Fair space crowdsourcing task recommendation method based on competition balance
By combining multi-stage probabilistic recommendation and supply-demand flow balancing algorithms with fairness-aware allocation, the problem of unfair competition among workers in spatial crowdsourcing task recommendation is solved, achieving high efficiency and fairness in task allocation, and improving platform profits and worker opportunity equality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-12
AI Technical Summary
Existing methods for recommending tasks in spatial crowdsourcing fail to effectively model the competitive relationship among workers, resulting in unfair task allocation, low completion rates, and damage to platform profits. Furthermore, traditional fairness mechanisms fail to reflect the balance of opportunities to obtain tasks.
We construct a fair crowdsourcing task recommendation method based on competitive balance. Through three mechanisms—multi-stage probabilistic recommendation, supply and demand flow balance, and fair perception allocation—we dynamically adjust worker-task weights to resolve conflicts and optimize overall profit and opportunity fairness.
It significantly improved the success rate of task assignment and the fairness of opportunity among workers, resulting in an 11.6% increase in platform profits, an 18.3% increase in assignment success rate, and a 24.5% increase in Jain's Fairness Index, validating the technical feasibility of balancing coordination efficiency and fairness.
Smart Images
Figure CN122022345A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of spatial crowdsourcing and intelligent task recommendation technology, and in particular to a fair spatial crowdsourcing task recommendation method based on competitive balance. Background Technology
[0002] With the widespread adoption of wireless edge devices (such as smartphones), spatial crowdsourcing has rapidly developed in the fields of urban services and data management. Typical application scenarios include ride-hailing platforms (such as Didi Chuxing) and on-demand delivery services (such as Meituan Delivery), whose core mechanism is to distribute geographically specific tasks (such as picking up and dropping off passengers, delivering meals or packages) to widely distributed individual crowdsourcing workers. To improve service quality and task completion rates, platforms typically need to intelligently recommend task sets to workers that match their location, abilities, and preferences. However, if the recommendations deviate significantly from the workers' actual intentions, it may lead to negative responses or even job withdrawals, thereby significantly reducing overall task completion efficiency and user experience.
[0003] Current mainstream methods for recommending tasks in space crowdsourcing primarily focus on optimizing the preference utility, geographical coverage, or diversity of the task recommendation set, but generally overlook a crucial real-world factor: the intense competition among workers after task recommendation. In practice, platforms often employ a "multiple push, one selection" strategy, simultaneously pushing the same task to multiple nearby workers, with only one ultimately accepting the order. While this mechanism helps improve matching efficiency, frequent rejections can easily lead to worker frustration, potentially resulting in the loss of high-value workers in the long run, thereby weakening the stability and profitability of the platform ecosystem.
[0004] In this context, existing technologies reveal several key shortcomings. First, most studies fail to effectively model and maximize the overall system profit because they often focus only on static indicators (such as coverage) during the recommendation phase, neglecting the high uncertainty introduced by subsequent worker autonomous choices. This makes it difficult to accurately assess and optimize actual returns. Second, maintaining high profits while increasing the allocation success rate (i.e., the ratio of the number of times a worker is successfully assigned a task to the number of times they choose a task) is extremely challenging. Excessive recommendations intensify internal competition, lowering the allocation success rate and harming worker satisfaction; while insufficient recommendations may lead to unresponsive tasks, affecting platform operational efficiency. Therefore, striking a balance between stimulating moderate competition to ensure task completion and avoiding excessive competition to maintain perceived fairness has become a core issue that urgently needs to be addressed. Third, although some studies attempt to incorporate fairness considerations, they typically use only the profit difference between workers as a fairness metric, ignoring the more fundamental dimension of fairness in the task recommendation scenario—the equality of opportunity acquisition and the consistency of allocation success rates. This narrow view of fairness fails to truly reflect workers' perception of the system's fairness and cannot effectively alleviate internal conflicts caused by unequal resource competition.
