Data-driven recruitment process efficiency optimization analysis method

Abstract

The present application belongs to the technical field of human resource data analysis, and relates to a data-driven recruitment process efficiency optimization analysis method. The present application processes historical recruitment data, extracts candidate features and discretizes them, calculates the category lift of single features and the joint lift of combined features, generates self-adaptive one-dimensional mask values and synergistic two-dimensional mask arrays. In this way, redundant feature signals are automatically intercepted and strong synergistic feature paths are retained. Finally, the feature encoding of the new resume is calculated by combining the network convergence weight and the mask array, and the matching probability is output. The present application solves the problems of low screening efficiency of massive resumes, easy introduction of feature noise and neglect of feature synergistic gain, effectively reduces the computing power consumption of the system, and significantly improves the accuracy and forward reasoning performance of the human-post matching evaluation.

Description

Technical Field

[0001] This invention relates to the field of human resource data analysis technology, and in particular to a data-driven method for optimizing the efficiency of the recruitment process. Background Technology

[0002] With the deepening of enterprise digital transformation and the intelligent development of human resource management, online recruitment platforms receive a massive number of job applications every day. Faced with such a vast and complex volume of recruitment data, how to quickly and accurately filter out candidates who are highly matched to the complex requirements of specific positions has become a core pain point that urgently needs to be addressed in the field of human resource management. Accurately assessing the degree of person-job fit is not only related to the overall efficiency of recruitment screening, but also crucial for enterprises to acquire high-quality talent.

[0003] Traditional resume screening and job matching methods primarily rely on manual review or conventional matching systems based on preset rules. These conventional techniques typically employ a one-size-fits-all approach to roughly filter candidates based on fixed logical conditions. With technological advancements, some systems have begun using conventional machine learning algorithms, indiscriminately inputting all extracted resume features into the model for probability prediction. However, these existing methods lack in-depth analysis of the actual contribution of various features in historical recruitment data; they neither establish targeted feature filtering mechanisms nor introduce logic for feature cross-selection.

[0004] In practical engineering applications, the aforementioned existing technologies have significant limitations. On the one hand, resume data is typically high-dimensional and sparse. Due to the failure to effectively evaluate and eliminate redundant and inefficient features, computing nodes face not only enormous computational and memory overhead when processing this full set of feature data, but are also highly susceptible to introducing feature noise. This interference from invalid signals leads to a significant decrease in the accuracy of the final job-person matching assessment. On the other hand, existing solutions often process various resume indicators in isolation during data modeling, ignoring the potential strong synergistic gain effects between different feature combinations. In real-world recruitment scenarios, the combined presence of features such as specific educational backgrounds and industry work experience often generates job competitiveness far exceeding the simple superposition of single features. Existing conventional models cannot accurately capture and quantify the core value of this cross-dimensional approach. Therefore, existing technologies cannot achieve accurate and dynamic assessment of the matching probability of high-potential candidates while effectively reducing the consumption of underlying computing resources. Summary of the Invention

[0005] The purpose of this invention is to propose a data-driven recruitment process efficiency optimization analysis method to solve the technical problems of traditional resume screening methods, such as huge computational overhead, easy introduction of feature noise, and neglect of collaborative gain between features when processing massive high-dimensional sparse data.

[0006] Therefore, the technical solution of the data-driven recruitment process efficiency optimization analysis method provided by this invention is as follows: Data-driven methods for optimizing recruitment process efficiency include: Obtain historical recruitment data and divide it into category subsets. Extract various features from each category subset and count the total number of applicants and the total number of hires for each category subset. For each feature, calculate the second-order difference of the ordered features, determine the position of the smallest second-order difference as the marginal split point, and divide the ordered features into low-value intervals and high-value intervals based on the marginal split point; Calculate the category lift of each feature and the joint lift of any two features based on the total number of applicants and the total number of hires. Calculate the lift difference between the high-value interval and the low-value interval and generate an adaptive threshold. Generate a one-dimensional mask value based on the lift difference and the adaptive threshold. A two-dimensional mask matrix is ​​generated based on the category lift and the joint lift; The original feature vector of the new resume is processed based on the edge splitting point to obtain the feature encoding value. The feature encoding value is then calculated by combining the one-dimensional mask value and the two-dimensional mask matrix to obtain the matching probability and output it.

[0007] This invention generates a corresponding mask structure by calculating feature lift, enabling targeted interception and release of features in new resumes. It can automatically block redundant noise and invalid cross paths, reducing the computational cost of the evaluation system while improving the accuracy of job matching assessment. It overcomes the shortcomings of manual screening, which is subjective and inefficient. At the same time, it clearly presents the discriminative power of individual features and the synergistic gain of combined features, realizing the dynamic tilt of computing resources towards high-value features and improving the reliability of the final forward inference output matching probability.

[0008] Furthermore, the calculation of the second-order difference for the ordered features among the features includes: The number of applicants and the number of hires at each level are statistically analyzed for the ordered characteristics. The ratio of the number of hires to the number of applicants is used as the smoothed hiring rate. The smoothed difference between the smoothed hiring rates of two adjacent levels is calculated, and the change in the smoothed difference between two adjacent levels is determined as the second-order difference.

