An intelligent recruitment reach decision method and device
By combining causal inference models and resource optimization scheduling models with exploratory intervention strategies, a recruitment outreach necessity score and action sequence are generated, which solves the problem of insufficient decision-making in existing systems under multiple constraints and realizes intelligent and adaptive optimization of the recruitment process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ENGLISH SHI INTERCONNECTION BEIJING INFORMATION TECH CO
- Filing Date
- 2026-02-05
- Publication Date
- 2026-06-16
Smart Images
Figure CN121660402B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent recruitment technology, and in particular to an intelligent recruitment outreach decision-making method and apparatus. Background Technology
[0002] In large-scale recruitment scenarios, the recruitment process typically involves multiple consecutive stages, including resume screening, initial communication, scheduling interviews, sending offer letters, and confirming employment. Recruiters need to proactively contact job seekers via phone, instant messaging, and other means to guide them to the next stage. However, existing recruitment management systems lack sufficient intelligence in reaching decision-makers and executing strategies.
[0003] First, most existing systems train predictive models based on historical data to determine the probability of a job seeker advancing to the next stage. However, such correlation-based predictions fail to answer the core business question: does this proactive contact truly and significantly increase the probability of progress? In other words, such correlation-based predictions lack an assessment of causality.
[0004] Secondly, existing recruitment outreach decision-making models typically assume no limitations on outreach resources (such as the number of calls that can be made per day), failing to consider real-world business constraints such as recruiters' daily work quotas, limits on the frequency of contact with the same job seeker, and minimum coverage rates for job seekers at each stage to ensure process health. Therefore, while these models can evaluate the value of a single recruitment outreach, they cannot generate a globally optimal, immediately executable daily recruitment outreach list under multiple constraints.
[0005] Finally, the long-term operation of existing recruitment outreach strategies is prone to strategy solidification, that is, the system repeatedly selects the same type of high-scoring job seekers, which narrows the distribution of model training data, distorts the estimation of causal effects, makes it unable to explore and adapt to new recruitment strategies or job seeker behavior patterns, and gradually degrades the system performance.
[0006] In summary, existing technical solutions suffer from technical deficiencies in recruitment outreach decisions, such as incomplete causal inference, lack of resource constraint handling capabilities, and the absence of proactive exploration mechanisms leading to long-term performance degradation of the recruitment outreach decision-making system. There is an urgent need for a more intelligent, robust, and practically business-constrained automated decision-making solution for the recruitment process. Summary of the Invention
[0007] In view of the above-mentioned defects or deficiencies in the prior art, the present invention provides an intelligent recruitment outreach decision-making method and apparatus to solve the technical problems mentioned in the background art.
[0008] One aspect of the present invention provides an intelligent recruitment outreach decision-making method for a decision server, comprising the following steps:
[0009] Retrieve job seeker attribute data, historical recruitment behavior logs, and recruitment stage information from the database server, and construct a daily updated state vector for each job seeker;
[0010] Based on the state vector, the probability difference between performing recruitment outreach actions on job seekers and not performing recruitment outreach actions within a specified time in the future is estimated using a causal inference model.
[0011] Based on the probability difference, the recruitment outreach necessity score of the job seeker is calculated according to the fusion calculation rules;
[0012] The recruitment outreach necessity score, preset recruitment business constraints, and decision status information are input into the resource optimization scheduling model. During the decision-making process of the resource optimization scheduling model, an exploratory intervention strategy is used to perturb its decision to generate a recruitment outreach action sequence. Among them, some outreach actions in the recruitment outreach action sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation of the exploratory intervention strategy.
[0013] The recruitment outreach action sequence is sent to the terminal device for execution, and outreach feedback data is collected from the application server associated with the terminal device and stored in the database server;
[0014] Based on the reach feedback data, the causal inference model is incrementally trained, the weight parameters in the fusion calculation rules are optimized, and the strategy of the resource optimization scheduling model is updated through reinforcement learning.
[0015] In another aspect, the present invention provides an intelligent recruitment outreach decision-making device for a decision server, comprising:
[0016] The vector building module is used to retrieve job seeker attribute data, historical recruitment behavior logs, and recruitment stage information from the database server, and build a daily updated state vector for each job seeker;
[0017] The causal inference module is used to estimate, based on the state vector, the probability difference between performing recruitment outreach actions on job seekers and not performing recruitment outreach actions within a specified time in the future through a causal inference model.
[0018] The recruitment outreach scoring module is used to calculate the recruitment outreach necessity score of job seekers based on the probability difference and according to the fusion calculation rules.
[0019] The recruitment outreach action sequence generation module is used to input the recruitment outreach necessity score, preset recruitment business constraints, and decision state information into the resource optimization scheduling model. During the decision-making process of the resource optimization scheduling model, an exploratory intervention strategy is used to perturb its decision to generate a recruitment outreach action sequence. Among them, some outreach actions in the recruitment outreach action sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation of the exploratory intervention strategy.
[0020] The execution and feedback module is used to send the recruitment outreach action sequence to the terminal device for execution, and to collect outreach feedback data from the application server associated with the terminal device and store it in the database server;
[0021] The model optimization module is used to incrementally train the causal inference model based on the reach feedback data, optimize the weight parameters in the fusion calculation rules, and update the strategy of the resource optimization scheduling model through reinforcement learning.
