A financial credit data processing method, apparatus, storage medium, and electronic device.

Abstract

This invention provides a financial credit data processing method, apparatus, storage medium, and electronic device, applied in the field of data processing technology. The invention acquires high-dimensional tabular data containing structured features related to credit applicants and their behavior, then sequentially performs low-rank feature compression, dimensionality compression, and quantization processing to generate a quantum-enhanced feature vector. This fused target feature vector is then input into a pre-trained credit scoring model to obtain the scoring result. This achieves both efficient feature compression and enhanced non-linear expressive power of the features, significantly improving the feature extraction effect of complex high-dimensional data in financial credit scoring, and effectively ensuring the accuracy and stability of financial credit scoring.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a financial credit data processing method, apparatus, storage medium, and electronic device. Background Technology

[0002] Feature extraction is a key data preprocessing step in machine learning. Traditional methods mainly include principal component analysis (PCA) and autoencoders with nonlinear transformations.

[0003] However, while PCA is suitable for linear feature extraction, it performs poorly on high-dimensional nonlinear data. Autoencoders, though capable of mining nonlinear features, are complex to train and prone to overfitting. With the development of quantum computing, quantum neural networks (QNNs) have been introduced to enhance nonlinear feature extraction capabilities, but their high requirements for the number of qubits and quantum gates limit their practical applications. Tensor networks, as an efficient low-rank feature compression technique, perform well in multidimensional data analysis, but their ability to handle complex nonlinear relationships is limited. Existing methods struggle to balance feature representation power and computational efficiency, especially performing poorly in high-dimensional nonlinear scenarios such as financial credit scoring.

[0004] Therefore, how to achieve efficient low-rank feature compression while enhancing the ability to express nonlinear features, thereby improving the feature extraction effect in high-dimensional and complex scenarios such as financial credit scoring, has become a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides a financial credit data processing method, apparatus, storage medium, and electronic device that overcomes or at least partially solves the above problems. The technical solution is as follows:

[0006] A financial credit data processing method, comprising:

[0007] Obtain high-dimensional tabular data, wherein the high-dimensional tabular data includes structured features related to credit applicants and credit application behavior;

[0008] The high-dimensional tabular data is subjected to low-rank feature compression to obtain the first feature vector;

[0009] The first feature vector is compressed in dimension to obtain the second feature vector;

[0010] The second eigenvector is quantized to obtain a quantum-enhanced eigenvector;

[0011] The second feature vector and the quantum-enhanced feature vector are fused to obtain the target feature vector;

[0012] The target feature vector is input into a pre-trained credit scoring model to obtain a financial credit scoring result.

[0013] Optionally, the step of performing low-rank feature compression on the high-dimensional tabular data to obtain the first feature vector includes:

[0014] Based on knowledge of the financial lending business domain, the feature dimensions of the high-dimensional tabular data are split to obtain multiple sub-features;

[0015] For any of the sub-features: perform tensor shrinkage operation on the sub-feature and the corresponding kernel tensor to obtain an intermediate tensor;

[0016] Merge the intermediate tensors to obtain low-rank features;

[0017] Flatten the low-rank features to obtain the first feature vector.

[0018] Optionally, the step of dimensionality compression of the first feature vector to obtain the second feature vector includes:

[0019] The first feature vector is linearly transformed by at least one fully connected layer;

[0020] The first eigenvector after linear transformation is compressed to a dimension that matches the preset number of qubits to obtain the second eigenvector.

[0021] Optionally, the step of quantizing the second eigenvector to obtain a quantum-enhanced eigenvector includes:

[0022] The second feature vector is encoded into a quantum state according to the preset number of qubits;

[0023] The quantum state is input into a parameterized quantum circuit for evolution;

[0024] The evolved quantum state is measured to obtain the quantum enhancement eigenvector.

[0025] Optionally, encoding the second feature vector into a quantum state includes:

[0026] Each component of the second feature vector is used as a rotation angle parameter;

[0027] Based on the rotation angle parameter, a Y-axis rotation gate operation is applied to each qubit to obtain the encoded quantum state.

[0028] Optionally, fusing the second feature vector and the quantum-enhanced feature vector to obtain the target feature vector includes:

[0029] The second feature vector and the quantum-enhanced feature vector are concatenated to obtain the concatenated feature vector;

[0030] Perform a linear transformation on the spliced ​​feature vector;

[0031] An activation function is applied to the concatenated feature vector after linear transformation to obtain the target feature vector.

[0032] Optionally, before obtaining the high-dimensional tabular data, the method further includes:

[0033] Obtain a training dataset that includes sample features and corresponding credit risk labels;

[0034] The sample features are sequentially subjected to low-rank feature compression, dimensionality compression, quantization processing, and fusion to obtain a predicted feature vector;

[0035] Calculate the loss function value based on the predicted feature vector and the credit risk label;

[0036] Based on the loss function value, the trainable parameters associated with the low-rank feature compression, the dimensionality compression, the quantization process, and the fusion process are updated through backpropagation.

[0037] A financial credit data processing device includes: a high-dimensional tabular data acquisition unit, a low-rank feature compression unit, a dimensionality compression unit, a quantization processing unit, a fusion unit, and a credit scoring unit;

[0038] The high-dimensional table data acquisition unit is used to acquire high-dimensional table data, wherein the high-dimensional table data includes structured features related to the credit applicant and credit application behavior;

[0039] The low-rank feature compression unit is used to perform low-rank feature compression on the high-dimensional table data to obtain a first feature vector;

[0040] The dimension compression unit is used to perform dimension compression on the first feature vector to obtain the second feature vector.

