Intelligent recruitment process control method and system based on task state and action type

Abstract

This invention provides an intelligent recruitment process control method and system based on task status and action type. The method includes the following steps: acquiring multi-dimensional input data of the recruitment process and converting it into a standardized task status feature vector, wherein the task status feature vector includes at least a process progress quantification value, candidate interaction feedback parameters, resource occupancy rate, and target achievement degree; constructing a collaborative decision-making system of preset decision rules and reinforcement learning model; this invention dynamically adjusts the execution strategy by collecting multi-dimensional task status features in real time and combining them with the adaptation logic of candidate action types, which can flexibly respond to various changes in the recruitment scenario and achieve accurate matching between the process and the real-time business scenario.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent recruitment process control method and system based on task status and action type. Background Technology

[0002] Existing recruitment automation systems mostly adopt fixed processes or simple rule-driven approaches, with process nodes pre-set. This makes it difficult to dynamically adjust the process based on candidate behavior, Q&A results, or platform feedback, resulting in rigid processes, low automation levels, and an inability to adapt to the ever-changing communication and decision-making needs in real recruitment scenarios.

[0003] For example, Chinese invention patent CN202111560554.3 discloses a recruitment control method, system, and readable storage medium based on remote technology. Although it introduces technologies such as remote audio and video data acquisition, interviewee authenticity verification, and neural network model-assisted interview question recommendation, which optimizes the information interaction efficiency of the interview process to a certain extent, its core is still limited to the local functional optimization of the interview stage. It has not broken through the constraints of the preset process framework. It does not involve the dynamic connection and scheduling logic of multiple stages such as reaching, screening, and inviting in the entire recruitment process, nor can it dynamically adjust the process direction and execution actions based on data such as real-time feedback from candidates and task execution status. It still suffers from insufficient process flexibility and weak full-process automated collaboration capabilities, making it difficult to cover the complex decision-making needs of multi-module collaboration and dynamic switching of multiple states in real recruitment scenarios. Summary of the Invention

[0004] In view of this, the present invention provides an intelligent recruitment process control method and system based on task status and action type, which has the advantages of being able to dynamically adjust the recruitment process according to real-time task status and action execution feedback, realizing full-process automated connection and intelligent scheduling, and improving recruitment efficiency and adaptability.

[0005] To solve the above-mentioned technical problems, the present invention adopts a technical solution as follows: an intelligent recruitment process control method based on task status and action type, comprising the following steps: S1, collecting multi-dimensional input data of the recruitment process and extracting task status features reflecting the process execution status, wherein the multi-dimensional input data includes process progress data, candidate interaction data, resource consumption data, and target achievement data; S2, constructing a collaborative decision-making system, wherein the collaborative decision-making system is based on a collaborative mechanism of preset decision rules and a reinforcement learning model that integrates task status and action type, and dynamically adjusts the decision logic according to the task status features; S3, based on the adaptation logic of the task status features and candidate action types, outputting the target action type and execution priority, wherein the candidate action type is configured as a result. S4. Construct action units, each action type corresponds to at least one set of instantiable task execution nodes; S5. Generate a dynamically iterative task execution sequence based on the target action type, execution priority, and process dependencies, wherein the task execution sequence is composed of the set of task execution nodes and includes dependencies between nodes; S6. Collect action execution feedback data, update task state characteristics, evaluate task state changes through a state transition judgment mechanism, and trigger a structural-level reconstruction of the current task execution sequence when a state mutation is determined, wherein the structural-level reconstruction includes at least regenerating at least one of the task execution node set, dependencies between nodes, or execution order; if the process termination condition is not met, return to S2 for iterative execution, and if it is met, archive the execution results.

[0006] As a further preferred embodiment of this technical solution: In S1, the construction logic of the task status features includes: defining feature dimensions, which include, but are not limited to, process stage identifiers, candidate response characteristics, operation time characteristics, anomaly occurrence, resource load level, and target completion progress; standardizing the multi-dimensional input data to eliminate the differences in the dimensions of the data; and generating unified task status features by weighted integration based on the influence weight of each feature dimension on the recruitment process control.

[0007] As a further preferred embodiment of this technical solution, the construction steps of the collaborative decision-making system include: pre-setting decision rules for handling deterministic recruitment scenarios, and using a reinforcement learning model that integrates task states and action types to handle dynamically changing scenarios; the collaborative decision-making system is configured with decision switching logic, and when a pre-set trigger condition is met, the decision weight ratio of the pre-set decision rules and the reinforcement learning model that integrates task states and action types is adjusted.

[0008] As a further preferred embodiment of this technical solution: In S3, the adaptation logic based on the task state features and candidate action types is as follows: based on the attribute correlation between the task state features and candidate action types, the degree of adaptation between the two is quantified, and the attribute correlation is determined based on the feature semantic adaptability and functional matching; the evaluation logic of the execution priority includes: comprehensively considering the action execution success rate, resource consumption cost, target contribution and execution time indicators to determine the execution order of candidate action types.

[0009] As a further preferred embodiment of this technical solution: In S4, the dynamic iterative task execution sequence generation logic includes: constructing a task execution topology based on the dependency relationship and jump rules of the target action type; embedding an elastic scheduling mechanism, activating the alternative action set and adjusting the task execution topology and node order when an abnormal execution state is detected; the task execution sequence is dynamically adjusted following the update of task state characteristics.

