Intrinsic interpretable model construction method based on Gaussian process in financial science and technology scene
By adopting an intrinsic interpretable model construction method based on Gaussian processes in machine learning models in the field of financial technology, the complex and difficult problem of model decision-making process is solved, achieving higher accuracy and interpretability, and reducing training costs.
Patent Information
- Application Number
- CN202510130201.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-05
- Publication Date
- 2025-06-20
AI Technical Summary
The existing machine learning models in the financial technology field have complex and difficult to explain the decision process, which leads to low trust in the model by users and regulators and may introduce bias, leading to unfair decision-making.
The intrinsic interpretable model construction method based on Gaussian process is adopted, and the Gaussian process kernel function is constructed through the random Fourier feature method, the Sobol index is calculated in combination with the functional variance analysis method, the characteristics and interactive features are screened, and an intrinsic interpretable neural additive model is constructed in the generalized additive model framework.
It improves the accuracy and interpretability of the model, reduces the training cost and computational complexity, and enables the model to be closer to the accuracy of the black box model, while retaining the good interpretability of the linear model.
Smart Images
Figure CN120181248A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of interpretable models, and more particularly to a method for constructing an intrinsically interpretable model based on Gaussian processes in the context of fintech scenarios. Background Art
[0002] In the field of fintech, the application of machine learning and other complex models is rapidly changing the face of financial services. The role and performance of machine learning models in the fintech field are remarkable. By analyzing large amounts of data, these models can discover patterns hidden in the data, thereby improving the accuracy and efficiency of financial decision-making. For example, credit scoring models can quickly assess the credit risk of borrowers, making the loan approval process more efficient; risk management models can predict market changes, helping financial institutions better avoid risks; and fraud detection models can identify and prevent suspicious transactions in real time, protecting the funds of customers.
[0003] However, these models also have some significant drawbacks. First, the decision-making processes of many machine learning models, especially deep learning models, are extremely complex and difficult to explain using traditional methods. This "black box" nature leads to an important problem: when the output results of the model cannot be understood and explained, users and regulatory agencies may be skeptical of these results, affecting the trust and acceptance of the model. In addition, complex models may inadvertently introduce biases, resulting in unfair decision-making, such as systematic discrimination against certain groups in credit scoring.
[0004] Therefore, the need for model interpretability has become even more urgent. Interpretability means that the decision-making process of the model can be understood and traced by humans, and can clearly explain why a particular decision is made. For machine learning models, interpretability has many important implications. For example, enhancing the transparency of the model can increase the trust of users and regulatory agencies; interpretability also helps to identify and correct biases and errors in the model, ensuring fair and accurate decision-making, and improving and optimizing the model based on this; good interpretability also helps to meet legal and regulatory requirements and ensure the rights of users. Therefore, how to make the model interpretable and enhance the transparency and credibility of the model is the focus of current research.
[0005] According to different interpretation methods, existing interpretability methods can be roughly divided into two types: intrinsic interpretability and post-hoc interpretability.
[0006] Intrinsic Interpretability: Intrinsic interpretability means that the model itself has the ability to be interpreted. The structures and behaviors of these models are designed to be simple and transparent enough for users to directly understand and interpret their decision-making processes. However, compared with relatively simple intrinsically interpretable models, their accuracy may not be comparable to that of complex models. The existing main intrinsically interpretable models mainly include linear regression, logistic regression, decision trees, rule sets, GAMs, etc.
[0007] Linear Regression: The linear regression model predicts the target as a weighted sum of the feature inputs. As the most basic linear model, the interpretation result based on weights it provides is very simple and clear. However, linear regression needs to meet a series of strict assumptions, such as normality, homoscedasticity, independence, fixed features, and no multicollinearity, etc.
[0008] Logistic Regression: Logistic regression is the probability modeling for binary classification problems, which is an extension of the linear regression model for classification problems. It restricts the output range through the logistic function to obtain the classification probability; however, relatively speaking, the multiplicative weights make the interpretation of logistic regression more difficult.
