Intelligent account checking system and method based on multi-feature engineering

Abstract

The invention relates to the technical field of intelligent reconciliation, in particular to an intelligent reconciliation system and method based on multi-feature engineering, and the method comprises the steps: obtaining large sample data of a historical bill, carrying out the data preprocessing of the large sample data of the bill, extracting historical data features from the preprocessed large sample data, building a multi-layer weight network, and carrying out the data preprocessing of the historical data features; historical data features are fused into a comprehensive feature vector by using a multi-layer weight network, the comprehensive feature vector is trained, parameters of the weight network are updated through back propagation, a self-attention mechanism model containing a three-layer architecture and used for a Transform architecture is constructed, the self-attention mechanism model of the Transform architecture is trained by using the fused features, and the self-attention mechanism model of the three-layer architecture is constructed. The method has the advantages that multiple features of the bill are obtained, the bill can be matched in multiple aspects, the comprehensiveness of bill reconciliation is guaranteed, matching can be rapidly and accurately carried out along with the bill, abnormal records and unmatched records can be rapidly found out, and clients can conveniently find out problem bills.

Description

Technical Field

[0001] This invention relates to the field of intelligent reconciliation technology, specifically to an intelligent reconciliation system and method based on multi-feature engineering. Background Technology

[0002] Currently, businesses retain both external and internal invoices during transactions. These invoices need to be reconciled periodically to ensure the accuracy of internal funds. Existing reconciliation methods typically rely on manual processes, which are extremely tedious, inefficient, and prone to errors. Furthermore, reconciling errors after they occur is very troublesome. Some automated electronic reconciliation technologies cannot consider all the characteristics of invoices, leading to discrepancies during reconciliation. Therefore, we propose an intelligent reconciliation system and method based on multi-feature engineering. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an intelligent reconciliation method based on multi-feature engineering, including:

[0004] S1: Obtain large sample data of historical bills, preprocess the large sample data of bills, and extract historical data features from the preprocessed large sample data.

[0005] S2: Establish a multi-layer weighted network, use the multi-layer weighted network to fuse historical data features into a comprehensive feature vector, train the comprehensive feature vector, and update the parameters of the weighted network through backpropagation;

[0006] S3: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the Transformer architecture self-attention mechanism model with the fused features, and obtain a trained feature similarity calculation model.

[0007] S4: Calculate the bill similarity using the trained feature similarity calculation model, automatically match reconciliation records based on similarity, mark records with high similarity as successfully reconciled, and use anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

[0008] Furthermore, the historical data features include text features, numerical features, time series features, and image features; the text features are converted into numerical values ​​through natural language processing technology, the numerical features are features extracted directly from numerical fields such as amount and date, and the image features are QR codes, stamps, or handwritten signatures in bills extracted through convolutional neural networks.

[0009] Furthermore, the multi-layer weighted network includes: a network layer and a weighted fusion layer; the network layer includes a fully connected layer whose input and output are both feature numbers, the fully connected layer is followed by a normalization layer, and the weighted fusion layer.

[0010] Furthermore, a multi-layer weighted network is used to fuse historical data features into a comprehensive feature vector, including:

[0011] Obtain features from historical billing data, using the feature set of a single bill as input features;

[0012] Each feature is assigned a weight value through the network layers, and the weight values ​​of each feature set are normalized to ensure that the sum of the weights of each sample is 1.

[0013] The weighted feature summation formula is used to sum the features of each feature set, and the final result is used as the comprehensive feature vector.

[0014] The formula for weighted feature summation is:

[0015]

[0016] Where n is the feature, f is the feature vector, and w is the weight.

[0017] Furthermore, step 5, training the comprehensive feature vector, includes:

[0018] Obtain the comprehensive feature vector and divide the comprehensive feature vector into a training set and a validation set;

[0019] Iterate through the entire training set and set the number of samples to process each time;

[0020] The loss function is used to calculate the difference between the model's predicted output and the true label;

[0021] The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm;

[0022] The optimizer updates the parameters of the weight network based on the gradient.

[0023] Furthermore, the self-attention mechanism model of the Transformer architecture includes an embedding layer in the first layer, six Transformer encoder modules in the second layer, and an output layer in the third layer. The Transformer encoder consists of six repeating encoders, each of which contains a self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by normalization and residual connections.

