Recommender systems provide suggestions for products, services, or information that match users' interests and/or needs. However, not all recommendations persuade users to select or use the recommended item. The Elaboration Likelihood Model (ELM) suggests that individuals with low motivation or ability to process the information provided with a recommended item could eventually get persuaded to select/use the item if appropriate peripheral cues enrich the recommendation. The purpose of this research is to investigate the persuasive effect of certain influence strategies and the role of personality in the acceptance of recommendations. In the present study, a movie Recommender System was developed in order to empirically investigate the aforementioned questions applying certain persuasive strategies in the form of textual messages alongside the recommended item. The statistical method of Fuzzy-Set Qualitative Comparative Analysis (fsQCA) was used for data analysis and the results revealed that motivating messages do change users' acceptance of the recommender item but not unconditionally since user's personality differentiates the effect of the persuasive strategies.
Persuasive Technologies utilize several techniques in order to
shape, reinforce or/and change humans’ attitudes and behaviours
without coercion or deception
In information-theoretical terms, persuasion is modeled by the
Elaboration Likelihood Model
Along the above lines, the first objective of this research is to investigate the persuasive effect of the influence strategies proposed by Cialdini (1993), namely Reciprocation, Consistency, Social Proof, Liking, Authority, Scarcity, which are implemented as persuasive messages in the form of recommendations explanations in a movie recommender system developed for the purposes of this study.
Moreover, previous studies
The mainstream of research in Recommender Systems has
traditionally focused on designing and developing accurate
recommendation algorithms
A number of persuasive (or influence) strategies have been
proposed in the literature and can be eventually be utilized in the
design of persuasive messages. For example, Fogg (2002)
describes 42 persuasion strategies and Cialdini (2001) 6 influence
strategies (also known as Six Weapons of Influence) In this study,
we rely upon Cialdini’s influence strategies since they have been
broadly used and verified there are evidences that if influence
strategies are implemented in a system then they increase its
persuasive effect
people who are in positions of authority.
Cialdini (1993) suggested that when a compliance professional (e.g. salesperson) uses the above six influence strategies (Reciprocity, Commitment, Social proof, Scarcity, Liking and Authority) in his/her strategy then (s)he managed to influence more successfully the customer to consume a product/service/information. In the same vein, Kaptein et al. (2012) suggests that applying the influence strategies on text messages people get persuaded to reduce snacking consumption. We adopted Cialdini’s influence strategies because they have already been tested and validated in other domains such as in ecommerce (Kaptein, 2011), use of credit cards (Shu and Cheng, 2012). They also provide a solid framework in order to investigate the persuasive power of messages as peripheral cues in recommender systems. The above leads to following hypothesis of our study: H1: Influence strategies (applied as peripheral cues through messages in recommendations explanations) will have a positive persuasive effect on individuals’ disposition towards the recommended item.
The examination of the above hypothesis will allow us to demonstrate (if validated) that when the preference matching level of the recommended item is low (i.e. when the recommended item is not close to the user’s preferences and interests), then enhancing the recommendation by applying influence strategies in the form of short explanatory messages, the user will be persuaded to use the recommended item, thus changing his/her original negative behavior towards the recommended item to positive intention to use item.
Influence strategies rely upon different psychological principles that may lead to persuasion and therefore it is expected that they will present different degrees of persuasive effect on the recipients of the respective persuasive messages. Thus, the second hypothesis of our research is: H2: Influence strategies lead to different degrees of persuasive effect on individuals’ disposition towards the recommended item.
