Pseudo-eating behavior of service robot to improve the trustworthiness of product recommendations

Service robots are used in a range of service situations, including those where they are required to recommend products. However, when robots use subjective statements for product recommendations, users may not place sufficient trust in the robot's recommendations which may degrade its recommendation capability. This study proposes a method to improve the perceived trustworthiness of the robot's subjective statements and thus promote the user's food purchase behavior. The proposed recommendation method involves explicitly sharing the robot's experience with the user and improving the perceived trustworthiness of the robot's subjective statements by performing a pseudo-eating behavior. The results of two online surveys show that the robot's statements and recommendations accompanied by the pseudo-eating behavior improved the trust in the robot's subjective statement and consequently, the user's willingness to purchase products. Further, the results of a field experiment in a real bakery store showed that the proposed method is comparable to recommendation methods proposed in previous works. The results obtained through the online surveys and the field experiment indicate that robots are trusted even when they express subjective statements for recommendations by explicitly sharing the experience with the user. GRAPHICAL ABSTRACT


Introduction
We come across product recommendations while daily shopping.It is generally accepted that product recommendations help users make the purchase decision concerning the recommended product, and various recommendation methods have been investigated [1].Among these, there are many studies on product recommendations using service robots.For example, there are studies where the sales have been improved by simple recommendations from robots [2][3][4], robot recommendations based on user preferences [5], recommendations by the cooperation of multiple robots [6], recommendation methods for long-term effects [7], and recommendations by robots collaborating with human clerks [8].
When a robot recommends a product, there have been cases where the product recommendation is accompanied by the robot's subjective statements, such as 'This product is delicious' or 'That dress looks good on you' (e.g.[5]).On the other hand, it has been pointed out that the subjective statement expressed by the robot causes problems: the subjective statement is difficult to attribute to the robot depending on the type of statement [9].Subjective statements that are not attributed to the robot do not increase the user's willingness to interact with the CONTACT Yuki Okafuji okafuji_yuki_xd@cyberagent.co.jp robot because the user does not feel a sense of reality [10], and the lack of reality does not increase the user's trust in the robot's statement [11,12].In general, trustworthy information has been shown to promote the user's purchase behavior [13] and to also have a strong influence on the robot's recommendation capability [14,15].Thus, the range of products that can be recommended by service robots may narrow down, owing to reduced trust in the robot's subjective statement.If the trustworthiness of the robot's subjective statement could be improved for all types of products, the recommendation target would not be restricted.Then, the robot's recommendations can be expected to promote the user's purchase behavior.Therefore, in this study, we focus on the perceived trustworthiness and recommendation capability when the service robot recommends food accompanied by its subjective statements.An example of a food recommendation with subjective statements is 'This product is delicious, so I recommend it to you.'However, since the robot cannot feel the deliciousness of the product, it is difficult for users to feel the reality of 'the delicious taste' and the trustworthiness of the statement may decrease.For this reason, in previous studies on food recommendation by the service robot, the robot only provides objective information such as the time when bread is freshly baked [2] and the degree of popularity of products [7], and the robots avoid expressing their subjective statements.Here, we propose a recommendation method where the robot gives a subjective statement combined with a pseudoeating behavior.The pseudo-eating behavior, as shown in Figure 1, consists of bringing a product attached to the robot's hand close to its mouth, making an onomatopoeic sound, and saying 'delicious'.It is well known that the tasting of food by the customer promotes purchasing behavior [16], but here we intend to increase the trust in the robot's subjective statement by the pseudoeating behavior of the robot.The lack of reality in the statement is one of the reasons why the trustworthiness of recommendations does not improve [11,12].The pseudo-eating behavior is expected to improve the perceived trustworthiness of the statement because it explicitly shares a pseudo-eating experience of a robot with the user.In addition, a previous study [17] has shown that the relationship between a user and a robot can be strengthened by giving the robot a simulated meal.Although this previous study is for robot owners, it may be equally relevant for the third-party users who are recommended products by service robots.By sharing the robot's pseudo-eating experience with the user, the trust in the robot's subjective statement may be improved; then, the recommendation capability may be enhanced.
In this study, we verified the effectiveness of the proposed method through two online surveys and one field experiment.The first online survey investigated the effect of the robot's pseudo-eating behavior on subjective statements independent of recommendations.The second online survey investigated the effect of the pseudo-eating behavior on product recommendations.These surveys are described in Section 2. Finally, we conducted a field experiment to recommend products in a real bakery store and verify the effectiveness of the proposed method, described in Section 3.

