HCRide: Harmonizing Passenger Fairness and Driver Preference for Human-Centered Ride-Hailing

1 Introduction

In recent years, ride-hailing services (e.g., Uber, Lyft, Ola Cabs, and DiDi Chuxing) have become indispensable to our daily transportation needs. By 2023, the global ride-hailing market size was valued at $109.3 billion, and it is expected to expand at a growth rate of 12.70% from 2024 to 2033 [RESEARCH, 2023]. One of the most important components of ride-hailing services is the order dispatch system, which directly impacts the revenue of platforms, the work experience of drivers, and the user experience of passengers.

Due to its importance, order dispatch has attracted significant attention from both industry and academia [Chen et al., 2019; Yuan and Van Hentenryck, 2021; Xu et al., 2018;

∗Corresponding author

Wang et al., 2022b]. However, most existing works focus on maximizing system revenue, which can potentially compromise the driver and passenger experience. Although some recent studies consider passenger fairness [S¨uhr et al., 2019; Zhou et al., 2023; Nanda et al., 2020; Wang et al., 2021; Wang et al., 2023a; Jiang et al., 2023], most adopt an absolute fairness setting [Wang et al., 2022a], assuming all passengers should experience equal waiting times regardless of location, which overlooks the dynamic nature of supply and demand across regions. Furthermore, driver preferences are frequently ignored, leading to poor experiences when drivers are dispatched to unfamiliar or undesired areas.

Hence, in this work, we aim to design a human-centered ride-hailing order dispatch system that harmonizes passenger fairness and driver preference. However, it is nontrivial to achieve this due to the following two reasons. (i) It is challenging to formally define passenger fairness and driver preference since they have highly spatial and temporal dynamics. (ii) Harmonizing passenger fairness and driver preference is also challenging due to their potential conflicts since improving passenger fairness may not align with drivers' preferences. For example, to ensure passenger fairness, drivers might be dispatched to high-demand areas, which could be distant from their preferred locations or lead to extended working hours beyond their preferred schedules.

To address the above challenges, we propose a HumanCentered Ride-hailing framework, called HCRide, which aims to minimize total passenger waiting time and enhance fairness without compromising driver preferences. In HCRide, we formulate the order dispatch problem as a Constrained Markov Decision Process (CMDP), where the passenger fairness-aware reward serves as the optimization objective and the accumulated driver preference-based cost is treated as a constraint. Passenger fairness is formally defined based on the divergence of waiting times across different spatial-temporal contexts, considering both inter-region and intra-region levels. Driver preferences are modeled using each driver's historical visitation frequency to various regions, reflecting their working habits and regional familiarity. To solve this CMDP, we develop a novel multi-agent reinforcement learning (RL) algorithm called Harmonizationoriented Actor-Bi-Critic (Habic). Habic consists of three key components. First, a multi-agent competition mechanism transforms the large joint action space into smaller distributed action spaces among a limited number of candidate agents. This enables a micro-level decision process [Qin et al., 2022]

that helps manage competition among proximate heterogeneous drivers with varying preferences. Second, a Bi-Critic module incorporates two evaluation networks: one estimates the reward value reflecting system efficiency and passenger fairness, while the other assesses the cost related to driver preferences. Third, an Actor module utilizes the outputs of the Bi-Critic networks to generate dispatch decisions that balance both passenger fairness and driver preferences.