Kaptein and Eckles (2012) in their study demonstrated that
influence strategies do not always lead to persuasion. They
indicate that in case a consumer receives a message with ‘wrong’
principle then this can bring undesired effects. The above suggests
that there are also other factors that should be taken into
consideration when a persuasive message is used, one of which is
individual’s personality. A human’s personality is defined as ‘a
dynamic organisation, inside the person, of psychophysical
systems that create the persons’ characteristic patterns of
behaviour, thoughts and feelings’
There is a variety of personality taxonomies one of which is Big 5
Dimensions of Personality
The first study that examined message-person congruence effects with a comprehensive model of personality traits is that of Hirsh et al. (2012). Since then message-person congruence effects have been examined in relation to a variety of psychological characteristics (Dijkstra, 2008). Hirsh et al. (2012) demonstrated that persuasive messages are more effective when they are custom-tailored to their interests and concerns. Moreover, Tintarev et al. (2013) demonstrated that people who are characterized from Open to Experience (one of the Big 5 personality traits) tend to prefer diverse recommendations. Additionally, Halko and Kientz (2010) combined persuasive strategies with user’s personality using Big Five Dimensions of Personality and the results of their study revealed relationships between individuals’ personalities and persuasive technologies which means that not all people are affected from the same persuasive means. Finally, Smith et al. (2016) examined the impact of patients personality on Cialdini’s influence strategies in the form of reminders. The research indicated that patient’s with high emotional stability seem to be more responsive to all strategies of persuasion, while patients with low agreeableness rated all Cialdini’s strategies higher than those with high. Finally, the research demonstrated that the reminders of “Authority” and “Liking” are the most popular.
For the execution of the experiment we had first to design the persuasive explanations that would accompany each recommended movie. For this task we followed the methodology proposed by Kaptein et al. (2012). More specifically, a group of three researchers familiar with Persuasive Technology, created thirty (30) textual explanations, i.e. five (5) for each Cialdini’s influence strategies. The content of each explanation was developed in order to comply with the main purpose of each principle in the movie domain. For instance, for the influence strategy of Social Proof, the five possible persuasive explanations that were constructed are: (1) The 85% of this research’s users rated the recommended movie with four (4) or five (5) stars. (2) The recommended movie is on ‘to watch’ list of 85% of this research’s users. (3) Most of the users with the same age and sex as yours, rated the recommended movie with 4 stars! (4) The recommended movie’s video trailer on youtube has more than 550,000 views. (5) The recommended movie’s video trailer on youtube has more than 1600 likes and only 200 dislikes. Seventeen (17) experts in the field of Information Systems and Marketing were invited in order to evaluate each explanation in terms of its compliance with the respective influence strategy. First, a brief presentation of the strategies was given to the evaluators so as to be more familiar with the influence strategies and then they were asked to evaluate the set of persuasive explanations. Each evaluator declared the compliance of each explanation to the respected influence strategy through a 1 to 5 rating scale (from “Completely Disagree” to “Completely Agree”). The persuasive explanation with the highest average was considered as the best-matching explanation for this particular influence strategy.
The six (6) best-matching persuasive explanations (one for each strategy), were chosen for the experiment are the following: Reciprocity: A Facebook friend, who saw the movie that you suggested him/her in past, recommends you this movie. Scarcity: The recommended movie will be available to view from 15/1/2014 to 31/1/2014 on cinemas.
Social Proof: The 87% of users in this survey rated the recommended movie with 4 or 5 stars!
A within subjects experimental design was followed in this research. One of our main concerns in the execution of the experiment was to manage participants’ burden by avoiding extensive exposure to treatments and questionnaires (only the psychographic questionnaire consisted of 44 items) while preserving the validity of the experiment. One option to deal with problem was to expose different groups to different cues (i.e. follow a between subjects design). However, this would significantly reduce the sample size within each group and also taking into account the anticipated low number different personality types represented in each of the groups it would have limited our ability to produce valid statistical results. Thus, we selected the within subjects design.
