In the field of recommendation, there have been many eforts to help users interact with recommender systems in ways that appropriately elucidate user preferences. To let users interact with recommender systems, it is desirable that recommender systems are as transparent as possible. However, it is dificult to achieve complete transparency even with a simple method for interactive recommender systems because the relationship between the item features and the user preference is not intuitive when there is a utility function to generate recommendations. We focus on multiattribute utility theory (MAUT) as one of the simplest methods for recommender systems and clarify the dificulties with its usage in interactive environments. Then, to overcome the dificulties, we propose an algorithm to generate natural language hints to let users understand ways of operation and see more items that match their preferences. The results of an ofline simulation demonstrate that our method can efectively recommend a diverse range of items to users. As future work, we will conduct empirical experiments to evaluate the performance of our method in online situations.
• Information systems → Recommender systems.
Interactive recommender systems have been developed to
elucidate user preferences [
To achieve this ideal interaction between recommender systems
and users, the systems should be transparent in that users can
understand the relationship between item feature space and user
preference space. This is dificult, however, due to the relationship
between the two spaces with utility functions, and according to
previous research, users do not tend to obtain a detailed enough
understanding of the intelligent system without a technical
explanation [
In this paper, we propose a new algorithm to help users
understand the specific characteristics of the relationship between
preference space and item space with one of the simplest utility functions
and conduct an ofline experiment. As a simple method to relate
item space and preference space, we utilize Multi-Attribute Utility
Theory (MAUT) [
The remaining of this paper is organized as follows: In Section 2, we provide a brief overview of previous works related to interactive recommender systems and MAUT. We describe the dificulties in operating MAUT-based interactive recommender systems and propose our algorithm in Section 3. In Section 4, we explain the detailed settings of our experiment. In Section 5, we report the results of our experiment. Additionally, in Section 6, we discuss the results in more detail and outline future work. Finally, the conclusion of the paper shown in Section 7. Our contribution to the community is two-fold: first, we point out the critical dificulties in one of the simplest utility functions, MAUT, when it is used in interactive recommender systems, and second, we ofer an algorithm to generate natural language hints to overcome these dificulties. 2
RELATED WORK AND BACKGROUND
There have been many eforts to develop interactive recommender
systems [
Relating to MAUT, there are many studies that utilize this method
to support human decision making [
In this section, first, we explain the functions of the MAUT-based interactive recommender system in detail. Then, we describe the dififculty experienced by users and present the method we developed to overcome the dificulty. 3.1
3.1.1 MAUT-based Interactive Recommendation. First, we briefly explain the characteristics of our MAUT-based interactive recommender system. As a method of recommendation, we utilize MAUT to recommend top-N items (top-10 items in practice). To rank items, the personalized utility (V u(i)) of each item (i) for each user (u) is calculated on the basis of MAUT, as follows:
Õ Vu (i) = f β (u)xf (i) f (1) where a feature for an item is denoted as f and the normalized value for each feature is denoted as xf (i). In this paper, we use the word ‘feature’ in the same meaning of ‘attribute’ of MAUT. With this function, we calculate a user’s preference as the values of coeficient ( βf (u)), which show to what extent the feature f is considered an important one by the user u. An item i is denoted as K-dimensional feature vector x (i) : x1(i), ..., xK (i).
