Bias-Aware Design for Informed Decisions: Raising Awareness of Self-Selection Bias in User Ratings and Reviews

It has become common for people to make decisions based on others’ opinions posted online, such as buying a product or booking a hotel online. However, it takes a lot of time for people to dig through large numbers of online reviews  (De Maeyer, 2012). Existing works in the natural language processing (NLP) domain proposed automatic algorithms to extract representative information and/or general user attitudes from reviews, such as generating summaries of reviews  (Suhara et al., 2020; Angelidis et al., 2021; Tsai et al., 2020) and classifying emotions in reviews by sentiment analysis  (Binder et al., 2019; He et al., 2017; Boiy et al., 2007). Recent studies tried to define the features of helpful reviews  (Du, 2020; Yadollahi et al., 2017) by mining the factors related to user satisfaction  (Lee et al., 2019) or analyzing the helpfulness by combining reviews and quantitative ratings jointly  (Chatterjee, 2020).

Based on these automated approaches, another body of work offered a visual interface or an interactive system for users to dig through online reviews in a more detailed manner  (Bjørkelund et al., 2012; Chen et al., 2015; Wu et al., 2010). They tend to present user reviews from multiple semantic perspectives and assist users in analyzing reviews from a higher level aspect. For example, Chang et al. visualized aspect-level data in hotel reviews for users to gain insights and analyze different types of customers  (Chang et al., 2019). ExtremeReader generates abstract visual summaries by offering a high-level structure of opinions in user reviews  (Wang et al., 2020). Considering the subjective aspects of user reviews, recent work designed an interactive visualization system with multiple views for domain experts to mine people’s opinions  (Zhang et al., 2020). These works provide users more space to explore and customize insights from user reviews. However, most of these systems or interfaces are complex for ordinary users and they are often designed for experts to analyze large-scale data.

The above works mainly focus on how to efficiently extract and represent user reviews, while they do not give much consideration to the potential bias hidden behind them. Considering the pervasive subjectivity in online reviews  (Halevy, 2019; Li et al., 2019), there is a strong need to reduce the impact of the self-selection bias on people’s decision-making and help them make informed decisions. In this work, we leverage automatic approaches to analyze and extract the critical aspects of user ratings and reviews, and integrate the output into a bias-aware design for raising people’s awareness.

Previous researchers have indicated that user ratings and reviews are biased, not only caused by the companies (i.e., the organization/person who can manipulate ratings or reviews), but also by people who gave feedback  (Aral, 2014; Karaman, ; Cicognani et al., 2016). One typical type of bias caused by people is the self-selection bias, which leads to more extreme positive and negative feedback in user ratings/reviews. While there are several other types of bias caused by people or platform interventions, like anchoring bias, social influence bias, or rating bias by algorithms in the online community  (Mayzlin et al., 2014b; Chevalier et al., 2018; Thebault-Spieker et al., 2017), in this paper, we mainly focus on the self-selection bias in user ratings and reviews  (Bhole and Hanna, 2017; Hu et al., 2009a, b; Karaman, ; Sterne et al., 2008).

To improve the representativeness of user feedback, researchers in the field of business and marketing designed different strategies to mitigate biases in data  (Aköz et al., 2020; Lim and Tucker, 2017; Wu et al., 2017b; Nagtegaal et al., 2020b). These strategies include (1) sending emails to a random selection of users and encouraging them to write reviews  (Askalidis et al., 2017; Karaman, ), (2) offering a relative comprehensive framework for users to give feedback (e.g., commenting on the pros and cons of a subject separately)  (Lim and Tucker, 2017; Wu et al., 2017b), and (3) selectively displaying a representative user feedback online by manipulating the orders  (Eslami et al., 2017). However, these approaches are primarily designed for businesses with the aim of maintaining the reputation of a platform. They paid little attention to the end-users who refer to user ratings or reviews to make decisions. Additionally, biases may also be brought by such strategies, as their processes may include financial or social incentives  (Karaman, ).

Some other researchers proposed technical methods to handle biases in user ratings and reviews  (Zheng et al., 2021a; Sikora and Chauhan, 2011; Zhang et al., 2019; Calmon et al., 2017b; Barrett et al., 2019; Walker and Buttinger, 2017). These works aim to detect  (Calmon et al., 2017b; Walker and Buttinger, 2017) or model  (Zheng et al., 2021a; Sikora and Chauhan, 2011; Zhang et al., 2019) biases in online ratings or reviews to mitigate the effects caused by biases. For instance, a recent work by Zheng et al.  (Zheng et al., 2021a) identified and modeled biased user ratings based on textual reviews using deep learning models. Riyaz and Kriti used the Kalman filtering technique to estimate the sequential bias in user reviews  (Sikora and Chauhan, 2011). There are also plenty of works that focus on building or designing unbiased algorithms in recommendation systems  (Joachims et al., 2017; Schnabel et al., 2016) or rating systems  (Bishop, 2015; Hickey, 2015), which tend to deal with biases in algorithms rather than data. Although these approaches are proven to be useful to some extent, they still suffer issues of scalability, and they cannot reduce the bias completely due to the diversity of bias as well as the difficulty of designing unbiased algorithms  (Zheng et al., 2021b). Furthermore, they keep biases detected by the back-end algorithms that are opaque to end-users, and thus may cause issues of trust and explainability  (Damak et al., 2021).

