Human Resources Data Management Method Based on Big Data

Abstract

This application discloses a human resources data management method based on big data, which relates to human resources management and includes: extracting gender and age from a sample resume dataset as bias features; semantically encoding the sample resume dataset to obtain a set of semantically encoded feature vectors of the resumes; performing dimensionality reduction processing on the set of semantically encoded feature vectors of the resumes based on the bias features to obtain a set S of feature vectors that reduce the impact of bias; performing cluster analysis on the set S of feature vectors that reduce the impact of bias to obtain multiple feature vector center points representing different feature combinations; using an improved hash algorithm to semantically encode the resume of an object to be evaluated to obtain a hash value of the object to be evaluated; calculating the similarity between the hash value of the object to be evaluated and the center points of each feature vector, and performing human resources data management based on the similarity. This application addresses the gender and age biases that exist in human resources data management and reduces the impact of bias on evaluation.

Description

Technical Field

[0001] The present application relates to the field of human resource management, and in particular to a human resource data management method based on big data. Background Art

[0002] With the deepening of digital transformation, human resource management is undergoing a significant shift from traditional, empirical decision-making to data-driven decision-making. The application of technologies such as big data and artificial intelligence is becoming increasingly widespread in HR areas such as recruitment, talent assessment, and performance management. Talent screening and matching based on the analysis of massive resume data has become an industry standard. However, with increasing automation, algorithmic bias in human resource data management systems has become increasingly prominent. Research shows that these systems often inadvertently learn and amplify historical biases in the data, particularly those related to gender and age.

[0003] Existing techniques for addressing bias in human resources data management primarily mitigate it by adjusting the distribution of training datasets. However, this approach often results in information loss and struggles to address the complex interactions of multiple bias factors.

[0004] For example, the related patent CN118278902A discloses a human resources data management method based on big data, including: semantically encoding the resume data of employees labeled as successful employees to obtain a set of semantically encoded feature vectors of the successful employee resume data; performing autocorrelation resume data saliency characterization on the set of semantically encoded feature vectors of the successful employee resume data to obtain a semantically fused feature vector of the commonality of the successful employee resume features; semantically encoding the resume of the obtained object to be evaluated to obtain a semantically encoded feature vector of the object to be evaluated resume; and calculating the hash similarity between the semantically encoded feature vector of the object to be evaluated resume and the semantically fused feature vector of the commonality of the successful employee resume features to determine whether to filter out the object to be evaluated resume. However, this solution is directly based on the modeling of "successful employee" resumes. If the "successful employees" in the historical data themselves have gender or age biases, the model will inherit and amplify these biases. Summary of the Invention

[0005] In response to the gender and age biases that exist in human resource data management in the existing technology, this application provides a human resource data management method based on big data, which reduces the impact of gender and age bias on talent evaluation through semantic encoding and bias feature dimensionality reduction processing.

[0006] The purpose of this application is achieved through the following technical solutions.

[0007] One aspect of the present application provides a human resources data management method based on big data, including: S1, obtaining a sample resume data set and a resume of an object to be evaluated; S2, extracting gender and age from the sample resume data set as bias features; S3, semantically encoding the sample resume data set to obtain a set of semantically encoded feature vectors of the resume; S4, performing dimensionality reduction processing on the set of semantically encoded feature vectors of the resume based on the bias features to obtain a set S of feature vectors that reduce the influence of bias; S5, performing cluster analysis on the set S of feature vectors that reduce the influence of bias to obtain multiple feature vector center points representing different feature combinations; S6, using an improved hash algorithm to semantically encode the resume of the object to be evaluated to obtain a hash value of the object to be evaluated; S7, calculating the similarity between the hash value of the object to be evaluated and the center points of each feature vector, and performing human resources data management based on the similarity.

[0008] Furthermore, S4, obtaining a feature vector set S for reducing the influence of bias, includes: respectively calculating the correlation coefficient between each dimension in the semantic coding feature vector set of the resume and the gender feature and the age feature; marking the dimensions corresponding to the absolute value of the correlation coefficient greater than the preset threshold as bias-sensitive dimensions, and obtaining a gender bias dimension set and an age bias dimension set; respectively calculating the attenuation coefficients λ1 and λ2 of the gender bias dimension set and the age bias dimension set; generating a bias adjustment weight matrix according to the absolute value of the correlation coefficient of each dimension and the attenuation coefficients λ1 and λ2 of the bias-sensitive dimensions; performing dimensionality reduction processing on the semantic coding feature vector set of the resume according to the bias adjustment weight matrix, and obtaining a feature vector set S for reducing the influence of bias.

[0009] Furthermore, the attenuation coefficients λ1 and λ2 of the gender bias dimension set and the age bias dimension set are calculated respectively, including: Among them, μ1 and μ2 represent the mean, σ1 and σ2 represent the standard deviation; c1 represents the gender bias suppression parameter, c2 represents the age bias suppression parameter, and b1 and b2 are adjustment parameters.