[0041] The quantization processing unit is used to quantize the second feature vector to obtain a quantum-enhanced feature vector;

[0042] The fusion unit is used to fuse the second feature vector and the quantum-enhanced feature vector to obtain the target feature vector;

[0043] The credit scoring unit is used to input the target feature vector into a pre-trained credit scoring model to obtain a financial credit scoring result.

[0044] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the aforementioned financial credit data processing method.

[0045] An electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the financial credit data processing method.

[0046] By employing the above technical solutions, this invention provides a financial credit data processing method, apparatus, storage medium, and electronic device. The method acquires high-dimensional tabular data, which includes structured features related to credit applicants and credit application behavior. The high-dimensional tabular data undergoes low-rank feature compression to obtain a first feature vector. The first feature vector is then subjected to dimensionality compression to obtain a second feature vector. The second feature vector is quantized to obtain a quantum-enhanced feature vector. The second feature vector and the quantum-enhanced feature vector are fused to obtain a target feature vector. The target feature vector is then input into a pre-trained credit scoring model to obtain a financial credit scoring result. This invention improves the feature extraction effect of complex high-dimensional data in financial credit scoring by sequentially performing low-rank feature compression, dimensionality compression, and quantization on high-dimensional tabular data containing structured features related to credit applications, and by fusing traditional and quantum-enhanced feature vectors. This constructs a target feature that combines efficient compression and non-linear expression capabilities, thereby enhancing the feature extraction effect of complex high-dimensional data in financial credit scoring and ensuring the accuracy and stability of financial credit scoring.

[0047] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0049] Figure 1 A flowchart illustrating one embodiment of the financial credit data processing method provided by this invention is shown.

[0050] Figure 2 This is a flowchart illustrating a specific implementation of step S110 in the financial credit data processing method provided in this embodiment of the invention.

[0051] Figure 3 This is a flowchart illustrating a specific implementation of step S120 in the financial credit data processing method provided by an embodiment of the present invention;

[0052] Figure 4 This is a flowchart illustrating a specific implementation of step S130 in the financial credit data processing method provided in this embodiment of the invention.

[0053] Figure 5 This is a flowchart illustrating a specific implementation of step S140 in the financial credit data processing method provided in this embodiment of the invention.

[0054] Figure 6 A schematic diagram of the structure of the financial credit data processing device provided in an embodiment of the present invention is shown;

[0055] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present invention is shown. Detailed Implementation

[0056] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0057] In machine learning and data processing, feature extraction is a crucial step in data preprocessing, and its quality directly impacts the performance of subsequent models. Traditional feature extraction methods primarily rely on linear transformation techniques, such as Principal Component Analysis (PCA), and nonlinear transformation methods, such as autoencoders. PCA achieves dimensionality reduction through linear projection and can effectively extract the main linear features in the data, but its performance is limited when faced with complex nonlinear data structures, making it difficult to capture deep-seated patterns. Autoencoders, as a neural network-based nonlinear feature extraction method, can uncover richer nonlinear features, but their training process is typically computationally complex and resource-intensive, and they are prone to overfitting when the sample size is insufficient, affecting the model's generalization ability.

[0058] With the rapid development of quantum computing technology, quantum neural networks (QNNs) have been proposed to enhance the expressive power of traditional machine learning models. QNNs utilize the properties of quantum superposition and entanglement, and have the potential to handle high-dimensional, complex, and nonlinear features, improving the quality and efficiency of feature extraction. However, in practical applications, QNNs face limitations in the number of qubits and quantum gate operations, leading to high demands on quantum computing resources and restricting their widespread application in large-scale, high-dimensional data processing.

[0059] On the other hand, Tensor Networks (TNNs), as an efficient low-rank feature compression technique, have been widely applied in the field of multidimensional data analysis. TNNs reduce computational complexity and improve storage efficiency by decomposing and compressing high-dimensional data, exhibiting good performance in certain specific scenarios. However, traditional TNNs have limited capabilities in handling complex nonlinear relationships and cannot fully express the deep nonlinear structures in the data, affecting their effectiveness in complex tasks.

[0060] In summary, while existing feature extraction methods each have their advantages, they also have significant limitations, especially in practical applications such as financial credit scoring with high-dimensional, highly nonlinear, and multimodal data, where it is difficult to simultaneously balance feature representation capabilities, computational efficiency, and resource consumption.

[0061] Based on this, this invention provides a financial credit data processing method. By acquiring high-dimensional tabular data containing structured features related to credit applicants and their behavior, the method sequentially performs low-rank feature compression, dimensionality compression, and quantization processing to generate a quantum-enhanced feature vector. The fused target feature vector is then input into a pre-trained credit scoring model to obtain the scoring result. This method achieves efficient feature compression and enhances the nonlinear expressive power of features, thereby significantly improving the feature extraction effect of complex high-dimensional data in financial credit scoring and effectively ensuring the accuracy and stability of financial credit scoring.

[0062] like Figure 1 The diagram shows a flowchart of one embodiment of the financial credit data processing method provided by this invention. The method may include:

[0063] S100. Obtain high-dimensional tabular data, which includes structured features related to credit applicants and credit application behavior.

[0064] High-dimensional tabular data refers to tabular data containing a large number of structured features and with high feature dimensions. The high-dimensional tabular data provided in this embodiment of the invention can be data information related to credit applications, such as: demographic information, loan history, repayment behavior, and other complex data sets composed of multiple modules, fields, and derived features.

