One-way data migration method based on deep feature synthesis and migration component analysis

Abstract

The application discloses a one-way data migration method based on deep feature synthesis and migration component analysis, and belongs to the technical field of big data processing and artificial intelligence. First, source domain and target domain data are acquired, and key attributes are screened through feature importance analysis; common data and unique data are separated by comparing indexes; feature derivation is performed on the unique data by using a deep feature synthesis algorithm to generate a deep feature matrix; the common data and the deep feature matrix are fused to construct isomorphic new source domain and new target domain; through a migration component analysis algorithm, a kernel function is introduced and a maximum mean difference is minimized, and an optimal mapping is solved by combining a centralization matrix and a regularization parameter to project the isomorphic data to a public latent feature space; and finally, prediction is performed in the public space by using a K nearest neighbor classifier. The application effectively solves the problem of insufficient model fitting caused by the scarcity of default samples in the target domain, and significantly improves the prediction accuracy and robustness of the credit evaluation model.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data mining technology, specifically to a one-way credit data migration method based on deep feature synthesis (DFS) and transfer component analysis (TCA). Background Technology

[0002] In the financial lending industry, the accuracy of credit rating models heavily relies on a large amount of high-quality labeled data. However, emerging lending platforms or companies in their early stages of development (Company A) often face a "cold start" problem: the accumulated customer data volume is small, especially the number of defaulting customers (negative samples) is extremely small. This sample imbalance and data sparsity lead to insufficient fitting of traditional machine learning models, making it difficult to accurately identify credit risk.

[0003] To address the problem of insufficient data, transfer learning has been extensively studied. Traditional transfer learning assumes that the feature spaces of the source domain (e.g., Company B, with a large amount of data) and the target domain (Company A, with a small amount of data) are the same. However, in real-world cross-platform environments, credit feature datasets exhibit significant heterogeneity due to differences in domain affiliation and platform architecture.

[0004] Inconsistent dimensionality: The number of user attribute fields collected by different platforms varies.

[0005] Inconsistent evaluation indicators: There are differences in the calculation methods and statistical indicator definitions of credit scoring cards.

[0006] Inconsistent data correlation: The logical relationships and dependency structures between data are different.

[0007] While existing shallow domain adaptation methods based on sample reweighting can handle some distributional differences, they cannot directly handle heterogeneous data with inconsistent feature spaces. Therefore, how to transform heterogeneous credit feature data from different sources into homogeneous data and achieve efficient knowledge transfer on this basis is a pressing technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0009] The purpose of this invention is to provide a one-way credit data migration method based on deep feature synthesis and migration component analysis, in order to solve the problems of inability to directly migrate credit data due to heterogeneity in cross-platform environments, low availability of target domain data, and low accuracy of default sample prediction.

[0010] To achieve the above objectives, this invention provides the following technical solution: a one-way data migration method based on deep feature synthesis and transfer component analysis, comprising the following steps: S1, acquiring source domain credit datasets and target domain credit datasets, cleaning and desensitizing the data, and screening key attribute features based on the correlation between attribute features and credit evaluation indicators; S2, comparing the indicators of the source domain credit dataset and the target domain credit dataset, extracting common data subsets with the same indicators and unique data subsets with different indicators; S3, using the deep feature synthesis (DFS) algorithm to derive features from the key attribute features in the unique data subsets, generating a deep feature matrix containing multi-layer transformation features; S4, ... S5. A transfer model based on Transfer Component Analysis (TCA) is constructed. By introducing a kernel function, the maximum mean difference (MMD) matrix, and a centered H matrix, the optimal projection matrix that maps the data to the common latent feature space is solved. S6. The optimal projection matrix is ​​used to map the new source domain dataset and the new target domain dataset to the common latent feature space, resulting in the dimensionality-reduced source domain and target domain feature representations. S7. In the common latent feature space, a classifier is trained using source domain features and labels, and the target domain features are predicted to output the credit prediction result.

[0011] As a preferred embodiment of the one-way data migration method based on deep feature synthesis and migration component analysis described in this invention, step S1 of selecting key attribute features specifically includes: calculating the correlation coefficient between each attribute feature in the dataset and the credit evaluation index; generating an attribute feature importance analysis chart; sorting the attributes from high to low correlation, selecting the top 10 attribute features with the highest correlation as key attribute features for subsequent deep feature synthesis steps.