Online surveys
In this section, two online surveys were conducted to verify the influence of the robot's pseudo-eating behavior on the effect of its subjective statements on users.The first online survey investigated the influence of subjective statement with pseudo-eating behavior on the impression of the robot and user experience by focusing on the degree of shared experience, independent of recommendations.The second online survey investigated the influence of subjective statement with pseudo-eating behavior on the effectiveness of the product recommendations as compared with other recommendation methods.
A video was created for online surveys using the humanoid robot 'Sota', which is manufactured by Vstone Co., Ltd.We used Sota in this experiment because it has been used in several previous studies on product recommendation [2,3,[6][7][8].The robot makes subjective statements about bread because the field experiment was conducted in a real bakery store.Questionnaire evaluation is often used for assessing the impression of robots [18], and there are also some studies that report on the use of video questionnaires for such evaluations [19][20][21].
In particular, a previous study [21] evaluates the influence of robots on users' purchasing behavior by using a video questionnaire.Therefore, we consider the evaluation of purchase behavior using the results of online surveys reliable to a certain degree.
All the participants in these online surveys were recruited through the crowd-sourcing service Lancers, and the online surveys were conducted using Google Forms.The online questionnaires took approximately 5 minutes to complete, and the participants were paid 55 JPY as a compensation.

Experimental overview
In this section, we aim to investigate the influence of subjective statements combined with pseudo-eating behavior on the impression of the robot and user experience by focusing on the degree of the shared experience.We prepared three types of videos in which the robot expressed subjective statements about bread, and we evaluated them through an online survey.

Experimental conditions
In this experiment, we set up three experimental situations: 'Without Experience,' 'Explicit Experience,' and 'Eating Behavior.'It has been shown that the trustworthiness of the statement is enhanced by providing reality in specific ways [11,12], and personal narratives are important as a way to enhance reality [22].Therefore, this study also focuses on sharing the robot's experiences to improve the reliability of the robot's statement regarding product recommendations, and we propose three types of conditions: Without Experience shows no specificity, Explicit Experience may present a weak specific experience, and Eating Behavior may give a strong specific experience because the experience is displayed in front of the user.These conditions were compared by focusing on the amount of shared experience.The assumption was that pseudo-eating behavior improves the trustworthiness of statements by strongly sharing the experience.
In the Without Experience condition as the baseline, the appearance of the robot is as shown in Figure 2(a).The robot suddenly says, 'Croissant is delicious,' and then presents its impressions of the taste and texture for about 20 seconds.Since this condition does not present a specific experience, it is assumed that improving the trustworthiness of its subjective statements is difficult.
In the Explicit Experience condition, the appearance of the robot is the same as in the Without Experience condition, as shown in Figure 2(a).The robot says 'I'll tell you about the croissant I ate a while ago.' before the statements in the Without Experience condition.Stating the robot's experience explicitly provides a sense of trustworthiness by presenting a weak concrete experience.
Finally, in the Eating Behavior condition, the appearance of the robot is as shown in Figure 2(b).The robot in this condition has bread in both hands to perform the pseudo-eating behavior.The pseudo-eating behavior was bringing the bread in both hands closer to its mouth and making an onomatopoeic sound 'GABU', which is a familiar onomatopoeia to the Japanese, to simulate eating.Two pseudo-eating behaviors were inserted between the statements of the Without Experience condition.For example, the robot said, 'Croissant is delicious!','GABU,' and 'The sweetness of the butter fills my mouth.'Although eating behavior is known to be a pseudo-behavior, it is assumed that the strong explicit experience provides a strong sense of trustworthiness to the subjective statement.
The robot's subjective statements are the same in all conditions, but the presence or absence of the pseudoeating behavior and the initial statement differ between conditions.

