbations to the actual recommendations for each subject, effect in Study 2 is symmetric at the High and Low end. There ranging from -1.5 to 1.5, the values used in Study 2, rather than was a marginally significant effect of the recommendation being just using a single perturbation per subject. lowered compared to being accurate (t(96) = 1.305, p = .098, Cohen’s d = .30), and a marginally significant effect at the High 5.1. Methods end compared to receiving Accurate recommendations (t(100) = 61 people received a fixed fee for completing the study. They 1.366, p = .088, Cohen’s d = .23). Similar effects are observed were solicited from the same paid subject pool used for Studies 1 when comparing High/Low to Control condition. In summary, and 2 with no overlap across the three studies. the Anchoring Hypothesis is supported in Study 2, consistently with Study 1. However, the anchoring effects were symmetric in As with Study 2, the anchors received by subjects were based on the recommendations of a true recommender system. The item- 40 based collaborative filtering technique was used to maintain analyses. The mean magnitude of the relationship is 0.37, with consistency with Study 2. The same list of 100 jokes was used values ranging from -0.27 to 0.87. during the study, though the order of the jokes was randomized between subjects. The jokes and the rating data for training the Overall, the analyses strongly suggest that the effect of recommendation algorithm were taken from the Jester Online perturbations on rating drift is not discrete. Perturbations have a Joke Recommender System repository [12]. Specifically, we used continuous effect upon ratings with, on average, a drift of 0.35 their Dataset 2, which contains 150 jokes. To get to our list of rating points occurring for every rating point of perturbation (e.g., 100, we removed those jokes that were suggested for removal at mean rating drift is 0.53 for a perturbation of +1.5). the Jester website (because they were either included in the “gauge set” in the original Jester joke recommender system or because they were never displayed or rated), jokes that more than 0.53 one of the coauthors of our study identified as having overly objectionable content, and finally those jokes that were greatest in 0.28 MeanRatingDrift length (based on word count). Control Ͳ0.04 0.07 The procedure paralleled that used for Study 2 with changes adapted to the new context. Subjects first evaluated 50 jokes, Ͳ1.5 Ͳ1 Ͳ0.5 0.5 1 1.5 randomly selected and ordered from the list of 100, as a basis for Ͳ0.20 providing recommendations. The same 5-star rating scale with Ͳ0.23 half-star ratings from Studies 1 and 2 was used, affording a 9- Ͳ0.41 point scale for responses. Next, the subjects received 40 jokes Ͳ0.53 with a predicted rating displayed. Thirty of these predicted ratings were perturbed, 5 each using perturbations of -1.5, -1.0, - PerturbationofRecommendation 0.5, +0.5, +1.0, and +1.5. The 30 jokes that were perturbed were determined pseudo-randomly to assure that the manipulated Figure 2. Mean Rating Drift as a Function of the Amount of ratings would fit into the 5-point rating scale. First, 10 jokes with Rating Perturbation and for Control Condition in Study 3. predicted rating scores between 2.5 and 3.5 were selected randomly to receive perturbations of -1.5 and +1.5. From the Table 6. Mean (SD) Rating Drift, in the Comparable remaining, 10 jokes with predicted rating scores between 2.0 and Conditions Used in Study 2 (±1.5, 0, Control), for Study 3. 4.0 were selected randomly to receive perturbations of -1.0 and Group N Mean (SD) +1.0. Then, 10 jokes with predicted rating scores between 1.5 and High 305 0.53 (0.94) 4.5 were selected randomly to receive perturbations of -0.5 and Control 320 -0.04 (1.07) +0.5. Ten predicted ratings were not perturbed, and were Accurate 610 -0.20 (0.97) displayed exactly as predicted. These 40 jokes were randomly Low 305 -0.53 (0.95) intermixed. Following the first experimental session (3 sessions were used in total), the final 10 jokes were added as a control. A 6. DISCUSSION AND CONCLUSIONS display was added on which subjects provided preference ratings We conducted three laboratory experiments and systematically for the 10 jokes with no predicted rating provided, again in examined the impact of recommendations on consumer random order. Finally in all sessions, subjects completed a short preferences. The research integrates ideas from behavioral demographic