[0024] Furthermore, obtaining the trained feature similarity calculation model includes:

[0025] Using the determined feature weights, calculate the comprehensive feature vector using the weighted feature summation formula, and establish a comprehensive feature vector set;

[0026] Initialize the three-layer Transformer architecture self-attention mechanism model, use the comprehensive feature vector set as the input of the self-attention mechanism model, train the self-attention mechanism model, and obtain the trained self-attention mechanism model.

[0027] Furthermore, the step of calculating bill similarity using the trained feature similarity calculation model includes:

[0028] Two sets of corresponding fused features are obtained, and the two sets of fused features are input into the Transformer model respectively to obtain two sets of feature representations;

[0029] Cosine similarity is used to measure the similarity between two sets of feature representations;

[0030] The formula for calculating similarity is:

[0031]

[0032] Where xi and yi are the i-th components of vectors x and y, respectively, and n is the dimension of the vector.

[0033] A multi-feature engineering intelligent reconciliation system includes:

[0034] Feature extraction module: Extracts historical data features from large sample data;

[0035] Feature fusion module: It uses a multi-layer weighted network to fuse historical data features into a comprehensive feature vector, trains the comprehensive feature vector, and updates the parameters of the weighted network through backpropagation;

[0036] Similarity calculation module: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the self-attention mechanism model of the Transformer architecture with the fused features, and obtain a trained feature similarity calculation model;

[0037] Reconciliation module: Calculates the similarity of model bills, matches reconciliation records, marks records with high similarity as successfully reconciled, and uses anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

[0038] The beneficial effects of this invention are reflected in:

[0039] 1. Obtain historical reconciliation records, preprocess the records to remove duplicates and errors to ensure data integrity, extract textual features, numerical features, time series features, and image features from the records, and convert the features into numerical values. Then, input the numerical value of each feature into a multi-layer weighted network, assign weight values ​​to each feature, and normalize the weight values ​​to obtain multiple features of the bills, enabling multi-faceted bill matching and ensuring the comprehensiveness of bill reconciliation. Then, use a weighted formula to sum the weights of each feature set to obtain a comprehensive feature vector set. Then, train the comprehensive feature vector and adjust the feature weight values ​​based on the training results to obtain the final feature weight values. Through multi-layer weighted network training, adjust and obtain the final feature vector. The system uses feature weights to ensure accuracy, guaranteeing the accuracy of each feature. Then, it uses the final feature weight values ​​and feature vectors to perform a weighted summation to obtain a final comprehensive feature vector set. This comprehensive feature vector set is then trained using a Transformer architecture to obtain a trained feature similarity calculation model. Reconciliation records are then input into the trained model, and the model matches the records based on similarity. Records with high similarity are marked as successfully reconciled, while low-similarity and unmatched records are further analyzed using an anomaly detection algorithm and marked as abnormal records requiring manual review. This system can quickly and accurately match invoices, rapidly identifying abnormal and unmatched records, making it easier for customers to find problematic invoices. Attached Figure Description

[0040] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale.

[0041] Figure 1 This is a flowchart illustrating the intelligent reconciliation method based on multi-feature engineering in Example 1. Detailed Implementation

[0042] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0043] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0044] like Figure 1 As shown, an intelligent reconciliation method based on multi-feature engineering includes,

[0045] S1: Obtain large sample data of historical bills, preprocess the large sample data of bills, and extract historical data features from the preprocessed large sample data.

[0046] S2: Establish a multi-layer weighted network, use the multi-layer weighted network to fuse historical data features into a comprehensive feature vector, train the comprehensive feature vector, and backpropagate the parameters of the weighted network.

[0047] S3: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the Transformer architecture self-attention mechanism model with the fused features, and obtain a trained feature similarity calculation model.

[0048] S4: Calculate the bill similarity using the trained feature similarity calculation model, automatically match reconciliation records based on similarity, mark records with high similarity as successfully reconciled, and use anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

[0049] In this embodiment, the historical data features include text features, numerical features, time series features, and image features; the text features are converted into numerical values ​​through natural language processing technology, the numerical features are features extracted directly from numerical fields such as amount and date, and the image features are QR codes, seals, or handwritten signatures extracted from bills through convolutional neural networks.

[0050] Specifically, feature extraction can be used to extract the data features of the bill content.

