Labor scheduling processing method, device, equipment, medium and program

Abstract

This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, medium, and program for labor scheduling. The method includes: acquiring labor demand input from the labor client; inputting the labor demand into a pre-trained language model, which outputs a corresponding directed acyclic graph (DAG), wherein the pre-trained language model performs semantic understanding to decompose the labor demand into at least one sub-task, and generates the DAG based on the dependencies between the at least one sub-task; identifying at least one worker's data; and selecting the optimal scheduling path from the DAG based on the worker's data and pushing it to the corresponding worker's end, wherein the worker's end receives and displays the pushed optimal scheduling path and provides a confirmation instruction. This solves the problem that related technologies can only handle standardized, discrete, single-point tasks with structured form input, failing to meet the needs of flexible labor scenarios and resulting in low work efficiency.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment, medium and program for labor scheduling. Background Technology

[0002] In related technologies, the task posting of flexible employment platforms heavily relies on structured forms, requiring employers to manually break down complex projects. It cannot automatically handle ambiguous natural language requirements through semantic understanding, resulting in high posting thresholds and low efficiency. The scheduling algorithms are mostly limited to point-to-point path planning for a single task, completely ignoring data such as the commuting of flexible workers, which cannot meet the needs of flexible employment scenarios and leads to low work efficiency. Summary of the Invention

[0003] This application provides a method, apparatus, equipment, medium, and program for labor scheduling, in order to solve the problems in related technologies that can only handle standardized, discrete single-point tasks with structured form input, which cannot meet the needs of flexible employment scenarios and result in low work efficiency.

[0004] The first aspect of this application provides a method for labor scheduling, comprising the following steps: obtaining labor demand input from a labor client; inputting the labor demand into a pre-trained language model, wherein the language model outputs a corresponding directed acyclic graph of tasks, wherein the pre-trained language model performs semantic understanding to decompose the labor demand into at least one sub-task, and generates a directed acyclic graph of tasks based on the dependencies between the at least one sub-task; identifying at least one worker data, wherein the worker data includes geographical location, idle time window, skill tags, and ability verification materials; selecting the optimal scheduling path from the directed acyclic graph of tasks based on the worker data and pushing it to the corresponding worker client, wherein the worker client receives and displays the pushed optimal scheduling path and provides a confirmation instruction for accepting the order.

[0005] Optionally, the step of selecting the optimal scheduling path from the directed acyclic graph of tasks based on the practitioner data and pushing it to the corresponding practitioner includes: selecting a matching set of candidate subtasks from all subtasks in the directed acyclic graph of tasks based on the ability verification materials, the skill tags, and the free time window; constructing a target spatiotemporal constrained directed acyclic graph based on any subtask in the candidate subtask set, using the geographical location as the starting and ending points of the path; and generating the optimal scheduling path by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph.

[0006] Optionally, the step of constructing a target spatiotemporally constrained directed acyclic graph based on any subtask in the candidate subtask set includes: identifying the end time of the first subtask in the candidate subtask set, the sum of the commuting time and buffer time from the location of the first subtask to the location of the second subtask, and the start time of the second subtask; if the end time of the first subtask and the sum of the commuting time and buffer time from the location of the first subtask to the location of the second subtask are less than or equal to the start time of the second subtask, then establishing a directed edge between the first subtask and the second subtask; and constructing a target spatiotemporally constrained directed acyclic graph based on at least one directed edge.

[0007] Optionally, before constructing the target spatiotemporal constrained directed acyclic graph based on any subtask of the candidate subtask set, the method further includes: calculating the straight-line distance between the geographical locations corresponding to any two candidate subtasks; if the straight-line distance between the geographical locations corresponding to any two candidate subtasks is less than or equal to a preset threshold, then the two candidate subtasks are determined to be connected, and the commuting time from the location of the first subtask to the location of the second subtask is calculated.

[0008] Optionally, the optimal scheduling path is generated by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph. This includes: obtaining the reward amount of the second sub-task, the commuting cost from the location of the first sub-task to the location of the second sub-task, and the skill matching degree between the worker and the second sub-task for each directed edge in the target spatiotemporal constrained directed acyclic graph; calculating the sum of the comprehensive weights of the directed edges traversed by each complete closed path based on the reward amount, commuting cost, and skill matching to generate the total weight of the corresponding path; and selecting the path with the largest total weight from all complete closed paths to generate the optimal scheduling path.

[0009] Optionally, the pre-trained language model processing flow is as follows: Generate employment intent based on the natural language text corresponding to the employment demand; determine the corresponding employment demand category based on the employment intent, and guide the employer to complete the missing key information slots in the employment demand through multi-turn dialogue interaction to generate a complete demand description; use thought chain prompts to guide the language model to decompose the complete demand description into at least one sub-task; establish corresponding dependencies based on the logical relationships between at least one sub-task, and construct the task directed acyclic graph based on at least one sub-task and its corresponding dependencies.

[0010] A second aspect of this application provides a workforce scheduling processing apparatus applied to a cloud-based scheduling server. The apparatus includes: an acquisition module for acquiring workforce requirements input by a workforce client; a processing module for inputting the workforce requirements into a pre-trained language model, wherein the language model outputs a corresponding directed acyclic graph (DAG), wherein the pre-trained language model performs semantic understanding to decompose the workforce requirements into at least one sub-task and generates the DAG based on the dependencies between the at least one sub-task; an identification module for identifying at least one worker data point, wherein the worker data includes geographical location, available time window, skill tags, and capability verification materials; and a filtering module for filtering the optimal scheduling path from the DAG based on the worker data and pushing it to the corresponding worker client, wherein the worker client receives and displays the pushed optimal scheduling path and provides a confirmation instruction.

[0011] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the labor scheduling processing method as described in the above embodiments.

[0012] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to perform the labor scheduling processing method as described in the above embodiments.

[0013] A fifth aspect of this application provides a computer program product, including a computer program or instructions, which, when executed, implement the labor scheduling processing method as described in the above embodiments.

[0014] Therefore, this application has at least the following beneficial effects: This application embodiment can obtain natural language employment requirements input from the employer and input them into a pre-trained language model. The model performs semantic understanding and logical reasoning on the unstructured requirements, automatically decomposing them into multiple atomic subtasks with dependencies and generating a directed acyclic graph of tasks. This transforms fuzzy and complex employment requirements into a set of structured tasks that can be parsed and scheduled by machines, completely eliminating the reliance on manual decomposition and form filling by the employer, and significantly reducing the threshold for publishing complex tasks and the risk of errors. On this basis, the system identifies the worker's geographical location, free time window, skill tags, and ability verification materials, using the worker's residence as the starting and ending point and the matched subtasks as intermediate nodes, based on the spatiotemporal relationship between commuting time and task time window. The system constructs a personalized, spatiotemporally constrained directed acyclic graph based on connectivity conditions. By selecting the path with the highest global weight, it generates the optimal scheduling path that balances task rewards, commuting costs, and skill matching. This path is then pushed to the worker's end to complete order confirmation. This upgrades the traditional fragmented, discrete order-grabbing model to a closed-loop, 24 / 7 continuous scheduling model based on residence location. This effectively solves the problem of lost revenue due to commuting and idle waiting for workers. At the same time, a multi-dimensional verification mechanism for capability verification materials ensures the authenticity and accuracy of person-job matching. Ultimately, it achieves full-process automation and intelligence from demand understanding, task decomposition, capability assessment, path planning to order dispatch and execution, significantly improving the efficiency of supply and demand matching, task delivery quality, and overall worker benefits in flexible employment scenarios.