β 15 10 0
UP (+)
UP
(+) Feature 1
DOWN
Feature 2
DOWN
We developed an interactive recommender system to collect real user data by using this method. In our system, the users’ preferences are displayed as bar graphs that indicate the value of β . The concept of our interface is shown in Figure 1, where there are two bars
Item space ) 2 F ( 2 e r u t a e F : Items
E : Direction users change their preference in UI 3rd E
D2nd
C A 1st
B User preference (a) representing the user preference (β ) for each feature of items. The values in Figure 1 correspond to the states of user preference in Figure 2. This user has a value of 15 for the preference of Feature 1 (F1) and a value of 10 for the preference of Feature 2 (F2). In our system, users can see the bar graphs as well as the recommended item list and change their preferences with a fixed length by clicking the buttons above and below the graphs. Moreover, they can change preferences for multiple features at a time. In the following, the number of times a user changes her/his preference (i.e., the number of interactions) is denoted as t . 3.1.2 Dificulties in MAUT. Users are confronted with dificulty when MAUT is used in interactive recommender systems. With MAUT, the scores of the items are calculated as the inner products of user preference and the position vectors of items in the item feature space. This means that the ranking of the items is the same as the order of the perpendicular feet drawn from the items to the user preference vector, as shown in Figure 2. This leads to unintuitive behavior of the recommender system and to dificulties in operating preference for the user. For example, as we show in Figure 3, when there are two features and the top-3 item list is displayed to a user, if the user wants to see items with a higher value of Feature 2, s/he naturally increases the preference for Feature 2. Ideally, s/he will see the item with the highest value for Feature 2 (item E) by increasing the preference for Feature 2 just once. However, it is dificult to see item E by increasing only feature 2, as shown in Figure 3(b). To see item E, the user has to decrease the preference for feature 1 while simultaneously increasing that for Feature 2 (Fig. 3(c)). This implicates that it is necessary for users to change their preferences in unintuitive manners. 7 items with higher values of Feature 1 and with lower values of Feature 2 can be shown. (Your preference for Feature 3 will decrease.) Jump in this direction To overcome the dificulties in operating the MAUT-based interactive recommender system described above, we propose a technology that generates hints to achieve more efective operation for users. The hints will be shown in the form of natural language1. As indicated in Figure 4, the hint suggests a direction users can take to see more diverse items and users can move toward this direction by clicking the ‘jump in this direction’ button. Therefore, users can decide to use or not to use hints whenever the hints are displayed. Our algorithm calculates the appropriate direction to be suggested to users. The appropriate direction is determined on the basis of the tendency of the user’s movement. First, our method calculates preferences that make as many items Top-1 as possible. With MAUT, items that can be top-1 are vertices of the convex hull of items. Therefore, we calculate every preference that makes each vertex of the convex hull of items Top-1 respectively. We call these preferences ‘top-1 vectors’. After that, we calculate which direction a user wants to move toward and which features s/he does not take into account. We call the direction in which the user moves ‘favored direction’ or more simply, ‘direction’, and the features that are not taken into account by users ‘ignored features’. Then, 1While we do not show the natural language hints to real users in our ofline simulation, we plan to conduct an online experiment by showing natural language hints as future work. we suggest the appropriate operation that enables the user to see items that have values matching the user’s tendency to move based on the top-1 vectors. Additionally, we ofer the user a way of operation that lets him/her see more items by adjusting the features that have been ignored until then. The details of this algorithm are described in the following subsection.
Item space 3.2.1 Calculation of Top-1 Vectors. To generate hints, we utilize the characteristics of MAUT that only vertices of the convex hull can be top-1 and calculate preferences making each vertex top-1. One of the preference vectors that makes a vertex top-1 (top-1 convex) is orthogonal to the plane (or hyper-plane) passing through vertices that are next to the top-1 vertex. This is because only vertices can be top-1 and rankings of items are decided as the order of the perpendicular feet drawn from items; therefore, when making a vertex top-1, it is enough that the perpendicular feet drawn from vertices that are next to the top-1 vertex are the same order. A 2-dimensional version is shown in Figure 5, where the top-1 vector for a vertex, A, is drawn. The vector is orthogonal to the line passing through the vertices B and D, which are next to A.
To generate top-1 vectors, we calculate the hyper-plane passing
through vertices that are next to a vertex that is made top-1 with
the vector. We can calculate this hyper-plane as the support vector
of the hard margin support vector classification for two classes: the
class of the top-1 vertex and the class of the other vertices [
minimize ||z||2 subject to yi (zT ai + b) ≥ 1 for i = 1, ..., M . zT ai + b = 0 is the equation of the hyper-plane that is parallel to the hyper-plane passing through next vertices to top-1 vertex and z is a perpendicular line for the hyper-plane. Therefore, we obtain the top-1 vector as z. 3.2.2 Tendency of Users’ Behavior. Next, we calculate which directions a user should move toward. A direction is defined as the combination of features and the movement of increasing or decreasing of the value for the features. We name the value that represents (2) to what extent a user wants to make a value of a feature up or down T endency. To identify the direction favored by the user, we calculate a Score to generate the values of T endency. When calculating the Score, we consider the more recent movement of the user as more important. First, we set the weight wt = wt , which indicates the importance of the movement of t th interaction as the number that increases in proportion to the number of interactions (t ) between the recommender system and the user. This means that if w1 = w = 1.2, then w2 = w ∗ w = 1.44, that indicates the importance of the movement of second interaction. The vector consisting of all weights is denoted as wt and the history of choice for up (+) or down (−) for a certain feature value is denoted as hf,±. Here, we define the Score as
Scoref ,± = hf,± · wt.