[0009] Decision Tree: The tree model divides the dataset into different groups by splitting the data through thresholds in the dataset. The interpretation of the tree model can be directly carried out using the thresholds within the nodes, which is simple and comparative, and at the same time has a natural visualization effect, namely nodes and edges; however, the tree model is also difficult to handle linear relationships, lacks smoothness, and at the same time, as the depth of the tree increases, the number of leaf nodes will increase exponentially.
[0010] Decision Rule: The decision rule consists of a set of conditional judgments in the form of IF-THEN statements. The dataset passes through the judgments of the decision rule in sequence and is divided into different groups, which is similar to the effect of the decision tree. The interpretation of the decision rule is close to natural language and is easy to understand; however, the decision rule can hardly handle regression problems and at the same time requires a certain degree of classification in the dataset itself.
[0011] Generalized Additive Models (GAMs): Generalized additive models are an extension of linear models, aiming to capture the nonlinear relationships in the data while maintaining the interpretability of the model. Its internal linear combination of multiple variable sub-models results in a variety of variants, with extremely strong extensibility; but on the contrary, the more extensions are added, the further the interpretability of the model is reduced.
[0012] Post-hoc Interpretability: Post-hoc interpretability refers to the ability to interpret and analyze complex or "black-box" models (such as deep neural networks, random forests, etc.). Through various technical means, post-hoc interpretability tools and methods help users understand the decision-making logic of these complex models. Common post-hoc interpretability methods include SHAP, LIME, model visualization methods, surrogate models, etc.
[0013] SHAP (SHapley Additive exPlanations): The SHAP method is based on the concept of Shapley values in cooperative game theory and is used to explain the output of machine learning models. This is a method of fairly distributing benefits and is used in cooperative games to determine the contribution of each participant to the total benefit. The SHAP method has complete mathematical theory support, but has a high computational complexity and strong assumptions about data features.
[0014] LIME (Local Interpretable Model-agnostic Explanations): The core idea of the LIME method is to explain the predictions of complex models by constructing a simple and interpretable surrogate model in the local area of the model. LIME is a model-agnostic method, which means it can be applied to any type of machine learning model. At the same time, as a local explanation, the LIME method has flexible adjustment space. In contrast, the LIME method has high requirements for linear assumptions of the dataset, application neighborhood selection, etc.
[0015] Model Visualization: Model visualization enables users to intuitively observe the structure, training process, prediction results, and their influencing factors through graphics and visualization techniques. Different visualization methods can be adopted according to different models and interpretation requirements, such as displaying neural network architectures, activation maps, and feature maps.
[0016] Surrogate Models: Surrogate models are a technique for interpreting complex machine learning models. It approximates the behavior of complex models by training a simple and interpretable model on the output of complex models. Usually, intrinsically interpretable models can be directly selected as surrogate models, which makes the surrogate model method simple and allows for flexible selection of different models; however, there is likely to be information loss between the surrogate model and the original model, making it difficult to accurately explain all details and relationships.
[0017] The above are the existing common interpretability methods. The existing interpretability methods have various disadvantages, which are as follows:
[0018] First, intrinsically interpretable models usually require a relatively low level of internal structural complexity, which makes it difficult for their accuracy to compare with that of fully complex models such as neural networks.
[0019] Second, existing models have insufficient ability to capture the internal relationships in the dataset and lack interpretability; it is possible that high-order interactions of features are not captured, or misidentified interaction relationships increase the complexity of the model.
[0020] Third, some interpretable models and their corresponding improved variants attempt to make improvements, but a large number of computational parameters are generated during the operation process, or multiple iterations are required to obtain a better model, all of which greatly increase the training cost of the model and reduce the training efficiency.