[0015] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0016] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a labor scheduling processing method provided according to an embodiment of this application; Figure 2 This is a framework diagram for intent recognition provided according to embodiments of this application; Figure 3 This is a schematic diagram of a spatiotemporal scheduling module provided according to an embodiment of this application; Figure 4 This is an overall flowchart provided according to the embodiments of this application; Figure 5 This is a block diagram of a workforce scheduling processing device provided according to an embodiment of this application; Figure 6 This application provides a flexible workforce automation system based on language models and directed acyclic graphs, according to embodiments of the present application. Figure 7This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0017] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0018] With the rapid development of the gig economy, flexible employment has become an important form of employment. Traditional flexible employment models mainly rely on offline intermediaries or basic information publishing platforms, resulting in low information matching efficiency and a lack of trust mechanisms. In recent years, with the development of mobile internet and artificial intelligence technologies, O2O (Online to Offline) flexible employment platforms based on mobile terminals have emerged. These platforms typically involve basic technologies such as LBS (Location-Based Service), natural language processing keyword extraction, collaborative filtering recommendation algorithms, and directed acyclic graph task scheduling.

[0019] LBS location services are used to obtain the geographic coordinates of users and tasks and calculate distances. Natural language processing is used to extract simple keywords from task descriptions (such as extracting tags like "cleaning" and "programming"). Based on collaborative filtering or logistic regression models, tasks are recommended according to the user's historical behavior. Directed acyclic graphs and topological sorting, as classic data structures and algorithms, are often used in fields such as computer task scheduling and engineering project management to handle task sequences with dependencies.

[0020] Currently, the closest existing technical solution to this invention is a crowdsourcing task distribution system based on tag matching and single-point path planning, specifically: Step 1: Structured Task Posting (Manual Breakdown): The task initiator (employer) posts the task by filling out a pre-set form on the client side. The system requires the user to manually select the task category (such as "food delivery" or "tutoring"), fill in the specific time period and location, and select the required hard skill tags.

[0021] However, for complex needs (such as "I want to hold a promotional event this weekend"), the existing system cannot understand them automatically. Users must break them down into multiple independent tasks such as "distributing flyers", "organizing goods", and "holding up signs" based on their experience and issue them separately, which has a very high operational threshold.

[0022] Step 2: Static Person-Job Matching (Keyword Matching): The system searches the database for candidates who possess the corresponding skill tags (e.g., "English Level 4" is selected in the user profile) and are within the task radius. However, relying solely on user-filled tags or simple keyword matching lacks in-depth verification of the user's actual abilities (e.g., certificate authenticity, work quality) and cannot identify complex semantic intent.

[0023] Step 3: Point-to-point scheduling for a single task: The system calculates the distance between the candidate's current location and the task location. For instant delivery tasks, algorithms such as Dijkstra's algorithm (a classic graph theory algorithm for solving the single-source shortest path problem) are used to plan the shortest path from the pickup point to the delivery point. For skill-based tasks, only a task list is displayed for users to bid on. This scheduling is typically discrete; the system only cares about how to deliver the current order, without considering the worker's overall daily schedule, resulting in a poor user experience.

[0024] Step 4: Simple task assignment mechanism: Usually adopts a bidding mode (first come, first served) or an instant assignment mode (assigned to the nearest person), lacking a complex team formation mechanism, and unable to automatically assign different sub-tasks of a large project to a "virtual team" composed of multiple people for collaboration.

[0025] In summary, existing flexible staffing platforms face four core bottlenecks when handling complex and non-standardized staffing needs, the specific manifestations and causes of which are as follows: Because the system lacks deep semantic understanding of natural language, it cannot automatically parse ambiguous and complex requirements. As a result, users must manually break down the entire project into multiple structured sub-tasks (such as electrical work, soundproofing, and lighting adjustments) and release them one by one. This process is highly dependent on the user's own project management experience and information technology operation capabilities, which significantly raises the release threshold. It is not only inefficient, but also prone to subsequent execution deviations due to incomplete breakdown or unclear description. In the candidate screening process, the system generally relies solely on user-filled skill tags or simple historical ratings for matching, lacking online verification mechanisms for the authenticity of qualification certificates and intelligent evaluation capabilities for unstructured ability proofs (such as design works and code samples). This forces employers to invest a significant amount of time in manual interviews and qualification reviews after accepting an order, which not only prolongs the screening cycle but also makes it difficult to effectively prevent the risk of resume inflation or mismatched abilities, seriously affecting the reliability of service delivery. Current scheduling logic is mostly limited to point-to-point path planning for a single task or static recommendations based on the current location, failing to model the closed loop of practitioners' all-day spatiotemporal behavior. Therefore, the system cannot predict the feasibility of spatiotemporal connections between tasks, leading to practitioners frequently encountering problems such as cross-regional travel, ineffective back-and-forth trips, or long gaps, resulting in high commuting costs and difficulty in maximizing the number of tasks that can be completed per day and total income; Because the platform lacks the ability to atomically manage dependencies between tasks, all orders are treated as independent events unrelated to each other, making it impossible to construct a directed acyclic graph of tasks that reflects the sequence and collaborative logic. This means that even if a complex project has been successfully broken down, the system cannot automatically assign it to multiple practitioners with complementary skills and coordinate the execution order, making it difficult to support multi-role collaborative delivery and severely restricting the expansion of flexible employment models to high-value, high-complexity service scenarios.

[0026] This application achieves intelligent understanding and automated decomposition of natural language requirements through a large language model, enabling users to publish complex tasks without cumbersome operations, thus lowering the publishing threshold. Through a dual mechanism of interface verification and scoring, it achieves accurate quantitative profiling of candidates' qualifications and abilities, reducing manual screening costs and ensuring matching quality. Utilizing a directed acyclic graph based on spatiotemporal constraints and a topological sorting algorithm, it plans personalized all-day routes for practitioners with the shortest commuting time and the highest continuous income, optimizing scheduling benefits. By constructing task dependencies and forming virtual teams, it enables flexible employment models to handle complex project delivery with time-dependent dependencies.

[0027] The following description, with reference to the accompanying drawings, describes the labor scheduling processing method, apparatus, equipment, medium, and program according to embodiments of this application.

[0028] Specifically, Figure 1 This is a flowchart illustrating a labor scheduling method provided in an embodiment of this application.

[0029] like Figure 1 As shown, the labor scheduling processing method includes the following steps: In step S101, the labor demand input by the labor user is obtained.

[0030] It is understood that the embodiments of this application can use natural language processing technology to semantically analyze the unstructured employment demand text input by the employer. First, the core task intent and key elements are identified. Then, based on preset business rules and knowledge graphs, it is automatically decomposed into several sub-tasks with clear execution standards. For each sub-task, the required skill tags, time windows, geographical locations, and pre- and post-dependent relationships are extracted, and finally, structured task sequence data is formed, providing a complete, accurate, and executable input foundation for subsequent intelligent matching, qualification verification, and path collaborative scheduling.

[0031] It should be noted that the demand for labor includes task objectives or content, skill or qualification requirements, time requirements, geographical location information, quantity or scale, inter-task dependencies or other constraints.

[0032] For example, the task objectives or content should clearly define the specific tasks to be completed, such as "designing promotional posters," "installing 50 air conditioners," or "hosting an offline promotional event." Skills or qualification requirements should specify the professional abilities, certificates, or experience required to perform the task, such as "must hold an electrician's license," "proficient in Photoshop," or "more than 3 years of experience in on-the-ground promotion." Time requirements should include the start time, end time, duration, or time period limits of the task, such as "March 1, 2026, 9:00 AM to 5:00 PM" or "must be completed within 48 hours." Geographical location information should specify the location or area where the task will be performed, such as "an office building in Chaoyang District, Beijing" or "within 5 kilometers of Pudong New Area, Shanghai." Quantity or scale should specify the quantitative indicators involved in the task, such as "recruiting 10 sales promoters," "delivering 200 meals," or "setting up 3 exhibition booths." Dependencies between tasks (for complex needs) should specify the order or logical constraints between multiple sub-tasks, such as "completing the site setup first, then debugging the equipment" or "printing can only proceed after the design draft is confirmed." Other constraints include additional details such as budget limits, availability of tools, whether personal transportation is required, and dress code.

