In recent years, the prevalence of smartphones and mobile apps has enabled many innovative technology-mediated mobility systems, e.g., ridesharing (e.g., Uber, Lyft, DiDi) [17,34,43], carsharing (e.g., Zipcar, car2go) [62,88], and bikesharing (e.g., Capital Bikeshare, Mobike) [87,91,99]. Among all these technology-mediated systems for more convenient mobility, ridesharing may be the most popular one. For example, over 31 million DiDi drivers provided 48.8 billion kilometers of services for 550 million passengers in 2018. Similarly, Uber is available in over 600 cities across 65 countries in the world. As of 2021, there were 118 million Uber passengers, who were served by a total of five million drivers [39].
As a technology-mediated mobility system involving interactions between the passenger, the driver, and the platform [44,80], ridesharing has attracted great interest from the research community, and multiple efforts have studied it from different perspectives, e.g., labor opportunities [32,70], norm and regulations [17], trust and privacy [9,12,78], values of time [18], economy and pricing [34,35], vehicle electrification [89], individual and collaborative behaviors of drivers [2], low-resource populations [24,92], and support for passengers with vision impairments [11,44] or drivers with hearing impairments [56]. Even though many existing Human-Computer Interaction (i.e., HCI) studies have focused on the ridesharing experience of the passengers, few of them have conducted a comprehensive study of the passenger wait time in ridesharing systems, which is reported as an important aspect of system design in HCI to enhance the user experience when the interaction between the system and its users is temporarily interrupted (e.g., waiting for the arrival of live agents at call centers and waiting for drivers to pick you up) [4]. Some other works also show that wait time directly impacts user experience. For instance, Dillahunt et al. [24] uncovered that accurate wait time estimation was an important benefit to passengers as it allowed them to make better use of their wait time and prepare for the ride. Yu et al. [100] unveiled that passengers' likelihood to abandon the service increases with a long initial wait time. Li et al. [58] also showed that long wait time is generally viewed as the primary negative emotion source from the user interface. Therefore, understanding wait times in technology-mediated mobility systems and identifying factors that may impact them is significant for the design of next-generation interactive mobility systems.
In this paper, we consider ridesharing as one of the most important and representative technologymediated mobility systems and examine DiDi passengers' wait times for ridesharing applications in China from both qualitative and quantitative perspectives. Due to the on-demand and technologymediated characteristic of ridesharing usage, there are two wait times interacting between passengers and the ridesharing system, i.e., (i) the promised wait time shown on the app after entering the origin and destination (i.e., OD) without sending requests and (ii) the actual wait time got by passengers. In particular, we study the following two research questions (RQ) in this work based on both qualitative and quantitative data:
• RQ1: How do people perceive the promised and actual wait times in one of the largest technology-mediated mobility systems in the world (i.e., Chinese DiDi)? • RQ2: What factors may impact the unequal actual wait time in DiDi's technology-mediated mobility system?
While some existing studies have focused on the actual wait time [30,49,106], we focus more on the discrepancy between the promised wait time and actual wait time in ridesharing, as well as people's perception and potential reasons of the difference, which, to our knowledge, has not been studied before. This wait time discrepancy issue is common across different platforms (e.g., Uber [64,72], DiDi [92], and Lyft [71,84]). As a first step toward answering the above two research questions, we adopted a mixed-methods approach by combining semi-structured interviews and data-driven analysis. We interviewed 102 participants (52 passengers and 50 drivers) to understand people's perceptions towards the promised and actual wait times, where both passengers and drivers reported discrepancy between the promised wait time and the actual wait time. We further examine participants' attitudes to the great discrepancy and the unequal actual wait time. Next, A Mixed-Methods Study of Wait Time Perception and Discrepancy CSCW023:3 we performed a data-driven analysis based on large-scale real-world ridesharing operation data (including over 36.6 million rides) to quantitatively triangulate the finding of the unequal wait time.
Based on our study, we found the two above-mentioned wait times impact the user experience in different ways, e.g., (i) long actual wait time or inaccurate promised wait times on DiDi mobile apps may lead to poor user experience and potential loss of users to other platforms. (ii) Inaccurate promised wait times to passengers may disturb passengers' schedules and cause negative emotions for passengers. (iii) Inaccurate promised wait times to passengers may also cause more payment for passengers if they make their schedules based on the promised wait times, i.e., promised wait times < driver pickup time = passenger actual wait time. Hence, a comprehensive understanding of passengers' perceptions of the two wait times in ridesharing systems is significant, but it has not been fully studied by existing works potentially due to the challenge in accessing large-scale multi-source data. Fortunately, with the wide deployment of ubiquitous sensing and communication devices on machines (e.g., mobile phones and vehicles), massive accurate log data can be collected, which, combined with qualitative interviews, provides us a great opportunity to comprehensively dissect and understand passenger wait times in technology-mediated mobility systems.
In this work, we also identify some design implications to enhance ridesharing systems and user experience. We summarize three highlighted lessons learned here to provide insights for other similar systems and future work. More detailed discussions will be shown in Section 6.1.
• Lesson learned 1: Value of Mixed-methods Approach in Human-centered Mobility Research. Our work highlights the value of the mixed-methods approach in HCI and humancentered technology-mediated mobility research. The qualitative analysis from the interviews could motivate detailed (scale and granularity) quantitative analysis of logged data. The quantitative analysis using logged data can also complement the results from the qualitative interviews. More importantly, both the qualitative and quantitative analysis can help us identify the important factors that impact wait times in ridesharing platforms. • Lesson learned 2: Algorithmic Fairness and Misconceptions. Our work highlights a new dimension across which differential treatment is met out to female users by algorithms in the ridesharing platform DiDi. The lack of algorithmic transparency in technology-mediated ridesharing could lead to frustration for users, which goes against the original intention of ridesharing for providing convenient mobility. Also, the perception of the wait time is often just as important as the actual wait times, as the perception in itself could drive the users away from the platforms, but people may experience cognitive dissonance when expressing perceptions based on past experiences.
In this section, we discuss prior related works about HCI studies on ridesharing (e.g., papers from CHI and CSCW), HCI studies on wait time, passenger wait time, and algorithmic awareness.
The emerging technology-mediated ridesharing systems (e.g., Uber, Lyft, DiDi) play a more and more important role in our daily traveling. Past HCI studies on ridesharing systems have examined and understood different aspects of them, e.g., dynamic pricing [34,35], labor opportunity [32], and trip plan for ridesharing [85]. Guo et al. [34] studied passengers' reactions to dynamic prices in ridesharing services from a user-behavior perspective based on both systematic data analysis and theoretical modeling. Kameswaran et al. [44] tried to understand independence through a case study of ridesharing use by people with visual impairments in metropolitan India. Brewer et al. [11] did similar work to examine how drivers support people with vision impairments during rides.
In addition to passengers with disabilities, some works also focused on disabled drivers, e.g., Lee et al. [56] identified and discussed design and product opportunities to improve the hard-of-hearing driver experience on Uber. Glöss ey al. [32] investigated how the design of new technology can create new labor opportunities in services like ridesharing. Almoqbel et al. [2] conducted interviews with drivers in the U.S. to understand how they, individually and collaboratively, address safetyrelated issues they face conducting their job. Svangren et al. [85] investigated how passengers use ridesharing systems to plan their trips through a qualitative study.
