Users are faced with an increasing information overload problem in large, complex data collections. Recommender systems reduce the data set to a manageable size by providing suggestions to the user. Research in the last years has primarily focused on the quality of the underlying algorithms. Recent research started to focus on the user experience in recommender systems. The main challenges are transparency, controllability, explorability, and context-awareness. Interactive visualizations have the potential to address all of these issues. In this paper, we present three user interface concepts for diferent usage scenarios: movie, activities, and travel search. We propose a taxonomy of user interface building blocks to evaluate these concepts with regards to the visualization challenges. CCS Concepts: • Information systems → Recommender systems; Search interfaces; • Human-centered computing → Interaction paradigms; Information visualization; Additional Key Words and Phrases: Recommender Systems, Information Visualization, Human Computer Interaction In large, complex data collections such as product catalogues, users are faced with an increasing information overload problem and it is dificult to find suitable items. To this end, recommender systems ofer an appropriate mechanism to reduce the data set to a manageable size by identifying items that may be interesting for the particular user [15]. During the last years, the quality of these systems has been considerably improved. However, most research has primarily focused on the algorithms to enhance performance and accuracy [13]. Latest research started to go beyond the optimization of algorithms by focusing on the user experience with recommender systems. Previous studies have shown that visual features and enhanced interaction improve the user engagement with the system and the acceptance of the recommended items [14]. Our literature review revealed many issues that visualizations can help to solve.In this paper, we focus on the following issues with regards to recommender systems that are easy to use for end users and support them in exploring and understanding the information space: • Transparency. Supporting the user in understanding the reasons behind the recommendations • Controllability. Providing user control over features that influence the recommender algorithm • Explorability. Presenting visualizations to browse the entire information space, e.g. related items that are not recommended • Context-Awareness. Considering diferent situations, such as mood, time, individual, or collaborative scenarios In this paper, we present three diferent user interface concepts tackling these issues by interactive visualizations. The paper is structured as follows: the next section considers related work regarding to the introduced visualization challenges. Subsequently, we illustrate interface concepts for recommender systems in three diferent scenarios:
VisBIA 2018 – Workshop on Visual Interfaces for Big Data Environments in Industrial Applications. Co-located with AVI 2018 – International Conference on Advanced Visual Interfaces, Resort Riva del Sole, Castiglione della Pescaia, Grosseto (Italy), 29 May 2018 © 2018 Copyright held by the owner/author(s).
movies, activities, and travel search. Each concept will highlight visualization and interaction techniques to address the challenges. In the last sections, we will discuss these approaches and give an outlook of future work. 2
User experience in recommender systems can be considerably improved using interactive visualizations [
Recommender systems lack transparency, when they appear as "black boxes" to the user, making it
incomprehensible how recommendations are generated and why a specific list of items is presented [
Further research focuses on the possibilities to influence the recommendation algorithms and increase the
controllability over the recommendation process by providing the user control over the user model and the
personalization [
A third aspect afects the explorability. The widely used presentation of results in the form of ranked lists
impedes getting an overview of the item space and hide relations to unrecommended items. Whilst in the field
of information retrieval and information visualization various methods exist for visualizing large dataset such
as document collections [
Furthermore, it is important to incorporate contextual information into the recommendation process in order
to recommend items to users in certain circumstances [
The application starts with a configuration view that allows adding new groups and editing existing groups by adding or removing users. As soon as all group members are chosen, each person is represented by his name and a unique color. Furthermore, diferent priorities can be assigned to selected group members. This might be necessary, when a family with children involved wants to choose a movie. Then the age and preferences of the children have a higher influence on the recommendation algorithm, excluding movies with certain age restrictions.
After the group is selected and adjusted, the recommendations are calculated based on the WARP (Weighted
Approximate-Rank Pairwise) algorithm [
The use case in this scenario is the search for activities while walking through a potentially unknown city. Mobile apps are perfect to provide recommendations for locations, events, or restaurants on the go. Every step of the recommendation process is supported by suitable visualizations that are explained in the next subsections. 4.1
Users first provide explicit feedback about their preferences using a dialog based card game comparable to the popular dating app "Tinder" (see Figure 2a). The following categories are provided in our prototype of the application, which can also be extended using open linked data: Food (e.g. Pizza, Pasta, Sushi), Drinks (e.g. Beer, 26 • Cocktails, Wine), Music (e.g. Electronic, Hip-Hop, Jazz, Rock), Culture (e.g. Museums, Sights, Theatre, Cinema), Activities (e.g. Dancing, Billiard, Dart).