[0022] The intelligent recruitment outreach decision-making method and device provided by this invention have the following beneficial effects:
[0023] (1) Significantly improved the accuracy of decision-making
[0024] By using a causal inference model to remove confounding factors, the incremental effect of recruitment outreach behavior is directly estimated. Furthermore, multi-dimensional dynamic adjustments are made based on the value of the recruitment stage, recruitment costs, job seeker fatigue, recruitment urgency, and the value of future natural progression, generating a comprehensive recruitment outreach necessity score. This upgrades the decision-making basis from "potentially effective" to "whether it is necessary and highly valuable in the current business context," significantly improving the business relevance and accuracy of individual recruitment outreach decisions and resolving the technical problem of one-sided decision-making basis in existing technologies.
[0025] (2) Improved global optimization capability under complex constraints
[0026] By modeling daily recruitment outreach decisions as a constrained Markov decision process or a similar resource optimization scheduling problem, and solving it using reinforcement learning methods, this invention can simultaneously consider and strictly satisfy multiple real-world business constraints. The recruitment outreach decision system outputs executable recruitment outreach actions that maximize overall expected revenue under given constraints, achieving a leap from value ranking to optimal scheduling under limited resources, and solving the technical challenge that existing decision systems cannot directly generate executable solutions under real constraints.
[0027] (3) Improved the system’s long-term self-evolution and anti-degeneration capabilities.
[0028] This invention creatively introduces an active exploration mechanism. By allocating exploration quotas and implementing strategies such as diversity exploration, strategy perturbation exploration, and constraint-aware exploration, the system can proactively acquire sparse regions, boundary cases, and strategy deviation samples. This exploratory data is continuously fed back to update the causal inference model, optimize the weight parameters in the fusion calculation rules, and iteratively update the resource optimization scheduling model's strategy, forming a reinforced closed loop of decision-making, execution, exploration, and learning. This effectively prevents causal estimation distortion and strategy rigidity caused by narrowing data distribution, ensuring the system can adapt to changes in job seeker behavior patterns and discover new effective strategies, significantly improving the system's long-term robustness and adaptability. Attached Figure Description
[0029] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:
[0030] Figure 1 This is a flowchart illustrating an intelligent recruitment outreach decision-making method provided in one embodiment of this application;
[0031] Figure 2 This is a system logic diagram of an intelligent recruitment outreach decision-making method provided in one embodiment of this application;
[0032] Figure 3 This is a schematic diagram of the structure of an intelligent recruitment outreach decision-making device provided in another embodiment of this application;
[0033] Figure 4 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] One embodiment of this application provides an intelligent recruitment outreach decision-making method. The core of this method lies in constructing a complete intelligent decision-making closed loop integrating state awareness, causal evaluation, business correction, constraint scheduling, proactive exploration, and continuous learning. In this embodiment, "recruitment outreach" refers to the proactive and recorded communication initiated by the recruiter, such as a recruitment consultant, HR, or automated tool, to target job seekers in order to advance the recruitment process.
[0036] like Figure 1 and Figure 2 As shown, this intelligent recruitment outreach decision-making method includes the following steps:
[0037] Step S101: Constructing the job seeker's state vector.
[0038] This step aims to address the problem that traditional models cannot distinguish the recruitment stage and dynamic context of job seekers, resulting in single feature representations and a lack of temporal and stage information. By adopting a three-dimensional structured modeling method based on job seeker-stage-date, the instantaneous state of a job seeker during the recruitment process is encoded into a high-dimensional, computable state vector, providing a unified feature input for subsequent causal inference and decision-making.
[0039] Specifically, the system retrieves multi-source data daily from associated database servers, including:
[0040] 1) Static data of job seekers, such as work experience, education, and skill tags;
[0041] 2) Job seeker dynamic data, such as recent platform login frequency and resume update behavior;
[0042] 3) Historical recruitment behavior logs, such as every phone call from HR, communication records on instant messaging software with their timestamps and communication summaries;
[0043] 4) Recruitment stage information, which accurately records the current stage of the job seeker (such as the initial interview) and the time when they entered that stage.
[0044] Based on the above data, a state vector S(c, stage, t, X) is constructed for each job seeker c on each day t. Here, stage represents the current stage, and X is a feature vector that integrates multiple information sources. The construction of the feature vector X includes, but is not limited to:
[0045] 1) Stage-specific temporal characteristics. For example: duration of stay in the current stage, time taken to advance from the previous stage to the current stage, etc.
[0046] 2) Reaching historical characteristics. For example: the number of days since the last time being reached, the total number of times reached in the last N days, and the distribution of reach methods.
[0047] 3) Inherent characteristics of job seekers. For example, the aggregation of static attributes and recent dynamic behaviors.
[0048] 4) Contextual features. Such as weekdays / weekends, peak recruitment seasons / off-peak seasons, etc.
[0049] Through this step, the system transforms unstructured, multi-source, and time-series-related recruitment process data into machine-understandable, standardized state representations, providing a high-quality and information-complete input foundation for subsequent algorithm models.
[0050] Step S102: Evaluation of the incremental effect of recruitment outreach.