The key contributions of this work are as follows:

1. To our knowledge, this is the first study on humancentered ride-hailing order dispatch that considers both passenger fairness and driver preference. Our design is motivated by social studies and data-driven analysis, from which we observe: (i) a notable discrepancy in waiting times among passengers, both within and across regions; (ii) drivers show distinct preferences for operational regions-some favor smaller, confined areas, while others are willing to cover broader locations.
2. Based on the data-driven findings, we design a humancentered ride-hailing order dispatch system called HCRide to improve passenger fairness without compromising driver preferences. Spatio-temporal-aware fairness and preference are defined. The core of HCRide is a novel multi-agent RL algorithm called Habic, which includes a multi-agent competition mechanism and an Actor-Bi-Critic module to harmonize passenger fairness and driver preference.
3. More importantly, we implement and extensively evaluate our HCRide based on two real-world ride-hailing datasets. Experiment results show our HCRide effectively improves system efficiency by 1.77% and 2.02%, inter-region fairness by 5.29% and 5.28%, intra-region fairness by 7.65% and 5.39%, and driver preference by 7.77% and 10.21% compared to baselines on the Shenzhen and NYC datasets, respectively. To verify our work, we have the code available at GitHub 1 .

2 Socially Informed Fairness and Preference Formulation

2.1 Data-driven Findings

In our previous project, we conducted qualitative studies to understand people's perceptions of current ride-hailing services, focusing on their views on fairness and their personal preferences [Wang et al., 2025]. In this paper, we also utilize real-world data with over one million ride-hailing orders to verify findings from a quantitative perspective, and finally, we have the following conclusions:

1 https://github.com/LinJiang18/HCRide

1. There is strong demand among passengers for equitable waiting times, particularly relative to others in close spatialtemporal proximity. However, substantial disparities in waiting times exist both within and across regions, driven by various spatial-temporal factors. As shown in Fig. 1, we visualize average passenger waiting times across 491 regions in Shenzhen during three periods: morning and evening rush hour, and noon non-rush hour. The results show that (i) spatially, significant differences exist between regions, such as between the Central Business District (CBD) and suburban areas; and (ii) temporally, even within the same region, average waiting times vary considerably within and across time periods.
2. Drivers have also reported having individual preferences for operating in specific areas at different times, such as locations near their homes, airports, or downtown districts. However, they are often assigned orders outside these preferred regions or in unfamiliar areas, which can negatively affect their satisfaction and operational efficiency. As shown in Fig. 2, darker colors represent areas frequently visited by a given driver. We observe that some drivers (e.g., Driver 1) tend to operate within limited regions, while others (e.g., Driver 3) cover a broader range. The preferred region entropy analysis in the upper right of Fig. 2 further highlights the diversity in drivers' operational areas.

2.2 Design of Fairness and Preference

Passenger Fairness

Motivated by the above findings, we define passenger fairness from both inter-region and intra-region perspectives, incorporating spatial and temporal patterns. In particular, intra-region fairness is defined as:

where WT (·) denotes the passenger waiting time, and p1 and p2 are two passengers located in the same region u during a specific time period v (e.g., 8:00-9:00). In this study, each time period is set to one hour, implying that passengers within the same region and period should experience comparable waiting times. Depending on the application, the temporal granularity can be adjusted. Under Eq. 1, fairness is achieved when the waiting times for both passengers are equal.

For passengers in different regions, we define a fairness benchmark, denoted as WTc(u, v), representing the expected average waiting time in region u during time period v. A passenger's waiting time in region u is considered fair if it aligns closely with this benchmark WTc(u, v). Inspired by the concept of Demographic Parity [Singh and Joachims, 2018], we evaluate fairness across regions by comparing their respective fairness benchmarks as follows:

where |Ci| = βi × N i passenger N i driver ; Npassenger is the historical average number of passengers in region u during period v, Ndriver denotes the number of drivers who prefer to operate in region u, and βi is an adjustable hyperparameter. Under this setting, the inter-region fairness benchmark WTc(u, v) is inversely proportional to the supply-demand ratio. The underlying rationale is that the expected waiting time in a region should decrease when driver supply exceeds passenger demand.

