A kind of salary data automatic checking method based on multidimensional feature modeling

Abstract

The application discloses a kind of based on multi-dimensional feature modeling's salary data automatic checking method, including the following steps: obtaining enterprise salary management related data and pre-processing, obtain pre-processing dataset;Multi-dimensional feature set is constructed based on pre-processing dataset;Sample contrast grouping key is generated based on multi-dimensional feature set, contrast sample set is formed based on sample contrast grouping key, feature coding is carried out to multi-dimensional feature set and is converted into input sequence;Attention constraint information is generated based on multi-dimensional feature set and pre-processing dataset;Input sequence, attention constraint information, sample contrast grouping key and contrast sample set are input into improved SAINT network, and generate checking result information;Abnormal score in checking result information and threshold strategy are used to complete abnormal determination and output checking result.The application uses improved SAINT network, realizes salary data automatic checking.

Description

Technical Field

[0001] This invention relates to the field of enterprise human resource management and salary data processing technology, and in particular to an automatic verification method for salary data based on multidimensional feature modeling. Background Technology

[0002] As enterprises improve their IT infrastructure, their human resource management systems are gradually centralizing payroll data management. In the payroll management process, enterprises need to handle detailed payroll data, basic employee information, job and grade information, organizational hierarchy information, and regional and currency-related information. This data is typically stored and processed through the payroll management system.

[0003] To ensure the accuracy of payroll payments, existing technologies typically use rule-based validation or statistical analysis to detect anomalies in payroll data. For example, fixed rules are used to determine the range of payroll fields, and simple statistical methods are used to identify data that deviates from the average level, thereby enabling the validation of payroll data.

[0004] However, existing technologies have certain limitations when dealing with complex salary structures and multidimensional data relationships. Traditional rule-based validation methods rely on preset rules, making it difficult to simultaneously conduct correlation analysis by comprehensively considering multidimensional information such as employee attributes, job levels, organizational structure, and historical cyclical changes. Statistical methods typically identify anomalies based on the overall data distribution, making it difficult to combine job groups, job levels, and regional conditions to form effective control sample relationships, resulting in limited ability to identify abnormal data in complex salary structure scenarios.

[0005] Therefore, how to provide an automatic verification method for salary data based on multidimensional feature modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an automatic verification method for salary data based on multidimensional feature modeling. This invention preprocesses enterprise salary management data and constructs a multidimensional feature set including employee attribute features, job and grade features, organizational structure features, salary structure features, and historical cycle change features. Combining a sample comparison grouping mechanism and attention constraint information, an improved SAINT network is introduced to automatically verify salary data and generate anomaly scores and verification evidence chains. This invention can comprehensively analyze salary data based on multidimensional feature correlations, achieving automated verification in complex salary structures and multi-condition comparison scenarios. It has the advantages of high automation, full utilization of multidimensional information, and high accuracy in anomaly identification.

[0007] An automatic verification method for salary data based on multidimensional feature modeling according to an embodiment of the present invention includes the following steps: Acquire relevant data on enterprise payroll management and preprocess it to obtain a preprocessed dataset; Construct a multidimensional feature set based on the preprocessed dataset; A sample control grouping key is generated based on a multidimensional feature set, a control sample set is formed based on the sample control grouping key, and the multidimensional feature set is feature-encoded and converted into an input sequence. Attention constraint information is generated based on a multidimensional feature set and a preprocessed dataset. The input sequence, attention constraint information, sample control grouping key, and control sample set are input into the improved SAINT network. The network generates verification result information through the constraint attention module, the attention structure between grouped samples, and the structure and time enhancement module. Anomaly detection and verification results are output based on the anomaly score and threshold strategy in the verification result information.

[0008] Optionally, the preprocessing includes field unification processing, format standardization processing, and data cleaning processing.

[0009] Optionally, constructing a multidimensional feature set based on the preprocessed dataset specifically includes: Based on the preprocessed dataset, the employee basic information field, job position and rank field, and organizational hierarchy field are extracted, and feature mapping processing is performed on the fields to construct employee attribute feature vectors, job position and rank feature vectors, and organizational structure feature vectors. Construct a salary structure feature vector based on the salary details data in the preprocessed dataset; Historical cycle change features are constructed based on the salary cycle field in the preprocessed dataset and historical salary data records. Construct caliber consistency features based on currency label vectors and caliber label vectors in the preprocessed dataset; By combining employee attribute feature vectors, job and rank feature vectors, organizational structure feature vectors, salary structure feature vectors, historical cycle change features, and consistency features, a multi-dimensional feature set is constructed.

[0010] Optionally, the step of generating a sample-control grouping key based on a multidimensional feature set, forming a control sample set based on the sample-control grouping key, and performing feature encoding on the multidimensional feature set and converting it into an input sequence specifically includes: Based on a multidimensional feature set, we extract job and rank features, organizational structure features, and consistency features, and combine currency labeling vectors and caliber labeling vectors to construct a sample comparison grouping field set. A sample control grouping key is generated based on the sample control grouping field set. The salary data records in the multidimensional feature set are grouped based on the sample control grouping key to form a control sample set. The various features in the multidimensional feature set are subjected to feature encoding processing to obtain a set of encoded feature vectors; The set of encoded feature vectors is concatenated and combined according to a preset feature order to construct the input sequence.

[0011] Optionally, the generation of attention constraint information based on the multidimensional feature set and the preprocessed dataset specifically includes: Based on the caliber consistency feature in the multidimensional feature set, as well as the missing value label matrix, field activation status matrix, currency label vector, and caliber label vector, a set of constraint fields is constructed. Generate an attention mask based on the missing value labeling matrix and the field-enabled state matrix; An attention bias is generated based on caliber consistency features, currency label vectors, and caliber label vectors. Attention mask and attention bias are combined to form attention constraint information.