[0010] As a further preferred embodiment of this technical solution: In S5, the state transition determination mechanism includes: a preset state change threshold, and a determination of state stability by quantifying the degree of difference between the updated task state characteristics and the current state characteristics; wherein: if the degree of difference is lower than the state change threshold, the current task execution sequence is maintained; if the degree of difference is not lower than the state change threshold, it is determined to be a state mutation, and a structural-level reconstruction of the current task execution sequence is triggered.

[0011] As a further preferred embodiment of this technical solution: In S5, the process termination condition includes at least one of the following: the target completion progress in the task status characteristics reaches a preset target threshold; the difference in task status characteristics in multiple consecutive iterations is lower than a preset stability threshold; an unrecoverable exception is triggered, and the action execution condition cannot be met even after exception handling; the core objective of the recruitment process is achieved, including obtaining a preset number of qualified candidates or completing a preset number of effective interactions.

[0012] Another technical solution adopted in this invention is: an intelligent recruitment process control system based on task status and action type, comprising: a status feature extraction module: used to collect multi-dimensional input data of the recruitment process, extract and standardize task status features; a collaborative decision-making module: configured with a collaborative decision-making system, outputting target action type and execution priority based on task status features; a task scheduling module: used to generate a dynamically iterative task execution sequence, and schedule the function module to execute the target action; a feedback update module: used to collect action execution feedback data, update task status features, and perform state transition determination; and a result archiving module: used to terminate the process and archive the execution results when the process termination conditions are met.

[0013] As a further preferred embodiment of this technical solution, the collaborative decision-making module further includes: a rule configuration unit for storing and calling preset decision rules to handle deterministic recruitment scenarios; a model learning unit for training a reinforcement learning model that integrates task states and action types using historical data to optimize decision logic in dynamic scenarios; and a weight adjustment unit for adjusting the decision weights of the preset decision rules and the reinforcement learning model that integrates task states and action types according to triggering conditions.

[0014] As a further preferred embodiment of this technical solution, the task scheduling module also includes a backup action activation unit, which is used to call the backup action set and dynamically adjust the task execution sequence when an abnormal execution state is detected.

[0015] The embodiments of the present invention have the following advantages due to the adoption of the above technical solutions:

[0016] I. This invention collects multi-dimensional task status features in real time and dynamically adjusts the execution strategy by combining the adaptation logic of candidate action types. It can flexibly respond to various changes in recruitment scenarios and achieve accurate matching between processes and real-time business scenarios.

[0017] Second, this invention adopts a collaborative decision-making system that combines preset decision-making rules with a reinforcement learning model that integrates task states and action types. This system ensures efficient and accurate decision-making in deterministic scenarios, while also adapting the learning model to dynamically changing complex scenarios. This avoids the limitations of a single decision-making mode, enabling recruitment decisions to balance standardization and flexibility, and improving the overall rationality of the decision-making process.

[0018] Third, this invention clarifies action dependencies and execution priorities through the generation of dynamic task execution sequences and elastic scheduling mechanisms, reduces redundant steps and invalid operations in the process, and at the same time reduces the risk of process interruption by using exception handling and alternative action activation logic, ensuring the continuous progress of the recruitment process and improving the overall efficiency of process operation.

[0019] Fourth, this invention evaluates the correlation and priority of task status and action type attributes, and directs HR resources and channel resources toward actions with high adaptability and high goal contribution, avoiding resource waste in inefficient operations, achieving precise matching of resources and recruitment needs, and improving overall resource utilization efficiency.

[0020] Fifth, this invention identifies abnormal states and potential risks in the process in advance through a state transition judgment mechanism and flexible scheduling logic. With the help of alternative action sets and execution sequence reconstruction functions, it effectively resolves various unexpected problems in process execution, reduces process stagnation or deviation caused by abnormal situations, and ensures the overall stability and predictability of the recruitment process.

[0021] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the collaborative decision-making switching process according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the state transition and task chain adjustment process according to an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0027] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0028] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0029] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0030] Traditional recruitment automation systems typically employ fixed processes or simple rules, resulting in rigid processes, low automation levels, and difficulty in dynamically adjusting based on candidate behavior or platform feedback. While existing technologies have improved certain functionalities, they remain constrained by pre-defined process frameworks, failing to achieve dynamic connection and scheduling across the entire recruitment process. This leads to insufficient process flexibility and weak full-process automation and collaboration capabilities, making it difficult to cope with complex and ever-changing recruitment scenarios.