[0021] In the field of fintech, machine learning models have a wide range of application scenarios, such as credit scoring, fraud detection, investment advice, risk management, etc. These scenarios all involve a large number of decisions and require transparent and interpretable models to ensure fairness, compliance, and user trust. At the same time, the transparency of interpretable models can also help users analyze data associations, obtain market information, manage potential risks, etc., or further improve the model's performance based on the discovered vulnerabilities. In summary, researching intrinsically interpretable models is not only beneficial for providing transparent, fair, and compliant model services, obtaining more accurate information and more reliable decisions, but also conducive to improving the machine learning models widely used in the financial field currently and providing a basis for subsequent model research. Summary of the Invention
[0022] The present invention is made to solve the above problems, and its purpose is to provide a method for constructing an interpretable model with better interpretability and lower training cost. The present invention adopts the following technical solutions:
[0023] The present invention provides a method for constructing an intrinsically interpretable model based on Gaussian processes in a fintech scenario. This method has the following technical features, which include the following steps: Step S1, constructing a Gaussian process kernel function using the random Fourier feature method; Step S2, calculating the Sobol index of the input dataset according to the functional variance analysis method; Step S3, using the Sobol index to screen the importance of features and interaction features of the input dataset, taking the Gaussian process kernel function constructed by the random Fourier feature method as the shape function, and constructing a generalized additive model based on the screened features and interaction features; Step S4, training the generalized additive model on the input dataset to learn the weight matrix, thereby obtaining an intrinsically interpretable neural additive model.
[0024] The method for constructing an intrinsically interpretable model based on Gaussian process in the fintech scenario provided by the present invention may further have the following technical features. Among them, step S1 includes the following sub-steps: Step S1-1, by minimizing the KL divergence between the standard Gaussian distribution and the input data, convert the continuous input features of the input data set into an approximate Gaussian density; Step S1-2, based on the orthogonal constraint, use the random Fourier feature method to construct the Gaussian process kernel function for each dimension of the input data set as the shape function of this dimension.
[0025] The method for constructing an intrinsically interpretable model based on Gaussian process in the fintech scenario provided by the present invention may further have the following technical features. Among them, in step S1-2, the shape function of each dimension satisfies the following constraints:
[0026]
[0027] In the formula, X i is the sample space, p i is the sample density. Under the said constraints, the Gaussian process kernel function is:
[0028]
[0029] Where
[0030] E[S i f i (·)] = ∫p i (x i )k i (x i , ·)dx i
[0031]
[0032] In the formula, E[S i f i (·)] and are the first-order constraint function and the second-order constraint function under the orthogonal constraint respectively, is the orthogonal constraint kernel.
[0033] The method for constructing an intrinsically interpretable model based on Gaussian process in the fintech scenario provided by the present invention may further have the following technical features. Among them, in step S2, the Sobol index associated with the input data set is:
[0034]
[0035]
[0036] Where:
[0037] where \(K(X,X)\) is the covariance of the input training set, and is the correlation coefficient of the \(u\)-th order interaction.
[0038] The method for constructing an intrinsically interpretable model based on Gaussian processes in the fintech scenario provided by the present invention may further have the following technical feature: in step S3, for each sample point of the input data set, a feature vector \(\varphi(x)\) is generated:
[0039]
[0040] where \(z\) i , \(c\) i are randomly generated parameters, and \(S\) is the feature dimension.
[0041] At this time, the kernel function calculation can be transformed into:
[0042] \(k(x,x ′ )=\varphi(x) T \varphi(x ′ ).
[0043] The method for constructing an intrinsically interpretable model based on Gaussian processes in the fintech scenario provided by the present invention may further have the following technical feature: in step S4, for the generalized additive model, a linear layer is used to fit the additive kernel, and training is performed on the input data set to obtain the intrinsically interpretable neural additive model.
[0044] Functions and effects of the invention
[0045] According to the method for constructing an intrinsically interpretable model based on Gaussian processes in the fintech scenario provided by the present invention, compared with various interpretable models in the prior art, the intrinsically interpretable model constructed by the method of the present invention has the following advantages in multiple aspects:
[0046] First, improve accuracy: Many current intrinsically interpretable models have low complexity and accuracy because they need to ensure the transparency inside the model to provide explanations, and it is difficult to compare with fully complex models. In the method of the present invention, improvements are made on the GAMs framework, and the high-order interaction capture ability is utilized to make the model accuracy approach the black-box model while retaining the good interpretability of the linear model.