Measurement
In this study, the results of the surveys were tested using Scheffe's paired comparison method [23] that evaluates which of the two presented conditions out of three is preferred by the participants.The participants were randomly assigned to one of six combinations and asked to complete a questionnaire about their impressions of the robot and user experience after watching two videos.
To evaluate robot impressions and user experience, the questionnaire items refer to the previous study [8,24].In terms of the impressions of the robot, the participants were asked about seven items: Trustworthy, Uncomfortable, Intelligent, Human-like, Incredible, Influential, and Friendly.The main aim of Survey I, whose hypothesis is described in more detail in the following Section 2.1.5, is to assess the trustworthiness of the robot's statement.Therefore, in addition to trustworthiness, we used other evaluation items that are expected to influence trustworthiness.In order to use the pseudo behavior in this study, two new items, 'Uncomfortable' and 'Incredible', were added to account for the possible negative impact of pseudo behavior.Credibility is considered one of the criteria for trustworthiness [25] and comfort has also been shown to improve trustworthiness [26].We consider that user comfort is more likely to influence trustworthiness since this experiment is about food.We believe that these negatively inverted items allow for a more multifaceted evaluation of trustworthiness.
In terms of the user experience, they were asked about five items: Interested, Understandable, Obtrusive, Looks Delicious, and Hope to Eat.We included the user experience items in the questionnaire because they are considered to influence purchase behavior.In particular, 'Looks Delicious' and 'Hope to Eat' are used as indicators of the robot's recommendation capability in the online survey.The other items were added as items that influence the process leading to purchase behavior because service robots are often ignored by users [27,28].While these items are particularly important in Survey II, high trustworthiness promotes purchase behavior [13][14][15]; thus, trustworthiness can be indirectly assessed using these items in Survey I.In the two videos presented (A and B), these items were rated on a five-point scale (strongly true for A, truer for A, similar, truer for B, and strongly true for B).In addition to these, we prepared some dummy questions to improve the reliability of answers.If the participants answered the dummy question, their questionnaire was excluded from the analysis.

Participants
A total of 514 answers were obtained through crowdsourcing.Of these, 340 responses answered none of the dummy questions, resulting in a valid response rate of 66.1%.Since it was necessary to equalize the number of responses for each combination to use Scheffe's method, the earliest responses were extracted, and a total of 228 answers (male respondents: 146, female respondents: 81, no answer: (1) were used for the analysis.It is assumed that most of the participants' backgrounds are Japanese because this survey was conducted in Japanese.

Hypothesis
Previous studies have shown that providing reality such as sharing personal experience increases the trustworthiness of statements (e.g.[11,12]).Therefore, we assume that a strong sense of the trustworthiness of the robot's statement in the Eating Behavior condition is promoted because the explicit experience accompanied by the concrete behavior is shared with the user.In addition, we assume that even in the Explicit Experience condition, since the experience is clarified verbally, it provides a sense of the trustworthiness of the robot's statement.We, therefore, propose the following hypotheses in this study.
Hypothesis 1: The trustworthiness of a robot's subjective statement is improved by shared experience.In particular, the pseudo-eating behavior of the robot improves the trustworthiness more than sharing the experience verbally.

Results
Figure 3 shows the average preference between each condition and the results of the analysis of variance (ANOVA).The ANOVA was performed with a significance level of 5% for each item.The results reveal the significant differences in the items: 'Trustworthy, Humanlike, Influential, and Friendly' of the impression of the robot and 'Interested, Understandable, Obtrusive, Looks Delicious, and Hope to Eat' as the user experience.
Next, the post-hoc test with a yardstick was performed on the items with significant differences.The results show that the Eating Behavior condition was higher in 'Trustworthy, Human-like, Influential, and Friendly' and 'Interested, Understandable, Obtrusive, Looks Delicious, and Hope to Eat' as compared to the other two conditions.On the other hand, there are no items that show a significant difference between the Without and Explicit Experience conditions.