Even though many of these papers [24,85,92] mentioned passenger wait time directly impacts user experience, they did not perform an in-depth investigation to understand passenger wait times in ridesharing systems, especially in a combination of a qualitative and quantitative analysis way. In addition, most of the literature on ridesharing systems is focused on Uber, Lyft, or Ola, but our research on a ridesharing system is in a different context, i.e., Chinese DiDi, which has the potential to advance the knowledge of ridesharing in different countries or cultures.
There is also a long history of the HCI community studying wait time for better user experience in different services [4,15,38,47,53], and different techniques have been developed to estimate wait times, e.g., light barrier sensors [7], smartphone WiFi [94], crowdsensing [13], and crowdsourcing [33]. Asthana et al. [4] reported that wait time in HCI is an important aspect of system design to enhance the user experience when the interaction between the system and its user is temporarily interrupted (e.g., file download, waiting for the arrival of live agents at call centers, and waiting for drivers to pick you up). Wang et al. [92] reported that passengers developed their own strategies to shorten the wait time, e.g., sending requests in advance, but it is challenging for passengers to decide when they should do so since both requesting too early and too late may cause unsatisfactory results (e.g., order cancellation or long wait time). Lallemand et al. [53] indicated that wait times were often sources of anxiety and irritation, and they reported that understanding wait time perception may provide valuable information for user interface design. Li et al. [58] showed that long wait time is generally viewed as the primary negative emotion source from the user interface. Kim et al. [47] also uncovered that users' satisfaction is related to their perception of wait time. All these works show that wait time directly impacts user experience [47,51,58], and understanding wait time in services has the potential to enhance user experience and interactive systems, but the wait time in ridesharing is still rarely studied.
Some existing works studied passenger wait time in public transportation systems (e.g., bus, taxi, or subways) [20,26,30,49,57,63,69,75,95,106]. Kjaerup et al. [49] reported that passengers may encounter a long wait time for public transportation in a rural context with interviews. Fang et al. [30] intuitively inferred the actual wait time of taxi and bus passengers by the time duration of two successive available taxis or buses, which is decided by the number of passengers in the waiting queue. Collins et al. [20] concentrated on delivering wait time information to transit bus passengers for reducing uncertainty about wait times and improving user experience. Zhou et al. [106] estimated the actual wait time of bus passengers by predicting bus arrival time via bus passengers' participatory sensing. However, for public transportation, it is challenging to track passengers' spatiotemporal information (i.e., location and time) when they have travel demand, so it is often hard to analyze the precise actual wait time of passengers on a large scale.
Different from these conventional public transportation systems, technology-mediated ridesharing provides us with a good opportunity to log the actual wait time of each passenger for analysis. Furthermore, there are different wait times when passengers interact with ridesharing systems, which makes it challenging to comprehensively understand them in a single qualitative way or a quantitative way. Hence, we combine both interview data and large-scale operation data to obtain an in-depth understanding of passenger wait times in ridesharing services, which extends existing HCI work on ridesharing services and provides new findings for future studies.
A growing body of research in HCI is exploring how users make sense of the algorithm features (e.g., algorithmic fairness and algorithmic transparency) and how the awareness of these algorithms affects users themselves [27,29,37]. They provide a rich understanding of how users react to algorithms on technology-mediated systems and again highlight the potential for uncertainty and frustration. Lee et al. [55] explored the impact of this algorithmic, data-driven management on human workers and work practices in the context of Uber and Lyft. Eslami et al. [27] measured algorithmic awareness with a mixed-methods study and they found users' awareness of the algorithm led to more active engagement with the system.
However, little research has focused on the impact of algorithmic awareness on wait times in technology-mediated ridesharing systems. In this work, we build on existing work by exploring algorithmic awareness and understanding its impact on the wait times of ridesharing passengers. Our choice of mixed methods provides an in-depth perspective into how passengers' perceptions and actions to the algorithm of the DiDi, and the sensemaking also forms the core of our findings.
In this section, we introduce the methodology of our paper for further analysis. Fig. 1 describes the overall structure of our work, where orange parts represent data, blue squares correspond to methodology, and green squares demonstrate the analysis results and findings. Our research context is described in Sec. 3.1. We first recruited and interviewed 102 participants (52 passengers and 50 drivers) to gain an in-depth understanding of their perceptions towards promised and actual wait times in ridesharing systems, as shown in Sec. 3.2. We then conducted a comprehensive data-driven analysis to quantitatively triangulate the findings of unequal wait time based on large-scale logged data (Sec. 3.3). The results of interviews and logged data analysis are shown in Sec. 4 and Sec. 5, respectively. The final findings and implications of this study are summarized in Sec. 6.
In this work, we focus on the Chinese DiDi, which is one of the largest ridesharing companies in the world. Founded in 2012, DiDi operates in over 400 cities with over 550 million users and tens of millions of drivers. In China, DiDi has 377 million annual active users and 13 million drivers providing 25 million rides each day [14], so understanding its operation is very important for large-scale users, and can also potentially provide guidelines for other ridesharing operators. We utilize Fig. 2 to illustrate the wait times on the DiDi app from three parallel dimensions, i.e., time, action, and app display.
Vehicle ID Input OD Request a Ride Search for Drivers Pick you Up t 0 t 1 t 2 Time Action Promised Wait Time 2 minutes App Display Pickup Time 6 minutes Time for Seeking a driver Actual Wait Time 1 ○ At the time 𝑡 0 , passengers input their origin and destination on the DiDi app, (suppose they plan to go to the Shenzhen North Railway Station by taking a DiDi ride after a visit at the Windows of The World), there will be a promised wait time shown on the app before they send the request (e.g., the 2 minutes in Fig. 2).
2 ○ At the time 𝑡 1 , they press the confirm button and send the ride request. The platform will then search for drivers for the passengers and the app will count the seeking time until it matches a driver. In some cases, e.g., in rush hours or remote rural areas, there are few drivers around the passengers, so the app will show "Sorry, Few drivers near you", which means the potential seeking time will be longer. After matching a driver for the passengers, the driver will go to pick the passengers up.
3 ○ At the time 𝑡 2 , the driver meets the passengers and picks them up. Here, 𝑡 2 -𝑡 1 (i.e., the time spent waiting for the ride once it is requested) is defined as the actual wait time of the passengers, so it consists of two components, i.e., (i) the time spent by the platform for searching for a driver (i.e., seeking time) and (ii) the time for the assigned driver pick the passengers up (i.e., pickup time, the 6 minutes in Fig. 2). It should be noted that during the waiting process, both passengers and drivers may cancel the requests for some reasons (e.g., passengers change their minds or drivers cannot arrive at the pickup locations due to heavy traffic congestion). Hence, the period from 𝑡 0 to 𝑡 2 is defined as a successful and complete ride request process. Each successful ride will be recorded by the DiDi ridesharing platform when the driver picks the passengers up.