[0047] Second, improve interpretability: In the past, the research on interpretable models was insufficient in differentiating high-order interactions. There were many cases where high-order interactions of features were not captured by many models, or the interaction relationships were misidentified, increasing the complexity of the models. To solve this problem, the method of the present invention combines the calculation method of the Sobol index on the GP model, uses a regularization algorithm to accurately distinguish feature interactions, and limits the interaction order to a lower level, providing more accurate and concise interpretation results and improving the interpretability of the model.
[0048] Third, reduce the training cost: Many existing interpretable models and improvement methods require a large number of calculation parameters or multiple iterative operations, greatly increasing the training cost and reducing the training efficiency. In the method of the present invention, by combining the posterior parameter characteristics of the GP model, the number of training parameters of the model is greatly reduced on the GAMs architecture, while also ensuring the training effect of the model and improving the training efficiency of the model. Brief Description of the Drawings
[0049] Figure 1 is a flowchart of an intrinsically interpretable method based on Gaussian process in a fintech scenario according to an embodiment of the present invention;
[0050] Figure 2 is a flowchart of step S1 according to an embodiment of the present invention;
[0051] Figure 3 is a flowchart of a construction method of an existing interpretable model in a comparative example of the present invention. Detailed Embodiments
[0052] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the following specifically describes the construction method of the intrinsically interpretable model based on Gaussian process in a fintech scenario of the present invention in combination with embodiments and drawings.
[0053] <Embodiment>
[0054] This embodiment provides a construction method of an intrinsically interpretable model based on Gaussian process in a fintech scenario. In a fintech scenario, common data sets have the following characteristics: First, the data sets are mostly tabular data sets, so linear model frameworks similar to GAM are often used for analysis and interpretation; Second, the data sets generally involve multi-dimensional information such as users, banks, and transactions, with a large number of feature dimensions, and there are relatively complex relationships between features, and high-order interaction analysis is required to obtain a complete and reliable interpretation. At the same time, most of the existing interpretation methods and models have very low computational efficiency and incomplete coverage of high-order interactions.
[0055] Therefore, based on the demand characteristics of model applications in the fintech scenario, in this embodiment, methods such as the GP model are adopted to optimize the computational complexity on the GAM linear framework, and utilize its posteriority, high precision and other characteristics to improve the training efficiency and interpretability of the model.
[0056] Figure 1 It is the flowchart of the method for constructing an intrinsically interpretable model based on Gaussian process in the fintech scenario in this embodiment.
[0057] As Figure 1 shown, the method for constructing an intrinsically interpretable model based on Gaussian process in the fintech scenario includes the following steps:
[0058] Step S1, construct a Gaussian process kernel function (GP kernel function) using the random Fourier feature method (RFF) to be used as the shape function of the generalized additive model (GAM).
[0059] Step S2, calculate the Sobol index of the input data set according to the functional analysis of variance (FANOVA) method.
[0060] Step S3, use the Sobol index to screen the importance of features and interaction features, use the GP kernel function constructed by the above RFF method as the shape function, and construct a GAM model based on the screened features and interaction features.
[0061] Step S4, for the GAM model, use a linear layer to fit the additive kernel, and train on the input data set to learn the weight matrix, so as to obtain an intrinsically interpretable neural additive model.
[0062] The above steps will be described in detail below.
[0063] Step S1, construct a GP kernel function using the RFF method, which is used as the shape function of the GAM model.
[0064] Figure 2 It is the flowchart of step S1 in this embodiment.
[0065] As Figure 2 shown, step S1 specifically includes the following sub-steps:
[0066] Step S1-1, by minimizing the KL divergence between the standard Gaussian distribution and the input data, transform the continuous input features of the input data set into an approximate Gaussian density to meet the data distribution assumption of the GP model.
[0067] Step S1-2, based on the orthogonality constraint, use the random Fourier feature method (RFF) to construct a GP kernel function for each dimension of the input data set as the shape function of this dimension.