[0010] Further, S5, obtains multiple feature vector center points representing different feature combinations, including: setting the optimal cluster number N, where N is a positive integer greater than 3; calculating the local density value ρ(v) of each feature vector v in the feature vector set S, where the local density value ρ(v) is defined as the number of feature vectors in the R-neighborhood of the feature vector v, where the R-neighborhood refers to an area with a Euclidean distance less than a radius R; randomly selecting a feature vector v1 from the feature vector set S as the first initial center point, and adding v1 to the center point set C; for each feature vector v in the feature vector set S that is not selected as a center point, i , calculate v i The distance D to the nearest center point in the center point set C i, and calculate the eigenvector v i The local density value ρ(v i );According to the distance D i and the local density value ρ(v i ) Calculate each unselected eigenvector v i The probability of selection {P i}; According to the selection probability {P i}, select the feature vector v with the highest selection probability max as the next center point and v max Add to the center point set C; if the newly selected center point v max If the distance from any selected center point is less than the preset minimum distance threshold MinDist, then v is abandoned. max , and select the feature vector with the second highest selection probability from the remaining unselected feature vectors as the center point; repeat until the center point set C contains N initial cluster center points.

[0011] In particular, traditional clustering algorithms such as K-means usually use random initialization or select center points based on a single distance metric, which performs poorly on high-dimensional, non-uniformly distributed datasets such as human resources data, easily leading to suboptimal clustering results and affecting the accuracy of talent classification and matching.

[0012] This application introduces a cluster center selection method that combines local density and distance factors. Specifically, it introduces a local density calculation mechanism. By counting the number of eigenvectors in the R-neighborhood, the system can identify dense areas of data distribution. These areas usually represent groups of talents with common characteristics. A comprehensive selection probability formula is designed. The weight coefficients α and β are used to achieve a flexible balance between distance and density factors, solving the limitation of single factor consideration in traditional methods. The outlier detection and penalty mechanism is innovatively introduced, and the outlier index O i The penalty coefficient γ effectively reduces the likelihood of abnormal resume data being selected as center points, improving the representativeness and stability of the clustering results. The MinDist mechanism, which sets a minimum distance threshold, avoids excessive clustering of center points and ensures that the selected center points fully cover all important areas of the feature space, thereby more comprehensively representing the characteristic combinations of different types of talents.

[0013] Further, Among them, α and β are weight coefficients used to balance the influence of distance factor and density factor, γ is the outlier penalty coefficient, O i is an outlier indicator, when the eigenvector v i When the average distance between its K nearest neighbors exceeds the preset multiple of the global average distance, O i The value is 1, otherwise it is 0.

[0014] Furthermore, S6 uses an improved hash algorithm to semantically encode the resume of the evaluation object, including: constructing m independent independent hash function families {h1,h2,.....,h m}, where each hash function family h l Represents a set of functions that map feature vectors to b-bit binary hash codes; using m independent independent hash function families {h1,h2,.....,h m Perform feature mapping on the resume of the object to be evaluated to obtain m groups of b-bit binary hash codes; combine the m groups of b-bit binary hash codes in sequence as the hash value of the object to be evaluated.

[0015] In particular, existing human resources data processing techniques suffer from high computational complexity and storage overhead for high-dimensional feature vectors, as well as inefficient similarity calculations. Traditional methods typically directly calculate Euclidean distance or cosine similarity based on the raw feature vectors. This presents a significant performance bottleneck when processing massive resume data, making it incapable of meeting the demands of real-time talent matching. Furthermore, single hash encoding schemes are prone to information loss and hash collisions, reducing matching accuracy.

[0016] This application introduces an improved hash algorithm based on multiple hash function families and a specialized similarity calculation method. In detail, it designs a coding framework based on m independent hash function families. Each function family is responsible for generating part of the hash code. This multi-function family architecture significantly reduces the information loss and conflict probability that may be caused by a single hash function.

[0017] A hash construction method based on random unit vector projection is adopted. By calculating the inner product of the feature vector and the random vector and binarizing it, the similarity relationship in the high-dimensional feature space is retained, and complex vector operations are converted into simple bit operations.

[0018] Converting high-dimensional feature vectors into compact binary hash codes significantly reduces storage requirements and is particularly suitable for the construction and maintenance of large-scale human resources databases.

[0019] Designed a similarity calculation formula based on Hamming distance

[0020] The complex vector similarity calculation is simplified into efficient bit operations, reducing the computational complexity from O(d) to O(b×m), where d is the original dimension, which is usually much larger than b×m.

[0021] By synthesizing the results of multiple hash function families, a "voting mechanism" for similarity calculation is implemented, which improves the stability and robustness of similarity evaluation and effectively solves the randomness problem in a single encoding scheme.

[0022] Furthermore, we construct a hash function family h i , including: randomly generating b d-dimensional unit vectors {r1,r2,.....,r d}, where d represents the dimension of the feature vector; for the feature vector v of the object to be evaluated e , calculate v e With each random unit vector r k The inner product of k =v e ×r k ; When the projection value p k When it is greater than or equal to zero, the corresponding hash bit is set to 1, otherwise it is set to 0; combine the b hash bits in sequence to form a b-bit binary hash code as the hash function family h l .

[0023] Furthermore, S7, human resources data management is performed based on the similarity, including: using the improved hash algorithm in step S6 to semantically encode the center point of each feature vector to obtain the hash value of each feature vector center point; respectively calculating the similarity between the hash value of the object to be evaluated and the hash value of each feature vector center point; classifying the object to be evaluated based on the calculated similarity, and classifying the object to be evaluated into the category corresponding to the feature vector center point with the highest similarity; based on the classification result, performing human resources management on the object to be evaluated, and the management includes talent matching and job recommendation.