It should be noted that sometimes people only want to check how long it will take for them to wait for a ridesharing vehicle, so they can decide whether they need to request a ride in advance or not. In this case, they will open an app, and then input their origin and destination information for checking the promised wait time without sending requests, so the promised wait time is what they care about. However, few existing works pay enough attention to the promised wait time shown on an app, which may have a large difference from the actual wait time. Hence, in this paper, we consider promised wait time as an important component of user-system interaction in ridesharing systems and comprehensively examine the wait times, and show their significance to user experience.
A Mixed-Methods Study of Wait Time Perception and Discrepancy CSCW023:7
To understand users' perceptions towards the promised and actual wait times in DiDi's ridesharing systems, we first conducted interviews with 52 passenger participants in a semi-structured format for around 40 minutes. We then further recruited 50 driver participants, who were interviewed to supplement the findings obtained from the interviews with passenger participants. The format of interviews with drivers is the same as the passenger participants. All interviews were conducted in Chinese, and our native co-authors translated the interview recordings into English for analysis.
Recruitment and Interview Process. We looked for passenger participants who had previous or ongoing experience in using the ridesharing service operated by the DiDi. We did not specify how much experience one should have for ridesharing services. We provided the participants with a detailed explanation of the purpose and procedure of the interview, which they should understand before starting the interview. The recruitment process started in June 2019 by posting digital flyers on the most visited ride-sharing discussion forums in China. We also advertised our study on our social networks, and some other people also advertised the study for us via their social media. To make the sample more representative, we recruited participants from diverse demographic backgrounds, including various age groups, genders, occupations, and geographic locations. In particular, we also tried to make the age distribution of participants similar to the overall age distribution of DiDi users [65]. After careful selection of our participants, we finally recruited 52 passenger participants (26 males and 26 females). Their average age was 30, and the standard deviation value was 5.3. The age distributions of our samples and all DiDi passengers are shown in Fig. 3, and we found they are similar. Our participants have different job occupations, including financial analysts (5.8%), software engineers (9.6%), professors (3.8%), graduate students (11.5%), doctors (3.8%), unemployment(11.5%), etc. Participants are from 18 cities covering all city tiers (i.e., tier 1 to tier 4). We posted digital flyers on two ridesharing discussion forums in China to recruit driver participants, and we also went to different ridesharing driver lounges to recruit driver participants. All driver participants were also informed about our interview purposes in the same manner. There were 38 male drivers and 12 female drivers recruited. Their average age was 37, and the standard deviation value was 3.2.
Interviews with passengers were conducted through in-person (25%), video chat (10%), or phone call (65%), depending on the interviewee's location and preference. The passenger interviewer began with questions about the participant's demographic information, last ride, frequency of ridesharing use, and motivation for using ridesharing. Followup questions probed participants' perceptions of the promised wait time and actual wait time in ridesharing and how their understanding influenced their feelings and ride request strategies. Passenger participants were also asked about their perceptions of the factors that may impact their wait times. We did a similar interview procedure with drivers with different questions. The percentages of interviews with driver participants through in-person, video chat, or phone call are 22%, 2%, and 76%, respectively. The detailed questions made at interviews are shown in Appendix A.
To analyze the interview data, we transcribed the interview recordings and conducted a thematic analysis [10], which includes the following four phases: (i) read transcripts while making notes; (ii) go over the notes to generate initial codes and collate; (iii) sort different codes into potential themes; (iv) revise themes and then finalize themes. We use a qualitative data coding software tool called AXQDA [77] to analyze and preprocess the interview data, which can help us find initial codes and potential themes behind the notes.
We created a coding system that includes codes agreed by coders such as satisfaction, disappointment, and anger. Coders received thorough training on the coding framework and the specific criteria for each code, and they also practiced coding on a subset of interviews and provided feedback to improve consistency before coding the whole dataset. Coders assessed the initial Kappa coefficient and identified areas of disagreement based on the practice, and they used the results from the practice to refine the codebook and improve the training. Coders then apply the codes to label the data, e.g., highlight sections of the text in interview transcripts that correspond to codes. After coding interviews, they review the coded segments to ensure consistency and accuracy. The final coding scheme had good reliability across two coders with a Kappa value of 0.93. Conflicts between coders were resolved through discussion. In total, there were 525 codes finalized by the end of the coding phase. They captured passenger participants' motivations and purposes for using ridesharing, all participants' perceptions towards wait times in ridesharing, emotions toward ridesharing and wait time, explanations for inaccurate and unequal wait time, etc.
We then started to analyze the initial codes and consider how different codes may combine to form an overarching theme. Finally, all data and codes gradually converged on 18 interrelated themes. We focused on three themes relevant to our research questions around passenger perception of wait times: (i) inconsistency between the promised wait time and actual wait time, (ii) attitudes of participants to the inconsistency, and (iii) unequal actual wait time for different participants.
From our interview results, we found accurate promised wait time is very important for user experience, and many factors may impact the unequal wait times of passengers. To further delve and have a better understanding of the impact factors for the ridesharing passengers' unequal wait times, during this project, we also conduct a quantitative analysis based on large-scale real-world data. Our dataset used in this work includes over 36.6 million ridesharing rides in the Chinese city Shenzhen, which was collected in 2019. In order to verify more factors that might impact passenger wait time, we utilize the official urban analysis zone data in this work, which can help to capture more fine-grained spatial features. The details of the two datasets are shown below.
(i) Ridesharing Logged Data. After successfully completing a ride, an order record will be generated, which consists of over 20 key fields describing the order (related to both passenger & driver), e.g., the order ID, order time, passenger ID, passenger gender, pick-up time, pick-up location, drop-off time, drop-off location, ride distance, pick-up distance, vehicle ID, etc. The actual wait time can be obtained by (pick-up time -order time). Note that gender information was limited to male and female, which we consider a limitation of this work. Fig. 4 shows the passenger actual wait time distribution based on our logged data, which indicates 678,300 (12.5%), 266,600 (4.91%), and 98,200 (1.81%) rides have wait times longer than 10 minutes, 15 minutes, and 20 minutes per week in Shenzhen, respectively. Some passengers' wait time is even longer than 60 minutes. More details of the logged data will be shown in Appendix B.
(ii) Urban Analysis Zone Data. It describes the urban partition for the population census of the studied city (i.e., the Chinese city Shenzhen). In this data, Shenzhen city is divided into 491 regions, and each region has a region ID and longitudes & latitudes of its boundary. In particular, we recognize nine representative areas (e.g., residential area, central business districts (CBD), industrial area, and airport) using the method in [101] to investigate the area-based wait time.
In this section, we present the details of our key findings from our interview data analysis. From our interviews, we found most passenger participants use ridesharing in a hurry for time-saving purposes. For example, a 35-year-old male participant said, "I usually use ridesharing in a hurry [...]"(P03, M, 35) 1 . Another 28-year-old female participant also noted "Sometimes I used it when there was an emergency"(P11, F, 28). Many passenger participants perceived short wait time as a reason to prefer ridesharing to taxi, but still, the discrepancy between the promised wait time to passengers after inputting origin-destination and the actual wait time creates problems for them. "I use ridesharing because it is convenient for me to order a ride via mobile apps. The promised wait time declared by the DiDi app was always very short. However, after I sent my request, it sometimes changed into a very long wait time estimation and I finally waited much longer than the promised wait time. [...] "(P08, M, 31).
