Effect of “Similarity” and “Friend” Framing on Recommendation Click-throughs: Influence of Gender, Category Experience and Privacy
MSc. in Business Administration – Digital Marketing Track
Student: Ela Akbulut Student number: 13299867
Date: 25.06.2021 - Final Submission Supervisor: Shan Chen
EBEC Approval Number: 20210408080416
Statement of Originality
This paper is written by student, Ela Akbulut who declares the full responsibility for the content of this paper.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Table of Content
Abstract: ... 3
1. Introduction: ... 4
2. Literature Review ... 8
2.1 Recommendation Systems and Consumer Behavior ... 10
2.2 Framing the Recommendation Systems... 11
2.3 Impact of social influence on framings ... 13
2.3.1 Friend factor in framing ... 14
2.3.2 Similarity factor in framing ... 16
2.4 Demographic Impacts ... 18
2.5 Impact of category experience on recommendation systems ... 19
2.6 Impact of privacy on recommendation systems ... 22
3. Research Design ... 24
4. Analysis: ... 27
5. Results: ... 29
5.1 Study 1: Impact of friend and similarity framing on click-throughs ... 30
5.2. Study 2 & Study 3: Impact of gender, cat. experience, privacy on click-throughs ... 33
6. Summary & Conclusion ... 39
7. Managerial Contribution ... 40
8. Future Research and Limitations ... 41
References: ... 44
This study explored explanation framings used in recommendation systems and other factors like gender, category experience and privacy, which may influence recommendation effectiveness. As framing types, similarity and friend factors have been used to understand if tie strength has a different influence on consumer's behavior. The study was conducted as an experiment survey by designing a music streaming platform experience for participants. Click-throughs of recommendations were measured based on different framing types and information provided by customers about their gender, category experience, and privacy concerns. The results showed that recommendation explanations with each framing type do not have any influence on recommendation click-throughs. However, gender and category experience has shown some significant direct effect on click-throughs and interaction effect with framing types. The findings also give important contributions for companies using recommendations.
Have you ever purchased or clicked on an item that was offered to you by the platform you use as a recommendation? Even if you haven't bought or clicked the item, it is highly possible that you have interacted with recommendations in some way. By definition, recommendation systems are machine learning algorithms that generate predictions according to data about the recent behavior of consumers (Lü et al., 2012). Recommendation systems became a major issue, specifically in e-commerce, and gained a critical value to be indeed more significant in the future.
Netflix’s recommendations through customers’ watch history and other customers’ given ratings, and Amazon’s: “Customers who bought this, also bought this” are just a few outstanding examples of recommendation systems (Sharma & Dutta, 2020; Linden et al., 2003). Recommendation systems are an important example of personalization. They create user engagement and build customer loyalty by offering added value with recommendations (Schafer et al., 1999). Customer loyalty is also directly proportional to recommendation quality (Yoon et al., 2013). Therefore, firms build different recommendation system approaches to make the best offerings for their customers.
In recent years, firms also started to focus on different explanation framings to make recommendations more effective. This also created a new era for researchers that studies recommendation explanations and other factors that affect recommendation effectiveness which can be beneficial for online platforms.
There are many types of framings as a part of marketing communications tactic of companies that are used in recommendations. We also know these as explanations of recommendations. The type of framings used in the explanations can impact the effectiveness of recommendations. For instance, when Amazon prefers framings like "Customers who bought this, also bought this", Spotify uses the "friend" factor and frames its recommendations as: "Your friends
have listened to these songs". Besides, many book store websites use framings like "Other readers with similar tastes also read these". The potential importance of recommendation system framings is discussed by Gai and Klesse (2019) when they studied framing a recommendation system and found out that collaborative filtering algorithms result in more click-throughs compared to framing it as item-based. This finding creates new research opportunities to study whether some common framing types used in recommendation explanations can be effective by using social influence factors (Shang et al., 2011).
Social influence has always been an essential player in the online environment. It has been an important factor to influence people's decision-making process, and in fact, it is a very extended area to do research (Dahl, 2013). Therefore, in this study, I will mainly discuss two very well- known factors of social influence, friends and other people who are similar to us (Kouki et al.
2019). The study will mainly look at: While communicating the recommendation systems to users, using social influence factors like similarity and friend can influence the effectiveness of recommendations. Gkika and Lekakos (2014) stated that recommendations with social proof are more persuasive for people, and because of this, these recommendations are more effective. When we individually evaluate similarity and friend, both of them are effective factors in consumer behavior as the literature indicates. In this study, friends are perceived as strong ties who are two parties and familiar with each other and interact with each other via a platform when other people who are similar to us are being weak ties. According to this, this study will measure the effect of strong and weak tie influence on recommendation click-throughs. People trust the preferences of others who are similar to them mainly because they feel like a part of a group (Fogg, 2002). Similar effect is also present with friends since consumers find references from strong ties as more reliable (Zhou & Tian, 2010). However, there are also counter arguments in the literature saying that people
would prefer to choose more distinctive choices instead of selecting similar options or their friends' preferences which they may already know about (Argo & White, 2011; Brown & Reingen, 1987).
Thus, it creates the research question of whether recommendations framed using strong vs. weak ties result in any positive or negative effects on recommendation click-throughs.
Besides the recommendation framing, there are other factors that can have a moderating influence on recommendation click-throughs. The reason is, consumers’ behavior towards recommendations depends on their decision-making process which could be affected by demography like gender, age, marital status, experience and interest and their perception towards some concerns. Therefore, one important factor which may influence recommendation click- throughs is gender. The effectiveness of recommendation systems can also be dependent on gender factors. The reason is, in the decision-making process, male and female behaviors have always differed from each other (Bakshi et al., 2012). For example, in the shopping context, men are more evasive when women search for pleasure and, therefore, more critical during the process of evaluation (Bakshi et al., 2012). Consequently, it is expected that men and women will evaluate the recommendation framings differently. Furthermore, their evaluations can also differ per framing type based on strong and weak ties. For instance, women feel more obliged to buy something compared to men when they go shopping (Kolyesnikova, 2009).