Driver Preference

To quantify driver preferences, we introduce the following definitions: (i) U: the set of all regions in the city; (ii) H+ k : the set of regions that provide positive feedback when driver k serves a passenger there; (iii) H0k : the set of regions that are neutral-i.e., do not yield positive feedback but are still acceptable to driver k; (iv) Hk : the set of regions that result in negative feedback for driver k. The formal representations of these sets are defined as follows:

where Vk(u) denotes the historical visitation frequency of driver k to region u. A region u is classified into the Positive region set H+ k for driver k if the visitation frequency Vk(u) exceeds a threshold d. dis(u, u1) represents the distance between the region u and u1, which should be shorter than κVk(u1). Here, κVk(u1) defines the radius of influence for a positive region u1. This means that if region u falls within the influence radius of the positive region u1, it is deemed acceptable for driver k to operate in.

In Fig. 2, we illustrate examples of the influence radius for drivers 1 and 3. The negative region set Hk comprises all other regions that are not included in the positive region set H+ k or the neutral region set H0k .

3 HCRide SystemDesign

In this part, we introduce the detailed design of the humancentered ride-hailing order dispatch system HCRide, which prioritizes passenger fairness and accommodates driver preferences, considering various spatio-temporal factors.

3.1 Order Dispatch ProblemFormulation

Formally, we model the fairness-oriented, preference-aware ride-hailing order dispatch problem as a Constrained Markov Decision Process (CMDP) G, defined as an 8-tuple: G = {S,A,P,R, C, µ, γr, γc}, where S is the state space; A the action space; P the transition probability function, P : S × A × S → [0, 1]; R the reward function; C the set of cost functions; µ : S → [0, 1] the initial state distribution; and γr, γc ∈ (0, 1] are discount factors for future rewards and costs. Let Π denote the set of all stationary policies, and let πθ(a | s) ∈ Π be a policy parameterized by θ that maps state s to a distribution over actions a. We divide each day into consecutive time slots (e.g., one minute per slot) and perform state transitions from time slot t to t + 1. Dispatch decisions are executed at the beginning of each slot. The detailed formulation of CMDP G is presented below.

• Agent: In our problem, we define each driver as an agent. An agent is marked as inactive while fulfilling an order, which renders it temporarily unable to accept new orders until the current one is completed. As a result, the number of active agents Nt varies across time slots.
• State S: To support feasible order dispatch decisions, we define the state S from three dimensions. The state of agent k at time slot t is defined as skt = {STkt ,DVkt ,CONkt }, where STkt is the spatial-temporal state, including the current region r, time period p, and coordinates. DVkt represents the driver state, which encodes the driver's preferences. CONkt is the context state, capturing global supply-demand conditions, weather, and traffic information.
• Action A: Agents in our system can perform one of the three action types: accepting an order ar, moving to complete an order am, or cruising ac (when the vehicle is unoccupied) based on the driver's preferences.
• Reward R: The reward Ri for order i includes both passenger waiting time and fairness, denoted as:

The reward function in our system consists of two components. The first term, WT(i | u, v), reflects system efficiency by representing the waiting time of order i in region u during period v. Shorter waiting times yield higher rewards. The second term, 1 Ku,v ∑Ku,v k=1 (WT(k | u, v) -WTc(u, v))2 , serves as a fairness regularization term. Here, Ku,v denotes the total number of passengers in region u during period v. This term captures the variance between actual waiting times WT(k | u, v) and the dynamic fairness benchmark WTc(u, v), encouraging equitable service across different contexts. To balance efficiency and fairness, we introduce a hyperparameter α to modulate the weight between these two components. In this framework, only the agent fulfilling order i receives the reward ri t .

• Cost C: In our setting, each agent k will be assigned a positive region set H+ k , a neutral region set H0k , and a negative region set Hk based on its historical operating locations (i.e., preference). Drawing from the negativity effect principle [Rozin and Royzman, 2001], negative experiences often exert a stronger influence on our psychological state than equally significant positive or neutral ones. Hence, we impose a cost cit on agent k when an order dispatch leads it to a destination within its negative region set Hk . The magnitude of this cost cit is determined by the distance between the order's destination and the nearest preferred location within H+ k .