survey. decision theory and recommender systems, both from practical and theoretical standpoints. The results provide strong evidence 5.2. Results that biased output from recommender systems can significantly As with Study 2, the main response variable for Study 3 was influence the preference ratings of consumers. Rating Drift (i.e., Actual Rating – Predicted Rating). As an illustration of the overall picture, Figure 2 shows the mean Rating From a practical perspective, the findings have several important Drift, aggregated across items and subjects, for each perturbation implications. First, they suggest that standard performance used in the study. In the aggregate, there is a linear relationship metrics for recommender systems may need to be rethought to both for negative and positive perturbations. For comparison account for these phenomena. If recommendations can influence purposes, Table 6 shows the mean (standard deviation) values consumer-reported ratings, then how should recommender across the four perturbation conditions of Study 3 that were systems be objectively evaluated? Second, how does this comparable to those used in Study 2 (aggregating across all influence impact the inputs to recommender systems? If two relevant Study 3 responses). The general pattern for Study 3— consumers provide the same rating, but based on different initial using jokes and within-subjects design—parallels that for Study recommendations, do their preferences really match in identifying 2—using TV shows and a between-subjects design. future recommendations? Consideration of issues like these arises as a needed area of study. Third, our findings bring to light the The within-subjects design also allows for analyses of the potential impact of recommender systems on strategic practices. Anchoring Hypothesis at the individual level. We began by If consumer choices are significantly influenced by testing the slopes across subjects between negative and positive recommendations, regardless of accuracy, then the potential arises perturbations, and no significant difference was observed (t(60) = for unscrupulous business practices. For example, it is well- 1.39, two-tailed p = .17). We also checked for curvilinearity for known that Netflix uses its recommender system as a means of each individual subject for both positive and negative inventory management, filtering recommendations based on the perturbations. No significant departures from linearity were availability of items [26]. Taking this one step further, online observed, so all reported analyses use only first-order effects. As retailers could potentially use preference bias based on an indicator of the magnitude of the effect, we examined the recommendations to increase sales. distribution of the correlation coefficients for the individual 41 Further research is clearly needed to understand the effects of [13] Jacowitz, K.E., and Kahneman, D. 1995. "Measures of recommender systems on consumer preferences and behavior. Anchoring in Estimation Tasks," Personality and Social Issues of trust, decision bias, and preference realization appear to Psychology Bulletin, 21, 1161-1166. be intricately linked in the context of recommendations in online [14] Johnson, J.E.V., Schnytzer, A., and Liu, S. 2009. "To What marketplaces. Additionally, the situation-dependent asymmetry Extent Do Investors in a Financial Market Anchor Their of these effects must be explored to understand what situational Judgments Excessively?" Evidence from the Hong Kong characteristics have the largest influence. Moreover, future Horserace Betting Market," Journal of Behavioral Decision research is needed to investigate the error compounding issue of Making, 22, 410-434. anchoring: How far can people be pulled in their preferences if a [15] Komiak, S., and Benbasat, I. 2006. "The Effects of recommender system keeps providing biased recommendations? Personalization and Familiarity on Trust and Adoption of Finally, this study has brought to light a potentially significant Recommendation Agents," MIS Quarterly, 30, (4), 941-960. issue in the design and implementation of recommender systems. [16] Koren, Y., Bell, R., and Volinsky, C. 2009. "Matrix Since recommender systems rely on preference inputs from users, Factorization Techniques for Recommender Systems," IEEE bias in these inputs may have a cascading error effect on the Computer Society, 42, 30-37. performance of recommender system algorithms. Further [17] Ku, G., Galinsky, A.D., and Murnighan, J.K. 2006. "Starting research on the full impact of these biases is clearly warranted. Low but Ending High: A Reversal of the Anchoring Effect in Auctions," J. of Personality and Social Psych, 90, 975-986. 