Knowing me knowing you'-using profiles and social networking to improve recommender systems P Bonhard and m a sasse The internet has not only brought us more information and choice, but has also increased the burden of making a choice. Recommender systems aim to address this problem by providing personalised recommendations in areas such as music films or books. Research on recommender systems has focused on improving the matching algorithms. The research presented in this paper takes a user-centred approach. Since recommendations are usually presented as lists of items, without explanations or justifications, users struggle to find out how appropriate recommendations are for them. Our research has shown that the relationship between advice-seeker and recommender is extremely important, so ways of indicating social closeness and taste overlap are required. We thus suggest that drawing on similarity and familiarity between the user and the persons who have rated the items can aid judgement and decision making. This was tested in an experiment, which carefully controlled familiarity, profile similarity and rating-overlap between the user and those rating items. The results help us understand the decision-making processes in an on line context, and form the basis of a user centred recommender system approach. We suggest that recommender systems can be improved by combining the benefits of social networking applications such as explicit networks of trust -with the matching capabilities of recommender systems. 1 Introduction provide the user with a ranked list of items, based on the Internet users today are confronted with an abundance ratings overlap with other users. they provide little or of information. Whether one is comparing prices for no information about where these choices come from CDs and DVDs, searching for film reviews, or looking for and why some items have been prioritised over others a restaurant for an important date, the available information at hand exceeds the amount of information most users want to consider before making a choice. In These systems aim to emulate the process of effect, choice has become a burden [1 seeking advice from trusted sources, yet users still struggle to convert the data they receive from such systems into meaningful information. When people seek Recommender systems(RSs)have been developed advice from peers or trusted sources in the real world to help people deal with this problem by filtering the there are a variety of interpersonal cues, an interaction available options according to their preferences. history and opportunities to question the advice given Amazon, for instance, recommends items based on our which are missing in an on-line system therefore when purchase and search history, while MovieLens [2] lets us users receive recommendations from an RS without any rate films and generates recommendations based hose ratings. At the heart of each RS is a large cues associated with one from a real person and appears more like a search result items(such as books, films, etc), This is a problem an rs should solve rather than create. Similarly, since people inherently know which of their ratings of those items their friends to trust for a particular recommendation RS should imitate this naturally occurring evaluation Recommendations are derived by calculating and decision-making process. It is therefore necessary imilarity of ratings across different users. Currently, rs to understand how people arrive at their decisions in 84 BT Technology Journal·Wol24No3·July2006
84 BT Technology Journal • Vol 24 No 3 • July 2006 ‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems P Bonhard and M A Sasse The Internet has not only brought us more information and choice, but has also increased the burden of making a choice. Recommender systems aim to address this problem by providing personalised recommendations in areas such as music, films or books. Research on recommender systems has focused on improving the matching algorithms. The research presented in this paper takes a user-centred approach. Since recommendations are usually presented as lists of items, without explanations or justifications, users struggle to find out how appropriate recommendations are for them. Our research has shown that the relationship between advice-seeker and recommender is extremely important, so ways of indicating social closeness and taste overlap are required. We thus suggest that drawing on similarity and familiarity between the user and the persons who have rated the items can aid judgement and decision making. This was tested in an experiment, which carefully controlled familiarity, profile similarity and rating-overlap between the user and those rating items. The results help us understand the decision-making processes in an on-line context, and form the basis of a usercentred recommender system approach. We suggest that recommender systems can be improved by combining the benefits of social networking applications — such as explicit networks of trust — with the matching capabilities of recommender systems. 1. Introduction Internet users today are confronted with an abundance of information. Whether one is comparing prices for CDs and DVDs, searching for film reviews, or looking for a restaurant for an important date, the available information at hand exceeds the amount of information most users want to consider before making a choice. In effect, choice has become a burden [1]. Recommender systems (RSs) have been developed to help people deal with this problem by filtering the available options according to their preferences. Amazon, for instance, recommends items based on our purchase and search history, while MovieLens [2] lets us rate films and generates recommendations based on those ratings. At the heart of each RS is a large database of: • items (such as books, films, etc), • users, • their ratings of those items. Recommendations are derived by calculating similarity of ratings across different users. Currently, RS provide the user with a ranked list of items, based on the ratings overlap with other users. They provide little or no information about where these choices come from, and why some items have been prioritised over others. These systems aim to emulate the process of seeking advice from trusted sources, yet users still struggle to convert the data they receive from such systems into meaningful information. When people seek advice from peers or trusted sources in the real world, there are a variety of interpersonal cues, an interaction history, and opportunities to question the advice given, which are missing in an on-line system. Therefore when users receive recommendations from an RS without any meaningful explanation, the recommendation lacks the cues associated with one from a real person and appears more like a search result. This is a problem an RS should solve rather than create. Similarly, since people inherently know which of their friends to trust for a particular recommendation, RS should imitate this naturally occurring evaluation and decision-making process. It is therefore necessary to understand how people arrive at their decisions in
Knowing me knowing you'-using profiles and social networking to improve recommender systems such domains and which recommender characteristics HCI approaches to rs research have been limited to are important to them examining existing rss in order to establish interaction Human-computer interaction(HCi)approaches RS research have primarily focused on evaluating the When comparing how advice from friends and from user interface of existing RSs [3-5] RS is perceived, Swearingen and Sinha [4, 9] found that while people overall preferred recommendations from Our research, however took a step back and asked their friends, they appreciated the ability of an rs to when users seek recommendations, and what provide serendipitous recommendations that information they hope to receive. We translated our broadened their horizons. In that context, Swearingen findings into a mock-up design of an RS which and Sinha (5, 9] identified two factors as fundamentally incorporates social networking features, and then important in the overall usefulness of an RS-familiar valuated those designs through experiments. This recommendations and system transparency paper first presents a brief review of rs research to date followed by an overview of the psychology Familiar recommendations can come in different literature on advice-seeking and decision making We forms items previously consumed or items that are then consider social networking applications and how related to known items(e. g. books by the same author) they fit into the rs context this is followed by a section Familiar items can generate trust in the system but it presenting the qualitative and quantitative studies can also make the recommendations seem too exploring the integration of Rs and social networking The last section then discusses the implications of this inference leading to a recommendation(and agreeing research for rs design and future research with it)not only increases trust in the recommendation and the system providing it but also makes it more likely that the user will follow the recommendation 2. Background RSs have been deployed contexts, such as book or music shopping, e.g. Amazon conducted an extensive study examining what effect (www.amazon.com),generalratingsites(www.explanationsforcollaborativefilteringresultshaveon ratingzone. com), and specific rating sites such as the user's perception of the system. In testing different Movielens(http:/iMovielens.umnedu explanation interfaces, they found that explanations are important to users, because their own reasoning often RSs aggregate the information received and redirect does not match the inference mechanism of the system it to the appropriate recipients. While many strategies Users were less likely to trust recommendations when for computing recommendations have been explored they did not understand why certain items were such as item-based collaborative filtering [6, Bayesian recommended to them Herlocker et al [ 3] suggest that networks [7 and factor analysis [8, user-user a rating histogram of the user' s closest neighbours is collaborative filtering(CF)comes closest to emulating the most effective wa real-world recommendations because they are based on collaborative filtering y of explaining the results of the user rather than item matching. Recommendations are generated for a given user by comparing the These above studies took an evaluation approach to existing ratings to those of all other users in the improving an existing RS In our view, however, the user database. In doing so, a neighbourhood of similar users requirements for RSs have not been effectively is established, and based on that, rating predictions are investigated. Rather than continuing the current cycle computed for items that users have not yet rated, but of deploying a matching algorithm and seeing how users closest neighbours have respond, we took a step back and examined existing literature on advice-seeking and decision-making 2.1 Recommender systems research strategies, to identify what support users seek during RS research to date has focused on designing these activities algorithms for more effective and efficient computation of rating predictions The former aims to increase the 2.2 Trust research and recommender systems precision of predicting ratings. This is tested through Previous Hcl studies on rs aimed to increase user trust existing rating data sets, where part of the rating set is in these systems by helping users understand how the deleted and the prediction results from algorithms are system calculates recommendations. Related HCI compared against the real ratings. Prediction efficiency research in recent years has investigated trust in is concerned with the computational cost in terms of e Commerce systems [10, social networking virtual time and resources for calculating these predictions communities and recommender systems [11, 12]. Each BT Technology Journal.Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems BT Technology Journal • Vol 24 No 3 • July 2006 85 such domains and which recommender characteristics are important to them. Human-computer interaction (HCI) approaches to RS research have primarily focused on evaluating the user interface of existing RSs [3—5]. Our research, however, took a step back and asked when users seek recommendations, and what information they hope to receive. We translated our findings into a mock-up design of an RS which incorporates social networking features, and then evaluated those designs through experiments. This paper first presents a brief review of RS research to date, followed by an overview of the psychology literature on advice-seeking and decision making. We then consider social networking applications and how they fit into the RS context. This is followed by a section presenting the qualitative and quantitative studies exploring the integration of RS and social networking. The last section then discusses the implications of this research for RS design and future research. 2. Background RSs have been deployed in various eCommerce contexts, such as book or music shopping, e.g. Amazon (www.amazon.com), general rating sites (www. ratingzone.com), and specific rating sites such as MovieLens (http://movielens.umn.edu). RSs aggregate the information received and redirect it to the appropriate recipients. While many strategies for computing recommendations have been explored, such as item-based collaborative filtering [6], Bayesian networks [7] and factor analysis [8], user–user collaborative filtering (CF) comes closest to emulating real-world recommendations because they are based on the user rather than item matching. Recommendations are generated for a given user by comparing their existing ratings to those of all other users in the database. In doing so, a neighbourhood of similar users is established, and based on that, rating predictions are computed for items that users have not yet rated, but closest neighbours have. 2.1 Recommender systems research RS research to date has focused on designing algorithms for more effective and efficient computation of rating predictions. The former aims to increase the precision of predicting ratings. This is tested through existing rating data sets, where part of the rating set is deleted, and the prediction results from algorithms are compared against the real ratings. Prediction efficiency is concerned with the computational cost in terms of time and resources for calculating these predictions. HCI approaches to RS research have been limited to examining existing RSs in order to establish interaction design guidelines [3—5]. When comparing how advice from friends and from RS is perceived, Swearingen and Sinha [4, 9] found that, while people overall preferred recommendations from their friends, they appreciated the ability of an RS to provide serendipitous recommendations that broadened their horizons. In that context, Swearingen and Sinha [5, 9] identified two factors as fundamentally important in the overall usefulness of an RS — familiar recommendations and system transparency. Familiar recommendations can come in different forms — items previously consumed, or items that are related to known items (e.g. books by the same author). Familiar items can generate trust in the system, but it can also make the recommendations seem too simplistic. As for transparency, understanding the inference leading to a recommendation (and agreeing with it) not only increases trust in the recommendation, and the system providing it, but also makes it more likely that the user will follow the recommendation. Using a similar approach, Herlocker et al [3] conducted an extensive study examining what effect explanations for collaborative filtering results have on the user’s perception of the system. In testing different explanation interfaces, they found that explanations are important to users, because their own reasoning often does not match the inference mechanism of the system. Users were less likely to trust recommendations when they did not understand why certain items were recommended to them. Herlocker et al [3] suggest that a rating histogram of the user’s closest neighbours is the most effective way of explaining the results of collaborative filtering. These above studies took an evaluation approach to improving an existing RS. In our view, however, the user requirements for RSs have not been effectively investigated. Rather than continuing the current cycle of deploying a matching algorithm and seeing how users respond, we took a step back and examined existing literature on advice-seeking and decision-making strategies, to identify what support users seek during these activities. 2.2 Trust research and recommender systems Previous HCI studies on RS aimed to increase user trust in these systems by helping users understand how the system calculates recommendations. Related HCI research in recent years has investigated trust in eCommerce systems [10], social networking, virtual communities and recommender systems [11, 12]. Each
Knowing me, knowing you'-using profiles and social networking to improve recommender systems of these aims to address different aspects of the source credibility concept of trust depending on the definition applied advice personalisation, Massa and Avesani [1l] addressed the problem of data sparseness in collaborative filtering -in a huge database, on average two users are unlikely to have rated the same items in a similar fashion To overcome Source credibility refers to the overall impression a site, personalisation to whether the information is this problem, they create a trust graph based on the tailored to the user 's needs, and predictability to degree of connectedness of users, using the distance whether the information presented reflects the users terms of arcs between two users on this graph as a knowledge, and prior experience with this and other measure of trust. This transitive trust graph is then used to increase the number of comparable users sites. However, Briggs et al actually examine infor mation seeking, rather than advice-seeking, and which O'Donovan and Smyth [12] take a different trustworthy elements of a site design make the information appear approach by using past rating reliability to generate trust values that increase or decrease the weight given to predictions of neighbouring users. They divide trust considering established decision-making and advice- into two possible categories- profile level and item level trust. On the profile level they compare all eeking research. For an rs to address users'needs in this domain, however, the requirements of the users common existing ratings of an advice- seeker and a task need to be considered, recommender. and examine whether the recommender would have been able to predict the correct ratings for a given advice-seeker The percentage of correct pre- The next section gives a brief overview of the dictions across all common items is their profile- level evant psychology literature on this issue, and trust value. Item-level trust is more fine-grained and identifies the support requirements which can be only measures the percentages of recommendations for inferred from it. A more detailed description can be one item that were correct found in an earlier paper [14 Each of these approaches addresses a different 2.3 Decision making and advice seeking hallenge for collaborative filtering (CF)algorithms. The main purpose of an RS is to aid its users in their Whereas the Massa and Avesani [11] approach can be decision making by presenting a reduced set of options used to overcome a technical problem in order to in the form of advice or recommendations. To [12] mainly aims to increase accuracy(precision)of such other sources it is instructive to examine the decision a CF algorithm. However, both approaches address a making and advice-seeking literature in psychology partici echnical problem without considerin user's perspective on trust. 