[0051] This step employs a causal inference model, transforming the modeling objective into the incremental impact of recruitment outreach on the outcome of stage progression. Specifically, based on the state vector, the causal inference model estimates the probability difference between performing recruitment outreach to job seekers and not performing recruitment outreach within a specified future timeframe, thus advancing the recruitment stage.
[0052] The input to this step is the state vector S(c,t) output from step S101. Its core is training and applying a causal inference model that estimates the following probability differences:
[0053] Uplift(c,t) = P(Y=1 | contact=1, S(c,t)) - P(Y=1 | contact=0, S(c,t))
[0054] Where Y=1 indicates that the job seeker has progressed to the next stage of recruitment (or finally joined the company) within a specified time window (e.g., within T days); contact=1 indicates that an active contact action was performed on the job seeker; contact=0 indicates that no active contact action was performed.
[0055] Since traditional machine learning models learn P(Y|X), their high probability predictions may stem from the job seeker's strong intention. In other words, the process may proceed even without contacting the job seeker, rather than due to the outreach behavior. The causal inference model described above, however, directly estimates the net increase probability brought about by the outreach behavior by distinguishing the differences in outcomes between the treatment group (outreaching / contacting the job seeker) and the control group (not outreaching / not contacting the job seeker) under the same state vector S, thus eliminating the confounding effects of the job seeker's own attributes.
[0056] The causal inference model for this step outputs the Uplift(c,t) value for each job seeker daily, typically ranging from (-1,1), where:
[0057] Uplift > 0: This indicates that contact can increase the probability of advancement, and it is recommended to reach out.
[0058] Uplift ≈ 0: This means that whether or not there is a connection has little impact, and resources can be saved.
[0059] Uplift < 0: This indicates that contact may cause aversion, reducing the probability of advancement, and contact should be avoided.
[0060] This provides the most basic causal effect signal for subsequent decision-making models, namely: probability difference.
[0061] Step S103: Calculate the necessity score for recruitment outreach.
[0062] This step aims to address the issue that the causal inference signal output in step S102 is too simplistic and fails to incorporate actual business considerations such as value, cost, and urgency. Specifically, by designing a fusion calculation rule, the causal effect (probability difference) is dynamically weighted and integrated with multi-dimensional business factors to generate a comprehensive recruitment outreach necessity score. This score directly reflects, within the current business context, whether it is necessary and worthwhile to contact the job seeker today.
[0063] The rules for fusion computing are defined as follows:
[0064] ;
[0065] = ;
[0066] in, Indicates the date For job seekers Score the necessity of recruitment outreach; >0 indicates a strong recommendation to reach out; ≈0 indicates that the target may or may not be reached; <0 indicates that access should be avoided. Transform abstract causal effects into decision scores that can be directly used for business prioritization and resource trade-offs.
[0067] This indicates the value estimated using a causal inference model, at the date... For job seekers The difference in probability that the recruitment process will advance within a specified time window if an outreach action is taken compared to if no outreach action is taken; Indicates based on job seeker Current recruitment stage Pre-defined business value weights, such as the V(s) in the stage of sending recruitment intentions, are much higher than those in the resume screening stage; This indicates a causal effect through stage-weighted value. It addresses the issue of different business importance at different stages, but the Uplift did not weight the results, allowing even small gains at critical stages to achieve high scores.
[0068] This indicates the reach cost set based on the reach method used. It depends on the method of delivery; for example, telephone calls are more expensive than instant messaging. It introduces a resource consumption mindset to avoid using high-cost methods for low-value objectives.
[0069] This indicates that the job seeker's level of fatigue is being punished. Indicates the decision deadline t, based on the job seeker's... The job seeker fatigue level is calculated based on the frequency of being reached within a time window; the higher the frequency of being reached, the higher the fatigue level. It is the fatigue penalty weight. The simulation addressed the potential negative effects of excessive contact, thus protecting the job seeker experience.
[0070] This indicates an increase in rewards based on the urgency of the stage. Indicates that on decision date t, based on job seeker The urgency factor is calculated based on the current stay duration and the remaining validity period of the current phase. The longer the stay or the shorter the validity period, the higher the urgency. It is the weight of the urgency reward. This ensures that job seekers at risk of being lost or stagnant for an extended period receive priority attention.
[0071] This indicates the deduction of future opportunity value. This represents an estimate of the expected benefits for a specific job seeker c on decision date t, if they are not proactively reached, and if they are to progress to the next stage through natural processes in the future. This indicates the configuration weight.
[0072] This indicates that without proactive outreach, given job seeker c is in state s on date t, in the future... The predicted probability of naturally progressing to the next stage within a day;
[0073] Indicates the next stage of recruitment The stage value coefficient;
[0074] This represents the discount function that decays over time, i.e., for a given stage... For job seekers, a natural advancement benefit expected to materialize in τ days is discounted to its present value factor on the current decision date t.
[0075] This step introduces the concept of opportunity cost, which means that if a job seeker has a high probability of naturally making progress, then the added value of proactive outreach should be reduced, thereby avoiding wasting valuable outreach resources on job seekers who would have succeeded anyway, and allocating them to job seekers who need more intervention.