Discussion
The results show that the Eating Behavior condition has more positive results in many items than the other conditions.In particular, the results partly support Hypothesis 1 because the robot can increase the trustworthiness of its subjective statement by performing the pseudoeating behavior.However, the trustworthiness between the Without and Explicit Experience conditions shows no significant differences.This is because the pseudoeating behavior made the robot feel more human-like and friendly.These factors influence the trustworthiness of the statement [29].Therefore, it is possible that simply sharing experiences may not be enough to increase the trustworthiness of the robot's subjective statements.
On the other hand, although the pseudo-eating behavior may be a type of deception, there were no significant differences between the items on Uncomfortable and Incredible across all conditions.This also shows that the pseudo-eating behavior does not negatively affect trustworthiness.
In addition, it is known that increasing trust in robots increases their influence [14,15].Therefore, the results of this study imply that the robot's influence has improved; then, user experiences that are directly related to purchase behavior, such as Looking Delicious and Hope to Eat, have improved in the Eating Behavior condition.These results also indirectly support Hypothesis 1.
These results show that the pseudo-eating behavior of robots, which shares experiences with users, improves the perceived trustworthiness of the robot's subjective statements and may be applied as a recommendation behavior.

Experimental overview
In this section, we aim to investigate the influence of subjective statements, together with the pseudo-eating behavior of the robot, on product recommendations as compared to other recommendation methods.We prepared three types of videos in which the robot recommended bread and evaluated them through an online survey.
Since the measurement is the same as for online survey I, the description is omitted in this section.

Experimental conditions
In this experiment, we set up three experimental conditions: 'Eating Recommendation,' 'Expert Recommendation,' which is an existing recommendation method, and 'Combination Recommendation,' which combines Eating and Expert Recommendations.
The Eating Recommendation is a condition for recommending bread with subjective statements accompanied by the pseudo-eating behavior shown in the Eating Behavior condition in the previous section.The appearance of the robot is as shown in Figure 2(b).In the Eating Recommendation condition, in addition to the statements ('Croissant is delicious!','GABU,' and 'The sweetness of the butter fills my mouth.')shown in the previous section, the robot says 'I recommend it, you should try it!' at the end of the statement for recommendation.
The Expert Recommendation is an existing method to improve the persuasiveness of the robot's statements by showing expertize [30].The appearance of the robot is as shown in Figure 2(a).Expert Recommendation is previously applied to product recommendation and the robot provides objective statements such as the degree of popularity and efficacy of food [7].The robot offers the recommendation statement 'I recommend it, you should try it!' after stating some trivial knowledge about the product combined with objective expertize such as 'It is the most popular and rich in iron.' The Combination Recommendation is a method that combines the subjective statements by the Eating Recommendation and the objective statements by the Expert Recommendation.The appearance of the robot is as shown in Figure 2(b).The robot's statements are the same as in the Expert Recommendation condition, but the two pseudo-eating behaviors were inserted between the statements.For example, the robot said 'GABU' and 'It is the most popular and rich in iron.' In the Eating and Combination Recommendations, the robot has bread in both hands, and the robot in Expert Recommendation does not have bread.The video of each condition was about 20 seconds.

Participants
A total of 513 answers were obtained through crowdsourcing separately from Survey I. Of these, 290 responses answered none of the dummy questions, resulting in a valid response rate of 56.5%.Since it was necessary to equalize the number of responses for each combination to use Scheffe's method, the earliest responses were extracted, and a total of 198 answers (male respondents: 119, female respondents: 77, no answer: 2) were used for the analysis.It is assumed that most of the participants' backgrounds are Japanese because this survey was conducted in Japanese.

Hypothesis
The results of Section 2.1 show that the pseudo-eating behavior increases the trustworthiness of the robot's subjective statement.Therefore, we assume that the recommendation accompanied by the robot's subjective statement with the pseudo-eating behavior can be comparable to the existing recommendation method proposed.In addition, by combining existing recommendation methods with the pseudo-eating behavior, it is possible to further improve the trustworthiness of the robot's statement and the recommendation.This study indirectly evaluates the robot's recommendation capability through 'Look Delicious' and 'Hope to Eat' in the questionnaire because the online survey cannot directly measure the recommendation capability.