[0024] Furthermore, the similarity Sim (H e ,H j ), using the following formula: Among them, H e Represents the hash value of the object to be evaluated, H j Represents the hash value of the center point of the jth eigenvector, h q (*) represents the hash code part generated by the i-th hash function family, Hamming(*) represents the Hamming distance, and b represents the number of hash code bits generated by each hash function family.

[0025] Another aspect of the present application also provides a human resources data management system based on big data, which is used to execute a human resources data management method based on big data of the present application.

[0026] Compared with the existing technology, the advantages of this application are:

[0027] Existing human resource management systems suffer from significant gender and age bias, primarily manifested in an overreliance on non-ability-related factors during resume screening. Traditional solutions typically simply remove gender- and age-related feature dimensions or apply uniform weights. This approach suffers from two key flaws: first, it fails to accurately identify which dimensions truly contain biased information, resulting in excessive loss of useful information; second, it lacks adaptive adjustment capabilities for different types of bias, making it impossible to differentiate based on the degree of bias.

[0028] This application uses the correlation coefficient analysis method to accurately quantify the degree of correlation between each feature dimension and gender and age, and objectively identifies bias-sensitive dimensions through preset thresholds, avoiding information loss caused by subjective judgment. The attenuation coefficient calculation formula is innovatively designed. and This formula considers the statistical properties of the bias dimension set (mean μ and standard deviation σ), enabling adaptive adjustment based on bias strength. By using a bias-adjusted weight matrix, rather than simply removing dimensions, it achieves refined dimensionality reduction of feature vectors. This maximizes capability-related information while effectively reducing the weight of bias. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The present application will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same numbers represent the same structures, wherein:

[0030] Figure 1 is an exemplary flow chart of a human resources data management method based on big data according to some embodiments of the present application;

[0031] Figure 2 is an exemplary flow chart of generating a feature vector set S for reducing bias effects according to some embodiments of the present application;

[0032] Figure 3 The hash function family h is constructed according to some embodiments of the present application. i An exemplary flow chart of . DETAILED DESCRIPTION

[0033] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0034] like Figure 1As shown, a sample resume data set and the resume of the object to be evaluated are obtained; gender and age are extracted from the sample resume data set as bias features; the sample resume data set is semantically encoded to obtain a set of semantically encoded feature vectors of the resume; based on the bias features, the set of semantically encoded feature vectors of the resume is dimensionality reduced to obtain a set S of feature vectors that reduce the influence of bias; cluster analysis is performed on the set S of feature vectors that reduce the influence of bias to obtain multiple feature vector center points representing different feature combinations; an improved hash algorithm is used to semantically encode the resume of the object to be evaluated to obtain a hash value of the object to be evaluated; the similarity between the hash value of the object to be evaluated and the center points of each feature vector is calculated, and human resources data management is performed based on the similarity.

[0035] S1. Obtain a sample resume dataset and the resumes of the candidates to be evaluated. In this example, the sample resume dataset was obtained by exporting 50,000 resumes from an online recruitment platform that had completed the recruitment process. These resumes covered 15 industry sectors, including IT, finance, and sales. These resumes were stored in a structured JSON format, and each resume contained fields such as personal information, education background, work experience, project experience, and skill certifications.

[0036] The raw resume data was preprocessed, including: removing irregularly formatted content, such as HTML tags and special symbols; standardizing time formats by converting time information in various formats to the "yyyy-mm-dd" format; completing missing fields by inferring and completing key information missing from resumes like education level using contextual information; and removing duplicate resumes by calculating content similarity between them to identify and remove duplicate submissions. After preprocessing, 43,682 valid resumes were retained.

[0037] Resumes of candidates to be evaluated are retrieved in real time through the company's recruitment system's API, or imported in batches through the file upload function. The system supports resume files in various formats, including PDF, Word, and HTML, and converts unstructured resumes into a unified structured data format using optical character recognition (OCR) and natural language processing technologies.

[0038] S2 extracts gender and age from a sample resume dataset as bias features. In this example, explicitly labeled gender values ​​are directly extracted from the resume's personal information field. Using a name gender inference algorithm, a character-level BiLSTM model trained on 120,000 Chinese names is used to infer gender from resumes without explicit gender labels, achieving an accuracy rate of approximately 94.3%. Pronouns (he / she) and gender-related descriptors used in the resumes are used to assist in the judgment. Ultimately, gender is represented as a binary value (0 for female, 1 for male).

[0039] Age information was extracted using the following methods: Prioritizing direct extraction of explicit birth dates or age values ​​from personal information; if direct age information was missing, an approximate age was calculated based on the earliest enrollment date of academic qualifications, based on statistical patterns in the Chinese education system (e.g., undergraduate enrollment is generally 18-19 years old). This was cross-validated with total years of work experience. The extracted age values ​​were maintained as continuous values ​​and standardized using the formula (actual age - mean age) / standard deviation, resulting in a distribution with a mean of 0 and a standard deviation of 1.

[0040] The sample dataset consists of a gender distribution of 63.2% male and 36.8% female. The age distribution ranges from 22 to 55 years old, with a mean age of 33.7 years and a standard deviation of 5.9 years. The extracted gender and age feature data are stored in a separate feature matrix to provide a foundation for subsequent bias analysis.