[0065] Among them, the credit applicant refers to an individual or organization applying for a loan from a financial institution. Their user information, credit history, income status, etc. are all included as structured features in the high-dimensional tabular data for risk assessment.

[0066] Among them, credit application behavior refers to the behavior of credit applicants in the credit process, including dynamic behavioral characteristics such as repayment history, number of overdue payments, and number of loans, which is an important dimension for assessing credit risk.

[0067] Specifically, embodiments of the present invention can extract structured tabular data related to credit applicants in batches from credit business systems or databases, organize it according to business dimensions, and supplement it with field-level and derived statistical features. Then, these raw data are standardized, missing value imputation and anomaly detection are performed to generate high-dimensional feature tensors in a unified format, thus obtaining high-dimensional tabular data.

[0068] As examples, embodiments of the present invention can extract multi-dimensional structured data of credit applicants from a credit business database to form a high-dimensional table. This table includes structured fields such as demographic information (e.g., age, occupation), historical credit records (e.g., number of loans, repayment status), behavioral data (e.g., number of recent inquiries), and third-party credit scores. Each credit applicant corresponds to a feature vector, with dimensions... The number of dimensions can reach several thousand (e.g., 3000), resulting in a batch size of... Input tensor .

[0069] S110. Perform low-rank feature compression on the high-dimensional tabular data to obtain the first feature vector.

[0070] Specifically, embodiments of the present invention can perform structured decomposition of high-dimensional features based on business knowledge, and use a set of preset kernel tensors to perform low-rank approximation and information extraction on each sub-feature through tensor operations. The extracted information is then merged and transformed into a unified vector representation, thereby compressing high-dimensional tabular data into a compact and structured first feature vector.

[0071] S120. Perform dimensionality compression on the first feature vector to obtain the second feature vector.

[0072] Specifically, in this embodiment of the invention, a multi-layer perceptron (MPL) can be used to perform nonlinear mapping and dimensional reduction on the first feature vector, compressing it to a target dimension that matches the subsequent quantum processing, thereby obtaining a second feature vector with a more compact structure and more concentrated information.

[0073] S130. The second eigenvector is quantized to obtain the quantum-enhanced eigenvector.

[0074] Among them, quantum-enhanced eigenvectors refer to low-dimensional eigenvectors processed by quantum neural networks.

[0075] Specifically, embodiments of the present invention can utilize the principles of quantum computing to encode the second feature vector as the initial state of a qubit, and then perform a unitary transformation evolution on it through a learnable quantum circuit. The evolved quantum state information is then extracted and mapped back to classical data form through quantum measurement, thereby obtaining a quantum-enhanced feature vector containing quantum superposition and entanglement properties.

[0076] S140. The second eigenvector and the quantum-enhanced eigenvector are fused to obtain the target eigenvector.

[0077] The target feature vector refers to the final feature representation after fusing the multilayer perceptron compressed feature vector and the quantum-enhanced feature vector. It contains both efficient compressed information and rich nonlinear expression, and serves as the input to the credit scoring model.

[0078] Specifically, in this embodiment of the invention, the second feature vector and the quantum-enhanced feature vector can be fused through feature splicing and nonlinear transformation to form a unified target feature vector with higher representational capability.

[0079] S150. Input the target feature vector into the pre-trained credit scoring model to obtain the financial credit scoring result.

[0080] Among them, the financial credit scoring result refers to the credit risk score generated after inputting the target feature vector into a pre-trained credit scoring model. It can be used to determine whether the credit applicant has the ability to repay and the level of credit risk.

[0081] Specifically, in this embodiment of the invention, the target feature vector can be input into a pre-trained credit risk scoring model through supervised learning, and forward inference can be performed to output a financial credit score containing a credit risk score or default probability. This financial credit score can be directly used for credit decisions, risk pricing, or as a core input to downstream risk control systems to guide actual business processes.

[0082] As examples, embodiments of the present invention can use the target feature vector as input, feeding it into a pre-trained credit scoring model (e.g., logistic regression, XGBoost, or a simple linear classifier). This credit scoring model can be trained based on historical training data (e.g., default status) and can map high-dimensional features to credit risk probabilities. Through forward propagation calculations, it outputs a credit score or risk probability (e.g., a value between 0 and 1) for each credit applicant. Finally, based on a set threshold (e.g., 0.5), it is transformed into a binary classification decision (e.g., approve / reject) or a continuous score for ranking, thereby supporting financial credit approval decisions.

[0083] This invention provides a financial credit data processing method, which includes: acquiring high-dimensional tabular data, wherein the high-dimensional tabular data includes structured features related to credit applicants and credit application behavior; performing low-rank feature compression on the high-dimensional tabular data to obtain a first feature vector; performing dimensional compression on the first feature vector to obtain a second feature vector; performing quantization on the second feature vector to obtain a quantum-enhanced feature vector; fusing the second feature vector and the quantum-enhanced feature vector to obtain a target feature vector; and inputting the target feature vector into a pre-trained credit scoring model to obtain a financial credit scoring result. This invention improves the feature extraction effect of complex high-dimensional data in financial credit scoring by sequentially performing low-rank feature compression, dimensional compression, and quantization on high-dimensional tabular data containing structured features related to credit applications, and by fusing traditional and quantum-enhanced feature vectors, constructing a target feature that combines efficient compression and non-linear expression capabilities. This ensures the accuracy and stability of financial credit scoring.

[0084] Optional, based on Figure 1 The method shown is as follows: Figure 2 The diagram shows a specific implementation of step S110 in the financial credit data processing method provided by this invention. Step S110 may specifically include:

[0085] S200: Based on knowledge of the financial credit business domain, the feature dimensions of high-dimensional tabular data are split to obtain multiple sub-features.