[0031] In this regard, such as Figure 1-3 As shown, this invention proposes an intelligent recruitment process control method based on task status and action type, including the following steps: S1, collecting multi-dimensional input data of the recruitment process and extracting task status features reflecting the process execution status. The multi-dimensional input data includes process progress data, candidate interaction data, resource consumption data, and target achievement data; S2, constructing a collaborative decision-making system. The collaborative decision-making system is based on a collaborative mechanism of preset decision rules and a reinforcement learning model that integrates task status and action type, dynamically adjusting the decision logic according to the task status features; wherein, the reinforcement learning model that integrates task status and action type can be a reinforcement learning model, and the following description uses a reinforcement learning model as an example; S3, based on the adaptation logic of task status features and candidate action types, outputting the target action type and execution priority, wherein, The candidate action types are configured as structured action units, and each action type corresponds to at least one set of instantiable task execution nodes; S4, generate a dynamically iterative task execution sequence based on the target action type, execution priority, and process dependencies, wherein the task execution sequence is composed of the set of task execution nodes and includes dependencies between nodes; S5, collect action execution feedback data, update task state features, evaluate task state changes through a state transition judgment mechanism, and trigger a structural-level reconstruction of the current task execution sequence when a state mutation is determined, wherein the structural-level reconstruction includes at least regenerating at least one of the task execution node set, dependencies between nodes, or execution order; if the process termination condition is not met, return to S2 for iterative execution, and if it is met, archive the execution results.

[0032] Specifically, this invention provides an intelligent recruitment process control method based on task status and action type. The method first acquires multi-dimensional input data of the recruitment process and transforms it into a standardized task status feature vector. For example, the multi-dimensional input data of the recruitment process can be manually entered, with recruiters periodically inputting process progress, candidate feedback, resource usage, and current goal completion rate. This raw data can then be directly concatenated into a non-standardized feature list as a simple representation of the task status. Process progress can be simply represented as the current stage number, candidate interaction feedback can be recorded as a Boolean value of "responded" or "not responded," resource utilization can be recorded as the current percentage of CPU or memory used, and goal achievement can be recorded as the number of candidates who have completed the process.

[0033] Furthermore, this method constructs a collaborative decision-making system based on preset decision rules and a reinforcement learning model. The preset decision rules are set as the first priority, and the reinforcement learning model is set as the adaptive supplementary priority. The model's decision is triggered when a switching condition is met. For example, the collaborative decision-making system can be constructed using simple conditional logic. The preset decision rules can consist of a series of hard-coded "if-then" statements, such as "If the candidate's response time exceeds 24 hours, send a reminder email." The reinforcement learning model can be a basic Q-learning model, whose decision is only invoked when the preset decision rules cannot provide a clear instruction. The switching condition can be set as follows: the reinforcement learning model is only activated when the system detects three consecutive failures of the preset decision rules. In this case, the preset decision rules always have the highest priority, and the reinforcement learning model serves only as a backup.

[0034] Furthermore, this invention proposes training steps for a reinforcement learning model, specifically including:

[0035] Collect task status feature vectors, action execution records, feedback results, and process goal achievement status from historical recruitment processes to build an experience replay pool;

[0036] Initialize Q network parameters and set the learning rate. Discount factor ;

[0037] Randomly sample training samples from the experience replay pool and calculate the target Q-value: ,in For instant rewards, For the next state, The action for the next state, For target network parameters;

[0038] The network parameters are iteratively updated until convergence by minimizing the loss function between the predicted Q value and the target Q value using gradient descent.

[0039] The parameters of the target network and the online network are synchronized every preset number of iterations to ensure model stability.

[0040] Building upon this foundation, this method, based on a pre-defined action type adaptation matrix and state-action mapping model, outputs the target action type and action selection priority according to the task state feature vector. Subsequently, based on the target action type, action selection priority, and the current task state feature vector, an iteratively updatable task chain is generated. This dynamic task chain defines the relationships and transition logic between execution nodes through a state transition matrix. For example, the task chain can be generated using a predefined template. For different target action types, the system can pre-define several fixed task execution sequences. When the target action type is "scheduling," the task chain might be fixed as "send interview invitation" -> "wait for candidate confirmation" -> "arrange interviewer." The state transition matrix can be a simple two-dimensional array, manually defining the fixed action node to which each action might jump after execution, without involving complex topology structures or dynamic adjustments.

[0041] Next, the corresponding functional modules are scheduled to execute the target action according to the execution sequence of the dynamic task chain, and action execution feedback data is collected in real time to update the task status feature vector. For example, the scheduling of functional modules can adopt a sequential execution method. When an action in the task chain is determined, the system will directly call the corresponding functional module to execute it, such as calling the email sending module to send an email, or calling the calendar management module to schedule a meeting. Action execution feedback data can be simply recorded as a Boolean value of "success" or "failure" and manually entered into the system. Updating the task status feature vector can be a simple addition or subtraction operation on one of its dimensions, such as incrementing the process progress by 1, without involving complex recalculation or standardization processes.

[0042] Finally, a similarity matching process is performed on the updated task state feature vector using a state transition determination mechanism. If a preset termination condition is met, the process terminates and the results are archived; otherwise, the process returns to the collaborative decision-making step for iterative execution. For example, the state transition determination mechanism can be based on a simple threshold comparison. The system can preset a fixed state change threshold. When the difference between the updated task state feature vector and the previous state vector in a certain key dimension (such as process progress) exceeds this threshold, it is determined that the state has changed. Similarity matching can be simplified to directly comparing whether the values ​​of a specific dimension in two vectors are equal. The preset termination condition can include only a single condition, such as terminating the process when the target completion progress of the recruitment process reaches 100%. If the termination condition is not met, the system unconditionally returns to the collaborative decision-making step for the next iteration.