[0005] In summary, existing spatial crowdsourcing task recommendation methods have significant shortcomings in balancing platform profits, allocation success rates, and multidimensional fairness. There is an urgent need for a new recommendation method that can synergistically optimize competitive balance, fairness of opportunity, and economic benefits. Summary of the Invention
[0006] This invention addresses the common technical problems in existing spatial crowdsourcing task recommendation systems, such as unfair task allocation, low task completion rates, and platform profit losses due to neglecting competition among workers. It provides a fair spatial crowdsourcing task recommendation method based on competitive equilibrium. Existing technologies typically focus only on preference matching, geographical coverage, or diversity optimization between tasks and workers, failing to adequately model the competitive conflicts arising from multiple workers simultaneously selecting a task and their negative impact on allocation success rate and worker satisfaction. Furthermore, traditional fairness mechanisms often use profit differences as the sole metric, failing to effectively reflect the balance of task acquisition opportunities. Therefore, this invention constructs a closed-loop task recommendation and allocation architecture integrating three mechanisms: multi-stage probabilistic recommendation, supply-demand flow balancing, and fairness-perceived allocation. This significantly improves task allocation success rate and opportunity fairness among workers while ensuring the platform's overall expected profit.
[0007] The purpose of this invention is to provide a fair space crowdsourcing task recommendation method based on competitive balance.
[0008] To achieve the above objectives, the present invention is implemented according to the following technical solution:
[0009] This invention includes the following steps:
[0010] Task recommendation steps: For each worker among multiple workers, determine and recommend a recommended task set from their valid task set; wherein, determining the recommended task set includes: constructing initial worker-task weights based on the utility that the worker can generate by performing the task and the probability that the worker selects the task; dynamically updating the weights through a multi-stage iterative process, and performing optimal matching based on the updated weights at each stage to gradually add tasks to the recommended task set of each worker;
[0011] Worker selection steps: Receive the subset of desired tasks that each worker autonomously selects from their recommended task set;
[0012] Task allocation steps: When multiple workers select the same task and a conflict occurs, a fairness-aware allocation algorithm is used to resolve the conflict and execute the final task allocation; wherein, the fairness-aware allocation algorithm dynamically adjusts the allocation weight according to the historical allocation success rate of each worker, and the workers with a low historical allocation success rate are given a higher allocation weight.
[0013] The task recommendation step also includes a supply and demand balancing sub-step:
[0014] Identify the first type of region where the number of workers is greater than the number of tasks, and the second type of region where the number of tasks is greater than the number of workers;
[0015] Construct a traffic network from the first type of region to the second type of region, and solve the minimum cost maximum flow problem to determine the transfer scheme of workers from the first type of region to the second type of region;
[0016] Based on the aforementioned transfer scheme, a corresponding recommendation preference region is determined for each worker;
[0017] When determining the set of recommended tasks, the tendency to recommend tasks located within the worker's corresponding preference area to that worker is increased.
[0018] The initial worker-task weight The calculation is as follows:
[0019]
[0020] in, Indicates workers Complete the task The profits generated Workers speed, It is the distance from the worker's position to the starting point of the task. It is the distance from the starting point of the mission to the destination. Workers Select task The probability of.
[0021] workers Select task probability Through exponential distribution Modeling is performed, among which For parameters.
[0022] In the aforementioned multi-stage iteration process, the first Worker-task weights in the next iteration The calculation is as follows:
[0023]
[0024] in, Indicates workers The probability of rejecting all previously recommended tasks. This indicates that no worker selected a task. The probability, For initial workers - task weights.
[0025] The optimal matching is achieved using the Hungarian algorithm.
[0026] workers Historical allocation success rate The calculation formula is:
[0027]
[0028] in, Workers The number of times a task has been successfully assigned. Workers The number of times a task was selected but not assigned.
[0029] The dynamic allocation weights used in the fair-aware allocation algorithm The calculation formula is:
[0030]
[0031] in, For initial worker-task weights, Assign the average historical success rate to all workers.