Another female passenger participant also expressed a similar perception."I used to check if there are vehicles available around me and how long I will wait when I need to use ridesharing. However, I found the promised wait time is really inaccurate. [...]" (P49, F, 29). Some participants also reported that they would not use the ridesharing service if the promised wait time before sending a request is longer than 10 minutes, so it may be a business strategy of ridesharing platforms to show a short promised wait time on the app to attract people to use it, which may potentially impair user experience of ridesharing passengers when they encounter the discrepancy.
4.1.2 Wait time and (dis)satisfaction. Passengers also expressed dissatisfaction about the inaccurate promised wait time shown on apps, and many participants reported encountering a discrepancy between the promised wait time and the actual wait time. One 33-year-old female participant remarked that her experience of the inaccurate promised wait time and long actual wait time in the interview, "[...] I inputted my origin-destination and tried to request a ride. It showed I needed to wait 5 minutes. However, after I submitted the request, I waited 25 minutes to have the vehicle pick me up. [...] It is a really terrible experience. " (P21, F, 33). This participant also mentioned she had several ridesharing experiences with a wait time longer than 10 minutes, mostly on bad weather days and late nights, although the app always indicated a relatively short wait time before I sent requests.
The discrepancy between the promised wait time and the actual wait time not only disturbed participants' plans but also caused extra costs for some of them when the promised wait time was longer than the actual wait time. One participant noted, "When I opened the DiDi app and inputted my origin and destination, it showed the wait time was 11 minutes [...] However after I sent the request, a vehicle was assigned to me immediately instead of waiting 11 minutes. Even though I hurried up to the pickup location, I still received an extra charge due to my lateness. " (P44, F, 35).
Between Promised Wait Time and Actual Wait Time 4.2.1 Wait time and passenger churn. Almost all participants perceived wait times are highly related to the user experience of ridesharing passengers. There are interactions between ridesharing services and passenger ride experience, and an intuitive reflection is the wait time. Poor ridesharing services or unreasonable dispatching will lead to long wait times or inaccurate wait time estimations, which potentially cause an unsatisfactory user experience, resulting in passengers switching to other ridesharing platforms. For example, one female passenger noted in the interview, "I will switch to take a taxi or another ridesharing platform if the wait time of DiDi is too long. Usually, after waiting for 15 minutes, I will cancel the request and try other alternatives. " (P11, F, 28). Participants always wished vehicles could pick them up as fast as possible, although they understood that the wait time is inevitable. They also mentioned that the wait time should be expectedly short or predicted accurately, so they can decide when to request their rides. For example, one female participant said, "The promised wait time by the platform should be accurate, so I can know when I should send the request and schedule my time better.[...] Although the promised wait time is usually accurate during non-rush hours, there may be a huge gap between the promised wait time and actual wait time during rush hours, especially in the CBD areas and residential areas. " (P06, F, 29) 4.2.2 Attitudes and strategies to the discrepancy of wait times. Most participants expressed negative attitudes towards a great discrepancy (over 10 minutes) between the promised wait time and the actual wait time, even though a few participants indicated that they did not mind it. We found that there were typically four moods of participants when they encountered a large discrepancy from our interview based on our thematic analysis: (i) Anxious, (ii) Mistrust and Disappointment, (iii) Boring, and (vi) Do not care. We also leverage the Circumplex Model of Emotions [73,74] to systematically code and interpret emotions in a structured way that takes into account both the intensity and the pleasantness of the emotional experience, which makes the process of identifying and categorizing moods more rigorous. It indicates that (i) Anxious is High Arousal and Negative Valence, (ii) Mistrust and Disappointment is Medium to Low Arousal and Negative Valence, (iii) Boring is Low Arousal and Slightly Negative Valence, and (iv) Do not care is Very Low Arousal and Neutral Valence. Some participants who used ridesharing for commuting or meetings indicated that they worried about being late, which may potentially disturb their schedules, and some of them also developed strategies to deal with the discrepancy. For example, a male passenger said, "I work at an IT company. [...] I usually use ridesharing for commuting. There are many times that I found the promised wait times had great discrepancy with the actual wait times in the morning rush hours. I felt anxious as I was worried about being late for work, so I usually left and sent requests 10 minutes earlier in the morning rush hour to avoid being late. " (P05, M, 32).
Another female passenger said, "Normally, I sent ride requests several minutes in advance before I went out of my house since I found the actual wait time was usually longer than the promised wait time. I do not want to waste my time waiting outside, especially on rainy and snowy days. " (P36, F, 29).
Strategies: Some passengers also stated they usually send requests in advance at some specific locations, e.g., railway stations. For example, one 46-year-old passenger said, "Ridesharing is more convenient when I need a ride at a railway station since I can send a request in advance. Every time I take a train, I will send a ride request several minutes before arriving at my destination. " (P45, M, 46). However, even though some strategies are adopted by passengers to deal with this discrepancy issue (e.g., in-advance-request), some of them also reported they have been charged a wait time fee since they cannot arrive at the pickup locations in time when the promised time is much longer than the actual wait time. For these participants, they complained about the extra charge and they were eager to have the precise promised wait time shown on the ridesharing app. For example, a male passenger noted, "In the university where I work, all ridesharing vehicles are not allowed into our campus and they normally stop outside the gates. When I opened the DiDi app and inputted my OD, it showed the wait time was eight minutes. However, after I sent the request, a vehicle was assigned to me immediately instead of waiting eight minutes. Even though I hurried up to the gate, I was unable to arrive at the pickup location on time. [...] I think the promised wait time on the app is unreliable and should be more precise. " (P35, M, 46).
In addition to the statements from the passenger participants, some driver participants also expressed similar situations. For example, a male driver indicated, "Some passengers sent requests too early, so I waited for more than 10 minutes. Even though there are some compensations for me, it is much less than the normal operating earnings. " (D01, M, 38) 2 .
Mistrust and Disappointment: Many passenger participants also expressed their mistrust and disappointment towards this discrepancy, especially for people who trusted the platform. One 29-year-old female passenger noted, "On a cold winter night, I planned to ride a DiDi home from my office. When I opened the app and inputted my destination, it showed the wait time would be 2 minutes, so I requested a ride when I arrived at a pickup location. However, when the platform assigned a driver for me, the wait time turned out to be 10 minutes, so I was freezing in the cold wind for a long time. I felt very disappointed with the DiDi app, and I had an extremely bad ridesharing experience since there was a large gap between the promised time and the actual wait time. " (P20, F, 29).
This discrepancy not only damages user experience but also causes loss of passengers since passengers may switch to other ridesharing platforms if they feel disappointed. For example, a 37-year-old female passenger said, "I unloaded the app after encountering the discrepancy several times. I did not trust it anymore. " (P22, F, 37).