[0012] Optionally, the step of inputting the input sequence, attention constraint information, sample-control grouping key, and control sample set into the improved SAINT network, and generating verification result information through the constraint attention module, the attention structure between grouped samples, and the structure and time enhancement module specifically includes: The input sequence is fed into the embedding encoding module of the improved SAINT network, and heterogeneous embedding processing is performed on each feature field in the input sequence to obtain a feature embedding vector sequence. A classification label vector is set in the feature embedding vector sequence, and the classification label vector is concatenated with the feature embedding vector sequence to obtain the feature label sequence; The feature label sequence and attention constraint information are input into the constraint attention module to obtain the feature label sequence after constraint attention processing. The feature label sequence after constrained attention processing is input into the backbone feature interaction module for feature interaction calculation. The backbone feature interaction module is composed of multiple cascaded network stages, and the output feature label sequence of the backbone feature interaction module is obtained. The attention structure between grouped samples introduced in the backbone feature interaction module selects samples within the control sample set based on the sample control grouping key and introduces a control reliability gating mechanism to obtain the interaction representation between samples after adjustment by the control reliability gating mechanism. A structure and time enhancement module is introduced into the backbone feature interaction module to obtain the fused classification label vector based on the output feature label sequence; The fused classification label vector is input into the multi-head output module to generate verification result information.

[0013] Optionally, the step of inputting the feature label sequence and attention constraint information into the constraint attention module to obtain the feature label sequence after constraint attention processing specifically includes: The feature label sequence and attention constraint information are input into the constraint attention module. In the constraint attention module, the query matrix, key matrix and value matrix are generated based on the feature label sequence. The attention score matrix is ​​calculated based on the query matrix and the key matrix, and an attention bias is introduced into the attention score matrix to obtain the attention score matrix after bias is introduced; The attention score matrix after bias is masked according to the attention mask to obtain the masked attention score matrix. The attention score matrix after masking is normalized to obtain the attention weight matrix; The value matrix is ​​weighted and summed based on the attention weight matrix to obtain the feature representation matrix after constrained attention processing, which is then output as the feature label sequence after constrained attention processing.

[0014] Optionally, the inter-sample attention structure introduced in the backbone feature interaction module, which selects samples within the control sample set based on the sample-control grouping key and introduces a control reliability gating mechanism to obtain the inter-sample interaction representation adjusted by the control reliability gating mechanism, specifically includes: In the attention structure between grouped samples in the main feature interaction module, the control sample set is retrieved and filtered based on the sample control grouping key to form the current control group; For each control sample in the current control group, obtain the corresponding feature label sequence after constrained attention processing, and perform a flattening operation to obtain the flattened representation vector; Attention calculation is performed on the flattened representation vector in the sample dimension to obtain the attention weights between samples. Then, the flattened representation vector is weighted and converged based on the attention weights between samples to obtain the interaction representation between samples. A control reliability gating mechanism is introduced. Based on the missing value label matrix, caliber consistency features, and historical periodic change features, the control sample reliability of each control sample in the current control group is calculated to obtain the control sample reliability set. The key matrix and value matrix in the inter-sample attention calculation are weighted and adjusted based on the set of confidence of the control samples. Based on the weighted key matrix and the weighted value matrix, the attention calculation for the sample dimension is re-performed to obtain the inter-sample interaction representation adjusted by the comparison reliability gating mechanism.

[0015] Optionally, the step of introducing a structure and temporal enhancement module into the backbone feature interaction module to obtain the fused classification label vector based on the output feature label sequence specifically includes: A structure and time enhancement module is introduced into the main feature interaction module to extract salary structure feature label subsequence and historical change feature label subsequence from the output feature label sequence of the main feature interaction module; Perform compositional enhancement processing on the subsequence of salary structure feature markers to obtain a structure-enhanced representation of the salary structure. Time augmentation processing is performed on the historical change feature marker subsequence to obtain a time-enhanced representation of historical change features; The time-enhanced historical change feature representation is fused with the classification label vector to obtain the fused classification label vector.

[0016] Optionally, the step of determining anomalies and outputting verification results based on the anomaly score and anomaly determination threshold in the verification result information specifically includes: Obtain the verification result information, extract the anomaly score from the verification result information, and determine the anomaly judgment threshold; The calculation is based on comparing the abnormal score with the abnormal judgment threshold. When the abnormal score is greater than or equal to the abnormal judgment threshold, the corresponding salary data record is judged as an abnormal sample. When the anomaly score is less than the anomaly detection threshold, the corresponding salary data record will be judged as a normal sample. Based on the anomaly determination result, the corresponding anomaly score, the anomaly type, and the verification evidence chain, a verification result is generated.

[0017] The beneficial effects of this invention are: This invention cleanses and standardizes enterprise salary management data to generate missing value markers, data source markers, currency markers, pre-tax and post-tax caliber markers, and field activation status markers. Based on this, it constructs a multi-dimensional feature set that includes employee attribute features, job and grade features, organizational structure features, salary structure features, historical cycle change features, and caliber consistency features. This allows salary data to form a structured input that can be used for feature interaction calculations before entering the verification stage, providing a unified data expression basis for subsequent automatic verification.

[0018] This invention generates sample comparison grouping keys based on multidimensional feature sets, and forms a comparison sample set from the sample comparison grouping keys, enabling salary data to build comparable comparison relationships in the dimensions of job groups, job levels, and regions or currencies. At the same time, attention constraint information is generated based on caliber consistency features and labeling information, introducing the restrictions of missing or disabled fields and the association enhancement conditions of caliber consistency fields into the feature interaction calculation process, so that the data missing state, field enabled state and caliber consistency information can participate in model calculation and affect the feature interaction results.