[0032] The beneficial effects of this invention are:
[0033] By modeling task recommendation as a multi-round dynamic weight matching process, explicitly considering worker rejection behavior and task rejection risk, the recommendation results are made closer to the actual execution scenario, thereby improving the feasibility of expected profits. By constructing a grid-based minimum-cost maximum flow model, an operable worker redirection strategy is generated, and a distance decay factor is embedded in the recommendation weights to effectively guide manpower to areas with scarce tasks, alleviating localized excessive competition. By using the consistency of allocation success rate as a fairness measure and dynamically amplifying the matching weight of workers with low success rates during the allocation phase, opportunity fairness is significantly improved without sacrificing platform profits. Experiments on real ride-hailing datasets (Didi, ride-hailing services) show that, compared to existing baseline methods, this invention maintains or even improves the platform's total profit (up to 11.6%), while simultaneously improving the task allocation success rate and Jain's Fairness Index by up to 18.3% and 24.5%, respectively. This verifies the technical feasibility and engineering practicality of this method in coordinating efficiency and fairness, alleviating internal competition conflicts, and improving the stability of the platform ecosystem. Attached Figure Description
[0034] Figure 1 This is a diagram illustrating the overall architecture of the fair spatial crowdsourcing task recommendation method based on competitive equilibrium of this invention.
[0035] Figure 2This is a flowchart illustrating the task recommendation and allocation process of this invention. Detailed Implementation
[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.
[0037] like Figure 1 , 2 As shown, this invention proposes a novel architecture based on a competitive equilibrium-based fair spatial crowdsourcing task recommendation method to address the problem of fair task allocation in spatial crowdsourcing. This architecture comprises two main modules: task recommendation and task allocation, aiming to maximize profit and allocation success rate while ensuring fairness. The method follows an integrated three-stage workflow: task recommendation, worker selection, and task allocation. Within each time interval, an effective task set is first constructed for each worker. (Saving spatiotemporal constraints), then recommend a task set. Then the workers choose a subset of tasks independently. Finally, the platform resolves the conflict and executes the task assignment. This invention can be applied to ride-hailing service scenarios, where drivers are considered workers and passengers' travel requests are considered spatial tasks. First, the system filters out a set of orders available to each driver based on their real-time location, service radius, and available time slots. Then, considering factors such as estimated income, expected completion time, and regional supply-demand ratio, the system intelligently recommends a set of high-potential orders to drivers. Drivers can freely choose their preferred orders from the recommendation list. The platform then uses a global optimization mechanism to resolve conflicts when multiple drivers simultaneously select the same order, dynamically adjusting task allocation weights to prevent excessive concentration of tasks in high-activity areas by a few high-frequency drivers, thus ensuring reasonable order-taking opportunities for new and inactive drivers. This invention achieves a task allocation success rate of over 90% and allocation fairness of over 95%, fully validating its ability to maximize overall platform profits while effectively coordinating and balancing competition among drivers (or service providers). By introducing a fairness constraint mechanism and intelligent task allocation strategy, it significantly alleviates internal conflicts caused by resource competition. This not only ensures a reasonable balance between efficiency and fairness in task allocation but also improves the overall scheduling efficiency and service quality of the system.
[0038] The task recommendation module aims to generate worker task sets. This study employs a combined strategy of a multi-stage probabilistic recommendation algorithm and a supply-demand flow balancing algorithm. The multi-stage probabilistic recommendation algorithm maximizes overall expected profit by recommending the optimal task set. It models task recommendation as a multi-stage matching process, where at most one task is added to each worker's recommended task set at each stage. The initial weights of the matched edges... It is calculated as the profit per unit of time multiplied by the probability of the worker's choice. (Worker) Execute the task initial weights The calculation is as follows:
[0039]
[0040] in Workers Execute the task The initial utility weight. Indicates task by workers The profit generated upon completion. Defined as task reward Subtract distance-related losses ,in It is the distance loss factor. Workers The speed. Workers Current location To the starting point of the mission The driving distance. It is the starting point of the mission. To the mission destination The driving distance. Workers Willing to choose tasks The probability, which follows an exponential distribution. Modeling.
[0041] To maximize the expected profit of all recommendations, the weights are dynamically updated in each iteration based on the results of previous iterations. Worker-task weights in the next iteration The calculation is as follows:
[0042]
[0043] in It is the first Worker-task weights for the next iteration. Indicates workers The probability of rejecting all previously recommended tasks. This indicates that no workers selected the task. The probability of. These are the initial weights. The optimal match is obtained by applying the Hungarian algorithm in each iteration. Iteratively added to workers Recommended Collection This maximizes expected profits and effectively avoids redundancy.