Interestingly, several participants specifically pointed out that they did not have large mood swings and they did not mind it if the actual wait time was less than half an hour. They reported that they would utilize this time to check their emails, and listen to music (e.g., "I check my emails or listen to music while waiting and I don't care about it. " (P12, M, 33)), or make phone calls (e.g., a male passenger said "I can listen to music or chat with my girlfriend when I wait for the driver. [...]" (P43, M, 26)). However, even for these wait time nonsensitive passengers, they also reported that they could not tolerate a very long wait time, e.g., they mentioned that they would also have a negative emotion if the actual wait time largely exceeded the promised wait time. Most people do not like waiting, so they will feel more dissatisfied and disappointed when they spend money to access those convenient services but get a long wait time.
Even though most passenger participants perceived different actual wait times when using ridesharing services, participants expressed diverse perceptions of the potential factors for the unequal wait time. Based on our interview results, we found, among all factors, the order time and location are considered the most significant two factors that impact the wait time by passengers. For example, similar statements are mentioned by two passengers: "The actual wait time is different when I order ridesharing services at different locations. It is also different in the morning and afternoon, weekdays and weekends, and especially it is long in holidays. " (P03, M, 35) (P51, F, 29).
Weather condition is also highly involved in the interview, and over half of the participants reported the wait time is largely prolonged on bad weather days (e.g., heavy rainy days) compared to good weather days. However, people expressed different ideas about the impact of temperature on the unequal wait time. One passenger noted, "The wait time would be much longer and uncertain on rainy days, but I did not see large differences between Summer and Winter. " (P10, M, 46) 4.3.3 Wait Time and Passenger Gender. One of the most surprising findings is that some participants perceived it is unequal for male and female passengers regarding the actual wait times even though most participants did not mention it. For example, a 29-year-old female passenger said, "I have tried to request a ride with my boyfriend at the same time and at the same location, but it showed 25 minutes waiting for me and 3 minutes waiting for my boyfriend. " (P15, F, 29).
Another 32-year-old female passenger also expressed similar experiences of her colleagues and gave a possible explanation, "My female colleagues always complained they needed to wait a longer time compared to the male colleagues when they requested rides at the same locations at late night time. " (P39, F, 32).
This may be an unexpected byproduct of the AI-based scheduler adopted by the ridesharing platform, where safety is the prerequisite for providing services. Platforms treat female users as a more vulnerable group, so they have higher standards to match drivers to female users compared to male users, e.g., male drivers with low ratings or complaints from female passengers in the past will not be assigned to female passengers. Interviews with some drivers also validated this finding. For example, one male driver noted, "After getting a poor rating from a female passenger, the platform has not dispatched female passengers to me anymore. " (D03, M, 33). Another male driver also said he has no female passengers for a long time, "I did not receive ride requests from female users for over three months. " (D07, M, 35). Some drivers also mentioned that their income has also been impacted since no orders from female passengers are assigned to them. "It is very strange that I did not receive orders from female passengers for two months. [...] My income also decreased since I had no orders from female passengers although I had the same operating time. [...]" (D25, M, 38).
Different from the statements from male drivers, no female drivers indicated that they lost female passengers. One female driver said, "I did see a decrease of female passengers, and I always have female passengers during nights. [...]" (D43, F, 35). This finding implies that the dispatching mechanism of ridesharing platforms may yield unequal treatment to female users, and an explicit perception of a long wait time compared to male users.
Distance. Drivers' preferences make the actual wait time of passengers more uncertain. Many participants reported that they waited a long time when their destinations were near to their origins. For example, one passenger noted, "Sometimes I waited a long time when my destination was near to my origin location. " (P17, M, 38). Some passengers thought drivers prefer to serve long trips. "[...] I think drivers normally prefer longer trips compared to short trips, especially at some locations with heavy traffic jams. " (P23, M, 29). Some drivers also expressed a similar view in terms of the trip distances. One 32-year-old driver said, "I will reject the request if the ride distance is short and traffic is heavy. " (D04, M, 32).
Locations. This may be because long rides could bring higher income for drivers since they do not need to spend extra time waiting for an order from the platform. In addition, we found rides that originate from rural areas usually have a long wait time, and some participants also mentioned that they waited over half an hour from rural areas to downtown CBD areas sometimes. For instance, a 28-year-old female passenger noted, "[...] I found it really hard to get a ridesharing service in rural areas. I waited for over 20 minutes. One reason is that there is a low ridesharing vehicle supply in rural areas, so drivers need to be dispatched from long distances. [...]" (P11, F, 28).
Furthermore, some areas with heavy traffic congestion were also hard to have access to ridesharing services, as indicated by both passengers and drivers. One passenger participant noted, "[...] It would be faster to take a taxi or other transportation modes in the CBD areas during rush hour or on weekends compared to ridesharing since drivers may cancel the order if the traffic is very heavy at that time.[...]" (P13, M, 37) The same situations are expressed by some drivers, e.g., "[...] Sometimes when I am stuck in heavy traffic, I will cancel the order in case I cannot pick the passenger up in time. " (D02, M, 43).
Summary: (i) Some participants perceived it is unequal for male and female passengers regarding the actual wait times, i.e., female passengers have longer wait times than male passengers during the night, which may be caused by the system design due to safety considerations. (ii) People use ridesharing in a hurry for time-saving purposes, and perceived short wait time and convenience are two important reasons to prefer ridesharing to taxi. (iii) The discrepancy between the promised wait time to passengers before sending a request (what they saw) and the actual wait time (what they got) potentially causes "broken expectations". (iv) Some strategies (e.g., in-advance-request) were adopted to deal with this discrepancy.
Even though participants expressed their perceptions of possible factors that impact actual wait time, findings from a sample of population may not be able to represent the public opinion or the fact due to possible cognitive dissonance. Hence, we further conducted a quantitatively statistical analysis based on large-scale logged data of over 36.6 million rides to study the factors mentioned by the participants.
Our most surprising finding regarding the unequal actual wait time from interviews is the perceived gender differences in wait times (reported in Sec. 4.3), so we first compare the actual wait time of female passengers and male passengers based on our large-scale logged data. As shown in Fig. 5a, we found the average wait time of female passengers during the late night (23:00-7:00) is noticeably longer than that of male passengers. Especially in 2:00-3:00, the average wait time of female passengers is 4-6 minutes longer than that of male passengers on average. But from 8:00-22:00, the average wait time difference between the female passengers and the male passengers is very small. To have a better understanding of the wait time of each ride from female and male passengers, we compare the wait time distributions of female and male passengers at late night and at other times in Fig. 5b. We found there is a huge difference in wait time between female and male passengers during late night (i.e., 23:00-7:00). Particularly, more than 31.6% and 18.6% of rides from female passengers have wait times longer than 10 and 12 minutes, and some females also wait for over 40 minutes.