At the first step of the experiment, a set of 20 movies where presented to participants (with no explanations besides the typical information provided by iMDB, such short description of the story, lead actors etc.), where they were asked to state (by checking the appropriate option) whether they have watched each movie and then provide their ratings (in 1-5 scale). Users were explicitly instructed to provide their intention to watch a movie (for all unwatched movies) in the form of a rating. For the movies they had already watched they provided their actual evaluation. Recommendations were drawn from the set of unwatched movies. The set of 20 movies was randomly selected from a pool of 60 movies from different genres and presented to the participants along with the typical information for each movie (movie’s genre, its plot, and the starring actors). The first criterion for the inclusion of a movie in the pool of 60 movies was its genre (action, drama, romance, etc.). In the pool of 60 movies there were at least three movies from each genre, although most of the movies belong to more than one genre. The second criterion was the popularity of the movie, With the term popular movie is meant a movie with high average rating (above 8.0) from a large amount of users (above 1000 users).. Since popular movies are more likely to collect higher ratings while unpopular ones may not be known to the experimental participants (and therefore attract lower ratings), we included in the sample both popular and unpopular movies according to their iMDB ratings. Although that the number of 20 movies was large enough to ensure that at least some of them wouldn’t have been watched by the participant, the system was designed to select from the pool of 60 movies and present to participants alternative movies in the extreme case that all 20 movies have been actually watched by the user. At the second (recommendation) step of the experiment (see Figure 1), the (unwatched) movie for which the participant has expressed the lowest intention to watch (note that if more than one movie was rated with the lowest score, then the recommended movie was selected randomly from the above set of low-rated movies) was presented to the user exactly as the original presentation but enhanced with persuasive explanations. Selecting to present users with the lowest rated movie, is in alignment with our theoretical ELM foundations, which suggest that when the preference matching of the user with respect to the recommended item is low then the peripheral route will be followed. Moreover, this choice enable us to track more easily any changes in the user’s intention to watch the movie since in computational terms it is much easier to identify changes in intentions from the lower to the higher levels of the 1-5 scale. It must be noted that the rating expresses the users’ intention to watch (or not) the recommended movie is considered in our study as a measure of persuasion (i.e. acceptance of the recommendation), which is operationalized by computing the difference between the original and the final ratings. However, the exact meaning of the “acceptance the recommendation” depends on the business objectives of a site. For example, in some cases (as in e-commerce) the desired behaviour may be to request more information, or to purchase the product and so on.
As mentioned above, the recommended movie was enriched with persuasive explanations, based on Cialdini’s Principles (i.e. the explanations designed in the first part of the experiment) and the participant was asked to assess the recommended movie in order to examine whether (and which) strategies influenced users in order to change their intention to watch the recommended movie. More specifically, the recommended movie was presented with the same set of information as the first step (title, actors, etc.) while participants were asked to declare their intention to watch the recommended movie, taking into consideration one of the 6 persuasive explanations each time, which were presented as a list below the recommended movie. The order of the persuasive explanations was appeared in a random way to each user but there were the same texts for all of them. For that reason the expressions that were used in the persuasive explanations were in a generic form, e.g. the wording ‘the recommended movie’ was used instead of the actual title of the recommended movie and so on.
The absolute difference between the original and the final rating was used to measure the persuasive effect. As the “final” rating with respect to the first hypothesis (examining if there are differences before and after the application of the persuasion strategies) we used the highest rating that users provided (independently of the strategy that corresponds to that rating). For the evaluation of the second hypotheses (examining if there are differences among strategies with respect to their persuasive effect), the rating given by the users’ as evaluation of each strategy was considered as the “final” rating.
At the third and last step of the experiment participants were
asked to complete the psychographic questionnaire that was used
to classify users into the Big 5 personality traits. The Big Five
Inventory- 44 (BFI) was used, constitutes from 44 questions
The experiment participants were invited through posts in University’s Facebook groups (e.g. undergraduate, postgraduate and PhD students) and authors’ personal mailing lists to participate in this research. The invitation message was asking recipients to participate in a research in which they would be asked to rate recommendations provided by an online application as well as to fill in a psychographic questionnaire. The link to access the system was provided and a clear suggestion concerning the anonymity of their participation was included in the message. The invitation did not specify that the research involved movies evaluation. The participants’ average scores for the items measuring the personality types in the 44-item psychographic questionnaire are (the standard deviation is included in the parentheses) Extraversion: 3.34 (0.49), Agreeableness: 3.47 (0.42), Conscientiousness: 3.34 (0.42), Neuroticism: 3.30 (0.48), Openness: 3.24 (0.46). The above descriptives showcase that the sample does not exhibit certain personality types more (or less) than others.