[0033] In step S102, the labor demand is input into the pre-trained language model, and the language model outputs the corresponding task directed acyclic graph. The pre-trained language model performs semantic understanding based on the labor demand and decomposes it into at least one sub-task, and generates the task directed acyclic graph based on the dependency relationship between at least one sub-task.

[0034] It is understood that, in the embodiments of this application, after inputting the labor demand into a pre-trained language model, the model automatically identifies and decomposes multiple sub-tasks constituting the demand based on a deep semantic understanding of natural language. At the same time, it analyzes the correlation between each sub-task in terms of execution logic, time sequence, or resource dependency, and then constructs a directed acyclic graph of tasks that reflects the pre- and post-constraints between tasks. This graph not only structurally expresses the internal composition of complex labor demand, but also provides a computable topological foundation for subsequent multi-role collaborative scheduling, path optimization, and dynamic allocation, thereby realizing the automated transformation from fuzzy, unstructured demand to an executable and schedulable task flow.

[0035] It should be noted that the pre-trained language model includes an input layer, a processing layer, and an output layer. The input layer receives and standardizes the natural language descriptions from the employers. The processing layer includes an intent recognizer, a slot filling engine, a thought chain inferencer, and a dependency extractor. The intent recognizer classifies employment needs from general inquiries and outputs intent labels. The slot filling engine generates follow-up question corpus based on a missing slot table, guiding employers to complete key information such as time, location, scale, and skills. The thought chain inferencer guides the model to break down the execution logic step by step using pre-set prompt word templates. The dependency extractor identifies the sequential logic between tasks and outputs a list of prerequisite tasks. The output layer outputs task dependency graph data that can be directly parsed by the machine.

[0036] The training method includes: collecting historical employment demand texts and corresponding standard task decomposition results generated by expert annotation or rule templates to construct a supervised fine-tuning dataset containing complete demand descriptions, structured atomic task lists, and dependencies; employing mind chain prompting engineering technology to explicitly define the task decomposition logic reasoning process and incorporate it into the training samples, guiding the model to learn a progressive reasoning path from fuzzy demands to ordered tasks; using an instruction fine-tuning framework with the "project manager" role prompt as a fixed prefix to force the model to learn to follow the format specifications of the preset JSON Schema for outputting structured data; introducing adversarial examples and error correction tasks during training to incorporate common format errors, logical loops, and other negative samples and their correction results into the training, enhancing the stability and self-correction capabilities of the model output. Finally, the pre-trained language model is trained to receive natural language employment demands and directly output structured task directed acyclic graph construction data that conforms to a predetermined JSON format and contains a list of atomic tasks and their prerequisite dependencies.

[0037] In this embodiment, the pre-trained language model processing flow is as follows: Employment intent is generated based on the natural language text corresponding to the employment demand; the corresponding employment demand category is determined based on the employment intent, and the employer is guided through multi-turn dialogue interaction to complete the missing key information slots in the employment demand to generate a complete demand description; thought chain prompts are used to guide the language model to decompose the complete demand description into at least one sub-task; corresponding dependencies are established based on the logical relationships between at least one sub-task, and a directed acyclic graph of tasks is constructed based on at least one sub-task and its corresponding dependencies.

[0038] Understandably, in this embodiment, the pre-trained language model first performs semantic parsing on the natural language text input by the employer to generate a clear employment intention. Then, based on this intention, it matches the corresponding employment demand category and actively guides the employer to fill in the missing key information slots through multi-round dialogue interaction, thereby forming a complete demand description with a complete structure and clear semantics. On this basis, the model uses a thought chain prompt mechanism to gradually decompose the complete demand description into several executable sub-tasks. It further analyzes the sequential dependencies of each sub-task in terms of time, resources, or logic, and constructs a directed acyclic graph of tasks that accurately reflects the task execution order and constraints. The entire process realizes end-to-end automated transformation from fuzzy, unstructured raw input to structured, schedulable task topology, significantly improving the understanding accuracy of complex employment demands and the executability of the system.

[0039] For example, the scenario is: Company A needs to hold an offline new product launch event: (1) Input natural language requirements: The user inputs on the platform: I want to hold a smart watch new product launch event for 30 people in Area A next Saturday. I need venue setup, product demonstration, on-site photography and subsequent promotional video editing.

[0040] (2) Generate employment intention and demand classification: The language model identifies the core intention as holding an offline new product launch event, which is classified as a marketing activity demand.

[0041] (3) Multi-round dialogue to fill in key information slots: The system guides the completion of missing information through interactive questions. For example, "What are the specific start and end times of the press conference?", which is filled in as "9:00–17:00"; "Do you need to provide equipment (such as projector, sound system)?", which is answered as "You need to bring your own sound system"; "What is the deadline for delivery of the promotional video?", which is filled in as "Within 48 hours after the event ends"; and finally a complete structured description is formed.

[0042] (4) Mind chain decomposition of subtasks: Guided by mind chain prompts (such as "Please break down the key steps required for this activity step by step in the execution order"), the model outputs the following subtasks: T1: Rent and set up the press conference venue (including tables, chairs, and backdrop); T2: Adjusting the sound system and demonstration equipment; T3: Conduct on-site product explanation and demonstration; T4: Multi-camera photography throughout; T5: Edit and deliver promotional short videos.

[0043] (5) Establish dependencies and construct a directed acyclic graph of tasks: Model analysis of logical dependencies: T1 must be completed before T2, T3, and T4 (the site cannot be set up and subsequent tasks cannot be carried out), T2 must be completed before T3 (the equipment needs to be debugged first), T4 can be executed in parallel with T3, and T5 must be started after T4 is completed; Based on this, construct a directed acyclic graph of tasks, where nodes are subtasks and directed edges represent constraints that must be completed first.

[0044] In step S103, at least one practitioner data is identified, wherein the practitioner data includes geographical location, available time window, skill tags, and competency verification materials.

[0045] It is understood that the embodiments of this application can construct a structured practitioner profile by collecting and integrating multi-dimensional data of practitioners, including their real-time geographical location, available time windows for accepting orders, skill tags, and capability verification materials. Based on this, the system can accurately determine the matching degree of practitioners in terms of both spatiotemporal availability and task competence, providing a highly reliable candidate pool for subsequent sub-task allocation based on the directed acyclic graph of tasks. This identification mechanism not only ensures the rationality and timeliness of person-job matching, but also enhances the objective assessment of the authenticity of practitioners' qualifications and professional level through the introduction of capability verification materials, thereby effectively reducing employment risks and improving the overall service delivery quality.

[0046] It should be noted that the geographical location refers to the practitioner's residential / permanent residence and real-time / dynamic location. The residential / permanent residence serves as the unique start and end point of the personalized spatiotemporal constraint DAG (Directed Acyclic Graph), and is the core constraint for realizing the closed-loop all-weather route planning of the residential location. The real-time / dynamic location is obtained by the system through LBS (Location Based Services) in the proactive order dispatch mode, and is used to calculate the commuting distance and commuting time between the practitioner and the task location.

[0047] A free time window can be a set of deterministic periods of acceptable working hours for a practitioner. Skill tags refer to the professional skills possessed by the practitioner and their quantified proficiency, which can be stored as key-value pairs of skill names and confidence scores. Ability verification materials refer to the original materials submitted by the practitioner to prove the authenticity of their skills and professional level. For certificate materials, OCR (Optical Character Recognition) can be used to identify key information (name, certificate number, issuing institution), and the authenticity can be verified by calling a third-party authoritative API (Application Programming Interface). For portfolio materials, a multimodal large model can be used to score quality according to dimensions such as creativity, technical compliance, and completion, without specific limitations.