[0012] As a preferred embodiment of the one-way data migration method based on deep feature synthesis and migration component analysis described in this invention, step S3 specifically includes: performing entity recognition on a unique subset of data to determine entity features for feature derivation; applying a deep feature synthesis algorithm to perform vertical transformation, horizontal transformation, and mathematical operations on the entity features to automatically construct a deep feature matrix; wherein, the feature matrix generated by deep feature synthesis is used to expand the sample capacity and unify the data standard, so that the data in the source domain and the target domain have an isomorphic basis at the feature level.

[0013] As a preferred embodiment of the unidirectional data migration method based on deep feature synthesis and migration component analysis described in this invention, wherein: the optimization objective function constructed in step S5 is: minTr(K T LK)+μTr(K T K), with the constraint: KHK=I m Where K is the kernel matrix, obtained through kernel function mapping; L is the MMD matrix, used to measure the distribution distance between the source and target domains; H is the centering matrix, used to maintain the original feature information of the dataset; μ is the regularization parameter; I m It is an m×m identity matrix.

[0014] As a preferred embodiment of the unidirectional data migration method based on deep feature synthesis and migration component analysis described in this invention, the MMD matrix L is constructed as follows: Let the total number of samples n = n s +n t Construct an n×n matrix L, whose elements L ij The calculation formula is:

[0015]

[0016] Among them, D s Let D be the set of samples from the source domain. t Let n be the sample set of the target domain. s n is the number of samples in the source domain. t The number of samples in the target domain.

[0017] As a preferred embodiment of the unidirectional data migration method based on deep feature synthesis and migration component analysis described in this invention, the centralized matrix H is constructed as follows:

[0018]

[0019] Among them, I n 1 is the identity matrix, and 11 is a vector of all 1s;

[0020] The role of the H matrix is ​​to minimize the MMD distance while preserving the variance information of the data and preventing data degradation.

[0021] As a preferred embodiment of the unidirectional data migration method based on deep feature synthesis and migration component analysis described in this invention, step S5, in which the optimal projection matrix is ​​solved, specifically includes: transforming the optimization objective function into a generalized eigenvalue decomposition problem: (KLK+μI)KHK=ΛKHK, solving the above equation, and selecting the eigenvectors corresponding to the first m smallest eigenvalues ​​to form the projection matrix; where m is the dimension parameter after migration, and Λ is the eigenvalue diagonal matrix.

[0022] As a preferred embodiment of the unidirectional data migration method based on deep feature synthesis and migration component analysis described in this invention, the kernel function includes a linear kernel function or a Gaussian kernel function (RBF); when a Gaussian kernel function is used, the formula for calculating the kernel function K is:

[0023]

[0024] Here, γ is the kernel bandwidth parameter of the Gaussian kernel, which needs to be tuned through grid search.

[0025] As a preferred embodiment of the one-way data transfer method based on deep feature synthesis and transfer component analysis described in this invention, the classifier used in step S7 is the K-nearest neighbor algorithm (KNN); the credit prediction result includes a predicted label, where 1.0 represents compliance and 0.0 represents default; the effect of transfer learning is verified by calculating the accuracy of the predicted label and the original target domain label.

[0026] This invention provides a one-way data migration method based on deep feature synthesis and migration component analysis, which has the following beneficial effects:

[0027] Unified credit rating standards: Companies A and B are located in similar regions and industries, but their credit rating standards may differ at the corporate governance level. The DFS deep feature synthesis algorithm is introduced into the model. By setting synthesis rules, the two datasets are made to have the same dimensional features, laying the data foundation for subsequent transfer learning.

[0028] Expanding sample size: The DFS algorithm not only unifies data standards during training, but also synthesizes new features to expand sample size. The generated sample data can also provide new evaluation standards for enterprises in future production activities.

[0029] Improving the accuracy of default sample labeling: This model improves the model's fit and robustness by utilizing other datasets to learn from negatively labeled samples, thereby improving the accuracy of default sample labeling. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0031] Figure 1 This is an overall flowchart of the method according to an embodiment of the present invention.

[0032] Figure 2This is a schematic diagram of the transfer learning training process based on sample selection.

[0033] Figure 3 This is a graph showing the importance of attribute features.

[0034] Figure 4 The figure shows the results of the migration experiment. Detailed Implementation

[0035] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0036] To address the heterogeneity issues of existing credit feature data, such as inconsistent dimensionality, credit evaluation indicators, and data correlation caused by differences in domain and platform affiliation, this invention proposes a source transfer model based on deep feature synthesis and transfer component analysis. This model solves the problems of small available source data volume and low data usability in cross-platform environments. Heterogeneous credit feature data from different sources are processed uniformly to form a homogeneous credit feature dataset with similar structure, dimensions, and indicators, which can then be used for research on transfer learning methods applied in various scenarios.