[0041] like Figure 2 As shown in Figure 3, S3 performs semantic encoding on the sample resume dataset to obtain a set of semantic encoding feature vectors for the resumes. Specifically, multimodal feature fusion and deep semantic representation techniques are used to encode each resume and generate a fixed-dimensional feature vector. First, the resume is parsed into the following five main information modules: Basic Information Module: Contains 25 basic attributes such as identity, educational level, and professional classification; Education Background Module: Extracts 12 educational features such as school level, professional ranking, and GPA; Work Experience Module: Encodes 30 professional features such as position level, industry attributes, and employer size; Skill Label Module: Extracts a 150-dimensional skill feature vector from skill descriptions using the TF-IDF method; Project Experience Module: Extracts 18 features such as project type, scale, and role from project descriptions. For categorical features, such as educational level, one-hot encoding is used to convert them into numerical vectors; for ordinal features, such as position level, ordinal encoding is used to digitize them; and for textual features, such as professional name, entity normalization is used to map them to a standardized classification system.

[0042] For the unstructured text content in resumes (such as job descriptions, project content, etc.), a hierarchical semantic encoding strategy is adopted: Sentence-level semantic representation: The text is divided into sentence units, and the semantic vector of each sentence is extracted through the pre-trained Chinese BERT-Large model. The model contains a 24-layer Transformer structure with a hidden layer dimension of 1024. The output of the [CLS] position is used as the sentence representation, and a 768-dimensional feature vector is obtained. Paragraph-level semantic aggregation: For sentence vectors belonging to the same semantic paragraph (such as the description of the same work experience), the importance weights of different sentences are calculated through the self-attention mechanism (Self-Attention), and the paragraph-level semantic vector is obtained by weighted averaging. Module-level semantic integration: Multiple paragraph vectors of the same module (such as all work experiences) are integrated using the Temporal Attention Network (TAN). This network considers the differences in the importance of experiences in different time periods and generates a 384-dimensional module-level semantic vector.

[0043] To capture the relationships between resume elements, we constructed the following relationship features: Temporal consistency: This feature calculates eight temporal features, including the temporal overlap between work and education history and a career continuity indicator. Domain consistency: This feature uses a cosine similarity algorithm to calculate 12 consistency indicators, including the correlation between education and job content and the matching between skills and project experience. Progression: This feature incorporates six developmental features, including the slope of job progression and the increasing skill complexity indicator, to reflect a candidate's growth trajectory.

[0044] Finally, these features are integrated through a multi-layer feature fusion network: Feature normalization: Feature vectors from different sources and dimensions are Z-score normalized to achieve a mean of 0 and a standard deviation of 1, eliminating dimensional differences. Low-order feature fusion: An attention-guided feature selection mechanism is used to initially fuse structural features with the semantic features of the corresponding module to generate a module-level fused feature vector. High-order feature fusion: A feature fusion network consisting of three fully connected layers (1024→512→256) is constructed to concatenate all module-level fused features and relational features and map them into a unified space. The network uses the Leaky ReLU activation function in the intermediate layers and the Tanh function in the output layer to ensure that the final feature vector has a range of [-1, 1] for each dimension. Dimensionality reduction and compression: To improve subsequent processing efficiency and remove redundancy, features are compressed into a final 256-dimensional semantic encoding feature vector using a regularized autoencoder. This embodiment finally obtains a resume semantic coding feature vector set E={e1,e2,......,e n}, where n is the size of the sample resume dataset, and each vector e i Represents the semantic encoding features of a resume. These feature vectors comprehensively capture the resume's structural information, semantic content, and internal relationship features, laying the foundation for subsequent bias processing and similarity calculation.

[0045] S4, using a combination of statistics and machine learning methods, identifies and precisely adjusts the bias dimensions in the resume semantic encoding feature vector set. First, calculate the Pearson correlation coefficient between each dimension in the feature vector E and the gender feature G and age feature A: Calculate separately: Gender correlation coefficient Age correlation coefficient Cov(*,*) represents covariance.

[0046] An analysis of the 256-dimensional feature vectors of 43,682 resume samples revealed that the correlation coefficient for gender ranged from -0.63 to 0.71, with an average absolute value of 0.14. The correlation coefficient for age ranged from -0.58 to 0.66, with an average absolute value of 0.19. This indicates that certain feature dimensions are indeed significantly correlated with gender or age. To verify the statistical significance of these correlations, a t-test was performed on each correlation coefficient, and a p-value was calculated. The Benjamini-Hochberg method was used to correct for multiple testing to ensure that the correlations identified were true rather than random associations.

[0047] Set the absolute value threshold of the correlation coefficient τ G = 0.35 (sex) and τ A = 0.30 (age), the dimensions whose absolute value of the correlation coefficient exceeds the threshold are marked as bias-sensitive dimensions: gender bias dimension set Age bias dimension set From the 256 feature dimensions, 43 gender-bias-sensitive dimensions and 37 age-bias-sensitive dimensions were identified. Further analysis revealed that gender-bias-sensitive dimensions are primarily concentrated in traits related to communication style and career field selection, while age-bias-sensitive dimensions are highly correlated with time-accumulated traits such as job level and management experience. If these dimensions are not addressed, the system can lead to unfair assessments of candidates based on gender and age information in resumes. For example, the system might underscore women's technical abilities or underscore the leadership abilities of young candidates.