Another factor that may influence recommendation click-throughs is category experience, which in other words, consumers' familiarity and knowledge with the product category. Consumers with high category knowledge are better at understanding and evaluating the product attributes (Mason & Bequette, 1998). Category experience and knowledge also influence the decision- making process of consumers. For instance, people with high category or product knowledge tend to show more impulsive buying behavior (Liang, 2012). However, in some circumstances like
recommendations, consumers can be more critical towards the recommended options if they are more experienced (Duhan et al., 1997). Therefore, category experience also holds as an exciting factor to measure its effect on recommendations.
Privacy concerns may also affect the click-throughs of recommendations. Recommendation systems are aimed to be personalized, and for this, they use customer data which can cause customers to be concerned about their privacy and usage of their data. The reason is, as Gedikli et al. (2014) mention, such explanations provide transparency and present reasoning of how the recommendation was made. Thus, the use of explanations is beneficial to establish customer trust for the recommendation system. Also, it can be helpful to reduce the information asymmetry caused by recommendation systems which can result in privacy concerns.
In general, this research will dig into explanation types framed by strong and weak ties as social influence and other factors such as gender, category experience, and privacy to see if they have any moderation or direct effect on recommendation click-throughs. As mentioned before, for weak ties "similarity" factor will be used, and for strong ties "friend" factor will be used. The findings will be based on whether different framing types result in different click-throughs or not and whether gender, category experience, and privacy impact the click-throughs. In addition, the study will also look at the interaction of gender, customer experience with the product, and privacy concerns with each framing type. Therefore, the study will address the question: How do the
"similarity" and "friend" factors used in recommendation explanations affect recommendation click-throughs and whether there's any type of effect of gender, category experience and privacy on click-throughs?
The contributions of this study may provide both academic and managerial guidance in how to communicate their recommendations by using which framing types to be more persuasive. As Gai and Klesse (2019) claimed, a backfiring effect of “friend” based framings may guide companies like Spotify to change their explanation of framings. Besides, the results may provide framing ideas to companies which would like to enrich their recommendation system effectiveness.
Thus, enhancing companies’ recommendation quality may increase click-throughs of recommendations, and this would fulfill long-term company objectives.
The study's academic contributions can create a new era for studying more creative framing types based on consumer behavior. Moreover, since the importance of recommendation systems is getting more important, changing focus to other factors besides recommendation algorithms can result in important findings in academic literature.
The rest of the paper will dig into the literature based on recommendation systems, consumer behavior, and all the factors which will be measured in the analysis. First, the research design will be explained and then analysis results will be discussed. Lastly, the contributions of the study and limitations, and future research ideas will be mentioned.
2. Literature Review
Recommendation Systems have been studied in the literature for more than 30 years since it is a trending area both for managerial and theoretical contributions. Recommendation systems hold a crucial position in customer experience by increasing personalization based on users' recent activities and behaviors by offering new alternatives accordingly (Ansari et al., 2000). The impact of recommendation systems on businesses has also been discussed in current literature.
Recommendation systems establish a long-term user engagement which can also benefit companies' long-term objectives (Jannach, Jugovac., 2019). They create not only user engagement, but also build customer loyalty by offering added value (Schafer et al., 1999). As an addition to building loyalty, in another article Schafer et al. (2001) mention two positive effects of recommendation systems on e-commerce. These are: Turning browsers into purchasers, increasing order size and cross-sell (Schafer et al., 2001). Besides, Pathak et al. (2010) stated recommendation systems and their positive influence on sales based on the effectiveness of recommendations. When a recommendation system is designed more effectively, it will return better business outcomes (Pathak et al., 2010). However, Lee and Hosanagar's (2019) research shows that recommendation systems can also negatively affect sales in terms of sales diversity by increasing the demand for already popular products. Thus, firms with extensive product categories may prefer different recommendation designs (Lee & Hosanagar, 2019). This also shows that the effectiveness of recommendation systems may differ between industries and require businesses to come up with their own recommendation designs.
Besides this already built general wisdom about recommendation systems, most of the existing studies are generally focused on how the recommendation systems work and ways of enhancing their algorithm. There are three different, well-known recommendation system algorithms, but there are various ways to frame each of them. First and most popular one is collaborative (user-based) filtering. It identifies customers with similar tastes based on their past product preferences and recommends one customer's preference to another similar customer. For collaborative filtering companies use framings like: "Customers who bought this, also bought this".
Second type is called content (item-based) filtering. Despite collaborative filtering, content filtering only considers the customer's prior preferences and recommends similar products accordingly by
not integrating it with other customers' preferences by framings like "You bought this item, you may also like these". Firms use both recommendation system types based on their market strategies.
However, since both filtering have some limitations -lack of historical behavior of consumer data for cold starters in collaborative filtering, and overspecialization of items in content filtering- a new type of approach called Hybrid is shown (Thorat et al., 2015). It incorporates both collaborative and content filtering by applying meta-paths. This approach holds more networks instead of focusing only on users or items (Zare et al., 2020). Thus, this approach enables more space for framings of recommendation explanations. These approaches consist of algorithms of recommendation systems. After applying the most appropriate algorithm, the further step for firms is to decide on how they will communicate these recommendations.
2.1 Recommendation Systems and Consumer Behavior
Existing literature also provides extensive studies about recommendation systems and consumer behavior, which is an area more associated with this paper’s research. Recommendation systems are a part of personalization. It is studied that, even though it may vary between cultures, notably young consumers desire personalization (Torico & Frank, 2019). Personalization causes consumers to develop effective responses and customer satisfaction (Torico & Frank, 2019).
However, the way the personalized content is proposed to consumers also matters and affects customers' decision-making processes. A significant point to consider while communicating recommendations are potential consumer biases. For instance, displaying the best attribute of a product at the top may increase the demand for that recommended item (Teppan & Zanker, 2015).
Likewise, presenting over one option, one of them being the poor option, leverages consumers to prefer the desired option for firm goals (Teppan & Zanker, 2015). Similarly, an article by Theocharous et al. (2019) also studies the effect of cognitive biases on personalized content. The
authors state that "Framing Effect" is a significant collective bias of customers, which is usually used by firms' strategic offerings to use in explanation framings of recommendations (Theocharous et al., 2019). For instance, Amazon's "Customers who bought this, also bought this'' approach is an example of explanation framing. These types of explanations which indicate the source of recommended items are important for consumers since it provides transparency (Gedikli et al., 2014).