[0009] This invention can effectively prevent gradient explosion or misjudgment caused by a few rare levels, and enhance the generalization ability of the analysis model when facing long-tail data. By calculating the second difference, it can automatically and objectively locate the marginal decreasing inflection point of the influence of various features on the recruitment results in massive and unevenly distributed recruitment data, avoiding the bias caused by manually setting feature thresholds and improving the scientific nature of feature interval division.

[0010] Furthermore, before calculating the second-order difference for the ordered features among the features, the analysis method further includes: detecting the number of valid levels of the ordered features in the category subset; in response to the number of valid levels being less than or equal to 2, skipping the step of calculating the second-order difference, aggregating all valid levels into a single interval, and directly calculating the category lift corresponding to the single interval.

[0011] This step avoids the loss of statistical significance in subsequent difference calculations and the waste of computing resources caused by overly simplistic feature values. It also prevents the introduction of feature noise by forcibly searching for edge split points, thereby improving data processing efficiency and ensuring the stability of matching probability evaluation.

[0012] Furthermore, the calculation of the class lift of an individual feature and the joint lift of any combination of two features includes: Obtain the number of single-feature recruits and the number of single-feature applications that possess the aforementioned single feature. Divide the product of the number of single-feature recruits and the total number of applications by the product of the number of single-feature applications and the total number of recruits to obtain the category lift of the aforementioned single feature. Obtain the number of dual-feature applications that have any two of the stated features. In response to the number of dual-feature applications being greater than 0, obtain the number of dual-feature hires that have any two of the stated features. Divide the product of the number of dual-feature hires and the total number of applicants by the product of the number of dual-feature applications and the total number of hires to obtain the joint lift.

[0013] This invention measures the actual impact of various features on the final recruitment results and explores the potential synergistic gain effect between different feature combinations. At the same time, by responding to the pre-judgment logic that the number of applications for two features is greater than 0, it filters out unpopular feature combinations that have never appeared simultaneously in historical recruitment data.

[0014] Furthermore, the generation of a one-dimensional mask value based on the lift difference and the adaptive threshold includes: Calculate the absolute value of the difference between the class lift of the single feature corresponding to the high value interval and the class lift of the single feature corresponding to the low value interval to obtain the lift difference value; calculate the average value of the lift differences of all ordered features and use it as the adaptive threshold; in response to the lift difference value being greater than or equal to the adaptive threshold, assign the corresponding one-dimensional mask value to 1; in response to the lift difference value being less than the adaptive threshold, assign the corresponding one-dimensional mask value to 0.

[0015] This invention can automatically and accurately cut off redundant signals with no significant distinguishability at the source of feature input, and greatly reduce the dimensionality and memory usage of subsequent data processing while ensuring the lossless transmission of key information.

[0016] Further, generating a two-dimensional mask matrix based on the class lift and the joint lift includes: calculating the product of the class lift corresponding to each feature in any two feature combinations, and adding the product to a preset tolerance value to obtain a collaborative baseline; In response to the joint lift of any two feature combinations being greater than the cooperative baseline, the element value of the corresponding feature combination in the two-dimensional mask array is set to 1; in response to the joint lift of any two feature combinations being less than or equal to the cooperative baseline, the element value of the corresponding feature combination in the two-dimensional mask array is set to 0.

[0017] Furthermore, before obtaining and outputting the matching probability, the analysis method further includes: Calculate the difference between the category lift of each feature and 1, and use it as the individual contribution. Calculate the sum of the absolute values ​​of the individual contributions of all features to obtain the global contribution sum. In response to the global contribution sum being greater than 0, divide the individual contribution by the global contribution sum to obtain the one-dimensional initial weight of the corresponding feature. In response to the global contribution sum being equal to 0, assign equal one-dimensional initial weights to all features using a mean distribution strategy.

[0018] This invention transforms objective historical recruitment business statistical patterns into a priori starting point for subsequent discrimination network parameters, effectively avoiding cold starts in the early stages of model processing and blind searches in a vast parameter space, thus accelerating the convergence speed of the entire evaluation link in complex data environments.

[0019] Furthermore, the analysis method also includes: extracting target feature pairs with an element value of 1 in the two-dimensional mask array, multiplying the one-dimensional initial weights corresponding to the two features in the target feature pair to obtain the cross initial weights; using the one-dimensional initial weights and cross initial weights as the starting point for the discrimination network parameters, training the category subset based on the recruitment labels in historical recruitment data to obtain the one-dimensional convergence weights and cross convergence weights.

[0020] Furthermore, the method for obtaining the matching probability includes: Multiply the one-dimensional convergence weight, the corresponding feature encoding value, and the one-dimensional mask value, and sum the product results of all individual features to obtain an independent feature term; multiply the cross convergence weight, the two feature encoding values ​​of the target feature pair, and the corresponding in-matrix element values, and sum the product results of all target feature pairs to obtain a collaborative feature term; sum the independent feature term, collaborative feature term, and network bias term, and then map them to obtain the matching probability.