Table 1. Wilcoxon Rank-Sum test for wait time of females and males Time h-value p-value Overall 1 2.9*10 -6 Late night 1 3.4*10 -22 Other Time 0 0.8798
We then utilize the Wilcoxon Rank-Sum testfoot_5 [96,103] to perform a statistical test in order to verify the differences between male and female passengers. This test can be used to determine whether two independent samples were selected from populations having the same distribution. The Wilcoxon test results are shown in Table 1. We found there is a significant difference (ℎ = 1, 𝑝 = 2.9 * 10 -6 ), which indicates female passengers have statistically different wait times from male passengers in general. More specifically, the difference is very significant during the late-night hours (ℎ = 1, 𝑝 = 3.4 * 10 -22 ), but there is no statistical difference during other daytime hours (ℎ = 0, 𝑝 = 0.8798). This result suggests that there actually exist unequal wait times for female and male passengers during the night, which uncovers the differential treatment of the DiDi ridesharing platform (or its algorithm) to female passengers.
One possible reason behind this gender disparity in wait times would be that DiDi orders of female passengers have to be served by female drivers at night for safety considerations, but there are fewer female drivers at night, which results in longer wait time of female passengers at nights. This explanation is also reflected in our interviewers. "[...] My husband and I talked to a male driver during a night trip, and he mentioned that orders from female passengers were typically served by female drivers instead of him, so he suggested my husband send requests if we traveled at night. [...] The driver also complained it was unreasonable since few female drivers were operating during the night"(P12, F, 30), a 30-year-old female passenger said. Some other participants also expressed similar situations, e.g., excerpts of interviews with P29 and P39 in Section 4.3. Even though it might be for safety considerations, the long wait time will still damage the user experience of female passengers.
As we reported in our interview results, some participants mentioned ride features (e.g., ride distance, ride origin, and destination) also related to the unequal wait time. Hence, we examine the passenger wait time with different ride features with large-scale logged data. We report some interesting/broken expectation results below.
Request Time of Rides. We first examine the correlation between passenger wait time and the ride origin location & order request time. We utilize the official urban analysis zone data described in Sec. 3.3 to divide the city into a set of regions. Fig. 6 visualizes the average wait time of passenger rides in the 491 analysis zones at different hours, including the late-night hour (0:00-1:00), morning rush hour (8:00-9:00), and evening rush hour (18:00-19:00). The darker red part means a longer average wait time in this region during this time slot, and the lighter yellow part means a shorter average wait time in this region during this time slot.
As shown in Fig. 6, we found that the average wait time in most zones is shorter than 7.1 minutes on both weekdays and weekends. In particular, the average wait time in the airport area is longer at the late-night hour. During the morning rush hour on weekdays, the average wait time in the CBD area and commercial business area is long. However, the wait time in these areas on weekends becomes shorter, but the average wait time in residential areas is long. A possible reason would be that the traffic in the CBD area and commercial business areas in the morning rush hour of weekdays is heavy, which potentially prolongs the wait time in these areas. However, on weekends, the traffic in the CBD area and commercial business areas are light since many people do not need to work, but more people may use ridesharing services to hang out on weekends, so the high passenger demand and heavy traffic may have a strong correlation with long wait time in residential areas on weekends. Furthermore, there is a large difference in wait time between the evening rush hour of weekdays and the evening rush hour of weekends. Specifically, the long wait time regions are in ports and some suburban areas on weekdays, but the wait time is long in most areas on weekends, including the CBD area, commercial area, airport areas, residential areas, etc.
Although participants of interviews perceived the time and location of sending ride requests can impact wait time, they hardly reported the in-depth relationships between them, so such a comprehensive statistical analysis may supplement participants' perceptions.
Pair Location of Rides. We further studied the passenger wait time from different origin areas to different destination areas since some participants indicated that they waited a long time from rural areas to downtown CBD areas sometimes. We utilize nine representative areas as mentioned in Sec. 3.3 to perform this examination, including airport, rail station, high-speed rail station, CBD, scenic, residential, industrial, commercial, and port areas. Fig. 7 shows the average wait time of rides from one type of area to another type of area, e.g., the average wait time of rides from the airport area to the rail station area is 15.104 minutes.
From Fig. 7, we found the average wait time for rides from the airport area is much longer than the rides from other areas. The average wait time for rides originating at transportation hubs (e.g., airport, rail station, high rail station, and port) is longer. One possible reason would be that passengers in these areas usually take subways or taxis since it is convenient for them to find taxis or subway lines there (there are fixed taxi pickup spots in these transportation hubs but ridesharing vehicles are not allowed to park there). Hence, ridesharing will lose its advantage of convenience in transportation hubs, which potentially lead to low demand & supply and long wait time. However, the wait time for rides from the CBD area is usually short as the average wait time to different areas is shorter than 6 minutes. The reason may be that there is a high and dense vehicle supply in the CBD area. Even though some participants perceived long wait times in the CBD area during rush hours, we found most rides have a relatively short wait time, which lowers the average wait time.
In general, the average wait time of passengers that originate at the transportation hubs (e.g., airport, rail stations, and port) is longer, while the wait time of passengers originating at the CBD
Airport RailStation HighRail CBD Scenic Resid. Indus. Commer. Port Airport RailStation HighRail CBD Scenic Resid. Indus. Commer. Port 4.121 15.104 18.028 20.407 16.106 13.817 19.463 13.978 10.504 8.308 5.551 5.268 5.0 4.898 6.628 4.751 4.451 4.446 7.265 5.837 4.82 5.331 6.006 4.843 5.1 5.407 6.013 5.901 4.988 5.74 4.772 5.301 5.095 5.041 5.882 5.465 6.154 5.881 6.926 5.389 4.942 5.278 4.903 5.234 6.141 6.507 5.374 5.515 6.005 5.187 4.808 5.196 6.759 5.437 6.55 5.544 5.497 6.109 5.294 5.162 5.016 4.874 5.907 5.822 5.397 6.231 5.751 5.412 6.171 5.202 5.964 5.366 6.799 4.317 6.324 5.307 5.062 7.195 6.302 5.464 4.82 5.00 area is usually short. Although some participants in interviews (e.g., P13 in Sec. 4.3) also indicated that different origin-destination pair locations impact their wait time, they only expressed subjective cognition based on their experiences. However, an in-depth finding was covered by the statistical analysis based on large-scale system operation data, which showed that the large-scale operation data can be leveraged to reveal more in-depth and comprehensive findings. 5.2.3 Ride Distance. Surprisingly, we found the Pearson correlation of passenger wait time and ride distance is always positive during different hours of a day with an overall value of 0.1617, and the results are significant (𝑝 < 0.001), as shown in Table 2. The results indicate that passengers for longer rides may need to wait a longer time. This is consistent with the findings from Sec. 5.2.1 and Sec. 5.2.2, e.g., the wait time for rides from the airport area to other areas is long because the airport is very far away from other areas as shown in Fig. 6. Also, since the Shenzhenwan Port is far away from residential areas and other transportation hubs, the average wait time for rides originating at the port is also long. However, our interview results show that some passengers reported long wait times when their rides were short, and drivers also preferred to serve long-distance rides and possibly rejected short-distance rides, which potentially indicates that passengers with longdistance rides may have a short wait time. One possible explanation for this disparity may be the memory lapses and biases that humans have [41,81]. People may have cognitive dissonance when they express their perceptions based on past experiences, and they are also easily suffering from cognitive limitations when conducting interviews. This interesting and unexpected finding underscores the value of complementary layers of data. Further interpretation and analysis of such dissonance could be a direction of interesting future work. Another possible reason would be that these passengers' short rides originate in the CBD area or residential areas in the morning rush hour of weekdays, where traffic congestion is heavy at these places, which potentially prolongs their wait time in these areas, as shown in Fig. 6, so they perceived a long wait time for those short trips in these areas at that time.
