In various application domains, recommender systems explicitly or implicitly act as virtual advice givers. They are not only used to lter large item sets or point users to unknown but relevant items, their recommendations can also help users to make a decision given a limited choice set. Such a system is usually considered e ective if the users adopt the recommendations because, for example, the system's suggestions match their preferences or because they generally trust in the system's benevolence and competence. With this work we aim to further explore the persuasive potential of automated recommendations. Our speci c goal was to investigate whether the mere presence of a recommendation has e ects on the user's choice process. We conducted two online studies in which participants received either no recommendation or a random recommendation for a given decision scenario. The obtained results showed that the pure existence of recommendations can, depending on the decision scenario, make users more con dent in their choices and reduce choice di culty. Furthermore, we observed that in both studies even random recommendations led to an anchoring e ect as the participants' choices were measurably biased by the characteristics of the recommended item.
Recommender systems (RS) can serve di erent purposes for their users. They, for example, help users to locate relevant items within large item collections or support them in discovering additional items of interest outside their typical preference patterns. A somewhat less explored role of recommenders is their capability of serving as virtual advice givers in scenarios where users make decisions given a limited set of choices.
Such systems are often more interactive and can implement
a number of persuasive cues to increase the users' con dence
in their decision. Additionally, providers can employ the
persuasive potential of such systems to \convince" users to
choose a certain recommended option, e.g., by providing
appropriate explanations or by helping them understand the
relevant decision factors [
Previous works on this topic focus, for example, on
analyzing the in uence of speci c decision-support functionalities,
like explanations, on persuasiveness [
To investigate the existence of such e ects, we conducted
user studies in which the participants had to make purchase
decisions on ctitious e-commerce shops. One participant
group received one randomly chosen element from the choice
set as a recommendation; the other group received no
recommendation at all. We decided to rely on random
recommendations in our studies as this allows us to rule out potential
e ects related to the (perceived) quality of the
recommendations themselves. Besides the question if a randomly chosen
recommendation can represent an anchor and bias the nal
user decisions, our expectation was that the mere presence of
recommendations has a positive e ect on choice con dence,
e.g., because the users are given a reference point for their
decision. Higher choice con dence might lead to higher choice
satisfaction, which in turn is supposed to increase the users'
intention to actually make a purchase [
In summary, our research questions are as follows: RQ1: To what extent has the mere presence of a recommendation an e ect on the customer's decision process? RQ2: Can the characteristics of a recommendation serve as an anchor for decision making? 2.
Research Model. Figure 1 shows our research model. The independent variables are the presence of a recommendation (RS) and the user's domain expertise (Dom. Exp.). We include the latter variable assuming that expertise may have an impact on the users' decision con dence (Dec. Conf.). We include choice di culty (Ch. Di culty) as a construct as we hypothesize that users { utilizing the recommendation as reference point { might focus on a subset of the items as choice set. In turn, lower choice di culty should also lead to higher decision con dence. We measure choice di culty with indicators variables that assess the degree to which making
Decision Confidence +
Decision Satisfaction the decision was perceived as (emotionally) challenging1. Finally, both decision con dence and choice di culty are assumed to impact the users' decision satisfaction (Dec. Sat.).
Study Environment and Procedure. We created ctitious online shops for two di erent domains: backpacks (Study A) and hotels (Study B). In both studies, the scenario for the participants was that they were searching for an item to purchase, and we asked them to select one of the available items on the online site. The choice set sizes were 18 (backpacks) and 24 (hotels), respectively. In each case, additional item information was provided. We presented the weight, dimensions, volume, and price of the backpacks and the star category, community rating, distance to the city center and the price for the hotels. Half of the participants of each study received one randomly selected item as a recommendation, which was clearly marked as being a recommendation as sketched in Figure 2.
We recruited 164 and 239 participants for Study A and Study B, by distributing the URL of the online shop via email and on social network groups. The average age of the respondents was about 22 for both groups; more than two thirds were female participants. When accessing the website, the participants read the scenario and task description, were instructed to selected one of the options, and answered a post-task questionnaire. The participants were randomly assigned to the treatment groups.
The recommended item was actually chosen by 6:7 % of the participants of Study A, and by 3:8 % in Study B. These numbers roughly correspond to the theoretical chance that the recommended item was indeed the preferred option for a 1The questionnaire items can be found at http://ls13-www. cs.tu-dortmund.de/homepage/intrs13q. user. Simply displaying a random recommendation therefore did not persuade users to adopt the recommendations. 3.1
We used Structural Equation Modeling (SEM) as an
analysis instrument to detect relationships between the variables
of our research model from Figure 1. Speci cally, we used the
PLS-SEM method, which is particularly recommended for
this type of exploratory research for which no strong theory
exists yet [
We applied di erent validity and reliability tests to our
models and excluded indicators that were not reliably
measuring a construct [
In SEM models, path coe cients ( ), which range from
1 and +1, express the strength of the relationships between
two variables. The empirical t-values obtained through
bootstrapping help us assess statistical signi cance. According
to the literature [
The middle columns of Table 1 show the -values and t-values for the backpack study. The results con rm the hypothesized e ects of the presence of a recommender on decision con dence and choice di culty, which in turn both a ect decision satisfaction. The main insight of this study is therefore that the mere existence of a random recommendation can (a) have a positive, statistically signi cant e ect on the user's decision con dence and (b) lead to lower choice di culty for users. This nding is relevant in practice as lower choice di culty contributes to higher decision con dence ( = 0:228) and decision satisfaction ( = 0:322). Likewise, decision con dence is strongly tied with decision satisfaction ( = 0:608). Notes: = Path coe cient with corresponding t-value, **p < :01, *p < :05, n.s.= not signi cant.