[0019] This invention inputs the input sequence, attention constraint information, sample control grouping key, and control sample set into an improved SAINT network. Through the constraint attention module, the intra-column self-attention structure and inter-group sample attention structure in the backbone feature interaction module, the control reliability gating mechanism, and the structure and time enhancement module, it achieves joint modeling of the consistency relationship of salary structure and the relationship of historical cycle changes, and outputs verification result information including anomaly scores. It can also output a verification evidence chain including the contribution field set and the control evidence set, thereby completing the automatic anomaly detection and traceable verification result output of salary data. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an automatic verification method for salary data based on multidimensional feature modeling proposed in this invention; Figure 2 This is a schematic diagram of the improved SAINT network structure in the automatic verification method for salary data based on multidimensional feature modeling proposed in this invention. Figure 3 This is a schematic diagram of the backbone feature interaction module and the attention structure between grouped samples in the automatic verification method for salary data based on multidimensional feature modeling proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 An automatic verification method for salary data based on multidimensional feature modeling includes the following steps: Acquire relevant data on enterprise payroll management and preprocess it to obtain a preprocessed dataset; Construct a multidimensional feature set based on the preprocessed dataset; A sample control grouping key is generated based on a multidimensional feature set, a control sample set is formed based on the sample control grouping key, and the multidimensional feature set is feature-encoded and converted into an input sequence. Attention constraint information is generated based on a multidimensional feature set and a preprocessed dataset. The input sequence, attention constraint information, sample control grouping key, and control sample set are input into the improved SAINT network. The network generates verification result information through the constraint attention module, the attention structure between grouped samples, and the structure and time enhancement module. Anomaly detection and verification results are output based on the anomaly score and threshold strategy in the verification result information.

[0023] In this embodiment, acquiring and preprocessing enterprise payroll management-related data to obtain a preprocessed dataset specifically includes: Obtain the raw salary data set from the enterprise's payroll management system. The raw salary data set consists of multiple salary data records. The salary data records include at least salary details, employee basic information, job and grade information, organizational hierarchy information, and information on region, currency, and pre- and post-tax caliber. The payroll data records are processed for unified field processing and format standardization. The unified field processing includes mapping field names from different source systems to unified field names, unifying the naming of fields with the same meaning but different names from different source systems, and organizing each field according to a unified field structure. The format standardization process includes standardizing the numerical format of fields related to salary amount in salary data records, standardizing the numerical representation of the amount corresponding to the currency field in salary data records, standardizing the date format of the period or date field in salary data records, and standardizing the encoding format of character fields in salary data records. A standardized set of salary records is obtained through unified field processing and format standardization. Data cleaning is performed on a standardized payroll record set. This cleaning process includes missing value detection, outlier detection, and format consistency detection for each field in the standardized payroll record set. Based on this, missing fields, outlier fields, and fields with format errors are identified in the standardized payroll record set. A missing value labeling matrix is ​​then generated based on the missing fields. The missing value labeling matrix consists of multiple label values, which are used to indicate whether a certain field in a payroll data record is missing. When a field is missing, the corresponding label value is recorded as one, and when the field is not missing, the corresponding label value is recorded as zero. A data source tag vector is generated based on the data source field in the standardized payroll record set. The data source tag vector consists of multiple data source tag values, which are used to represent the data source system identifier of the corresponding payroll data record in the standardized payroll record set. For each pay data record in the standardized pay record set, read the data source field and write the read data source system identifier as the corresponding data source tag value into the data source tag vector; A currency tag vector is generated based on the currency field in the standardized payroll record set. The currency tag vector consists of multiple currency tag values, which are used to represent the currency type of the corresponding payroll data record in the standardized payroll record set. For each payroll data record in the standardized payroll record set, read the currency field and write the read currency type as the corresponding currency tag value into the currency tag vector to obtain the currency tag vector; A caliber marker vector is generated based on the pre-tax and post-tax caliber fields in the standardized payroll record set. The caliber marker vector consists of multiple caliber marker values, which are used to represent the pre-tax or post-tax caliber type of the corresponding payroll data record in the standardized payroll record set. For each salary data record in the standardized salary record set, read the pre-tax or post-tax caliber field, and write the read pre-tax or post-tax caliber type as the corresponding caliber tag value into the caliber tag vector to obtain the caliber tag vector. A field activation status matrix is ​​generated based on the activation status of each field in a standardized payroll record set. The field activation status matrix consists of the field activation status of multiple payroll data records and is used to indicate whether each field in each payroll data record is in an activated state. For each pay data record in the standardized pay record set, read the activation status information of each field and arrange them according to the correspondence between pay data record number and field number to form a field activation status matrix; In the field enable / disable state matrix, when the corresponding field is enabled, the corresponding position is marked as one; when the corresponding field is disabled, the corresponding position is marked as zero. The standardized payroll record set, missing value labeling matrix, data source labeling vector, currency labeling vector, caliber labeling vector, and field enable status matrix are combined to form a preprocessed dataset.

[0024] In this embodiment, constructing a multidimensional feature set based on the preprocessed dataset specifically includes: Based on the preprocessed dataset, the employee basic information field, job title and grade field, and organizational level field are extracted from each salary data record. Feature mapping processing is then performed on the fields to construct employee attribute feature vectors, job title and grade feature vectors, and organizational structure feature vectors. The employee attribute feature vector, job title and grade feature vector, and organizational structure feature vector are all formed by arranging the corresponding field values ​​in the field order. A salary structure feature vector is constructed based on the salary details data in the preprocessed dataset. The salary structure feature vector is formed by arranging the monetary values ​​of multiple salary component fields in sequence. The salary component fields include at least the basic salary field, performance salary field, subsidy field, bonus field, and other salary component fields. For each salary data record in the preprocessed dataset, the monetary values ​​of each salary component field are read sequentially and arranged to form a salary structure feature vector; Historical cycle change features are constructed based on the salary cycle field in the preprocessed dataset and historical salary data records. The historical cycle change features are calculated by difference. For each payroll data record in the preprocessed dataset, determine the corresponding current payroll period, and retrieve the historical payroll data records of the same employee in the previous payroll period from the preprocessed dataset. After obtaining the salary amount for the current pay period and the salary amount for the previous pay period, the historical period change characteristics are calculated by subtracting the salary amount for the previous pay period from the current pay period's salary amount. Based on the currency label vector and caliber label vector in the preprocessed dataset, a caliber consistency feature is constructed. The caliber consistency feature consists of multiple caliber consistency feature terms, which are used to represent the currency information and pre-tax and post-tax caliber information corresponding to the same salary data record. For each payroll data record in the standardized payroll record set, the currency tag value in the currency tag vector and the caliber tag value in the caliber tag vector are read respectively, and the currency tag value and the caliber tag value are combined to form the caliber consistency feature item corresponding to the payroll data record; By combining employee attribute feature vectors, job and rank feature vectors, organizational structure feature vectors, salary structure feature vectors, historical cycle change features, and consistency features, a multi-dimensional feature set is constructed.