[0048] In step S104, the optimal scheduling path is selected from the directed acyclic graph of tasks based on the practitioner data and pushed to the corresponding practitioner's end. The practitioner's end receives and displays the pushed optimal scheduling path and provides a confirmation instruction for accepting the order.

[0049] It is understood that, based on the practitioner's geographical location, available time window, skill tags, and capability certification materials, the system can perform multi-dimensional constraint matching and path optimization on subtasks in the directed acyclic graph of tasks, and select the optimal scheduling path that meets the practitioner's spatiotemporal availability and capability requirements. This path not only satisfies the dependency order between tasks, but also takes into account commuting efficiency and profit maximization, forming a personalized continuous task execution plan. Subsequently, the system pushes the optimal scheduling path to the corresponding practitioner, who receives and visualizes it for confirmation of order acceptance. The practitioner completes the closed-loop interaction by providing feedback on order acceptance confirmation, enabling the system to dynamically lock resources and advance subsequent task scheduling, thereby achieving efficient, accurate, and responsive intelligent order dispatch from task topology to individual execution units.

[0050] It should be noted that the Task Directed Acyclic Graph (DAG) is a structured graph data model output by a pre-trained language model after semantic understanding and logical decomposition of the natural language employment requirements input from the employer. It is used to formally express the execution dependencies between multiple sub-tasks within a complex project. The Task DAG consists of two basic elements: nodes and directed edges, and contains no cycles (i.e., it is impossible to start from a node, traverse along a directed edge, and return to that node). A node is the smallest unit of work extracted from the complex employment requirements and independently dispatchable to a single worker; it can be the task name, an array of required skill tags, estimated working hours, and suggested number of workers. Directed edges represent the temporal constraints between preceding and subsequent tasks, and can identify the list of preceding task names that the current task depends on.

[0051] In this embodiment of the application, the optimal scheduling path is selected from the directed acyclic graph of tasks based on practitioner data and pushed to the corresponding practitioner. This includes: selecting a set of matching candidate subtasks from all subtasks in the directed acyclic graph of tasks based on ability verification materials, skill tags, and free time windows; constructing a target spatiotemporal constrained directed acyclic graph based on any subtask in the candidate subtask set, using geographical location as the starting and ending points of the path; and generating the optimal scheduling path by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph.

[0052] It is understood that this application's embodiments can conduct hard verification and soft scoring of practitioners based on competency verification materials, and accurately screen candidate subtask sets that meet the practitioner's qualifications and time constraints from a directed acyclic graph of tasks by combining skill tags and available time windows. This ensures that tasks entering the scheduling process have a truly feasible entry basis in terms of skill matching and time availability, fundamentally solving the risks of mismatch between people and positions and qualification fraud caused by traditional reliance on self-filled tags; furthermore, using the practitioner's place of residence as a mandatory start and end point, and taking candidate subtasks as intermediate nodes, based on commuting time and task time... The spatiotemporal connectivity inequality of windows constructs a personalized target spatiotemporal constrained directed acyclic graph, accurately transforming the spatiotemporal limitations of the physical world into the existence determination of directed edges in graph theory. This allows for the first time a structured mathematical expression of the combination of tasks that practitioners can continuously execute throughout the day. Finally, the optimal scheduling path is generated by selecting the path with the largest total weight in this directed acyclic graph through topological sorting and dynamic programming algorithms. This path comprehensively weighs the total task reward, total commuting cost, and total skill matching score, achieving a multi-objective joint optimization that maximizes the practitioner's daily income, minimizes commuting distance, and optimizes task suitability.

[0053] In this embodiment of the application, constructing a target spatiotemporal constrained directed acyclic graph based on any subtask in the candidate subtask set includes: identifying the end time of the first subtask in the candidate subtask set, the sum of the commuting time and buffer time from the location of the first subtask to the location of the second subtask, and the start time of the second subtask; if the end time of the first subtask, the sum of the commuting time and buffer time from the location of the first subtask to the location of the second subtask is less than or equal to the start time of the second subtask, then establishing a directed edge between the first subtask and the second subtask; and constructing the target spatiotemporal constrained directed acyclic graph based on at least one directed edge.

[0054] It is understood that the embodiments of this application can identify the end time of the first subtask, the commute time from the location of the first subtask to the location of the second subtask, the buffer time, and the start time of the second subtask. The sum of the end time of the first subtask, the commute time, and the buffer time is compared with the start time of the second subtask. A directed edge is established only when the sum is less than or equal to the start time of the second subtask. This accurately transforms the four real-world constraints faced by practitioners—time continuity constraints, spatial displacement costs, traffic uncertainty, and rigid task time requirements—into graph theory's rules for determining the existence of directed edges. The problem of whether two tasks can be sequentially executed, which originally relied on human experience to determine, has been quantified into a programmable, batch-computable, and dynamically updated mathematical discriminant. This allows for the automatic construction of a target spatiotemporal constrained directed acyclic graph with the worker's residence as the starting and ending point and feasible task sequences as nodes, based on at least one directed edge that meets the conditions. This enables second-level exhaustive search and global optimization of the feasibility of massive task combinations without relying on human scheduling experience. It fundamentally solves the technical bottleneck of traditional flexible employment systems that cannot predict the spatiotemporal feasibility of continuous execution of multiple tasks and can only rely on workers to combine tasks and try and fail.

[0055] It should be noted that intermediate nodes refer to all task nodes in the target spatiotemporal constrained directed acyclic graph constructed in this invention, excluding the starting point (worker's residence) and the ending point (worker's residence). Each intermediate node represents a specific task execution behavior during the worker's entire day's itinerary, including complete business attributes such as task content, location, start and end times, compensation amount, and skill requirements. For example, in a personalized spatiotemporal constrained directed acyclic graph with the worker's residence as the starting and ending points, the graph node located between the starting and ending points represents an executable atomic task; each intermediate node uniquely corresponds to a work task that matches the worker's skills, allows time, and is spatially accessible, carrying multi-dimensional attributes such as task compensation, geographical location, time window, and skill requirements.

[0056] First subtask: A task node located at the beginning of the timeline. After completing this task, the practitioner needs to go to the next task location. Second subtask: A task node located at the end of the timeline. After completing the first subtask, the practitioner needs to commute to the location and start the task on time.

[0057] The first and second subtasks are abstract representations of two sequentially adjacent task nodes when determining whether spatiotemporal connectivity exists between two atomic tasks. They constitute the basis for determining whether a directed edge exists. In the process of constructing a personalized spatiotemporally constrained directed acyclic graph for target practitioners, when performing connectivity determination on any two candidate atomic tasks, the system uses the abstract roles of a pair of task nodes located at the preceding and succeeding positions on the timeline. The system compares the sum of the end time of the first subtask, the real-time commute time from the location of the first subtask to the location of the second subtask, and the preset buffer time with the start time of the second subtask to uniquely determine whether a directed edge exists between them.

[0058] In this embodiment of the application, before constructing the target spatiotemporal constrained directed acyclic graph based on any subtask of the candidate subtask set, the method includes: calculating the straight-line distance between the geographical locations corresponding to any two candidate subtasks; if the straight-line distance between the geographical locations corresponding to any two candidate subtasks is less than or equal to a preset threshold, then the two candidate subtasks are determined to be connected, and then the commuting time from the location of the first subtask to the location of the second subtask is calculated.

[0059] The preset threshold can be set according to actual needs without specific limitations.

[0060] It is understood that, before constructing the target spatiotemporally constrained directed acyclic graph, the embodiments of this application can pre-calculate the straight-line distance between any two candidate sub-task geographical locations and compare it with a preset threshold. Only task pairs with a straight-line distance less than or equal to the threshold are triggered to call the map navigation API to obtain accurate commuting time, while task pairs with a straight-line distance exceeding the threshold are directly determined to be disconnected and pruned. This avoids invalid API (Application Programming Interface) calls and computational resource consumption. Under the premise of ensuring the accuracy of spatiotemporal connectivity determination, the scale of commuting time calculation is compressed from the theoretical full combination to a controllable range, thereby significantly reducing the system's computational load and response latency. This ensures that real-time scheduling for a large number of practitioners and task pools becomes possible, and solves the technical problem of the difficulty in engineering implementation of all-weather route planning caused by the bottleneck of commuting time calculation.