[0021] Furthermore, the matching probability satisfies the formula: ;in, This represents the matching probability. Represents a nonlinear mapping function. Corresponding to the independent feature term, Indicates the first One-dimensional convergence weight of the feature term, , as well as They represent the first Item, No. Item and the The feature encoding value corresponding to the item, Indicates the first One-dimensional mask value of the feature, Corresponding to the aforementioned collaborative feature term, Indicates by the first Item and the The cross-convergence weight of the cross-combination features composed of feature items. This represents the value of the corresponding element in the two-dimensional mask matrix. This refers to the network bias term.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: This invention addresses the problems of low efficiency in screening massive amounts of resumes, easy introduction of feature noise, and neglect of feature synergy gains. It effectively reduces the computational overhead of recruitment assessment systems and improves the accuracy of job matching assessment and forward inference performance. At the same time, it overcomes the shortcomings of manual screening, which is highly subjective and inefficient. It can automatically block redundant noise and invalid cross paths, realize the dynamic tilt of computing resources towards high-value features, and improve the reliability of forward inference output matching probability. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the data-driven recruitment process efficiency optimization analysis method provided in this embodiment. Figure 2 This is a comparison curve of robust degradation under high-dimensional sparse feature noise provided in this embodiment; Figure 3 This is a comparison curve of the convergence speed of the training process provided in this embodiment. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] The data-driven recruitment process efficiency optimization analysis method provided by this invention is mainly applied to real-world scenarios such as intelligent recruitment management platforms for enterprise human resources or online recruitment, to solve the problems of low efficiency in screening massive amounts of resumes and insufficient accuracy in job matching assessment. Subsequent steps are all based on this recruitment scenario.

[0026] like Figure 1 As shown, the data-driven recruitment process efficiency optimization analysis method provided in this embodiment includes the following steps: S1. Obtain historical recruitment data and divide it into category subsets. Extract the features of each category subset and count the total number of applicants and the total number of hires for each category subset.

[0027] Specifically, this step involves acquiring all historical recruitment data from the recruitment management platform over a preset period. To eliminate data bias caused by seasonal recruitment fluctuations and ensure data timeliness, the preferred preset period is the past 24 months. After acquiring the historical recruitment data, it is divided into multiple independent category subsets based on the business area to which the positions belong. For example, the historical recruitment data can be divided into R&D position subsets, sales position subsets, and functional position subsets.

[0028] Furthermore, for any specific category subset, various features of the candidates are extracted from the corresponding resume text. In this embodiment, the features include discrete features without size comparison relationships and ordered features with progressive or size relationships. Preferably, discrete features include the candidate's gender and type of graduating institution, while ordered features include educational level and years of work experience. After feature extraction, based on the database aggregation query operation, the total number of people who have submitted resumes within the category subset is counted, and this total number is taken as the total number of applicants; the total number of people whose final recruitment status is "offer notice issued" is counted, and this total number is taken as the total number of hired personnel. Both the total number of applicants and the total number of hired personnel are represented as one-dimensional scalar values.

[0029] In this way, by transforming unstructured historical recruitment records into category subsets under specific positions and extracting various features as well as the overall application and hiring scale, we can define the analytical boundaries of the data from a macro perspective, laying a reliable and high-quality data foundation for subsequent accurate assessment of the actual impact of various features on the hiring results.

[0030] S2. Calculate the second-order difference for the ordered features among the features, determine the position of the smallest second-order difference as the marginal split point, and divide the ordered features into low-value intervals and high-value intervals based on the marginal split point.

[0031] Before performing fine-grained interval partitioning on the ordinal features among the various features, it is necessary to verify the richness of the features to avoid the subsequent difference calculation losing statistical significance due to overly simplistic feature values ​​and thus wasting computational resources. Before calculating the second-order difference for the ordinal features among the various features, the number of valid ranks of the ordinal features in the aforementioned category subsets is detected; if the number of valid ranks is less than or equal to 2, the second-order difference calculation is skipped, all valid ranks are aggregated into a single interval, and the category lift corresponding to this single interval is directly calculated.

[0032] The effective level number refers to the number of non-repeating and non-empty value levels that an ordered feature actually appears in a category subset. For example, if a job position has extremely high educational requirements, and the educational background of past applicants only has two levels: Master's and Doctoral, then the effective level number is 2. Forcibly finding a split point can easily introduce noise. Therefore, it is better to treat it as a whole for subsequent calculations, which improves processing efficiency and ensures the robustness of subsequent evaluations.

[0033] For ordered features that pass the pre-verification (i.e., have more than 2 valid grade levels), in order to accurately capture the marginal change trend of the admission probability between different grades and thus locate the threshold point where the admission advantage undergoes a fundamental change, this step calculates the second-order difference for ordered features. The specific method is as follows: Count the number of applicants and the number of admitted students at each grade level for ordered features; use the ratio of admitted students to applicants as the smoothed admission rate; calculate the smoothed difference between the smoothed admission rates corresponding to two adjacent grades, and determine the change in the smoothed difference between two adjacent grades as the second-order difference.

[0034] To address the problem of probability extrema caused by a small number of outlier samples, this step introduces the Laplace smoothing concept to construct a calculation model for the smoothed acceptance rate. Specifically, the smoothed acceptance rate satisfies the following relationship:

[0035] In the formula, Indicates the ordered feature in the first... The smoothed acceptance rate corresponding to each level Indicates being in the first The number of people recruited at each level. Indicates being in the first The number of applicants for each grade level is calculated, with the addend 1 in the numerator and the addend 2 in the denominator serving as smoothing coefficients. The value of these smoothing coefficients is determined based on the standard binomial distribution Laplace smoothing calibration rule, aiming to smooth the acceptance rate towards the prior median value of 0.5 when the number of applicants for each grade level is extremely small.