[0068] Specifically, for the shape function of each dimension, it is desired that it satisfies the following constraints:
[0069]
[0070] In the formula, X i is the sample space, p i is the sample density, and this constraint represents that the integral of the function f i with respect to the input is 0. Under this constraint, the kernel function is:
[0071]
[0072] Where
[0073] E[S i f i (·)] = ∫p i (x i )k i (x i ,·)dx i
[0074]
[0075] In the formula, E[S i f i (·)] and are the first-order constraint function and the second-order constraint function under the orthogonal constraint respectively, is the constrained kernel.
[0076] Step S2, calculate the Sobol index of the input data set according to the Functional Analysis of Variance (FANOVA) method as the quantization value of the feature importance.
[0077] Among them, the Sobol index associated with the input data set u is:
[0078]
[0079] Where:
[0080]
[0081] In the formula, K(X,X) is the covariance of the input training set, is the correlation coefficient of the u-th order interaction.
[0082] The calculated Sobol index includes the results of single and second-order feature interactions. This index serves as both the basis for subsequent feature screening in the model and the interpretation result of the feature importance in the model.
[0083] Step S3: Use the Sobol index to screen the importance of features and interaction features, take the GP kernel function constructed based on the RFF method as the shape function above, and construct a GAM model based on the screened features and interaction features.
[0084] Among them, use the Sobol index to screen the importance of features and interaction features, so as to limit the features and interaction features of the input data set.
[0085] Specifically, for each sample point of the input data set, generate a feature vector φ(x):
[0086]
[0087] In the formula, z i , c i are randomly generated parameters, and S is the feature dimension.
[0088] At this time, the kernel function calculation can be transformed into:
[0089] k(x, x ′ ) = φ(x) T φ(x ′ )
[0090] Step S4: For the GAM model, use a linear layer to fit the additive kernel, and train on the input data set to learn the weight matrix, so as to obtain an intrinsically interpretable neural additive model.
[0091] Functions and effects of the embodiment
[0092] According to the intrinsically interpretable method based on Gaussian process in the fintech scenario provided by this embodiment, compared with various interpretable models in the prior art, the intrinsically interpretable model constructed by the method of this embodiment has the following advantages in many aspects:
[0093] First, improve accuracy: Many current intrinsically interpretable models have low complexity and accuracy because they need to ensure the transparency inside the model to provide explanations, and it is difficult to compare with full-complexity models. In the method of this embodiment, improvements are made on the GAMs framework, and the high-order interaction capture ability is used to make the model accuracy approach the black-box model while retaining the good interpretability of the linear model.
[0094] Second, improve interpretability: In the past, the research on interpretable models was insufficient in differentiating high-order interactions. In many models, high-order interactions of features were not captured, or the interaction relationships were misidentified, increasing the model complexity. To solve this problem, the method of this embodiment combines the Sobol index calculation method on the GP model, uses the regularization algorithm to accurately distinguish feature interactions, and limits the interaction order to a lower level, providing more accurate and concise interpretation results and improving the interpretability of the model.
[0095] Third, reduce the training cost: Many existing interpretable models and improvement methods require a large number of computational parameters or multiple iterative operations, greatly increasing the training cost and reducing the training efficiency. In the method of this embodiment, by combining the posterior parameter characteristics of the GP model, the number of training parameters of the model is greatly reduced on the GAMs architecture, while ensuring the training effect of the model and improving the training efficiency of the model.
[0096] <Comparative Example 1>
[0097] This comparative example provides two interpretable models based on the GAM architecture in the prior art and their construction methods for comparing the effects with the models and methods provided in the embodiment. The two existing models for comparison are NAM and NBM. The NAM model is the basic model, and the NBM model is the latest interpretable model with the GAM architecture. Both use neural networks as shape functions.