2.3.1 Decision making Riegelsberger et al [10] developed a trust model for Classic decision-making theory proposes interactions between two actors in a situation of strategies that decision makers might adopt information or advice. The model divides means of under the compensatory models of decision making signalling trustworthiness into symbols and symptoms, [16] would predict that people make decisions based both of which the technology involved can transmit on the expected utility of an option. Evidence, however Symbols can be e Commerce trust seals or a professional Suggests that decision makers often use various, less precise non-compensatory models. Decision makers be specifically created; rather they are a by-product of often do not make decisions in the rational or logica way proposed by utility theory but rather they are often recommender who repeatedly gives good advice to an influenced by a plethora of non-rational factors of tr who has given good advice to others may become a 2.3.2 Advice seeking symbol of trustworthiness The advice-seeking literature has predominantly focused on objective domains where advice can be Briggs et al [13 identified trust-warranting factors classified as right or wrong [17-20 Generally, advice for on-line advice. They concluded that three factors seeking is seen as a problem of combining and were paramount for users to perceive information as weighting different information sources to come to a trustworthy final conclusion 86 BT Technology Journal·vol24No3·July2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems 86 BT Technology Journal • Vol 24 No 3 • July 2006 of these aims to address different aspects of the concept of trust depending on the definition applied. Massa and Avesani [11] addressed the problem of data sparseness in collaborative filtering — in a huge database, on average two users are unlikely to have rated the same items in a similar fashion. To overcome this problem, they create a trust graph based on the degree of connectedness of users, using the distance in terms of arcs between two users on this graph as a measure of trust. This transitive trust graph is then used to increase the number of comparable users. O’Donovan and Smyth [12] take a different approach by using past rating reliability to generate trust values that increase or decrease the weight given to predictions of neighbouring users. They divide trust into two possible categories — profile level and item level trust. On the profile level they compare all common existing ratings of an advice-seeker and a recommender, and examine whether the recommender would have been able to predict the correct ratings for a given advice-seeker. The percentage of correct predictions across all common items is their profile-level trust value. Item-level trust is more fine-grained and only measures the percentages of recommendations for one item that were correct. Each of these approaches addresses a different challenge for collaborative filtering (CF) algorithms. Whereas the Massa and Avesani [11] approach can be used to overcome a technical problem in order to generate recommendations more quickly, O’Donovan’s [12] mainly aims to increase accuracy (precision) of such a CF algorithm. However, both approaches address a particular technical problem without considering the user’s perspective on trust. Riegelsberger et al [10] developed a trust model for interactions between two actors in a situation of uncertainty and risk, such as an exchange of goods, information or advice. The model divides means of signalling trustworthiness into symbols and symptoms, both of which the technology involved can transmit. Symbols can be eCommerce trust seals or a professional looking Web site. Symptoms, on the other hand, cannot be specifically created; rather they are a by-product of trustworthy actions. In the context of an RS, a recommender who repeatedly gives good advice to an individual shows evidence of trustworthiness. Someone who has given good advice to others may become a symbol of trustworthiness. Briggs et al [13] identified trust-warranting factors for on-line advice. They concluded that three factors were paramount for users to perceive information as trustworthy: • source credibility, • advice personalisation, • predictability. Source credibility refers to the overall impression of a site, personalisation to whether the information is tailored to the user’s needs, and predictability to whether the information presented reflects the user’s knowledge, and prior experience with this and other sites. However, Briggs et al actually examine information seeking, rather than advice-seeking, and which elements of a site design make the information appear trustworthy. In RS research, there has been a significant lack of considering established decision-making and adviceseeking research. For an RS to address users’ needs in this domain, however, the requirements of the user’s task need to be considered. The next section gives a brief overview of the relevant psychology literature on this issue, and identifies the support requirements which can be inferred from it. A more detailed description can be found in an earlier paper [14]. 2.3 Decision making and advice seeking The main purpose of an RS is to aid its users in their decision making by presenting a reduced set of options in the form of advice or recommendations. To understand how people normally deal with advice from other sources, it is instructive to examine the decisionmaking and advice-seeking literature in psychology. 2.3.1 Decision making Classic decision-making theory proposes various strategies that decision makers might adopt when examining options [15, 16]. Utility theory, which falls under the compensatory models of decision making [16], would predict that people make decisions based on the expected utility of an option. Evidence, however, suggests that decision makers often use various, less precise non-compensatory models. Decision makers often do not make decisions in the rational or logical way proposed by utility theory, but rather they are often influenced by a plethora of non-rational factors. 2.3.2 Advice seeking The advice-seeking literature has predominantly focused on objective domains where advice can be classified as right or wrong [17—20]. Generally, advice seeking is seen as a problem of combining and weighting different information sources to come to a final conclusion
Knowing me, knowing you'-using profiles and social networking to improve recommender systems Whereas the benefit of combining opinions in that there is a need for people to create matters of fact is both demonstrable and understood theoretically, the same is not true for matters of taste [18. Simple aggregation of opinions in matters of taste that using a network of existing connections is the raises conceptual difficulties as people are entitled to best way of doing this different opinions on movies, music or restaurants. that making the above easy is a great benefit. Yaniv argues that the most promising strategy to examine these factors in matters of taste would be in While users have various options to find other users considering the personal match between advice-seeker with whom to make connections the main one is still and recommender, assuming that the greater the through the networks of their friends. Other options similarity between them, the greater the impact and usually include some form of database search where benefit of advice received [18]. The argument is interests and demographics can be specified. In therefore not to find optimal solutions, but rather to use networks of potentially thousands of people both these heuristics that simply find good solutions options can be quite cumbersome and do not actually follow the advantages of real-world introductions 2.3.3 Social embedding of recommendations Researchers have pointed out that recommendations In real-world introductions, the person introducing have to be considered in the social context in which they two people usually has a motive for doing so, be it for are delivered [21]. For instance, what kind of occasion is professional or personal reasons. In an on-line social a restaurant recommended for? Is it a business dinner or network, it is easy for users to browse and contact a first date? Who are you going to the cinema with? Are people in their friends,. thus skipping the you buying a CD for yourself or as a present? introduction stage. Further, in the real world, social connections are typically revelatory about the social RSs operate under the assumption that the world status of a person, and seeing someone within the can be described objectively, and that optimal solutions context of their connections is a useful source of to problems can be deduced from these objective information about that person [27] descriptions [22]. Consequently, recommendations om such systems are de-contextualised thus ignoring In an on-line context such as a social networking the situatedness of a recommendation application, the situation is slightly different. Wherea real-world connections carry some significance(because To allow advice-seekers to judge the validity and both parties are tied to their real- world identities an their connection involves some common ground crucial piece of information that should be included in whether personal or professional), the same is not true often be easily altered and thus do not carry the same Situating recommendations in their social context in reliability factor that real-world identities would an on-line environment leads us to examine how people use tools to se on line. Social networkin Thus, an often seen phenomenon in on-line social applications pre networking is that a lot of users sign up and simply grow connect their f a platform for their users to their networks as much as possible. Subsequently, their through their profiles. Seeking network loses value in terms of situating that person recommendations from friends is a naturally occurring within a social circle social process, which has so far been under explored in this context. Section 2. 