[0036] Understandably, as the number of applicants for each grade increases, the impact of constants 1 and 2 on the overall result rapidly diminishes, and the smoothed acceptance rate will approach the actual acceptance ratio. This approach effectively prevents gradient explosion or misjudgment caused by a few rare grades, and enhances the generalization ability of the analysis model when dealing with long-tail data.

[0037] Furthermore, to quantify the growth rate and acceleration of acceptance rates between adjacent levels, this step constructs a difference formula. Specifically, the smoothed difference and the second-order difference satisfy the following relationship:

[0038]

[0039] In the formula, Indicates the ordered feature in the first... The smoothed acceptance rate corresponding to each level Indicates the first The level and the first Smoothed difference between each level Indicates the first The level and the first Smoothed difference between each level Indicates the first The second-order difference corresponding to each level.

[0040] Based on the above calculations, it can be understood that the smoothing difference reflects the first derivative, that is, the direct increase in the hiring rate brought about by each improvement in feature level; while the second difference reflects the second derivative, that is, the marginal change of this increase. By iterating through the second differences corresponding to all levels, the second difference with the smallest value is extracted. This means that from this level onwards, the increase in the hiring rate brought about by continuing to improve this feature level exhibits a significant diminishing marginal effect, i.e., the inflection point of cost-effectiveness has been reached. The level position corresponding to this point is determined as the marginal cut-off point. Using the marginal cut-off point as the boundary, ordered features are divided into low-value and high-value intervals. For example, if the levels of the years of service feature are successively 1 to 5 years, and the calculation shows that the second difference corresponding to the 3rd year is the smallest, then the 3rd year is determined as the marginal cut-off point. At this time, years 1 to 2 are classified as low-value intervals, and years 3 to 5 are classified as high-value intervals.

[0041] Thus, by processing ordered features through Laplace smoothing combined with the second-order difference algorithm, the marginal utility inflection point of each feature's impact on the recruitment results can be automatically and objectively located in massive and unevenly distributed recruitment data. This avoids the bias caused by manually setting feature thresholds and improves the scientific nature of feature interval division.

[0042] S3. Calculate the category improvement of each feature and the joint improvement of any two features based on the total number of applicants and the total number of hires.

[0043] After completing the initial processing of the feature intervals, in order to accurately quantify the actual impact of each feature on the final recruitment and hiring results, and to explore the potential synergistic gain effect between different feature combinations, thus providing reliable data for subsequent accurate evaluation, this step calculates the category lift of individual features and the joint lift of any two feature combinations. The specific method is as follows: Obtain the number of single-feature acceptances and single-feature applications. Divide the product of the single-feature acceptances and the total number of applications by the product of the single-feature applications and the total number of acceptances to obtain the category lift of the single feature. Obtain the number of dual-feature applications with any two features. If the number of dual-feature applications is greater than 0, obtain the number of dual-feature acceptances with any two features. Divide the product of the dual-feature acceptances and the total number of applications by the product of the dual-feature applications and the total number of acceptances to obtain the joint lift.

[0044] To measure the change in the acceptance probability relative to the global average acceptance probability when a single feature is present, this step constructs a calculation model for category lift. Specifically, category lift satisfies the following relationship:

[0045] In the formula, This represents the class lift corresponding to a single feature. This indicates the number of people recruited based on a single characteristic. Represents the number of deliveries with a single feature. This represents the total number of applications received, calculated from historical recruitment data. This represents the total number of people hired, as calculated from historical recruitment data.

[0046] Understandably, this formula, in an engineering physics sense, is equivalent to the ratio of the local acceptance rate to the global acceptance rate under this single feature. When the category lift is greater than 1, it indicates that the single feature has a positive promoting effect on the recruitment results; the larger the value, the higher the value of the feature in recruitment screening. Conversely, if it is less than 1, it indicates that the feature is at a relative disadvantage.

[0047] Furthermore, in actual recruitment scenarios, a candidate's multiple abilities often exhibit strong correlations. For example, the combined competitiveness of graduating from a top-tier university and gaining internship experience at a leading tech company typically far exceeds the simple sum of these two individual characteristics. To accurately capture the synergistic value of these cross-features, this step constructs a calculation model for joint lift. Specifically, the joint lift satisfies the following relationship:

[0048] In the formula, and These represent the index numbers of any two different features. Indicates by the first Item features and the first The joint lift degree corresponding to any two feature combinations consisting of feature items. This represents the number of dual-feature applicants corresponding to the combination of the above two features. This represents the number of dual-feature deliverables corresponding to the combination of the above two features. This refers to the total number of people who received the above deliveries. This refers to the total number of people recruited as mentioned above.

[0049] Understandably, before executing the above formula calculation, by responding to the pre-judgment logic that the number of dual-feature submissions is greater than 0, unpopular feature combinations that have never appeared simultaneously in historical recruitment data are filtered out. This avoids the program crashing due to illegal operations with a denominator of 0 at the underlying computing node, thus ensuring the high availability of the entire feature mining chain.

[0050] In this way, by calculating the category lift of individual features and the joint lift of any two feature combinations, it is possible to extract the core key indicators that truly determine the success or failure of recruitment and their potential combined advantages from massive and complex recruitment data. This provides business personnel with an intuitive and reliable data foundation to quickly screen out candidates with high potential, effectively overcoming the errors and biases that are easily generated when traditional manual resume screening is based on subjective experience.

[0051] S4. Calculate the lift difference between the high-value interval and the low-value interval and generate an adaptive threshold. Generate a one-dimensional mask value based on the lift difference and the adaptive threshold.