The self-selection bias in data are difficult to eliminate completely by automatic algorithms or data management strategies because it is caused by humans’ self-selected behaviors. Additionally, the previous approaches typically performed operations that were invisible to the end-users and thus cannot help them to realize the potential bias and thus directly aid their decision making process. Therefore, it is critical to make people informed of the potential biased data while making decisions with user rating and reviews. In this work, we aim to raise people’s awareness of the self-selection bias by putting the decision in the hands of users and reduce the risk of users being influenced by biased data rather than directly addressing biases.

Transparency can be used as a mechanism to help users make informed decisions with systems, and it usually comes with two goals: showing how and what  (Rader et al., 2018; Zhang and Chen, 2020). For the “how” situation, transparency can be used to make the process of how data is processed and analyzed visible to users. For example, in some recommendation systems, it is easier for users to make sense of the underlying algorithms when the process recommendation is shown  (Zanker and Ninaus, 2010; Gedikli et al., 2014). Transparency can greatly improve users’ perception of the credibility and usefulness of a system  (Wang and Benbasat, 2007), and empower the decision-making processes  (Eslami et al., 2019; Diakopoulos, 2016). The “what” type of transparency deals with the outcome of a system, revealing hidden data (if any) or displaying the reasons for the output  (Zhang and Chen, 2020). Researchers improved transparency by providing supplemental information, such as visualization or explanation, to help decipher the existing results. Typical examples are post-hoc explanations that help users to understand the information provided by algorithms  (Costa et al., 2018; Park et al., 2017). Facing with transparent information provide, users trust the system more and become more satisfied about their decisions  (Cramer et al., 2008). In this paper, we mainly focus on the second form of transparency (“what”) and explore how to use it to increase people’s awareness of the self-selection bias in user-generated data.

Transparent design of data plays a vital role in users’ decision-making in the area of visualization  (Bertino et al., 2019). Similar designs for users to track and analyze data include the Google Dashboard  (Google, 2009) and Mozilla’s Lightbeam  (Mozilla, 2013). Similar works have explored ways to give users back control of their data through tools that enhance transparency, as summarized in  (Janic et al., 2013). There also exist many works that leverage visualization to provide users with transparency about particular aspects of the data in different scenarios  (Angulo et al., 2015; Ghoniem et al., 2004; Kolter et al., 2010). For instance, Zavou et al. informed users via a “chord diagram” visualization that helps users understand how their data is treated by third-part cloud-hosted services  (Zavou et al., 2013).

In addition to enhancing data understanding and informing user decisions, transparency is also beneficial for raising users’ awareness of the hidden patterns behind data. A previous work pointed out a critical need for increased data transparency because the layperson might not be fully aware of how data aggregation was done by third-party services  (Rader, 2014). Some other works emphasized that users should be aware of how systems or algorithms process user data and called for higher transparency  (Rader et al., 2018; Eslami et al., 2017). Transparent features were also used to improve users’ awareness of their behaviors  (Stevens et al., 2018) and personal privacy issues  (Ebert et al., 2021; Kani-Zabihi and Helmhout, 2011). Visualization can be used to promote transparency and raise people’s awareness of biases for informed decisions. The latest work by Arpit et al. proposed to raise users’ awareness of their biased behaviors during data analysis for decision-making by showing the interaction traces  (Narechania et al., 2021).

In this paper, we set the goal of raising people’s awareness of the self-selection bias in user ratings and reviews for informed decisions. We propose a bias-aware design based on user ratings to reduce the impact of self-selection biases on people’s decisions.

In this work, we target the self-selection bias in user-generated data: people with extreme experience (i.e., positive or negative experience) are more likely to give their feedback online than those who have a moderate experience  (Bhole and Hanna, 2017; Li and Hitt, 2008). Online ratings and reviews widely suffer from the self-selection bias  (Aral, 2014; Baeza-Yates, 2018) and thus often fail to provide information about the quality of products/services that represents the general opinions of the entire user base (Li and Hitt, 2008; De Langhe et al., 2016).

The self-selection bias is mainly caused by people’s self-selected behaviors  (Li and Hitt, 2008; Bhole and Hanna, 2017) that can be unconscious or triggered by people’s fast, instinctive processing system  (Kahneman, 2011; Evans and Frankish, 2009). For example, people could self-select to report or not report based on their experience (extreme or not extreme), and choose to merely report the aspects that leave the strongest impression. Such kind of self-selection behavior is part of human nature  (Demartini et al., 2021), thus, the resulting bias in user-generated data is hard to prevent and remove completely by algorithms or data management  (Hettiachchi et al., 2021).

Furthermore, later users – who refer to the ratings/reviews – may found their personal experiences with the products/services inconsistent with their expectations established based on existing user reviews, as the biased feedback might not give a complete, up-to-date picture. The big expectation-experience disparity may cause a vicious circle of extreme feedback online, which hinders people’s decision-making process  (Askalidis et al., 2017).

