A model interpretable method for detecting interactive features in the field of financial risk control

Abstract

The present invention discloses a method for model interpretability of detecting interaction features in the field of financial risk control, which relates to the field of graph neural network technology and model interpretability. The method comprises the following steps: Step 1, training an additive model according to existing training features and training labels; Step 2, using an interaction feature detection module to detect existing interaction feature pairs, reconstructing a GAM model using the obtained interaction feature pairs, the training labels and the residuals of Step 1, and then adding the additive model and the GAM model to obtain a GAM<supgt;2< / supgt; model containing the interaction feature pairs; Step 3, according to the principle that a high-order interaction exists if and only if all its low-order interactions exist, performing multiple rounds of iteration on the GAM<supgt;2< / supgt> model until the order of interaction features no longer increases in a certain round of iteration, to obtain a GAM<supgt;n< / supgt> model; Step 4, realizing visualization and report export.

Description

Technical Field

[0001] The present invention relates to the field of graph neural network technology and model interpretability, and in particular to a model interpretability method for detecting interactive features in the field of financial risk control. Background Art

[0002] In recent years, the new generation of artificial intelligence technology, with machine learning, especially deep learning, as its guide, has been developing in a more advanced, complex and autonomous direction, bringing new opportunities for change to economic and social development. The application of AI has brought about a "species explosion", which is increasingly penetrating into all walks of life and all aspects of human life, and is expected to shape a new economic and social form. At the same time, scientific and technological ethics has increasingly become a "must-have" in the current development of AI technology and industrial applications, and all walks of life have explored the ethical principles, frameworks and governance mechanisms of AI.

[0003] A core issue in science and technology ethics is AI transparency and explainability. In November 2021, the United Nations Educational, Scientific and Cultural Organization (UNESCO) adopted the first global AI ethics agreement, the Recommendation on the ethics of artificial intelligence. The ten AI principles proposed include "transparency and explainability", which means that the working methods of algorithms and the data used to train them should be transparent and explainable.

[0004] Although not all AI systems are "black box algorithms" and are no more unexplainable than non-AI technologies, traditional software, or manual programs, machine learning models, especially deep learning models, are currently opaque and difficult for humans to understand. In the future, continued progress in AI is expected to bring autonomous systems that perceive, learn, make decisions, and act. However, the actual utility of these systems is limited by whether the machines can fully explain their thoughts and actions to human users. If users want to understand, trust, and effectively manage the new generation of artificial intelligence partners, the transparency and explainability of AI systems are crucial. Therefore, in recent years, Explainable Artificial Intelligence (XAI) has become an emerging field of AI research, and academia and industry have explored methods and tools to understand the behavior of AI systems. In particular, explainability is particularly critical in the fields of finance and medicine.

[0005] The rapid development of financial technology on the Internet has enabled people to easily access financial services provided by banks, including funds, deposits, transfers, etc. When it comes to lending business, customers can also obtain certain loans at a lower threshold. Therefore, while financial services are becoming more convenient, the fraud methods that arise with them are also rapidly upgrading and iterating, and the intelligence, virtualization, and concealment of fraud are increasing. From the earliest offline financial fraud to purposeful, organized group financial fraud, banks need to continuously iterate and update risk control systems to reduce bad debt losses.

[0006] In order to minimize bad debt losses, it is necessary to introduce a more accurate model. However, a more accurate model means higher model complexity and worse interpretability, that is, it is more difficult to give a reasonable explanation when judging a bad customer. In other words, business personnel cannot understand why the model rejects a customer, which is a relatively unacceptable problem for banks. In response to the above situation, the People's Bank of China issued the "Evaluation Specifications for Financial Applications of Artificial Intelligence Algorithms" in March 2021, which clearly stipulates the interpretability evaluation part. The specified features must meet the following interpretability requirements:

[0007] 1. Feature definition should meet relevant business logic and rules.

[0008] 2. Feature definitions should be clearly recorded in the system.

[0009] 3. The feature definition should have detailed records of the index processing process of data extraction, transformation, loading (ETL).

[0010] The following requirements must be met for feature derivation:

[0011] 1. Feature derivation should be reasonable. For capital scenarios, only discrete feature crossover is allowed, and complex feature crossover with no business significance is not allowed.