[0051] In this embodiment, the multi-layer weighted network includes: a network layer and a weighted fusion layer; the network layer includes a fully connected layer whose input and output are both feature numbers, the fully connected layer is followed by a normalization layer, and the weighted fusion layer.

[0052] Historical data features are fused into a comprehensive feature vector using a multi-layer weighted network, including:

[0053] Obtain features from historical billing data, using the feature set of a single bill as input features;

[0054] Each feature is assigned a weight value through the network layers, and the weight values ​​of each feature set are normalized to ensure that the sum of the weights of each sample is 1.

[0055] The weighted feature summation formula is used to sum the features of each feature set, and the final result is used as the comprehensive feature vector.

[0056] The formula for weighted feature summation is:

[0057]

[0058] Where n is the feature, f is the feature vector, and w is the weight.

[0059] Specifically, by using a multi-layer weighted network, weights can be assigned to each feature, and a weighted formula can be used to obtain a comprehensive vector feature.

[0060] In this embodiment, training the comprehensive feature vector includes,

[0061] Obtain the comprehensive feature vector and divide the comprehensive feature vector into a training set and a validation set;

[0062] Iterate through the entire training set and set the number of samples to process each time;

[0063] The loss function is used to calculate the difference between the model's predicted output and the true label;

[0064] The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm;

[0065] The optimizer updates the parameters of the weight network based on the gradient.

[0066] Specifically, a multi-layer network is trained by integrating vector features to obtain fixed feature weight values.

[0067] In this embodiment, the self-attention mechanism model of the Transformer architecture includes an embedding layer in the first layer, six Transformer encoder modules in the second layer, and an output layer in the third layer. The Transformer encoder consists of six repeating encoders, each of which contains a self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by normalization and residual connections.

[0068] In this embodiment, obtaining the trained feature similarity calculation model includes:

[0069] Using the determined feature weights, calculate the comprehensive feature vector using the weighted feature summation formula, and establish a comprehensive feature vector set;

[0070] Initialize the three-layer Transformer architecture self-attention mechanism model, use the comprehensive feature vector set as the input of the self-attention mechanism model, train the self-attention mechanism model, and obtain the trained self-attention mechanism model.

[0071] In this embodiment, calculating bill similarity using a trained feature similarity calculation model includes:

[0072] Two sets of corresponding fused features are obtained, and the two sets of fused features are input into the Transformer model respectively to obtain two sets of feature representations;

[0073] Cosine similarity is used to measure the similarity between two sets of feature representations;

[0074] The formula for calculating similarity is:

[0075]

[0076] Where xi and yi are the i-th components of vectors x and y, respectively, and n is the dimension of the vector.

[0077] A multi-feature engineering intelligent reconciliation system includes:

[0078] Feature extraction module: Extracts historical data features from large sample data;

[0079] Feature fusion module: It uses a multi-layer weighted network to fuse historical data features into a comprehensive feature vector, trains the comprehensive feature vector, and backpropagates the parameters of the weighted network.

[0080] Similarity calculation module: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the self-attention mechanism model of the Transformer architecture with the fused features, and obtain a trained feature similarity calculation model;

[0081] Reconciliation module: Calculates the similarity of model bills, matches reconciliation records, marks records with high similarity as successfully reconciled, and uses anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

[0082] It should be noted that the process involves acquiring historical reconciliation records, preprocessing these records to extract textual, numerical, time-series, and image features, converting these features into numerical values, and then inputting the numerical values ​​of each feature into a multi-layer weighted network. Each feature is assigned a weight value, which is then normalized. A weighted summation formula is then used to calculate a comprehensive feature vector set. This comprehensive feature vector set is then trained, and the feature weight values ​​are adjusted based on the training results to obtain the final feature weight values. These final feature weight values ​​are then weighted and summed with the feature vectors to obtain the final comprehensive feature vector set. This final comprehensive feature vector set is then trained using a Transformer architecture to obtain a trained feature similarity calculation model. Reconciliation records are then obtained and input into the trained feature similarity calculation model. Records are matched based on similarity, and those with high similarity are marked as successfully reconciled. Records with low similarity or no match are further analyzed using an anomaly detection algorithm and marked as abnormal records requiring manual review.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An intelligent reconciliation method based on multi-feature engineering, characterized in that: include, S1: Obtain large sample data of historical bills, preprocess the large sample data of bills, and extract historical data features from the preprocessed large sample data. S2: Establish a multi-layer weighted network, use the multi-layer weighted network to fuse historical data features into a comprehensive feature vector, train the comprehensive feature vector, and update the parameters of the weighted network through backpropagation; S3: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the Transformer architecture self-attention mechanism model with the fused features, and obtain a trained feature similarity calculation model. S4: Calculate the bill similarity using the trained feature similarity calculation model, automatically match reconciliation records based on similarity, mark records with high similarity as successfully reconciled, and use anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