The early research about explanations has primarily focused on explanation interfaces such as: Personalized tag cloud, neighborhood-based, Movilens, etc. to increase the explanations’
trustworthiness (Gedikli et al., 2014). Besides the existing research about explanation interfaces, some other researchers studied the different types of influences of explanations on users. Tintarev and Mastoff (2012) categorized these types as persuasiveness, effectiveness, satisfaction, scrutability, transparency, trust, efficiency. For instance, indicating how the recommendation is offered to customers like Amazon’s explanation example, enhances the transparency of recommendations. Besides, Karacapilidis et al. (2017) stated that additional information provided in explanations is relied on more by customers and strengthens the trustworthiness. In general, it is fair to say that explanations used in recommendations are important to customers for transparency, and using different framing types can improve the effectiveness of recommendation explanations.
2.2 Framing the Recommendation Systems
When we evaluate all these studies, none of them clearly acknowledges the marketing communication side of recommendation systems and explanations. However, communication of recommendation systems to customers also has a role in the effectiveness of recommendations.
The main reason is, it is important to be transparent with customers while communicating the
recommendations. Recommendation explanations are seen as transparency of offered recommendation since it gives a reasoning about the way that the recommendation was made.
Gretzel and Fasenemier (2011) point out the importance of transparency while designing the recommendation explanations. In other words, users should be able to understand how and from which source the recommendation was offered.
Besides, the way the recommendations are communicated to customers, with different framings, may affect the recommendation's effectiveness and persuasiveness. As Gai and Klesse (2019) discuss, framing of recommendation systems as user-generated or item-generated may have diverse effects on click-throughs independently from recommended content. Their conclusions indicate that regardless of the algorithm used (user-based vs item-based) framing it as user-based results in more click-throughs compared to item-based framing of used in explanations of the recommendations (Gai & Klesse, 2019). On the contrary, an article by Kouki et al. (2019) states that customers prefer item-based explanations more compared to user-based or social-centric ("your friends liked this") explanations when the system provides only one explanation. This difference might be due to industry differences since industry types can have influence on recommendations. For instance, for traveling, people rely on their friends and relatives' ideas a lot (Gitelson & Kerstetter, 1995). Besides, Kouki et al. (2019) added that specifically open customers prefer more explanation styles instead of one explanation since they like to encounter different things (Kouki et al., 2019). This finding implies that firms require enriching their explanation styles by using both item and user-generated explanations. Spotify is a wonderful example of this finding.
It uses a user-based algorithm with explanations like "your friends listened to these". At the same time, it recommends daily playlists based on an item-based recommendation algorithm and communicates it as "Here is a playlist for your taste". This study also favors this finding, and it will
contribute its research on different framing types' effectiveness on recommendations, which can be useful more if practiced with other recommendation explanations.
2.3 Impact of social influence on framings
Social influence has a big role in consumers' decision-making processes. Word of mouth (WOM) and influencer marketing are a few great examples that show the strength of social influence in today's world. The impacts of WOM and mass media have been studied for years, including Lazarsfeld et al. (1944) and two-way communication concepts. Their concept stated that mass media would influence the opinion leaders and opinion leaders would influence the society (Lazarsfeld et al., 1944). Especially in the online world, customers follow or consider other customers' ideas before and after taking any action. For instance, customers look at ratings or reviews shared by other customers before making a purchase. They are even influenced by other reviewers while sharing a review about a product to be consensus with other reviewers or not (Moe
& Schweidel, 2012). They decide between two main aspects of consumer psychology: following others with a bandwagon effect or differentiating themselves by being non-consensus (Moe &
Schweidel, 2012). Therefore, people can still be influenced by others even if they do not follow or share a common opinion with them. In their study, Matthew et al., (2011) state that positive messages by other customers about their shopping experience change people's attitudes towards online shopping or the product. They start to feel more positive towards online shopping (Lee et al., 2011). Opinion leaders or authorized people also have an important power on other people. For instance, a study by Haans et al. (2013) argues that in Google Search Results, ad-body text created by using expert evidence gets more click-throughs. Authors think that this is the case due to more reliability with expert opinion, so that people find it more persuasive (Haans et al., 2013). Similar findings were also shared by Gkika and Lekakos (2014) about recommendation systems. Their
study claimed that recommendation explanations that include authorized people or social proof tend to be more persuasive by customers (Gkika & Lekakos, 2014). A different point of view about social influence is how people feel forced by others. They find others as a pressure to be influenced especially while adopting new technologies (Dennehy & Sammon, 2015; Venkatesh et al., 2012 as cited in Singh et al. 2020). Therefore, the pressure can also trigger or shape people's preferences and decision making as a result of other people.
In line with all these studies, it might be interesting to identify main social influences. In their study, Kouki et al. (2019) distinguished recommendation system framings in a more extended way than dividing them based on their algorithm as user-based and item-based. They divide them in 5 categories: User-based which is framed as "customers who have similar tastes with you also looked at these", item-based framed as, "here are some other products based on your previous preferences" (Kouki et al., 2019). Others are, content which is based on tags and framed as, "this product has similar tags with the one that you viewed", social which frame recommendations by using "your friends like this product", and lastly a non-personalized explanation which is based on item-popularity and framed as "these items are very popular" (Kouki et al., 2019). In this research, this categorization will be used, but by merging Kouki's user-based and social categories together since underlying the similarity between other users are also a social influence factor as similarity attraction theory also suggests.
2.3.1 Friend factor in framing
For the friend factor, people feel higher self-esteem when they focus on their impression of strong ties such as friends and family and feel more satisfied than their shown image to weak ties (Wilcox & Stephen, 2013). Therefore, a recommendation system which indicates that it offers
products bought or searched by strong ties, may also encourage consumers' behaviors within the platform since they would act with more self esteem by knowing their behaviors will be recommended to their friends. A study by Ye et al., (2012) measured the degree of people's influence from their friends' opinions in decision making. They found out that consumers usually follow their own preferences. However, they also get affected by their friends' opinion (Ye et al., 2012). The authors also underline the benefit of social influence for recommendation systems by referencing users' friends since it is possible that users will most likely share common interests with their friends (Ye et al., 2012). In another study, people with narrower friend networks were found to be influenced by their friends more compared to people who are in larger friend groups (Chong et al., 2021). When many researchers study the influence of friends, Klepper et al., (2010) gave evidence about the opposite effect between social influence and friendship. They stated that friends become similar to each other when they are influenced by each other (Klepper et al., 2010).