[0086] Specifically, in this embodiment of the invention, the high-dimensional tabular data input can be dimensionally split according to the modular business structure based on the characteristics of financial credit business. For example, in this embodiment of the invention, the features can be divided into 10 main modules, each main module can be further subdivided into 30 field subclasses, and each field can be further subdivided into 10 derived statistics or coding features, ultimately forming a three-dimensional tensor format (B×10×30×10), where B is the batch size.

[0087] As examples, embodiments of the present invention can, based on knowledge in the field of financial lending, decompose the input 3000-dimensional feature vector into multiple levels according to the business logic structure. First, the overall feature vector is divided into several main modules, for example... The main modules include demographic information, loan history, repayment behavior, and third-party ratings. Each main module is then further subdivided into... Each field subclass contains specific fields such as "Age" and "Number of Historical Overdue Payments". Finally, each field subclass includes... Each derived statistic or encoded class feature. Through this modular decomposition, the original input tensor... Reorganized into a three-dimensional structure with clear business meaning. ,Right now This lays the foundation for subsequent hierarchical tensor processing.

[0088] S210. For any sub-feature: perform tensor shrinkage operation on the sub-feature and the corresponding kernel tensor to obtain the intermediate tensor.

[0089] Specifically, in this embodiment of the invention, for each of the decomposed sub-feature tensors, tensor shrinking is performed with the corresponding MPO (Matrix Product Operator) kernel tensor at each level to construct a three-layer MPO core tensor chain, corresponding to the main module, field subclass, and derived feature level, respectively. Through Einstein summation, the input tensor is multiplied sequentially by each layer's kernel tensor, and the key dimensions are combined to compress the sub-feature information layer by layer, outputting the corresponding intermediate tensor and capturing a low-rank feature representation containing business structure.

[0090] As examples, embodiments of the present invention may define a corresponding kernel tensor for each of the split business layers. For the first Layer, and its corresponding sub-feature tensor With kernel tensor Perform Einstein summation to achieve tensor shrinkage. For example, the first-layer kernel tensor. The dimension can be {1, 10, 5, d1}, which corresponds to the sub-features of the input (corresponding to...). We shrink each module to obtain an intermediate tensor. Similarly, the second-layer kernel tensor... The dimensions can be {d1, 30, 5, d2}, and the third-layer kernel tensor The dimension can be {d², 10, 4, 1}. The key to this operation lies in using a shared "key dimension". and To capture and fuse the correlation information of features across levels, while also reducing the original dimensions of the features. Compress to output dimension Each level of contraction operation produces an intermediate tensor whose dimensions contain the compressed feature information.

[0091] S220. Merge the intermediate tensors to obtain low-rank features.

[0092] Specifically, in this embodiment of the invention, after completing the tensor shrinkage of each sub-feature, all intermediate tensors can be merged and integrated according to the business dimension to form an overall low-rank feature representation.

[0093] As some examples, embodiments of the present invention can complete all Levels (such as) After the tensor shrinking operation, the intermediate tensors generated at each level are processed by iterative trace operation. The intermediate results from all levels are merged. For example, in the hierarchical shrinkage operation, the intermediate tensor output from the previous level can be used as part of the input, and then shrunk together with the core tensor of the current level. This process fuses and compresses the intermediate results from all levels along a shared key dimension chain.

[0094] ,

[0095] in, For input Individual characteristics, For the corresponding number A kernel tensor, Perform a trace operation (merging key dimensions). Output: ).

[0096] Finally, after three contraction operations, the original high-dimensional features are effectively compressed into a low-rank tensor. ,Right now This tensor is the low-rank feature, which, through the low-rank decomposition structure of the tensor network, preserves the core linear and low-order nonlinear relationships in the original high-dimensional data.

[0097] S230. Flatten the low-rank features to obtain the first eigenvector.

[0098] Specifically, in this embodiment of the invention, low-rank tensors can be batch-flattened to convert them into a two-dimensional matrix format, obtaining a first eigenvector. This first eigenvector retains the core information and low-order nonlinear correlations of the original data while significantly reducing dimensionality, providing efficient input for subsequent MLP compression and quantum enhancement modules.

[0099] As some examples, embodiments of the present invention can obtain multidimensional low-rank feature tensors. Perform a flattening operation. The flattening operation concatenates all feature dimensions (i.e., 5×5×4=100) except for the batch dimension B, forming a two-dimensional tensor. Each row (100-dimensional) of this two-dimensional tensor is the first feature vector corresponding to a sample. This vector is a highly compressed and abstracted low-rank representation extracted from the original 3000-dimensional features, providing input for subsequent further compression and quantization processing.

[0100] This invention achieves efficient and structured low-rank feature compression by splitting high-dimensional tabular data features based on financial credit business knowledge and combining kernel tensors for tensor shrinkage. This results in a first feature vector with strong expressive power and appropriate dimensionality, improving the accuracy and computational efficiency of subsequent credit scoring models.

[0101] Optional, based on Figure 1 The method shown is as follows: Figure 3 The diagram shows a specific implementation of step S120 in the financial credit data processing method provided by this invention. Step S120 may specifically include:

[0102] S300, Perform a linear transformation on the first feature vector through at least one fully connected layer.

[0103] Specifically, in this embodiment of the invention, at least one trainable fully connected layer can be constructed for the first feature vector. The weight matrix and bias vector of this layer are optimized using the backpropagation algorithm. After inputting the first feature vector, a linear transformation is achieved using matrix multiplication, converting it into a new feature representation. This adjusts and reconstructs the feature space structure, preparing for subsequent dimensionality compression and quantum encoding. This process ensures that the linear combination of features is fully learned, improving the effectiveness of feature representation and information condensation.