[0037] (1) Transfer learning theory

[0038] Given a source domain Heyuan Learning Tasks Target domain and target learning tasks ,exist or In the context of different source and target domains, transfer learning aims to leverage the source domain... Heyuan Learning Tasks Knowledge to help improve the target domain Medium target prediction function Its learning performance.

[0039] In early shallow domain adaptation methods, reweighting instances was used to overcome the bias in the marginal probability distributions of data between two domains, thus achieving domain adaptation. This is one of the most common methods. Its core idea is to calculate the likelihood ratio of a sample instance belonging to either the source or target domain, and use this ratio to weight the samples. Weighting data samples by calculating the maximum mean difference (MMD) between the data distributions of the two domains is a common method. The training process is as follows... Figure 2 As shown.

[0040] For example, in credit loan companies, the number of default samples is insufficient to support the number of default samples required for model fitting. Using a transfer learning method based on sample weights can increase the weight of negative samples (i.e. default samples) in the transfer process, amplify the features of the data with default records, and enable the model after transfer to have better discrimination and prediction capabilities for negative samples.

[0041] (2) Data processing based on deep feature synthesis

[0042] Similar to the augmentation of small sample credit feature data, the research on one-way transfer learning based on deep feature learning selects datasets D1 and D2 with the same indicators through comparison and verification, and then uses the remaining datasets D3 and D4 for the next step of deep feature synthesis.

[0043] (3) Migration model based on migration component analysis

[0044] The feature matrix (feature_matrix) synthesized through stage (2) is concatenated with datasets D1 and D2, which have the same indices, to form new source domains D5 and D6 and a new target domain. Transfer learning based on component transfer analysis is then performed using datasets D5 and D6. Let D5 be the source domain (src), and its attribute value be... , tag as D6 represents the target domain tar, and its attribute value is... , tag as Transfer component analysis (JBO) solves for the optimal mapping between the source and target domains by minimizing the MMD distance. This is transformed into a mathematical problem that requires the introduction of the L and H matrices:

[0045] L matrix:

[0046]

[0047] H matrix:

[0048]

[0049] The H matrix serves to preserve the original feature information of the dataset. Then, two kernel functions are selected for mapping (Linear kernel and Gaussian kernel RBF) to calculate K, and the matrix K is solved.

[0050]

[0051] Solving for the first m eigenvalues ​​of (KLK+μI)−1KHK yields a low-dimensional matrix K (Equation 3). The optimal mapping is then used to reduce the distance between the source and target domains, maximizing the accuracy of the predicted label output from the new target domain. The research framework is as follows: Figure 1 As shown.

[0052] (4) Validation of isomorphic credit transfer learning technology

[0053] Will The data in [the data source] is transferred to [the location] via linear transformation and spatial mapping. In the middle, we get X s_new and X t_new The K-nearest neighbor algorithm is used to call the predict method to predict the output label Y. pred and original target domain tags Accuracy analysis was performed to verify the effectiveness of transfer learning. The pseudocode for the model algorithm is shown below:

[0054]

[0055] Specifically, the one-way data migration method based on deep feature synthesis and migration component analysis includes the following steps:

[0056] S1. Obtain the source domain credit dataset and the target domain credit dataset, clean and desensitize the data, and select key attribute features based on the correlation between attribute features and credit evaluation indicators.

[0057] It should be noted that: The process involves obtaining the original credit datasets (e.g., the credit dataset from Alibaba Cloud's "Tianchi" data network) from both the source domain (e.g., the data-rich platform B) and the target domain (e.g., the data-scarce platform A). The data undergoes data anonymization (removing sensitive information such as names and ID numbers) and cleaning (filling in missing values ​​and handling outliers). Subsequently, feature importance analysis is performed. The correlation between each attribute feature and credit evaluation indicators (e.g., the "Approved" field, indicating whether the creditor has kept their promises) is calculated. Based on the analysis results, such as... Figure 3 As shown, the top 10 attribute features most correlated with the indicator "Approved" were identified for feature derivation in the next experimental step.

[0058] The dataset used in this model experiment comes from the credit dataset in Alibaba Cloud's "Tianchi" data network, and has undergone a certain degree of anonymization. The data first undergoes importance analysis to select some attribute features for deep feature synthesis. Based on the importance analysis chart of attribute features and evaluation indicators, as shown... Figure 3 As shown, the top 10 attributes most correlated with the "Approved" metric are identified, and feature derivation is performed in the next experimental step. As mentioned in the theoretical section above, this model uses the Entity Feature derivation method, that is, new features are derived from the feature attributes selected through importance analysis.