[0012] 2. Based on the business feature derivation, there should be a detailed ETL feature derivation processing process.

[0013] 3. Feature derivation based on the algorithm needs to be able to demonstrate the process and logic of algorithm feature derivation.

[0014] Therefore, if the model's algorithm is not interpretable, the choice of model will be greatly limited.

[0015] Existing methods are mainly used to explain the role of deep learning in the field of image analysis, and the interpretation accuracy is usually low. For applications with fuzzy boundaries, such as image recognition, relatively low interpretation accuracy is acceptable. However, for security applications, even a deviation in the interpretation of one byte can lead to serious misunderstandings or errors, and relatively low interpretation accuracy becomes unacceptable.

[0016] In addition, current interpretability methods are greatly limited by factors such as algorithms, model structures, and application scenarios. Although they can be used to explain the behavioral decisions and prediction results of deep learning models, they may not work properly in the following situations:

[0017] 1. If the model is an interactive model, such as a random forest, the current interpretability still cannot achieve real-time interpretation, so the model that explains the interactive modeling needs to be studied.

[0018] 2. Whether features are correlated with each other and the interaction between features will greatly increase the difficulty of model interpretation. It is necessary not only to consider the significance of the features, but also to evaluate the impact of the correlation between different features on model decisions.

[0019] 3. If the model does not correctly model causal relationships, since interpretable methods directly explain the model, there is a lack of prior testing of the correctness of the modeling.

[0020] 4. If the parameters of the interpretation method are not set correctly, the parameter settings will affect the interpretation results, and the stability of the parameters will directly affect the credibility of the interpretability.

[0021] Therefore, technical personnel in this field are committed to developing a model interpretable method for detecting interactive features in the field of financial risk control, to make up for the lack of interpretable and interactive feature detection solutions for AI models in the financial field in the prior art. Summary of the invention

[0022] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to optimize the existing model interpretation algorithms, such as Local Interpretable Model-Agnostic Explanations (LIME), SHAP (Shapley Additive exPlanation), Parzen (Parzen Window), Partial Dependency Plot (PDP), etc., so that the existing model interpretation algorithms can provide personalized interpretation results with interactive features according to the relevance of features, and maintain the same accuracy as the tree model. An example of an interactive feature is as follows:

[0023]

[0024] The mainstream interpretable algorithms for tabular data in the field of financial risk control include LIME, SHAP, Explainable Boosting Machine (EBM), etc. Most of these interpretable algorithms can only provide interpretation results for a single feature, but cannot detect the interaction between them. EBM uses GAM 2 The algorithm can detect the feature interaction between two pairs, and the comparison chart is as follows Figure 1 (LIME) and Figure 2 As shown in (EBM), the explanation result of EBM includes the explanation items of feature1×feature2, which are the interactive feature explanation items in the provided explanation results.

[0025] To achieve the above object, the present invention provides a model interpretable method for detecting interactive features in the field of financial risk control, comprising the following steps:

[0026] Step 1: Train an additive model based on the existing training features and training labels. The classifier of the additive model is represented by F(x):

[0027] F(x)=∑f i (x i )

[0028] where F(x)∈[0,1], x i is the training feature or the training label, f i (x i ) for each x i Contribution to the overall forecast results;

[0029] Step 2: Use the interactive feature detection module to detect the existing interactive feature pairs, use the obtained interactive feature pairs, the training labels and the residuals of step 1 to reconstruct a GAM model, and then add the additive model to the GAM model to obtain a GAM containing the interactive feature pairs. 2 Model, the GAM 2 The classifier of the model is denoted as F 2 (x):

[0030]

[0031] Among them, F 2 (x)∈[0,1],E n is a set of the interaction feature pairs detected by the interaction feature detection module, {i, j} represents any of the interaction feature pairs, and f ij For the interaction feature pair {i,j}∈En About x in i 、x j The transformation function of

[0032] Step 3: According to the principle that a high-order interaction exists if and only if all its low-order interactions exist, 2 The model is iterated for multiple rounds until the order of the interactive features in a certain round of iterations no longer increases, and a GAM is obtained. n Model, the GAM n The model contains a classifier of order n interactions, denoted as F n (x):

[0033]

[0034] Among them, F n (x)∈[0,1], the highest order of n-order interaction is n, (x i ,x j ,...x n ) represents a set of high-dimensional interactive features;

[0035] Step 4: Realize visualization and report export.