[0104] As some examples, embodiments of the present invention can use the first feature vector As input, a trainable fully connected layer f is fed in. This fully connected layer performs linear transformation operations. ,in, , is the bias vector. For the preset target dimension (e.g.) This linear transformation process maps the first eigenvector from the original 100-dimensional space to a new one. The 3D linear space aims to recombine and weight features through learnable parameters to capture more complex linear relationships between features, preparing for subsequent nonlinear compression.

[0105] S310. Compress the first eigenvector after linear transformation to a dimension that matches the preset number of qubits to obtain the second eigenvector.

[0106] Specifically, in this embodiment of the invention, after completing the linear transformation output, a trainable fully connected layer is used to compress the dimension of the transformed feature vector to a preset number of qubits Q, thereby achieving a mapping from a high-dimensional feature space to a low-dimensional quantum coding space. For example, this embodiment of the invention can use matrix multiplication to map the input features to a dimension of... The vector space satisfies the quantum simulator's limitation on the number of input qubits while preserving key feature information, facilitating subsequent angle embedding and parameterized quantum circuit processing. This compression process can be continuously optimized through training to balance information loss and compression efficiency.

[0107] As examples, embodiments of the present invention can directly achieve dimensionality compression of features through the output dimension Q of the fully connected layer after completing the linear transformation: by designing the number of columns of the weight matrix to a preset number of qubits Q (e.g., Q=8), the output of the linear transformation... The dimension is compressed to This process further compresses the 100-dimensional low-rank features extracted by MPO into a low-dimensional space that strictly matches the number of qubits in the subsequent quantum neural network. The final result is... This is the second eigenvector, and the value of each dimension (corresponding to one qubit) will be directly used as the rotation angle input for quantum state encoding, ensuring a seamless connection between classical features and quantum processing modules, and providing a feasible input dimension for quantum simulation under resource-constrained conditions.

[0108] The embodiments of the present invention employ at least one fully connected layer to linearly transform and compress the first feature vector to a dimension that matches the preset number of qubits. This effectively reduces the feature dimension, decreases the consumption of quantum simulation resources, and maintains the integrity of key information. It provides efficient and suitable input features for subsequent quantization processing, thereby improving the overall performance and computational efficiency of the credit scoring model.

[0109] Optional, based on Figure 1 The method shown is as follows: Figure 4 The diagram shows a specific implementation of step S130 in the financial credit data processing method provided by this invention. Step S130 may specifically include:

[0110] S400. Encode the second feature vector into a quantum state according to the preset number of qubits.

[0111] Specifically, in this embodiment of the invention, each component of the second feature vector can be used as a rotation angle parameter; based on the rotation angle parameter, a Y-axis rotation gate operation is applied to each qubit to obtain the encoded quantum state.

[0112] As some examples, embodiments of the present invention can use the second feature vector Each component As the rotation angle, it is mapped to a quantum state using an angle encoding method: for each sample's 8-dimensional feature vector, the corresponding... Each qubit is initially in its ground state. Subsequently, regarding the first Apply a Y-axis rotation gate to each qubit. This operation is based on the characteristic value. This serves as a rotation angle parameter. Ultimately, the prepared states of all qubits are combined through tensor products to form the initial quantum state. This process completes the lossless encoding of classical feature information (the second eigenvector) into the phase of the quantum state, providing input for subsequent quantum computing.

[0113] S410. Input the quantum state into the parameterized quantum circuit for evolution.

[0114] Specifically, in this embodiment of the invention, the encoded quantum state can be used as the initial input of the circuit to construct a D-layer parameterized quantum circuit (VQC). Each layer includes a single-qubit rotation gate with trainable parameters and a quantum bit entanglement gate (CNOT). The quantum state is transformed and evolved according to a preset network structure. The adjustable parameters are used to improve the ability of the quantum state space to express complex features and complete the nonlinear mapping of the quantum state.

[0115] As some examples, embodiments of the present invention can encode the initial quantum state. The input is fed into a parameterized quantum circuit consisting of D=3 layers of variable quantum circuits. Each layer of the circuit performs the following operations in sequence: First, a trainable single-bit rotation gate is applied to each qubit. ,in, Is with layer and bit The corresponding learnable parameters; secondly, in each pair of adjacent qubits and An entanglement gate is applied between the qubits to establish entanglement. The evolution of the entire circuit is described by unitary operators: This process uses trainable parameters. With a fixed entanglement structure, quantum states undergo complex nonlinear transformations in Hilbert space, thereby enhancing their characterization capabilities.

[0116] S420. Measure the evolved quantum state to obtain the quantum enhancement eigenvector.

[0117] Specifically, in this embodiment of the invention, the expected value on each qubit can be measured by Pauli-Z measurement operation after the quantum state has evolved through parameterized quantum circuits, so as to obtain a Q-dimensional quantum enhancement feature vector. This vector reflects the information characteristics of the quantum system and is used to characterize higher-order correlations and nonlinear features that are difficult to capture by classical methods.

[0118] As some examples, embodiments of the present invention can process the final quantum state after parameterized quantum circuit evolution. Measurements are performed independently on each qubit, and the observations are Pauli-Z operators. For the first Calculate the Pauli-Z expectation value of each qubit. The expected value is a real number between -1 and +1. This is repeated for all... Performing the above calculation on each qubit yields a vector consisting of 8 expected values. This vector is the quantum-enhanced feature vector, which encodes high-dimensional nonlinear mode information captured by superposition and entanglement in quantum states that is difficult to express by classical methods, providing input for subsequent feature fusion.