[0059] S2. Compare the metrics of the source domain credit dataset and the target domain credit dataset, and extract the common data subset with the same metrics and the unique data subset with different metrics.

[0060] It should be noted that by comparing the dataset structures of the source and target domains, feature indicators with identical field names and physical meanings were identified, including:

[0061] Common dataset extraction: Extract the data subsets containing the same metrics, and denote them as common datasets D1 (from the source domain) and D2 (from the target domain).

[0062] Unique Dataset Extraction: The remaining fields, i.e., heterogeneous features that exist only in the source domain or only in the target domain, are extracted into unique datasets D3 (unique to the source domain) and D4 (unique to the target domain), respectively. Since the dimensions and definitions of D3 and D4 are inconsistent, they cannot be directly jointly trained and require further deep feature synthesis.

[0063] S3. Based on the deep feature synthesis DFS algorithm, the key attribute features in the unique data subset are derived to generate a deep feature matrix containing multi-layer transformation features.

[0064] It should be noted that, to address the heterogeneity issue, this step employs a deep feature synthesis algorithm to process the unique datasets D3 and D4. DFS, based on entity features, derives new features by automatically applying a series of transformation operators (such as addition, subtraction, multiplication, division, and aggregation). This process not only unifies the data dimensionality but also enhances the data's expressive power through feature cross-referencing. Specifically, the DFS algorithm is executed on the key attribute features in D3 and D4 to generate a deep feature matrix containing a large number of derived features. This matrix maps the heterogeneous feature space to a unified derived feature space, achieving both sample size expansion and standardized evaluation criteria. This model uses the Entity Feature derivation method.

[0065] S4. Connect the common data subset with the deep feature matrix to construct a new source domain dataset and a new target domain dataset with consistent feature dimensions and structure.

[0066] It should be noted that the shared datasets D1 and D2 extracted in step two are concatenated with the deep feature matrices generated in step three. Specifically, the deep feature matrices generated from D1 and D3 are concatenated to form a new source domain dataset D5; the deep feature matrices generated from D2 and D4 are concatenated to form a new target domain dataset D6. At this point, D5 and D6 are completely identical in structure, dimensions, and metric definitions, meeting the conditions for proceeding to the next stage of the credit isomorphic transfer learning model experiment.

[0067] S5. Construct a migration model based on migration component analysis (TCA). By introducing a kernel function, the maximum mean difference (MMD) matrix, and a centered H matrix, solve for the optimal projection matrix that maps the data to the common latent feature space.

[0068] It should be noted that although D5 and D6 have the same dimensions, their data distributions still differ (the marginal probability distributions are different). This step uses the transfer component analysis algorithm to find a common latent feature space into which D5 and D6 are mapped, minimizing the difference in data distribution between the two domains within this space.

[0069] Further explanation is needed regarding the technical details of this solution, which are as follows:

[0070] Define input: Let D5 be the source domain src, and its attribute value be X. s The label is Y s D6 represents the target domain tar, and its attribute value is X. t The label is Y t .

[0071] Kernel function mapping: Two kernel functions are selected for mapping: linear kernel and Gaussian kernel (RBF). The kernel matrix K is calculated.

[0072] MMD Matrix Construction (L Matrix): Construct the maximum mean difference (MMD) matrix L to measure the distribution distance between the source and target domains.

[0073] Centralized matrix construction (H matrix): Constructing a centralized matrix:

[0074]

[0075] The purpose of the H matrix is ​​to maintain the original feature information of the dataset.

[0076] Optimizing the objective function: Converting this into a mathematical problem requires introducing the L and H matrices. minTr(K) T LK)+μTr(K T K)stKHK=I m

[0077] The first term is the MMD distance, the second term is the regularization term, μ is the balance parameter (i.e., lamb), and I m It is an identity matrix.

[0078] Solving the mapping: Solving the matrix (KLK+μI) −1 The first m eigenvalues ​​of KHK (corresponding to the smallest eigenvalues) are used to obtain the low-dimensional matrix of matrix K. The optimal mapping is then obtained to reduce the distance between the source domain and the target domain, so that the predicted label output of the new target domain has the best accuracy.

[0079] S6. Using the optimal projection matrix, map the new source domain dataset and the new target domain dataset to the common latent feature space to obtain the dimensionality-reduced feature representations of the source domain and the target domain.