[0044] Supply-demand flow balancing algorithms improve worker allocation success rates by adjusting supply and demand. Recommending a single task to multiple workers can lead to intense competition, thus reducing worker participation. The supply-demand flow balancing algorithm first uses a grid-based approach to analyze supply and demand distribution, identifying grid sets where the number of workers exceeds the number of tasks. Grid sets where supply exceeds demand and the number of tasks is less than the number of workers. (Supply is less than demand). The algorithm then constructs a traffic network graph. By definition arrive The capacity (number of workers that can be transferred) and cost (travel distance between grids) of the edges between them. (Through solving the diagram) The minimum cost and maximum flow problem is solved by generating a supply and demand transfer graph. The transition diagram Able to identify the most suitable task area for each worker This allows surplus workers to be efficiently redirected from areas of oversupply to areas of undersupply, thus striking a balance between encouraging sufficient competition to complete tasks and avoiding excessive competition.
[0045] During the task recommendation phase, in order to balance competition, the probability of workers choosing a task is determined. It will be adjusted to To make it biased towards its most suitable mission area. Internal tasks. Adjusted weights. as follows:
[0046]
[0047] in It is the adjusted selection probability weight, used only in the matching process. It is the original choice probability. and These are hyperparameters, which control the intensity and range of the adjustment respectively. Indicates task with workers The most suitable task area The distance between them.
[0048] The task allocation module allows workers to independently select task sets. Subsequently, the task allocation module uses a fairness-aware allocation algorithm to resolve conflicts. This algorithm aims to maximize the consistency of allocation success rates among workers. This achieves the goal of fairness. Workers Allocation success rate The definition is as follows:
[0049]
[0050] in Workers The success rate of allocation. Workers The number of times a task has been assigned. Workers The number of times a task was selected but not assigned.
[0051] Fairness The fairness index is used as a metric, based on the success rate of task assignment for each worker. Consistency. To promote fair outcomes, the platform introduces dynamic weights to adjust matching weights, prioritizing task allocation to workers with historically lower assignment success rates. Dynamic weights The calculation is as follows:
[0052]
[0053] in These are the dynamic weights used in the final allocation. It is the initial weight. It is the average success rate of all workers. Workers The success rate of allocation. If Below average, workers Dynamic weights This will be amplified, thus increasing its chances of obtaining tasks, and finally the Hungarian algorithm is applied to obtain the optimal allocation result based on dynamic weights.
[0054] The experiments used two real-world ride-hailing service datasets: Didi and Yueche. In both datasets, driver-passenger matching was used to simulate the spatial crowdsourcing problem: passengers were treated as tasks, and drivers as workers. The Didi dataset contained 766 workers and 9962 tasks, covering the period from 19:00 to 21:00, located in Chengdu. The Yueche dataset contained 598 workers and 11299 tasks, covering the period from 15:00 to 17:00, also located in Chengdu. All experiments were conducted on hardware configured with an Intel(R) Xeon(R) Silver 4214 CPU @ 2.20GHz, 256 GB RAM, and an NVIDIA GeForce RTX 3080. In the task recommendation and assignment phase, the default available time for workers was set to 2 hours, the default expiration time for tasks was set to 0.5 hours, the default waiting time for tasks was set to 30 seconds, and the default number of tasks selected by a worker was set to 2. All baseline models were evaluated over 10 rounds, and the average performance was used as the final metric.
[0055] The method of this invention was compared with several baseline methods: NOC: a non-competitive method that ensures each worker is recommended a unique task. OTA: an optimal task allocation algorithm that enforces task allocation based on the best match in each round, assuming workers must accept and complete the task. GCA: a greedy coverage-aware method that maximizes the number of unique points of interest in the recommendation set through a greedy and non-competitive approach. CRA: a completion-sensitive recommendation algorithm designed to obtain the optimal expected number of completed tasks. DYTOPK: a dynamic Top-k method that increases the variability of the number of recommended tasks to adapt to dynamic worker and task numbers. MSPR: the multi-stage probabilistic recommendation algorithm of this invention, using the Kuhn-Munkres algorithm for allocation. MSPRD: MSPR employs a supply-demand flow balancing algorithm to balance competition among workers. FairMSPR: MSPRD further employs a fairness-aware allocation algorithm to maximize the consistency of worker allocation success rates. We used four evaluation metrics to measure effectiveness and efficiency: Profit: measures the total utility of the entire recommendation-to-allocation process. Allocation success rate: the ratio of the number of times a worker is assigned a task to the number of times they select a task. Fairness: Consistency of allocation success rate as measured by the Jains Fairness Index. Central Processing Time: Measured by the total computational cost of the entire recommendation-to-allocation process.