[0119] By encoding the second feature vector into a quantum state and using parameterized quantum circuits for evolution and measurement, this invention can fully exploit the high-order nonlinear characteristics and complex correlations of the data, thereby obtaining richer and more discriminative quantum-enhanced feature vectors, significantly improving the accuracy and robustness of the credit scoring model.

[0120] Optional, based on Figure 1 The method shown is as follows: Figure 5 The diagram shows a specific implementation of step S140 in the financial credit data processing method provided by this invention. Step S140 may specifically include:

[0121] S500: Concatenate the second eigenvector and the quantum-enhanced eigenvector to obtain the concatenated eigenvector.

[0122] Specifically, in this embodiment of the invention, a concatenation operation can be used to concatenate a second feature vector and a quantum-enhanced feature vector of the same dimension along the feature dimension to form a concatenated feature vector, thereby preserving the complete information of the classical features and the quantum-enhanced features.

[0123] As some examples, embodiments of the present invention can divide the second feature vector along the feature dimension (i.e., the first dimension, after the batch dimension). and quantum-enhanced eigenvectors By concatenating the first and last parts, a concatenated feature vector is generated. .

[0124] S510. Perform a linear transformation on the concatenated feature vectors.

[0125] Specifically, in this embodiment of the invention, the concatenated feature vector can be transformed through a trainable fully connected (linear) layer. That is, matrix multiplication and bias addition operations are performed using the weight matrix and bias vector learned through backpropagation during model training to adjust and map the distribution and structure of the concatenated features, thereby achieving weighted integration and dimensional transformation of the features and enhancing the flexibility and discriminative ability of feature expression.

[0126] As some examples, embodiments of the present invention can concatenate feature vectors. The input is passed to a trainable fully connected layer (linear layer) for a linear transformation. This transformation is performed by the weight matrix. and bias vector parameterization Parameterization, execution operation This linear transformation globally reweights and combines the fused 16-dimensional features, aiming to learn the optimal linear correlation between different feature dimensions (including classical and quantum parts) and map the features to a new 16-dimensional vector space. This prepares for the introduction of nonlinear activation.

[0127] S520. Apply an activation function to the concatenated feature vector after linear transformation to obtain the target feature vector.

[0128] Specifically, in this embodiment of the invention, the concatenated feature vector after linear transformation can be input into a nonlinear activation function to capture complex feature relationships, enhance the expressive power of features, and avoid information loss caused by linear transformation, thereby outputting the final target feature vector as a high-quality input for credit scoring or downstream tasks.

[0129] As some examples, embodiments of the present invention can transform vectors into linear vectors. The ReLU (Rectified Linear Unit) activation function is applied element-wise to generate the target feature vector. The ReLU function, by introducing nonlinearity, enhances the model's representational power and promotes feature sparsity by suppressing negative values, thus helping the network learn more discriminative feature representations. The final result... This refers to the target feature vector extracted by the model, which integrates classical low-rank and quantum enhancement information. It can be directly used for downstream classification tasks (such as credit risk prediction through the Sigmoid function) or as input for other risk control models.

[0130] This invention achieves an effective fusion of classical and quantum features by concatenating the second feature vector with the quantum enhanced feature vector and processing it through linear transformation and activation function. This enhances the expressive power and nonlinear discriminative power of the features, thereby significantly improving the accuracy and robustness of the credit scoring model.

[0131] Optionally, in the above Figure 1 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, before step S100, the method may further include:

[0132] Obtain a training dataset including sample features and corresponding credit risk labels; sequentially perform low-rank feature compression, dimensionality compression, quantization, and fusion on the sample features to obtain a predicted feature vector; calculate the loss function value based on the predicted feature vector and the credit risk label; and update the trainable parameters associated with the low-rank feature compression, dimensionality compression, quantization, and fusion processes through backpropagation based on the loss function value.

[0133] Among them, sample features refer to the structured or numerical descriptive information corresponding to each credit application sample in the credit scoring task, including the applicant's demographic information, credit history, repayment behavior, asset status and third-party scoring, and other multi-dimensional original or derived indicators.

[0134] Among them, credit risk labels refer to target variables related to credit default or credit risk that are labeled by actual business or historical data for each credit application sample. They usually exist in a binary classification form (for example, 1 indicates that the user is a risky customer and 0 indicates that the user is a non-risky customer) and are used as "true values" in supervised learning to guide the model to learn the mapping relationship between features and risks.

[0135] Specifically, embodiments of the present invention can collect raw tabular data containing multidimensional user information from a credit business data warehouse or a third-party data source. Each data record represents a user sample, and its feature part is a 3000-dimensional vector, including 3000 feature variables under 10 main modules such as demographic information, loan history, and repayment behavior; the label part is a binary variable, with a value of 0 (non-risk customer) or 1 (risk customer), used to identify whether the user has experienced risk events such as credit default. Through a data preprocessing module, these raw data are standardized to eliminate the influence of dimensions and missing values ​​are filled. Finally, the processed feature data and label data are organized into a tensor format to form a training dataset.

[0136] The low-rank feature compression, dimensionality compression, quantization processing and fusion involved in the embodiments of the present invention can be referred to the relevant descriptions in steps S110 to S140, which will not be repeated here.