[0048] For different types of bias features, the adaptive attenuation coefficients λ1 and λ2 are calculated:

[0049] Where: μ1 and μ2 are the gender bias dimension set B Gand age bias dimension set B A The average of the absolute values ​​of the correlation coefficients in the dataset is calculated; σ1 and σ2 are the standard deviations of the absolute values ​​of the correlation coefficients in the corresponding set, respectively. c1 = 0.75 and c2 = 0.65 are the gender and age bias suppression parameters, reflecting the strength of suppression; b1 = 1.2 and b2 = 1.5 are adjustment parameters, controlling sensitivity to data dispersion. We obtain μ1 = 0.47, σ1 = 0.09, μ2 = 0.42, and σ2 = 0.08, which in turn yields λ1 = 0.63 and λ2 = 0.58. When the correlations between bias dimensions are generally high (large μ) or the distribution is more concentrated (small σ), the system considers the bias signal more clear and should strengthen suppression (small λ). Conversely, when the correlations are generally low or the distribution is more dispersed, the system adopts a more conservative suppression strategy (large λ) to avoid over-adjustment and loss of useful information.

[0050] Based on the correlation coefficient and the attenuation coefficient, a 256×256 diagonal weight matrix W is constructed: for j∈B G And j∈B A (Dimensions affected by both gender and age bias): For j∈B G and

[0051] (Dimensions affected only by gender bias): for And j∈B A (Dimensions affected only by age bias): for and (Dimensions without significant bias): W[j, j] = 1. The weights of bias-sensitive dimensions are precisely adjusted, with the adjustment amount proportional to the degree of bias correlation. Dimensions insensitive to bias retain their original weights to ensure that normal feature information is not affected. In this example, the weight matrix range is [0.35, 1.0], with an average value of 0.89. The weight distribution is skewed to the right, indicating that only a few dimensions are significantly adjusted, while the majority of dimensions retain their original information.

[0052] Finally, the weight matrix W is used to reduce the dimensionality of the original feature vector set E: the weight matrix W is a 256×256 diagonal matrix, and its diagonal elements W[j,j] represent the retention weights of the j-th dimension features. The characteristics of the diagonal matrix make the matrix multiplication essentially perform independent scaling operations on each dimension of the feature vector, rather than mixed transformations between dimensions, which ensures the semantic consistency of the feature space. The original feature vector set E is stored in the form of an n×256 matrix (n=43682 is the number of samples), and the weight matrix W is a sparse diagonal matrix of 256×256. Considering the sparse characteristics of W (all non-diagonal elements are 0), the system uses diagonal vector storage and only saves 256 diagonal element values, which significantly reduces storage requirements. For the original feature vector of sample resume i and the diagonal weight vector w=[w1,w2,......,w 256 ], the calculation formula for dimensionality reduction is: Right now:

[0053] S5: Cluster analysis is performed on the bias-reducing feature vector set S to obtain multiple feature vector centers representing different feature combinations. Specifically, based on an analysis of the diversity of human resources data, the optimal number of clusters, N, is set to 15. This value is determined by the Silhouette Coefficient (SVC) test. Testing was performed within the range of N = 5 to 25, with N = 15 reaching a maximum value of 0.68, indicating the most reasonable cluster structure.

[0054] The system sets the following key parameters: local density calculation radius R = 0.35 (approximately the 25% quantile of the Euclidean distance distribution between all pairs of eigenvectors); density factor weight β = 0.8; distance factor weight α = 1.2; outlier penalty coefficient γ = 0.9; K nearest neighbor value K = 20; outlier judgment multiple threshold = 2.5; minimum center point distance threshold MinDist = 0.6.

[0055] For each eigenvector v∈S, the system calculates its local density value ρ(v i ): Construct a ball tree data structure to store all feature vectors to optimize the efficiency of nearest neighbor search; for each vector v, query the ball tree to find the number of vectors whose Euclidean distance to v is less than R; this number is the local density value ρ(v i). In this embodiment, the analysis of the local density values ​​of 43,682 feature vectors shows that the density distribution range is 12 to 487, the average density is 156, and the standard deviation is 92. The density distribution presents an obvious multi-peak structure, reflecting the existence of multiple high-density areas in the resume data, which represent different types of talent combination characteristics. The system detects that approximately 25% of the vectors are located in high-density areas (ρ>200), 50% are located in medium-density areas (80<ρ<200), and 25% are located in low-density areas (ρ<80). This distribution characteristic confirms the necessity of adopting a clustering method that takes local density into account.

[0056] The system uses a random selection method to determine the first centroid: a probability distribution is constructed such that the probability of a feature vector being selected as the first centroid is proportional to its local density. Based on this probability distribution, a feature vector v1 is randomly sampled as the first centroid. v1 is then added to the initially empty set of centroids C. This density-based sampling strategy avoids selecting extreme outliers as the first centroid, improving clustering stability. The system records the random seed value during each selection process to ensure reproducible results.