Users interact with the main category cards by swiping left, right, or up. A swipe to the left is a dislike, which expresses little interest in a category. By swiping right, this category is added to the user interests as a like. Swiping up expresses a superlike, which means that this category is crucial to the user. Consequently, sub-categories are selected after performing a superlike in order to specify this interest in more detail. The dialog can be completed at any time using the "send" button. Otherwise, all main categories are presented to the user randomly. In the current prototype, the dialog must always be completed for each session. However, it is conceivable that the completed preference dialog can be saved and loaded each time a new session is started, e.g. in another city.
The current location of the user is automatically leveraged for the recommendations that are calculated based on the user preferences obtained in the initial dialog. Each location that is retrieved via external APIs such as Google Places, Yelp, or Facebook is ranked according to their related categories. A dislike yields a value of -1, no rating 0, like 1 and superlike 2. Since a location can have several categories, all values are summarized. Thus, locations that meet more user criteria receive a higher ranking. The prototype5 is implemented using Node.js6.
Recommendations are then visualized on a map as markers with numerical values for their ranking (see Figure 2b). User location is visualized as well in order to visualize the physical context (blue dot on the map). The top recommendation is visualized as a yellow marker. Common map interaction such as zooming and panning is supported.
By tapping on a marker, another view for the transparency and controllability of the recommendation results is displayed (see Figure 2c). In order to make the recommendation more transparent, a combination of a bubble chart and mind-map is used. Simultaneously, the map in the background is frozen and displayed with lower saturation to make the information visualization on top more prominent. Categories are displayed as bubbles in a radial grid around the selected location. Both the user preference and the location properties must be shown in this visualization. Hence, each bubble features two diferent radii. One radius shows the relevance to the user according to the initial preference dialog. The bubble is displayed with a full color inside this radius. The radius for the location is only displayed by an unfilled circle. Consequently, if the user preference is equal to the location value for this category, the bubble is displayed in full color. If the user preference was higher than the location relevance, the ring representing the location value is visible inside the bubble. If the location has a higher score, a ring is displayed around the user bubble. Tapping a category reveals the subcategories, which are displayed in the same way (see Figure 2d). On larger mobile devices or on desktop computers, it is also possible to display categories and subcategories simultaneously. On smaller screens, a semantic zoom can be used.
Controlling the recommendations is achieved by altering the weight of the selected categories. A tap-and-hold gesture on a node results in a continuous change of its size from small to large. Completing the tap-and-hold gesture results in a new calculation of the recommendations. Explorability is provided by panning the map to explore regions further away from the current location of the user or zooming into the map to show more recommendations in this specific area. Through changing the viewport, new recommendations are calculated for this specific area, hence controllabiliy is provided as well.
Most travel portals ofer the best deals for the desired holiday destination. But if the user does not know exactly,
where he or she can go and how to formulate a concrete search query with this vague information need, only
5video of the prototype: https://www.youtube.com/watch?v=O7is2SeSeuw, retrieved on 04.05.2018
6https://nodejs.org, retrieved on 04.05.2018
little guidance is provided. This approach focuses on travel search with vague information need and serves as
inspiration and comparison tool for diferent holiday destinations. The search approach allows a step-by-step
reduction of the result set by selecting visualized concepts such as "beach", "culture", and "relaxing". The introduced
visualization concept extends the interface "GetInspired" that uses a selection-based recommendation-driven
search, based on the principle of divide and conquer [
The interface is divided into three parts. The left-hand side provides nine diferent visual concepts that describe a holiday destination, such as "warm", "party" and "water sport" (see Figure 3, left part). Based on the selection of the user, nine new concepts are suggested that should lead to a desired result with the fewest number of refinement steps possible. The selection of a visual concept reduces the result set, which is visualized in the center of the interface (see Figure 3, middle). The interface element uses a sieve metaphor that is explained in Figure 4. Each visual concept or filter is represented by a horizontal line, sorted by their selection. The results are represented as vertical lines, which can be extended downwards. In the case that a filter criteria is not matching the particular result, the vertical line is stopped in this position. Hence, just the results that cross all horizontal lines, match all selected filters, which is comparable with a sieve. For this visualization concept, a radial layout is chosen, to use the given space eficiently and to avoid diferent interpretations that might be triggered through the position of the results. In the radial version (see Figure 3, middle), the filters are sorted from outside to the center of the circle. The results are represented as clockwise ordered lines. Through mouse-over interaction, the user can see previews of the holiday destination in the inner circle of the visualization. Interesting destinations can be collected in the right-hand side of the interface (see Figure 3, right part). This part uses a mood board metaphor and supports the organization and comparison of the collected results. A mood board is a common tool to collect pictures, materials and texts to give a general idea of a topic. Hence, the results are represented as circles. The outer circle serves as menu and the inner circle as preview to present diferent information about the holiday destination, such as prices and temperatures in the form of a graph and pictures. The elements can be t 1 2 3 4 5 6 7 8 9 10 11 12 F1 F2 F3 Result Set(F1) = {2, 4, 5, 7, 8, 10} Result Set(F2) = {5, 7, 8}