[0137] In this embodiment of the invention, the predicted feature vector can be used as input, passed through a credit scoring model (such as a fully connected layer and a sigmoid function), and the predicted probability can be output. , representing the probability that the sample is a high-risk customer, where B is the batch size. Using the true labels of the samples. Calculate the binary cross entropy (BCE) loss function:

[0138] ,

[0139] This loss function is used to measure the deviation between the model's predicted probability and the true label.

[0140] Furthermore, embodiments of the present invention utilize an automatic differentiation mechanism, focusing on the loss function. Calculate the MPO kernel tensor MPL fully connected layer weights, QNN variable quantum circuit parameters fusion layer weights W and bias vector The gradient information is used. Based on the gradient information, the Adam optimizer is employed with a set learning rate (e.g., 10). -3 The parameters are iteratively updated to optimize model performance. This process is carried out throughout the training epoch and batch, achieving end-to-end joint training, which enables the model to gradually improve the accuracy and generalization ability of credit risk prediction.

[0141] Before acquiring high-dimensional tabular data, this invention constructs a training dataset containing sample features and their credit risk labels. Using an end-to-end training process, the sample features are sequentially processed through low-rank feature compression, dimensionality compression, quantization, and fusion. Combined with loss function calculation and backpropagation optimization, the training effect of each learnable parameter can be effectively improved, thereby enhancing the expressive power of the target feature vector and significantly improving the prediction accuracy and generalization performance of the subsequent credit scoring model.

[0142] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0143] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0144] Corresponding to the above method embodiments, this invention also provides a financial credit data processing device, the structure of which is as follows: Figure 6 As shown, it may include: a high-dimensional table data acquisition unit 10, a low-rank feature compression unit 20, a dimension compression unit 30, a quantization processing unit 40, a fusion unit 50, and a credit scoring unit 60.

[0145] The high-dimensional table data acquisition unit 10 is used to acquire high-dimensional table data, which includes structured features related to credit applicants and credit application behavior.

[0146] The low-rank feature compression unit 20 is used to perform low-rank feature compression on high-dimensional tabular data to obtain the first feature vector.

[0147] The dimension compression unit 30 is used to compress the first feature vector to obtain the second feature vector.

[0148] The quantization processing unit 40 is used to quantize the second eigenvector to obtain a quantum-enhanced eigenvector.

[0149] The fusion unit 50 is used to fuse the second feature vector and the quantum-enhanced feature vector to obtain the target feature vector.

[0150] Credit scoring unit 60 is used to input the target feature vector into a pre-trained credit scoring model to obtain financial credit scoring results.

[0151] Optionally, the low-rank feature compression unit 20 can be used to split the feature dimensions of high-dimensional tabular data based on knowledge of financial credit business domains to obtain multiple sub-features; for any sub-feature: perform tensor shrinkage operation on the sub-feature and the corresponding kernel tensor to obtain an intermediate tensor; merge the intermediate tensors to obtain a low-rank feature; flatten the low-rank feature to obtain the first feature vector.

[0152] Optionally, the dimension compression unit 30 can be used to perform a linear transformation on the first feature vector through at least one fully connected layer; and to compress the linearly transformed first feature vector to a dimension that matches the preset number of qubits to obtain a second feature vector.

[0153] Optionally, the quantization processing unit 40 can be used to encode the second feature vector into a quantum state according to a preset number of qubits; input the quantum state into a parameterized quantum circuit for evolution; and measure the evolved quantum state to obtain the quantum-enhanced feature vector.

[0154] Optionally, the quantization processing unit 40 can be used to take each component of the second eigenvector as a rotation angle parameter; based on the rotation angle parameter, a Y-axis rotation gate operation is applied to each qubit to obtain the encoded quantum state.

[0155] Optionally, the fusion unit 50 can be used to concatenate the second feature vector and the quantum-enhanced feature vector to obtain a concatenated feature vector; perform a linear transformation on the concatenated feature vector; and apply an activation function to the linearly transformed concatenated feature vector to obtain the target feature vector.

[0156] Optionally, the financial credit data processing device may also include a parameter training unit.

[0157] The parameter training unit is used to obtain a training dataset including sample features and corresponding credit risk labels; the sample features are sequentially subjected to low-rank feature compression, dimensionality compression, quantization and fusion to obtain a predicted feature vector; the loss function value is calculated based on the predicted feature vector and the credit risk label; based on the loss function value, the trainable parameters associated with the low-rank feature compression, dimensionality compression, quantization and fusion processes are updated through backpropagation.

[0158] This invention provides a financial credit data processing device, which is used to: acquire high-dimensional tabular data, wherein the high-dimensional tabular data includes structured features related to credit applicants and credit application behavior; perform low-rank feature compression on the high-dimensional tabular data to obtain a first feature vector; perform dimensional compression on the first feature vector to obtain a second feature vector; perform quantization processing on the second feature vector to obtain a quantum-enhanced feature vector; fuse the second feature vector and the quantum-enhanced feature vector to obtain a target feature vector; and input the target feature vector into a pre-trained credit scoring model to obtain a financial credit scoring result. This invention improves the feature extraction effect of complex high-dimensional data in financial credit scoring by sequentially performing low-rank feature compression, dimensional compression, and quantization processing on high-dimensional tabular data containing structured features related to credit applications, and fuses traditional and quantum-enhanced feature vectors, thereby constructing a target feature that combines efficient compression and nonlinear expression capabilities. This ensures the accuracy and stability of financial credit scoring.

[0159] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0160] The financial credit data processing device includes a processor and a memory. The high-dimensional table data acquisition unit 10, low-rank feature compression unit 20, dimension compression unit 30, quantization processing unit 40, fusion unit 50 and credit scoring unit 60 are all stored in the memory as program units. The processor executes the program units stored in the memory to realize the corresponding functions.