Therefore, friendship is built not with selection but with influence (Klepper et al., 2010). Besides, Brown and Reingen (1987) discuss that, for word-of-mouth, weak ties have more impact on customers since people assume that they already share similar tastes with their strong ties. Gai and Klesse (2019) also support this argument by claiming that using "friends" in framings may backfire since customers may not prefer to click on their friends' choices since they already know about their friends' tastes. Erkan and Evans (2016) adds that in eWOM people prefer other unknown people's opinions more compared to their friends' opinions and because of this, they look at online reviews instead of asking their friends. Authors define major reasons for this finding as presence of more detailed information online, number of reviews and ease of access to others' opinions (Erkan & Evan, 2016). Even though this might be the case for friend framing compared to other framing types, this study argues that friend framing would still influence people about recommendations compared to no framing. According to research done by Forrester, 70% of
Americans were more likely to trust their friends and family's opinion (Stokes, Cooperstein, and Hayes 2013). Ye et al. (2012) also argued that influential effect of friends is higher when there is not too much similarity between people and their friends. Even with the existing literatures' findings, the exact impact of strong ties in recommendation systems has not been found yet.
Therefore, I hypothesize that referencing friends in recommendations would increase the perceived trustworthiness of recommendation systems for the second hypothesis. Thus formally,
H1: Using friend framing in recommendations results in a significant positive impact on click-throughs compared to no framing.
2.3.2 Similarity factor in framing
Similarity attraction theory suggests that people are influenced by others who they share similar attitudes with. This was also proven by Byrne and Nelson (1965) when they found out the positive linear relationship between attraction and similarity. People can be influenced by similarity for many reasons, such as relying more on similar people's ideas or sense of belongingness to a specific group with similar interests. Fu et al. (2020), stated the significant effect of similarity on trust. They also added that it enables customers to create groups on online platforms (Fu et al., 2020). Besides, sometimes they are attracted by similarity to be able to reach their ideals by expecting that similar others are similar to a person's ideal own version. (Wetzel & Insko, 1982).
Because of all these, similarity factor is also used in recommendations by framings like "people who have similar tastes with you".
Framing the recommendations with the reference groups that share similar attributes provide a descriptive social norm to customers. In fact, underlines the reference who is similar to the customer, which is an example of the provincial norm (Goldstein, 2008). Goldstein's (2008)
study claims that provincial norms motivate people more in their decision-making process. As Fogg (2002) states, the similarity factor can be more persuasive since people feel like they are part of a group or a team. Accordingly, it is fair to say that people can be influenced by others with who they share similar tastes. In fact, perceived similarity between people can also be increased if their interaction and emphasis on their similar tastes are continued for a while (Crandall et al., 2008).
Besides, using similarity can also benefit the algorithm system of platforms since it causes less user query and product search results, providing a better customer experience (Ricci et al., 2003).
However, at the same time, people like to imitate others in consumption, and this may result in distinctiveness in people's choices since they usually search for things disassociated with their taste (Argo & White, 2011). Thus, using similarity framing may have the opposite effect for consumers who are searching for more diversified items. Moreover, Naylor et al. (2011) says that, when the reviewer is not known, people find the reviewer close to themselves. Therefore, if people are able to establish similarity with others easily, emphasizing the similarity factor in recommendation framing may not result in a significant effect on recommendation click-throughs.
To understand its real effect, in this study, we will see whether emphasizing on "similarity" terms in recommendations has an influence on click-throughs to recommendations. Based on the literature, I argue that, the majority of users would be more likely to click on recommendations according to their similar tastes with others. Therefore,
H2: Using similarity framing in recommendations has a significant positive effect on click-throughs compared to no framing.
2.4 Demographic Impacts
The effect of tie strength can also depend on cultural and demographic factors. The study by Broeder and Hout (2019) compares online purchase behavior of Dutch and Russian people and states the finding as both cultures prefer recommendations from their strong ties but its effect is more for Dutch people.
Regarding the effect of demographics on recommendations, there was some literature about the impact of gender on the online environment. It was studied by Chang and Chin (2010), that since females are less experienced with online shopping and e-commerce, they are more likely to use recommendations compared to men. Similarly, Suan et al. (2015) stated that since females like social interaction more, they can be more likely to use recommendations or reviews too. However, Beel et al. (2013) found that there is not a significant difference between males and females on recommendation click-throughs. Therefore, even if the literature has the findings that online behaviors of males and females can differ, the study findings which were based on recommendation systems will be considered more since this study also looks at recommendation systems and user evaluations of people might not be the same with other online behaviors. Therefore, regarding the direct effect of gender on recommendation click-throughs, following hypothesis is prepared:
H3A: Gender does not have a direct effect on recommendation click-throughs.
However, gender's effect on click-throughs can change depending on the framing type used. An article by Argo and Dahl (2020) claims a different perspective about the social influence of friends. In the retail context, they mention that the presence of friends may affect men and women differently since men may find it better when women become a little modest about it. At the same time, while friends may encourage men during shopping, they can also cause more
expenditure (Kurt et al., 2011). The reason is, men are spending more with their friends since they want to improve their self-presentation and style from their friends' perspective (Kurt et al., 2011).
In this study, the focus will be more on gender's impact on recommendation framings and whether a similar result is found like in the retail context for friend framing. In hypotheses 3B and 3C, the moderating effect of gender on friend and similarity framing respectively will be measured.
However, since the existing literature only provides arguments about gender and friend interaction and does not provide any evidence about gender's moderation effect with similarity, this study is not expecting a moderation effect of gender on similarity. Therefore, the other hypotheses will be as follows:
H3B: Females click-through less on recommendation with friend framing compared to males.
H3C: Gender does have an effect on the relation between similarity framing and click-throughs.
2.5 Impact of category experience on recommendation systems
Customers' experience of a connection with the online platform or recommendation agents is important. If users were satisfied after their experience with a recommendation agent, and became familiar with the platforms' recommendations, they are more likely to find recommendations effective (Komiak & Benbasat 2006). Furthermore, the connection with the website or online platform also impacts the effectiveness of social influence (Iyengar et al., 2009).