The goal of the defined CMDP G is to optimize the longterm cumulative reward Jr(πθ) while ensuring that the cumulative cost Jc(πθ) remains below a predetermined threshold ξ. Given the centralized nature of the order dispatch system overseeing all agents, we adopt a strategy of centralized training with decentralized execution [Sharma et al., 2021] to reduce computational complexity. This strategy emphasizes the cumulative reward and cost across all agents, rather than focusing on the outcomes of individual agents. Therefore, our long-term cumulative reward Jr(πθ) and cumulative cost Jc(πθ) can be represented as:

where T denotes the total number of time slots within one episode, and Ot represents the total number of orders in time slot t. During each slot, we dispatch order i to an agent according to the strategy πθ, and the agent receives a corresponding reward ri t and cost cit .

Our objective of maximizing the cumulative reward Jr(πθ) over an episode and ensuring that the cumulative cost Jc(πθ) does not exceed a predetermined value ξ can be denoted as Eq. 9. It indicates we aim to enhance system efficiency (i.e., reduce the total waiting time of all passengers) and improve passenger fairness without disproportionately compromising driver preferences.

3.2 Optimization Objective Conversion

Directly solving the constrained problem in an MDP is challenging, so we further convert Eq. 9 into the Lagrangian form:

where the λ ∈ R+ is the Lagrange multiplier, which is a positive real number. The objective of the above Eq. 11 aims to find the global optimal saddle point (θ∗ , λ∗ ). Since θ∗ is the optimal value, the θ∗ should satisfy L(θ∗ , λ∗ ) ≥ L(θ, λ∗ ), ∀θ ∈ R. Similarly, λ∗ should satisfy L(θ∗ , λ∗ ) ≤ L(θ∗ , λ), ∀λ ∈ R+. Finally, ∀(θ, λ), we obtain:

However, optimizing the two parameters simultaneously is computationally intractable, especially for θ that is described by a deep neural network. Therefore, we alternatively optimize the two parameters by fixing one and updating the other until convergence. We obtain the final θ∗ and λ∗ when both of them satisfy that:

Where θ∗- and λ∗- are the values of the previous values before achieving convergence. In the next part, we will show how we solve this optimization problem with MARL.

3.3 Harmonization-oriented Actor-Bi-Critic

In this section, we design a new MARL algorithm called Habic (i.e., Harmonization-oriented Actor-Bi-Critic) to solve the above-defined problem. There are three key components in the Habic: (i) A multi-agent competition mechanism, which is designed to provide information for decision-making by generating matching features between orders and drivers. (ii) A dynamic Actor network, which is designed to alternately update the policy parameter θ and the Lagrange parameter λ based on the matching features to make decisions in the multi-agent environment. (iii) A Bi-Critic network, which is utilized to evaluate the values of accumulated reward Jr(πθ) and accumulated cost value Jc(πθ) simultaneously. An overall framework of Habic is shown in Fig. 3.

Multi-agent Competition Mechanism

We consider drivers within a certain range of an order to compete for it. The driver selected by the Actor will accept the order, and other drivers will keep their original actions, i.e., staying or cruising. Since the number of drivers around an order is dynamic, the candidate agent set and action space are also dynamic. Each agent in the set generates a matching feature mki = {skt , oit } to compete for the order, which includes the state skt of agent k and the state oit of order i. The details of skt = {ST kt ,DV kt ,CON kt } can be seen in Sec. 3.1, and oit = {ORi,DEi, u, v} describes order i's information, including the pickup location ORi and drop-off location DEi (represented by longitude and latitude), region u, and period v. As shown in Fig. 3, n matching features are fed into the Actor network to help make dispatch decisions.