[0061] Specifically, the geographical coordinates of two candidate subtasks are read from the task attributes, and the straight-line distance between the geographical locations of any two candidate subtasks is calculated based on the geographical coordinates.

[0062] In this embodiment of the application, the optimal scheduling path is generated by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph. This includes: obtaining the reward amount of the second sub-task of each directed edge in the target spatiotemporal constrained directed acyclic graph, the commuting cost from the location of the first sub-task to the location of the second sub-task, and the skill matching degree between the worker and the second sub-task; calculating the sum of the comprehensive weights of the directed edges traversed by each complete closed path based on the reward amount, commuting cost, and skill matching degree to generate the total weight of the corresponding path; and selecting the path with the largest total weight from all complete closed paths to generate the optimal scheduling path.

[0063] It is understood that the embodiments of this application can obtain the reward amount of each directed edge pointing to the task in the target spatiotemporal constrained directed acyclic graph, the commuting cost from the current task location to the next task location, and the skill matching degree between the practitioner and the task to be executed. These three factors are then integrated into a comprehensive weight of the directed edges using a weighting function. This unifies the three optimization objectives of maximizing conflicting benefits, minimizing costs, and optimizing skill matching during the practitioner's 24-hour scheduling process into a single quantifiable edge weight index. Based on this, using the practitioner's residence as a mandatory start and end point, all combinations of task nodes satisfying the spatiotemporal connectivity conditions are traversed, and the comprehensive weight of the directed edges traversed by each complete closed path is calculated. The sum of these weights is used as the total weight of the paths. Then, the path with the largest total weight is selected from all feasible paths as the optimal scheduling path. This precisely transforms the scheduling decision problem of flexible workers performing multiple tasks continuously around the clock into a known solvable longest path search problem on a directed acyclic graph. Under the premise of satisfying the triple hard constraints of time, space, and skills, the system can automatically and in real time find the globally optimal travel plan that maximizes the worker's daily comprehensive benefit, minimizes commuting loss, and has the highest skill matching degree from the exponentially growing task combination space. This fundamentally breaks through the technical bottleneck of traditional discrete matching mode, which can only provide locally feasible solutions and cannot guarantee global optimality.

[0064] It should be noted that the reward amount for the second subtask represents the positive benefits brought by the task, the commuting cost from the location of the first subtask to the location of the second subtask represents the time or monetary cost that must be paid to perform the task, and the skill matching degree between the practitioner and the second subtask represents the degree of fit between the task and the practitioner's ability. The higher the matching degree, the higher the expected execution efficiency and delivery quality.

[0065] Specifically, the formula for calculating the total weight is: ; Where W represents the total weight of the path, , and For configurable weighting coefficients, R represents the total reward amount for all sub-tasks in the path; C represents the total commuting cost calculated based on commuting time; and M represents the total skill matching score between all sub-tasks in the path and the practitioner profile.

[0066] In this embodiment of the application, after generating the optimal scheduling path by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph, the process includes: for each segment of the journey between adjacent nodes in the path, re-invoking the map navigation application interface to verify the commuting time based on the latest real-time traffic data; if the verification finds that the actual commuting time between any adjacent nodes exceeds the estimated commuting time due to changes in traffic conditions, and causes the spatiotemporal connectivity conditions between adjacent nodes to no longer be met, then starting from the node that currently does not meet the connectivity conditions, a replacement node that meets the spatiotemporal connectivity conditions and maximizes the total weight of the corrected path is found, and the original path is corrected and updated to generate the final executable scheduling path.

[0067] It is understood that, after generating the initial scheduling path with the highest total weight, this application embodiment can re-invoke the map navigation application interface for the journey between each adjacent node in the path, and use the latest real-time traffic data to accurately verify the commuting time. If it is found that the actual commuting time of a certain segment exceeds the estimated value due to changes in traffic conditions, thereby destroying the original spatiotemporal connectivity conditions, then starting from the node that does not meet the conditions, a replacement node that meets the time window constraints and can maximize the total weight of the corrected path is searched again in the remaining candidate sub-tasks. Through this dynamic verification and local replanning mechanism, the system effectively copes with the scheduling deviation caused by traffic uncertainty in the real environment, ensuring that the final output scheduling path is not only structurally feasible, but also has high reliability and timeliness in actual execution, thereby significantly improving the task fulfillment rate and the operator's experience.

[0068] According to the labor scheduling processing method proposed in this application, the natural language labor demand input by the labor demander is obtained and input into a pre-trained language model. The model performs semantic understanding and logical reasoning on the unstructured demand, automatically decomposes it into multiple atomic sub-tasks with dependencies, and generates a directed acyclic graph of tasks. This transforms the fuzzy and complex labor demand into a set of structured tasks that can be parsed and scheduled by the machine, completely eliminating the reliance on manual decomposition and form filling by the labor demander, and significantly reducing the threshold and error risk of publishing complex tasks. On this basis, the system identifies the laborer's geographical location, free time window, skill tags, and ability verification materials. Taking the laborer's residence as the starting and ending point and the matched sub-tasks as intermediate nodes, the system calculates the task based on commuting time and task time. The system constructs a personalized, spatiotemporally constrained directed acyclic graph based on the spatiotemporal connectivity conditions of the work window. By selecting the path with the highest global weight, it generates the optimal scheduling path that balances task rewards, commuting costs, and skill matching. This path is then pushed to the worker's end to complete order confirmation, thus upgrading the traditional fragmented, discrete order-grabbing model to a closed-loop, 24 / 7 continuous scheduling model based on residence. This effectively solves the problem of lost revenue due to commuting and idle waiting for workers. At the same time, a multi-dimensional verification mechanism of ability verification materials ensures the authenticity and accuracy of person-job matching. Ultimately, it achieves full-process automation and intelligence from demand understanding, task decomposition, ability assessment, path planning to order dispatch and execution, significantly improving the efficiency of supply and demand matching, task delivery quality, and overall worker benefits in flexible employment scenarios.

[0069] The following will combine Figures 2-4 The labor scheduling processing method described in this application is explained in detail, and the specific steps are as follows: Step 1: LLM-based identification of employment demand intent and decomposition of atomic tasks (employer side) Source data reception: The intelligent demand decomposition module receives unstructured natural language descriptions (such as voice or text, for example: "I want to hold a ground promotion event") input by the user through an instant messaging interface.

[0070] Data processing flow: 1) Intent recognition and completion understanding layer: The layer recognizes the intent of the input. If the information is incomplete (lacking scale, cycle, etc.), it guides the employer to fill in the demand slots through multi-turn dialogue. 2) Structured decomposition: The completed demand input is decomposed into multiple atomic tasks with temporal dependencies using a pre-trained language model based on preset prompts. 3) Dependency construction: The task list in JSON format (including task name, skill tag, working hours, and prerequisite relationships) is extracted, and a task dependency DAG is constructed using a graph theory library. 4) Result data: A directed acyclic graph of tasks is generated.

[0071] Step 2: Constructing and pooling practitioner competency profiles based on multi-dimensional verification (practitioner side) Source data reception: The user profile building and real-time analysis module receives practitioners' geographical location, available time windows, skill descriptions, and certificate / work images.