[0098] For existing models based on the GAM architecture, the existing corresponding interpretable model construction methods attempt to construct the following model:
[0099] g(E[y]) = ∑f i (x i ) +... + ∑f ij (x i , x j )
[0100] Compared with the simple linear regression model, the GAM model transforms the sparse matrix into a shape function matrix, that is, the f n function, to learn the non-linear relationship of the data set, and at the same time adds n second-order feature interaction terms to improve the performance of the model and make its performance approach that of the black box model.
[0101] Figure 3 is the flow chart of the construction method of the existing interpretable model in this comparative example.
[0102] As Figure 3 shown, the construction methods of the two models for comparison are similar and include the following steps:
[0103] Step R1: Select a suitable shape function for constructing the GAM model and construct the GAM model based on the selected shape function.
[0104] The key to constructing the GAM is to control the complexity and variance of the shape function to obtain better interpretability and accuracy.
[0105] For the EBM model, select the boosted bagged tree model based on gradient training as the shape function, and at the same time limit the number of leaf nodes of the tree model to 2 - 4 to obtain lower complexity and training cost.
[0106] For the NAM model, select the single - layer neural network model directly as the shape function, and use the model visualization technology to interpret the simple single - layer neural network shape as a feature.
[0107] Step R2: Train the GAM model. During the training process, enable the model to learn high - order feature interaction relationships through the high - order feature capture algorithm.
[0108] Whether the model can learn the relationships between features is an important feature that differentiates the full - complexity model from the simple model. Adding the high - order feature capture algorithm can make the model performance approach that of the full - complexity model while retaining interpretability.
[0109] The EBM model maintains two datasets S and Z, where the Z dataset is all possible pairwise feature interactions; in each iteration, based on the existing best model, select the best second - order interaction feature from Z and add it to the S set until the model performance no longer improves, and at the same time use pre - calculation to simplify the iteration process.
[0110] The NAM model directly constructs redundant neural networks to fit the relationships between multiple pairs of features, which is still interpretable and visualizable when the interaction order is relatively low. At the same time, to reduce the training cost, use the Pearson correlation coefficient, etc. to construct a regularization penalty to reduce the number of interactions.
[0111] In step R1 above, an additive sub - function was constructed, and on this basis, the GAM model was constructed; in step R2, the model learned high - order feature interaction relationships, which is the key to improving the model performance. In the construction process of existing interpretable models, similar to step R2 above, generally, interaction features are calculated during the model training process, and after the training is completed, the importance quantification values of the first - order and second - order interaction features are obtained. This requires traversing all high - order features during the model training, resulting in a high training cost.
[0112] Although there are also some optimization methods in the existing technology that attempt to change the interaction selection algorithm of the model in step R2, usually, such methods also require iterative confirmation of hyperparameters, increasing the model training parameters and training cost, so the effect is very limited.
[0113] Compared with step R1 of the above-mentioned existing method, in step S1 of the method of the embodiment, a GP model implemented based on the RFF method is selected as the shape function.
[0114] As shown in step R2 of the above-mentioned existing method, the interpretable model of the traditional GAM architecture screens interaction features during the training process. In contrast, in steps S2 to S4 of the method of the embodiment, the two processes of interaction feature screening and model training are separated, and first, interaction feature capture is processed, then model construction is performed, and then model training is carried out. Therefore, it is no longer necessary to traverse all high-order features during the model training process.
[0115] It can be seen by comparison that the method of the embodiment innovates in both steps of the construction method of the existing GAM architecture model to comprehensively improve the performance of the intrinsic interpretable model and reduce the model training cost.
[0116] In addition, for the above-mentioned NAM model, NBM model, and GP-NAM model constructed by the method of the embodiment, two identical datasets (CA Housing and LCD) were used to train the three models respectively. Table 1 below shows the internal parameter quantities of each model when training these three models using different datasets:
[0117] Table 1 Internal parameter quantity table when different interpretable models are trained
[0118] Model CA Housing LCD NAM 83K 32K NBM 64K 63K GP-NAM 1201 501
[0119] As shown by the data comparison in Table 1, the GP-NAM model of the embodiment applying the GP kernel is lighter, faster, and has a lower training cost.