[0025] In this embodiment, generating a sample-control grouping key based on a multidimensional feature set, forming a control sample set based on the sample-control grouping key, and performing feature encoding on the multidimensional feature set and converting it into an input sequence specifically includes: Based on the multidimensional feature set, the job and rank features, organizational structure features, and consistency features are extracted. Combined with currency label vectors and caliber label vectors, a sample comparison grouping field set is constructed. A sample comparison grouping key is generated based on the sample comparison grouping field set, which includes a job family field, a job grade field, and a region or currency field. For each field in the sample comparison grouping field set, the source of its corresponding value in the multidimensional feature set is determined. For each salary data record in the multidimensional feature set, the corresponding field value for the salary data record is extracted from the job family field, job grade field, and region or currency field included in the sample comparison grouping field set. The extracted field values ​​are then concatenated and combined in the order of job family field value - job grade field value - region field value or currency field value to generate the sample comparison grouping key corresponding to the salary data record. Based on the sample control grouping key, the salary data records in the multidimensional feature set are grouped to form a control sample set; Specifically, for each salary data record in the multidimensional feature set, the corresponding sample control grouping key is obtained, and salary data records with the same sample control grouping key are grouped into the same group. The salary data records contained in each group are then collected to form a control sample set consisting of multiple groups. The various features in the multidimensional feature set are processed by feature encoding, and the employee attribute features, job and grade features, organizational structure features, salary structure feature vectors, historical cycle change features and consistency features are converted into feature vectors of a unified dimension to obtain a set of encoded feature vectors. The set of encoded feature vectors is concatenated and combined according to a preset feature order to construct the input sequence; The preset feature order includes at least the coded feature vectors corresponding to employee attribute features, the coded feature vectors corresponding to job position and rank features, the coded feature vectors corresponding to organizational structure features, the coded feature vectors corresponding to salary structure features, the coded feature vectors corresponding to historical cycle change features, and the coded feature vectors corresponding to consistency features. The encoded feature vectors belonging to the above-mentioned feature categories in the encoded feature vector set are arranged and concatenated in a preset feature order to form an input sequence composed of multiple feature encoding results in sequence.

[0026] In this embodiment, generating attention constraint information based on a multidimensional feature set and a preprocessed dataset specifically includes: Based on the caliber consistency feature in the multidimensional feature set, as well as the missing value label matrix, field enable state matrix, currency label vector, and caliber label vector, a set of constraint fields for generating attention constraint information is constructed. An attention mask is generated based on the missing value label matrix and the field enable state matrix. The attention mask consists of multiple attention mask values, which indicate whether the corresponding feature field of the corresponding salary data record is allowed to participate in feature interaction calculation. Each corresponding position in the missing value label matrix and the field enable state matrix is ​​judged one by one. When the corresponding feature field in the missing value label matrix is ​​in a missing state, or the corresponding feature field in the field enable state matrix is ​​in an inactive state, the attention mask value of the corresponding position in the attention mask is set to zero. When the corresponding feature field in the missing value label matrix is ​​in a non-missing state and the corresponding feature field in the field enable state matrix is ​​in an enabled state, the attention mask value at the corresponding position in the attention mask is set to one, thereby obtaining the attention mask; Attention biases are generated based on caliber consistency features, currency label vectors, and caliber label vectors. The attention biases consist of multiple attention bias values, which are used to enhance the feature interaction calculations between caliber consistency feature fields. For any two feature fields in the input sequence, obtain the currency label value and the caliber label value corresponding to the two feature fields respectively, and determine whether the currency label value and the caliber label value of the two feature fields are consistent. When the currency label values ​​of two feature fields are the same and the caliber label values ​​of two feature fields are the same, the attention bias value corresponding to the two feature fields in the attention bias is set to a positive bias value. When the currency label values ​​of two feature fields are inconsistent or the caliber label values ​​of two feature fields are inconsistent, the attention bias value corresponding to the two feature fields in the attention bias is set to zero to obtain the attention bias. Attention mask and attention bias are combined to form attention constraint information.