[0036] Furthermore, the interactive feature detection module in step 2 is based on a graph neural network.

[0037] Furthermore, the graph neural network in the interactive feature detection module is a L0-SIGN model, including a L0-edge detection model and a SIGN model.

[0038] Furthermore, the input of the L0-SIGN model is a graph without edge information, wherein each of the training features is a node X n , the interactions between the training features correspond to the edges E n , that is, a data sample n is a graph represented as:

[0039] G n (X n ,E n )

[0040] in,

[0041]

[0042] (e n ) ij ∈{1,0}

[0043] Among them, 1 means that there is an edge between {i, j}, and 0 means that there is no interaction between {i, j}.

[0044] Furthermore, during the prediction process, the L0-edge detection model F ep (X n ; ω) is used to analyze whether the edge in the graph exists, where ω is F ep The output is an edge set E′ n .

[0045] Furthermore, the L0-edge detection model is a separate detection module that uses a matrix decomposition method.

[0046] Furthermore, the SIGN model is a graph classifier that provides n (X n ,E′ n ) is used to predict the results, and then L0 regularization is used to limit the number of the detected interaction feature pairs.

[0047] Furthermore, the L0-SIGN model interactively models the initial nodes connected by edges, and then updates the representation of each of the initial nodes by aggregating all corresponding modeling results. Finally, all updated representations of the initial nodes are aggregated to obtain a final prediction.

[0048] Furthermore, the SIGN model will obtain a prediction function f S , used to predict the classification results, and then use the L0 regularization in the process of fitting parameters by the stochastic gradient descent iterative method to obtain the final prediction function f for predicting the classification results LS , the general form of the SIGN model prediction function is:

[0049]

[0050] Among them, the prediction function f S is a prediction function without regularization, the prediction function f LS is the prediction function f after the L0 regularization process S , θ is the parameter set in the graph neural network.

[0051] Furthermore, the initial input does not require side information.

[0052] The model interpretable method for detecting interactive features in the field of financial risk control provided by the present invention has at least the following technical effects:

[0053] 1. The technical solution provided by the present invention proposes an interactive feature detection module based on a graph neural network, which can be embedded in the network to obtain interactive feature pairs with higher credibility, solving the problem that the current model has a high interpretation accuracy in the interpretable field but cannot obtain an interpretation with effective interactive features;

[0054] 2. The technical solution provided by the present invention improves the existing EBM model and embeds an interactive feature detection module based on a graph neural network, thereby solving the overfitting problem of the GAM linear model due to its simple structure and improving the credibility and effectiveness of the model interpretation;

[0055] 3. The technical solution provided by the present invention extends the EBM model to identify high-dimensional feature interactions, so that high-precision black box models such as deep learning can be introduced into fields such as financial risk control that require high interpretability.

[0056] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is the explanation result graph of a single sample explained by the LIME algorithm;

[0058] Figure 2 It is the explanation result diagram of a single sample explained by the EBM algorithm;

[0059] Figure 3 It is a schematic diagram of the interactive feature detection network L0-IGN algorithm. DETAILED DESCRIPTION

[0060] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0061] The present invention provides a model interpretable method for detecting interactive features in the field of financial risk control. It is a white box modeling method that can provide personalized explanations for samples with the same accuracy as the tree model. That is, the model itself is interpretable and is mainly divided into two parts: an interactive feature detection module based on graph neural network and a GAM-based interactive feature detection module. N The model fitting method.

[0062] 1. Interaction feature detection module based on graph neural network:

[0063] Graph Neural Networks (GNNs) can facilitate learning entities and their overall relationships. Existing work uses GNNs to perform relational reasoning in various fields, while in the embodiments of the present invention, relational reasoning is used for interaction feature detection. The graph neural network used in the embodiments of the present invention is L0-SIGN (L0-Norm Statistical Interaction Graph Neural Network), in which each data sample is regarded as a graph, the features are nodes, and the interactions between features are edges. The L0-SIGN neural network can detect the interaction feature pairs with the greatest contribution, and then use the Top-K interaction features as the next GAM N Part of the input to the model.