[0043] This invention achieves precise perception of the recruitment process status by acquiring multi-dimensional state data and converting it into standardized feature vectors. Through collaborative decision-making using pre-defined decision rules and a reinforcement learning model, it balances decision stability and adaptability, effectively addressing the complexities and variability of recruitment scenarios. The generation and iterative updates of dynamic task chains allow the recruitment process to flexibly adjust based on real-time feedback, overcoming the limitations of traditional fixed processes. Therefore, this invention significantly improves the automation and flexibility of the recruitment process, enhances the collaborative decision-making capabilities across the entire process, and better adapts to the changing communication and decision-making needs of real-world recruitment scenarios.

[0044] Specifically, in S1, the construction logic of task status features includes: defining feature dimensions, which include but are not limited to process stage identifiers, candidate response characteristics, operation time characteristics, anomaly occurrence, resource load level, and target completion progress; standardizing multi-dimensional input data to eliminate the difference in the units of measurement of data from different dimensions; and generating unified task status features by weighted integration based on the influence weight of each feature dimension on the recruitment process control.

[0045] More specifically, the multi-dimensional input data is standardized using normalization, and the normalization formula is:

[0046]

[0047] in These are the original eigenvalues. The minimum value of this dimension. The maximum value of this dimension is determined; weight coefficients are assigned to each normalized feature dimension, which are determined using the analytic hierarchy process (AHP), and the sum of all dimension weights is 1; the task state feature vector is obtained by weighted summation, and its vector expression is: ,in For the first Weight coefficients for each dimension For the first Normalized eigenvalues ​​in each dimension.

[0048] The above technical solution clearly defines several key dimensions of the recruitment process status and normalizes these dimensions to eliminate the impact of differences in units and numerical ranges, ensuring the comparability of each feature. Based on this, a scientifically reasonable weight is assigned to each dimension using the analytic hierarchy process (AHP), enabling the task status feature vector to more accurately reflect the actual contribution and importance of each dimension to the recruitment process status. This systematic and standardized approach to constructing task status feature vectors provides high-quality, high-precision input for the subsequent collaborative decision-making system, significantly improving the accuracy and robustness of preset decision rules and reinforcement learning models in judging the current recruitment process status. Simultaneously, it allows the action type adaptation matrix and state-action mapping model to output more accurate target action types and action selection priorities based on more realistic and refined state information, thereby optimizing the generation and execution of dynamic task chains. Ultimately, this improves the overall efficiency and decision-making quality of the intelligent recruitment process control method, enabling it to more effectively cope with the complexity and uncertainty of the recruitment process.

[0049] Specifically, the construction steps of the collaborative decision-making system include: pre-setting decision rules to handle deterministic recruitment scenarios, and using a reinforcement learning model that integrates task states and action types to handle dynamically changing scenarios; configuring decision switching logic in the collaborative decision-making system, and adjusting the decision weight ratio of the pre-set decision rules and the reinforcement learning model that integrates task states and action types when the pre-set trigger conditions are met.

[0050] Specifically, the preset decision rules include state threshold rules, action trigger rules, and process jump rules, used to handle deterministic recruitment scenarios. The reinforcement learning model uses the recruitment process conversion rate as the reward objective, constructs a training sample set based on historical task state data, action execution records, and feedback results, and generates a state-action value function through training. Quantitative parameters for switching conditions are set: the preset threshold for the amount of historical interaction data is no less than 1000 records, and the preset threshold for the process execution success rate is no less than 85%. When the switching conditions are not met, the decision result is output only through the preset decision rules. When the switching conditions are met, the results of the preset decision rules and the state-action value function output of the reinforcement learning model are integrated, and the final decision result is calculated according to the preset weight ratio threshold.

[0051] By clearly defining pre-defined decision-making rules, including state threshold rules, action triggering rules, and process jump rules, this invention can effectively handle deterministic scenarios in the recruitment process, ensuring the stability and controllability of the basic process. Simultaneously, the reinforcement learning model, with the recruitment process conversion rate as the reward objective, learns from historical data and generates state-action value functions, providing the system with adaptive optimization capabilities. When the amount of historical interaction data reaches a pre-defined threshold and the process execution success rate falls below a set threshold, the system can intelligently switch from single-rule decision-making to collaborative decision-making that integrates pre-defined decision rules and the reinforcement learning model. This allows for the introduction of reinforcement learning's exploration and optimization capabilities when data accumulation is sufficient and process performance is poor, effectively improving the adaptability and overall conversion rate of the recruitment process. This collaborative decision-making mechanism ensures process stability while achieving intelligent optimization for complex and dynamic recruitment scenarios, making recruitment process control more efficient and precise.

[0052] Specifically: In S3, the adaptation logic based on task state features and candidate action types is as follows: based on the attribute correlation between task state features and candidate action types, the degree of adaptation between the two is quantified, and the attribute correlation is determined based on the feature semantic adaptability and functional matching; the evaluation logic for execution priority includes: comprehensively considering the success rate of action execution, resource consumption cost, target contribution and execution time indicators to determine the execution order of candidate action types.