[0057] The system selects the remaining N-1 center points through the following iterative process: For each non-center point v i , calculate: distance factor D i :v i The Euclidean distance to the nearest point in the current center point set C; the density factor ρ(v i ): calculated local density value; outlier index O i :First calculate v i The average distance to its K=20 nearest neighbors Then calculate the global K nearest neighbor average distance when When, O i =1, otherwise O i = 0. Calculate the selection probability: This formula takes into account three factors: The term ensures that points farther away from the existing center are selected to increase the coverage; ρ(v i ) β The item tends to select points located in high-density areas to enhance representativeness;

[0058] (1-γ×O i ) item reduces the possibility of outliers being selected as center points and improves stability; select the one with the highest probability P i The eigenvector v max as candidate center points.

[0059] Test v maxDoes the distance to the existing center point meet the minimum distance constraint: calculate v max The distance to each center point in C; if there is a center point with a distance less than MinDist=0.6, then reject v m ; In the case of rejection, select the vector with the second highest probability from the remaining vectors; add the vector that passed the test to the center point set C; update the distance factor D of all remaining vectors i (Only the distance to the newly added center point needs to be updated); repeat until the center point set C contains N=15 vectors.

[0060] After initially selecting 15 centroids, the system performed the following optimization and analysis steps: Verification of centroid representativeness: The distance from each eigenvector to the nearest centroid was calculated. 98.3% of the vectors were within R×1.5 of the nearest centroid, demonstrating that the selected centroids provided good coverage of the dataset. Centroid feature interpretation: Principal component analysis (PCA) and explanatory feature analysis were performed on the eigenvectors corresponding to each centroid to extract the core feature combinations represented by each centroid. For example, centroid C1: a technical management feature combination representing mid- to senior-level management personnel with a technical background; centroid C2: a sales performance feature combination representing personnel with outstanding sales performance; centroid C3: a R&D innovation feature combination representing personnel with innovative R&D capabilities; and so on. Centroid stability testing: The clustering process was restarted 10 times (using different random seeds) to compare the degree of match between the resulting centroid sets. The results showed that, on average, 12.3 centroids found a match between runs, demonstrating the method's good stability. By using this method to identify feature vector center points while balancing density and distance, the system successfully extracted 15 center points representing different skill sets and talent characteristics from the feature vector set S, mitigating the impact of bias. This laid the foundation for subsequent similarity calculations and talent matching. These center points effectively capture the diversity and inherent structure of talent data, representing different combinations of talent characteristics and thus supporting more accurate human resources data management decisions.

[0061] like Figure 3As shown in step S6, an improved hash algorithm is used to semantically encode the resume of the object to be evaluated, obtaining the hash value of the object to be evaluated. Specifically, the number of hash function families, m, is 12, and each function family generates a b-8-bit hash code. The feature vector dimension, d, corresponds to the dimension of the feature vector after dimensionality reduction in step S4. The total length of the resulting hash value is m × b = 96 bits. These parameters have been optimized experimentally to achieve a balance between feature preservation and computational efficiency. Experiments show that when m < 8, the hash collision rate is too high, resulting in a decrease in discriminative power. When m > 16, the computational overhead increases, but the improvement in similarity accuracy is limited. Similarly, b = 8 provides sufficient local encoding capability, allowing a single function family to capture the distribution characteristics of the feature vector in a specific direction.

[0062] To construct each hash function family hi, the system uses an improved orthogonal random vector generation technology: First, a standard normal distribution (mean = 0, std = 1) is used to randomly generate b × d independent Gaussian random numbers to form b d-dimensional vectors; Gram-Schmidt orthogonalization is applied to ensure that the random vectors in each group are as orthogonal as possible, thereby enhancing the independence of each hash bit. Each vector is L2 normalized and converted into a unit vector {r1, r2, ..., r b To ensure independence among the m hash function families, each family uses a different random seed to generate its internal set of random vectors. The system persistently stores the generated m × b random unit vectors, ensuring that multiple hash operations on the same feature vector produce consistent results.

[0063] For the feature vector v of the object to be evaluated e , the system performs the following processing: For the hash function family h i Each random unit vector r in j , calculate the eigenvector v e With r j The inner product of: According to the projection value p j The sign of the hash bit value is determined: if p j ≥0, the hash bit is set to 1; if p j <0, the hash bit is set to 0. This processing method is geometrically equivalent to dividing the feature space with a random hyperplane. The side of the hyperplane on which the feature vector falls determines the value of the corresponding hash bit. When two feature vectors are similar, they are more likely to be on the same side, thus generating similar hash codes. The system uses a vectorized calculation method to convert the feature vector v e With the function family h i The matrix R consisting of all random vectors i By multiplying (b×d), b projection values ​​are obtained at one time, which significantly improves the calculation efficiency.

[0064] In this embodiment, for the feature vector v to be processed e , the system first processes each hash function family in turn: for the first hash function family h1, the system takes the 8 random unit vectors {r i1 ,r i2 ,.....,r i8}; Calculate the eigenvector v e The inner product value with these 8 vectors: p 1j =v e ×r 1j ; Based on the sign of the inner product value, generate an 8-bit hash code H1: bit j = 1 if p 1j ≥0; jth bit = 0 if p 1j <0; The system adopts parallel computing mode and uses matrix multiplication to obtain all projection values ​​at one time: [p 11 ,p 12 ,.....,p 18 ]=v e ×[r 11 ,r 12 ,.....,r 18 ] T When converting projection values ​​into hash bits, vectorized symbolic function operations are used to further improve computational efficiency.