Hence, it is critical to reduce the chance of people being “victims” of the self-selection bias when referring to online ratings/reviews for their decisions. Raising people’s awareness of the bias can guide us toward a solution to deal with biases in user-generated data  (Baeza-Yates, 2018; Demartini et al., 2021; Hettiachchi et al., 2021). In the next subsection, we introduce how we raise people’s awareness of the bias by making them realize that certain information is over-represented in the sample of user reviews.

We aimed to raise people’s awareness of the self-selection bias by making three types of information that potentially reflect self-selection bias transparent in user ratings/reviews. We distilled three types of information by reviewing the literature  (Li and Hitt, 2008; Askalidis et al., 2017; Gong et al., 2015; Baeza-Yates, 2018; Bhole and Hanna, 2017; Yadollahi et al., 2017; Karaman, ; Halevy, 2019): (I1) the distribution of reviewers with difference experience who choose to share their opinions  (Baeza-Yates, 2018; Gong et al., 2015), (I2) the distribution of emotion extremity  (Bhole and Hanna, 2017; Askalidis et al., 2017; Karaman, ), and (I3) the distribution of reported aspect(s) in user reviews  (Li and Hitt, 2008; Halevy, 2019). We introduce these pieces of information and why they need to be presented to users in details below.

• •

  We concerned the composition of reviewers (feedback providers) because those who leave comments online often cannot represent the entire population of people who have tried the products/services  (Gong et al., 2015). There exists the “silent majority” who select not to give ratings or reviews on the Internet  (Baeza-Yates, 2018). Thus, we need to keep readers of online ratings/reviews informed of the characteristics of the reviewers.

• •

  We cared about the emotion extremity of feedback because people tend to speak out when they have very strong feelings about something, which only constitutes a relatively small portion of the experiences of the majority  (Bhole and Hanna, 2017; Li and Hitt, 2008). Thus, it is necessary to make users aware of the polarity and intensity of emotions behind the user ratings/reviews.

• •

  We considered the reported aspect(s) of user reviews because people can self-select to give feedback based on their own preferences  (Halevy, 2019). They can be more sensitive to or impressed by one or several specific aspects of a product/service and write reviews (or give ratings) accordingly rather than equally weighting all aspects  (Li and Hitt, 2008). Showing the categories of aspect(s) that formed the basis of current user ratings/reviews could facilitate new users weight the feedback based on their own needs and interests.

Transparency can be an efficient way to raise people’s awareness of the biased information from the perspective of how users perceive user ratings and reviews  (Demartini et al., 2021). Before considering how to make the information transparent to raise awareness of the bias, we need to know exactly (1) how people perceive the three types of information and (2) what are the specific facets they would like to view about these three kinds of information when making decisions online. We conducted a formative study to collect users’ opinions for these questions to inform the bias-aware design.

We conducted an online survey with 206 participants who had booked a hotel online in the past two years. We performed the survey via Prolific³³3https://www.prolific.co/, which is a crowd-sourcing platform designed specifically for online academic research. Prolific has been shown to provide better quality data from diverse participants than other platforms  (Peer et al., 2017).

We firstly set up a screening session to find participants who have experience in booking hotels online in the last two years. There are 273 of 400 participants who passed the screening session and 219 of them chose to participate in the formal survey. We collected 206 valid responses after excluding the participants who finished the survey in an extremely short time (¡2 minutes).

The content of the survey can be divided into three parts.

1. (1)

   Participants first answered the questions about demographics and their habits of booking hotels, such as reasons of booking hotels, platforms they used, etc. We collected their habits as the guide for the later development of prototype systems with the bias-aware design.

2. (2)

   Then, we collected participants’ perceptions of the three types of information (I1, I2, I3) from two perspectives: how do they perceive the information and to what extent do they make decisions based on the information when booking a hotel. To do so, we provided statements for each type of information and let participants to rate their agreement or disagreement on a scale of 1 to 7 (from strongly disagree to strongly agree).
   For example, the statements for the extremity of emotion (I2) are expressed like: “The extreme positive ratings/reviews are accurate”, “The extremely positive ratings/reviews affect your likelihood of choosing a hotel”, etc. The word “positive” can be replaced by “negative” or “moderate” for different statements. To get the comprehensive perceptions, we also provide statements using different words beyond “accurate”, such as “trustworthy”, “reflect people’s opinions”, “reflect the quality of hotels” or “reflect different aspects of hotels”.

3. (3)

   In the third part, we collected the specific information participants care about when booking hotels online. We set questions like “How did you filter reviews when you book a hotel online?”, “What kind of information of the reviewers you would like to view?”, etc.

All questions are compulsory for participants and they need to return a complete code of Prolific after finishing the survey.

We controlled the acceptance rate (¿95%) of all participants in Prolific. Each participant was paid $2 for the survey that took eight minutes on average ($MAX=11,MIN=6$). These participants are all US-based (81 females, 122 males, 3 preferred not to disclose) with a mean age of 34 years ($MAX=19,MIN=71$). They came from different areas of the US (West 29.61%, Northeast 23.79%, Southeast 22.33%, Midwest 15.53%, Southwest 8.74%). 52.43% of them were full-time employed, 17.96% were students, 10.19% were employed part-time, 6.80% were unemployed, 4.85% were self-employed, 2.43% were retired, and 5.34% selected “other”. Participants selected their highest educational level from the following options: high school degree (7.77%), the college without a degree (17.96%), associate degree (7.28%), bachelor’s degree (41.26%), master’s degree (20.39%), professional degree (2.43%), and doctorate (2.91%).