For example, when a user change the preference for Feature 1 in the three interactions like U P → DOW N → U P , h1,+ = (1, 0, 1) and h1,− = (0, 1, 0). When w = 1.2, w3 = (1.2, 1.44, 1.73). Therefore, Score1,+ = (1, 0, 1) · (1.2, 1.44, 1.73) = 2.93 and Score1,− = (0, 1, 0) · (1.2, 1.44, 1.73) = 1.44.
Then, we define the T endency as
T endencyf ,± =
Scoref ,± . Ít wt When consider the example we explained above, T endency1,+ = Score1,+/Ít wt = 2.93/4.37 = 0.67, and similarly, T endency1,− = 0.33. We infer the direction in which the user wants to move by judging whether the T endency surpasses the upper threshold, UT . For example, if T endency1,+ > UT , we assume the user wants to increase the value of Feature 1. Additionally, the features ignored by the user can be known by judging whether the T endency of a feature for both up and down is lower than the lower threshold, LT . For example, if T endency2,+ < LT at the same time T endency2,− < LT , Feature 2 is ignored by the user. 3.2.3 Combinations of Directions. After inferring the direction toward which the user wants to move and the features ignored by her/him, we generate the combination of the directions favored and ignored by the user. Initially, we set the number of hints to display and the number of directions to consider. Then, we calculate the sum of T endencies for the combinations of directions that a user moves toward. For example, consider a case where there are three directions users tend to move toward, (Feature 1, +, Tendency = 0.6), (Feature 2, -, Tendency = 0.7 ), (Feature 3, +, Tendency = 0.55). When we take two directions into account, we obtain three combinations of directions and the sum of T endencies, i.e., (Feature 1, +, and Feature 2, -), (Feature 1, +, and Feature 3, +), and (Feature 2, - and Feature 3, +). For example, in terms of the sum of T endency, the top-1 in the combinations is (Feature 1, +, and Feature 2, -). The sum of T endency of this combination is 1.3. Now, we obtain the list of combinations of directions.
If there are any features ignored by the user, we add the direction related to the ignored features to the list. For example, when Feature 4 is ignored, (Feature 4, +) and (Feature 4, -) are added to the list. Therefore, the top-1 in the list of combinations is (Feature 1, +, Feature 2, -, and Feature 4, +) or (Feature 1, +, Feature 2, -, and Feature 4, -). The plus or minus of Feature 4 is decided on the basis of how many items will be changed by the hint. If the number of items that (3) (4) Algorithm 1: Calculate Candidate Vector input :d: Combination of directions, I : Vertices,
SI : Shown items, P : User preference output : Candidate vector 1 GenerateCandidateVector(d, I , SI ) 2 MaxItem ← i most matching d in SI ; 3 foreach j in I do 4 if j more matching d than MaxItem then 5 Add j to CandidateItem appear in the new recommendation list when the combination with Feature 4, + is more than that with Feature 4, -, the combination with Feature 4, + is applied. 3.2.4 Calculation of Candidate Vector. On the basis of the listed combinations of directions, we calculate the candidate vector to generate hints by using algorithm 1. First, the values of features in all vertices (I ) are compared with the values of displayed items (SI ) to judge whether there are items that have higher (or lower) values suited to the listed combinations of directions. Each item in SI is represented as i. If there are vertices that match the combination of directions (d), they are added to the list of candidate items. For example, when the combination of directions toward which the user wants to move is (Feature 1, +, and Feature 2, -), if there are vertices whose values of Feature 1 is higher than the highest value for Feature 1 in the displayed items, and at the same time, if the values of Feature 2 of the vertices are lower than the lowest value for Feature 2 in the displayed items, the vertices are added to the list of candidate items.
Next, for each candidate item, the top-1 vector for the item is selected and the top-N item list is calculated on the basis of the top-1 vectors. At the same time, the number of diferent items in the top-N list that is generated on the basis of the top-1 vector from the currently displayed top-N list is calculated. Then, the top-1 vector that generates the top-N list in which the number of diferent items from the currently displayed items is the biggest is chosen as the candidate vector. If there remain multiple candidate vectors at this time, we select the one candidate vector that has the biggest cosine similarity (cos sim) with the current user preference (P ). 3.2.5 Generating Hints. Finally, we display hints for users with natural language. Although multiple hints can be displayed, the number of hints should be limited to avoid information overload. When multiple hints are displayed, the hints generated based on top-N combinations in terms of the sum of T endency are displayed. To display hints in the form of natural language, we generate hint vectors (HV) from the chosen candidate vectors (CV) as (5) HV =
CV ||CV|| ∗ ||P||, where the user preference is denoted as P. Each hint vector is the unit vector of the candidate vector multiplied by the norm of the current user preference. This calculation is executed because the norm of candidate vectors is not fixed at this time and T endency is calculated on the basis of the diference between current preference and the hint vector, so the norm of the vector and the preference should be aligned.