Results
Figure 4 shows the average preference between each condition and the results of the ANOVA.The ANOVA was performed with a significance level of 5% for each item.The results reveal significant differences in all items.
Next, the post-hoc test with a yardstick was performed on all items.The results show that the Eating Recommendation was higher than the Expert Recommendation in 'Trustworthy, Intelligent, Human-like, Influential, and Friendly' and 'Interested, Understandable, Looks Delicious, and Hope to Eat.' In addition, the results show that the Eating Recommendation was higher than the Combination Recommendation in 'Trustworthy, Uncomfortable, Intelligent, Human-like, Incredible, and Friendly' and 'Interested, Understandable, Obtrusive, Looks Delicious, and Hope to Eat.' Finally, the results show that the Expert Recommendation condition was higher than the Combination Recommendation in 'Uncomfortable, Human-like, Incredible, Influential, and Friendly' and 'Obtrusive, Looks Delicious, and Hope to Eat.'

Discussion
The results show that the Eating Recommendation condition was higher in 'Looks Delicious and Hope Eat,' as compared to the other two conditions.These results are consistent with the results in the survey I, and we also assume that the trustworthiness of the robot's statement was improved because 'Human-like and Friendly' are improved.Hypothesis 2.1 assumed that the recommendation capability of the Eating Recommendation is as competent as that of the Expert Recommendations because the Eating Recommendation does not reduce the trustworthiness of the robot' statement.However, the results of the Eating Recommendation were more trustworthy, and the recommendation capability on such items as 'Looks Delicious and Hope to Eat' was higher.Therefore, Hypothesis 2.1 was not supported.The results show that the Eating Recommendation can achieve a higher recommendation capability than existing recommendation methods.
On the other hand, the Combination Recommendation was less effective than the Eating Recommendation, although the Combination Recommendation was more effective than the Expert Recommendation for many items.The reason for this may be that 'Uncomfortable, Incredible, and Obtrusive' are higher than the other conditions.This may be because the pseudo-eating behavior 'GABU', which reinforces the robot's subjective statement, was inserted in the middle of the expert recommendation 'It is the most popular and rich in iron.', which represents an objective statement.This is an example, but it shows that the robot understood the popularity of bread through its eating behavior.Thus, the robot's behavior became ambiguous and inconsistent.It is known that when the robot's behavior is inconsistent, the user tends not to continue the interaction [31].Thus, the results of this study are consistent with the results of the previous study.These results do not support Hypothesis 2.2.
Here, we investigated the recommendation capability of each recommendation method through an online survey, but it is unclear whether the same effect would be observed in a real store.Therefore, in the following sections, we present our examination of the effect of the pseudo-eating behavior of the robot on the user's purchase behavior in a real store.

Overview
This section presents the effectiveness of the Eating Recommendation as compared with two types of recommendation behavior of the robot shown in Section 2.2 in a real bakery store.For this verification, we installed a product recommendation robot in a bakery and conducted a recommendation experiment for five specific types of bread.The experiment lasted for 15 days.The number of sales of the recommended products was evaluated to compare the robot experiment period with the non-experiment period.This study was approved by the Research Ethics Committee of Ritsumeikan University (reference number: BKC-HitoI-2020-027-4).

Experimental environment
This experiment was conducted at a bakery at Ritsumeikan University.Figure 5 shows the appearance of the bakery.This store is open for 4.5 hours from 10:30 to 15:00 on weekdays and the number of daily customers visiting this store is approximately 70-160.During the experiment, the robot was operational during all business hours.

Interaction design
The robot used in this experiment was Sota which was also used in the online surveys.In addition, an RGB-D camera (ZED2i manufactured by Stereolabs Inc.) was installed next to the robot, and the behavior of customers around the robot was measured using a pose detection algorithm (OpenPose [32]).The robot behaved autonomously according to the measured human behavior.
Figure 6 shows the robot recommending bread in each condition.The robot was installed on the counter where the products were displayed so that it could be easily noticed by the visiting customers.In addition, the recommended bread was placed in front of the robot so that the customers could immediately recognize it.The behavior of the robot stopped when there were no customers in the bakery, but the robot started behaving and recommending products when customers visited the store.The robot could not communicate with the customer exhibiting passive media behavior where information was transmitted unilaterally [33].The robot continued to recommend bread for the duration of the customer's stay.
The scene during the field experiment is shown in Figure 7.