DR1: Disclose the distribution of the information behind user ratings. The current design of user ratings online merely provides the average rating and the number of different ratings (e.g., number of 1-5 stars). Therefore, it is hard for people to realize the self-selection bias hidden under user ratings and reviews with the current design. According to Section 3 and Section 4, we aimed to disclose the distribution of reviewers’ rating experience (I1), the extreme level of emotions (I2), and different reported aspects (I3) in user reviews to raise people’s awareness of the self-selection bias. We discuss the details on how to extract these three types of information in Section  6.3. Making people informed of how these aspects are distributed in user reviews can reduce the influence of bias on their decisions.

DR2: Use simple and common visualizations for lay individuals to support their decisions. The target audience of our work is people who refer to user ratings/reviews online to make decisions, and they usually have no expertise in data visualization. Hence, the design should be simple and make the best use of common visual representations that people commonly encounter in their daily lives. Additionally, we considered the practicality and generality of the bias-aware design by discussing with two visualization experts, as it aims to support people’s daily online decisions. Therefore, we proposed three design alternatives based on the conventional design of user ratings (i.e., bar chart) rather than creating an entirely new design  (Peck et al., 2019).

DR3: Enable filtering functions with the transparent information for exploring user reviews. When making purchase decisions, people spend the majority of their time exploring user reviews to mine suitable opinions for their personal decisions  (Wang et al., 2020). According to the formative study, the information we aimed to disclose in the bias-aware design is also critical for people’s decisions. Thus, providing users with filtering functions based on transparent information is necessary to facilitate efficient decision-making.

As shown in Fig.  2, we present the final design of the transparent information with user ratings. The design combines pie charts with user ratings that clearly show the distribution of information (e.g.,  e1 –  e5) vertically in user reviews behind each bar (DR1). It shows the categories across user ratings horizontally, leveraging the interactive design in Fig.  2(A3). The pie chart is a conventional chart showing distributed data, which is familiar to people in daily life (DR2). The legend on the right-hand side of the design shows the color scale associated with each category (i.e.,  e1 –  e5) of the information (i.e., the extremity of emotion). To coordinate the proportions of elements in the design and improve space utilization, we keep all pie charts the same size and use the thickness of the (light grey) lines linking pies and bars to represent the number of ratings/reviews for each pie.

We designed three interactive functions for the design in Fig.  2 according to the participants’ requirements in the pilot study. These functions could help users compare the detailed information with the design (DR4) while offering filtering functions to filter user reviews in a specific aspect (DR3). In Fig.  2 (A1), we added a zoom effect on each sector of a pie chart so that users can easily hover over them to view the details of each category in a floating text box and filter user reviews by clicking a sector. Similarly, in Fig.  2 (A2), users can check the information of all categories under a rating by hovering on the bar and clicking to filter reviews by a rating. To meet the DR4, we designed a mouse-over triggered animation with the legend. By hovering on the legend, users can view the distribution across different ratings horizontally related to one category (Fig.  2 (A3)). The corresponding sectors of the selected category (e.g.,  e1) in different pie charts will also be highlighted. In this way, users can easily check the distribution of one category across different user ratings. We added a smooth transition animation to make the changes between bar charts more natural. Moreover, users can also filter reviews by clicking the categories in the legend, such as filtering reviews that express extreme negative emotions (i.e., clicking  e1).

To evaluate our design effectively, we aimed to collect data from representative hotels online, which could mirror the opinions of ordinary users. Thus, we decided to select hotels from big cities in the world. Considering that participants of the formative study were all US-based, we retained this character in the user study. Hence, to reduce the impact of their previous impressions, we collected data from non-US cities that participants might not be familiar with. Two authors explored and compared the hotels’ data in different cities. Considering that the COVID-19 may influence the user ratings and reviews of hotels online, we only collected user ratings/reviews in the time interval of 30 June, 2016 to 31 January, 2020 (before COVID-19 became International Concern ⁶⁶6The World Health Organization declared the outbreak a Public Health Emergency of International Concern on 30 January 2020, and a pandemic on 11 March 2020.). Indeed, the pandemic is likely to change demographics of travelers, their emotions, and the aspects they concern about a hotel. It would be interesting to explore and compare the changes of content and the potential biases in the content of user feedback before and after COVID-19. We leave this topic as part of our future work, because COVID-19 made travel extremely difficult, if not impossible, in most parts of the world when we conducted the experiments. Finally, we chose London as it had the most hotels as well as user reviews.