[0076] It should be noted that the weighting parameters λ, γ, and η in the above formula can be dynamically adjusted according to the stage. For example, in the resume screening stage, more emphasis is placed on... During the stage of sending out offers of employment, more emphasis is placed on... During the onboarding confirmation phase, the threshold is significantly increased to avoid disturbing the already selected candidates.
[0077] Step S104: Generation and exploration of reach plans under constraints.
[0078] This step inputs the recruitment outreach necessity score, preset recruitment business constraints, and decision state information into the resource optimization scheduling model. During the decision-making process of this resource optimization scheduling model, an exploratory intervention strategy is used to perturb its decision to generate a recruitment outreach action sequence (a sequence containing outreach-related content such as recruiter ID, job seeker ID, outreach method, and planned time). Some outreach actions in the recruitment outreach action sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation of the exploratory intervention strategy.
[0079] The specific process is as follows:
[0080] Step 1: Constrained Optimization Scheduling Modeling
[0081] The problem is modeled as a Constrained Markov Decision Process (CMDP), whose core elements include:
[0082] 1) Status s_t: includes remaining reach quota (remaining call duration, remaining number of SMS messages, etc.), statistical characteristics of the pool of job seekers awaiting decision (distribution of the number of people in the recruitment stage, average score of job seekers, number of people who stayed longer in the recruitment stage, etc.), time information, and recent reach frequency of each job seeker.
[0083] 2) Action a_t: Select one candidate c from the job candidates and specify the contact method l. The contact method l refers to the specific contact channel or technical means that the system can schedule and execute, such as: telephone, SMS, email, etc.
[0084] 3) Reward r_t: the immediate recruitment outreach necessity score obtained by performing the recruitment outreach action.
[0085] 4) Constraint C_k: Define multiple constraint cost functions, for example:
[0086] Resource constraint: Total daily reach actions ≤ quota Q.
[0087] Frequency constraint: For the same job seeker c, the number of times contacted within N days must be less than or equal to the maximum threshold.
[0088] Attribution constraint: Job seekers can only be reached by recruiters they designate.
[0089] Coverage constraint: The percentage of job seekers who are at a critical stage (such as those awaiting an interview) must be reached more than the minimum threshold each day.
[0090] This step incorporates constraints into the optimization objective and uses a solution strategy to maximize the long-term cumulative reward, i.e., the recruitment reach necessity score, while satisfying all constraints.
[0091] The specific solution method involves constructing the Lagrange function:
[0092]
[0093] in It is the Lagrange multiplier of the k-th constraint, representing the penalty for violating the constraint; Indicates the scheduling strategy; Indicates policy-based Expectations; This represents the immediate reward received at time t, i.e., the immediate business value generated by performing a specific outreach action on the job seeker at time t; Indicates the first A constrained cost function in state and actions The value to be taken below.
[0094] Step Two: Active Exploration Mechanism
[0095] To prevent the decision outcome π (scheduling strategy) from converging to a local optimum in the long term, leading to narrowed data distribution and model degradation, this invention forcibly introduces an exploratory intervention strategy to perturb its decision-making. Specifically, when formulating the daily outreach plan, the system does not use all resources to execute the currently known, highest-yielding strategy. Instead, it reserves a portion of resources specifically for executing planned, exploratory actions that deviate from the current optimal strategy (referred to as exploration in this embodiment) to obtain new and diverse feedback data, thereby driving the long-term evolution of the system.
[0096] Specific exploratory interference strategies include:
[0097] (1) Diversity exploration
[0098] By clustering job seekers' state vectors, priority is given to reaching job seekers in clusters with sparse historical reach samples. This supplements the sparse region samples for the causal inference model, thereby improving its generalization ability.
[0099] (2) Strategy perturbation exploration
[0100] The system perturbs the action probability distribution output by the optimal strategy π. Specifically, during the generation of daily reach action sequences, for each reach opportunity to be assigned, the system follows the recommendation of the current optimal scheduling strategy π with a probability of 1-ε, while simultaneously selecting an action completely randomly from eligible candidate actions (such as from a pool of job seekers allowed to be reached) with a probability of ε. This forces the system to try non-optimal actions, obtaining new trajectory samples needed for strategy improvement.
[0101] (3) Constraint perception exploration
[0102] The intensity of exploration interference can be dynamically adjusted. For example, exploration can be reduced when the daily total reach quota is tight to ensure optimal performance, and exploration can be increased when the quota is ample to gather more information. This balances exploration and utilization, making exploration behavior adaptive to business risks.
[0103] In summary, the recruitment outreach action sequence ultimately output by the resource optimization scheduling model is actually a hybrid product of global optimization and proactive exploration. Most outreach actions are generated by optimization strategies, while a small portion of outreach actions within the exploration quota are exploratory and may deviate from the immediate optimal recommendation, but they provide crucial data for the long-term evolution of the system.
[0104] For example:
[0105] Step A: Generate a strategy-recommended sequence of recruitment outreach actions.
[0106] Based on the optimal strategy of the current causal inference model, recruitment outreach necessity scoring and resource optimization scheduling model, the system selects the 80 job seekers with the highest expected value from the job seeker pool and generates a preliminary outreach list. The goal of the decision is to maximize the known and certain benefits of the day.