[0064] Compared with the EBM model, which calculates the strength of all interaction features between each pair, the interactive feature detection module based on the graph neural network only uses the detected interactive features as input. In addition, EBM only captures the contribution of each individual feature interaction to the prediction, but not the overall contribution of a set of interactive feature pairs.

[0065] 2. GAM N Model fitting:

[0066] The EBM model uses a two-step method to calculate the contribution strength of the interaction features to the overall model. In an embodiment of the present invention, certain improvements are made to the two-step calculation method of the basic EBM model. The second step of calculating the interaction feature strength no longer uses the Fast Interaction Detection (FAST) algorithm of the EBM model, but instead integrates an interaction feature detection module based on a graph neural network.

[0067] At the same time, the GAM of the EBM model 2 Make some improvements and map the 2nd-order feature interactions to a higher order based on existing theoretical derivations.

[0068] 3. Visualization and report export module:

[0069] The local explainable module and the global explanation module are integrated in the same visualization window. Users can query the corresponding prediction results and explanation results by entering the ID number corresponding to each piece of data. At the same time, the explanation report of each customer can be exported with one click through the export function of the visualization window, indicating the suspicious points in the customer portrait.

[0070] In the field of bank risk control, the current technical difficulty and time cost of detecting customer fraud risk, credit risk, etc. are both high, and the existing models will have certain deviations. Manual review requires in-depth exploration of the original data of each customer and the discovery of abnormal points in the customer data, which is a considerable test for the energy and professionalism of practitioners. The technical solution provided by the present invention aims to reduce the manual review cost of practitioners, reduce the possibility of model misjudgment, and provide certain guidance for subsequent feature engineering.

[0071] An embodiment of the present invention provides a model interpretable method for detecting interactive features in the field of financial risk control, comprising the following steps:

[0072] Step 1: Train an additive model based on the existing training features and training labels. The classifier of the additive model is represented as F(x):

[0073] F(x)=∑f i (x i )

[0074] where F(x)∈[0,1], x i is the training feature or training label, f i (x i ) for each x i Contribution to the overall forecast results;

[0075] Step 2: Use the interactive feature detection module to detect the existing interactive feature pairs, use the obtained interactive feature pairs, training labels and the residuals of step 1 to construct a GAM model, and then add the additive model and the GAM model to obtain a GAM containing the interactive feature pairs. 2 Model, GAM 2 The classifier of the model is denoted as F 2 (x):

[0076]

[0077] Among them, F 2 (x)∈[0,1],E n is the set of interactive feature pairs detected by the interactive feature detection module, {i, j} represents any interactive feature pair, and f ij For the interactive feature pair {i,j}∈E n About x in i 、x j The transformation function of

[0078] Step 3: According to the principle that a high-order interaction exists if and only if all its low-order interactions exist, GAM 2 The model is iterated for multiple rounds until the order of the interactive features in a certain round of iterations no longer increases, and a GAM is obtained.n Model, GAM n The model contains a classifier of order n interactions, denoted as F n (x):

[0079]

[0080] Among them, F n (x)∈[0,1], the highest order of n-order interaction is n, (x i ,x j ,...x n ) represents a set of high-dimensional interactive features;

[0081] The principle of "a high-order interaction exists if and only if all its low-order interactions exist" is specifically: x1x2x3 exists if and only if its subsets {x1x2,x2x3,x1x3} all exist, and it is extended to high-dimensional space. However, due to the high dimension, its subsets are O(2 n ), so the requirements for interaction will increase with each extension to a higher dimension. For example, to extend the interaction to 4-dimensional space, it is necessary to The existence of all subsets will make the algorithm provided by the present invention not detect very high-dimensional feature interactions, but ensure the existence of important interactions.

[0082] Step 4: Realize visualization and report export.

[0083] The front-end and back-end are constructed using JAVA, and the interpretation operation results are rewritten based on the EBM part of the Microsoft open source framework Interpret in Python, embedded in the interactive feature detection module of the graph neural network, and finally realize the function of personalized query interpretation results.

[0084] Among them, the interactive feature detection module in step 2 is based on graph neural network.