We initially crawled 976 hotels’ pages under the tag “London” on Tripadvisor. To narrow down the scope and select representative hotels, we narrowed down the scope of hotels to 3-star hotels labelled by Tripadvisor as hotel-class because there are more 3-star hotels (40.7%) than others. These hotels had a price range from $40 to $200, with $88 as the median price, and we found that hotels with price ranges from $82 to $105 contained more than 80% of the user ratings/reviews within the setting time interval. Therefore, we selected hotels based on this price range and filtered out the hotels with no feedback for the last six months. Finally, we got 57 hotels and crawled the data of these hotels, including the hotels’ names, price, user ratings, reviews, etc. By controlling these variables, we could select the representative hotels and let users focus more on the user feedback to evaluate the bias-aware design effectively.

After selecting the representative hotels by control variables, such as prices and the hotel-class, we explored the distribution shapes of user ratings in these hotels as the distribution can be affected by the self-selection bias proven by previous works  (Schoenmueller et al., 2018; Hu et al., 2009b). We analyzed the user rating data of the 57 hotels and summarized three kinds of distribution in Fig.  3.

These three distributions include (1) the monotonic increasing distribution, (2) the J-shaped distribution, and (3) the positively skewed distribution, which are observed from the selected hotels’ ratings on Tripadvisor. (1) The monotonically increasing distribution means that the shape of the user rating bars is monotonically increasing from 1 point to 5 points. (2) The J-shaped distribution is the typical distribution caused by self-selection bias as it has an extreme distribution with more 1-point and 5-point ratings  (Hu et al., 2009b; Bhole and Hanna, 2017). (3) The positively skewed distribution has more middle points (from 2 to 4 points) with fewer 5-point ratings compared to the other two distributions. This distribution is close to real-world data under lab conditions  (Lim and Tucker, 2017).

We randomly choose 15 hotels from 57 hotels with five for each distribution shape, considering that the distribution shape may affect people’s perceptions of hotels, and their decisions and people usually narrow down to no more than 15 hotels for decision-making based on the pilot study. We further extracted and analyzed the user ratings and reviews for the selected 15 hotels, which is illustrated in the next subsection.

The data processing step consists of three parts corresponding to the three kinds of information mentioned in Section  4. We collected 5940 user ratings and reviews from the 15 hotels (MEAN = 368, MAX = 397, MIN = 320) in total between 30 June 2016 to 31 January 2020. We first obtained the reviewers’ information (I1) based on the reviewer badge on Tripadvisor. Then we got the information about emotion (I2) and reported aspects (I3) using two automatic algorithms in the natural language processing (NLP) domain, including sentiment analysis and topic extraction. Based on these approaches, we mapped the information (I1-I3) into the bias-aware design.

To evaluate the design with ordinary users, we simulated a hotel booking scenario by building a prototype website integrating the bias-aware design. To increase the scalability and user friendliness of the prototype, we implemented the basic functions of booking hotels and used a responsive framework that makes the interface adaptive to different screen sizes.

We first collected the interface design of many popular hotel booking websites on the Internet, such as Tripadvisor.com, Booking.com, Expedia.com, etc. We analyzed their interfaces and found that these websites use a similar framework of user interfaces, which shows the general information of each hotel in a card view and displays the detailed user reviews in a separate web page or a pop-up page. We adopted this design and integrated the bias-aware design into the user rating position in the detailed page of each hotel.

We show the two main interfaces of the prototype in Fig.  4, in which (A) shows the homepage with the introduction of the user study and a list of hotels. By clicking the button in a card view, users can open a new page in a pop-up window that displays the corresponding user ratings and reviews of one hotel in Fig. 4 (B). We display the bias-aware design with a navigation bar above, which is used to switch between different information types (Fig. 4 (B1)(B2)(B3)). There are four options in the navigation bar that include two types of rating experience of reviewers (i.e., written reviews (I1-1) and “helpful” votes (I1-2)), emotion (I2) and aspects (I3). Users can interact with the design to filter user reviews below or check the distribution of one category across all different user ratings, as illustrated in Section  5.3.

We display the user reviews below the bias-aware design and show each review together with the user name, time, reviewer, user rating, and reported aspects, referring to the design of Tripadvisor. Users can scroll up and down to read user reviews in this page. To avoid gender bias, we replaced all the original reviewers’ names with either abbreviations or random common names for both males and females.

We used the Prolific to conduct the user study online. To ensure that all participants had the experience of booking hotels online prior to our study, we set screening questions on the frequency and most recent activities of booking hotels. We initially recruited 144 participants and excluded the responses of eight participants who failed this quality control during the study.

We obtained responses from 136 participants after the quality check, of whom 68 (33 females, 34 males, and one prefer not to say) completed the experiment with our proposed prototype system, and the other 68 (35 females and 33 males) used the baseline system. The distribution of our participants’ education background is as follows: undergraduate (46.3%), graduate (29.4%), technical/community college (8.1%), high school diploma (8.8%), doctorate degree (5.9%) and unknown (1.5%). The average age of the participants was 31 (Max= 73, Min = 20, SD = 9.97, 95% CI [29.14, 32.49]). Following the rules of the Prolific platform, we paid all participants at a rate of £7.60/hour. The actual amount for each individual depended on the task completion time and the Prolific settings according to our predefined time range (30 min).