[0053] Based on the above solution, this invention further proposes an action type adaptation mechanism, wherein the action type adaptation matrix is ​​as follows: Among them, row dimension Clustering categories corresponding to task state feature vectors, column dimensions Corresponding action type, matrix elements Indicates the first Class state corresponds to the first The fitness score for action types is calculated using cosine similarity.

[0054] Specifically: In S4, the dynamic iterative task execution sequence generation logic includes: constructing a task execution topology based on the dependency relationship and jump rules of the target action type; embedding an elastic scheduling mechanism to activate the alternative action set and adjust the task execution topology and node order when an abnormal execution state is detected; and dynamically adjusting the task execution sequence according to the update of task state characteristics.

[0055] Based on the above solution, this invention further proposes a method for generating iteratively updatable task chains. The steps for generating this dynamic task chain specifically include:

[0056] Based on the output of the state-action mapping model, the core action and alternative action set of the current node are determined; a task chain topology is constructed through a graph neural network, where nodes are action execution units and edges represent the dependencies and jump probabilities between nodes; the nodes in the topology are sorted according to the action selection priority to generate an initial task chain; an elastic scheduling factor is embedded, which activates the alternative action set and adjusts the topology and node order when an abnormal state occurs during task execution, thereby realizing the dynamic iteration of the task chain.

[0057] Through the above technical solution, this invention, when generating the recruitment process task chain, not only considers the core actions in the current state but also pre-determines a set of alternative actions, thereby enhancing the robustness of the process. By introducing a graph neural network to construct the topology of the task chain, it can more comprehensively and intelligently capture the complex dependencies and jump probabilities between actions, making the generation of the initial task chain more accurate and optimized. More importantly, by embedding an elastic scheduling factor, this invention can monitor abnormal states during task execution in real time and promptly activate the set of alternative actions, dynamically adjusting the topology and node order of the task chain. This makes the recruitment process no longer a rigid linear execution but possesses a high degree of adaptability and fault tolerance, effectively coping with various unexpected situations and uncertainties that may arise during the recruitment process, avoiding process interruptions or inefficiencies, and ensuring that recruitment goals can be continuously advanced and ultimately achieved.

[0058] Specifically, in S5, the state transition determination mechanism includes: a preset state change threshold, which determines state stability by quantifying the degree of difference between the updated task state characteristics and the current state characteristics; wherein: if the degree of difference is lower than the state change threshold, the current task execution sequence is maintained; if the degree of difference is not lower than the state change threshold, it is determined to be a state mutation, triggering a structural-level reconstruction of the current task execution sequence.

[0059] Specifically, this state transition determination mechanism includes:

[0060] Preset state transition threshold The Euclidean distance between the updated task state feature vector and the current node state vector is calculated. Determine the direction of migration;

[0061] when When the state is deemed stable, the current task chain execution sequence is maintained.

[0062] when When this occurs, it is determined to be a state change, triggering dynamic task chain reconstruction to regenerate execution nodes and related relationships;

[0063] The formula for calculating Euclidean distance is:

[0064]

[0065] For the updated version 3D eigenvalues For the current node Dimensional eigenvalues.

[0066] By introducing the aforementioned state transition determination mechanism, this invention can accurately quantify and intelligently determine the changes in task states during the recruitment process.

[0067] Specifically, in S5, the process termination conditions include at least one of the following: the target completion progress in the task status features reaches a preset target threshold, which can be set to, for example, 95%. Specifically, when the target completion progress in the task status feature vector reaches 95% or higher, the system can promptly identify the near-completion state of the process, avoiding redundant operations when the target has been basically achieved, thereby saving computing resources; the difference in task status features over multiple consecutive iterations is lower than a preset stability threshold, which can be set to 0.1. For example, when the Euclidean distance d of the state feature vectors in three consecutive iterations is less than 0.1, the system can accurately determine whether the process has fallen into a stable or stagnant state, effectively preventing meaningless loop iterations and improving the convergence efficiency of the process; an unrecoverable exception is triggered, and the action execution conditions cannot be met even after exception handling; the core objectives of the recruitment process are achieved, including obtaining a preset number of qualified candidates or completing a preset number of effective interactions.

[0068] like Figure 4 As shown, another technical solution adopted by the present invention is: an intelligent recruitment process control system based on task status and action type, including: a status feature extraction module: used to collect multi-dimensional input data of the recruitment process, extract and standardize task status features; specifically, this module can be configured with a data collection interface to connect to recruitment management systems (such as BOSS Zhipin, Zhaopin API), HR management systems and candidate interaction platforms, and collect resume data, interview data, resource usage data, etc. in real time; furthermore, this module has a built-in feature processing unit, for example, using the Pandas library of Python to perform data cleaning, using the StandardScaler class or MinMaxScaler class in the Scikit-learn library to complete standardization processing (corresponding to the normalization formula in the initial solution), and then using feature selection algorithms (such as variance thresholding method, mutual information method) to filter core feature dimensions, and finally output a task status feature vector in a unified format.