[0107] Step B: Generate an exploratory recruitment outreach action sequence
[0108] The system initiates an exploratory intervention strategy, generating the following diverse possible attempts within the exploratory quota:
[0109] Scenario 1: Diversity Exploration
[0110] The system performed cluster analysis on job seekers and discovered a cluster of job seekers with skills in niche programming languages that had historically been rarely reached.
[0111] Scenario 2: Exploring Strategy Perturbations
[0112] The system randomly selects 5 people from the above clusters to join the final plan.
[0113] The system perturbs the reach probabilities of the 81st to 100th ranked job seekers recommended by the optimal strategy π. For example, instead of selecting job seeker X, the system may select job seeker Y, who is ranked slightly lower but has different characteristics, with a certain probability. This generates 15 exploratory reach actions.
[0114] Step C: Merging and Execution
[0115] The 80 job seekers with the highest expected value are combined with 20 exploratory outreach actions to form the final recruitment outreach sequence, which is then sent to the recruitment terminal for execution. These 20 exploratory outreach actions deviate from the model's optimal recommendation due to perturbations in the exploration strategy.
[0116] In this example, the 20 exploratory outreach attempts are hard-reserved. Even if the optimal strategy π considers there to be 100 highly rated job seekers, the system must still allocate 20 opportunities for exploration. This ensures the execution of exploratory actions and prevents a greedy strategy from completely crowding out the exploration space. The feedback data generated from these 20 exploratory outreach attempts (whether successful or not) will flow back to the database along with the data from another 80 utilization attempts. This effectively prevents the model from getting stuck in local optima due to long-term training on similar data, achieving the goal of preventing strategy solidification and enabling long-term self-evolution of the model.
[0117] Step S105: Reach and execute, and receive feedback.
[0118] This step aims to achieve a closed loop of "recruitment outreach decision-making - terminal execution - data feedback". See also Figure 2 The contact action sequence generated in step S104, containing recruiter ID, job seeker ID, contact method, and planned time, is sent to the corresponding recruitment terminals, such as PC workbench and mobile APP. Recruiters or AI robots then execute contact operations such as phone calls and IM messages. The execution process is monitored throughout the entire chain through enterprise communication gateways, CRM systems, or event tracking points. Feedback data is stored in the application server, where contact feedback data is collected in real time or within a work cycle. This contact feedback data includes, but is not limited to: whether the action was actually executed, execution timestamp, contact method, whether the call was answered / the message was read, the job seeker's expressed intention during the communication, whether the job seeker's progress has been made after this contact, and whether any complaints or negative feedback have been received from the job seeker. After structuring the above contact feedback data, it is written to the database server, thus obtaining a complete data chain of "recruitment contact decision-terminal execution-data feedback," providing a real source of supervisory signals for subsequent model updates.
[0119] Step S106: Model and policy update.
[0120] This step aims to leverage the newly fed-back data to drive continuous self-evolution of all models and strategies in the system, addressing the fundamental problem of long-term performance degradation. Specifically, based on the feedback data, the causal inference model is incrementally trained, the weight parameters in the fusion calculation rules are optimized, and the strategy of the resource optimization scheduling model is updated using reinforcement learning.
[0121] Model and policy updates include:
[0122] (1) Incremental training of causal inference model
[0123] The new state vector, reach action, and stage progress result triplet are added to the training set, and the causal inference model in step S102 is incrementally learned or fine-tuned periodically or triggered, so that the causal effect estimation can track changes in job seeker behavior patterns and market environment and maintain timeliness.
[0124] (2) Parameter optimization of fusion calculation rules
[0125] Based on historical data, the weight parameters λ, γ, and η, as well as the stage value V(s) and cost C(l) in the fusion calculation rules, are optimized. For example, fatigue is analyzed. The correlation with subsequent negative feedback (such as complaints) is used to adjust λ. This allows the recruitment reach necessity score to more accurately reflect current recruitment needs and cost considerations.
[0126] (3) Reinforcement learning update of scheduling strategy
[0127] The daily outreach decision trajectory (resource status, job seeker pool characteristics, outreach instructions, recruitment outreach necessity score, degree of constraint violation, etc.) is stored as experience in the replay buffer and used to continuously update the parameters of the resource optimization scheduling model in step S104. In this way, the scheduling strategy continuously adapts to new data distributions and learns to make better sequence decisions under constraints.
[0128] (4) Exploration effect monitoring and adjustment
[0129] The system continuously monitors the proportion of exploratory feedback data, the information gain of exploratory feedback data for model updates, and whether exploration leads to frequent violations of key constraints or a significant decrease in core benefits. Based on these monitoring indicators, the system dynamically adjusts the exploration quota ratio, perturbation coefficient, etc., thereby achieving adaptive management of the intensity of exploration interference.
[0130] This step enables the system to complete a closed loop from decision-making to learning. Each execution feedback continuously evolves the model's decision-making ability, effectively combating the problems of policy solidification and model degradation, and realizing the model's long-term self-evolution capability.
[0131] The intelligent recruitment outreach decision-making method provided in this embodiment achieves automated and precise screening of outreach targets and optimal resource allocation in real recruitment scenarios with multiple stages and multiple constraints by constructing a complete closed loop of state assessment, causal inference, constraint scheduling and proactive exploration. This effectively improves the overall efficiency and conversion rate of the recruitment process and has long-term robustness against continuous self-optimization and strategy degradation.