[0120] <Comparative Example 2>
[0121] This comparative example provides six interpretable models constructed by existing methods for comparing the effects with the GP-NAM model constructed by the method of the embodiment.
[0122] The six interpretable models are Linear, EBM, NAM, NODE-GAM, and NBM. These interpretable models were trained using the same multiple datasets, and the datasets used included CA Housing, Loan Data, LCD, Boston Housing, and German Credit. The accuracy performance of each interpretable model on each dataset was evaluated.
[0123] Table 2 below shows the accuracy performance evaluation of each interpretable model on different datasets.
[0124] Table 2 Accuracy performance evaluation table of different interpretable models
[0125]
[0126] As shown in Table 2, the values in the table vary according to different data sets. For the regression data set, the root mean square error (RMSE) is used as the evaluation index, and for the classification data set, the area under the curve (AUC) is used as the evaluation index. Through data comparison, it can be seen that compared with the existing six interpretable models, the GP-NAM model constructed by the method of the embodiment has significantly higher accuracy performance on both types of data sets, that is, the method of the embodiment can achieve the improvement of the performance of the interpretable model. Among them, the improvement effect on the regression data set is the most obvious.
[0127] Moreover, the method of the embodiment still achieves a very high precision when the feature order is limited to 2, which is sufficient to show that the method of the embodiment can enable the model to fully learn the correct feature relationships and improve the interpretability of the model.
[0128] As described above, through two comparative examples, the specific advantages of the method of the embodiment are as follows: First, the constructed interpretable model has more accurate explanations and higher interpretability, especially the improvement on the regression data set. Second, compared with other complex interpretable models, the training time is greatly shortened and the training efficiency is significantly improved.
[0129] The above embodiments are only used to illustrate the specific implementation manners of the present invention, and the present invention is not limited to the description scope of the above embodiments. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.
Claims
1. A method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario, characterized in that: The following steps are involved: Step S1, constructing a Gaussian process kernel function using a random Fourier feature method; Step S2, calculating the Sobol index of the input data set according to the functional variance analysis method; Step S3, using the Sobol index to screen the importance of the features and interactive features of the input data set, using the Gaussian process kernel function constructed by the random Fourier feature method as a shape function, and constructing a generalized additive model based on the screened features and interactive features; Step S4, training the generalized additive model on the input data set to learn the weight matrix, thereby obtaining an intrinsically interpretable neural additive model.
2. The method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario according to claim 1, characterized in that: in, Step S1 includes the following sub-steps: Step S1-1, converting the continuous input features of the input data set into an approximate Gaussian density by minimizing the KL divergence between the standard Gaussian distribution and the input data; Step S1-2, based on the orthogonal constraint, using the random Fourier feature method to construct the Gaussian process kernel function for each dimension of the input data set as the shape function of the dimension.
3. The method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario according to claim 2, characterized in that: in, In step S1-2, the shape function of each dimension satisfies the following constraints: Where, X i is the sample space, p i is the sample density, Under the constraints, the Gaussian process kernel function is: in, E[S i f i (·)]=∫p i (x i )k i (x i ,·)dx i In the formula, are the first-order constraint function and the second-order constraint function under orthogonal constraints, respectively. is the orthogonal constraint kernel.
4. The method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario according to claim 3, characterized in that: in, In step S2, the Sobol index associated with the input data set is: in: Where K(X,X) is the input training set covariance, is the correlation coefficient of the u-th order interaction.
5. The method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario according to claim 4, characterized in that: in, In step S3, for each sample point of the input data set, a feature vector φ(x) is generated: In the formula, z i 、c i is a randomly generated parameter, S is the feature dimension, At this point, the kernel function calculation can be transformed into: k(x,x ′ )=φ(x) T φ(x ′ )。 6. The method for constructing an intrinsically interpretable model based on Gaussian process in a financial technology scenario according to claim 1, characterized in that: in, In step S4, for the generalized additive model, a linear layer is used to fit the additive kernel and training is performed on the input data set to obtain the intrinsically interpretable neural additive model.