[0161] The processor contains a kernel that retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, high-dimensional tabular data containing structured features related to credit applicants and their behavior are acquired. This data is then sequentially processed through low-rank feature compression, dimensionality compression, and quantization to generate a quantum-enhanced feature vector. This fused target feature vector is then input into a pre-trained credit scoring model to obtain the scoring result. This approach achieves both efficient feature compression and enhanced non-linear expressive power of features, significantly improving the feature extraction performance of complex high-dimensional data in financial credit scoring and effectively ensuring the accuracy and stability of financial credit scoring.

[0162] This invention provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the financial credit data processing method.

[0163] This invention provides a processor for running a program, wherein the program executes the financial credit data processing method during runtime.

[0164] like Figure 7 As shown, this embodiment of the invention provides an electronic device 1000, which includes at least one processor 1001, at least one memory 1002 connected to the processor 1001, and a bus 1003. The processor 1001 and the memory 1002 communicate with each other via the bus 1003. The processor 1001 is used to call program instructions in the memory 1002 to execute the aforementioned financial credit data processing method. The electronic device in this document can be a server, PC, PAD, mobile phone, etc.

[0165] The present invention also provides a computer program product that, when executed on an electronic device, is suitable for executing a program that initializes a financial credit data processing method step.

[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] In a typical configuration, an electronic device includes one or more processors (CPUs), memory, and a bus. The electronic device may also include input / output interfaces, network interfaces, etc.

[0168] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0169] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0171] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0172] In the description of this invention, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0173] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0174] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0175] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the present invention.

Claims

1. A method for processing financial credit data, characterized in that, include: Obtain high-dimensional tabular data, wherein the high-dimensional tabular data includes structured features related to credit applicants and credit application behavior; The high-dimensional tabular data is subjected to low-rank feature compression to obtain the first feature vector; The first feature vector is compressed in dimension to obtain the second feature vector; The second eigenvector is quantized to obtain a quantum-enhanced eigenvector; The second feature vector and the quantum-enhanced feature vector are fused to obtain the target feature vector; The target feature vector is input into a pre-trained credit scoring model to obtain a financial credit scoring result.

2. The method according to claim 1, characterized in that, The step of performing low-rank feature compression on the high-dimensional tabular data to obtain a first feature vector includes: Based on knowledge of the financial lending business domain, the feature dimensions of the high-dimensional tabular data are split to obtain multiple sub-features; For any of the sub-features: perform tensor shrinkage operation on the sub-feature and the corresponding kernel tensor to obtain an intermediate tensor; Merge the intermediate tensors to obtain low-rank features; Flatten the low-rank features to obtain the first feature vector.

3. The method according to claim 1, characterized in that, The step of compressing the dimensionality of the first feature vector to obtain the second feature vector includes: The first feature vector is linearly transformed by at least one fully connected layer; The first eigenvector after linear transformation is compressed to a dimension that matches the preset number of qubits to obtain the second eigenvector.

4. The method according to claim 1, characterized in that, The step of quantizing the second eigenvector to obtain a quantum-enhanced eigenvector includes: The second feature vector is encoded into a quantum state according to the preset number of qubits; The quantum state is input into a parameterized quantum circuit for evolution; The evolved quantum state is measured to obtain the quantum enhancement eigenvector.

5. The method according to claim 4, characterized in that, Encoding the second feature vector into a quantum state includes: Each component of the second feature vector is used as a rotation angle parameter; Based on the rotation angle parameter, a Y-axis rotation gate operation is applied to each qubit to obtain the encoded quantum state.

6. The method according to claim 1, characterized in that, The process of fusing the second feature vector and the quantum-enhanced feature vector to obtain the target feature vector includes: The second feature vector and the quantum-enhanced feature vector are concatenated to obtain the concatenated feature vector; Perform a linear transformation on the spliced ​​feature vector; An activation function is applied to the concatenated feature vector after linear transformation to obtain the target feature vector.

7. The method according to any one of claims 1 to 6, characterized in that, Prior to obtaining the high-dimensional table data, the method further includes: Obtain a training dataset that includes sample features and corresponding credit risk labels; The sample features are sequentially subjected to low-rank feature compression, dimensionality compression, quantization processing, and fusion to obtain a predicted feature vector; Calculate the loss function value based on the predicted feature vector and the credit risk label; Based on the loss function value, the trainable parameters associated with the low-rank feature compression, the dimensionality compression, the quantization process, and the fusion process are updated through backpropagation.

8. A financial credit data processing device, characterized in that, include: The system includes a high-dimensional tabular data acquisition unit, a low-rank feature compression unit, a dimensionality compression unit, a quantization processing unit, a fusion unit, and a credit scoring unit. The high-dimensional table data acquisition unit is used to acquire high-dimensional table data, wherein the high-dimensional table data includes structured features related to the credit applicant and credit application behavior; The low-rank feature compression unit is used to perform low-rank feature compression on the high-dimensional table data to obtain a first feature vector; The dimension compression unit is used to perform dimension compression on the first feature vector to obtain the second feature vector. The quantization processing unit is used to quantize the second feature vector to obtain a quantum-enhanced feature vector; The fusion unit is used to fuse the second feature vector and the quantum-enhanced feature vector to obtain the target feature vector; The credit scoring unit is used to input the target feature vector into a pre-trained credit scoring model to obtain a financial credit scoring result.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the financial credit data processing method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the financial credit data processing method as described in any one of claims 1 to 7.

Images