For instance, less-connected people show less interaction with other people and pay less attention to others' opinions (Iyengar et al., 2009).
Besides this, user experience in a specific category or product also matters. In fact, category experience has been an important part of recommendation systems. The reason is, recommendation systems have been built with the idea of providing guidance and influence to users with recommendations (Resnick & Varian, 1997). Especially recommendations were expected to be more beneficial for less experienced users (Resnick & Varian, 1997). The opposite effect can be observed for highly experienced customers. Yoon et al. (2013), claims that customers with more product or category knowledge are less satisfied by recommendation quality since they are already familiar with the category and search for diversity. Mainly, in collaborative filtering recommendation systems, users with low category experience find recommendations more trustworthy and find them more effective (Yoon et al., 2013). Therefore, authors find a negative relation between category experience and customer satisfaction with recommendation systems (Yoon et al., 2013) In fact, this dissatisfaction with recommendation systems can also cause dissatisfaction towards the website or platform (Yoon et al., 2013). Thus, category experience is important to consider since it may have serious impacts on the success of online platforms. It is mainly because users with high category experience are more aware of recommended products and can evaluate differences between them in a detailed way, so that they can be more critical (Duhan et al., 1997). Gai and Klesse (2019) point out the importance of consumption experience by finding that the effectiveness of user-based framing vs. item-based framing decreases when consumption experience increases since more experienced people do not find taste matching factual. More specifically, Oramas et al. (2016) stated that, in music platforms, recommendation's effectiveness depends on familiarity of users with the recommendation system and also users' expertise on music.
Besides, a study about cognitive recommendation explanations in music streaming platforms found that rational thinking customers regard explanations more to understand the recommendation process, while intuitive customers need explanations only when they are unfamiliar with the
recommended product (Millecamp et al., 2020). Thus, it is expected that the cognitive style of users can also influence users' decision-making process and especially intuitive customers who are familiar with the recommended product would not click on recommendations (Millecamp et al., 2020). Therefore, I hypothesize that, regardless of the recommendation framing, recommendation click-throughs would be less when category experience is high. Therefore,
H4A: Category experience has a direct negative effect on recommendation click- throughs.
The way the recommendation is communicated through users may also show interesting results. Liu et al. (2021) stated that strong advertising could be more influential for experienced customers. However, less experienced customers can find a strong advertising attitude less trustworthy (Liu et al., 2021). Since recommendations can also be perceived as advertising, the way it is communicated can also change the click-through effects differently for experienced and not very experienced customers. In fact, the influence of category experience also differs between weak and strong ties. As Godes and Mayzlin (2009) found, when there is less knowledge about the product, people tend to be influenced by others who are not very close to them. The reason is, those people can reach more people and create a more extended awareness especially for users with less knowledge (Godes & Mayzlin, 2009). Besides, category experience can also affect difficulty level and length of decision-making process since prior knowledge exists (Duhan et al., 1997). Because of this, experienced customers with high knowledge finds decision making process less difficult and rely on their evaluations (Duhan et al., 1997). Thus, they prefer recommendations by weak ties when less experienced customers with a product find decision making processes more difficult and complex so that they are more influenced by strong ties due to more trust (Duhan et al). Also, Ye et al., (2012) found that most people rely on their own choices, but friends' opinions may also
influence their decisions but mainly if preferences of friends are familiar to them. This also brings us to users' experience with a certain category which means, if users are experienced with their friends' preferences, they might be more likely to click on the recommendations offered based on their friends. Therefore, to measure the impact of category experience on framing types, two hypotheses will be tested:
H4B: Higher category knowledge decreases the effectiveness of friend framing on click-through.
H4C: Higher category knowledge decreases the effectiveness of similarity framing on click-throughs.
2.6 Impact of privacy on recommendation systems
Anything including data usage in an online environment usually causes privacy concerns to customers due to the sharing of customer information (Malhotra et al., 2004). In fact, information privacy concern is a big threat for many online platforms (Malhotra et al., 2004). Hong and Thong (2013) specified consumer concerns on the internet level by calling them Internet Privacy Concerns which includes customers' concern for the flow of their information through websites and companies who play a part on the Internet (Hong & Thong, 2013). A very well-known reason for the information privacy concern for economists is information asymmetry. Information asymmetry causes trust issues between two parties when one party knows that the other has more information about him/her and the market (Calo, 2015). Therefore, people's attitudes towards customization or personalization activities used in online environments can be shaped negatively because of too much information asymmetry. Since recommendations also include the use of data and past
behavior of customers, customization, and personalization, similar effects can also be seen here.
Therefore, with hypothesis 5A the following will be tested.
H5A: Higher privacy concern results in less recommendation click-throughs.
Article by Knijnenburg et al. (2012) indicates that while communicating recommendation systems even the words used can have influence on people’s privacy concerns. For instance using
“your ratings data” instead of “all your activities” reduces privacy concerns, when the second one increases the privacy concern even though both framings mean the same meaning (Knijnenburg et al., 2012). Therefore, this study argues that users’ privacy concerns for differently framed recommendations can have effects on click-throughs.
Social influence factors which I distinguished as weak vs. strong ties may also shape privacy concerns differently. For instance, weak ties may result in privacy problems since the source of the recommendation is unknown (Ramakrishnan et al., 2001). Thus, it is important to reflect the recommender systems algorithm in an informative way for users by indicating the relation between the user and weak ties (Ramakrishnan et al., 2001). On the other hand, users' relationship boundaries may also influence their attitudes towards framing types because of different privacy concerns (Wisniewski et al., 2016). Therefore, since people have the autonomy to decide on others to have a connection within the platform, the usage of friend factor in recommendation framing may cause less privacy concern compared to similarity framing.
When the framing type is more personalized for the individual, this may also trigger privacy concerns. For instance, the study aims to capture the persuasiveness of framing with similarity factors but at the same time underlying "customers who have similar tastes with you" to make it more personalized may cause privacy concerns too since people may think it collects too much
data. For friend framing, concerns can be more since users' data and their friends' data are used by the recommendation systems (Toch et al., 2012). In brief, data collection is perceived as more extended since it tracks user information which causes less autonomy to the user and increases privacy concerns (Toch et al., 2012). Framing them as "other customers" or "your friends" will also indicate an adaptation of recommendation systems which is not only to the user but user's social environment and wider web (Toch et al., 2012). This would also decrease user control and increase privacy concerns (Toch et al., 2012). Thus, this can also result in a backfire in framing types.