4 therefore briefly examines social networking application Both the systems and its users have matured in terms of usage and specific usage goals. Through higher broadband Internet proliferation, more users 2.4 Social networking have access to effectively use these communication Social networking applications such as Friendster [23, platforms for a variety of purposes such as dating Facebook [24], hi5[25, and mySpace [26], have grown business networking and sharing of digital content significantly over the past few years. While early systems saw an initially enthusiastic take-up, this was While certain social networking platforms are still followed by an equally quick decline which was mainly completely open(e. g. Friendster), some opt for a semi due to a lack of clearly defined usage goals. Once closed approach(e. g. Facebook ) and others for a closed signed up, users simply did not know what to do with 'by-invitation-only' approach (e. g. Orkut). Open e applications. Donath and boyd point out [27, these networks such as Friendster are able to attract a large networking sites had the three following basic number of users but are also more vulnerable to underlying assumptions BT Technology Journal.Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems BT Technology Journal • Vol 24 No 3 • July 2006 87 Whereas the benefit of combining opinions in matters of fact is both demonstrable and understood theoretically, the same is not true for matters of taste [18]. Simple aggregation of opinions in matters of taste raises conceptual difficulties as people are entitled to different opinions on movies, music or restaurants. Yaniv argues that the most promising strategy to examine these factors in matters of taste would be in considering the personal match between advice-seeker and recommender, assuming that the greater the similarity between them, the greater the impact and benefit of advice received [18]. The argument is therefore not to find optimal solutions, but rather to use heuristics that simply find good solutions. 2.3.3 Social embedding of recommendations Researchers have pointed out that recommendations have to be considered in the social context in which they are delivered [21]. For instance, what kind of occasion is a restaurant recommended for? Is it a business dinner or a first date? Who are you going to the cinema with? Are you buying a CD for yourself or as a present? RSs operate under the assumption that the world can be described objectively, and that optimal solutions to problems can be deduced from these objective descriptions [22]. Consequently, recommendations from such systems are de-contextualised, thus ignoring the situatedness of a recommendation. To allow advice-seekers to judge the validity and appropriateness of a recommendation, situatedness is a crucial piece of information that should be included in an RS. Situating recommendations in their social context in an on-line environment leads us to examine how people use tools to socialise on line. Social networking applications provide a platform for their users to connect their friends through their profiles. Seeking recommendations from friends is a naturally occurring social process, which has so far been under explored in this context. Section 2.4 therefore briefly examines social networking applications. 2.4 Social networking Social networking applications such as Friendster [23], Facebook [24], hi5 [25], and mySpace [26], have grown significantly over the past few years. While early systems saw an initially enthusiastic take-up, this was followed by an equally quick decline, which was mainly due to a lack of clearly defined usage goals. Once signed up, users simply did not know what to do with the applications. Donath and Boyd point out [27], these networking sites had the three following basic underlying assumptions: • that there is a need for people to create connections, • that using a network of existing connections is the best way of doing this, • that making the above easy is a great benefit. While users have various options to find other users with whom to make connections, the main one is still through the networks of their friends. Other options usually include some form of database search where interests and demographics can be specified. In networks of potentially thousands of people both these options can be quite cumbersome and do not actually follow the advantages of real-world introductions. In real-world introductions, the person introducing two people usually has a motive for doing so, be it for professional or personal reasons. In an on-line social network, it is easy for users to browse and contact people in their friends’ networks, thus skipping the introduction stage. Further, in the real world, social connections are typically revelatory about the social status of a person, and seeing someone within the context of their connections is a useful source of information about that person [27]. In an on-line context such as a social networking application, the situation is slightly different. Whereas real-world connections carry some significance (because both parties are tied to their real-world identities and their connection involves some common ground, whether personal or professional), the same is not true for on-line social networking. On-line identities can often be easily altered and thus do not carry the same reliability factor that real-world identities would. Thus, an often seen phenomenon in on-line social networking is that a lot of users sign up and simply grow their networks as much as possible. Subsequently, their network loses value in terms of situating that person within a social circle. Both the systems and its users have matured in terms of usage and specific usage goals. Through higher broadband Internet proliferation, more users have access to effectively use these communication platforms for a variety of purposes such as dating, business networking and sharing of digital content. While certain social networking platforms are still completely open (e.g. Friendster), some opt for a semiclosed approach (e.g. Facebook), and others for a closed ‘by-invitation-only’ approach (e.g. Orkut). Open networks such as Friendster are able to attract a large number of users, but are also more vulnerable to misuse
Knowing me, knowing you'-using profiles and social networking to improve recommender systems Whereas Friendster lets users sign up without an encounter into a connection for which there must be knowing anybody within the system, Orkut only allows some common ground Finding a mutual acquaintance users access if they have been invited by an existing establishes a common ground this includes beliefs and member. The logic of this by- invitation-only strategy is interests, which through the mutual acquaintance are to have a trusted network of identifiable contacts and assumed to be shared to a certain extent [27] thus to stop people from abusing or attacking the system Apart from mutual acqui however. findin Facebook offers a different approach in being a like- minded people is still a very laborious search semi-closed network that only university students/staff/ process in which the user has to browse different alumni can sign up to(verified through their university networks, communities and profiles, and make e-mail extensions, e. g. ucl. ac uk or mit. edu). This semi- individual judgements about whether another person is closed approach carries the advantage that because the actually of any interest user accounts are tied to university e- mail addresses, it harder for users to switch identities by signing up with 2.5 Finding common ground-social different e-mail account recommender systems In addition to that. as a default Facebook limits While ratings are the essential ingredient of any R profile viewing to people within either the sa obtaining them is often difficult, for a number of university or the same geographical location In this way reasons users are primarily encouraged to seek contact within their existing and potential real-life social network based explicitly rating items or adding items to a database around their university this forcefully bridges the gap between people' s real life and on line networks in that there is an increased likelihood that two connected ed. some users quickly need a recommendation as a people actually know each other through university with such a system over a long period of time Users' profiles in these networks are not only about having an rs merely for information retrieval when depicting the users themselves, but also a way of needed is not going to motivate users to either communication, which helps to .. shape the contribute to, or use the system, in the long run representation of others within the system'[28.The context in which this communication takes place There is more to advice-seeking and decision owever,is very particular because(apart from their making than merely being presented with options. this immediate friends)users need to negotiate their information needs to be qualified on different levels profile's representation to an often largely unknown such as: audience. Whereas in real life people have clear distinctions among their public, professional and private how the recommendation was computed ives. in an on-line context this distinction is lost and thus 'people must be prepared to explain both their if done by matching, who it is based performance and that of their friends'[28 information about the recommended item itself, of ourse Social networks are a new form representation and communication, and sul Thus, the usefulness of such systems needs to social behaviour that is different to the real world. For extend that of mere information retrieval or has to be instance, while for some users the main purpose of embedded into another structure that complements the social networking sites has been to amass social uses of it. currency through a huge network of friends a large number of whom they do not actually know in person) Advice-seeking research has so far focused on for others there is a more goal-driven approach (e. g. objective, factual domains as opposed to domains of business in LinkedIn), where accountability is valued taste, with which personalised RSs are predominantly more highly albeit at the cost of reduced privacy concerned. Simple aggregation of opinions for the because people use their real names [27. purpose of decision making might help as an indicator of