[0072] Data Processing Flow: 1) Triggering and Collection: Through proactive push notifications (geofencing) or passive interaction (user-initiated search), collect practitioners' confirmed free time slots and desired positions using multi-turn dialogues. 2) Hard Qualification Verification: For tasks requiring high skills, verify the authenticity of certificates in real time by recognizing certificates using OCR and calling authoritative third-party API interfaces (such as the China Higher Education Student Information System). 3) Soft Quality Assessment: For non-standard works (such as design drawings and code), use a multimodal large-scale model to perform quality scoring and generate quantified ability scores. 4) Result Data: Generate high-dimensional qualified candidate request objects and send them to the user candidate pool.

[0073] Step 3: Personalized DAG Path Planning for Scheduling Based on Spatiotemporal Constraints (Core Algorithm Engine) Source data reception: The spatiotemporal scheduling module simultaneously obtains atomic tasks to be assigned from the task candidate pool and qualified candidate objects from the user candidate pool.

[0074] Data Processing Flow: 1) Initial Task Screening (Coarse Ranking): Based on skill tags and time overlap, a preliminary subset of matching tasks is selected for each practitioner. 2) Spatiotemporal DAG Construction: Using residence as the starting and ending points and tasks as intermediate nodes, the map API is used to obtain the commuting time between each pair of nodes. Directed edges are established based on spatiotemporal constraint inequalities: the sum of the end time of the first sub-task, the commuting time from the first sub-task location to the second sub-task location, and the buffer time is less than or equal to the start time of the second sub-task. 3) Weight Assignment and Calculation: Edges are assigned a comprehensive weight that integrates task amount, commuting cost, and skill matching degree. The maximum valid path with the highest total weight is calculated using topological sorting and the longest path algorithm (or dynamic programming).

[0075] 4) Results data: Obtain the optimal one-day / multi-day personalized scheduling plan.

[0076] Step 4: Two-stage game-theoretic order dispatch and order status closure (execution layer) Source data reception: The two-stage game dispatch module receives the scheduling plan generated in step 3.

[0077] Data Processing Flow: 1) First Stage (Passive Matching): The system pushes the scheduling path to the workers. If the worker confirms, the system automatically attempts to lock all relevant atomic tasks on the path. 2) Second Stage (Active Search): For tasks that have not been selected for a long time, a scheduled task is triggered. The system expands the radius from the task point as the center, and performs a game-like ranking of nearby users based on reputation, distance, and urgency, actively pushing the task assignment. 3) Status Transition Management: The entire lifecycle of a task is monitored through an order status machine, from pending matching to locking, pending execution, and completion. 4) Delivery Method: Finally, the finalized scheduling plan and assignment results are synchronized in real time to both workers and employers via the App client, enabling cross-role collaborative delivery monitoring within the virtual team.

[0078] Next, the labor scheduling processing apparatus proposed according to the embodiments of this application is described with reference to the accompanying drawings.

[0079] Figure 5 This is a block diagram of the labor scheduling processing device according to an embodiment of this application.

[0080] like Figure 5 As shown, the employment scheduling processing device 10 applied to the cloud scheduling server includes: an acquisition module 100, a processing module 200, an identification module 300, and a filtering module 400.

[0081] The system comprises the following modules: an acquisition module 100 for acquiring employment needs input from the employer; a processing module 200 for inputting employment needs into a pre-trained language model, which outputs a corresponding directed acyclic graph (DAG), wherein the pre-trained language model performs semantic understanding and decomposes the employment needs into at least one sub-task, and generates the DAG based on the dependencies between the at least one sub-task; an identification module 300 for identifying at least one worker data point, which includes geographical location, available time window, skill tags, and competency verification materials; and a filtering module 400 for filtering the optimal scheduling path from the DAG based on the worker data and pushing it to the corresponding worker, wherein the worker receives and displays the pushed optimal scheduling path and provides a confirmation instruction.

[0082] According to the labor scheduling processing device proposed in this application, the system acquires the natural language labor demand input by the labor client and inputs it into a pre-trained language model. This model performs semantic understanding and logical reasoning on the unstructured demand, automatically breaking it down into multiple atomic sub-tasks with dependencies and generating a directed acyclic graph of tasks. This transforms the fuzzy and complex labor demand into a set of structured tasks that can be parsed and scheduled by the machine, completely eliminating the reliance on manual breakdown and form filling by the labor client, and significantly reducing the threshold and error risk of publishing complex tasks. Based on this, the system identifies the worker's geographical location, free time window, skill tags, and ability verification materials. Using the worker's residence as the starting and ending point and the matched sub-tasks as intermediate nodes, it considers commuting time and task time... The system constructs a personalized, spatiotemporally constrained directed acyclic graph based on the spatiotemporal connectivity conditions of the work window. By selecting the path with the highest global weight, it generates the optimal scheduling path that balances task rewards, commuting costs, and skill matching. This path is then pushed to the worker's end to complete order confirmation, thus upgrading the traditional fragmented, discrete order-grabbing model to a closed-loop, 24 / 7 continuous scheduling model based on residence. This effectively solves the problem of lost revenue due to commuting and idle waiting for workers. At the same time, a multi-dimensional verification mechanism of ability verification materials ensures the authenticity and accuracy of person-job matching. Ultimately, it achieves full-process automation and intelligence from demand understanding, task decomposition, ability assessment, path planning to order dispatch and execution, significantly improving the efficiency of supply and demand matching, task delivery quality, and overall worker benefits in flexible employment scenarios.

[0083] The following will combine Figure 6 The technical solution of this application is described below. Specifically, this application employs multiple modules to process the method, including an intelligent demand decomposition module (employer side), a user profile construction and real-time analysis module based on multi-turn dialogue, a spatiotemporal scheduling module, and a two-stage game-theoretic order dispatch module, as detailed below: The intelligent demand decomposition module (for employers) is divided into four layers: interaction layer, understanding layer, reasoning layer, and structuring layer. Each layer collaborates to complete the end-to-end transformation from raw natural language input to a schedulable task flow. Specifically: At the interaction layer, the system interacts naturally with employers through instant messaging and multi-turn dialogue mechanisms. Each session maintains an independent employment demand session, dynamically recording user intent, filled key information slots (such as objectives, deliverables, time requirements, scale, skill preferences, location, etc.), and original input content. This layer follows clear dialogue principles: first, it determines whether the user input constitutes an employment demand; if confirmed, it does not rush to break down the task, but prioritizes identifying missing information; it only conducts precise follow-up questions on key missing slots; the entire process avoids exposing model limitations, ensuring a smooth and natural user experience. For example, when a user enters "I want to organize a shopping mall promotion," the system recognizes this as a valid employment intention, but due to the lack of key information such as the period, format, and specific content, it proactively guides the user to complete the information, laying the foundation for subsequent processing.

[0084] At the understanding layer, the system performs intent recognition and information completeness assessment based on the completed description. Using a pre-defined intent classification framework, the system categorizes requirements into specific business types (such as marketing campaigns, technical services, etc.) and verifies that all necessary information slots are filled, ensuring that the input meets the preconditions for decomposability.

[0085] At the inference layer, the system invokes a pre-trained language model for deep semantic reasoning. This process employs a highly constrained cue word design, explicitly assigning the model the role of a senior project manager, forcing it to output atomic-level subtasks in a structured manner. Each subtask must meet requirements such as independent assignment, no circular dependencies, and inclusion of skill tags and work hours. Guided by thought chains, the model progressively breaks down ambiguous requirements into logically clear execution units and explicitly labels the pre- and post-delay dependencies between tasks. For example, for a 3-day mall promotion, the model outputs tasks including scheme confirmation, material design, booth construction, and on-site execution, establishing a strict execution sequence.

[0086] Finally, at the structured layer, the system rigorously validates and processes the raw output of the model. First, it verifies format compliance, then checks the uniqueness of task names, the existence of dependent tasks, and whether the overall structure constitutes a directed acyclic graph (DAG). If parsing failures, circular dependencies, or missing fields are detected, the system initiates an automatic correction strategy, such as returning an error example for retry or re-entering the conflict context into the model for repair. After successful validation, the system formally constructs the task DAG using graph tools and further generates task instantiation models, identifying whether each task can be executed in parallel (e.g., without prerequisites or sharing the same prerequisite task), providing a basis for subsequent virtual team formation and order distribution.