In total 117 users participated in our research. 61 (52%) participants of our sample were males while the rest 56 (48%) were females. Additionally, the 46% of the sample was aged between 18 and 24 years old, the 52% was between 25 and 34 years old and the 2% at the age of 35-44 years old.
This research employs the prescriptions of the fuzzy-set
qualitative comparative analysis methodology (fsQCA) to explore
which personality traits explain the effectiveness of each
persuasion strategy. Opposed to variance-based statistical
methods (e.g. structural equation modelling or partial-least
squares based regression models) in which the independent
variables ‘compete’ with each other to explain one or more
dependent variables, fsQCA treats the hypothesized causal factors
as conditions that may be related to the phenomenon under
investigation either by themselves or in combination with one
another
Fuzzy-set QCA identifies conditions or combinations of conditions that are necessary or sufficient to explain an outcome. In this research, a combination of conditions reflects the personality profile of an individual. Such profile would include specific membership values to each personality trait following the calibration procedure. As such, a value close to 1 in a particular personality trait implies that the individual exhibits this trait. In contrast, membership values close to zero imply that the individual does not exhibit the said personality trait. Necessity of a condition implies that an outcome may not derive without the presence of the condition; nevertheless, the condition alone is not able to produce the outcome. Sufficiency of a condition implies that the condition alone is capable of producing the outcome. In practice, if a solution includes the presence of only one condition (i.e., a solution requires the presence or absence of only one personality trait),, then this condition is sufficient to produce the outcome. To estimate the sufficiency and/ or necessity of hypothesized conditions, fsQCA follows a Boolean minimization process based on truth table analysis. The outcome of this process includes the generic combinations of conditions that are sufficient for the outcome whilst remaining logically true. These are encapsulated in three solutions that differ based on their complexity, named as complex, intermediate, and parsimonious. Of interest is the parsimonious solution, which reduces the causal recipes to the smallest number of conditions possible. This research explores how individuals’ personality traits, in the form of five alternative dimensions, fit with different persuasive strategies. Nevertheless, an individual may not be exclusively categorized under a unique personality trait. Instead, individuals may exhibit elements of multiple traits, which collectively form their personality. Moreover, these personality traits are not fixed within all individuals; a particular persuasive strategy may be perceived as equally appropriate to individuals that exhibit completely dissimilar values on their fundamental personality qualities. As a result, we cannot assume that there is a single, universal, personality profile that explains the impact of a given persuasion strategy, which would call for the application of traditional statistical analysis methods based on regression models, but we need to examine how the different combination of the personality traits interweave in order to explain the suitability of a given persuasion strategy. The modus operandi of fsQCA covers this requirement, thus warrants us to adopt it as our guiding analysis methodology.
The first step of our analysis is involved investigating effect of each influence strategy on individuals’ attitude towards watching a movie that they, initially, were unmotivated to watch. We performed two different comparisons to examine the persuasive effect of the influence strategies. In the first test, we measured the difference between the maximum of the ratings that each user provided for the six influence strategies and the original rating. The t-test results suggested that on average there are significant differences (p< .001) between the original rating and most persuasive (for each user) strategy (original and final ratings average scores: 1.49 and 3.05 respectively with standard deviation 0.50 and 1.23). In the second statistical test, we performed a t-test analysis that compares their initial beliefs and the ones formulated after the application of the strategy. The results suggest that all influence strategies were successful in increasing the likelihood of individuals to watch the movie (Table 1) nevertheless, this increase is marginal in absolute figures.. Moreover, a one-way ANOVA test between the attitude changes of individuals for each influence strategy (see Table 2). The results of this analysis indicate that there are statistical differences among the six strategies at the p<.05 level (F= 14.941, p= .000). To probe for differences between the strategies we performed a Games-Howell Post Hoc Test. Based on these results we accept H1. H2 was evaluated through the application of fsQCA methodology. We used the five personality traits as possible conditions that influence the acceptance of each influence strategy. As a first step, the prescriptions of fsQCA require for calibration of the cases into membership sets. Calibration was performed using the corresponding function provided by fsQCA 2.0 software. The function demands as input three threshold points; a fullmembership value, a non-membership value and a cutoff point. Because the dataset consists of subjective cases, we used cluster analysis following the k-means algorithm (k=3) to calculate the three membership sets. More specifically, high values are correlated with the full-membership set, medium values are correlated with the crossover point set and finally low values are correlates with the non-membership set.