Conditions. We compare the passenger wait time in three typical weather conditions, i.e., sunny, rainy, and typhoon. Even though typhoons are always accompanied by rain, they are more adverse than normal rain. To avoid the interference of other factors (e.g., temperature and day of the week), we choose the types of days carefully. All the datasets for the weather condition comparison are from workdays (i.e., Thursday), during which have similar temperatures (i.e., about 23-27 • C). From Fig. 8, we found that there is a large difference in wait time distributions between days of different weather conditions. More trips have longer wait times on typhoon days and rainy days compared to sunny days. Quantitatively, the average wait time on typhoon days is 2.7 minutes longer than on sunny days, and the standard deviation is also larger on typhoon days. 5.2.5 Temperature. We divide the temperature into three levels to investigate the passenger wait time in different temperatures, i.e., high (27 -35 • C), mild (24 -27 • C), low (11 -15 • C).
Similarly, to prevent the interference of other factors, we also choose the there types of days carefully. All three kinds of days for the temperature comparison are workday (i.e., Thursday), and they have the same weather conditions (i.e., sunny). From Fig. 9, we found there is no obvious difference between the wait time distribution on mild days and cold days, even though the hot weather may cause a little longer wait time compared to them. In general, the average wait time in mild sunny days is 5.1 minutes with a standard deviation of 3.2. The average wait time in hot days is 1 minute longer than in mild days.
Summary: (i) Our large-scale data analysis indicates the differential treatment of the DiDi ridesharing platform (or its algorithm) to female passengers, which may be related to the safety considerations of ridesharing services. This matches with the findings in Sec. 4.3.
(ii) Different factors that impact passenger wait time are identified from our qualitative study and quantitative study. Some findings are consistent in the two studies, e.g., wait time is highly relevant to order time, origin and destination locations, weather conditions, and passenger gender. However, some conclusions from the two studies are conflicted, e.g., the relationship between wait times and ride distance. Also, the impact of some other factors on wait time (e.g., temperature) is not obvious.
Understanding wait time is crucial for technology-mediated systems such as on-demand mobility (e.g., ridesharing [42,100]), instant delivery [93,104], last-mile delivery [1,105], and telemedicine [46]. When individuals have a clear understanding of how long they need to wait for a service or response, they can better organize their tasks within the wait time and manage their expectations. For example, if ridesharing passengers can know the accurate wait times when they input their origins and destinations, they can decide when or if they should send a ride request or go to the pickup location. However, these platforms always provide an inaccurate wait time perceived by users. Prior HCI literature has studied the importance of "broken expectations" in the smart home, mobile apps, and digital device scenarios [8,50,82]. In this work, we highlight the discrepancy between the promised wait time and the actual wait time in the technology-mediated mobility systems, which potentially leads to poor user experience, mistrust, and abandonment of the systems or seeking alternative solutions [22]. Although ridesharing users actively try to make sense of the algorithm and interpret its actions, they may have cognitive limitations since these algorithms are highly complex and dynamic, which makes it challenging for users to fully comprehend how the algorithms work. Especially when they are in a hurry (a common scenario when users use ridesharing), users may not have the time or mental bandwidth to delve deeply into the workings of a ridesharing algorithm. In addition, people may have cognitive dissonance [6] and cognitive biases [16] when they express their perceptions based on past experiences, and a possible explanation for this dissonance may be the memory lapses and biases that humans have [41,81], leading to perceptions that may not align with the algorithm's actual behavior. This may be considered a drawback of using small-scale qualitative studies to understand user perceptions or experiences [21]. However, the perception of the wait time is often just as important as the actual wait times, as the perception in itself could drive the users away from the platforms. This is also the reason why we need to conduct in-depth interviews with users with diverse backgrounds and experiences.
Although existing work has studied passenger wait times in public transportation systems (e.g., bus [61], taxi [49], or subways [30]) through interviews or surveys, there has been limited exploration into comprehending the wait times in technology-mediated mobility frameworks like ridesharing combining qualitative and quantitative perspectives, which is more important since those relatively young users may be more sensitive to wait time [68]. Also, the work presented here highlights the significance of large-scale logged data in validating qualitative findings from a statistical perspective, which can provide a statistical value of passenger wait time discrepancy and impact levels of different factors on wait times based on millions of trips. Technology-mediated systems are becoming increasingly reliant on software algorithms to manage the information needs of their users and deal with the increasing amounts of data generated for satisfying user demand [55]. In recent years, algorithmic fairness has gained significant interest in the HCI community [27,28,90,98]. Prior research adopted different methods such as surveys [90], interviews [54,98], workshops [3], and content analysis [23] for in-depth investigation. Different from them, our work highlights the value of mixed methods in algorithmic fairness and humancentered research. The qualitative analysis from the interviews motivated detailed (scale and granularity) quantitative analysis of logged data. The quantitative analysis using logged data can also complement the results from the qualitative interviews. The initial qualitative analysis space was useful in hypothesis generation, which could be tested at scale (and the nuances explored) using quantitative analysis. Also, this is the first work to study wait time perception and discrepancy in technology-mediated mobility systems with a mixed-methods approach, which can potentially lay a foundation for future research on algorithmic fairness in different areas like Social Media and Online Platforms [52], Healthcare [19], Recruitment [79], and Education [48]. Existing work on understanding user experience usually conducts in-depth qualitative studies such as interviews [66,67,97]. This work reveals that large-scale operational logged data analysis is effective in triangulating the findings from qualitative studies for a more comprehensive study. Although ridesharing platforms had female passengers' interests (safety) in their minds when they designed the intervention of selective matching of female passengers and drivers, the design interventions may cause unintended consequences such as unequal wait time between male users and female users [86]. This unfair design will not only damage passenger user experience (e.g., increase wait time) but also have a negative impact on driver income and vehicle utilization efficiency.
Our work provides novel design implications for the HCI community. We demonstrate the importance of actual wait time and perceived wait time discrepancies in user experience when using ridesharing systems. Specifically, to improve the user experience in technology-mediated mobility systems, we delineate the typical features of promising technology-mediated mobility platforms.
Algorithm Fairness and Transparency. The lack of algorithmic transparency manner in which the matching of passengers and drivers is implemented could lead to frustration for both the passengers and the drivers. The disparate impact principle in algorithmic fairness suggests that large differences in outcomes for different demographic groups are considered unacceptable irrespective of the intent of the designers [5]. Hence, our results motivate more transparent and more equitable approaches by ridesharing apps in their matching algorithms in the future, e.g., considering the travel demand of female users at late night to reduce their wait time. Ridesharing platforms should ensure that algorithms used for driver allocation are fair and do not discriminate against certain demographics or areas. Platforms should also make wait time prediction algorithms transparent and explainable to build trust with users and allow them to understand how predictions are made and the factors influencing their wait time to manage passenger expectations. This has implications for user experience design and is important from the perspectives of value-sensitive design [28,31].