[0052] To filter out redundant features with minimal impact on the hiring results from massive amounts of candidate information and extract core features with significant discriminative power, thereby reducing the computational overhead of subsequent nodes, this step calculates the lift difference between the class lift corresponding to the high-value interval and the low-value interval and generates an adaptive threshold. A one-dimensional mask value is then generated based on the lift difference and the adaptive threshold. Specifically, the method is as follows: calculate the absolute value of the difference between the class lift corresponding to the high-value interval and the low-value interval to obtain the lift difference; calculate the average of the lift differences of all ordered features and use it as the adaptive threshold; in response to a lift difference greater than or equal to the adaptive threshold, assign the corresponding one-dimensional mask value to 1, otherwise assign the one-dimensional mask value to 0.

[0053] To quantify the jump in acceptance probability before and after crossing the marginal cut-off point for each feature, and thus evaluate the effectiveness of the feature, this step constructs a calculation model for the lift difference and adaptive threshold. Specifically, the lift difference and adaptive threshold satisfy the following relationship:

[0054]

[0055] In the formula, Indicates the first The difference in lift corresponding to each ordered feature Indicates the first The class boosting degree corresponding to the high value range of each ordered feature Indicates the first The class boosting degree corresponding to each ordered feature in the low value range Indicates an adaptive threshold. The summation symbol represents the total number of ordered features, and the summation symbol indicates that all lift differences are summed.

[0056] Understandably, since the high-value and low-value intervals were determined in the preliminary step using second-order differences, the absolute value of the difference between the categories corresponding to these two intervals is now extracted, and this difference is the lift difference. The larger this absolute value, the more dramatic the increase in a candidate's job competitiveness when that feature crosses the boundary split point. By using the average of the lift differences of all ordered features as an adaptive threshold, a baseline that dynamically fluctuates with the overall data distribution can be provided, avoiding the problem of poor model generalization caused by manually hard-coding a fixed value as the threshold.

[0057] It should be noted that the class boosting degree corresponds to the high value range of ordered features. And the class boosting degree corresponding to the ordered features in the low value range The calculation logic and the overall category lift mentioned above Completely identical, only the formula needs to be changed. and Replace them with the number of local features adopted and the number of local features submitted within the high-value or low-value range for that feature.

[0058] Furthermore, to achieve hard screening of feature dimensions and mask feature filtering, a numerical comparison mechanism was established. The values ​​of the one-dimensional mask satisfy the following relationship:

[0059]

[0060] In the formula, Indicates the first One-dimensional mask value corresponding to each ordered feature Indicates the first The difference in lift corresponding to each ordered feature This indicates an adaptive threshold.

[0061] Understandably, when the lift difference is greater than or equal to the adaptive threshold, it means that the discriminative power of this feature exceeds the global average level of the current business scenario, and it belongs to a high-value feature. Therefore, the signal transmission of this feature is retained, and the one-dimensional mask value is assigned to 1; otherwise, it is assigned to 0. In the actual engineering physics level, this processing method is equivalent to directly cutting off the data propagation path of this redundant feature at the input of the neural network.

[0062] In this way, by dynamically calculating the differences in the probability of recruitment brought about by different intervals of each feature, and automatically generating a one-dimensional mask value containing only the values ​​0 and 1 using the global average level, in engineering practice, not only can inefficient feature noise be effectively shielded, and the computing power and memory usage of computing nodes when processing high-dimensional sparse features be greatly reduced, but also the accuracy of the final recruitment matching evaluation results is improved by intercepting redundant features in advance.

[0063] like Figure 2As shown, this embodiment provides a robustness degradation comparison curve under high-dimensional sparse feature noise to verify the above-mentioned technical effect. Real-world recruitment scenarios are often accompanied by massive, high-dimensional, and unevenly distributed sparse data. In this comparison curve, the horizontal axis represents the continuously increasing proportion of feature noise, and the vertical axis represents the accuracy of the job matching assessment. The feature noise proportion in the range of 0.6 to 1 corresponds to a typical high-dimensional sparse data environment. As redundant and invalid signals in the input features increase, existing technologies, due to their strategy of undifferentiated input of all features, easily introduce a large amount of feature noise, leading to a significant degradation in assessment accuracy, which drops sharply from approximately 0.9 to below 0.2. In contrast, this invention uses an adaptive threshold that fluctuates globally with the data to construct a one-dimensional masking mechanism, automatically and accurately cutting off redundant signals with no significant distinguishing power at the source of feature input. Even in the harsh high-noise range, this invention can maintain stable and high prediction accuracy, with the accuracy consistently above 0.8.

[0064] S5. Generate a two-dimensional mask matrix based on the category lift and the joint lift.

[0065] The purpose of this step is to accurately identify feature combinations that truly possess strong synergistic effects in actual recruitment operations, eliminating false associations caused solely by a high basic acceptance rate, thereby providing a reliable topological connection basis for subsequent cross-feature network construction. Specifically, the method for obtaining the two-dimensional mask matrix is ​​as follows: calculate the product of the class lift of each feature in any two feature combinations, add a preset tolerance value to this product to obtain the synergistic baseline. In response to the joint lift of any two feature combinations being greater than the synergistic baseline, the element value of the corresponding feature combination in the two-dimensional mask matrix is ​​assigned to 1; in response to the joint lift of any two feature combinations being less than or equal to the synergistic baseline, the element value of the corresponding feature combination in the two-dimensional mask matrix is ​​assigned to 0.