[0085] The graph neural network in the interactive feature detection module is a L0-SIGN model, which includes an L0-edge detection model and a SIGN model, such as Figure 3 shown.

[0086] The input of the L0-SIGN model is a graph without edge information, where each training feature is a node X n , the interaction between the training features corresponds to each edge E n , that is, a data sample n is a graph represented as:

[0087] G n (X n ,E n )

[0088] in,

[0089]

[0090] ( n ) ij ∈{1,0}

[0091] Among them, 1 means there is an edge between {,j}, and 0 means there is no interaction between {i,j}. The initial input does not require edge information.

[0092] In the prediction process, the L0-edge detection model F ep (X n ; ω) is used to analyze whether the edge in the graph exists, where ω is F ep The output is an edge set E′ n .

[0093] The L0-edge detection model is a separate detection module that uses the matrix factorization (MF) method.

[0094] The SIGN model is a graph classifier that provides n (X n ,E′ n ) and then use L0 regularization to limit the number of detected interactive feature pairs.

[0095] The L0-SIGN model interactively models the initial nodes connected by edges, then updates the representation of each initial node by aggregating all corresponding modeling results, and finally aggregates all updated initial node representations to obtain the final prediction.

[0096] The SIGN model will get a prediction function f S , used to predict the classification results, and then use L0 regularization in the process of fitting parameters by stochastic gradient descent iteration method to obtain the final prediction function f for predicting the classification results LS , the general form of the SIGN model prediction function is:

[0097]

[0098] Among them, the prediction function f S is the prediction function without regularization, the prediction function f LS is the prediction function f after L0 regularization S , θ is the parameter set in the graph neural network.

[0099] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A model interpretable method for detecting interactive features in the field of financial risk control, characterized in that: The method comprises the following steps: Step 1: Train an additive model based on the existing training features and training labels. The classifier of the additive model is expressed as : in, , is the training feature or the training label, For each Contribution to the overall forecast result; Step 2: Use the interactive feature detection module to detect the existing interactive feature pairs, and use the obtained interactive feature pairs, the training labels and the residuals of step 1 to reconstruct a model, and then the additive model and the Models are added to obtain a pair containing the interaction features. Model, The model’s classifier is represented as : in, , is a set of the interaction feature pairs detected by the interaction feature detection module, represents any of the interaction feature pairs, For the interaction feature pair About The transformation function of Step 3: According to the principle that a high-order interaction exists if and only if all its low-order interactions exist, The model is iterated for multiple rounds until the order of the interactive features in a certain round of iterations no longer increases, and a Model, Model contains The classifier of order interaction is expressed as : in, , The highest order of interaction is , Represents a set of high-dimensional interacting features; Step 4: Realize visualization and report export; The interactive feature detection module in step 2 is based on a graph neural network; The graph neural network in the interactive feature detection module is Model, including a -Edge detection model with a Model; Said The input of the model is a graph without edge information, where each training feature is a node , the interactions between the training features correspond to the edges , that is, a data sample Is a graph represented as: in, Among them, 1 means There is an edge between them, 0 means There is no interaction between them; In the prediction process, the -Edge Detection Model It is used to analyze whether the edge in the graph exists, where for The parameter is output as an edge set ; Said The model is a graph classifier that provides The prediction results are then Regularization, limiting the number of interacting feature pairs detected; Said Interaction modeling is performed on the initial nodes connected by the edges, and then the representation of each initial node is updated by aggregating all corresponding modeling results, and finally, all updated representations of the initial nodes are aggregated to obtain the final prediction; Said The model will get a prediction function , used to predict the classification results, and then used in the process of fitting parameters by stochastic gradient descent iterative method Regularization obtains the final prediction function for predicting classification results , The model prediction function is of the form: Among them, the prediction function is a prediction function without regularization. For the The prediction function of the regularization process , is the parameter set in the graph neural network.

2. The model interpretable method for detecting interactive features in the field of financial risk control according to claim 1, characterized in that: Said -The edge detection model is a separate detection module that uses a matrix factorization approach.

3. The model interpretable method for detecting interactive features in the field of financial risk control according to claim 1, characterized in that: The initial input does not require side information, .

Images