7. ACKNOWLEDGMENT [18] Lichtenstein, S., and Slovic, P. (eds.). 2006. The This work is supported in part by the National Science Foundation Construction of Preference. Cambridge: Cambridge grant IIS-0546443. University Press. [19] Mcnee, S.M., Lam, S.K., Konstan, J.A., and Riedl, J. 2003. REFERENCES "Interfaces for Eliciting New User Preferences in [1] Adomavicius, G., and Tuzhilin, A. 2005. "Toward the Next Recommender Systems," in User Modeling 2003, Generation of Recommendation System: A Survey of the Proceedings. Berlin: Springer-Verlag Berlin, 178-187. State-of-the-Art and Possible Extensions," IEEE [20] Mussweiler, T., and Strack, F. 2000. "Numeric Judgments Transactions on Knowledge and Data Engineering, 17, (6), under Uncertainty: The Role of Knowledge in Anchoring," 734-749. Journal of Experimental Social Psychology, 36, 495-518. [2] Bell, R.M., and Koren, Y. 2007. "Improved Neighborhood- [21] Northcraft, G., and Neale, M. 1987. "Experts, Amateurs, and Based Collaborative Filtering," KDD Cup'07, San Jose, Real Estate: An Anchoring-and-Adjustment Perspective on California, USA, 7-14. Property Pricing Decisions," Organizational Behavior and [3] Bennett, J., and Lanning, S. 2007. "The Netflix Prize," KDD- Human Decision Processes, 39, 84-97. Cup and Workshop, San Jose, CA, www.netflixprize.com. [22] Pu, P., and Chen, L. 2007. "Trust-Inspiring Explanation [4] Breese, J.S., Heckerman, D., and Kadie, C. 1998. "Empirical Interfaces for Recommender Systems," Knowledge-Based Analysis of Predictive Algorithms for Collaborative Systems, 20, (6), Aug, 542-556. Filtering," 14th Conf. on Uncertainty in Artificial [23] Russo, J.E. 2010. "Understanding the Effect of a Numerical Intelligence, Madison, WI. Anchor," Journal of Consumer Psychology, 20, 25-27. [5] Chapman, G., and Bornstein, B. 1996. "The More You Ask [24] Sarwar, B., Karypis, G., Konstan, J.A., and Riedl, J. 2001. for, the More You Get: Anchoring in Personal Injury "Item-Based Collaborative Filtering Recommendation Verdicts," Applied Cognitive Psychology, 10, 519-540. Algorithms," 10th International WWW Conference, Hong [6] Chapman, G., and Johnson, E. 2002. "Incorporating the Kong, 285 - 295. Irrelevant: Anchors in Judgments of Belief and Value.," in [25] Schonfeld, E. July 2007. "Click Here for the Upsell." Heuristics and Biases: The Psychology of Intuitive Judgment, CNNMoney.com, from T. Gilovich, D. Griffin and D. Kahneman (eds.). Cambridge: http://money.cnn.com/magazines/business2/business2_archiv Cambridge University Press, 120-138. e/2007/07/01/100117056/index.htm. [7] Cosley, D., Lam, S., Albert, I., Konstan, J.A., and Riedl, J. [26] Shih, W., S., K., and Spinola, D. 2007. "Netflix," Harvard 2003. "Is Seeing Believing? How Recommender Interfaces Business School Publishing, (case number 9-607-138). Affect Users’ Opinions," CHI 2003 Conference, Fort [27] Strack, F., and Mussweiler, T. 1997. "Explaining the Lauderdale FL. Enigmatic Anchoring Effect: Mechanisms of Selective [8] Deshpande, M., and Karypis, G. 2004. "Item-Based Top-N Accessibility," Journal of Personality and Social Recommendation Algorithms," ACM Trans. Information Psychology, 73, 437-446. Systems, 22, (1), 143-177. [28] Swearingen, K., and Sinha, R. 2001. "Beyond Algorithms: [9] Epley, N., and Gilovich, T. 2010. "Anchoring Unbound," J. An Hci Perspective on Recommender Systems," ACM SIGIR of Consumer Psych, 20, 20-24. 2001 Workshop on Recommender Systems, New Orleans, [10] Flynn, L.J. January 23, 2006. "Like This? You'll Hate That. Louisiana. (Not All Web Recommendations Are Welcome.)." New York [29] Thorsteinson, T., Breier, J., Atwell, A., Hamilton, C., and Times, from Privette, M. 2008. "Anchoring Effects on Performance http://www.nytimes.com/2006/01/23/technology/23recomme Judgments," Organizational Behavior and Human Decision nd.html. Processes, 107, 29-40. [11] Funk, S. 2006. "Netflix Update: Try This at Home." 2010, [30] Tversky, A., and Kahneman, D. 1974. "Judgment under from http://sifter.org/~simon/journal/20061211.html. Uncertainty: Heuristics and Biases," Science, 185, 1124- [12] Goldberg, K., Roeder, T., Gupta, D., and Perkins, C. 2001. 1131. "Eigentaste: A Constant Time Collaborative Filtering Algorithm," Information Retrieval, 4, (2), 133-151. 42