[0065] The system processes all 12 hash function families in a predetermined order: Repeat the above process to obtain function families h1 to h 12 The 8-bit hash code H1, H2, ....., H 12 ; Each function family uses a different set of random vectors to ensure that different aspects of the eigenvector information are captured; to ensure processing speed, the system uses batch processing technology to combine 96 random vectors from 12 function families into a large matrix (96×256), and obtain all projection values ​​simultaneously through a single matrix multiplication.

[0066] After obtaining 12 groups of 8-bit hash codes, the system performs precise ordered connection: according to the order of the function family, the 12 groups of hash codes are sequentially spliced: H=H1|H2|.....|H 12 , where “|” represents the bit sequence concatenation operation; during the concatenation process, the internal bit order of each group of hash codes remains unchanged to ensure that the position encoding meaning of the final hash value is consistent; the concatenation result is a binary sequence H with a total length of 96 bits.

[0067] To optimize storage and computing efficiency, the system converts the 96-bit binary sequence into a more compact data structure: the 96-bit sequence is divided into 12 bytes in units of 8 bits, recorded as B1, B2, ..., B 12; Each byte represents an unsigned integer value in the range of 0 to 255; finally, the 12 bytes are stored as an integer array of length 12 [B1, B2, ....., B 12 ]; arrays are stored in contiguous memory to optimize access performance; in memory-constrained scenarios, the system provides an optional compression scheme that compresses 12 bytes into three 32-bit integers with minimal information loss. The combination of multiple independent hash function families reduces the fluctuations caused by a single random hyperplane partition, enhancing the stability of the hash representation.

[0068] S7, calculate the similarity between the hash value of the object to be evaluated and the center point of each feature vector, and manage human resources data based on the similarity. Specifically, in this embodiment, the system first applies the same hash algorithm as S6 to the 15 feature vector center points identified in step S5: using exactly the same m=12 hash function family and random vector set to ensure the consistency of the coding space; for each center point feature vector c j ,j=1,2,.....,15 respectively calculate the 96-bit binary hash value H j ; These center point hash values ​​are stored in the memory index structure to support fast similarity queries; the pre-calculation of center point hash values ​​greatly improves the system operation efficiency, because the number of center points is far less than the number of resumes to be evaluated, and the center points are relatively stable and do not need to be updated frequently.

[0069] The system uses a segmented weighted hash similarity calculation method: the 96-bit hash value is divided into 12 segments, each segment corresponds to an 8-bit hash code output by a hash function family; e =[h1(H e ),h2(H e ),.....,h 12 (H e )],

[0070] H j =[h1(H j ),h2(H j ),.....,h 12 (H j )]; where h i (H) represents the ith byte, corresponding to the 8-bit output of the ith hash function family. Segment-by-segment comparison: Calculate the Hamming distance for each segment. For i = 1 to 12, calculate the Hamming distance. i =Hamming(h i (H e ),h i (H j )); each hamm iThe value range is [0, 8], which indicates the number of different bits in the two 8-bit hash codes. Segment similarity calculation: convert the Hamming distance of each segment into similarity. Each sim i The value range is [0, 1], indicating the similarity of the segment. Weighted combination: Calculate the average value of the similarities of the 12 segments as the overall similarity. Equivalent to This segmented calculation method enables the overall similarity to fully consider the similarity of the feature vectors in different random projection directions, providing a more comprehensive similarity measure.

[0071] The actual calculation uses bitwise optimization: an XOR operation is performed on the two hash codes to obtain the difference bit; hardware-level instructions (such as POPCNT) are used to quickly calculate the number of set bits; for hash code segments with b = 8 bits, a pre-computed lookup table is used to directly obtain the Hamming distance. This calculation method considers the combined results of multiple groups of hash functions, and each group of hash functions contributes equally to the similarity. The similarity range is [0, 1], with larger values ​​indicating more similarity between the two feature vectors.

[0072] The system adopts the nearest center point classification strategy with threshold: calculate the hash similarity between the object to be evaluated and all 15 center points; basic rule: classify the object into the category represented by the center point with the highest similarity; enhanced rule: if the highest similarity exceeds the threshold T high =0.75, then directly classify; if it is lower than the threshold T low =0.45, it is marked as "no clear match"; for intermediate situations, the second highest similarity is combined for auxiliary judgment; Diversity considerations: For multiple center points with similarity differences less than 0.1, the system records all possibilities and provides diversified recommendations.

[0073] Based on the classification results, the system implements the following human resource management strategies: analyzing the matching success rates of various categories of talents with different positions in historical data; establishing a correlation matrix between center point categories and position requirements; and recommending 3-5 positions with the highest matching degree based on the category of the object to be evaluated.