popularity, but does not contain the same statistical As the main point of social networking sites is benefit of improving judgement accuracy as in factual making new connections, the underlying assumption is domains. Indeed Yaniv [18 points out that a theory in that having a mutual acquaintance or being connected matters of taste is needed Additionally, it needs to be via a chain of acquaintances, provides context for examined whether consulting others'opinions actually connecting. People need a focus to rally around to turn helps decision quality(assuming a more precise 88 BT Technology Journal·vol24No3·July2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems 88 BT Technology Journal • Vol 24 No 3 • July 2006 Whereas Friendster lets users sign up without knowing anybody within the system, Orkut only allows users access if they have been invited by an existing member. The logic of this ‘by-invitation-only’ strategy is to have a trusted network of identifiable contacts and thus to stop people from abusing or attacking the system. Facebook offers a different approach in being a semi-closed network that only university students/staff/ alumni can sign up to (verified through their university e-mail extensions, e.g. ucl.ac.uk or mit.edu). This semiclosed approach carries the advantage that because the user accounts are tied to university e-mail addresses, it is harder for users to switch identities by signing up with a different e-mail account. In addition to that, as a default Facebook limits profile viewing to people within either the same university or the same geographical location. In this way, users are primarily encouraged to seek contact within their existing and potential real-life social network based around their university. This forcefully bridges the gap between people’s real life and on-line networks in that there is an increased likelihood that two connected people actually know each other through university. Users’ profiles in these networks are not only about depicting the users themselves, but also a way of communication, which helps to ‘... shape the representation of others within the system’ [28]. The context in which this communication takes place, however, is very particular because (apart from their immediate friends) users need to negotiate their profile’s representation to an often largely unknown audience. Whereas in real life, people have clear distinctions among their public, professional and private lives, in an on-line context this distinction is lost and thus ‘people must be prepared to explain both their performance and that of their friends’ [28]. Social networks are a new form of selfrepresentation and communication, and subject to a social behaviour that is different to the real world. For instance, while for some users the main purpose of social networking sites has been to amass social currency through a huge network of friends (a large number of whom they do not actually know in person), for others there is a more goal-driven approach (e.g. business in ‘LinkedIn’), where accountability is valued more highly, albeit at the cost of reduced privacy because people use their real names [27]. As the main point of social networking sites is making new connections, the underlying assumption is that having a mutual acquaintance or being connected via a chain of acquaintances, provides context for connecting. People need a focus to rally around to turn an encounter into a connection, for which there must be some common ground. Finding a mutual acquaintance establishes a common ground. This includes beliefs and interests, which through the mutual acquaintance are assumed to be shared to a certain extent [27]. Apart from mutual acquaintances, however, finding like-minded people is still a very laborious search process in which the user has to browse different networks, communities and profiles, and make individual judgements about whether another person is actually of any interest. 2.5 Finding common ground — social recommender systems While ratings are the essential ingredient of any RS, obtaining them is often difficult, for a number of reasons: • explicitly rating items or adding items to a database can be a cumbersome process, • some users quickly need a recommendation as a one-off, and do not see any reason for interacting with such a system over a long period of time, • having an RS merely for information retrieval when needed is not going to motivate users to either contribute to, or use the system, in the long run. There is more to advice-seeking and decision making than merely being presented with options. This information needs to be qualified on different levels, such as: • how the recommendation was computed, • if done by matching, who it is based on, • information about the recommended item itself, of course. Thus, the usefulness of such systems needs to extend that of mere information retrieval, or has to be embedded into another structure that complements the uses of it. Advice-seeking research has so far focused on objective, factual domains as opposed to domains of taste, with which personalised RSs are predominantly concerned. Simple aggregation of opinions for the purpose of decision making might help as an indicator of popularity, but does not contain the same statistical benefit of improving judgement accuracy as in factual domains. Indeed Yaniv [18] points out that a theory in matters of taste is needed. Additionally, it needs to be examined whether consulting others’ opinions actually helps decision quality (assuming a more precise
Knowing me, knowing you'-using profiles and social networking e recommender systems definition of quality) in reducing some of the inherent or CDs to buy, which films to see, in which restaurant to weaknesses: these include the decision maker's failure eat or even how to find a reliable dentist. to generate sufficient alternatives for choice and the tendency to try to confirm rather than challenge prior 3. 1.1 Method views. Yaniv suggests that the best way to study this Twelve semi-structured interviews and five fo domain is to examine the role of a personal match were conducted with a total number of 44 pa between the advice giver and seeker this assumes that with an age range from 21 to 46, including the greater the similarity between them the greater the students as well as professionals. the aim was to elicit benefit from receiving advice concepts and priorities that are important to decision makers when seeking advice, which would then guide This is exactly what collaborative filtering rss do he subsequent focus groups matching users according to their tastes as expressed through item ratings, and basing their recom- 3.1.2 Results mendations on that similarity correlation To analyse the data and model the advice-seeking strategies and processes, we drew on an established widespread forms of getting to know each other is approach from social psychology called grounded theory establishing some form of common ground in our likes 30]. We present the key findings here, summarised in and dislikes. This is a trust-building exercise as it helps Fig l, while readers can refer to our previous paper for a evaluate our counterpart's personality. Indeed, social more detailed account [14 psychology has shown that people like others who are, 3.1.3 Domains of interest among other things, familiar, similar to themselves and with whom they have a history of interaction [(291 Participants clearly differentiated between objective and taste domains. Objective item domains items are Friends from whom we seek recommendations are characterised by measurable and comparable not just a source of information for us - we know their specifications. These are perceived as being neutral to tastes,views and they provide not only different ideas of taste. Items that fall into this category recommendations, but also justification are things like electronic goods, computer hardware and explanations for them. If in doubt we can always software, and cars. Items in taste domains are seen question their recommendations by simply asking about far harder to find advice on. One item may be rated their reasoning and referring back to previous differently by two recommenders, yet both can justi recommendations we might have received from them their ratings. Examples for such items are music, books films and restaurants Seeking and receiving a recommendation is a social activity that often involves the discussion of a particular 3.1.4 Item considerations -risk item. Why did the recommender like it? Would the Regardless of the item domain, a key consideration is recommender want to experiencelbuy it again? Will the the potential risk and the consequences in making a experience change after a while? hoice. At its most basic and common level one of the first considerations is a financial one('lt also comes Our main goal was therefore to explore in more down to the value of the thing ) Participants generally detail the factors that drive people's decision-making said the greater the financial risk involved in a particular improve rS design 9 and how this can be applied to choice, the more detailed and careful their research and advice. seek would be before coming to a decision Risk is associated with the following factors Empirical studies The studies presented in this paper aimed to address bought goods(e. g. books/CDs), that can easily be the above questions from different per rspectives returned bear a lower risk than experience first study was to lay the basis in establishing how consumed goods (e. g. cinema/restaurant visits), people actually seek advice in taste domains. The results would then inform the design of the second when another party is involved the risk is higher (e. g. romantic dinner date) study, which aimed to test the findings from the first study in an rs context. for services (e.g. lawyers, plumbers, dentists), people look for reassurance in a recommendation 3.1 Study 1 -looking for a good restaurant: 3.1.5 Recommender considerations known who to ask versus unknown Since rss aim to give advice in taste domains, we Across all interviews and focus groups, the relationship decided to investigate how people choose what books with the recommender was seen to be extremely BT Technology Journal.Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems BT Technology Journal • Vol 24 No 3 • July 2006 89 definition of quality) in reducing some of the inherent weaknesses; these include the decision maker’s failure to generate sufficient alternatives for choice and the tendency to try to confirm rather than challenge prior views. Yaniv suggests that the best way to study this domain is to examine the role of a personal match between the advice giver and seeker. This assumes that the greater the similarity between them, the greater the benefit from receiving advice. This is exactly what collaborative filtering RSs do — matching users according to their tastes as expressed through item ratings, and basing their recommendations on that similarity correlation. When meeting new people, one the most widespread forms of getting to know each other is establishing some form of common ground in our likes and dislikes. This is a trust-building exercise as it helps evaluate our counterpart’s personality. Indeed, social psychology has shown that people like others who are, among other things, familiar, similar to themselves and with whom they have a history of interaction [29]. Friends from whom we seek recommendations are not just a source of information for us — we know their tastes, views and they provide not only recommendations, but also justification and explanations for them. If in doubt we can always question their recommendations by simply asking about their reasoning and referring back to previous recommendations we might have received from them. Seeking and receiving a recommendation is a social activity that often involves the discussion of a particular item. Why did the recommender like it? Would the recommender want to experience/buy it again? Will the experience change after a while? Our main goal was therefore to explore in more detail the factors that drive people’s decision-making and advice-seeking and how this can be applied to improve RS design. 3. Empirical studies The studies presented in this paper aimed to address the above questions from different perspectives. The first study was to lay the basis in establishing how people actually seek advice in taste domains. The results would then inform the design of the second study, which aimed to test the findings from the first study in an RS context. 3.1 Study 1 — looking for a good restaurant: who to ask Since RSs aim to give advice in taste domains, we decided to investigate how people choose what books or CDs to buy, which films to see, in which restaurant to eat or even how to find a reliable dentist. 3.1.1 Method Twelve semi-structured interviews and five focus groups were conducted with a total number of 44 participants with an age range from 21 to 46, including university students as well as professionals. The aim was to elicit concepts and priorities that are important to decision makers when seeking advice, which would then guide the subsequent focus groups. 3.1.2 Results To analyse the data and model the advice-seeking strategies and processes, we drew on an established approach from social psychology called grounded theory [30]. We present the key findings here, summarised in Fig 1, while readers can refer to our previous paper for a more detailed account [14]. 3.1.3 Domains of interest Participants clearly differentiated between objective and taste domains. Objective item domains items are characterised by measurable and comparable specifications. These are perceived as being neutral to different ideas of taste. Items that fall into this category are things like electronic goods, computer hardware and software, and cars. Items in taste domains are seen as far harder to find advice on. One item may be rated differently by two recommenders, yet both can justify their ratings. Examples for such items are music, books, films and restaurants. 3.1.4 Item considerations — risk Regardless of the item domain, a key consideration is the potential risk and the consequences in making a choice. At its most basic and common level, one of the first considerations is a financial one (‘It also comes down to the value of the thing’). Participants generally said the greater the financial risk involved in a particular choice, the more detailed and careful their research would be before coming to a decision. Risk is associated with the following factors: • bought goods (e.g. books/CDs), that can easily be returned, bear a lower risk than experienced/ consumed goods (e.g. cinema/restaurant visits), • when another party is involved the risk is higher (e.g. romantic dinner date), • for services (e.g. lawyers, plumbers, dentists), people look for reassurance in a recommendation. 3.1.5 Recommender considerations — known versus unknown Across all interviews and focus groups, the relationship with the recommender was seen to be extremely
Knowing me, knowing you'-using profiles and social networking to improve recommender systems 回回回网回 effort recommender decision onsiderations advice seeking Fig 1 Advice-seeking modeL. important when seeking advice in taste domains but know that this source is either very knowledgeable it might seem common sense that people would ce or is known to give good advice this can increase trust their friends for recommendations for cds or in a first-time encounter participants clearly pointed out that the relation to the recommender alone is not sufficient. In addition to 3.2 Study 2-profile similarity conditions has to be fulfilled before an advice-seeker will wanted to examine what effects different recommender trust a recommender: characteristics would have on people' s choices in an RS simulation. More specifically, what combination of either the advice seeker knows that the recom- familiarity profile similarity and rating overlap would mender has similar or the same tastes (taste have an influence on the choices people make in an rS context? Would a visualisation of profile similarity o or both the advice- seeker and recommender have between the decision maker and recomm sufficient mutual knowledge about each others, influence the decision maker's choice as suggested in tastes. so that even with taste differences . the social psychology and our previous study [ 14? Following recommender will be able to predict what the Perugini et als idea of modelling the user [3l and advice seeker will like representing their preferences and interests, we aimed to visualise a recommender in a way that would help the dge the appropriate 3.1.6 Decision process - trust and reliance recommendation We present an overview of the Past experience, source reputation and expertise have a experiment here while a more detailed account can be act on the final judgement of recommendation both from a known or unknown source. They tend to increase or decrease the level of 3.2.1 Method trust in, or reliance on, any given advice In this context, Since every participant would be different in terms of we define trust as faith in a known advisor in a first time demographic data, interests and tastes, we had to context, whereas reliance is based on past experience. create an experiment that would adapt to each Past experience simply means once advice seekers have individual participant, while conceptually remaining received good recommendations, they tend to stick with consistent for everyone To do this, we devised a film a particular recommender both the reputation and the festival scenario where participants receive fictitious expertise of a recommender can increase the trust in a movie recommendations from recommenders (gen first-time encounter. Equally, even if they have not erated on the fly) that were familiar or unfamiliar, similar received any advice from a particular advisor in the past or dissimilar, and either had the same or different film 90 BT Technology Journal. Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems 90 BT Technology Journal • Vol 24 No 3 • July 2006 important when seeking advice in taste domains. While it might seem common sense that people would consult their friends for recommendations for CDs or films, participants clearly pointed out that the relation to the recommender alone is not sufficient. In addition to knowing the recommender, one of two important conditions has to be fulfilled before an advice-seeker will trust a recommender: • either the advice seeker knows that the recommender has similar or the same tastes (taste overlap), • or both the advice-seeker and recommender have sufficient mutual knowledge about each others’ tastes, so that even with taste differences, the recommender will be able to predict what the advice seeker will like. 3.1.6 Decision process — trust and reliance Past experience, source reputation and expertise have a significant impact on the final judgement of a recommendation, both from a known or unknown source. They tend to increase or decrease the level of trust in, or reliance on, any given advice. In this context, we define trust as faith in a known advisor in a first time context, whereas reliance is based on past experience. Past experience simply means once advice seekers have received good recommendations, they tend to stick with a particular recommender. Both the reputation and the expertise of a recommender can increase the trust in a first-time encounter. Equally, even if they have not received any advice from a particular advisor in the past, but know that this source is either very knowledgeable or is known to give good advice, this can increase trust in a first-time encounter. 3.2 Study 2 — profile similarity With the results from the qualitative study in mind, we wanted to examine what effects different recommender characteristics would have on people’s choices in an RS simulation. More specifically, what combination of familiarity, profile similarity and rating overlap would have an influence on the choices people make in an RS context? Would a visualisation of profile similarity between the decision maker and recommender influence the decision maker’s choice as suggested in social psychology and our previous study [14]? Following Perugini et al’s idea of modelling the user [31] and representing their preferences and interests, we aimed to visualise a recommender in a way that would help the decision maker judge the appropriateness of a recommendation. We present an overview of the experiment here while a more detailed account can be found in an earlier paper [32]. 