[0132] The following simulated recruitment scenario serves as a standalone example to demonstrate the specific application process of the method of the present invention.
[0133] Scenario: The recruitment department of a large internet company uses this system to manage the external recruitment process for technical positions. The process stages are defined as: S1 - Resume entry, S2 - Initial interview, S3 - Technical interview arrangement, S4 - Issuance of offer letters, and S5 - Confirmation of employment. Currently, there are thousands of job seekers at different stages.
[0134] (1) System initialization and daily startup
[0135] Every morning at midnight, the decision server synchronizes the latest data from the company's HR database and recruitment CRM system. A state vector S is constructed for each active job seeker for the day, including the job seeker, current stage, date, and features X. For example, for job seeker Zhang San (currently in stage S3 for 5 days), his features X include: 5 days in stage S3, 1 phone call from HR in the last 3 days, a skill-to-target-job match of 0.9, and a record of logging into the recruitment platform last week.
[0136] (2) Causal assessment and necessity scoring
[0137] The system invokes a trained causal inference model, estimating a probability difference of 0.15 for contacting Zhang San by phone today. This means that contacting Zhang San increases the probability of him entering stage S4 within 3 days by 15%. Based on this probability difference of 0.15, a recruitment outreach necessity score is calculated:
[0138] Uplift × V(S3): 0.15 × 80 (S3 stage value coefficient is preset to 80) = 12.
[0139] C: The preset telephone cost is 2.
[0140] λ×F: Zhang San's recent contact frequency is low, F=0.1, λ=5, the penalty is -0.5.
[0141] γ×U: Stay in stage S3 for 5 days, with urgency level U=0.8 and γ=10, and the reward is +8.
[0142] η×FOV: Estimated probability of Zhang San's natural progression The value is relatively low (0.05), with an FOV of approximately 1.2 after calculation. η = 1, minus -1.2.
[0143] Zhang San's recruitment outreach necessity score is 12-2-0.5+8-1.2=16.3.
[0144] At the same time, the NeedScore is calculated for all other job seekers to determine the necessity of recruitment outreach.
[0145] (3) Constrained scheduling and exploratory plan generation
[0146] On that day, the recruitment team had a total phone quota of 200. Constraints included: a daily limit of 30 calls per person, a maximum of 2 contacts per job seeker within 3 days, and a daily coverage rate of ≥30% for job seekers in Phase S2.
[0147] The system inputs 200 quotas, the recruitment outreach necessity scores of all job seekers, and the aforementioned constraints into a resource optimization scheduling model. Under Lagrange multiplier regulation, an initial outreach list of 200 people is calculated, aiming to maximize the total recruitment outreach necessity score while satisfying the constraints. Simultaneously, the system designates 20% (40) of the quota as an exploration quota.
[0148] Of these 40 quotas, 20 are used for diversity exploration, meaning that after the system clusters job seekers, it identifies a group of job seekers with unique skills but few historical connections, and randomly selects 20 from this group to add to the list. 15 are used for strategy perturbation exploration, meaning that from the other 190 people recommended by the optimal strategy (190 reserved for exploration), 15 of them are randomly replaced with other job seekers whose recruitment reach necessity scores are slightly lower, with a probability of ε=0.08. The remaining 5 quotas serve as a buffer.
[0149] Finally, a daily plan containing 200 outreach tasks is generated and assigned to specific recruitment terminals for execution.
[0150] (4) Execution and feedback
[0151] The recruitment terminal receives tasks through the workbench and begins executing phone calls. After the call ends, the results are recorded in the system, for example: Zhang San, the call was connected, the communication went smoothly, the interview time was confirmed, and the stage progresses to S4. These results are fed back to the database in real time or within a given period.
[0152] (5) Model update
[0153] At night, the system adds all access records from the day, including exploration missions, to the causal inference model training set and initiates an incremental training.
[0154] Analyzing today's data reveals a strong correlation between frequent phone calls and receiving "reconsider" feedback for job seekers in the S4 stage. The system automatically adjusts the fatigue penalty weight λ for the S4 stage slightly upwards.
[0155] The scheduling trajectory of today (state, action, reward, constraint consumption, etc.) is used to update the parameters of the resource optimization scheduling model. If the coverage constraint of the S2 stage is occasionally violated, its corresponding multiplier is slightly increased to increase the penalty.
[0156] Monitoring shows that the exploration sample accounts for 22%, which has a significant information gain and does not seriously affect the core indicators. Therefore, the current exploration quota ratio will be maintained.
[0157] Through this example, the company's recruitment department achieved the following: It can directly provide an optimal task list daily without manually sifting through massive amounts of resumes and determining contact priorities; the system prioritizes contacting job seekers who "might stagnate if not contacted, but are likely to advance the process if contacted," avoiding harassment of those with high organic conversion intentions, thus improving the overall stage progress rate; it strictly ensures a balanced workload and job seeker experience; and by exploring interference mechanisms, the system accidentally contacted some job seekers with niche skills, discovering that some of them had high conversion rates. Subsequently, the system automatically increased the necessity score and scheduling weight for recruiting these job seekers, forming a new and effective recruitment strategy.