Before launching the formal experiment, we conducted a pilot study to test our task design and configuration with eight participants (four females), who were undergraduate or postgraduate students recruited from a local university. Their ages ranged from 21 to 27 (Mean = 24, SD = 2.51, 95% CI [21.90, 26.10]), and their backgrounds included computer science, finance, psychology, and biology. These participants were randomly divided into two groups; half completed the task with our prototype, while the rest used the baseline system. After finishing the task, they filled out a post-study questionnaire and participated in a brief interview to share their study experiences.

Based on the feedback from the pilot study, we adjusted several designs in both conditions. Seven out of eight participants (87.5%) reported that they usually inspect and compare fewer than 10 hotels in detail in real life, because it became overwhelming and time-consuming beyond that number. Examining 15 hotels fully deviated from their common practice and they were not able to remember and compare the information of all these hotels at once. Thus, we reduced the number of hotels in the candidate pool from 15 to 9 for our formal study by randomly removing two hotels from each distribution type mentioned in Fig.  3. Furthermore, as informed by the pilot study, we further streamlined our study procedure (presented in the next subsection) and enriched the details of the instruction videos.

In our formal study, 144 participants were randomly split between the baseline group and the experimental group. All participants first signed a consent form, agreeing to join the experiment and allowing us to collect basic information about them. We recorded their operations during our study and obtained their demographic data from the Prolific platform. Then, following the three steps presented in the task flow (Section  7.2.3), each participant watched an introduction video, browsed the candidate hotels, specified the top three choices, wrote down the reasons for the choice, and submitted in the prototype system. After sending their selections and reasons, each participant filled out a post-study questionnaire, which assessed their awareness of potential bias and collected qualitative feedback on our design and systems (RQ1). The average time for the whole study was 34 minutes (baseline: 29 minutes, prototype: 39 minutes). We did not control the time of our user study because we aimed to simulate the real-world scenario of booking hotels so that the study results could truly reflect people’s decisions and their decision-making process. We concerned that strictly controlling the time may add stress to the participants and affect their behavior during the experiment and consequently the observed results.

We measured people’s awareness of the self-selection bias in user ratings and reviews by the post-study questionnaire. Note that the self-selection bias is implicit, which needs to be measured carefully without revealing the goal of our study (testing awareness) during the experiment. To effectively test whether the bias-aware design works or not, we instructed each participant to book a hotel using the given prototype system under a concrete scenario (going to London on vacation) as how they would do it in the real life. We followed a psychological method called implicit association test (IAT)  (Nosek et al., 2005) and adapted the questionnaire of a famous IAT project called Project Implicit⁹⁹9https://implicit.harvard.edu/implicit/education.html to design our questions for accessing awareness. IAT is widely used to assess people’s implicit bias, stemming from attitudes or stereotypes that mostly occur outside of people’s consciousness and control  (Nosek et al., 2005).

As shown in Fig.  6, we present our five adapted IAT questions (Q1–Q5) with the corresponding participant ratings that measure their overall awareness (Q1) and the awareness of the three types of information (Q2-Q3: emotion [I2], Q4: aspects [I3], Q5: reviewers [I1]) caused by the self-selection bias. We used the Mann-Whitney test (McKnight and Najab, 2010) to compare the result of both conditions and calculated the 95% bootstrap confidence intervals for each question (DiCiccio and Efron, 1996). We display the responses of other post-study questions (Q6-Q12) that evaluate the information provided by the transparent design as well as its usability on the right side of Fig.  6.

The results of Q1–Q5 show that people in the experimental condition disagreed more with the descriptions, which indicated that they were significantly more aware of the self-selection bias in user ratings and reviews. Q1 measures the general awareness of participants between two conditions, which indicates that people booked hotels with our design were more aware of the diverse information behind user ratings. Among these cases, participants were more aware of the diverse experience of reviewers behind user ratings (Q5, p-value ¡ 0.001) more than other cases using the transparent design. This result echoes the last strategy used by participants when selecting hotels in Section  8.2. For the responses of Q4, participants’ awareness of the different aspects of a hotel behind user ratings (p-value = 0.0247) had the least difference between the two conditions. The reason is that participants using the baseline could also view the tags of diverse aspects for filtering reviews under the user rating as the ordinary hotel booking websites. Q2 and Q3 evaluate the level of participants’ awareness of the extreme emotions behind user ratings, including the positive and negative emotions. Participants who used the transparent design were relatively more aware of the negative emotions (Q3, p-value ¡ 0.001) than positive ones (Q2, p-value ¡ 0.01).

We collected participants’ feedback on the information and usability of our design (on the right side of Fig. 6) and found that our design was perceived to be easy to understand (Q6, Mean = 5.28, SD = 1.51, 95% CI [4.92, 5.64]) and interact with (Q7, Mean = 5.96, SD = 1.30, 95% CI [5.64, 4.26]). In general, participants would like to recommend our design to other people (Q8, Mean = 5.46 , SD = 1.48, 95% CI [5.12, 5.82]). We also asked the participants who used the transparent design about the helpfulness of the information provided (Q9-Q12). Overall, participants agreed that showing the proportion of aspects (Q12, Mean = 5.24, SD = 1.63, 95% CI [4.84, 5.61]) and emotions (Q9, Mean = 5.24, SD = 1.62, 95% CI [4.86, 5.64]) behind user ratings is useful for their decision-making. Showing the information about the reviewers, such as the number of reviews they have published (Q10, Mean = 4.75, SD = 1.89, 95% CI [4.29, 5.18]) and the number of “helpful” votes those reviews received (Q11, Mean = 4.34, SD = 1.81, 95% CI [3.90, 4.75]) are deemed relatively less useful than other information. Several participants reflected that they got used to filtering reviews by positive or negative aspects rather than using the information related to the reviewers.