Claims

1. A human resources data management method based on big data, characterized in that: include: S1, obtain the sample resume dataset and the resume of the subject to be evaluated; S2, extracts gender and age from the sample resume dataset as bias features; S3, semantically encode the sample resume dataset to obtain a set of semantic encoding feature vectors of the resume; S4, based on the bias characteristics, the dimensionality reduction processing is performed on the semantic encoding feature vector set of the resume to obtain the feature vector set S with reduced bias influence; S5, cluster analysis is performed on the feature vector set S that reduces the influence of bias, and multiple feature vector center points representing different feature combinations are obtained; S6, using an improved hash algorithm to semantically encode the resume of the subject to be evaluated, and obtaining a hash value of the subject to be evaluated; S7, calculating the similarity between the hash value of the object to be evaluated and the center point of each feature vector, and performing human resource data management based on the similarity; The feature vector set S that reduces the influence of bias is obtained, including: Calculate the correlation coefficients between each dimension in the semantic encoding feature vector set of the resume and the gender feature and age feature respectively; Dimensions whose absolute values ​​of correlation coefficients are greater than a preset threshold are marked as bias-sensitive dimensions, and a set of gender bias dimensions and a set of age bias dimensions are obtained; Calculate the attenuation coefficients λ1 and λ2 of the gender bias dimension set and the age bias dimension set respectively; Generate a bias adjustment weight matrix based on the absolute value of the correlation coefficient and the attenuation coefficients λ1 and λ2 of each dimension; According to the bias adjustment weight matrix, the dimensionality reduction process is performed on the semantic encoding feature vector set of the resume to obtain the feature vector set S with reduced bias influence; Calculate the attenuation coefficients λ1 and λ2 of the gender bias dimension set and the age bias dimension set respectively, including: Among them, μ1 and μ2 represent the average values ​​of the absolute values ​​of the correlation coefficients in the gender bias dimension set and the age bias dimension set, σ1 and σ2 represent the standard deviations of the absolute values ​​of the correlation coefficients in the gender bias dimension set and the age bias dimension set; c1 represents the gender bias suppression parameter, c2 represents the age bias suppression parameter, and b1 and b2 are adjustment parameters; An improved hash algorithm is used to semantically encode the resume of the person to be evaluated, including: Construct m independent hash function families {h1,h2,.....,h m }, where each hash function family h l represents a set of functions that map feature vectors to b-bit binary hash codes; Using m independent hash function families {h1,h2,.....,h m Perform feature mapping on the resume of the person to be evaluated and obtain m groups of b-bit binary hash codes; Combine m groups of b-bit binary hash codes in sequence as the hash value of the object to be evaluated.

2. The human resources data management method based on big data according to claim 1, characterized in that: S5, obtain multiple feature vector center points representing different feature combinations, including: Set the optimal number of clusters N, where N is a positive integer greater than 3; Calculate the local density value ρ(v) of each eigenvector v in the eigenvector set S. The local density value ρ(v) is defined as the number of eigenvectors in the R-neighborhood of the eigenvector v, where the R-neighborhood refers to the area with a Euclidean distance less than the radius R; Randomly select a feature vector v1 from the feature vector set S as the first initial center point, and add v1 to the center point set C; For each eigenvector v in the eigenvector set S that is not selected as the center point i , calculate v i The distance D to the nearest center point in the center point set C i , and calculate the eigenvector v i The local density value ρ(v i ); According to the distance D i and the local density value ρ(v i ) Calculate each unselected eigenvector v i The probability of selection {P i }; According to the selection probability {P i }, select the feature vector v with the highest selection probability max as the next center point and v max Add to the center point set C; If the newly selected center point v max If the distance from any selected center point is less than the preset minimum distance threshold MinDist, then v is abandoned. max , and select the feature vector with the second highest selection probability from the remaining unselected feature vectors as the center point; Repeat until the center point set C contains N initial cluster center points.

3. The human resources data management method based on big data according to claim 2, characterized in that: Among them, α and β are weight coefficients used to balance the influence of distance factor and density factor, γ is the outlier penalty coefficient, O i is an outlier indicator, when the eigenvector v i When the average distance between its K nearest neighbors exceeds the preset multiple of the global average distance, O i The value is 1, otherwise it is 0.

4. The human resources data management method based on big data according to claim 1, characterized in that: Construct a hash function family h l ,include: Randomly generate b d-dimensional unit vectors {r1,r2,.....,r d }, where d represents the dimension of the feature vector; For the feature vector v of the object to be evaluated e , calculate v e With each random unit vector r k The inner product of k =v e ×r k ; When the projection value p k When it is greater than or equal to zero, the corresponding hash bit is set to 1, otherwise it is set to 0; Combine b hash bits in sequence to form a b-bit binary hash code as the hash function family h l .

5. The human resources data management method based on big data according to claim 1, characterized in that: S7, calculating the similarity between the hash value of the object to be evaluated and the center point of each feature vector, including: Using the improved hash algorithm in step S6 to perform semantic encoding on the center point of each feature vector to obtain the hash value of the center point of each feature vector; Calculate the similarity Sim(H) between the hash value of the object to be evaluated and the hash value of the center point of each feature vector e ,H j ); where H e Represents the hash value of the object to be evaluated, H j Represents the hash value of the center point of the j-th eigenvector; Classify the object to be evaluated according to the calculated similarity, and classify the object to be evaluated into the category corresponding to the center point of the feature vector with the highest similarity; Based on the classification results, human resource management is performed on the subject to be evaluated, including talent matching and job recommendation.

6. The human resources data management method based on big data according to claim 5, characterized in that: Calculate the similarity Sim(H) between the hash value of the object to be evaluated and the hash value of the center point of each feature vector e ,H j ), using the following formula: Among them, h q (*) represents the hash code part generated by the qth hash function family, Hamming(*) represents the Hamming distance, and b represents the number of hash code bits generated by each hash function family.

Images