[0158] This example demonstrates the automation, intelligence, and self-evolution capabilities of the method of the present invention in real-world recruitment processes.
[0159] See Figure 3 Another embodiment of the present invention provides an intelligent recruitment outreach decision-making device 200, including a vector construction module 201, a causal inference module 202, a recruitment outreach scoring module 203, a recruitment outreach action sequence generation module 204, an execution and feedback module 205, and a model optimization module 206. This intelligent recruitment outreach decision-making device 200 is capable of executing the intelligent recruitment outreach decision-making method in the method embodiment.
[0160] Specifically, the intelligent recruitment outreach decision-making device 200 includes:
[0161] The vector construction module 201 is used to obtain job seeker attribute data, historical recruitment behavior logs and recruitment stage information from the database server, and to construct a daily updated state vector for each job seeker.
[0162] The causal inference module 202 is used to estimate, based on the state vector, the probability difference between performing recruitment outreach actions on job seekers and not performing recruitment outreach actions within a specified time in the future through a causal inference model.
[0163] The recruitment outreach scoring module 203 is used to calculate the recruitment outreach necessity score of job seekers based on the probability difference and according to the fusion calculation rules.
[0164] The recruitment outreach action sequence generation module 204 is used to input the recruitment outreach necessity score, preset recruitment business constraints, and decision state information into the resource optimization scheduling model, and to apply perturbation to the decision of the resource optimization scheduling model through an exploratory intervention strategy during the decision-making process of the resource optimization scheduling model, so as to generate a recruitment outreach action sequence; wherein, some outreach actions in the recruitment outreach action sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation of the exploratory intervention strategy;
[0165] The execution and feedback module 205 is used to send the recruitment outreach action sequence to the terminal device for execution, and to collect outreach feedback data from the application server associated with the terminal device and store it in the database server;
[0166] The model optimization module 206 is used to incrementally train the causal inference model based on the reach feedback data, optimize the weight parameters in the fusion calculation rules, and update the strategy of the resource optimization scheduling model through reinforcement learning.
[0167] It should be noted that the intelligent recruitment outreach decision-making device 200 provided in this embodiment, which can be used to execute the technical solutions of various method embodiments, has similar implementation principles and technical effects to the methods, and will not be repeated here.
[0168] See Figure 4 In another embodiment of the present invention, a schematic diagram of an electronic device 300 is provided. This electronic device is used to implement the intelligent recruitment outreach decision method in the method embodiment. The electronic device 300 in the embodiments of the present invention may include, but is not limited to, a PC, a server, a laptop computer, and a smart terminal. Figure 4 The electronic device 300 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.
[0169] like Figure 4 As shown, the electronic device 300 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes to implement the methods of the embodiments described herein, based on a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing device 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.
[0170] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.
[0171] The above description is merely a preferred embodiment of the present invention. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to the specific combination of the above-described technical features, but should also cover other technical solutions formed by any combination of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.
Claims
1. An intelligent recruitment outreach decision-making method, used in a decision server, characterized in that, Includes the following steps: Retrieve job seeker attribute data, historical recruitment behavior logs, and recruitment stage information from the database server, and construct a daily updated state vector for each job seeker; Based on the state vector, the probability difference between performing recruitment outreach actions on job seekers and not performing recruitment outreach actions within a specified time in the future is estimated using a causal inference model. Based on the probability difference, the job seeker's recruitment outreach necessity score is calculated according to the fusion calculation rule; the fusion calculation rule is as follows: ； = ； in, Indicates on date For job seekers Score the necessity of recruitment outreach; This indicates the value estimated using a causal inference model, at the date... For job seekers The difference in probability that the recruitment process will advance within a specified time window if an outreach action is taken compared to if no outreach action is taken; Indicates based on job seeker Current recruitment stage Pre-defined business value weights; This indicates the reach cost set based on the reach method used. Indicates on date Based on job seekers Job seeker fatigue level is calculated based on the frequency of being reached within a time window; Indicates based on job seeker The stage urgency factor is calculated based on the current stage's duration of stay and the remaining duration of the stage's validity. Indicates on date job seeker Expected returns from natural progression in the future; This indicates that job seekers will not be reached without being proactively contacted. On the current date The Future The predicted probability of naturally progressing to the next stage within a day; Indicates the next stage of recruitment The stage value coefficient; This represents a discount function that decays over time. They correspond to , and Configurable weight parameters; The necessity score for recruitment outreach, preset recruitment business constraints, and decision state information are input into a resource optimization scheduling model. During the decision-making process of this resource optimization scheduling model, an exploratory intervention strategy is used to perturb its decisions to generate a recruitment outreach action sequence. Some outreach actions in this sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation by the exploratory intervention strategy. The exploratory intervention strategy allocates a preset proportion of exploration quotas from the daily available recruitment outreach resources. Cluster analysis is performed on the job seeker's state vector. From job seeker categories with historical outreach frequencies below a preset threshold, job seekers to be reached are randomly selected based on the exploration quota. And / or, a random perturbation is applied to the probability distribution of the optimal recruitment outreach actions recommended by the resource optimization scheduling model. Job seekers to be reached are reselected according to the perturbed probabilities, and the intensity of the random perturbation is dynamically adjusted based on the proportion of remaining outreach resources each day. The recruitment outreach action sequence is sent to the terminal device for execution, and outreach feedback data is collected from the application server associated with the terminal device and stored in the database server; Based on the reach feedback data, the causal inference model is incrementally trained, the weight parameters in the fusion calculation rules are optimized, and the strategy of the resource optimization scheduling model is updated through reinforcement learning.