Then, natural language hints are generated in two diferent forms. The first one is hints without ignored features. In this form, we focus on making the user preference move to the hint vector. For the hint, we display a natural language hint and ofer a ‘Jump in this direction’ button to make the current preference jump to the hint vector as shown in Figure 4. This means that, when the user clicks the ‘Jump in this direction’ button, displayed user preference changes to the values of the selected hint vector and the displayed list of items changes at the same time. Additionally, we display the information on features that are not included in the combinations of directions if the diference of values for the feature between the current user preference and the hint vector is large enough. For example, when the chosen combination is (Feature 1, + and Feature 2, -), seven items will be newly shown with the selected hint vector, and at the same time, the diference of Feature 3 between current user preference and the hint vector is large enough, the hint is shown as “7 items with higher values of Feature 1 and with lower values of Feature 2. (You need to decrease the preference for Feature 3.)” The second form is hints with ignored features. In this form, while we generate the same natural language hint as the first one, the user can move toward the suggested direction with the same value by clicking the ‘Jump in this direction’ button, as users can move with the normal operation. By “normal operation,” we mean the operation with the up or down button in the interface displayed regardless of the existence of hints. This is because the hint vectors considering the ignored features can change the user preference in radically diferent directions from the direction in which the user wants to move, and so the movement toward the hint vectors should be gradual. 4
To conduct the ofline experiment with our method, we have to define the simulation environment. In this section, we explain the dataset we use, the settings of the parameters, the policy of the simulated user behaviors, and the metrics to evaluate the performance. 4.1
To conduct the ofline experiment, we use the data gathered from our MAUT-based interactive recommender system to let users find their favorite areas to relocate to. There are two reasons why this application domain is chosen. First is that there is a trade-of relation between convenience and safety in this domain. This trade-of is needed because it makes users search for their appropriate preference between conflicting characteristics. Otherwise, users will just make the preference for each attribute maximum or minimum. Second is that users interact with recommender systems more seriously when making important decisions related to their own lives.
When using our recommender system, users input their demographic characteristics to the system and are recommended areas that match them. After that, users can adjust their preferences and approach their final preferences. The recommender system recommends blocks in a city (areas from one street to the next street) as items that are characterized by six features: i). Transportation: the number of trains arriving at the station in the area. ii). Shopping: the scale of the shopping facilities. iii). School proximity: the time it takes to walk to the nearest elementary school. iv). Neighborhood activity: the attendance ratio to events of neighborhood associations. v). Hospital: the number of hospitals.
vi). Safety: crime rate.
We normalize these features as 0 for the minimum number and 1 for the maximum number when calculating users’ utility for each area. The number of 0 is set to show that there are not any positive or negative values related to the feature. With our system, 163 blocks can be recommended, and 83 blocks are vertices. Therefore, we generated 83 top-1 hint vectors. While the system was open publicly on the Internet, we use the data of users who came into our ofice to use the system because the motivation to use the system should be aligned. The users are interested in relocating to the same or similar cities that the system recommends. They are asked to look at the recommended areas and mark the ones they think they might like to move to as favorite areas. We used the data from 49 users who marked at least one favorite area. Users who could not ifnd any favorite areas or who did not use the system seriously are not included in the data. We evaluated the seriousness of users with a 5-scale questionnaire of impressions about the system after using it. If a user checked the same point for all questions or the answers were not coherent, we excluded the user from the data set. 4.2
We set the parameters for our experiment as follows. First, the number of features (f in Equation 1) is six. Therefore, we conduct our experiment in the six-dimensional feature space. We consider the situation where the top-10 ranking of recommended items is shown to users. This is the same number used in our collection of the data set. Next, we set the value for the weight (w) when calculating the Score and T endency of the favored direction as 1.2 (= w = w1 in Equation 3 and 4). We set the upper threshold UT to judge the direction a user wants to move toward as 0.5 and the lower threshold LT to judge the ignored features as 0.3. Then, we set the number of displayed hints to 1. In our experiment, this means that when the hints are displayed, users always choose the hint at the top of the list. Additionally, when we generate the favored direction, we take three features into account. This is because we want to display hints with natural language, and more than three hints would be too many to take in at a glance. 4.3
To conduct the ofline experiment, we need a policy for the
simulation. The ofline simulation of interactive recommendation methods
is considered dificult because the reactions of real users cannot
be known. While simulation methods are investigated for
critiquebased recommender systems [
Can hints be generated?