Dynamic Actor Network for Decision

In this part, we will first introduce the decision process in Actor, and then show how we update the parameter θ for the policy function πθ(a|s) and λ for the Lagrange multiplier. We regard the Actor to be equivalent to the policy function πθ(a|s) to make order dispatch decisions. Assuming there are n candidate agents in the agent set and the matching feature set is M = {m1i ,m2i , · · · ,mni }, the probability of choosing the kth agent can be represented as:

Where Q(mki ) is the matching degree neural network to calculate the matching degree between the order i and the agent k. πθ(ak |s) means the probability of selecting agent k from the candidate agent set. After calculating the probability of selecting each agent, we can obtain the policy function πθ(a|s) for the order dispatch decisions.

After executing our order dispatch strategy based on the πθ(a|s), we can collect the a set of transitions ((s1t , · · · , snt t ), akt , rk t , ckt , skt+1 ) and use them as the training data. Based on the gradient search procedure, we can obtain the updating rules for θ and λ as follows:

Where ηθ and ηλ represent the update step sizes for parameters θ and λ. As we described in Sec. 3.2, the parameters θ and λ will be updated alternately until they reach the optimal values θ∗ and λ∗ . However, in Eq. 15 and Eq. 16, the values of J R(πθ) and J C (πθ) still remain unknown. Therefore, we further design the Bi-Critic network to evaluate their values.

Bi-Critic Network for Evaluation

This section introduces how we evaluate the value of J R(πθ) and J C (πθ). As shown in Eq. 7 and Eq. 8, the J R(πθ) and J C (πθ) share the same structure, so we will only show the evaluation process for J R(πθ), and the evaluation process of J C (πθ) is the same. According to [Schulman et al., 2017], J R(πθ) can be rewritten as:

Where Dπ (s) is the state visitation distribution, which can be described as the average probability of the state s appearing at each moment in the trajectory. πθ- is the old strategy in the last cycle and πθ is the new strategy waiting to be updated in this cycle, which means πθ- and πθ are equivalent to πθn and πθn+1 in Eq. 15. ARπθ (s, a) is the Advantage function, which can be considered as another version of Q-value with lower variance by taking the state-value off as the baseline. In Habic, we calculate ARπθ (s, a) by utilizing the Generalized Advantage Estimation (GAE) method [Schulman et al., 2015], which can be described as:

Where ψ ∈ [0, 1] is a hyper-parameter in GAE, and V R πθ (st) is the value function to describe the value of state st when following a policy πθ.

In Eq. 17, we adopt the off-policy strategy [Brandfonbrener et al., 2021] by using the old strategy πθ- to collect the data and update the new strategy parameter θ, so the difference between the new strategy and the old strategy will not be too large. We also leverage the KL-divergence [Kullback, 1951] to restrict the updating range of θ, which is shown in Eq. 19:

By combining Eq. 17 and Eq. 19, we can obtain the final describe of J R(πθ) based on the PPO-Clip [Jayant and Bhatnagar, 2022]:

Where clip(x, a, b) = max(min(x, b), a), which means restricting x within the range [a, b]. Using the Eq. 20, we can estimate the value of J R(πθ) based on the Advantage function ARπθ (st, at), which can be represented by V R πθ (st) based on Eq. 18. Therefore, we build the Reward Critic network to calculate V R πθ (st) with the parameter ψr. We denote the Reward Critic network as V R ψr (st). Similarly, the Cost Critic network can be denoted as V C ψc (st) with the parameter ψc. The update rules for parameter ψr and ψc are:

To summarize, we build a Reward Critic network V R ψr (st) and a Cost Critic network V C ψc (st) to estimate V R πθ (st) and V C πθ (st), respectively. By updating ψr and ψc using Eq. 21 and Eq. 22, we obtain the accumulated estimates. The values of V R πθ (st) and V C πθ (st) are then used to compute the accumulated reward J R(πθ) and cost J C (πθ) based on Eq. 18 and Eq. 20. After obtaining these estimates, we use them to update the parameters θ and λ, where the converged θ∗ serves as the optimal solution to our optimization objective in Eq. 9, improving passenger fairness without disproportionately compromising driver preferences.