[0080] It should be noted that: using the optimal mapping matrix obtained by solving, the data from the new source domain D5 and the new target domain D6 are projected onto the common latent feature space to obtain the dimensionality-reduced feature representation X. s_new and X t_new In this process, a grid search method is used to optimize key parameters. The key parameters of the TCA algorithm include four: the choice of the kernel function, the distance *dim* after transfer, the λ value *lamb* in the equation, and the kernel bandwidth *gamma* of the Gaussian kernel function. The source and target domain datasets, after being cleaned by the deep feature synthesis algorithm, are structurally and dimensionally identical, meeting the conditions for proceeding to the next stage of the credit isomorphic transfer learning model experiment. The key parameters involved in the transfer component analysis algorithm include four: the choice of the kernel function, the distance *dim* after transfer, the λ value *lamb* in the equation, and the kernel bandwidth *gamma* of the Gaussian kernel function. The ultimate goal of the transfer component analysis algorithm is to optimize the distance between the source and target domains to obtain the result with the optimal transfer accuracy, i.e., the *dim* and *lamb* values ​​are optimal; if a Gaussian kernel function is used, the *gamma* value must also be optimal. During the experiment, the linear kernel "linear" and the Gaussian kernel "rbf" were used to verify the accuracy of isomorphic credit transfer. After parameter tuning, the TCA algorithm will output a new data feature matrix. The K-nearest neighbor classifier is used to predict labels on the new feature matrix, and its transfer accuracy is calculated. After experimental debugging, the parameters with the best transfer effect are shown in Table 1:

[0081] Table 1. Parameters for Optimal Migration Performance

[0082]

[0083] Linear kernel function: dim = 30, lamb = 1, transfer accuracy = 51.23%.

[0084] Gaussian kernel function rbf: dim = 30, lamb = 1, gamma = 4, transfer accuracy = 51.31%.

[0085] S7. In the common latent feature space, train the classifier using source domain features and labels, and predict the target domain features to output the credit prediction result.

[0086] It should be noted that the data in D5 and D6 are transferred to the common space through linear transformation and spatial mapping to obtain X. s_new and X t_new The K-Nearest Neighbors (KNN) algorithm is used to predict the label Y using the `predict` method. pred and the original target domain label Y tAccuracy analysis was performed to verify the effectiveness of transfer learning. An additional set of experiments using the KNN baseline classifier for data transfer accuracy prediction served as a control. The basic idea of ​​KNN is to input unclassified new data, first extract features from the new data and compare them with each feature in the test set; then, extract the K most similar data feature labels from the test set, and count the most frequent category among the K nearest neighbors, taking that as the new data category. This is similar to the concept of "birds of a feather flock together" in real life. Based on the simplicity of the KNN classifier's classification, it is suitable as the baseline classifier for verifying the transfer accuracy in this experiment, objectively demonstrating the feasibility and effectiveness of the transfer component analysis algorithm in solving practical problems. The results of the control experiments are shown in Table 2.

[0087] Table 2. Control Experiment Table

[0088]

[0089] KNN baseline classifier: 47.30%

[0090] Homogeneous credit transfer model (linear kernel): 51.23%

[0091] Homogeneous credit transfer model (RBF kernel): 51.31%

[0092] The above transfer experiment results demonstrate that the isomorphic credit transfer model, combining deep feature synthesis and transfer component analysis algorithms, can, to some extent, address issues of mutual trust such as inconsistent dimensionality of credit feature space datasets, inconsistent credit evaluation indicators, and inconsistent data correlation. Comparative experiments using the baseline classifier KNN show that the isomorphic credit transfer model can improve transfer accuracy by nearly 4%, which is significant for improving the accuracy of predicting whether credit customers will default. Because in most real-world credit default scenarios, the number of defaulters (the negative label) is usually small, leading to insufficient model fitting and stagnant prediction accuracy, this source transfer model improves transfer accuracy by expanding the data sample. This improvement in transfer accuracy indirectly proves the improved prediction accuracy of the model.

[0093] This technical report also integrates the information source transfer model into the "Multi-Source Credit Data Management System," simulating transfer learning between companies A and B by inputting two sets of information source data to supplement the actual situation of insufficient data. The transfer experiment results are as follows: Figure 4As shown, the red area displays the migration results. The field "Approved" indicates whether the predicted tuple customer defaults, with 1.0 representing compliance and 0.0 representing default. The other fields describe some of the customer's own attributes, such as loan amount, employment status, credit score, etc.

[0094] Furthermore, the specific steps in step S1 for selecting key attribute features include:

[0095] Calculate the correlation coefficients between each attribute feature in the dataset and the credit rating indicators;

[0096] Generate an attribute feature importance analysis chart;

[0097] Sort the attributes by relevance from high to low, and select the top 10 attributes with the highest relevance as key attributes for use in the subsequent deep feature synthesis steps.