[0069] Collaborative Decision-Making Module: Configures the collaborative decision-making system and outputs the target action type and execution priority based on task status characteristics; this module includes:

[0070] Rule configuration unit: Used to store and call preset decision rules to handle deterministic recruitment scenarios; the rule configuration unit uses a database to store preset decision rules, for example, a rule table is built through a MySQL database, with fields including "trigger condition", "action type", "weight ratio", etc., and supports adding, deleting, modifying and querying rules through a visual interface. Rule matching is implemented using regular expressions or decision tree algorithms;

[0071] Model Learning Unit: Used to train a reinforcement learning model that integrates task state and action type using historical data, optimizing decision-making logic in dynamic scenarios. The model learning unit can use the reinforcement learning model in the initial scheme as the core implementation. For example, a DQN (Deep Q-Network) model can be built based on the TensorFlow framework, with task state features as input, action type as output, and process conversion rate as reward signal. The model parameters can be trained using historical recruitment data (such as job recruitment data from the past year) to optimize the Q-value calculation logic (corresponding to the Q-value calculation formula in the initial scheme).

[0072] Weight Adjustment Unit: Used to adjust the decision weights of preset decision rules and the reinforcement learning model that integrates task states and action types based on trigger conditions. The weight adjustment unit dynamically outputs the weight ratio of rules and models by calculating trigger condition indicators (such as historical data scale and process execution effect) in real time. For example, when the historical sample size reaches 1000, the weight ratio is automatically adjusted to 0.2 for rules and 0.8 for models, and the weight parameters are synchronized to the decision calculation unit through the API interface.

[0073] The task scheduling module is used to generate dynamically iterative task execution sequences and schedule the target actions. Specifically, this module has a built-in topology building unit and a backup action activation unit. For example, the topology building unit can use the NetworkX library to visualize the task chain and clarify the dependencies between each action node. The backup action activation unit has a pre-set set of backup actions (such as backup actions for "resume screening failure: readjusting screening rules and expanding recruitment channels"). When the anomaly detection unit (such as an anomaly recognition algorithm based on threshold judgment) detects an abnormal state such as "interview scheduling conflict", it automatically calls the set of backup actions and adjusts the task execution sequence through a greedy algorithm or a genetic algorithm to ensure continuous execution of the process.

[0074] Feedback Update Module: This module collects action execution feedback data, updates task state features, and performs state transition determination. Specifically, it is configured with data acquisition probes to monitor the execution results of functional modules in real time (such as whether the interview invitation was successfully sent or whether the candidate confirmed their participation in the interview). The feedback data is cleaned by a data preprocessing unit (such as outlier removal and missing value filling). Furthermore, the state transition determination unit evaluates the state by calculating the degree of feature difference. For example, it uses Manhattan distance or Chebyshev distance to quantify the feature changes before and after the update, compares them with a preset threshold, and outputs the instruction to "maintain sequence" or "reconstruct sequence," which is then synchronously fed back to the collaborative decision-making module and the task scheduling module.

[0075] The results archiving module is used to terminate the process and archive the execution results when the process termination conditions are met. Specifically, this module can connect to enterprise cloud storage services (such as Alibaba Cloud OSS and Tencent Cloud COS) to archive process execution data (such as recruitment progress reports, candidate information, and decision logs) in preset formats (such as Excel and JSON). At the same time, it generates a process execution report, which includes core indicators (such as total process time, conversion rate, and number of exceptions handled) and optimization suggestions (such as "the interview response rate for a certain position is low, and it is recommended to optimize the invitation script"), providing data support for subsequent recruitment process optimization.

[0076] For other details regarding the implementation technical solutions of each module in the above embodiment system, please refer to the description in the intelligent recruitment process control method based on task status and action type in the above embodiment, which will not be repeated here.

[0077] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system-type embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0078] The following example will provide a more detailed explanation of the above technical solution:

[0079] In a smart recruitment scenario, a company needs to hire a senior software engineer. Traditional recruitment systems typically follow a pre-defined process, such as posting a job, resume screening, phone interviews, technical interviews, and HR interviews. However, in practice, factors such as candidate backgrounds, feedback, and the availability of recruitment resources are dynamic and change, making fixed processes ineffective. This method aims to address the problems of rigid processes, low automation, and inability to adapt to changing recruitment needs in existing technologies.

[0080] First, the system acquires multi-dimensional input data from the recruitment process, such as job descriptions, candidate resumes, historical recruitment data, the current recruitment stage (e.g., "initial resume screening"), and available interviewer resources. This raw data is transformed into a standardized task state feature vector. Specifically, the system defines a set of state feature dimensions, including process stage identifiers, candidate response rates, operation completion times, anomaly frequency, resource load coefficients, and target completion progress. For example, in the "initial resume screening" stage, the process stage identifier might be 0.1, the initial candidate response rate 0%, the anomaly frequency 0%, the resource load coefficient low, and the target completion progress 0%. The system normalizes these dimensional features and assigns weight coefficients to each normalized feature dimension. These weight coefficients are determined using the analytic hierarchy process (AHP). Finally, the current task state feature vector S is calculated through weighted summation.

[0081] Next, the system constructs a collaborative decision-making system combining preset decision rules and a reinforcement learning model. In the early stages of the recruitment process, due to the limited amount of historical interaction data (e.g., less than the preset threshold of 1000 records) and the lack of clear statistics on the success rate of process execution, the system primarily relies on preset decision rules for decision-making. These preset decision rules include preset state threshold rules, action triggering rules, and process jump rules, used to handle deterministic recruitment scenarios. For example, in the "resume initial screening" stage, the preset decision rules will trigger actions such as "resume parsing" and "preliminary screening."