For the independent variables (personality traits) no cluster analysis was conducted due to the fact that the differences among the personality traits’ scores were imperceptibly small. Thus, for this case we calculated the independent variables (personality traits) through frequencies with cut points for 4 equal groups, in SPSS. The percentiles that emerged correspond to the fullmembership set for the high values, the crossover point set for medium values and finally the non-membership set for low values.
The results of fsQCA indicate 3-7 alternative solutions per influence strategy comprising of alternative combinations of the personality traits that lead to high acceptance of each influence strategy. Black circles indicate the required presence of a personality trait in a solution. White circles indicate the required absence of a personality trait from the solution. Blank cells indicate that in that particular solution, the presence or absence of that personality trait is indifferent. Each solution is accompanied by two additional measurements of fitness, which express the ‘predictive power’ of each solution, namely the consistency and coverage indexes. Consistency presents how consistent is the empirical evidence with the outcome which is investigated while coverage estimates the proportion of cases that address the outcome which is under investigation. Table 3 illustrates the results of fsQCA for the Reciprocity influence strategy. The methodology, identified four solutions leading to high influence of an individual by the application of the respective strategy. The results indicate that the absence of even one personality trait is sufficient to individuals in order to be influenced by the Reciprocity strategy The methodology identified 6 alternative paths leading to high acceptance of the Scarcity influence strategy. The majority of paths require two personality traits to be present in an individual’s personality in order to be influenced by Scarcity strategy (Table 4). For example, individuals that are both agreeable and conscious, but do not exhibit traits of neuroticism are likely to be influenced by the scarcity influence strategy (solution 6). The remaining Tables present the different paths, consisting of combinations of personality traits, which lead to high acceptance of the remaining four influence strategies. These tables may be interpreted as an atypical personality profile of individuals (one per produced fsQCA solution) in order to be influenced by each strategy (Table 5 – Table 8). Similar to the previous solutions, each table should be interpreted as a combination of mandatory personality traits (indicated with black circles) coupled with the mandatory absence of one or more personality traits (indicated with white circles). Hence, each solution represents a unique combination of the personality traits that should exist in order to explain the acceptance of a persuasive strategy. 2 0.48 0.31 0.64 0.192 3 0.64 0.192
0.47 0.41 1 0.47 0.41
This research emphasizes on two elements of persuasive/ recommender systems. First, we empirically validate that the application of an influence strategy may indeed positively shift the attitude of an individual towards a specific recommended item. Nevertheless, not all influence strategies have the same persuasive effect. We attribute this deviation to the personality traits of the recommender system users. Hence, the second contribution of this study reflects on the development of personality profiles per influence strategy. Each profile, measured as a combination of personality traits that need to be present or absent from the personality mix, reflects the set of traits that fit most with each influence strategy (i.e., individuals sharing the same profile would indeed be persuaded following the application of the respective strategy). It must be noted that an important issue in utilizing recommendation explanations is that persuasive messages may be perceived as promotional ones and therefore impact users’ trust in the recommender systems. For this reason we used a control variable measuring (in an 1-5 scale) users’ trust in the system, which has shown that no such effect occurred (i.e. no significant differences were found between the trust levels before and after the presentation of the persuasive messages, which was on average 2.96 for the users with low intention to watch the movie and 3.27 for the users with high intention to watch a movie).