Above all, we observed that participants are more aware of the potential bias after viewing data distribution behind user ratings with the help of the bias-aware design than those in the baseline condition. They were more aware of the difference of reviewers’ experiences than other kinds of information based on the responses of questionnaire. In addition, participants who experienced our design gave positive feedback on its usability and usefulness. Most participants in the experimental group reported that the distribution of extreme emotions behind user ratings was the most useful information for them to make decisions.

To understand how people leverage our design to make decisions, we asked participants to share their strategies for selecting, eliminating and ranking hotels in the post-questionnaire. Together with the uploaded reasons in the systems (Fig.  5 (D)), two authors coded the strategies using thematic analysis  (Braun and Clarke, 2006) and finally reached consensus through two rounds of discussion. In general, we found that participants in both conditions share similar strategies of choosing hotels while they put different emphasis on several aspects. We extracted five typical strategies from their feedback and summarized them in Table  1. We calculated the percentage of every strategy among all participants in each condition; four out of the five strategies occurred in both. We derive and report four salient findings regarding these strategies below.

We summarize the participants’ final selections of hotels in Fig.  7 (A). The bar chart shows the percentage of each hotel selected in both conditions respectively. We grouped the hotels by the distribution type of user ratings (Fig.  3) and compared the participants’ selections in the two conditions. We distilled several findings from participants’ selection reasons by coding and analyzing the reasons provided by the participants using thematic analysis  (Braun and Clarke, 2006).

We decided the exact transparent information by a formative study that collected the critical information people care about and people’s perceptions of the transparent information. In the study, we showed the information combining user ratings with reviews to people and explore how they exploit the information and make decisions. The results indicate that the information provided are helpful for decision-making in general (Fig.  6) while changing people’s behaviors and their strategies of referring user ratings and reviews for decision-making (Section  8.2).

However, it is necessary to mention that people in the group with the bias-aware design spend 10 more minutes on average than the baseline in the study. The reason was that, on the one hand, the experimental group needed to watch 1 more minute of the introduction video to understand our design and task before the study (Section  7.4), and they answered more questions in the post-questionnaire. On the other hand, the operation logs shows that they interacted more with the bias-aware design than the baseline (Fig.  7). Hence, how much and the granularity of transparent information to be shown are all needed to be considered. If we disclose too much information, users may abandon checking the details as they can be overwhelmed. For example, while other aspects of user ratings and reviews may be biased (e.g., reviewers’ gender and ethnic group), the formative study results suggest that they are not the key concerns of most crowd respondents during hotel selection. Rather than presenting all potentially biased dimensions of the user-generated data, we only emphasize on the most salient ones that have strong impact on people’s decisions.

In addition, the process of making information transparent is not just about the information itself. It also concerns how to present the information that was previously “invisible” to users without overloading users. It is thus necessary to consider and reflect the information needs of users in the design. We investigated how to present transparent information intuitively based on the formative study and the discussion with two visualization experts. Then, we proposed three alternatives and compared how people perceive and explain the information in these designs. While some of the alternatives could better support visual analytic tasks, they are more visually complex, more challenging to interact with, more space-consuming, and harder to scale. Based on user feedback in the pilot study, we selected the one design that is the most acceptable to our target users (i.e., lay public) and improved it by adding user-friendly interactions. This process ensured that at least most ordinary users could easily learn to use our design and grasp the conveyed information quickly and freely in their decision-making process. In general, designs related to information transparency should always consider the needs and characteristics of users and strike a balance between information transparency and information complexity.

Certain biases in user-contributed contents (e.g., imbalance in gender) could be alleviated by applying a better sampling technique proactively, such as the methods used by previous works (Aköz et al., 2020; Lim and Tucker, 2017; Wu et al., 2017b; Nagtegaal et al., 2020b; Askalidis et al., 2017), as discussed in the Related Work (Section  2.2). However, some other types of biases in user-generated data are caused by unconsciousness, such as people’s extreme emotions  (Schoenmüller et al., 2019), their knowledge/expertise  (Halevy, 2019), and selection of individuals  (Bhole and Hanna, 2017). They tend to be more implicit and may not be easily detected, measured, and mitigated by algorithms. Therefore, in this work, we chose one kind of implicit bias (i.e., self-selection bias) and proposed to raise people’s awareness of the bias as the goal of our work  (Baeza-Yates, 2018), on the assumption that such biases could not be technically eliminated from the data.