[0066] To quantify the expected theoretical joint enhancement of any two features under the assumption of mutual independence, this step constructs a mathematical model for the synergistic baseline. Specifically, the synergistic baseline satisfies the following relationship:

[0067] In the formula, Indicates by the first Item features and the first The collaborative baseline corresponding to any two feature combinations composed of feature items. and They represent the first Item features and the first The category lift of each individual feature. This indicates the preset tolerance value.

[0068] It should be noted that the preset tolerance value is determined using a grid search method based on historical data. To balance the recall and precision of strong cross-features in business operations, the specific steps for determining the tolerance value include: obtaining a validation subset independent of the feature statistical benchmark, and labeling the true cross-features based on feature combinations that have generated significant real recruitment gains in actual business operations; setting the candidate interval and search step size for the tolerance value, and generating a candidate parameter set containing multiple candidate tolerance values. For example, the candidate range is set to 0.01 to 0.2, and the search step size is 0.01. Each candidate tolerance value in the candidate parameter set is substituted into the calculation formula of the collaborative baseline to generate a corresponding test two-dimensional mask matrix. The test two-dimensional mask matrix is ​​used to perform cross-path filtering on the validation subset. The number of correctly retained strong collaborative feature pairs, the number of incorrectly retained weak collaborative feature pairs, and the number of incorrectly blocked strong collaborative feature pairs are counted respectively. Based on this, the feature evaluation accuracy and feature evaluation recall under the current candidate tolerance value are calculated. The harmonic mean of the feature evaluation accuracy and feature evaluation recall is calculated to obtain a comprehensive evaluation index for measuring the overall filtering performance. The candidate tolerance value that makes the comprehensive evaluation index reach the global maximum value is extracted and determined as the final tolerance value.

[0069] Understandably, based on the above data-driven optimization process, when the candidate tolerance value is between 0.05 and 0.15, the comprehensive evaluation index converges and remains consistently high and stable. Therefore, the optimal range for the tolerance value is limited to 0.05 to 0.15. Preferably, in a typical R&D position recruitment evaluation scenario, when this value is 0.1, the feature evaluation accuracy and feature evaluation recall reach the optimal mathematical balance point. Thus, by introducing data-driven traversal optimization, the high generalization ability and parameter adaptation ability of the two-dimensional mask filtering mechanism when switching between different job scenarios are ensured.

[0070] Understandably, lift, in statistical semantics, represents the ratio of a feature's local acceptance rate to its global average acceptance rate; it is a dimensionless, purely numerical value. Assuming the two features are independent, their theoretical joint lift should equal the product of the lifts of their respective categories. Multiplying the individual category lifts of the two features and adding a tolerance value as a redundancy buffer yields a threshold line for measuring whether there is a true synergistic gain between the two features.

[0071] Furthermore, the generation of the two-dimensional mask matrix satisfies the following logical expression:

[0072]

[0073] In the formula, Represents the corresponding number in the two-dimensional mask array. Item features and the first The matrix element values ​​of the feature combination, This represents the joint lift of any two feature combinations obtained in the preceding steps. This represents the collaborative baseline.

[0074] The actual calculated value of the joint lift is compared with the theoretical expected value of the cooperative baseline. When the actual calculated value of the joint lift is greater than the theoretical expected value of the cooperative baseline, it indicates that the simultaneous presence of the two features produces a cooperative gain effect that is significantly higher than the sum of the effects of the two features acting independently. In this case, the element value in the two-dimensional mask matrix corresponding to the feature combination is assigned a value of 1 to preserve the cross-computation topology path corresponding to the feature combination; conversely, the corresponding element value is assigned a value of 0 to directly block the cross-data flow corresponding to feature combinations with no significant cooperative value from a mathematical operation perspective.

[0075] Thus, this invention can adaptively prune invalid cross feature pairs in a high-dimensional feature space, which not only avoids the dimensionality curse and computational waste caused by fully connected crosses during deep network training, but also effectively improves the forward inference efficiency and evaluation accuracy of the recruitment matching model after engineering deployment.

[0076] S6. Based on the edge splitting points, process the original feature vector of the new resume to obtain the feature encoding value. Combine the one-dimensional mask value and the two-dimensional mask matrix to calculate the feature encoding value, obtain the matching probability, and output it.

[0077] Before proceeding to the final calculation of the new resume matching probability, it is necessary to first use historical data to initialize the neural network weights and train them to converge.

[0078] First, calculate the difference between the category lift of each feature and 1, as the individual contribution; then, sum the absolute values ​​of the individual contributions of all features to obtain the global contribution sum; if the global contribution sum is greater than 0, divide the individual contribution by the global contribution sum to obtain the one-dimensional initial weight of the corresponding feature; if the global contribution sum is equal to 0, assign equal one-dimensional initial weights to all features using a mean distribution strategy. The above calculation process satisfies the following relationship:

[0079]

[0080]

[0081]

[0082] In the formula, Indicates the first The individual contribution of each feature. This indicates the class lift calculated in the preceding steps. Indicates the global contribution. Indicates the total number of features. Indicates the first The absolute value of the individual contribution of each feature. Indicates the first The one-dimensional initial weights corresponding to each feature. By summing the absolute values ​​and using the sum as the denominator for division, the initial influence factors of all features can be compressed into a strict relative proportion space, providing an excellent physical business benchmark for the linear layers of the neural network.