[0087] (2) User Profile Construction and Real-Time Analysis Module Based on Multi-Turn Dialogue (for Practitioners) This module serves as the user-side front-end entry point of the system. It is primarily responsible for accurately capturing user intent, completing the structured collection of unstructured information through multi-turn dialogue, and dynamically verifying qualifications based on task attributes. The module's operational logic is divided into active triggering mode and passive triggering mode. The specific implementation process is as follows: A dual-mode triggering mechanism is adopted to cover different user behavior scenarios. In proactive mode, the system monitors the task pool around the user's geographical location in real time based on LBS. Once a high-priority or high-yield popular task is detected and the user is reachable, a task inquiry card is proactively pushed. If the user expresses interest, the information verification process begins immediately. In passive mode, when the user actively inputs natural language, the system uses an intent classifier to determine whether the user intends to seek part-time employment. If it is job-related (such as mentioning free time or skills), a multi-turn dialogue process is initiated. If it is only a general inquiry, the knowledge base provides a direct response without entering the scheduling chain.

[0088] The system then enters a multi-turn dialogue information collection phase based on slot filling. Targeting the intent to "find a part-time job," the dialogue engine focuses on three core slots: available time window, skill tags, and desired job area. The system analyzes the information already provided by the user within the context, precisely asks follow-up questions about missing or ambiguous content, and automatically maps colloquial expressions such as "after get off work" to standardized time intervals, ensuring clear time boundaries for subsequent scheduling.

[0089] Dynamic qualification verification and information completion mechanism. The system not only checks the completeness of basic fields such as real-name authentication, but also queries the "Skill-Certificate Rule Base" in real time based on the user's declared skills or the category of the task being responded to, to identify whether there are mandatory entry qualifications (such as a driver's license for drivers and a health certificate for nannies). Once a rule is matched and there is no valid verification record in the user profile, the system immediately guides the user to upload a certificate photo in the dialogue and triggers an automated backend processing flow: extracting key information from the certificate through OCR, calling authoritative third-party interfaces to verify authenticity and validity, providing real-time feedback and updating the user profile. For non-standard skills such as design and programming, the system further introduces a multimodal large model to conduct expert scoring on the user's submitted works or code, generating quantitative ability indicators from dimensions such as creativity, standardization, and completion, enriching the soft assessment dimensions.

[0090] Once a user's available time, skill tags, and all associated core qualifications have been verified, the system encapsulates this into a qualified candidate request object, injects it into the backend dynamic matching pool, and immediately triggers the directed acyclic graph (DAG) scheduling engine. Based on the candidate's spatiotemporal availability and capability profile, the scheduling engine searches the DAG for the optimal scheduling path that satisfies dependency constraints and commuting feasibility, thereby achieving a seamless connection from user intent expression to executable task assignment.

[0091] (3) Spatiotemporal scheduling module: In the initial screening of candidates and the generation of scheduling paths, the system first uses skill tag matching and idle time period filtering to quickly select a set of candidate tasks that meet the preliminary conditions from the global task pool using a distributed search and analysis engine or vector database, thus completing coarse-grained matching.

[0092] Based on this, the system constructs a personalized spatiotemporal scheduling directed acyclic graph for individual workers. This graph uses the worker's residence as the starting and ending point, and all candidate tasks selected in the initial screening as intermediate nodes. The establishment of edges in the graph relies on a key connectivity judgment logic: for any two task nodes A and B, the system calls the map navigation API to obtain the actual commuting time from task A to task B, and verifies this time by combining it with the end time of A and a preset buffer time; only when task A is completed and can arrive at the start time window of task B in time is a directed edge from A to B allowed to be established. This process essentially transforms the physical commuting constraints of the real world into topological feasibility conditions in graph theory.

[0093] To support optimal path decision-making, each directed edge is assigned a comprehensive weight, which is a linear combination of three parts: task reward (positive incentive), commuting cost (negative penalty), and skill matching degree (ability adaptability). The weight is flexibly adjusted by configurable coefficients to allow for adjustments based on platform strategies or user preferences.

[0094] After the graph is constructed, the system performs a longest path search on the DAG (since the DAG is acyclic, dynamic programming can be used for efficient solution). The goal is to find the feasible path with the largest total weight from the starting point to the ending point. This path represents the continuous task execution sequence with the highest daily revenue, the most reasonable commute, and the best match to the user's capabilities.

[0095] Finally, the system renders this optimal scheduling route in natural language (such as "Execute task A from 9:00 AM to 12:00 PM, execute task B from 1:00 PM to 5:00 PM, and return home after completion") and pushes it to the worker for them to confirm and accept the order.

[0096] The core of the entire process lies in the construction logic of the spatiotemporal DAG—which successfully maps multi-dimensional real-world constraints such as geospatial constraints, time windows, and task dependencies to a computable graph structure. The biggest challenge, however, lies in performance optimization: real-time API calls for commuting time between pairs of tasks are costly and can easily become a bottleneck in scenarios with a large number of candidate tasks. Therefore, the system needs to introduce efficient pruning strategies, such as quickly eliminating obviously unreachable task pairs using a straight-line distance threshold (e.g., 5 kilometers) to avoid invalid API requests; or pre-calculating commuting matrices between high-frequency areas using historical data, thereby significantly improving response speed and system scalability while ensuring scheduling accuracy.

[0097] (4) Two-stage game order dispatch module (execution layer) The order state machine clearly defines the key state transition path from order generation to completion: initially, it is in the "pending matching" state, indicating that the task has been published but not yet included in any scheduling scheme; when a practitioner confirms acceptance of the recommended scheduling path, all tasks included in the path enter the "locked" state to prevent them from being duplicated by other users; subsequently, when execution begins at the agreed time, the state is updated to "pending execution"; after the task is completed, it finally becomes "completed". This state machine provides fundamental guarantees for task concurrency control, exception rollback, and data consistency.

[0098] Based on this, the system implements intelligent matching logic in two stages. The first stage is passive matching: after the practitioner confirms the scheduling path generated by the personalized spatiotemporal DAG algorithm through the front end, the system automatically attempts to lock all related tasks on that path at once, and simultaneously reserves idle resources for the corresponding time period of the practitioner, ensuring the atomicity and executability of the scheduling plan.

[0099] The second phase is a proactive search mechanism for handling long-tail or less popular tasks. The system deploys a scheduled monitoring task that continuously scans the task pool. If a task has not been selected by any scheduling path for a period greater than or equal to a preset threshold (e.g., 30 minutes) since its publication, a radius expansion strategy is triggered: the original matching radius is dynamically expanded (e.g., increased by 5 kilometers) centered on the task's geographical location, proactively searching for available workers within a larger area and pushing task notifications to them. Once the user responds, the system immediately generates a local scheduling path containing the task and locks it.

[0100] During the proactive search phase, to improve matching quality and task success rate, the system introduces game-theoretic ranking logic to prioritize candidate practitioners. The ranking considers not only the fundamental factor of geographical distance but also integrates a user's historical order acceptance rate as a reputation indicator (reflecting fulfillment reliability) and the urgency of the task as a business weight (e.g., tasks nearing their deadlines are given higher priority). Through multi-dimensional weighted evaluation, the system prioritizes pushing urgent tasks to users with high reputation and high suitability, thereby expanding coverage while ensuring timely delivery of critical tasks.