4 Evaluation

4.1 EvaluationMethodology

Data: We evaluate our HCRide on two real-world ridehailing datasets from the Chinese City Shenzhen and New York City (NYC). The Shenzhen dataset includes 1.07 million orders served by 1,200 ride-hailing vehicles from 03/2021 to 06/2021. The NYC dataset includes 214k orders served by 800 ride-hailing vehicles from 01/2024 to 02/2024. Baselines: We compare our HCRide with five different categories of baselines: (1) Myopic dispatching method: MD [Zhang et al., 2017]. This method aims to minimize the total waiting time for all the passengers in one slot without future consideration. The method MD will be considered the benchmark to be compared with all other methods in Table 1. (2) Single-agent RL methods: DQND [Mnih et al., 2013], AC-bgm [Wang et al., 2023b]. (3) Multi-agent RL methods: IPPO [De Witt et al., 2020], MAPPO [Yu et al., 2022]. (4) Constrained RL methods: CPO [Achiam et al., 2017], Lag-TRPO [Ray et al., 2019]. Compared to the previous two types of methods, the constrained RL methods introduce the cost and constraint. (5) Variants of our HCRide considering different fairness definitions: HCRide-AF with absolute fairness [S¨uhr et al., 2019], HCRide-MMF with max-min fairness [Sun et al., 2022].

Metrics: We define three categories of metrics to evaluate the performance of system efficiency (Average Passenger Waiting Time APWT), passenger fairness (inter-region fairness PFinter and intra-region fairness PFintra based on the variance of passenger waiting time), and driver preference (Preference Violation Rate PV R), which evaluates the proportion of orders assigned to non-preferred regions of drivers, i.e., the negative region set Hk .

4.2 Overall Performance

As shown in Table 1, we compare our HCRide with all eight baselines on the two datasets.

System Efficiency

We evaluate the system efficiency using the Average Passenger Waiting Time APWT . As shown in Table 1, our HCRide outperforms all other baselines. Using the Shenzhen dataset as an example, our HCRide reduces the average passenger waiting time by 5.48% compared to the benchmark baseline MD and outperforms the state-of-the-art method Lag-TRPO by 1.77% = (5.48%-3.78%)/(100%-3.78%) in the whole day and 1.53% in the morning rush hour. Compared to LagTRPO, HCRide utilizes a more efficient updating method PPO-clip [Jayant and Bhatnagar, 2022], thereby achieving better convergence. HCRide-AF and HCRide-MMF are variants of HCRide with different fairness settings. Since there are no changes to the efficiency settings, their efficiency performances are similar. In particular, single-agent methods DQND and AC-bgm can achieve better performance than multi-agent methods IPPO and MAPPO. One possible reason is that we use the discrete extension for the single-agent RL methods, which incorporates our multi-agent competition mechanism. This allows these methods to learn the value of order-passenger pair from every discrete order dispatch behavior, providing abundant training transactions to guide learning the matching degree between orders and passengers. However, multi-agent RL methods like IPPO and MAPPO can learn from the operation trajectories of each driver. When there are a large number of agents (e.g., over 1,000), it will bring a high variance for training. Additionally, it is challenging for the Actor to learn the competition among different drivers since they are trained independently.

Passenger Fairness

We evaluate passenger fairness from both inter-region and intra-region levels with metrics DPFinter and DPFintra. As shown in Table 1, our HCRide notably improves both metrics. For single-agent RL algorithms such as AC-bgm and constrained RL algorithms like Lag-TRPO, although they utilize the same fairness-based reward function as HCRide, they are less effective due to poor exploration and update ability. We also compare the performance of our spatio-temporalaware fairness definition with two other widely used fairness definitions: absolute fairness [Zhou and Sethu, 2002] and max-min fairness [Sun et al., 2022]. The results show that our HCRide can achieve better fairness performance compared to HCRide-AF and HCRide-MMF. A possible reason is that our spatio-temporal-aware fairness definition focuses on more fine-grained local information across different spatiotemporal contexts. We also provide the visualization results for inter-region fairness in Fig. 4, which shows the distribution of the average passenger waiting times of all regions on each day during the training process. We find that the variance of average passenger waiting time between different regions decreases during the training process, and the system eventually converges to be close to the fairness benchmark.