Experimental conditions
The experimental conditions of this field experiment used 'Eating Recommendation', 'Expert Recommendation', and 'Combination Recommendation' which are the same as the online survey II described in Section 2.2.In the Eating Recommendation, the robot presented subjective statements such as the taste and texture of the recommended product along with the pseudo-eating behavior and said 'I recommend it, you should try it!' at the end.In the Expert Recommendation, after explaining objective expertize such as the popularity, trivial knowledge, and efficacy of the recommended product, the recommendation statement was presented at the end.Finally, in the Combination Recommendation, the robot's statement was the same as in the Expert Recommendation, but the pseudo-eating behaviors were inserted in the middle of the statements.
Several dialog sentences were prepared and randomly used for each condition so that the robot did not repeat the same content even if the customer stayed for a long time.All the prepared sentences lasted for about 20 seconds.In the Eating and Combination Recommendations, the robot had bread in both hands as shown in Figure 6, and the robot in the Expert Recommendation condition did not have bread.

Experimental schedule
Table 1 shows the experimental schedule.The experiment was conducted for 5 days under each condition, for a total of 15 days.The experiments were conducted under one condition per day, in the order of Expert, Eating,  and Combination Recommendation.Then, specific conditions were assigned once for each day of the week, and the order of the conditions within the same day of the week is randomized.In addition, in this experiment, the robot recommended only one type of bread throughout the day, but we recommended a total of 5 types of bread by changing the recommended bread every 3 days.This is because recommending the same product throughout the experimental period would cause customers to get bored.
In addition, to compare the recommendation capability of the robot, data were also collected for a total of 20 days, 10 days (two days to each day of the week) before and 10 days after the experiment, as the non-experiment period.(5) Comb

Measurement
We used, as the evaluation indices, the sales amount, the number of customers, the number of all types of bread sold, and the sales ratio of recommended bread, for a total of 35 days (15 days during the robot experiment period and 20 days during the non-experiment period).
In terms of the average number and the sales ratio of recommended bread sold, only one type of bread is recommended during the robot experiment period, while all five types of bread are targeted during the nonexperiment period.Therefore, the average number of sales P of recommended bread is defined as follows: where M is the number of the types of target bread per day, N is the number of target days, and K j i is the number of recommended breads j sold on day i.We use M = 1, N = 5 for each robot condition, and M = 5, N = 20 for the non-experiment period.
The average sales ratio of the recommended bread indicates the ratio of the recommended bread sold to the total number of all bread sold.Therefore, in this study, the sales ratio S of the recommended bread is defined as follows: where A i represents the number of all breads sold on day i.In this study, since the robot only recommended bread, other products such as drinks were excluded from the evaluation.

Participants
In this experiment, customers who purchased at least one bread were selected as research participants.During the 15-day period of the robot experiment, a total of 1594 customers (Expert Recommendation: 562 customers, Eating Recommendation: 521 customers, and Combination Recommendation: 511) were evaluated.There were 2,007 customers during the non-experiment period (10 days before the experiment: 978, and 10 days after the experiment: 1029).Because this study was conducted in a bakery at Ritsumeikan University, it is assumed that most of the participants' backgrounds are Japanese (approximately 92% of the students are Japanese according to data from the university in 2023 1 ).

Hypothesis
The results of the online survey II in Section 2.2 show that the Eating Recommendation promoted the user's willingness to purchase products as compared with the Expert and Combination Recommendation.Therefore, we assume that the recommendation capability of the Eating Recommendation is also high when recommending products in the real store; thus, the following hypothesis was proposed.
Hypothesis 3: Recommendation using the pseudoeating behavior of the robot has a higher recommendation capability than other recommendation methods.