[0083] Furthermore, in order to initialize the network weights of the cross features and complete the full parameter fitting of the model, the analysis method also includes: extracting target feature pairs with an element value of 1 in the two-dimensional mask matrix, multiplying the one-dimensional initial weights corresponding to the two features in the target feature pair to obtain the cross initial weights; using the one-dimensional initial weights and cross initial weights as the starting point for the discrimination network parameters, training the category subset based on the recruitment labels in historical recruitment data to obtain the one-dimensional convergence weights and cross convergence weights.

[0084] Understandably, for the strong collaborative features activated by the two-dimensional mask array, i.e., the target feature pair, their initial weights are considered as the physical superposition of the weights of the two basic features. After obtaining the complete starting point of the parameters, a standard adaptive moment estimation optimization algorithm is used to iteratively train on historical recruitment data. For example, the loss function selected in this training step is the standard binary cross-entropy loss function, and the initial learning rate of the optimizer is preferably 0.001. After multiple rounds of forward and backward propagation, the network parameters gradually stabilize, and the weight tensors in the network layers are extracted, i.e., the solidified one-dimensional convergence weights and cross-convergence weights are obtained.

[0085] like Figure 3As shown, this embodiment provides a comparison curve of convergence speed during the training process to verify the advantages of model initialization. In this comparison curve, the horizontal axis represents the training epochs of the model, and the vertical axis represents the binary cross-entropy loss. Compared with existing technologies that lack feature filtering mechanisms and prior knowledge guidance, this invention assigns reasonable one-dimensional initial weights and cross-entropy initial weights to the network based on the statistically derived global contribution sum in the early stages of training, thereby providing a high-quality prior parameter starting point with significant business value for the discriminant network. This strategy effectively avoids the problem of blind search and cold start in the vast parameter space during the early stages of model processing. The curve comparison clearly shows that existing technologies exhibit slow loss value decreases and severe oscillations during training, while this invention significantly accelerates convergence speed in complex data environments. The loss curve of this invention is not only smoother, but the final converged loss value is also much lower than that of existing technologies. This indicates that this invention effectively overcomes the training oscillation problem caused by the easy introduction of feature noise in traditional deep learning methods, improving the efficiency of model training and the stability of parameter fitting.

[0086] After the network parameters are solidified, for the new candidate resumes input in real time, based on the edge segmentation points obtained from the optimization in step S2, the original feature vector of the new resume is discretized into continuous values ​​and normalized hierarchically, transforming it into computer-readable feature encoding values. Subsequently, the feature encoding values ​​are calculated using a one-dimensional mask value and a two-dimensional mask matrix to obtain the matching probability and output it. The specific process is as follows: multiply the one-dimensional convergence weight, the corresponding feature encoding value, and the one-dimensional mask value, and sum the product results of all individual features to obtain the independent feature term; multiply the cross convergence weight, the two feature encoding values ​​of the target feature pair, and the corresponding matrix element values, and sum the product results of all target feature pairs to obtain the collaborative feature term; sum the independent feature term, collaborative feature term, and network bias term, and then map them to obtain the matching probability.

[0087] In order to complete the nonlinear transformation from the high-dimensional discrete feature space to the final scalar probability space, this step constructs a calculation model for the matching probability, which satisfies the following relationship:

[0088] In the formula, Indicates the probability of a match; This represents a nonlinear mapping function, preferably a standard logistic function. Corresponding to the above independent feature terms, Indicates the first The one-dimensional convergence weights obtained by training the features are... , as well as They represent the first Item, No. Item and the The feature encoding value corresponding to the item, Indicates the first One-dimensional mask value of the feature, Corresponding to the above collaborative feature terms, Indicates by the first Item and the The cross-convergence weights are obtained through training from the cross-combination features composed of feature items. This represents the value of the corresponding element in the two-dimensional mask matrix. This represents the network bias term. Both the one-dimensional mask value and the element value within the matrix act as physical valves; when the value is 0, the signal transmission of that dimension's feature is forcibly shut off, enabling sparse computation of the graph structure.

[0089] It should be noted that each category subset corresponds to an independent discriminant network and its output matching probability. Because different business domains (such as the R&D category subset and the sales category subset) have significantly different preference weights for candidate features and strong collaborative cross-paths between features, for the same new resume, the system can perform forward inference by parallelly calling the one-dimensional convergence weights, cross-convergence weights, and mask matrices specific to each category subset, thereby outputting a set of matching probabilities for the candidate under different job categories. This not only accurately assesses the fit between the candidate and the currently targeted job, but also automatically mines the candidate's potential second suitable job based on the numerical ranking of the matching probability set.

[0090] In this way, by transforming the lift index obtained from traditional business statistical calculations into the prior initialization weights of the deep learning network, and forcibly introducing one-dimensional and two-dimensional mask matrices for dot multiplication interception in the final forward inference formula, not only is the convergence efficiency of the calculation model accelerated and the performance fluctuations caused by cold start avoided, but also the hard truncation characteristic of the mask eliminates the multiplication and addition operations of invalid features from the bottom layer, so that the overall evaluation link has both evaluation accuracy and response performance when dealing with massive high-dimensional recruitment resume streams.