[0098] Furthermore, step S3 specifically includes:

[0099] Entity identification is performed on a unique subset of data to determine the entity features used for feature derivation;

[0100] By applying a deep feature synthesis algorithm, vertical and horizontal transformations and mathematical operations are performed on entity features to automatically construct a deep feature matrix.

[0101] Among them, the feature matrix generated by deep feature synthesis is used to expand the sample capacity and unify the data standards, so that the data in the source domain and the target domain have a homogeneous basis at the feature level.

[0102] Furthermore, the optimization objective function constructed in step S5 is: minTr(K) T LK)+μTr(K T K), with the constraint: KHK=I m Where K is the kernel matrix, obtained through kernel function mapping; L is the MMD matrix, used to measure the distribution distance between the source and target domains; H is the centering matrix, used to maintain the original feature information of the dataset; μ is the regularization parameter; I m It is an m×m identity matrix.

[0103] Furthermore, the MMD matrix L is constructed as follows: Let the total number of samples be n=n s +n t Construct an n×n matrix L, whose elements L ij The calculation formula is:

[0104]

[0105] Among them, D s Let D be the set of samples from the source domain. tLet n be the sample set of the target domain. s n is the number of samples in the source domain. t The number of samples in the target domain.

[0106] Furthermore, the centralization matrix H is constructed as follows:

[0107]

[0108] Among them, I n 1 is the identity matrix, and 11 is a vector of all 1s;

[0109] The role of the H matrix is ​​to minimize the MMD distance while preserving the variance information of the data and preventing data degradation.

[0110] Furthermore, the solution to the optimal projection matrix in step S5 specifically includes: transforming the optimization objective function into a generalized eigenvalue decomposition problem: (KLK+μI)KHK=ΛKHK, solving the above equation, and selecting the eigenvectors corresponding to the first m smallest eigenvalues ​​to form the projection matrix;

[0111] Where m is the dimension parameter after the transfer, and Λ is the eigenvalue diagonal matrix.

[0112] Furthermore, kernel functions include linear kernel functions or Gaussian kernel functions (RBF); when using a Gaussian kernel function, the formula for calculating the kernel function K is:

[0113]

[0114] Here, γ is the kernel bandwidth parameter of the Gaussian kernel, which needs to be tuned through grid search.

[0115] Furthermore, the classifier used in step S7 is the K-Nearest Neighbors (KNN) algorithm; the credit prediction result includes a predicted label, where 1.0 represents compliance and 0.0 represents default; the effect of transfer learning is verified by calculating the accuracy of the predicted label versus the original target domain label.

[0116] Additionally, to better understand the technical solution of the present invention, the following Embodiment 1 and Embodiment 2 are provided:

[0117] Example 1: One-way credit data migration method

[0118] This example simulates two different companies in the same region and industry. During the production process, Company A (target domain) needs to use data from Company B (source domain) to perform transfer learning and build a credit rating model.

[0119] 1. Data Preprocessing and Feature Selection

[0120] First, two source datasets (Credit1, Credit2) are read. After data cleaning, feature importance analysis is performed. The correlation between each feature and the label "Approved" (1.0 represents compliance, 0.0 represents default) is calculated. Figure 3 As shown, the top 10 most relevant attributes (such as loan amount, employment status, credit score, etc.) are selected as key features, and irrelevant features are removed. This model uses the Entity Feature derivation method.

[0121] 2. Feature set segmentation

[0122] Comparing the field names of Credit1 and Credit2, assuming both contain fields such as "Age" and "Income," these constitute the shared datasets D1 (from Credit1) and D2 (from Credit2). The remaining fields, such as "Field_X" unique to Credit1 and "Field_Y" unique to Credit2, constitute the unique datasets D3 and D4, respectively. Since D3 and D4 have different structures, the distance cannot be directly calculated; a further deep feature synthesis step is required.

[0123] 3. Deep Feature Synthesis (DFS)

[0124] Perform the Depth-First Search (DFS) algorithm on D3 and D4. The algorithm applies transformation operators to the first 10 key features based on the entity (e.g., user ID). For example, it generates "Loan Amount / Monthly Income" as a new debt repayment capability feature, and "Number of Years Residing at Current Address * Employment Status" as a stability feature. After DFS processing, D3 and D4 generate deep feature matrices containing the same derived feature definitions.

[0125] 4. Data Restructuring

[0126] Horizontally connect the depth feature matrices of D1 and D3 to form a new source domain D5; horizontally connect the depth feature matrices of D2 and D4 to form a new target domain D6. At this point, D5 and D6 have the same number of columns and data format.

[0127] 5. Transmitted Component Analysis (TCA)

[0128] Building the TCA model:

[0129] Input: D5(source domain X) s ), D6 (target domain X) t ).