Results
Table 2 shows the average sales amount and the average number of customers for the day.The results of the robot experiment period (15 days) and the results of the non-experimental period (20 days) are also shown.We compare the periods between the robot experiment and the non-experiment using the t test.The results show no significant difference for either index (average sales amount: (t(28) = 0.48, p = 0.64), average number of customers: (t(28) = 0.73, p = 0.47)).Table 3 shows the results of the average number of all bread units sold for the day, the average number of recommended bread units sold, P, and the average sales ratio of recommended bread units, S. We used the chisquared test in the average sales ratio of recommended bread among four conditions: Expert, Eating, Combination, and non-experiment period.We used Cramer's V as the effect size in the chi-squared test.The results show a significant difference in the average sales ratio (χ 2 (3) = 23.33,p < .01,V = 0.06).In addition, the residual analysis shows that the average sales ratios in the Expert and Eating Recommendations are significantly higher than the overall average.Whereas the results in the nonexperiment periods are significantly lower than the overall average.

Discussion
Regarding the recommendation capability of the robot, the sales ratios of the recommended bread in the Expert and Eating Recommendations increased significantly.This implies that the pseudo-eating behavior of the robot can influence the purchase behavior of customers and has the same recommendation capability as the Expert Recommendation, which is the current method.While the results of online survey II showed that the Eating Recommendation was more effective than the Expert recommendation, no difference was found between the two conditions in the field experiment.Therefore, Hypothesis 3 is not supported.One of the reasons for this is that there is a risk in purchasing decisions that customers have to pay in the real store.Since the customers may be disappointed if they purchase bread that does not meet their taste, it is possible that the customers tend to rely on information based on objective facts rather than the robot's subjective statement.Hypothesis 3 was not supported, though the robot's subjective statement can be accepted by explicitly sharing the robot's experience with the user.The experimental results show that the recommendation by the robot's subjective statement using the pseudo-eating behavior may be able to achieve a recommendation capability equivalent to that of the existing methods.This result shows that the robot overcomes the problem of unacceptability of its subjective statements, as explained in the introduction.
In addition, although the Combination Recommendation resulted in a higher sales rate of the recommended bread than the non-experiment period, it was lower than the Eating and Expert Recommendations.This seems to be because the result of Section 2.2 shows that it is more uncomfortable, incredible, and obtrusive than the other conditions.The robot's statement includes both subjective statements 'GABU' and objective statements 'It is the most popular and rich in iron.'.Since the robot understands the popularity of bread from its eating behavior, the robot's behavior was unnatural; therefore, its recommendation capability was not high.In the future, it will be necessary to combine both types of statements for the recommendation to make the robot's behavior consistent.
We observed customer behavior during the experiment, but a significant difference between conditions was not observed.The robot was placed in the customer flow line in the store, and the robot provided information from next to where the customer was choosing bread.Hence, many customers saw the robot at least once while in the store, recognized its presence, and heard the recommended information.Rather than stopping in front of the robot to listen intently, many customers heard audio information from the robot while they were choosing bread.No differences in these customer behaviors were observed between conditions.In this experiment, the placement of the robots was designed to be easily noticed by customers; therefore, we assume that this is a reason for the lack of differences in customer behavior between conditions despite showing a significant difference in user experience in Online Survey II.It is known that in public spaces, differences in robot behavior can change the rate of passersby who listen to the robot [31].Thus, customer behavior might have differed between conditions if the environment in which the robot approaches users is different.
Finally, in this experiment, there were no significant differences in the average sales amount and the average number of customers for the day between the robot experiment period and the non-experiment period.That is, the robot recommendations used in this study were not for the customer to purchase an additional item, but to switch the purchased item from a certain bread to the recommended bread.Therefore, to improve the sales of the store, it is necessary to further investigate and improve the method.