The right-most columns of Table 1 show the results for the hotels. The obtained path coe cients indicate similar trends but the e ects did not reach signi cant levels in this scenario. This suggests that the hypothesized e ects depend on speci cs of the domain or the decision scenario.
In both scenarios, the expertise of the users { in contrast to our expectations { had no measurable direct or moderating e ect on decision con dence. This indicates that the observed mere-presence e ects in the backpack domain applied equally to both experienced and less-experienced participants.
Overall, the results indicate that the mere presence of
recommendations can measurably impact the user's decision
process in terms of decision con dence and choice di culty.
However, while these e ects were signi cant in the backpack
domain, they did not reach signi cance in the hotel domain.
Further research is therefore required to understand which
factors cause these di erences. One explanation could be that
comparing item characteristics might be inherently easier
in one of the domains. In fact, an analysis of the construct
values related to choice di culty revealed that choosing an
item was considered signi cantly (p < 0:05) more di cult in
the backpack domain, which could, e.g., be due to
domainspeci c trade-o s, such as the backpack's weight vs. its
volume. This suggests that the presence of a recommender
has more e ects when the choice situation is more di cult.
Another possible factor contributing to the perceived choice
di culty can be the size of the choice set [
Besides the question to what extent a recommendation in uences the decision making process, we aimed to investigate if the recommendations also had an e ect on the actual choices of the participants. We have already mentioned above that participants did not blindly adopt the random recommendations. However, we hypothesize that the participants could (unconsciously) be biased in their nal choice by the presented recommendation, i.e., that the recommendation served as an anchor for their decision. Speci cally, we assume that participants select options that have similar features compared to the recommendation. To our knowledge, the existence of attribute-level anchoring e ects has not been explored in the recommender systems literature so far.
Technically, we performed several univariate regression analyses to quantify to what extent the attributes of the chosen items were dependent on the features of the recommended item. Furthermore, as a simpler form of analysis, we calculated the correlations between the attribute values. The results of these analyses are shown in Table 2 and clearly show the existence of anchoring e ects for both scenarios.
In the backpack domain, all features of the nally chosen item, i.e., weight, volume, and price, were positively and statistically signi cantly related to the recommended item. All correlation values ( ) were also positive and signi cant at p < 0:05. Similar e ects were observed for the hotel domain. On average, participants chose items that were similar to the recommended item in terms of the distance to the city center, the community rating, and the price. No anchoring e ect was, however, observed for the star category in this domain. A possible explanation for this phenomenon could lie in the comparably coarse grained scale of the star category and that the participants might have already had a comparably strong mindset before the experiment regarding the star category of hotels they would possibly book.
To illustrate the strength of these e ects, we looked at the item attributes and created di erent subsamples of the data (e.g., light vs. heavy backpacks). For example, when the system recommended a light backpack with a weight between 1.7 and 2.8 kilograms, the average weight of the chosen backpack was at 2.26 kilogram. When the weight of the \anchor" was higher and between 2.9 and 4.0 kilograms, the average weight of the selected backpacks went up by 13 % to 2.60 kilogram. Similar e ect strengths were observed for other item features, which we nd remarkable, given that the recommendations were randomly selected.
In an additional analysis we tested if domain expertise had an impact on the strength of the anchoring e ect and incorporated these aspects into our regression models. We could, however, not observe any statistically signi cant main or interaction e ects, which suggests that both novice and expert users seem to be equally susceptible to anchoring e ects. 3.3
Our research is mostly based on responses from students of our university. While the group is homogeneous and students are potential customers in both tested domains, we cannot state with certainty that the ndings are representative for other societal groups. Furthermore, the participants did not actually make a purchase in the end, and our scenario was purely ctitious. On the other hand, as the participation was voluntary, no strong motivators exist for the participants to act dishonestly during the study. 4.
Anchoring e ects, as observed in both of our studies, were
rst discussed in the 1970s in the context of research on
human judgment under uncertainty [
In the RS literature only few works on anchoring e ects
exist. To what extent displaying predicted ratings for
unfamiliar items in uences the ratings assigned by users is
discussed in [
In a broader context, anchoring e ects can be seen as
one of several possible approaches to implement persuasive
recommender systems [
In contrast to these works, our studies indicate that the
mere presence of random recommendations can have a
persuasive and biasing e ect. More research is however required
to understand the underlying reasons of these e ects. Past
research showed that users see (personalized) recommendations
as a decision aid that can reduce the perceived e ort and
choice overload [
Our work suggests that the mere presence of random recommendations can measurably a ect the choice processes of users. In both studies reported in this paper we could observe anchoring e ects on the attribute level, i.e., the participants exhibited a tendency to select items that had similar characteristics compared to the recommended reference item. In one of the tested domains, the presence of the recommender furthermore led to lower perceived choice di culty and higher choice con dence.
Overall, our work therefore contributes additional evidence of the persuasive capabilities of recommender systems and their potential as decision-making aids. In terms of practical implications, the observed anchoring e ects emphasize that recommenders can be valuable instruments for providers to guide the customer choice toward a desired direction. Since not all e ects could be observed in both studies, more work is required to understand the underlying factors that determine the e ective persuasiveness of recommendations in di erent scenarios.