[0130] Parameter settings: Select RBF (Gaussian kernel) kernel function, and set dim=30, lamb=1, gamma=4.

[0131] Calculate the matrix:

[0132] Calculate the n×n kernel matrix K, where n=n s +n t .

[0133] Calculate the MMD matrix L.

[0134] Calculate the centering matrix H.

[0135] Solution: Solve the eigenvalue problem (KLK+μI)KHK=ΛKHK, and take the first 30 eigenvectors.

[0136] Mapping: Project K using these 30 feature vectors to obtain 30-dimensional common feature space data.

[0137] 6. Model Prediction and Evaluation

[0138] The model is trained using the KNN algorithm (K=5) in a 30-dimensional common space.

[0139] Input: Mapped source domain data and labels.

[0140] Prediction: Predicting the target domain data after mapping.

[0141] Output: Predicted default probability labels (0 or 1) for customers in the target domain.

[0142] Validation: Calculate the prediction accuracy (Acc). As shown in Table 2, the TCA-RBF model achieves an accuracy of 51.31%, higher than the KNN benchmark of 47.30%. Results are... Figure 4 The data is displayed in the form of a "Approved" field, which represents the predicted compliance / default label, while the other fields describe the customer's own attributes.

[0143] Example 2: Credit Data Migration System

[0144] This invention also provides a one-way credit data migration system based on deep feature synthesis and migration component analysis, comprising:

[0145] Data Acquisition and Preprocessing Module: Responsible for accessing data from multiple sources, performing cleaning, desensitization, and feature importance screening, and outputting key attribute features.

[0146] Deep feature synthesis module: Receives heterogeneous and unique datasets, performs feature derivation through the DFS algorithm, generates a deep feature matrix, and solves the problem of inconsistent dimensions.

[0147] Homogenization module: Connects the common dataset with the deep feature matrix to output D5 and D6 datasets with uniform structure.

[0148] Migration Component Analysis Module: The core computational module, embedding the TCA algorithm. It is responsible for constructing the MMD and H matrices, performing grid search to optimize parameters (dim, lamb, gamma), calculating the optimal mapping matrix, and realizing the transformation from the original space to the common latent feature space.

[0149] Prediction and Output Module: Uses the KNN classifier for training and prediction in the public space, and visualizes the transfer accuracy and prediction results (such as customer credit scores and default determination).

[0150] This system is integrated into the "Multi-Source Credit Data Management System." By inputting two sets of credit data, it simulates transfer learning research between companies A and B to supplement the actual situation where its own data is lacking. In the transfer experiment results, the red area shows the transfer results. The field "Approved" indicates whether the predicted tuple customer has defaulted, with 1.0 representing compliance and 0.0 representing default.

[0151] This experiment explores optimal parameters by using a grid search (GridSearchCV) method to continuously traverse the optimal regions, ultimately obtaining a set of optimal parameters. A control group is also included, using the KNN algorithm as a baseline classifier for comparison. Experimental data shows that the transfer accuracy of the proposed source transfer model is improved compared to the baseline classifier KNN, intuitively demonstrating the feasibility and practicality of this model in practice. Since the experimental data simulates different companies in the same industry across platforms, homogenization was performed using the DFS algorithm, resulting in some differences between the data, leading to a less significant improvement in transfer accuracy. However, the experiment demonstrates that introducing transfer learning algorithms effectively improves enterprise credit data in terms of data volume and prediction label accuracy, providing a new approach to solving three major problems encountered in establishing credit indicator systems: single and incomplete data sources, unbroken information silos, and inconsistent standards.

[0152] The dataset used in this experiment simulates two different companies in the same region and industry. During the production process, Company A needs to use Company B's data for transfer learning to build a credit rating model.

[0153] 1) Unified credit rating standards. Companies A and B are located in similar regions and industries, but their credit rating standards may differ at the corporate governance level. The DFS deep feature synthesis algorithm is introduced into the model. By setting synthesis rules, the two datasets are made to have the same dimensional features, laying the data foundation for subsequent transfer learning.

[0154] 2) Expanding sample size. The DFS algorithm not only standardizes data during training, but also synthesizes new features to expand the sample size. The generated sample data can also provide enterprises with new evaluation standards for future production activities.

[0155] 3) Improve the accuracy of default sample labeling. As mentioned above, an important role of transfer learning is to increase the labeling of target data. In fields like credit, where the ratio of positive to negative samples is severely imbalanced, transfer learning improves the model's fit and robustness by utilizing other datasets to learn from negative labeled samples, thereby improving the accuracy of default sample labeling.