2. The intelligent recruitment outreach decision-making method according to claim 1, characterized in that, Also includes: Continuously monitor the stability of the output of the causal inference model, the distribution deviation trend of the recruitment outreach necessity score, the cumulative benefit of the resource optimization scheduling model, the outreach coverage and frequency of each recruitment stage, and the proportion of exploratory outreach feedback data. When the monitoring indicators exceed the preset threshold, the allocation ratio of the exploration quota is automatically adjusted, the weight parameters in the fusion calculation rules corresponding to job seeker fatigue, stage urgency factors and expected benefits of future natural progress are automatically adjusted, the causal inference model is retrained, and the random perturbation intensity of the exploratory intervention strategy in the resource optimization scheduling model is adjusted.
3. The intelligent recruitment outreach decision-making method according to claim 2, characterized in that, Also includes: The distribution statistics of the probability difference, the overall offset trend of the recruitment reach necessity score, and the violation frequency of each recruitment business constraint are calculated periodically. If the calculated indicators exceed a preset threshold, at least one of the following operations is automatically triggered: Increase the proportion of the exploration quota, adjust the weight parameters in the fusion calculation rules, initiate incremental learning for the causal inference model, and retrain the resource optimization scheduling model.
4. An intelligent recruitment outreach decision-making device, used in a decision server, characterized in that, include: The vector building module is used to retrieve job seeker attribute data, historical recruitment behavior logs, and recruitment stage information from the database server, and build a daily updated state vector for each job seeker; The causal inference module is used to estimate, based on the state vector, the probability difference between performing recruitment outreach actions on job seekers and not performing recruitment outreach actions within a specified time in the future through a causal inference model. The recruitment outreach scoring module is used to calculate the recruitment outreach necessity score of job seekers based on the probability difference and according to the fusion calculation rules; the fusion calculation rules are as follows: ； = ； in, Indicates on date For job seekers Score the necessity of recruitment outreach; This indicates the value estimated using a causal inference model, at the date... For job seekers The difference in probability that the recruitment process will advance within a specified time window if an outreach action is taken compared to if no outreach action is taken; Indicates based on job seeker Current recruitment stage Pre-defined business value weights; This indicates the reach cost set based on the reach method used. Indicates on date Based on job seekers Job seeker fatigue level is calculated based on the frequency of being reached within a time window; Indicates based on job seeker The stage urgency factor is calculated based on the current stage's duration of stay and the remaining duration of the stage's validity. Indicates on date job seeker Expected returns from natural progression in the future; This indicates that job seekers will not be reached without being proactively contacted. On the current date The Future The predicted probability of naturally progressing to the next stage within a day; Indicates the next stage of recruitment The stage value coefficient; This represents a discount function that decays over time. They correspond to , and Configurable weight parameters; The recruitment outreach action sequence generation module is used to input the recruitment outreach necessity score, preset recruitment business constraints, and decision state information into a resource optimization scheduling model. During the decision-making process of the resource optimization scheduling model, an exploratory intervention strategy is used to perturb its decision to generate a recruitment outreach action sequence. Some outreach actions in the recruitment outreach action sequence deviate from the optimal recommendation of the resource optimization scheduling model due to the perturbation of the exploratory intervention strategy. The exploratory intervention strategy allocates a preset proportion of exploration quota from the daily available recruitment outreach resources; performs cluster analysis on the job seeker's state vector; randomly selects job seekers to be reached from job seeker categories with historical outreach frequencies below a preset threshold based on the exploration quota; and / or applies random perturbation to the probability distribution of the optimal recruitment outreach actions recommended by the resource optimization scheduling model, reselects job seekers to be reached according to the perturbed probabilities, and dynamically adjusts the intensity of the random perturbation based on the proportion of remaining outreach resources each day. The execution and feedback module is used to send the recruitment outreach action sequence to the terminal device for execution, and to collect outreach feedback data from the application server associated with the terminal device and store it in the database server; The model optimization module is used to incrementally train the causal inference model based on the reach feedback data, optimize the weight parameters in the fusion calculation rules, and update the strategy of the resource optimization scheduling model through reinforcement learning.
5. The intelligent recruitment outreach decision-making device according to claim 4, characterized in that, It also includes a monitoring and adjustment module, used for: The system continuously monitors the stability of the causal inference model output, the distribution trend of the recruitment outreach necessity score, the cumulative revenue of the resource optimization scheduling model, the outreach coverage and frequency at each recruitment stage, and the proportion of exploratory outreach feedback data. When the monitored indicators exceed the preset threshold, the system automatically adjusts the allocation ratio of the exploratory quota, automatically adjusts the weight parameters in the fusion calculation rules corresponding to job seeker fatigue, stage urgency factors, and expected revenue of future natural progress, retrains the causal inference model, and adjusts the random perturbation intensity of the exploratory intervention strategy in the resource optimization scheduling model.