[0056] When analyzing the impact of increased task or worker numbers, the proposed method consistently achieves higher profits compared to CRA and NOC, reaching up to 11.6% (compared to CRA) and 44.9% (compared to NOC). This indicates that the strategy of maximizing expected profits through dynamically updating matching weights is effective. Furthermore, the profit levels of MSPRD and FairMSPR are comparable to MSPR, demonstrating that the core profit generation capability remains robust when supply-demand balancing and fairness mechanisms are introduced. Among the competing algorithms, the FairMSPR algorithm achieves the highest fairness. Compared to MSPR and CRA, the FairMSPR algorithm improves fairness by up to 13.2% and 24.5%, respectively. Even with increased task waiting time, the FairMSPR algorithm still shows a 6.2% improvement over MSPR. The results confirm that incorporating fairness-aware constraints during the allocation phase can significantly improve the consistency of allocation success rates among workers.
[0057] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.
Claims
1. A fair space crowdsourcing task recommendation method based on competitive equilibrium, characterized in that, Includes the following steps: Task recommendation steps: For each worker among multiple workers, determine and recommend a recommended task set from their valid task set; wherein, determining the recommended task set includes: constructing initial worker-task weights based on the utility that the worker can generate by performing the task and the probability that the worker selects the task; dynamically updating the weights through a multi-stage iterative process, and performing optimal matching based on the updated weights at each stage to gradually add tasks to the recommended task set of each worker; Worker selection steps: Receive the subset of desired tasks that each worker autonomously selects from their recommended task set; Task allocation steps: When multiple workers select the same task and a conflict occurs, a fairness-aware allocation algorithm is used to resolve the conflict and execute the final task allocation; wherein, the fairness-aware allocation algorithm dynamically adjusts the allocation weight according to the historical allocation success rate of each worker, and the workers with a low historical allocation success rate are given a higher allocation weight.
2. The fair spatial crowdsourcing task recommendation method based on competition balance according to claim 1, characterized in that: The task recommendation step also includes a supply and demand balancing sub-step: Identify the first type of region where the number of workers is greater than the number of tasks, and the second type of region where the number of tasks is greater than the number of workers; Construct a traffic network from the first type of region to the second type of region, and solve the minimum cost maximum flow problem to determine the transfer scheme of workers from the first type of region to the second type of region; Based on the aforementioned transfer scheme, a corresponding recommendation preference region is determined for each worker; When determining the set of recommended tasks, the tendency to recommend tasks located within the worker's corresponding preference area to that worker is increased.
3. The fair spatial crowdsourcing task recommendation method based on competitive balance according to claim 1, characterized in that: The initial worker-task weight The calculation is as follows: in, Indicates workers Complete the task The profits generated Workers speed, It is the distance from the worker's position to the starting point of the task. It is the distance from the starting point of the mission to the destination. Workers Select task The probability of.
4. The fair spatial crowdsourcing task recommendation method based on competition balance according to claim 3, characterized in that: workers Select task probability Through exponential distribution Modeling is performed, in which For parameters.
5. The fair spatial crowdsourcing task recommendation method based on competition balance according to claim 1, characterized in that: In the aforementioned multi-stage iteration process, the first Worker-task weights in the next iteration The calculation is as follows: in, Indicates workers The probability of rejecting all previously recommended tasks. This indicates that no worker selected a task. The probability, For initial workers - task weights.
6. The fair spatial crowdsourcing task recommendation method based on competitive balance according to claim 1 or 5, characterized in that: The optimal matching is achieved using the Hungarian algorithm.
7. The fair spatial crowdsourcing task recommendation method based on competitive equilibrium according to claim 1, characterized in that: workers Historical allocation success rate The calculation formula is: in, Workers The number of times a task has been successfully assigned. Workers The number of times a task was selected but not assigned.
8. The fair spatial crowdsourcing task recommendation method based on competitive balance according to claim 7, characterized in that: The dynamic allocation weights used in the fair-aware allocation algorithm The calculation formula is: in, For initial worker-task weights, Assign the average historical success rate to all workers.