[0156] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A one-way data migration method based on deep feature synthesis and migration component analysis, characterized in that, Includes the following steps: S1. Obtain the source domain credit dataset and the target domain credit dataset, clean and desensitize the data, and select key attribute features based on the correlation between attribute features and credit evaluation indicators. S2. Compare the metrics of the source domain credit dataset and the target domain credit dataset, and extract the common data subset with the same metrics and the unique data subset with different metrics. S3. Based on the deep feature synthesis DFS algorithm, the key attribute features in the unique data subset are derived to generate a deep feature matrix containing multi-layer transformation features. S4. Connect the common data subset with the deep feature matrix to construct a new source domain dataset and a new target domain dataset with consistent feature dimensions and structure. S5. Construct a migration model based on migration component analysis (TCA). By introducing a kernel function, the maximum mean difference (MMD) matrix, and a centered H matrix, solve for the optimal projection matrix that maps the data to the common latent feature space. S6. Using the optimal projection matrix, map the new source domain dataset and the new target domain dataset to the common latent feature space to obtain the dimensionality-reduced feature representations of the source domain and the target domain. S7. In the common latent feature space, train the classifier using source domain features and labels, and predict the target domain features to output the credit prediction result.

2. The unidirectional data migration method based on deep feature synthesis and migration component analysis according to claim 1, characterized in that, Step S1, which involves filtering key attribute features, specifically includes: Calculate the correlation coefficients between each attribute feature in the dataset and the credit rating indicators; Generate an attribute feature importance analysis chart; Sort the attributes by relevance from high to low, and select the top 10 attributes with the highest relevance as key attributes for use in the subsequent deep feature synthesis steps.

3. The unidirectional data migration method based on deep feature synthesis and migration component analysis according to claim 2, characterized in that, Step S3 specifically includes: Entity identification is performed on a unique subset of data to determine the entity features used for feature derivation; By applying a deep feature synthesis algorithm, vertical and horizontal transformations and mathematical operations are performed on entity features to automatically construct a deep feature matrix. Among them, the feature matrix generated by deep feature synthesis is used to expand the sample capacity and unify the data standards, so that the data in the source domain and the target domain have a homogeneous basis at the feature level.

4. The unidirectional data migration method based on deep feature synthesis and migration component analysis according to claim 3, characterized in that, The optimization objective function constructed in step S5 is: minTr(K) T LK)+μTr(K T K), with the constraint: KHK=I m Where K is the kernel matrix, obtained through kernel function mapping; L is the MMD matrix, used to measure the distribution distance between the source and target domains; H is the centering matrix, used to maintain the original feature information of the dataset; μ is the regularization parameter; I m It is an m×m identity matrix.

5. The one-way data migration method based on deep feature synthesis and migration component analysis according to claim 4, characterized in that, The MMD matrix L is constructed as follows: Let the total number of samples be n = n s +n t Construct an n×n matrix L, whose elements L ij The calculation formula is: Among them, D s Let D be the set of samples from the source domain. t Let n be the sample set of the target domain. s n is the number of samples in the source domain. t The number of samples in the target domain.

6. The one-way data migration method based on deep feature synthesis and migration component analysis according to claim 5, characterized in that, The centralization matrix H is constructed as follows: Among them, I n 1 is the identity matrix, and 11 is a vector of all 1s; The role of the H matrix is ​​to minimize the MMD distance while preserving the variance information of the data and preventing data degradation.

7. The one-way data migration method based on deep feature synthesis and migration component analysis according to claim 6, characterized in that, The specific steps for solving the optimal projection matrix in step S5 include: transforming the optimization objective function into a generalized eigenvalue decomposition problem: (KLK+μI)KHK=ΛKHK, solving the above equation, and selecting the eigenvectors corresponding to the first m smallest eigenvalues ​​to form the projection matrix; Where m is the dimension parameter after the transfer, and Λ is the eigenvalue diagonal matrix.

8. The one-way data migration method based on deep feature synthesis and migration component analysis according to claim 7, characterized in that, Kernel functions include linear kernel functions or Gaussian kernel functions (RBF). When using a Gaussian kernel function, the formula for calculating the kernel function K is: Here, γ is the kernel bandwidth parameter of the Gaussian kernel, which needs to be tuned through grid search.

9. The one-way data migration method based on deep feature synthesis and migration component analysis according to claim 8, characterized in that, The classifier used in step S7 is the K-Nearest Neighbors (KNN) algorithm; the credit prediction result includes a predicted label, where 1.0 represents compliance and 0.0 represents default; the effect of transfer learning is verified by calculating the accuracy of the predicted label and the original target domain label.

Images