3.2.1 Method Since every participant would be different in terms of demographic data, interests and tastes, we had to create an experiment that would adapt to each individual participant, while conceptually remaining consistent for everyone. To do this, we devised a film festival scenario where participants receive fictitious movie recommendations from recommenders (generated on the fly) that were familiar or unfamiliar, similar or dissimilar, and either had the same or different film Fig 1 Advice-seeking model. A1 A2 A3 A4 A5 A6 advice weighting A1 A2 A3 A4 A5 own influencing factors - own expertise - advisor expertise - advice confirms/ contradicts own opinion receiving advice 2 3 4 objective domain choice experience • cinema/theatre • restaurant consumption • books • CDs risk • financial • other people • consequences known - personally 1) taste overlap 2) mutual knowledge unknown - no personal contact - reviews - experts - people low cognitive effort trust/ reliance 1 decision maker ? 2 3 5 high cognitive effort past experience with source • source reputation • source expertise influencing factors choice taste domain advice seeking advice seeking item considerations recommender considerations decision process
Knowing me, knowing you'-using profiles and social networking e recommender systems tastes(in terms of rating overlap). the individua would see one of four 'buddy profiles and be told characteristics of the recommenders were adapted on whether these buddies had rated those films the the fly to each participant same way or not · Profile similarity 3.2.2 Participants and procedure A total of 100 participants completed the study whicl Profile similarity was based on the demographic lasted 5 minutes, in a computer laboratory. The data, film genre preferences and interests and had varied from 18-44, with a variety of two levels . similar or dissimilar. Thus a simil backgrounds, including students and professionals. ( two years), same profession and have Each participant encountered four phases(see Fig 2 significant overlap in their film genre preferences for an overview). In the first phase each pant was hobbies and interests. Profile similarity was required to provide a basic profile, consisting of visualised through highlighting of the interests demographic data such as age, gender and profession preferences the participant and recommender had but also preferences and interests such as preferred film genres, hobbies, leisure activities and music tastes. In phase 2, participants rated 20-30 films, which would a dissimilar profile on the other hand, would show little or no overlap in terms of interests and serve as a basis for generating their recommendations preferences and significantly differ in terms of demographic data. In phase 3, participants chose from a series of 48 pairs of films recommended by people who -the Rating Overlap scenario made out- had already seen those films Rating overlap was based on the film ratings Phase 4 consisted of a post-study questionnaire participants had previously supplied (phase 2) where participants rated some of the profiles they had Thus recommenders could either have high or low previously seen in terms of familiarity, profile similarity rating overlap, which was visualised in one of two rating overlap and trust, and provided qualitative data vays (consistent for each participant (in text fields)about their decision-reasoning text visualisation (similar to the profile similarity visualisation) showed by explicit highlighting which 3.2.3 Independent variables- recommender films the participant and the recommender had profile characteristics rated in a similar fashion The recommender profile characteristic variables are symbol visualisation showed a Venn diagram listed below with two overlapping circles, a large middle circle representing a large rating overlap and a small middle circle representing a small rating overlap Considering that people consult known sources for Film recommender profiles recommendations, we aimed to simulate familiarity with a recommender through repeated exposure as With these three independent variables, each with suggested in social psychology [29]and the advice. two levels (e. g. familiar versus unfamiliar), this seeking and decision-making literature 33 experiment was a2×2 Exposing participants to a limited number of combinations of these resulted in eight recom- profiles before actually receiving recommendations mender profiles as shown in Fig 3 simulates the process of getting to know that particular profile in relation to one's own. This 3.2.4 Dependent variable-choice variable had two levels, familiar and unfamiliar. In phase 3(see Fig 2 for details), participants saw the Thus during the rating phase of the experiment titles of completely fictitious films and the profile (phase 2, see Fig 2), after each rating participants information of the recommender. The reason for not register rate films choose films create profile with familiarisation with 4 48 pairs of films demographic data 8 different types of erification hobbies/interests recommender BT Technology Journal.Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems BT Technology Journal • Vol 24 No 3 • July 2006 91 tastes (in terms of rating overlap). The individual characteristics of the recommenders were adapted on the fly to each participant. 3.2.2 Participants and procedure A total of 100 participants completed the study, which lasted 30—45 minutes, in a computer laboratory. The age range varied from 18—44, with a variety of backgrounds, including students and professionals. Each participant encountered four phases (see Fig 2 for an overview). In the first phase, each participant was required to provide a basic profile, consisting of demographic data such as age, gender and profession, but also preferences and interests such as preferred film genres, hobbies, leisure activities and music tastes. In phase 2, participants rated 20—30 films, which would serve as a basis for generating their recommendations. In phase 3, participants chose from a series of 48 pairs of films recommended by people who — the scenario made out — had already seen those films. Phase 4 consisted of a post-study questionnaire, where participants rated some of the profiles they had previously seen in terms of familiarity, profile similarity, rating overlap and trust, and provided qualitative data (in text fields) about their decision-reasoning. 3.2.3 Independent variables — recommender profile characteristics The recommender profile characteristic variables are listed below. • Familiarity Considering that people consult known sources for recommendations, we aimed to simulate familiarity with a recommender through repeated exposure as suggested in social psychology [29] and the adviceseeking and decision-making literature [33]. Exposing participants to a limited number of profiles before actually receiving recommendations simulates the process of getting to know that particular profile in relation to one’s own. This variable had two levels, familiar and unfamiliar. Thus during the rating phase of the experiment (phase 2, see Fig 2), after each rating, participants would see one of four ‘buddy’ profiles and be told whether these buddies had rated those films the same way or not. • Profile similarity Profile similarity was based on the demographic data, film genre preferences and interests and had two levels, similar or dissimilar. Thus a similar profile would be the same gender, similar age (± two years), same profession and have a significant overlap in their film genre preferences, hobbies and interests. Profile similarity was visualised through highlighting of the interests or preferences the participant and recommender had in common. A dissimilar profile, on the other hand, would show little or no overlap in terms of interests and preferences and significantly differ in terms of demographic data. • Rating Overlap Rating overlap was based on the film ratings participants had previously supplied (phase 2). Thus recommenders could either have high or low rating overlap, which was visualised in one of two ways (consistent for each participant): — text visualisation (similar to the profile similarity visualisation) showed by explicit highlighting which films the participant and the recommender had rated in a similar fashion, — symbol visualisation showed a Venn diagram with two overlapping circles, a large middle circle representing a large rating overlap and a small middle circle representing a small rating overlap. • Film recommender profiles With these three independent variables, each with two levels (e.g. familiar versus unfamiliar), this experiment was a 2 × 2 × 2 design. All possible combinations of these resulted in eight recommender profiles as shown in Fig 3. 3.2.4 Dependent variable — choice In phase 3 (see Fig 2 for details), participants saw the titles of completely fictitious films and the profile information of the recommender. The reason for not Fig 2 Experiment overview. register – create profile with demographic data hobbies/interests rate films – familiarisation with 4 specific profiles choose films – 48 pairs of films – 8 different types of recommender post-study questionnaire – variable verification – profile judging instructions instructions instructions 1 2 3 4