Wait Time Prediction-based Vehicle Repositioning. Large-scale logged data allows for the development of predictive models that anticipate demand fluctuations and wait time to optimize resource allocation accordingly. Ridesharing platforms should proactively dispatch more vacant ridesharing vehicles to some regions that will have high demand based on wait time prediction and provide incentives for those drivers to reduce passenger wait time, which will benefit drivers and passengers, as well as the long-term profits of platforms. The vehicle repositioning should also be dynamic and real-time since the passenger wait time is highly related to the order time, order location, weather conditions, trip distance, traffic conditions, and the origin & destination locations.
Feedback Mechanisms. Understanding user experience is important for platforms to improve user experience. Our study emphasizes the importance of mixed-methods approaches. Platforms can maintain a feedback mechanism that allows passengers to provide input on their experience, including punctuality of drivers and wait times, and use this feedback and large-scale logged data to continually improve the platform's performance.
User Experience Design. Future technology-mediated mobility app designers can consider designing a user-friendly interface creating visualizations and sharing information that can provide clear and easily accessible information about wait times, driver details, and trip progress, and help remove user misconceptions such as short-distance rides have longer wait times.
Compared to conventional user study works based on surveys and/or interviews, one advantage of our work is that we collect large-scale real-world technology-mediated system operation data to triangulate findings from our interviews. We believe that such large-scale system operation data is quite meaningful for HCI studies. Further, there is a growing acknowledgment of the need to move the HCI literature beyond the "WEIRD", i.e., western, educated, industrialized, rich, and democratic realm [36,40,59,76,83]. Hence, we believe our work has the potential to advance the non-WEIRD HCI literature on wait times, ridesharing interactive systems, and algorithmic fairness.
To benefit the research community, we plan to release real-world ridesharing operation data with proper privacy protection mechanisms in the camera-ready version. This large-scale ridesharing data has the potential to promote other research related to user behavior and ridesharing interactive system design.
Like any study, this work has some limitations. This work particularly focuses on a specific cultural context instead of a global setting, so our findings may not generalize to all other countries or technology-mediated platforms, which is also a common limitation of many CSCW studies [2,11,25,45,60,92,102]. Our results are from interviews with passengers and drivers. We could not interview algorithm developers or official representatives of the DiDi companies as it was against company policy. Our findings should be complemented by future research that uses different research methods such as ethnography or surveys. The research findings presented in this study are mainly derived from one ridesharing platform. Therefore, they might be biased due to the platform designs, dispatching algorithms, and the cultural background of the user population, so we are cautious to make any claims about the generalization capabilities of our findings to other countries or platforms. However, the problem identified and the features we considered in this study are based on the settings shared by similar platforms, which potentially improves their generalizability. We are actively seeking research opportunities on other ridesharing platforms like Uber and Lyft and perform a more comprehensive cross-platform analysis in future work. Another limitation is our interviews are conducted in China, but people in different countries may have different perceptions about wait times. Also, different conditions in different countries may also affect the actual result. For example, easiness of finding a taxi may be an important factor that affects how users consider the wait time. We plan to run our interviews in parallel in the U.S. to see how passengers' perceptions of wait times are different between China and the U.S. However, it is challenging and time-consuming to collect similar datasets used in this paper, and this parallel comparative study is also not the focus of this paper, so we will leave it for our future work. Although we conducted interviews with both passengers and drivers, in our work, we focus on passengers' perception of wait times, and future work can also focus on drivers' perceptions. Moreover, while studying wait time in ridesharing helped us understand the real-world impact factors, finding-based applications will be useful for verifying potential benefits. Hence, in the future, we will integrate our findings into later system design to enhance user experiences in ridesharing services. Our interview data contain participants' self-reported memories of past ridesharing experiences and subjective interpretations of their experiences. Some participants might memorize more details of certain kinds of experiences than other people, which could bias our findings. Another possible direction would be follow-up interviews with some participants for long-term perception understanding. Also, the percentage of participants (drivers and passengers) who leave the platform over different times of the year can be interesting to study. Finally, future work will also explore contextual details about why certain discrepancies (e.g., 4-6 minutes longer wait time for female passengers than males at night) were observed. In addition, although we have identified many factors (e.g., ride features, time, and locations) influencing wait time disparity, there may be some other factors that have not been uncovered, so we will further investigate those potential factors with additional contextual information (e.g., other transportation modes and socioeconomic status of users) in the future.
In this paper, we comprehensively examined the promised and actual wait times in one of the largest technology-mediated mobility systems with a mixed-methods approach. We first presented a qualitative study with 102 participants (including 52 passengers and 50 drivers) to understand their perceptions towards the two wait times in one of the largest ridesharing systems in the world. We found the wait time discrepancy between the promised wait time and actual wait time led to psychological discomfort and negative emotions for passengers, and there were some unexpected factors for the unequal wait time, e.g., the gender of passengers and ride distance. We then performed a quantitative analysis based on large-scale real-world ridesharing logged data to triangulate the findings from interviews. Finally, we shared some lessons learned (e.g., people may have cognitive dissonance when they express their perceptions based on past experiences) and design implications (e.g., minimizing expectation mismatch, supporting algorithmic transparency and fairness), which may potentially guide designers to improve their systems for better user experience and pave the way for the next generation interactive mobility systems. We show more details of our logged data here. As shown in Fig. 10, we show the demographic information of all ridesharing drivers in Shenzhen. We found the age range of drivers is between 20 to 64 years old, and about 80% of drivers are 30-49 years old. Among all drivers, male drivers account for 98.86% and only 1.14% of them are female. The extremely unbalanced ratio between male and female drivers provides good evidence to the finding of unequal wait time between male and female passengers.
In Fig. 7, we only show the average wait time of passengers from different origin areas to different destination areas, which may miss some detailed information. Hence, we also show the wait time distributions of passengers from different origin areas to different destination areas, as shown in Fig. 11. Received January 2024; revised July 2024; accepted October 2024
Proc. ACM Hum.-Comput. Interact., Vol. 9, No. 2, Article CSCW023. Publication date: April 2025.
(P03, M, 35) means the passenger participant number 3, Male, 35 years old. Proc. ACM Hum.-Comput. Interact., Vol. 9, No.
2, Article CSCW023. Publication date: April 2025.
(D01, M,
38) indicates the driver participant number 1, Male, 38 years old Proc. ACM Hum.-Comput. Interact., Vol. 9, No. 2, Article CSCW023. Publication date: April 2025.
h = 1 indicates a rejection of the null hypothesis, and h = 0 indicates a failure to reject the null hypothesis. The null hypothesis is the two groups have the same distribution. If the p-value is lower than the significance level (e.g., 0.05), then we can say that we have statistically significant evidence to reject the null hypothesis.Proc. ACM Hum.-Comput. Interact., Vol. 9, No. 2, Article CSCW023. Publication date: April 2025.