Therefore, to test this, the hypotheses are:
H5B: Higher privacy concern would result in less click-throughs for friend framing
H5C: Higher privacy concern would result in less click-throughs for similarity framing
3. Research Design
For this study, 2x2 between-subjects design was created and an online survey experiment by using the tool Qualtrics was formed. The survey experiment was designed by using vignettes with a scenario to provide real-life experience to participants to elicit their click-through behaviors on different recommendation framings. In this scenario, a music streaming platform was used by encouraging participants to imagine themselves listening to some music on a platform like Apple Music or Spotify and then receiving some song recommendations.
Different industries or platforms also have an influence on recommendations. The reason I used a music streaming platform was to be able to measure the online consumption experience of
users with the usage of recommendations on the platform. It was also possible to use other platforms like e-commerce. For example, e-commerce platforms like Amazon use many framings like "others bought these" as a recommendation. In most of these platforms using "friend" as a framing is not applicable since the platform does not include such a friend connection network and lack of this data. Thus, it would not be realistic to apply this survey using e-commerce platforms.
Therefore, in order to see the effect of social interaction, music streaming platforms were more ideal. Since users can also have networks on music streaming platforms, using both "friend" and
"similarity" framings is more applicable to use in this study.
On the other hand, music streaming platforms like Spotify provide a connection with friends feature. Besides, platforms that offer utilitarian product categories were not applied in this study.
The reason is, customers are less sensitive to recommendations for utilitarian products (Gai &
Klesse, 2019). Therefore, the survey experiment was similarly designed with Spotify or Apple Music, since it has data about users' friends, and does not include utilitarian products, and is ideal to provide an experience to participants by using both framing factors in the survey.
In the survey experiment, customers were first provided by a sample song and then accordingly they were randomly assigned to 4 conditions with different framing types as Table 1 shows. Then, 4 different songs and not clicking on any of the songs option were provided for customers to make a selection. After their selection, they indicated their preference and level of category experience and privacy concerns.
Specifically, the survey experiment started by making randomly chosen participants listen to a certain song like they are listening to this song on a normal occasion on the music streaming
platform they use. Then, they were randomly assigned to one condition out of four conditions which included different framing types.
Table 1: Experiment Design
Similarity Framing No Yes
No No framing Friend Framing
Yes Similarity Framing Friend & Similarity Framing
The song played at the beginning of the survey was The Box by Roddy Ricch and chosen according to participants' affinity. Since participants' different interests based on music genre, music artist or song could have had impacts on their decisions for recommendations, the music genre hip-hop was chosen based on the most played music genre of 2020 results. Similarly, The Box was chosen as the sample song for this survey since it was the most played Hip-Hop song of 2020 (Newsroom 2020). Then, participants assigned one condition from four different conditions (see table 1) which included different recommendation framings. One was only using the friend framing as "Your friends have also listened to these songs", one used only the similarity framing as "Others who have similar tastes with you also listened to these songs". Third condition included both factors as "Your friends who have similar tastes with you also listened to these songs" and the last one was the baseline and control variable and didn't include any framing, instead only said
"some other songs". In each of these conditions, the same four songs were offered as recommendation but participants were also provided with the "I wouldn't click on any of the songs"
option. Thus, it was planned to measure the click throughs according to different framing types.
While choosing the recommended songs, a pre-test was conducted. In this pre-test, songs with the same genre with the sample song, were selected from the most played songs of 2020 list to avoid affinity of interest and provide songs that most of the people would like. However, the pre-test results indicated that participants tend to pay less attention to recommendation framing and instead immediately choose one of the songs. This was understood when participants were asked to choose the recommendation framing that was shown at the beginning. In Particular, the majority of them mixed "others who have similar tastes with you" and "other songs" from each other. Therefore, for similarity framing "people" was used in the main survey instead of "other". Also, to have more cognitive elaboration during the decision-making process, less well-known songs in Hip-Hop genre were randomly chosen from an existing Spotify's newly released Hip-Hop songs playlist. The genre was kept the same both in sample song and recommended songs to avoid affinity of interest and measure category experience in the following parts of the research. Recommendations offered the same four songs which are Future & Drankin N Smokin - Lil Uzi Vert, What You Know Bout Love - Pop Smoke, Try Me - Lil Mosey Headshot - Lil Tjay, Polo G & Fivio Foreign with the same order. After participants chose their preferences, they were asked to give a score to their familiarity with the music genre from 1-10 (1-not familiar, 10-very familiar). Similarly, they rated their privacy concerns rose after the given recommendation. Lastly, participants were asked to indicate their gender.
Qualtrics data was exported to SPSS for the analysis. In total, 160 participants participated in the survey. Since 22 of them did not finish the survey, I didn’t include their data and conducted
the analysis with 138 participants. Due to ethical sensitivity, the survey included 4 options for gender by also including “other” and “prefer not to say”. There were three participants who chose either “other” or “prefer not to say”. Therefore, for this specific analysis these three participants were not included. Thus, the analysis was done with a sample of 135 participants.
The aim of providing recommended songs and expecting customers to choose their preference was to design the survey as an experiment and consider each chosen songs as click- throughs. Similarly, the 5th option, "I wouldn't click on any of the songs" was to consider this option as no click-through. To conduct the analysis, I had to simplify the data. Therefore, participants who chose any of the songs were recoded into dichotomous variables, which indicate click-through as Yes, 1. Others who chose the option of "I would not click on any of the recommendations were recoded as No, 0. Four conditions (see Table 1) were also recoded as being present (1) or absent (0) and two columns which include two independent variables (friend &
similarity) were created by using "Recode into different variables '' option is SPSS. As a result, both the predictors and outcomes recoded as dichotomous, categorical variables. Similar process is used for the gender category. Females who were coded 2 were coded as 1 and Males who were coded as 1, coded as 0.
Category experience and privacy were used as moderators. Category experience included 4 similar questions in the survey and participants were asked to rate it out of 10 (1-not familiar, 10-very familiar). None of them were reversed so that no recording was done for it. To measure the consistency of scales, a reliability analysis was conducted. The Cronbach’s Alpha was considered.