[0027] In this embodiment, the input sequence, attention constraint information, sample control grouping key, and control sample set are input into the improved SAINT network. The network generates verification result information through a constraint attention module, an attention structure between grouped samples, and a structure and time enhancement module. Specifically, this includes: The input sequence is fed into the embedding encoding module of the improved SAINT network, and heterogeneous embedding processing is performed on each feature field in the input sequence to obtain a feature embedding vector sequence. The input sequence is formed by arranging the input values ​​corresponding to multiple feature fields in order, and the feature embedding vector sequence is formed by arranging multiple feature embedding vectors in the same order as the input sequence. For each feature field in the input sequence, if the feature field is a categorical feature field, then vectorization is performed by setting an embedding vector mapping function for each feature field, so that the feature embedding vector corresponding to the feature field is equal to the result calculated by the corresponding embedding vector mapping function after the input value of the feature field is processed. If the feature field is a continuous feature field, then the vectorization process is carried out by setting the linear projection matrix and the bias term separately for each feature field. The feature embedding vector corresponding to the feature field is equal to the product of the input value of the feature field and the corresponding linear projection matrix, plus the corresponding bias term. The feature embedding vectors corresponding to each feature field obtained through the above processing are arranged sequentially according to the feature field order of the input sequence, thus obtaining the feature embedding vector sequence; A classification label vector is set in the feature embedding vector sequence, and the classification label vector is concatenated with the feature embedding vector sequence to obtain the feature label sequence; The feature label sequence is composed of the classification label vector and each feature embedding vector in the feature embedding vector sequence in sequence, and the concatenation method is to place the classification label vector at the beginning of the feature embedding vector sequence; The feature labels in the feature label sequence are categorized by type. Feature labels corresponding to employee attribute features, job and grade features, and organizational structure features are categorized into context feature label subsequences. Feature labels corresponding to salary structure feature vectors are categorized into salary structure feature label subsequences. Feature labels corresponding to missing value label matrices, data source label vectors, currency label vectors, and caliber label vectors are categorized into meta-information feature label subsequences. Feature labels corresponding to historical cycle change features are categorized into historical change feature label subsequences. Thus, the feature label sequence and each feature label subsequence are obtained. The feature label sequence and attention constraint information are input into the constraint attention module to obtain the feature label sequence after constraint attention processing. The feature label sequence after constrained attention processing is input into the backbone feature interaction module for feature interaction calculation. The backbone feature interaction module is composed of multiple cascaded network stages. For each network stage in the backbone feature interaction module, the feature label sequence output by the previous network stage is used as the input feature label sequence of the network stage. Under the constraints of attention mask and attention bias, the in-column self-attention structure calculation is first performed to obtain the interaction representation between each feature label within a single sample. The output of the in-column self-attention structure is used as input, and combined with the sample control grouping key and the control sample set, attention calculation between samples within the group is performed under the group sample attention structure introduced in the backbone feature interaction module to obtain the interaction representation between samples within the same control group. The output of the attention structure between grouped samples is used as the output feature label sequence of the network stage, and then used as the input feature label sequence of the next network stage. Repeat the above process until the calculation of all network stages in the backbone feature interaction module is completed, thereby obtaining the output feature label sequence of the backbone feature interaction module; The attention structure between grouped samples introduced in the backbone feature interaction module selects samples within the control sample set based on the sample control grouping key and introduces a control reliability gating mechanism to obtain the interaction representation between samples after adjustment by the control reliability gating mechanism. A structure and time enhancement module is introduced into the backbone feature interaction module to obtain the fused classification label vector based on the output feature label sequence; The fused classification label vector is input into the multi-head output module to generate verification result information; The multi-head output module includes at least an anomaly score output head, an anomaly type output head, and an evidence chain output head. The anomaly score output head calculates the anomaly score from the fused classification label vector, the anomaly type output head calculates the anomaly type from the fused classification label vector, and the evidence chain output head calculates the verification evidence chain from the fused classification label vector. The verification evidence chain includes at least a contribution field set and a comparison evidence set to obtain verification result information.

[0028] In this embodiment, the feature label sequence and attention constraint information are input into the constraint attention module to obtain the feature label sequence after constraint attention processing, specifically including: The feature label sequence and attention constraint information are input into the constraint attention module. In the constraint attention module, the query matrix, key matrix and value matrix are generated based on the feature label sequence. The query matrix represents the query representation of the feature tag sequence, the key matrix represents the key representation of the feature tag sequence, and the value matrix represents the value representation of the feature tag sequence. The attention score matrix is ​​calculated based on the query matrix and the key matrix, and an attention bias is introduced into the attention score matrix to obtain the attention score matrix after bias is introduced; The attention score matrix after introducing the bias is calculated by multiplying the transpose of the query matrix and the key matrix to obtain the product result, dividing the product result by the square root of the attention calculation dimension to obtain the normalized product result, and then summing the normalized product result with the attention bias to obtain the attention score matrix after introducing the bias. The attention score matrix after bias is masked according to the attention mask, so that the attention scores corresponding to the positions set to zero in the attention mask are masked before normalization, and the masked attention score matrix is ​​obtained. The masking process involves judging and assigning values ​​to each position in the biased attention score matrix based on the attention mask. When the attention mask value of the corresponding position in the attention mask is zero, the attention score of the corresponding position in the biased attention score matrix is ​​set as the mask value, so that the position does not participate in the attention weight allocation in the subsequent normalization calculation. When the attention mask value at the corresponding position in the attention mask is one, the attention score at the corresponding position in the attention score matrix after introducing the bias remains unchanged. After completing the above judgment and assignment, the masked attention score matrix is ​​obtained. The attention score matrix after masking is normalized to obtain the attention weight matrix; The normalization calculation includes exponentializing the attention score in each row of the masked attention score matrix to obtain the corresponding exponentialized result, and summing all the exponentialized results in the same row to obtain the normalized denominator. Divide the indexed result of each attention score in the row by the normalized denominator to obtain the attention weight corresponding to each attention score in the row. Arrange the attention weights obtained from each row in their original positions to form an attention weight matrix. The value matrix is ​​weighted and summed based on the attention weight matrix to obtain the feature representation matrix after constrained attention processing, which is then output as the feature label sequence after constrained attention processing.