Result Set(F3) = {7} grouped by placing them close to each other, visualized by a surrounding border that connects these elements and also synchronizes the presented properties of each items. This allows the comparison of the holiday destinations, e.g. finding the cheapest time to travel according to the prices represented in the graph for each month. The prototype7 is implemented using jQuery8, D3js and Sylvester9. 5.2
The sieve metaphor afords explorability, since filtered results are not hidden from the user. Rather, he or she is able to hover over the results in order to get explanations about which filter is responsible for the exclusion from the current result set. Hence, transparency is supported with respect to the result sets, but not the filters that are influenced by the recommendation algorithm. By removing filters that exclude interesting results, the user is able to reformulate the search query. Hence the recommendation algorithm provides a new set of visual concepts to filter the result set. The inclusion and exclusion of visual elements gives controllability over the recommendation algorithm that is recalculated depending on the users’ interaction. There is also the possibility to move results to a trash bin in order to influence the recommended filters and results. 6
In this paper, we presented a taxonomy of visualization challenges for recommender systems. The presented interface concepts tackle these challenges by using specific building blocks. Figure 5 summarizes the interface building blocks of each usage scenario with regards to transparency, controllability, explorability, and specific aspects of context-awareness. In addition, the suitability of each approach with regard to diferent physical, social, media and modal context are assessed.
The taxonomy is suitable to reflect diferent solutions with regard to their support for user experience issues in recommender systems. Furthermore, an analysis of more visualization approaches with the help of this taxonomy can provide building blocks to tackle the visualization challenges that serve as an inspiration for future developments.
In terms of transparency, glyphs have the potential to show the impact of single preferences on the recommendation algorithms and allow the comparison of the user preferences with the current result set, shown in all three scenarios. Furthermore, the visualization techniques Radial View and Bubble Chart can support in analyzing and comparing the position of items in the ranking algorithm, whereas the visualization technique Sieve supports the user in understanding which preferences match or don’t match the viewed result. 7video of the prototype: https://www.youtube.com/watch?v=3LixVxYHGPI, retrieved on 04.05.2018 8https://jquery.com, retrieved on 04.05.2018 9http://sylvester.jcoglan.com, retrieved on 04.05.2018
Challenges Transparency Controllability Explorability
Physical Social Media
Modal Context-Awareness
Color Coded Users Group Selection Zoom into Cluster
Group
Controllability can be applied by adapting the user preferences, illustrated by the tap-and-hold gesture in the activity scenario and the selection and excluding of results in the travel scenario. Also, the feedback on the presented result can ofer controllability by collecting interesting or exclude uninteresting results that can influence the recommender algorithm as well, what is shown in the mood board of the travel scenario.
Explorability was also ofered in all scenarios. The movie scenario ofers a navigation concept that allows the exploration of diferent regions of interest. The activities scenario supports the well-known interaction techniques panning and zooming to explore diferent regions of the map. The travel scenario supports a visualization concept that highlights the suggested results but is still showing the filtered items in the context for further exploration and analysis.
The fourth row in the table is related to context-awareness and presents interface building blocks that show context-specific information. In the movie scenario, diferent users are represented by a color-coded glyph segment, whereas the activities scenario presents the user position with a map marker.
The lower part of the table addresses the aspects from Adomavicius et al.[
Our three demo applications address the modal context with regards to the goals of the user only in a static way. So far, no adaptations of the interfaces are ofered if an information need changes from vague to concrete. Both movie and travel scenarios support a vague information need that is mostly associated with ofering extensive explorability support in the application. In contrast, the activities scenario is focused on concrete information, which is addressed by the controllability features of the application and the initial preference dialog. 30 • 7
From the experience in designing and implementing our demo applications, all three main challenges can be addressed in a given application without sacrificing either transparency, controllability, or explorability. However, there is a rather obvious need for more screen real estate when trying to support explorability or transparency in an interface. Hence, small devices in mobile scenarios usually provide weaker support in these aspects.
The proposed taxonomy is a work-in-progress and should be extended with more building blocks in the future
and applied to more solutions. So far, our taxonomy can be seen as a subset of the taxonomy by [
The usage scenarios in this contribution have been developed in a student workshop (Movie scenario: Kristian Kyas and Elisabeth Baudisch, Activities scenario: Julian Haluska and Finn Schlenk) and a diploma thesis (Travel scenario: Marcus Kirsch).