As a result, the Category Experience scale consisting of 4 different scales has been found to be
reliable (α=0,931) which means there’s an excellent internal consistency. For privacy, I’ve also done a reliability analysis. Privacy scale consisted of 3 similar questions which none of them were a reverse question. The Cronbach’s Alpha for privacy was also found reliable (α=0,899) After the reliability analysis, to use this data for the analysis, all scales was gathered as one Product/Category experience scale and Privacy Scale respectively, by using Mean Computation.
To measure the results of the data, three analysis models were built. Table 2 shows the models in a detailed way. The main effects of recommendations with similarity and friend framing have been tested in the first analysis model using both Chi-Square and Binary Logistic Regression. In the second model, direct effects of moderators on click-throughs have been added to the first model to see whether gender, category experience, and privacy directly affect recommendation click- throughs regardless of the framing type. For the third model, I aimed to measure the interaction of moderators with different framings and whether they result in any difference in click-throughs or not. Therefore, interaction between each framing type with gender, category experience, and privacy was added to Model 2.
Table 2: Significant Analysis of all variables with different models
Model 1 Model 2 Model 3
Variables Friends Similarity (Main Effect)
Friends, Similarity, Gender, Cat.Exp., Privacy
Friends, Similarity, Gender, Cat.Exp., Privacy + all interactions
Similarity -0,503 -0,521 -0,524
Friend -0,072 -0,055 3,143**
Gender -0,753* -0,863
Category Experience 0,230** 0,636***
Privacy -0,048 -0,158
Similarity x Gender 1,312
Friend x Gender -1,761*
Similarity x Prod. Exp -0,339
Friend x Prod. Exp -0,410*
Similarity x Privacy 0,309
Friend x Privacy -0,211
5.1 Study 1: Impact of friend and similarity framing on click-throughs Chi-Square:
Since two independent variables and the dependent variable are categorical, two chi-tests were conducted to see whether there's an association between each framing type with the result.
SPSS Crosstabs analysis resulted in Chi-Square Tests and Symmetric Measures to see the main relation effect by looking at Phi Coefficient. The expected frequency rate in both results was more than 5. Thus, loss of statistical power is not a case. To measure if friend framing has any influence on click-through rates, I mainly looked at Pearson Chi-Square and Asymptomatic Significance.
The results showed no significant association between friend framing and click-throughs (χ2 (1,135) = ,068a, p=0,795). The Phi coefficient also indicated the same that there's a very weak but negative association between two variables 0,022. The case was similar for the Similarity framing.
It was found that similarity framing has no significant influence on click-throughs (χ2 (1,135) = 2,136a, p=0,144). Also, φ= -0,126 meant that there's a very weak negative influence of Similarity framing on click-throughs. In fact, more than friends framing. To sum up, each of the framing types are independent from the outcome. Even though there's a little negative association, it is not enough to accept the first two hypotheses since Chi-Square is statistically insignificant. In each analysis, Spearman correlation was also checked. For the friend factor it was 0,086 (p<0,05) and for the Similarity factor it was 0,085 (p<0,05) which again showed that there's no correlation.
Binary Logistic Regression
As an addition for Chi Square Test, a binary logistic regression was also applied to test the two hypotheses. Both for friend framing and similarity framing, the results were insignificant with a p-value more than 0,05. Therefore, the first two hypotheses are rejected. Table 3 shows all these findings. The results were surprising since previous studies gave me the reference that similarity and friend framing would have an influence on recommendation click-throughs. However, there were also studies which contradict the idea that similarity and friendship can positively influence people. So, the existing literature was not consensus about the effect of these framing types on recommendation click-throughs. The possible reasons of this result can be participants’ mood at the moment when they encounter with the recommendations. First of all, if they do not have an interest to listen Hip-Hop genre at that moment, it is likely that they would prefer not to click on any of the recommended songs regardless of the framing type. Recommended songs may also have caused the same effect if participants don’t feel in the mood of listening to those songs. Some others
might be searching for more distinctive content and may also prefer to not to click on recommendations of people who are similar to them (Argo and White, 2011). They may also prefer to do this to differentiate themselves from others like Moe & Schweidel (2012) mentioned when they studied people's desire to differentiate themselves while writing reviews. Specifically for friend framing, if participants’ friend network includes the ones, they are very close with, it is highly possible that they already know their friends’ preferences and therefore, show no interest to recommendations (Brown & Reingen; Gai & Klesse, 2019). However, the opposite effect can also be seen if they are not very close to their friend network on the platform, they may find their recommendation as not very important. As the existing research showed evidence both for friend and similarity framings’ possible positive and negative effect as social influence, this study found that friend and similarity framings used in recommendation systems do not result in more click- throughs than no framing. Thus, H1: "Using similarity framing in recommendations has a significant positive effect on click-throughs compared to no framing.", and H2: "Using friend framing in recommendations results in a significant positive impact on click-throughs compared to no framing." are rejected.
Table 3: Logistic Regression
Variable B SE Wald sig Exp(B) 95% CI
Friend -0,072 0,347 0,043 0,836 0,931 0,471-1,929
Similarity -0,503 0,347 2,101 0,147 0,604 0,306-1,194
Constant 0,277 0,297 0,871 0,351 1,320
5.2. Study 2 & Study 3: Impact of gender, cat. experience, privacy on click-throughs
To test the rest of the hypotheses, I used the results of Model 2 and 3. In study 2, as an addition to the main effect, direct effects of moderators used in this study have been measured using Model 2. In other words, the effect of gender, category experience and privacy on click- throughs were found. In Study 3, I used Model 3 and added interaction effects of gender, category experience, and privacy with friend and similarity framing to measure if their interaction results in any change in recommendation click-throughs. For this, I again used a binary logistic regression.
For H3A, B and C, a frequency test was first applied to see the number of male and female participants, since one of them being very high may lead to inaccurate results. Table 4 shows frequency test results for male and female. After finding out that frequencies are not very different from each other (male=56,3%, female=43,3%), to test H3A, Study 2 has been conducted by taking the baseline as males (0). The results stated that there is a significant negative direct effect of gender on recommendation click-throughs (p<0,05), meaning that females click less on recommendations.