[0029] In this embodiment, the inter-sample attention structure introduced in the backbone feature interaction module selects samples within the control sample set based on the sample-control grouping key and introduces a control reliability gating mechanism to obtain the inter-sample interaction representation adjusted by the control reliability gating mechanism, specifically including: In the attention structure between grouped samples in the main feature interaction module, the control sample set is retrieved and filtered based on the sample control grouping key, and the control sample with the same sample control grouping key as the current salary data record is selected to form the current control group. For each control sample in the current control group, obtain the corresponding feature label sequence after constrained attention processing, and perform a flattening operation on the feature label sequence after constrained attention processing to obtain the flattened representation vector; The flattening operation includes sequentially concatenating each feature label in the constrained attention-processed feature label sequence into a one-dimensional vector according to a preset order to obtain the flattened representation vector. Attention calculation is performed on the flattened representation vectors of each control sample within the current control group in the sample dimension to obtain the inter-sample attention weights. Based on the inter-sample attention weights, the flattened representation vectors of each control sample within the current control group are weighted and converged to obtain the inter-sample interaction representation corresponding to the current salary data record. A control reliability gating mechanism is introduced. Based on the missing value label matrix, caliber consistency features, and historical periodic change features, the control sample reliability of each control sample in the current control group is calculated to obtain the control sample reliability set. The set of control sample confidence scores consists of multiple control sample confidence scores, which are used to represent the confidence level of the corresponding control sample. For each control sample in the current control group, the missing value, caliber consistency, and historical stability of the corresponding control sample are obtained, and the control sample confidence score of the control sample is determined based on the missing value, caliber consistency, and historical stability. Arrange the confidence scores of each control sample within the current control group according to the order of the control samples to form a set of control sample confidence scores. The key matrix and value matrix in the inter-sample attention calculation are weighted and adjusted based on the set of confidence of the control samples. Based on the weighted key matrix and the weighted value matrix, the attention calculation for the sample dimension is re-performed to obtain the inter-sample interaction representation adjusted by the control reliability gating mechanism; The weighted adjustment includes: for each control sample in the current control group, obtaining the control sample confidence score in the control sample confidence score set, and applying the control sample confidence score to the key matrix and value matrix corresponding to the control sample, so that the weighted adjusted key matrix is ​​equal to the product of the control sample confidence score and the key matrix before gating adjustment, and the weighted adjusted value matrix is ​​equal to the product of the control sample confidence score and the value matrix before gating adjustment. After weighting adjustment, the key matrix and value matrix before gating adjustment are replaced with the weighted key matrix and value matrix before gating adjustment. Attention weights are recalculated and weighted convergence is performed in the sample dimension to obtain the inter-sample interaction representation adjusted by the control reliability gating mechanism.

[0030] In this embodiment, a structure and temporal enhancement module is introduced into the backbone feature interaction module. Based on the output feature label sequence, the fused classification label vector is obtained, specifically including: A structure and time enhancement module is introduced into the main feature interaction module to extract salary structure feature label subsequence and historical change feature label subsequence from the output feature label sequence of the main feature interaction module; Perform compositional enhancement processing on the subsequence of salary structure feature markers to establish a consistency relationship between the total salary field and each salary item field; The composition structure enhancement process includes determining the total feature marker of the corresponding total salary field and the sub-item feature marker set of each salary sub-item field from the salary structure feature marker sub-sequence; The set of sub-item feature labels is aggregated to obtain sub-item aggregated feature labels. The aggregation calculation is the result of processing each sub-item feature label in the set of sub-item feature labels in at least one preset method. The preset method includes any one or a combination of vector concatenation, weighted summation, and linear mapping. The total amount feature label and the itemized aggregate feature label are fused and calculated to obtain the structure-enhanced salary structure representation. The structure-enhanced salary structure representation is the output result obtained by inputting the total amount feature label and the itemized aggregate feature label into the structure enhancement mapping. The structure enhancement mapping is any one or a combination of fully connected mapping, gated mapping or attention mapping. The historical change feature marker subsequence is subjected to time enhancement processing. The time enhancement processing includes generating corresponding time series codes for each historical change feature marker in the historical change feature marker subsequence based on historical periodic change features, and fusing the time series codes with the corresponding historical change feature markers to obtain the time-enhanced historical change feature representation. Among them, the time-enhanced historical change features are represented by the result of fusing the corresponding time series codes of each historical change feature label in the historical change feature label subsequence; The time-enhanced historical change feature representation is fused with the classification label vector to obtain the fused classification label vector; The fused classification label vector satisfies the following relationship: the fused classification label vector is equal to the output result obtained by the fusion function in fusing the classification label vector with the time-enhanced historical change feature representation; The fusion function is used to fuse the historical change feature representation and the classification label vector, and the classification label vector is the original classification label vector.

[0031] In this embodiment, the process of determining anomalies and outputting verification results based on the anomaly score and anomaly determination threshold in the verification result information specifically includes: Obtain the verification result information, extract the anomaly score from the verification result information, and determine the anomaly judgment threshold; The anomaly detection threshold is either a preset threshold or an adaptively updated threshold based on the sample control grouping key. When the preset threshold is used, the anomaly detection threshold is set as a constant threshold. When the adaptively updated threshold based on the sample control grouping key is used, the adaptive threshold is calculated based on the set of anomaly scores of control samples with the same sample control grouping key. The calculation is based on comparing the abnormal score with the abnormal judgment threshold. When the abnormal score is greater than or equal to the abnormal judgment threshold, the corresponding salary data record is judged as an abnormal sample. When the anomaly score is less than the anomaly detection threshold, the corresponding salary data record will be judged as a normal sample. Based on the anomaly determination result, the corresponding anomaly score, the anomaly type, and the verification evidence chain, a verification result is generated. The verification evidence chain is the evidence output information generated based on the anomaly score. It includes at least a set of contributing fields and a set of control evidence. The set of contributing fields is used to indicate the feature fields that contribute significantly to the anomaly score, and the set of control evidence is used to indicate the control sample information used for comparison.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an automatic salary data verification scenario in the human resource management system of a large enterprise. This enterprise centrally manages employee salary information in its human resource management system, which includes various data sources such as basic employee information, job and grade information, organizational structure information, and salary structure information. As the enterprise expands, the number of salary records to be processed in each salary cycle gradually increases. The salary structure includes multiple components such as basic salary, position allowance, performance bonus, and subsidies, and also contains salary data from different regions and in different currencies. In actual management, salary data verification mainly relies on rule-based verification and manual sampling. When the salary structure is complex or there are many data sources, manual verification is inefficient, and the ability to compare and analyze data within the same job group or grade level is limited, which easily leads to the problem of untimely identification of abnormal salary records.