Driver Preference

For driver preference, we focus on the percentage of dispatched orders that differ from driver preferences. Using the Shenzhen dataset as an example, the driver preference violation rate reaches 18.21% for the benchmark MD. In this experiment, we set a predetermined violation rate of 15% and expect it to decrease during the training process, eventually converging below this value. The predetermined violation rate can also be set to other values based on operators' goals.

From Table 1, we find that HCRide achieves the best performance on the two datasets. In contrast, baselines such as DQND, AC-bgm, IPPO, and MAPPO, which fail to consider driver performance in order dispatch decisions, show performance similar to the myopic dispatch strategy. Constrained RL baselines such as CPO and Lag-TRPO outperform other non-constrained RL baselines but still fall short compared to our HCRide. A key reason is that our HCRide has a multiagent competition mechanism to improve the sampling efficiency and a Bi-Critic to provide direct reward value and cost value estimation for the matching degree.

5 RelatedWork

We divide order dispatch work into efficiency-oriented order dispatch and human-centered order dispatch.

Efficiency-oriented Order Dispatch: Most existing studies prioritize efficiency without considering human factors such as fairness and preference. Xu et al. [Xu et al., 2018] use the DRL to solve sequential dispatch problems by building the global Q-function for orders and passengers. Sadeghi et al. [Sadeghi Eshkevari et al., 2022] propose a scalable RL dispatching algorithm and conduct both offline evaluation and online evaluation. Recently, some works have begun to pay attention to the transferability of algorithm efficiency across various platforms. Wang et al. [Wang et al., 2018] use the transfer learning method to make DRL-based order dispatch algorithms more adaptive in different cities. Wang et al. [Wang et al., 2022b] propose a federated learning algorithm to improve the reliability of dispatching data during cross-platform processes.

Human-centered Order Dispatch: In recent years, human-centered design has attracted much interest, and more and more works focus on fairness. S¨uhr et al. [S¨uhr et al., 2019] propose an order dispatch method considering twosided fairness for both driver and passenger. Lu et al. [Lu et al., 2021] introduce the queueing theory to solve the long waiting time problem for passengers and make a trade-off between efficiency and fairness. There are also some other works focusing on human preference. Carvalho et al. [de Carvalho and Golpayegani, 2022] propose a multi-agent multiobjective optimization approach to satisfy user preferences in ridesharing services. Li et al. [Li et al., 2021] introduce the mutual information-based approach to solve the preferenceaware group task assignment in spatial crowdsourcing.

To our knowledge, our HCRide is the first order dispatch system that harmonizes both passenger fairness and driver preference for human-centered ride-hailing services.

6 Conclusion

Motivated by insights from our previous qualitative study and data-driven analysis, in this paper, we design a humancentered ride-hailing order dispatch system called HCRide, which aims to improve both system efficiency and passenger fairness in terms of waiting time without compromising driver preferences. In HCRide, spatio-temporal-aware fairness and preference are formally defined, and we design a novel multi-agent reinforcement learning algorithm called harmonization-oriented Actor-Bi-Critic, which includes a multi-agent competition mechanism, a dynamic Actor network, and a Bi-Critic network to optimize system efficiency and passenger fairness with driver preferences as constraints. Extensive evaluations on two datasets show our HCRide effectively improves system efficiency by 2.02%, inter-region fairness by 5.28%, intra-region fairness by 5.39%, and driver preference by 10.21% compared to baselines.

Acknowledgments

We sincerely thank all anonymous reviewers for their insightful comments and valuable suggestions. This work is partially supported by Florida State University, National Science Foundation under Grant Numbers 2411152, 2427915, and 2318697.