Therefore, H3A which was: “Gender does not have a direct effect on recommendation click- throughs.” was rejected. This was an interesting finding since the article by Chang and Chin (2010) stated the opposite. However, one possible reason for that can be a female's increased experience in an online environment compared to the time when the authors made the research. Therefore, it is possible that females may not need recommendations as they used to. It might also be because of trust issues. An article by Suon et al. (2015) stated that online experience increases trust more in males compared to females. Therefore, females who still have trust issues may also find recommendations skeptical and prefer not to click on them. On the contrary Beel et al. (2013) stated
no influence of gender on recommendations. However, the difference of their findings and this study’s findings in results can also be because of other factors like platforms used in the research, time and participants.
For H3B and H3C, I have used the results of Study 3, to see the interaction effect between gender and different framing types on click-throughs. Figure 1 also shows the interaction effect between gender and friend framing. For gender baseline was again taken as males (0) and for framings baseline was taken as absent (0) between absent vs present. Results showed that, there’s a significant interaction effect between gender and friend framings on customer click-throughs (p<0,05). Besides, Beta coefficient result (B=-1,761) with a negative value also tells the predictor variable (female) would less likely result in click-throughs compared to the baseline (male).
Therefore, H3B is accepted with statistical evidence. This also supports Argo and Dahl’s (2020) finding which was about women feeling insecure while shopping with their friends in a retail context. For recommendation systems, women also are less influenced by their friends compared to males.
However, there was no similar relation found between gender and similarity framing since their interaction was insignificant. Since literature also did not provide any specific relation between similarity framing and gender, it is a support for hypothesis 3C.
Table 4: Frequencies of Males and Females
Gender N %
Male 76 56,3%
Female 59 43,7%
Interaction between Friend Framing and Gender
5.2.2. Category Experience
Hypotheses 4A, 4B and 4C were about category experience. Similar with H3A, to measure the direct effect of category experience on click-throughs, model 2 has been used. The results showed a significant positive relationship between category experience and click-throughs. In other words, people click on recommendations more, when they are more experienced with the category.
Therefore, since the hypothesis stated a negative direct effect, H4A is rejected. There might be possible explanations of this result. For instance, some people can be experienced with the category but can be not very interested at the moment when a recommendation is offered. On the other hand, there might be experienced users who are still interested in the products which are in the same category. Therefore, users' mood and interests can influence their decisions on recommendation click-throughs at the moment of recommendation. A study by Kwon et al. (2009) also stated that users are not always stable with their preferences on recommendations. Besides, users with high category experience can find recommended products more familiar and relevant compared to less experienced people (Swaminathan, 2003). Because of this, depending on the recommended products, users can find recommendations qualified and beneficial for a quicker decision-making process (Swaminathan, 2003). Therefore, recommended songs used in the experiment survey may have also impacted highly experienced users' click-throughs if they have positively evaluated the songs.
For hypothesis 4B and 4C, Category Experience used as a moderator in the study.
Therefore, analysis results of Model 3 have been used. Results showed that there's a significant negative effect between category experience and friend framing (p<0,05). This means that, when category experience is higher, click-throughs of recommendation with friend framing decreases.
Therefore, H4B, which was: “Higher category knowledge decreases the effectiveness of friend
framing on click-through” is accepted since a negative effect (β<0) was observed. The results also correspond with my arguments based on existing literature. As it is discussed, one of the main advantages of friend framing compared to no framing and similarity framing is its clear advantage due to more reliable reference. Even if the main effect between friend framing and click-throughs did not show significant evidence, it is not surprising that people are influenced less by their friends when confident about a category. As Ye et al. (2012) also mentioned, people mainly rely on themselves while making decisions and sometimes are influenced by their friends. When they have enough knowledge about the category, they have less needs for their friends. Figure 2 shows the interaction effect of friend framing and category experience. Category experience has three variables as mean - std, mean, and mean + std. Therefore, three lines were created, which also gives the interaction effect.
However, for H4C, no statistical evidence was found for relation between category experience and recommendation with similarity framing. Since existing literature says that the importance of social influence decreases in general when the user is more experienced in a category, a similar result for H4C like H4B was expected. To the author’s knowledge, existing literature also does not provide any information about consumers’ reaction towards other people who are similar to them when they have high or low knowledge within a category. Therefore, it is possible to say that product category knowledge or experience do not influence consumers’
perception towards other people who are similar to them.
Interaction between Friend Framing and Category Experience
For hypothesis 5A, again, findings of Model 2 which showed no significant relationship between privacy and recommendation click-throughs was used. Similar results were found for H5B and H5C since Model 3 did not result in any statistical evidence between interaction of privacy and framing types and their effect on recommendation click-throughs. However, I expected a negative relation in all hypotheses since higher privacy concerns would cause people to have trust issues and therefore, not prefer to click on the recommendations even if they are framed differently.
Therefore, H5A, H5B and H5C are all rejected. One possible explanation of this, can be a privacy paradox. The studies about privacy paradox mainly state that even if people are concerned about
their data in an online environment, they still continue to share their data, usually because they find the benefits of companies more than the cost of giving their data (Barth & Jong, 2017). Therefore, even privacy concerns might be a problem for customized recommendations, the privacy paradox may be the reason behind the insignificance since people still click-through on recommendations regardless of their privacy concern level.
6. Summary & Conclusion
In this study, I have studied the impact of social influence on recommendation click- throughs by dividing them as strong vs. weak ties and framing with friend and similarity factors in recommendation explanations and whether other factors like gender, category experience and privacy influence the recommendation click-throughs in anyway or not. For this, an experiment survey was built by getting an inspiration of a music streaming platform and by randomly assigning different framing types and measuring customers’ gender, experience in the category and privacy concerns. The analysis was done by binary logistic regression by using three different models and results were discussed. According to results, it was found that, using either similarity or friend framing do not show any statistical effect on recommendation click-throughs. Therefore, we can say that, using other groups as reference to benefit from social influence, does not result in any significant effects in recommendation systems like it does in many other online activities as the literature stated. However, when we look at, direct and interaction effect, there are some findings which may have implications. For instance, it is found that both gender and category experience have direct effect on recommendation click-throughs when no similar evidence was found for privacy concern. Females result in less click-throughs compared to men and highly experienced users turned out to click more to recommendation systems compared to less experienced users. An