[0033] In this scenario, the automatic salary data verification method based on multidimensional feature modeling of this invention is deployed in the salary verification module of a human resource management system. The system first acquires relevant salary management data from the enterprise and performs unified preprocessing. Salary details, employee basic information, job and grade information, and organizational hierarchy information are cleaned and formatted uniformly, generating missing value markers, currency markers, and pre- and post-tax caliber markers. Subsequently, a multidimensional feature set is constructed based on the preprocessed dataset, including employee attribute features, job and grade features, organizational structure features, salary structure features, and historical periodic change features. The system further generates sample comparison grouping keys based on job groups, grades, and regional or currency information, forming a comparison sample set under the same grouping key conditions, enabling salary data with the same job level and regional conditions to form a comparison relationship. After feature encoding is completed, the system inputs the input sequence, attention constraint information, and a set of control samples into the improved SAINT network. The system uses a constraint attention module to limit the impact of missing fields on feature interactions and performs comparative analysis of samples within the same group using the attention structure between grouped samples. Simultaneously, it combines a structure and time enhancement module to comprehensively model salary structure relationships and historical changes, thereby generating anomaly scores and verification evidence chains. The system automatically determines anomalies based on the anomaly scores and threshold strategies, and returns the anomaly records to the salary management interface for managers to review.

[0034] In actual operation, the company's payroll system processes 12,850 payroll records within a payroll cycle, including 14 categories of payroll component fields and 8 categories of employee attribute fields. Before applying this invention, the system mainly used fixed rules for payroll anomaly detection, with an overall verification processing time of approximately 146 minutes per payroll cycle. After rule verification, 436 suspected anomaly records required manual review, of which 96 were ultimately confirmed as anomalies. After applying the method of this invention, under the same data scale, the system automatically completes multidimensional feature analysis and sample comparison calculations through an improved SAINT network. The automatic verification processing time per payroll cycle is reduced to 52 minutes, and the system automatically identifies 102 anomaly records, of which 94 are confirmed as anomalies after review. Simultaneously, the system can output a set of contribution fields and a set of comparison evidence, enabling managers to quickly locate the causes of anomalies. For example, some records may significantly deviate from the average pay range of the same job group and grade, or show abnormal fluctuations in historical cycle trends. Actual operation results show that when the scale of salary data is large, the present invention can effectively complete multi-dimensional feature modeling and automated verification, and reduce the workload of manual review while ensuring the ability to identify anomalies.

[0035] To further illustrate the application effect of the present invention during implementation, key data during the implementation process were statistically analyzed, and the relevant data are shown in Table 1.

[0036] Table 1. Statistical Table of the Implementation Effect of Automatic Salary Data Verification

[0037] As can be seen from the table above, this invention demonstrates significant differences compared to traditional rule-based verification and manual sampling in several key indicators of automatic salary data verification. Regarding anomaly detection, this invention achieves an accuracy of 96.8%, precision of 94.6%, and recall of 93.1% on the same dataset, with a false positive rate of 3.2% and a false negative rate of 6.9%. In contrast, the rule-based verification in the comparison group achieves an accuracy of 88.4%, precision of 72.5%, recall of 61.8%, a false positive rate of 11.6%, and a false negative rate of 38.2%. These results indicate that by introducing a sample comparison grouping key to form a comparison sample set and combining it with attention constraint information, the model can establish a more stable benchmark within job groups, job levels, and regional or currency dimensions, significantly improving the coverage of identification for issues such as inconsistent definitions, historical fluctuation anomalies, and structural inconsistencies.

[0038] In terms of efficiency, the average time for single-batch verification in this invention is 8.7 minutes, a significant reduction compared to the 43.5 minutes in the control group. Manual review time decreased from 18.4 hours per batch to 5.6 hours, while the manual review pass rate increased from 62.0% to 85.7%, indicating that the anomaly score and verification evidence chain in the verification results provide clearer clues for anomaly localization. Regarding stability, the threshold adaptive strategy of this invention under different control grouping keys narrows the false alarm rate fluctuation range within a group to 2.6% to 3.8%, while the false alarm rate fluctuation of the control group at the grouping dimension is 8.9% to 14.7%. This reflects that the control reliability gating mechanism can reduce the impact of low-quality control samples on the judgment when there are differences in missing data, consistency of criteria, and historical stability. In summary, this invention reduces processing time and review costs while ensuring verification accuracy and improving the traceability of anomaly localization.

[0039] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic verification method for salary data based on multidimensional feature modeling, characterized in that, Includes the following steps: Acquire relevant data on enterprise payroll management and preprocess it to obtain a preprocessed dataset; Construct a multidimensional feature set based on the preprocessed dataset; A sample control grouping key is generated based on a multidimensional feature set, a control sample set is formed based on the sample control grouping key, and the multidimensional feature set is feature-encoded and converted into an input sequence. Attention constraint information is generated based on a multidimensional feature set and a preprocessed dataset. The input sequence, attention constraint information, sample control grouping key, and control sample set are input into the improved SAINT network. The network generates verification result information through the constraint attention module, the attention structure between grouped samples, and the structure and time enhancement module. The anomaly determination is completed based on the anomaly score and threshold strategy in the verification result information, and the verification result is output.

2. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The preprocessing includes field unification processing, format standardization processing, and data cleaning processing.

3. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The construction of a multidimensional feature set based on the preprocessed dataset specifically includes: Based on the preprocessed dataset, the employee basic information field, job title and rank field, and organizational hierarchy field are extracted, and feature mapping processing is performed on the fields to construct employee attribute feature vectors, job title and rank feature vectors, and organizational structure feature vectors. Construct a salary structure feature vector based on the salary details data in the preprocessed dataset; Historical cycle change features are constructed based on the salary cycle field in the preprocessed dataset and historical salary data records. Construct caliber consistency features based on currency tag vectors and caliber tag vectors in the preprocessed dataset; By combining employee attribute feature vectors, job and rank feature vectors, organizational structure feature vectors, salary structure feature vectors, historical cycle change features, and consistency features, a multi-dimensional feature set is constructed.

4. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The process of generating sample-control grouping keys based on multidimensional feature sets, forming control sample sets based on sample-control grouping keys, and encoding and converting the multidimensional feature sets into input sequences specifically includes: Based on a multidimensional feature set, we extract job and rank features, organizational structure features, and consistency features, and combine currency labeling vectors and caliber labeling vectors to construct a sample comparison grouping field set. A sample control grouping key is generated based on the sample control grouping field set. The salary data records in the multidimensional feature set are grouped based on the sample control grouping key to form a control sample set. The various features in the multidimensional feature set are subjected to feature encoding processing to obtain a set of encoded feature vectors; The set of encoded feature vectors is concatenated and combined according to a preset feature order to construct the input sequence.

5. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The generation of attention constraint information based on a multidimensional feature set and a preprocessed dataset specifically includes: Based on the caliber consistency feature in the multidimensional feature set, as well as the missing value label matrix, field activation status matrix, currency label vector, and caliber label vector, a set of constraint fields is constructed. Generate an attention mask based on the missing value labeling matrix and the field-enabled state matrix; An attention bias is generated based on caliber consistency features, currency label vectors, and caliber label vectors. Attention mask and attention bias are combined to form attention constraint information.

6. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The process of inputting the input sequence, attention constraint information, sample-to-sample grouping key, and control sample set into the improved SAINT network, and generating verification result information through the constraint attention module, the attention structure between grouped samples, and the structure and time enhancement module specifically includes: The input sequence is fed into the embedding encoding module of the improved SAINT network, and heterogeneous embedding processing is performed on each feature field in the input sequence to obtain a feature embedding vector sequence. A classification label vector is set in the feature embedding vector sequence, and the classification label vector is concatenated with the feature embedding vector sequence to obtain the feature label sequence; The feature label sequence and attention constraint information are input into the constraint attention module to obtain the feature label sequence after constraint attention processing. The feature label sequence after constrained attention processing is input into the backbone feature interaction module for feature interaction calculation. The backbone feature interaction module is composed of multiple cascaded network stages, and the output feature label sequence of the backbone feature interaction module is obtained. The attention structure between grouped samples introduced in the backbone feature interaction module selects samples within the control sample set based on the sample control grouping key and introduces a control reliability gating mechanism to obtain the interaction representation between samples after adjustment by the control reliability gating mechanism. A structure and time enhancement module is introduced into the backbone feature interaction module to obtain the fused classification label vector based on the output feature label sequence; The fused classification label vector is input into the multi-head output module to generate verification result information.

7. The automatic verification method for salary data based on multidimensional feature modeling according to claim 6, characterized in that, The step of inputting the feature label sequence and attention constraint information into the constraint attention module to obtain the feature label sequence after constraint attention processing specifically includes: The feature label sequence and attention constraint information are input into the constraint attention module. In the constraint attention module, the query matrix, key matrix and value matrix are generated based on the feature label sequence. The attention score matrix is ​​calculated based on the query matrix and the key matrix, and an attention bias is introduced into the attention score matrix to obtain the attention score matrix after bias is introduced; The attention score matrix after bias is masked according to the attention mask to obtain the masked attention score matrix. The attention score matrix after masking is normalized to obtain the attention weight matrix; The value matrix is ​​weighted and summed based on the attention weight matrix to obtain the feature representation matrix after constrained attention processing, which is then output as the feature label sequence after constrained attention processing.

8. The automatic verification method for salary data based on multidimensional feature modeling according to claim 6, characterized in that, The inter-sample attention structure introduced in the backbone feature interaction module, based on the sample-control grouping key, performs in-group sample selection on the control sample set and introduces a control reliability gating mechanism to obtain the inter-sample interaction representation adjusted by the control reliability gating mechanism, specifically includes: In the attention structure between grouped samples in the main feature interaction module, the control sample set is retrieved and filtered based on the sample control grouping key to form the current control group; For each control sample in the current control group, obtain the corresponding feature label sequence after constrained attention processing, and perform a flattening operation to obtain the flattened representation vector; Attention calculation is performed on the flattened representation vector in the sample dimension to obtain the attention weights between samples. Then, the flattened representation vector is weighted and converged based on the attention weights between samples to obtain the interaction representation between samples. A control reliability gating mechanism is introduced. Based on the missing value label matrix, caliber consistency features, and historical periodic change features, the control sample reliability of each control sample in the current control group is calculated to obtain the control sample reliability set. The key matrix and value matrix in the inter-sample attention calculation are weighted and adjusted based on the set of confidence of the control samples. Based on the weighted key matrix and the weighted value matrix, the attention calculation for the sample dimension is re-performed to obtain the inter-sample interaction representation adjusted by the comparison reliability gating mechanism.

9. The automatic verification method for salary data based on multidimensional feature modeling according to claim 6, characterized in that, The introduction of a structure and temporal enhancement module into the backbone feature interaction module, and the obtaining of the fused classification label vector based on the output feature label sequence, specifically includes: A structure and time enhancement module is introduced into the main feature interaction module to extract salary structure feature label subsequence and historical change feature label subsequence from the output feature label sequence of the main feature interaction module; Perform compositional enhancement processing on the subsequence of salary structure feature markers to obtain a structure-enhanced representation of the salary structure. Time augmentation processing is performed on the historical change feature marker subsequence to obtain a time-enhanced representation of historical change features; The time-enhanced historical change feature representation is fused with the classification label vector to obtain the fused classification label vector.

10. The automatic verification method for salary data based on multidimensional feature modeling according to claim 1, characterized in that, The process of determining anomalies and outputting verification results based on the anomaly score and anomaly determination threshold in the verification result information specifically includes: Obtain the verification result information, extract the anomaly score from the verification result information, and determine the anomaly judgment threshold; The calculation is based on comparing the abnormal score with the abnormal judgment threshold. When the abnormal score is greater than or equal to the abnormal judgment threshold, the corresponding salary data record is judged as an abnormal sample. When the anomaly score is less than the anomaly detection threshold, the corresponding salary data record will be judged as a normal sample. Based on the anomaly determination result, the corresponding anomaly score, the anomaly type, and the verification evidence chain, a verification result is generated.

Images