In effect, most studies in the field of recommender systems have primarily focused on the algorithmic perspective through the proposition of algorithms that provide recommendations tailored to users’ interests and preferences. In contrast, this study provides insights indicating that the provision of properly selected (i.e. taking into account users’ personality) motivating messages have a persuasive effect on users intention to “use” the recommended item, e.g. to watch a movie.
According to the Elaboration Likelihood Model (ELM), when an
individual has low motivation (or ability) to process a
recommendation then she will not proceed through the central
route of persuasion, i.e. he will not thoroughly assess the quality
of argumentation in order to get persuaded. Instead, if appropriate
peripheral cues are implemented (such as persuasive strategies
applied in the form of messages, as suggested in our study) then
she will eventually be influenced (i.e. motivated) to elaborate the
recommendation following the peripheral route to persuasion.
Such peripheral cues act as extra motivating triggers that
influence a user by “diverting attention, reallocating cognitive
resources, and evoking affective responses and behaviours”
Current recommendation applications typically disregard items with low degrees of fitness with the users’ current interests. The confirmation of the first hypothesis of this study indicate that even for such items, there is strong possibility that they may be favoured by the users if they are presented with the appropriate motivating peripheral cue. Moreover, not all people are influenced from the same persuasive messages. This study provides empirical evidence that there is a relationship between personality and Persuasive Strategies. People with specific combination of personality traits are affected more from particular persuasive messages.
The results of the experiment that was conducted surfaced that
motivating messages are not uniformly applied to all recipients of
recommendations. Users’ personality traits are an important
factor that differentiates the effect of influence strategies applied
as persuasive explanations. More specifically, a person who is
characterized by high extraversion seems to be influenced by all
Six Persuasive Strategies. This is reasonable if we take into
consideration that they enjoy interacting with the environment
whilst such people have the tendency to seek for stimulation
Individuals with high agreeableness are eager to help other people
People with high conscientiousness are dutiful (Major et al.,
2006). Ιn other words, they are careful to fulfil obligations, and
thus when someone helps them they feel obligated so they become
more motivated when a persuasive explanation implementing
“Reciprocity” is presented to them. Despite our expectations,
humans with low conscientiousness changed their intention to
watch the movie influenced by the “Consistency” strategy rather
than humans with high conscientiousness. This may be attributed
to the fact that individuals with high conscientiousness avoid to
take risks because that might make them feel uncertain or cause
unexpected delays to their work
On the other hand, people with high openness tend to be
characterized by creativity, sophistication, and curiosity
The findings of the study must be interpreted taking into account its limitations. The sampling frame (students) and the relatively low sample size restrict the possibility of having an actual representation of the population in the sample in terms of personality types. By extending the experiment, in future research, to a larger sample of users we would also have the opportunity to avoid possible self-selection bias as well as to follow a between subjects design, showing not only more movies to each user but most importantly avoiding the learning effect associated with the presentation of all six strategies to all experiment participants. It must be noted that we tried to control the learning effect bias by showing to users recommendations with persuasive explanations in a random way, i.e. the mix of recommendations representing different persuasive strategies was presented in varying order to each of the participants. It is clear that this study provides insights concerning the movie recommendation domain in which it was applied. The generalization of our findings would be enabled only if this research is extended to other application domains. In our future research plans, besides the extension of our research to other domains (e.g. e-commerce) we aim to investigate additional factors that may influence persuasive communication, as for example the need for cognition, which is a personality variable and reflects people’s intrinsic motivation to engage in and enjoy thinking (Cacioppo and Petty, 1982, p. 116).
The first author acknowledges the financial support of the Department of Management Science and Technology and the third author the financial support of the Research Center of the Athens University of Economics and Business for the presentation of this work.