YES
Generate hints.
Change preference in the offered direction.
Is the direction in which the user preference move similar to that in the last movement? YES NO
Have any hints generated so far?
YES Preference changes in the same way as the real user’s.
Apply simulation mode.
NO
The detailed flow of our simulation is as follows. We summarize the flow in Figure 6. In our scenario, there are three ways of movement. The first is the movement that is the same as a real user’s movement. When hints have not been displayed yet, the movements of preference follow the log data of real users’ movement. The second is movement when the hints are displayed. The simulated users always click the ‘jump in this direction’ of the hints at the top of the displayed hints button whenever the hints are displayed in our scenario. If the number of items that are diferent from previous recommended items is less than four, we start to 2 We use the deterministic approach, not stochastic approaches mainly because we do not know how often users keep directions suggested by hints and how the final preferences changes with the existence of hints. generate hints. If we generate the hints to recommend items that match a user’s preference successfully, we make the preference of the user jump to the hint vector. If there is a hint vector based on the ignored features, the user preference moves toward the hint vector generated on the basis of the ignored feature regardless of the existence of the hint vectors generated on the basis of the favored directions.
As the third type of movement, we set the simulation mode. If the user’s T endency changes radically resulting from the jump to hint vectors, it is not appropriate to continue to generate hints. Therefore, if the sign of favored direction for the same feature becomes opposite to the previous tendency resulting from the application of hint, we stop generating hint vectors and start to apply the simulation mode. In this simulation mode, the preference is made to move straight toward the final preference. We calculate the distance between the current preference and the final preference, and if the distance is more than that the user can move at one time, we make the preference move toward the final preference. If the distance from the current value and the final value is more than the distance a user can move at one time, we select the top-3 feature whose values are distant from those of final preferences more than the distance that a user can move at one time, and after that, make the preference moves toward the final preference at one time. In the simulation mode, hints can be generated. 4.4
We evaluate the performance of our method with several metrics. First, we identify the percentage of users provided with hints. In our method, a user needs to move her/his preference in the same direction to some extent to be provided with hints. Naturally, there are some users who do not have any specific direction they tend to move in. Therefore, by dividing users into those who are given hints and those who are not, we investigate the possibility of the provision of hints. We count a user as one who is shown hints if s/he is given hints at least just once. Then, we investigate the performance of our method only for the users provided with hints.
Next, for the users provided with hints, we use a dependent sample t-test to compare the number of interactions needed to move close to the final preference and the number of favorite items shown for each user. The number of interactions is evaluated to confirm the behaviors of the simulated users is not far from that of original users. With our method and simulation policy, it is possible for user preferences to approach final preferences in a diferent way from the original one. By evaluating the number of interactions, we can analyze the efects of our method on the simulated user behavior. Additionally, the number of favorite items is investigated to evaluate to what extent users become unable to see their favorite items with the existence of hints. When our method is applied, the user preference sometimes changes radically and skip the items that are marked as favorite items originally. Therefore, by investigating the favorite items, we can evaluate whether or not users overlook their favorite items in our simulated environment.
Finally, we investigate the diversity and novelty of the
recommended items. Our research question is whether users can see
more diverse items when there are hints generated on the
basis of our method. Therefore, the diversity and the novelty are
our main results. As the metric to evaluate the diversity and
novelty achieved with our method, we apply temporal diversity [
N
This equation can only calculate the diversity between two lists.
Therefore, we calculate novelty to compare new recommendations
to the set of all items that have been recommended (At ) on the
basis of [
N If both the diversity and novelty increase significantly with the existence of the hints, it is clarified that our method has positive effects on the recommendation. We calculate the average of diversity and novelty for the number of users’ interaction (i.e. how many times a user renewed the displayed recommendations) as MD and M N respectively. 5
In this section, we report our results. We initially explain the ratio of simulated users who are provided with hints. After that, the number of interactions and shown favorite items are analyzed. Finally, the diversity and the novelty of the recommendation is described. 5.1
Percentage of Users Provided with Hints First, we show the number of users who are provided with hints. Data from 49 users are used. A total of 20 users (40.8%) are provided with hints generated on the basis of users’ favored direction, and only one user (2.0%) is provided with hints generated on the basis of ignored features. This latter user is included in the users who are shown hints based on the favored directions. 5.2
The Number of Interactions and Displayed Favorite Items We check the results of the number of interactions and of shown favorite items to analyze the efects of our method to the user behaviors. The results are summarized in Table 1. The number of users analyzed (N ) is 20 because we only consider the users provided with hints. Between the original data and the simulated data (with the existence of hints), while the number of interactions does not decrease significantly ( p = 0.67), the number of favorite items shown to users decreases with the existence of hints, the p-value is more than but close to 0.05 (p = 0.067).