[0091] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A data-driven method for optimizing the efficiency of the recruitment process, characterized in that: include: Obtain historical recruitment data and divide it into category subsets. Extract various features from each category subset and count the total number of applicants and the total number of hires for each category subset. For each feature, calculate the second-order difference of the ordered features, determine the position of the smallest second-order difference as the marginal split point, and divide the ordered features into low-value intervals and high-value intervals based on the marginal split point; Calculate the category lift of each feature and the joint lift of any two features based on the total number of applicants and the total number of hires. Calculate the lift difference between the high-value interval and the low-value interval and generate an adaptive threshold. Generate a one-dimensional mask value based on the lift difference and the adaptive threshold. A two-dimensional mask matrix is ​​generated based on the category lift and the joint lift; The original feature vector of the new resume is processed based on the edge splitting points to obtain the feature encoding value. The feature encoding value is then calculated by combining the one-dimensional mask value and the two-dimensional mask matrix to obtain the matching probability and output it. The analytical method includes: Extract target feature pairs with an element value of 1 from a two-dimensional mask array. Multiply the one-dimensional initial weights corresponding to the two features in the target feature pair to obtain the cross initial weights. Use the one-dimensional initial weights and cross initial weights as the starting point for the discriminant network parameters. Train the category subset based on the recruitment labels in historical recruitment data to obtain the one-dimensional convergence weights and cross convergence weights. The method for obtaining the matching probability includes: Multiply the one-dimensional convergence weight, the corresponding feature encoding value, and the one-dimensional mask value, and sum the product results of all individual features to obtain an independent feature term; multiply the cross convergence weight, the two feature encoding values ​​of the target feature pair, and the corresponding in-matrix element values, and sum the product results of all target feature pairs to obtain a collaborative feature term; sum the independent feature term, collaborative feature term, and network bias term, and then map them to obtain the matching probability.

2. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, The calculation of the second-order difference for the ordered features among the features includes: The number of applicants and the number of hires at each level are statistically analyzed for the ordered characteristics. The ratio of the number of hires to the number of applicants is used as the smoothed hiring rate. The smoothed difference between the smoothed hiring rates of two adjacent levels is calculated, and the change in the smoothed difference between two adjacent levels is determined as the second-order difference.

3. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, Before calculating the second-order difference for the ordered features in each feature, the analysis method further includes: detecting the number of valid levels of the ordered features in the category subset; in response to the number of valid levels being less than or equal to 2, skipping the step of calculating the second-order difference, aggregating all valid levels into a single interval, and directly calculating the category lift corresponding to the single interval.

4. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, The calculation of the category lift of an individual feature and the joint lift of any combination of two features includes: Obtain the number of single-feature recruits and the number of single-feature applications that possess the aforementioned single feature. Divide the product of the number of single-feature recruits and the total number of applications by the product of the number of single-feature applications and the total number of recruits to obtain the category lift of the aforementioned single feature. Obtain the number of dual-feature applications that have any two of the stated features. In response to the number of dual-feature applications being greater than 0, obtain the number of dual-feature hires that have any two of the stated features. Divide the product of the number of dual-feature hires and the total number of applicants by the product of the number of dual-feature applications and the total number of hires to obtain the joint lift.

5. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, The generation of a one-dimensional mask value based on the lift difference and an adaptive threshold includes: Calculate the absolute value of the difference between the class lift of the single feature corresponding to the high value interval and the class lift of the single feature corresponding to the low value interval to obtain the lift difference value; calculate the average value of the lift differences of all ordered features and use it as the adaptive threshold; in response to the lift difference value being greater than or equal to the adaptive threshold, assign the corresponding one-dimensional mask value to 1; in response to the lift difference value being less than the adaptive threshold, assign the corresponding one-dimensional mask value to 0.

6. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, The step of generating a two-dimensional mask matrix based on the class lift and the joint lift includes: calculating the product of the class lift corresponding to each feature in any two feature combinations, and adding the product to a preset tolerance value to obtain a collaborative baseline; In response to the joint lift of any two feature combinations being greater than the cooperative baseline, the element value of the corresponding feature combination in the two-dimensional mask array is set to 1; in response to the joint lift of any two feature combinations being less than or equal to the cooperative baseline, the element value of the corresponding feature combination in the two-dimensional mask array is set to 0.

7. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, Before obtaining and outputting the matching probability, the analysis method further includes: Calculate the difference between the category lift of each feature and 1, and use it as the individual contribution. Calculate the sum of the absolute values ​​of the individual contributions of all features to obtain the global contribution sum. In response to the global contribution sum being greater than 0, divide the individual contribution by the global contribution sum to obtain the one-dimensional initial weight of the corresponding feature. In response to the global contribution sum being equal to 0, assign equal one-dimensional initial weights to all features using a mean distribution strategy.

8. The data-driven recruitment process efficiency optimization analysis method according to claim 1, characterized in that, The matching probability satisfies the formula: ;in, This represents the matching probability. Represents a nonlinear mapping function. Corresponding to the independent feature term, Indicates the first One-dimensional convergence weight of the feature term, , as well as They represent the first Item, No. Item and the The feature encoding value corresponding to the item, Indicates the first One-dimensional mask value of the feature, Corresponding to the aforementioned collaborative feature term, Indicates by the first Item and the The cross-convergence weight of the cross-combination features composed of feature items. This represents the value of the corresponding element in the two-dimensional mask matrix. This refers to the network bias term.

Images