[0082] Based on the current task state feature vector S, the system utilizes a pre-defined action type adaptation matrix and a state-action mapping model to output the target action type and action selection priority. The action type adaptation matrix is ​​an Mm×n matrix, where the row dimension m corresponds to the clustering category of the task state feature vector, and the column dimension n corresponds to the action type (including parsing, calculation, scheduling, and reach). Matrix element Mij represents the adaptation score of the j-th action corresponding to the i-th state, calculated using cosine similarity. For example, for the "resume initial screening" state, the adaptation matrix might indicate a high adaptation between "parsing" actions (such as "resume information extraction") and "scheduling" actions (such as "arranging initial screening"). Simultaneously, the system constructs an action priority evaluation index system, including action execution success rate, resource consumption cost, target contribution weight, and execution time, and calculates the comprehensive priority score for each candidate action using the weighted TOPSIS method. For example, the "resume information extraction" action might receive a high priority due to its low resource consumption and high target contribution.

[0083] Based on the target action type, action selection priority, and current task state feature vector, the system generates an iteratively updatable task chain. For example, in the "initial resume screening" stage, the system might generate an initial task chain: "resume information extraction," "keyword matching and filtering," and "preliminary candidate scoring." This dynamic task chain uses a graph neural network to construct a topology, where nodes are action execution units and edges represent dependencies and jump probabilities between nodes. The nodes in the topology are sorted according to action selection priority to generate the initial task chain.

[0084] Subsequently, the system schedules the corresponding functional modules to execute the target actions according to the execution sequence of the dynamic task chain. For example, the "resume information extraction" module is executed first to automatically extract key information from a large number of resumes. During the execution of the actions, the system collects action execution feedback data in real time, such as "resume parsing success rate" and "information extraction completeness", and updates the task status feature vector accordingly.

[0085] After updating the task state feature vector, the system performs similarity matching through a state transition determination mechanism. The system presets a state transition threshold δ and determines the transition direction by calculating the Euclidean distance d between the updated task state feature vector and the current node's state vector.

[0086] Suppose that during the "keyword matching and screening" phase, the system discovers that a large number of resumes are eliminated due to keyword mismatch, resulting in a far lower than expected number of qualified candidates. At this point, the "process progress quantification value" in the updated task status feature vector may stagnate, "candidate interaction feedback parameters" (such as the proportion of valid resumes) may significantly decrease, and the "goal achievement rate" may be low. The calculated Euclidean distance d may be greater than or equal to a preset threshold δ, indicating a state abrupt change. This contrasts with existing technologies that cannot dynamically adjust the process based on real-time feedback; existing systems would continue to perform subsequent screenings, leading to inefficiency.

[0087] When a state change is detected, the system triggers a dynamic task chain reconstruction. At this time, the system activates a set of alternative actions, such as "adjusting keyword weights" or "expanding resume sourcing channels," and adjusts the topology and node order of the task chain to achieve dynamic iteration. For example, the new task chain might become: "Adjust keyword weights" → "Re-match and screen keywords" → "Initial candidate scoring." This dynamic adjustment capability significantly improves the flexibility and adaptability of the process.

[0088] As the recruitment process progresses, the system continuously collects historical interaction data. When the amount of historical interaction data reaches a preset threshold (e.g., no less than 1000 records) and the process execution success rate (e.g., the conversion rate from resume screening to interview) is lower than a set threshold (e.g., no less than 85%), the switching condition is met. At this point, the system will integrate the results of the preset decision rules with the state-action value function output of the reinforcement learning model, and calculate the final decision result according to the preset weight ratio. The reinforcement learning model uses the recruitment process conversion rate as the reward objective, constructs a training sample set based on historical task state data, action execution records, and feedback results, and generates the state-action value function through training. For example, at a certain stage, the preset decision rules may suggest conducting a "telephone interview," but the reinforcement learning model, based on historical experience, finds that for a specific type of candidate, conducting an "online written test" directly can achieve a higher conversion rate. In this case, the system will integrate the suggestions of both to make a better decision. This enables the system to handle more complex and uncertain recruitment scenarios, overcoming the problem of "weak full-process automated collaboration capabilities" in existing technologies.

[0089] The entire process continues iteratively until preset termination conditions are met. These preset termination conditions include: the target completion progress in the task state feature vector reaches or exceeds 95%; the Euclidean distance d of the state feature vectors for three consecutive iterations is less than 0.1 (indicating the process has stabilized); an unrecoverable exception is triggered, and the action execution conditions cannot be met even after exception handling; or the recruitment process goal has been achieved, such as obtaining a preset number of qualified candidates and completing their onboarding. Once any termination condition is met, the system will terminate the process and archive the results.

[0090] Using the methods described above, this system can dynamically adjust decisions and actions based on the real-time status and feedback of the recruitment process. This effectively solves the technical problems of existing recruitment automation systems, such as rigid processes, low automation levels, and difficulty in adapting to changing recruitment scenarios, and achieves intelligent and adaptive control of the recruitment process.