[0101] In summary, the intelligent demand decomposition module (employer side) stably and verifiably transforms the natural language demands of the employer side into a structured task dependency DAG, and further maps it into an executable set of virtual team orders, and incorporates subtasks into a pool of candidates that still have no candidates, providing a foundation for subsequent order dispatching, matching, and scheduling. The user profile building and real-time parsing module based on multi-turn dialogue: This module serves as the user-side front-end entry point of the system. Its main responsibility is to accurately capture user intent, complete the structured collection of unstructured information through multi-turn dialogue, and perform dynamic qualification verification in conjunction with task attributes. The module's operational logic is divided into active triggering mode and passive triggering mode. Users with clear intent and complete information (such as available time and desired part-time job type) who pass the qualification verification are added to the user candidate pool, providing a foundation for subsequent order assignment, matching, and scheduling.

[0102] Spatiotemporal Scheduling Module: This invention's spatiotemporal scheduling module jointly models the skills, free time, and geographical location of flexible workers. It first performs initial screening of a massive number of tasks, then constructs a personalized spatiotemporal directed acyclic graph (DAG) around the user's start and return points as boundaries. Based on real-world map commuting times, it determines the reachability between tasks. Furthermore, it introduces a comprehensive weighting function that integrates task benefits, commuting costs, and skill matching for connected edges. Through topological sorting and the longest path search algorithm on the DAG, it automatically calculates the executable task combinations and execution order with the optimal total benefit under real-world time and space constraints, thereby generating feasible one-day or multi-day scheduling plans for workers. Simultaneously, it locks the user's corresponding time window and matching tasks, and generates orders to provide data support for subsequent business settlements.

[0103] The two-stage game-theoretic task assignment module, acting as the execution layer, uses an order state machine to finely manage the entire process of tasks from pending matching, locking, pending execution, to completion. In the first stage, a passive matching mechanism is employed. Once a worker confirms the optimal scheduling path recommended by the system, multiple tasks within that path are automatically locked, ensuring the feasibility and consistency of the scheduling plan. In the second stage, an active search mechanism is introduced. By periodically monitoring tasks that have not been selected for an extended period, a dynamic search radius expansion strategy is triggered, proactively assigning tasks to available workers within a wider range who meet the time and skill requirements. During this process, a game-theoretic ranking and decision-making process is implemented, considering distance, historical order acceptance rate, and task urgency. This significantly improves the overall fulfillment rate of high-time-sensitive tasks while ensuring a positive user experience.

[0104] Figure 7 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 701, the processor 702, and the computer program stored on the memory 701 and executable on the processor 702.

[0105] When the processor 702 executes the program, it implements the labor scheduling processing method provided in the above embodiments.

[0106] Furthermore, electronic devices also include: Communication interface 703 is used for communication between memory 701 and processor 702.

[0107] The memory 701 is used to store computer programs that can run on the processor 702.

[0108] The memory 701 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0109] If the memory 701, processor 702, and communication interface 703 are implemented independently, then the communication interface 703, memory 701, and processor 702 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0110] Optionally, in a specific implementation, if the memory 701, processor 702, and communication interface 703 are integrated on a single chip, then the memory 701, processor 702, and communication interface 703 can communicate with each other through an internal interface.

[0111] The processor 702 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0112] This application also provides a computer-readable storage medium storing a computer program or instructions thereon, which, when executed by a processor, implements the above-described labor scheduling processing method.

[0113] This application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described labor scheduling processing method.

[0114] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0115] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0116] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0117] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0118] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A labor scheduling method, characterized in that, The method is applied to a cloud-based scheduling server, and includes the following steps: Obtain the employment requirements input by the employer; The labor demand is input into a pre-trained language model, and the language model outputs a corresponding task directed acyclic graph. The pre-trained language model performs semantic understanding based on the labor demand and decomposes it into at least one sub-task, and generates a task directed acyclic graph based on the dependency relationship between at least one sub-task. Identify at least one practitioner data point, wherein the practitioner data includes geographical location, available time window, skill tags, and competency verification materials; Based on the practitioner data, the optimal scheduling path is selected from the directed acyclic graph of tasks and pushed to the corresponding practitioner's end. The practitioner's end receives and displays the pushed optimal scheduling path and provides a confirmation instruction for accepting the order.

2. The labor scheduling processing method according to claim 1, characterized in that, The step of selecting the optimal scheduling path from the directed acyclic graph of tasks based on the practitioner data and pushing it to the corresponding practitioner includes: Based on the capability verification materials, the skill tags, and the idle time window, a set of matching candidate subtasks is selected from all subtasks in the directed acyclic graph of the task. Using the geographical locations as the starting and ending points of the path, construct a target spatiotemporal constrained directed acyclic graph based on any subtask in the candidate subtask set; The optimal scheduling path is generated by selecting the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph.

3. The labor scheduling processing method according to claim 2, characterized in that, The step of constructing a target spatiotemporally constrained directed acyclic graph based on any subtask of the candidate subtask set includes: Identify the end time of the first subtask in the candidate subtask set, the sum of the commute time and buffer time from the location of the first subtask to the location of the second subtask, and the start time of the second subtask; If the sum of the end time of the first subtask, the commute time from the location of the first subtask to the location of the second subtask, and the buffer time is less than or equal to the start time of the second subtask, then a directed edge is established between the first subtask and the second subtask. Construct a target spatiotemporally constrained directed acyclic graph based on at least one directed edge.

4. The labor scheduling processing method according to claim 3, characterized in that, Before constructing the target spatiotemporally constrained directed acyclic graph based on any subtask from the candidate subtask set, the following steps are included: Calculate the straight-line distance between the geographical locations corresponding to any two candidate subtasks; If the straight-line distance between the geographical locations corresponding to any two candidate subtasks is less than or equal to a preset threshold, then the two candidate subtasks are determined to be connected, and the commute time from the location of the first subtask to the location of the second subtask is calculated.

5. The labor scheduling processing method according to claim 4, characterized in that, The process of generating the optimal scheduling path by filtering the path with the largest total weight in the target spatiotemporal constrained directed acyclic graph includes: Obtain the reward amount of the second subtask for each directed edge in the target spatiotemporal constrained directed acyclic graph, the commuting cost from the location of the first subtask to the location of the second subtask, and the skill matching degree between the practitioner and the second subtask. The total weight of the corresponding path is generated by summing the comprehensive weights of the directed edges traversed by each complete closed path based on the reward amount, commuting cost, and skill matching. The optimal scheduling path is generated by selecting the path with the highest total weight from all complete closed paths.

6. The labor scheduling processing method according to claim 1, characterized in that, The processing flow of the pre-trained language model: Generate employment intent based on the natural language text corresponding to the employment demand; Based on the employment intention, the corresponding employment demand category is determined, and the employment client is guided to fill in the missing key information slots in the employment demand to generate a complete demand description through multi-round dialogue interaction. The language model is guided by thought chain prompts to break down the complete requirement description into at least one sub-task. Establish corresponding dependencies based on the logical relationships between at least one subtask, and construct the task directed acyclic graph based on at least one subtask and its corresponding dependencies.

7. A labor dispatching and processing device, characterized in that, The device is used in a cloud-based scheduling server, wherein the device includes: The acquisition module is used to acquire the employment requirements input by the employer. The processing module is used to input the employment demand into a pre-trained language model, and the language model outputs a corresponding task directed acyclic graph. The pre-trained language model performs semantic understanding based on the employment demand and decomposes it into at least one sub-task, and generates a task directed acyclic graph based on the dependency relationship between at least one sub-task. An identification module is used to identify at least one practitioner data, wherein the practitioner data includes geographical location, free time window, skill tags, and ability verification materials; The filtering module is used to filter the optimal scheduling path from the directed acyclic graph of tasks based on the practitioner data and push it to the corresponding practitioner's end. The practitioner's end receives and displays the pushed optimal scheduling path and provides a confirmation instruction for accepting the order.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the workforce scheduling processing method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they are used to implement the labor scheduling processing method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the labor scheduling processing method as described in any one of claims 1-6.

Images