profile 2 profile 3 familia familiar similar profile similar profile similar profile similar profile high rating overlap high rating overlap low rating overlap low rating overlap profile 5 profile 6 profile 7 profile 8 unfamiliar dissimilar profile dissimilar profile dissimilar profile dissimilar profile high rating overlap high rating overlap-low rating overlap low rating overlap Fig 3 Recommender profiles providing any fictitious reviews or synopses was to →- profile choser inimise noise through a possible participant bias towards any film based on the films properties. Each recommendation pair was a forced choice, thus the dependent variable was which film the participant 1。6号 would choose and in turn which recommender they would trust (see the screenshot in Fig 4) 0.4 OMeM.mt出 serbia Eagles wing Oscar winner. 34. Film Bult.24 profile number Fig 5 Profile choice ratios/trust rating significant influence of profile similarity and rating How cmntldest are you b te reo Figure 6 shows the above trend mapped out with the individual profile characteristics Fig 4 Film recommendation Overall profile similarity and high rating overlap had In addition to the actual choice we recorded a an effect on the choice of which recommender to trust confidence rating on a likert scale from l to 5 for each choice Participants rated the same profiles in terms of trust phase 4 of the experiment. Further. we examined the ratio of the number of times a particular profile was chosen and the number of Interestingly, the trend for the trust ratings is the times a profile was seen(choice ratio same as the one for the profiles chosen, except for profiles 7 and 8. Here participants trusted the familiar 3.2.5 Results -overall profiles chosen profile more, whereas they chose the unfamiliar one more often Figure 5 shows the overall trend of choice ratio, as wel as how these profiles were rated in the end in terms of trust 3.2.6 Individual variable analysis Since in each condition recommenders differed in one Trust was elicited through a Likert scale(1-5)in characteristic, while the other characteristics were kept phase 4 in the post-study questionnaire It is interesting constant it is worth examining the independent to note the trend shown in the graph in Fig 5. When variables individually within each condition type. the mapped against profile characteristics, it points to a results are summarised below and detailed in Fig 7 BT Technology Journal. Vol 24 No 3 July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems 92 BT Technology Journal • Vol 24 No 3 • July 2006 providing any fictitious reviews or synopses was to minimise noise through a possible participant bias towards any film based on the film’s properties. Each recommendation pair was a forced choice, thus the dependent variable was which film the participant would choose and in turn which recommender they would trust (see the screenshot in Fig 4). In addition to the actual choice, we recorded a confidence rating on a Likert scale from 1 to 5 for each choice. Further, we examined the ratio of the number of times a particular profile was chosen and the number of times a profile was seen (choice ratio). 3.2.5 Results — overall profiles chosen Figure 5 shows the overall trend of choice ratio, as well as how these profiles were rated in the end in terms of trust. Trust was elicited through a Likert scale (1—5) in phase 4 in the post-study questionnaire. It is interesting to note the trend shown in the graph in Fig 5. When mapped against profile characteristics, it points to a significant influence of profile similarity and rating overlap. Figure 6 shows the above trend mapped out with the individual profile characteristics. Overall profile similarity and high rating overlap had an effect on the choice of which recommender to trust. Participants rated the same profiles in terms of trust in phase 4 of the experiment. Interestingly, the trend for the trust ratings is the same as the one for the profiles chosen, except for profiles 7 and 8. Here participants trusted the familiar profile more, whereas they chose the unfamiliar one more often. 3.2.6 Individual variable analysis Since in each condition recommenders differed in one characteristic, while the other characteristics were kept constant, it is worth examining the independent variables individually within each condition type. The results are summarised below and detailed in Fig 7. Fig 4 Film recommendation. Fig 3 Recommender profiles. profile 1 – familiar – similar profile – high rating overlap profile 2 – unfamiliar – similar profile – high rating overlap profile 3 – familiar – similar profile – low rating overlap profile 4 – unfamiliar – similar profile – low rating overlap profile 5 – familiar – dissimilar profile – high rating overlap profile 6 – unfamiliar – dissimilar profile – high rating overlap profile 7 – familiar – dissimilar profile – low rating overlap profile 8 – unfamiliar – dissimilar profile – low rating overlap 0 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 1 2 5 3 4 6 8 7 trust + profile chosen ratio profile number profile chosen trust Fig 5 Profile choice ratios/trust rating
Knowing me, knowing you'-using profiles and social networking to improve recommender systems profile 1 profile 2 file 5 rofile 3 iliat similar profile similar profile dissimilar profile similar profile high rating overlap high rating overla high rating overlap low rating overlap profile 4 profile 6 profile 8 profile 7 unfamiliar unfamiliar similar profile dissimilar profile dissimilar profile dissimilar profile low rating overlap high rating overlap low rating overlap low rating overlap Fig 6 Chosen profiles with ratios and characteristics familiarity conditions similarity conditions rating overlap conditions profile 1 profile 1 5575%|vs44.25% 58375%V1625% 15.5% unfamiliar dissimilar high rating overlap profile 7 6875%6512.5% 88.5% 1.5 familiar dissimilar high rating overlap low rating overlap 35675%s43.25% 784.75%s15.25% 11855% 14.5% familiar unfamiliar similar dissimilar high rating overlap low rating overlap 48 875%vs51.25% 8 75%|vs 525% 815%vs 18:5% familiar dissimilar igh rating overlap Fig 7 Profiles chosen- individual percentages per condition · Familiarity 3.2.7 Choice confidence In this experiment familiarity did not prove to be a For every choice, participants provided a confidence significant influence on participants'choices. rating on a Likert scale from 1 to 5. We considered a Overall only condition 3 turned out to be significant deviation of the mean from 3(p <0.05)as statistically significant examinable. With the exception of condition 10, rating means significantly deviated from 3 (see Fig 8) · Profile similarity Profile similarity had a significant impact on the. Familiarity conditions recommender choice he only condition where participants were Rating overlap confident in their choices was condition 1. where Rating overlap also had a significant impact on the the choice was between a familiar and unfamiliar recommender choice. Thus participants generally recommender that was similar and had high rating chose films from recommenders with whom they had a high rating overlap Examining the different visualisations of rating overlap, text and symbol, it Profile similarity conditions is worth noting that there was a higher tendency to In those conditions it was interesting to observe choose films from recommenders with high rating that participants were more confident in their overlap in the symbol conditions (87.6 %)rather choices in conditions 5 and 7 where recommenders than in the text conditions(82.5 %) had a high rating over lap than in conditions 6 and BT Technology Journal.Vol 24 No 3. July 2006
‘Knowing me, knowing you’ — using profiles and social networking to improve recommender systems BT Technology Journal • Vol 24 No 3 • July 2006 93 • Familiarity In this experiment familiarity did not prove to be a significant influence on participants’ choices. Overall only condition 3 turned out to be statistically significant. • Profile similarity Profile similarity had a significant impact on the recommender choice. • Rating overlap Rating overlap also had a significant impact on the recommender choice. Thus participants generally chose films from recommenders with whom they had a high rating overlap. Examining the different visualisations of rating overlap, text and symbol, it is worth noting that there was a higher tendency to choose films from recommenders with high rating overlap in the symbol conditions (87.6 %) rather than in the text conditions (82.5 %). 3.2.7 Choice confidence For every choice, participants provided a confidence rating on a Likert scale from 1 to 5. We considered a significant deviation of the mean from 3 (p <0.05) as examinable. With the exception of condition 10, all rating means significantly deviated from 3 (see Fig 8). • Familiarity conditions The only condition where participants were confident in their choices was condition 1, where the choice was between a familiar and unfamiliar recommender that was similar and had high rating overlap with the participant. • Profile similarity conditions In those conditions it was interesting to observe that participants were more confident in their choices in conditions 5 and 7 where recommenders had a high rating overlap, than in conditions 6 and Fig 6 Chosen profiles with ratios and characteristics. Fig 7 Profiles chosen — individual percentages per condition. profile 1 familiar similar profile high rating overlap profile 2 unfamiliar similar profile high rating overlap profile 5 familiar dissimilar profile high rating overlap profile 3 familiar similar profile low rating overlap profile 4 unfamiliar similar profile low rating overlap profile 6 unfamiliar dissimilar profile high rating overlap profile 8 unfamiliar dissimilar profile low rating overlap profile 7 familiar dissimilar profile low rating overlap profile 1 55.75% familiar profile 2 44.25% unfamiliar profile 3 50% familiar profile 5 56.75% familiar profile 7 48.75% familiar profile 4 50% unfamiliar profile 6 43.25% unfamiliar profile 8 51.25% unfamiliar profile 1 83.75% similar profile 5 16.25% dissimilar profile 3 87.5% similar profile 2 84.75% similar profile 4 84.75% similar profile 7 12.5% dissimilar profile 6 15.25% dissimilar profile 8 15.25% dissimilar profile 1 84.5% high rating overlap profile 3 15.5% low rating overlap profile 5 88.5% high rating overlap profile 2 85.5% high rating overlap profile 6 81.5% high rating overlap profile 7 11.5% low rating overlap profile 4 14.5% low rating overlap profile 8 18.5% low rating overlap vs vs vs vs vs vs vs vs vs vs vs vs 1 2 3 4 5 6 7 8 9 10 11 12 familiarity conditions similarity conditions rating overlap conditions