[0091] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for intelligent recruitment process control based on task status and action type, characterized in that, Includes the following steps: S1. Collect multi-dimensional input data of the recruitment process and extract task status features that reflect the execution status of the process. The multi-dimensional input data includes process progress data, candidate interaction data, resource consumption data and target achievement data. S2. Construct a collaborative decision-making system, which is based on a collaborative mechanism of preset decision rules and a reinforcement learning model that integrates task state and action type, and dynamically adjusts the decision logic according to the task state characteristics; S3. Based on the adaptation logic of the task state characteristics and candidate action types, output the target action type and execution priority, wherein the candidate action type is configured as a structured action unit, and each action type corresponds to at least one set of instantiable task execution nodes. S4. Generate a dynamically iterative task execution sequence based on the target action type, execution priority, and process dependencies, wherein the task execution sequence consists of the set of task execution nodes and includes the dependencies between nodes; S5. Collect action execution feedback data, update task status characteristics, evaluate task status changes through a state transition judgment mechanism, and trigger structural reconstruction of the current task execution sequence when a state mutation is determined. The structural reconstruction includes at least regenerating at least one of the following: task execution node set, node dependencies, or execution order. If the process termination condition is not met, return to S2 for iterative execution; if it is met, archive the execution results. The state transition determination mechanism includes: A preset state change threshold is used to determine state stability by quantifying the difference between the updated task state characteristics and the current state characteristics. in: If the degree of difference is lower than the state change threshold, the current task execution sequence is maintained. If the degree of difference is not lower than the state change threshold, it is determined to be a state mutation, triggering a structural-level reconstruction of the current task execution sequence; The process termination conditions include at least one of the following: The target completion progress in the task status characteristics has reached the preset target threshold; The degree of difference in task state characteristics across multiple consecutive iterations is lower than a preset stability threshold; An unrecoverable exception is triggered, and the conditions for action execution cannot be met even after exception handling; The core objectives of the recruitment process are achieved, including obtaining a predetermined number of qualified candidates or completing a predetermined number of effective interactions.

2. The intelligent recruitment process control method based on task status and action type according to claim 1, characterized in that: In S1, the logic for constructing task state features includes: Define feature dimensions, which include, but are not limited to, process stage identifier, candidate response characteristics, operation time characteristics, anomaly occurrence, resource load level, and target completion progress; Standardize multi-dimensional input data to eliminate the differences in the units of measurement between different dimensions of data; Based on the influence weights of each feature dimension on recruitment process regulation, a unified task status feature is generated through weighted integration.

3. The intelligent recruitment process control method based on task status and action type according to claim 1, characterized in that: The steps for constructing the collaborative decision-making system include: Pre-defined decision rules are used to handle deterministic recruitment scenarios, while reinforcement learning models that integrate task states and action types are used to handle dynamically changing scenarios. The collaborative decision-making system is configured with decision-switching logic. When preset triggering conditions are met, the decision weight ratio of the preset decision rules and the reinforcement learning model that integrates task states and action types is adjusted.

4. The intelligent recruitment process control method based on task status and action type according to claim 1, characterized in that: In step S3, the adaptation logic based on the task state features and candidate action types is as follows: Based on the attribute correlation degree between task state features and candidate action types, the degree of fit between the two is quantified. The attribute correlation degree is determined based on the feature semantic adaptability and functional matching. The evaluation logic for execution priority includes: taking into account the success rate of action execution, resource consumption cost, target contribution, and execution time, to determine the execution order of candidate action types.

5. The intelligent recruitment process control method based on task status and action type according to claim 1, characterized in that: In S4, the dynamic iterative task execution sequence generation logic includes: Based on the dependencies and jump rules of the target action type, construct the task execution topology; An embedded elastic scheduling mechanism is used to activate a set of alternative actions and adjust the task execution topology and node order when an abnormal execution state is detected. The task execution sequence is dynamically adjusted in accordance with the updates of task state characteristics.

6. An intelligent recruitment process control system based on task status and action type, characterized in that, include: Status feature extraction module: used to collect multi-dimensional input data of the recruitment process, extract and standardize task status features; Collaborative Decision Module: Configures the collaborative decision-making system and outputs the target action type and execution priority based on task status characteristics; Task scheduling module: Used to generate dynamically iterative task execution sequences, and the scheduling function module executes the target action; Feedback Update Module: Used to collect action execution feedback data, update task status characteristics, and perform state transition determination; Results archiving module: Used to terminate the process and archive the execution results when the process termination conditions are met.

7. The intelligent recruitment process control system based on task status and action type according to claim 6, characterized in that: The collaborative decision-making module also includes: Rule configuration unit: used to store and recall preset decision rules to handle deterministic recruitment scenarios; Model learning unit: used to train a reinforcement learning model that integrates task state and action type using historical data, and to optimize decision-making logic in dynamic scenarios; Weight adjustment unit: used to adjust the decision weights of the reinforcement learning model that integrates the preset decision rules and the task state and action type according to the triggering conditions.

8. The intelligent recruitment process control system based on task status and action type according to claim 6, characterized in that: The task scheduling module also includes a backup action activation unit, which is used to call a set of backup actions and dynamically adjust the task execution sequence when an abnormal execution state is detected.

Images