Conclusion
The aim of this study was to investigate how the pseudoeating behavior of the robot influences the perceived trustworthiness of the robot's subjective statement and its capability to make good recommendations.Then, we verified the effectiveness of the proposed method through two online surveys and one field experiment.
Two online surveys were conducted to verify the influence of the robot's pseudo-eating behavior on the effect of its subjective statements on users.The first online survey focused on the degree of the shared experience, independent of recommendations (Hypothesis 1).The second online survey focused on the effectiveness of the product recommendations (Hypothesis 2.1 and 2.2).The results of two online surveys found that the pseudoeating behavior of the robot had a positive effect on the user's impression and experience in many aspects, including the perceived trustworthiness of subjective statements.In addition, we conducted the field experiment to present the effectiveness of the pseudo-eating behavior for recommendation in a real store (Hypothesis 3).The results of the field experiments show that the recommendations accompanied by the robot pseudoeating behavior can achieve the same recommendation capability as other existing recommendation methods.These series of results imply that the robot is trusted even when they express subjective statements for recommendations by explicitly sharing its experience with the user.
In this study, we proposed pseudo-eating behavior to explicitly share the experience while recommending food products.However, depending on other recommended non-food products, it is unclear what kind of pseudobehaviors should be expressed for shared experiences.Furthermore, it is unclear whether the results of this study are applicable to other robots.We assume that the proposed method would not be applicable to robots that do not have a mouse design because the pseudoeating behavior would be more fake.Therefore, in the future, it is necessary to verify the effectiveness of the pseudo-eating behavior for each product category and robot type in the future.
We should also keep in mind that the results of this study are highly dependent on the background of the participants.Previous studies have shown that people from different cultures behave differently [34,35].The results of this study may be shown by participants from many Japanese cultures.Therefore, experiments that take into account the diversity of participants' backgrounds, such as culture and experience with the robot, are needed for general findings.
In addition, in this study, the robot recommendation was conducted in a bakery for a total of only 15 days, but it is expected that customers will get bored once the novelty effect wears off, when conducting long-term studies [36].In such cases, it is also unclear whether the robot should establish relationships with customers and recommend products according to the relationship.In the future, we will verify a recommendation method that considers customer relationships.

Note
and rich understanding of user experiences.He is a member of ACM and IPSJ.
Jun Baba received his ME degree in informatics from Kyoto University, Kyoto, Japan in 2014.He was a data scientist at CyberAgent, Inc. in Tokyo, Japan from 2014 to 2017.He has been a research scientist at CyberAgent AI Lab and a visiting researcher at Osaka Universit since 2017.His research interests include teleoperation for social robots, humancomputer interaction in service encounter, and artificial intelligence.
Junya Nakanishi received the Ph.D. degrees in engineering from Osaka University, Osaka, Japan, in 2018.In graduate school, he was a Student Intern, Advanced Telecommunications Research Institute (ATR, 2012-2018) and a Research Fellow of the Japan Society for the Promotion of Science (JSPS fellow, DC2, 2016-2018).Currently, he is a Research Assistant Professor at Frontier Intelligent System Laboratory, Osaka University.His research interests include human-robot touch interaction, human-robot interaction in service encounter and advertisement.

Figure 2 .
Figure 2. Appearance of robot in online surveys.(a) Without and Explicit Experience conditions in Survey I and Expert Recommendation condition in Survey II and (b) Eating Behavior condition in Survey I and Eating and Combination Recommendation conditions in Survey II.

Figure 3 .
Figure 3.The results of the average preferences in an online survey I.The top 7 items involve the impressions of the robot, and the bottom 5 items involve the user experience.The larger the positive value, the truer to the condition, and * indicates the significant differences at the 5% significance level between conditions.

Hypothesis 2 . 1 :Hypothesis 2 . 2 :
The recommendation capability of the Eating Recommendation is as competent as the Expert Recommendation.The recommendation capability is improved by combining the expert recommendation with the pseudo-eating behavior.

Figure 4 .
Figure 4.The results of the average preferences in an online survey II.The top 7 items involve the impressions of the robot, and the bottom 5 items involve the user experience.The larger the positive value, the truer to the condition, and * indicates the significant differences at the 5% significance level between conditions.

Figure 6 .
Figure 6.Robot recommending bread.(a) Expert Recommendation condition and (b) Eating and Combination Recommendation conditions.

Figure 7 .
Figure 7. Experimental scenes in a bakery with customers listening to the robot.

Table 1 .
Experiment schedule.Numbers in parentheses indicate types of recommended bread.Exp, Eat, and Comb indicate Expert, Eating, and Combination Recommendation conditions, respectively.

Table 2 .
Average sales amount and the average number of customers for the day.

Table 3 .
Average number of all bread sold for the day, average number of recommended bread units sold P, and average sales ratio of recommended bread S.