2. The intelligent reconciliation method based on multi-feature engineering according to claim 1, characterized in that: The historical data features include text features, numerical features, time series features, and image features; the text features are converted into numerical values ​​through natural language processing technology, the numerical features are features extracted directly from numerical fields such as amount and date, and the image features are QR codes, seals, or handwritten signatures extracted from bills through convolutional neural networks.

3. The intelligent reconciliation method based on multi-feature engineering according to claim 2, characterized in that: The multi-layer weighted network includes: a network layer and a weighted fusion layer; the network layer includes a fully connected layer whose input and output are both feature numbers, the fully connected layer is followed by a normalization layer, and the weighted fusion layer.

4. The intelligent reconciliation method based on multi-feature engineering according to claim 3, characterized in that: Historical data features are fused into a comprehensive feature vector using a multi-layer weighted network, including: Obtain features from historical billing data, using the feature set of a single bill as input features; Each feature is assigned a weight value through the network layers, and the weight values ​​of each feature set are normalized to ensure that the sum of the weights of each sample is 1. The weighted feature summation formula is used to sum the features of each feature set, and the final result is used as the comprehensive feature vector. The formula for weighted feature summation is: Where n is the feature, f is the feature vector, and w is the weight.

5. The intelligent reconciliation method based on multi-feature engineering according to claim 4, characterized in that: The training of the comprehensive feature vector includes, Obtain the comprehensive feature vector and divide the comprehensive feature vector into a training set and a validation set; Iterate through the entire training set and set the number of samples to process each time; The loss function is used to calculate the difference between the model's predicted output and the true label; The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm; The optimizer updates the parameters of the weight network based on the gradient.

6. The intelligent reconciliation method based on multi-feature engineering according to claim 5, characterized in that: The self-attention mechanism model of the Transformer architecture includes an embedding layer in the first layer, six Transformer encoder modules in the second layer, and an output layer in the third layer. The Transformer encoder consists of six repeating encoders, each of which contains a self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by normalization and residual connections.

7. The intelligent reconciliation method based on multi-feature engineering according to claim 1, characterized in that: The process of obtaining the trained feature similarity calculation model includes: Using the determined feature weights, calculate the comprehensive feature vector using the weighted feature summation formula, and establish a comprehensive feature vector set; Initialize the three-layer Transformer architecture self-attention mechanism model, use the comprehensive feature vector set as the input of the self-attention mechanism model, train the self-attention mechanism model, and obtain the trained self-attention mechanism model.

8. The intelligent reconciliation method based on multi-feature engineering according to claim 7, characterized in that: The calculation of bill similarity using a trained feature similarity calculation model includes: Two sets of corresponding fused features are obtained, and the two sets of fused features are input into the Transformer model respectively to obtain two sets of feature representations; Cosine similarity is used to measure the similarity between two sets of feature representations; The formula for calculating similarity is: Where xi and yi are the i-th components of vectors x and y, respectively, and n is the dimension of the vector.

9. A multi-feature engineering-based intelligent reconciliation system, applied to the multi-feature engineering-based intelligent reconciliation system described in any one of claims 1-8, characterized in that, include: Feature extraction module: Extracts historical data features from large sample data; Feature fusion module: It uses a multi-layer weighted network to fuse historical data features into a comprehensive feature vector, trains the comprehensive feature vector, and updates the parameters of the weighted network through backpropagation; Similarity calculation module: Construct a self-attention mechanism model with a three-layer architecture using the Transformer architecture, train the self-attention mechanism model of the Transformer architecture with the fused features, and obtain a trained feature similarity calculation model; Reconciliation module: Calculates the similarity of model bills, matches reconciliation records, marks records with high similarity as successfully reconciled, and uses anomaly detection algorithms to further analyze records with low similarity and unmatched records, marking them as abnormal records that require manual review.

Images