MD (Mean of temporal diversity) p < 0.001
MN (Mean of temporal novelty) p < 0.001
Finally, we compare the average of the mean of the temporal diversity for the times of interactions (MD) and that of the temporal novelty (M N ) per user. In Figure 7, we show the average of MD and of M N for the number of data (users) with bar graphs. We obtain the results that both MD and M N increase significantly with the existence of hints from when there are not any hints. The representative values are as follows: For MD, M = 0.245, SD = 0.190 when there are not any hints, and M = 0.425, SD = 0.100 with the existence of hints. With the t-test of MD, t = 7.07, d f = 19, and p = 1.00 ∗ 10−06. For M N , M = 0.120, SD = 0.098 when there are not any hints, and M = 0.242, SD = 0.096 with the existence of hints. With the t-test of MD, t = 6.52, d f = 19, and p = 3.03 ∗ 10−06. These results implicate that the generation of hints results in the efective recommendation of new items that a user has not seen before in our experiment. 6
We discuss the results and their implications in this section. Related to the ratio of users provided with hints, more than half of the simulated users are not provided with the hints. This implies that more than half the users do not have a definite tendency of movement. One possible explanation for this is that users search items at the same time as they change their preference, so sometimes there is not a strong tendency in the behavior. On the other hand, this means that it is possible that we can not provide hints more than half of the users when we conduct an online study. This implicates that we need to set some devices to make users have specific tendencies as fast as possible with intelligent user interfaces.
With the change in the number of interactions, it is indicated that our method does not have any positive or negative efects on the simulated users. While it is possible that the number of interaction decreases because the user preferences can jump to other values radically with our method, there is no significant efect on the number of interaction with our simulation policy. For the number of shown favorite items, when our method is applied, users sometimes can not find items that they originally marked as favorite items. Skipping the favorite items can generate the missing of opportunities for both users and the item providers, so it is necessary to display skipped items in some ways. However, this may occur the problem of information overload and devices to overcome this problem is needed. Finally, the results of the average of temporal diversity and of temporal novelty indicate that our method can display more items for each user. This means that, as the answer to our research question, our method can provide more diverse range items for users than when there are not any hints.
However, we cannot confirm these implications are valid when we conduct an online experiment due to the limitations of our simulation policy. In our experiment, we adopt a simulation policy that assumes the final preference of each user that collected in the situation where there are not any hints does not change regardless of the existence of hints. However, in reality, the final preference of the same users can change when there are hints because diferent items are recommended to users and the cognitive load will change as a result of this additional information. It is possible that these diferences will lead to negative efects, such as users exit the system earlier or dificulties in identifying a clear tendency of user movement. There may also be other efects from the hints, such as favorite items changing or the final preference changing.
Therefore, as future work, we need to conduct an online user experiment by designing real user interfaces. In order to minimize the negative efects, we should design the parameters for the hints on the basis of user observations while, in our simulation, we set parameters without user observations using real user interfaces. Additionally, efective user interfaces have to be developed. In addition to the assumed efect of the existence of hints described above, we will investigate whether displayed hints are actually chosen by users. In the experiment reported in this paper, we assumed that users always utilize the hints whenever they are displayed. However, in a real situation, users might ignore the hints for several reasons, e.g., if the calculated tendency does not show the user’s preferred directions or if the hints are not highlighted enough. At the same time, the impact of hints on the user experience of the recommender systems should be investigated in the online experiment. 7
In this paper, to achieve transparent interactive recommender systems, we ofered an algorithm to generate natural language hints to let users know the appropriate way of movement that matches users’ preferences. We used MAUT as one of the simplest methods to relate user preference to the item features and clarified the dificulties in operating MAUT-based interactive recommender systems. With an ofline experiment, it is found that our algorithm achieves more diverse recommendations to the users. This will leads to the more transparent interactive recommender systems in that users can understand the detailed algorithmic behaviors of the systems. As future work, we will conduct online experiments to evaluate the performance of our algorithm in the real interactive environment.