One challenge we met in this work is the evaluation of the people’s awareness, as it is not proper to directly ask whether they are aware of the bias or not. A previous study by psychologists has shown that people tend to appear knowledgeable when their awareness is explicitly tested  (Srivastava, 2016). We explored the literature and found that the way of evaluating the awareness should be associated with the concrete task and goal. For example, a recent work by CSCW researchers used the reliability scores in the pre- and post-survey to measure the awareness of the echo chamber effect  (Jeon et al., 2021). Another work by Eslami et al.  (Eslami et al., 2017) tested users’ awareness of the bias in algorithms by labeling whether they could articulated a discrepancy between their intended review score and the system output. We designed the questions in the post-questionnaire based on the implicit association test (IAT) in psychology  (Nosek et al., 2005) and the questions are adapted with the defined transparent information related to the self-selection bias. Therefore, the method of evaluation of people’s awareness should be carefully considered and designed, especially for the awareness of the implicit bias.

Additionally, there is an interesting finding of the awareness testing result in Fig.  6. People were more aware of the potential biased reviewers’ experience than the extreme emotions and various aspects in user reviews while the emotion is the most care than the other two types of information according to the formative study (Section  4). We attribute this to people’s belief updating problems as people have a demand of the extreme user reviews (e.g., negative reviews) to some extent for making decisions  (Ambuehl and Li, 2018). Hence, although people may be aware of the bias in the extreme feedback, they still tend to associate the user rating score (e.g., 1-point) with the corresponding emotions (negative only). Moreover, we also attribute this to one limitation of our work in the evaluation of people’s awareness. The questions for evaluating awareness in the post-questionnaire could be improved by gathering people’s reactions and perceptions of the aware-testing questions in the pre-study session or pilot study. Therefore, evaluating people’s awareness of bias needs to consider implicit question design, specific application scenarios and pre-estimate people’s responses to the evaluation method.

In our design process, we decided to use a unintrusive way to raise people’s awareness of the self-selection bias. The reason is that the direct warning of potential bias that explicitly informs people can be a kind of inducement for users which may cause unwanted emphasis on the data rather than the decision  (Law et al., 2021). Therefore, we did not add any explicit prompts or warnings of the self-selection bias in the design, and we even did not reveal the goal of our study or design during the experiment. The study results show that this implicit approach can effectively increase people’s knowledge and perception of possible biases in user ratings and reviews without distracting them from their decision process during the experiment. However, in the study results (Fig.  6), at least for the awareness of the different aspects (I3), it seems that people’s general level of awareness has not reached the ideal level. One reason may be the implicit way of the evaluation and the design. Hence, it deserves further exploration of the effects of implicit design, as well as how to increase awareness in the implicit way.

In addition, it remains to be studied on how to efficiently offering the awareness-raising design for informed decisions in daily life. The latest research in the visualization domain proposed a method to help people realize their biased behavior during the data analysis process by showing the interaction logs explicitly in visual analytic systems  (Narechania et al., 2021; Wall et al., 2021). We could adopt this way in the online decision-making scenario by making people’s behaviors transparent (e.g., scanning extreme reviews) and informing them about the potentially biased behavior on the fly. From the perspective of applying visualization, the literacy of visualization is key point that should be considered by designers as ordinary users may need some efforts to understand and interpret the information of the design.

We acknowledge that this work has several limitations. One limitation is that the sentiment analysis results and the keywords from user reviews involved in our experiment were all processed by automatic algorithms. Even though we chose a widely used tool  (Gardner et al., 2017) and a powerful model  (Grootendorst, 2020), we still cannot guarantee the algorithms are unbiased. These are emergent tools for checking model fairness  (Bellamy et al., 2019), and we hope to ensure the use of unbiased algorithms with these tools before integrating them into the backend of our system in the future.

Secondly, for the purpose of our controlled study, variables other than reviews and derived information are confined to a narrow range (e.g., price) to limit potential confounding effects. In a real world situation, users may face a wide variety of available hotels and their choice may be largely influenced by other factors, such as price, location, and even images of the hotel rooms. Moreover, considering the duration of the experiment and the burden on users, we only provided nine alternative hotels, while people may browse much more candidates in reality. We consider assessing people’s awareness of the self-selection bias in more complex situations, while using more hotels to evaluate the design in future work.

Thirdly, as we conducted our study using an online crowdsourcing platform, there are some uncertainties in this setup, such as unstable internet connections, browser incompatibilities, and other technical problems. To reduce the impact of this, we implemented some detailed functions to optimize the user experience, such as the adaptive interface design, prompt windows in the system, etc. While these helped in the experiment, several participants said they experienced compatibility and network issues. Future work needs to provide a more robust system for evaluating the design with a broader audience. Additionally, a long-term study is also necessary for observing how users leverage the bias-aware design for making decisions based on online ratings/reviews in real situations.

Despite these limitations, the user study does help us to learn more about users and their experiences, identify possible shortcomings in our design, and pinpoint improvements for our future work. In addition, exploring the impact of COVID-19 on the user-generated content and the possible bias behind (e.g., ratings and reviews of products or services) could be an interesting future direction to explore. We will also explore if our proposed approach can be generalized to other real-life scenarios in which user-generated content may affect users’ decisions (e.g., buying a product online) in the future. As such, we will continuously improve our work that aims to raise people’s awareness of various biases, and reflexively apply the design in other scenarios (e.g., buying products) based on the broader HCI and CSCW community’s feedback.