A method and device for creating a risk control model

Abstract

The application discloses a method and device for creating a risk control model, and relates to the technical field of credit risk control. The main technical scheme of the application is as follows: sample data of each target sample in a target sample set is acquired, the sample data includes category type feature data and numerical type feature data; the category type feature data and the numerical type feature data are processed into a first multi-dimensional feature data set, the first multi-dimensional feature data set is composed of a plurality of first multi-dimensional numerical type feature data after a feature correlation relationship has been constructed; the category type feature data and the numerical type feature data are processed into a second multi-dimensional feature data set; a third model is trained by using the first multi-dimensional feature data set and the second multi-dimensional feature data set, the third model is a logistic regression risk control model for predicting a scoring result of the target sample. The application is used for creating a risk control model.

Description

Technical Field

[0001] This invention relates to the field of credit risk control technology, and in particular to a method and apparatus for creating a risk control model. Background Technology

[0002] Risk control technology plays a crucial role in the field of credit risk control. Currently, the industry practice is to build machine learning models to assess the risk of individual or corporate users. For machine learning risk control models, the mainstream approaches in the industry are to use three types of algorithms: logistic regression, decision tree-based algorithms, and deep learning-based algorithms.

[0003] Since the associated data of individual or enterprise users contains both categorical and numerical feature data, existing technologies generally employ one-hot encoding algorithms and mapping encoding algorithms (Woe encoding, Badrate encoding, etc.) to process the categorical feature data, converting it into numerical feature data. This numerical feature data is then used as training samples to build a risk control model. However, using one-hot encoding algorithms converts numerical feature data into high-dimensional sparse feature data, which is unsuitable for ensemble tree models and deep neural network models with better performance. This can easily lead to overfitting, limiting the use to simple logistic regression risk control models and resulting in poor model training performance. Using mapping encoding algorithms, on the other hand, results in the loss of some feature information, also leading to poor model training performance and impacting the accuracy of subsequent evaluations. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method and apparatus for creating a risk control model, the main purpose of which is to solve the problem of poor model training effect when there is categorical feature data, so as to ensure the accuracy of subsequent model evaluation.

[0005] To solve the above-mentioned technical problems, the present invention proposes the following solution:

[0006] In a first aspect, the present invention provides a method for creating a risk control model, the method comprising:

[0007] Obtain sample data for each target sample in the target sample set, wherein the sample data includes categorical feature data and numerical feature data;

[0008] The categorical feature data and the numerical feature data are processed into a first multidimensional feature dataset, which is composed of multiple first multidimensional feature data after feature association has been established;

[0009] The categorical feature data and the numerical feature data are processed into a second multidimensional feature dataset, which is composed of multiple second multidimensional feature data after the high-dimensional feature mapping has been constructed.

[0010] A third model is trained using the first multidimensional feature dataset and the second multidimensional feature dataset. The third model is a logistic regression risk control model used to determine the scoring results of the target sample.

[0011] Secondly, the present invention provides a risk control model creation apparatus, the apparatus comprising:

[0012] The acquisition unit is used to acquire sample data of each target sample in the target sample set, wherein the sample data includes categorical feature data and numerical feature data;

[0013] The first processing unit is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit into a first multidimensional feature dataset, wherein the first multidimensional feature dataset is composed of multiple first multidimensional feature data after feature association has been constructed.

[0014] The second processing unit is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit into a second multidimensional feature dataset. The second multidimensional feature dataset is composed of multiple second multidimensional feature data after the feature high-dimensional mapping has been constructed.

[0015] The training unit is used to train a third model using the first multidimensional feature dataset obtained by the first processing unit and the second multidimensional feature dataset obtained by the second processing unit. The third model is a logistic regression risk control model used to predict the scoring results of the target sample.

[0016] To achieve the above objectives, according to a third aspect of the present invention, a storage medium is provided, the storage medium including a stored program, wherein, when the program is executed, the device where the storage medium is located executes the method for creating the risk control model of the first aspect.

[0017] To achieve the above objectives, according to a fourth aspect of the present invention, a processor is provided for running a program, wherein the program executes the method for creating the risk control model of the first aspect.

[0018] By means of the above technical solution, the present invention provides a method and apparatus for creating a risk control model. When it is necessary to create a risk control model, the method first obtains sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data. The categorical feature data and numerical feature data are processed into a first multidimensional feature dataset. At the same time, the first multidimensional feature dataset is composed of multiple first multidimensional feature data after the feature association relationship has been constructed. The categorical feature data and numerical feature data are processed into a second multidimensional feature dataset. The second multidimensional feature dataset is composed of multiple second multidimensional feature data after the feature high-dimensional mapping has been constructed. Finally, a third model is trained using the first multidimensional feature dataset and the second multidimensional feature dataset. The third model is a logistic regression risk control model used to determine the scoring result of the target sample. The risk control model creation scheme provided by this invention can process categorical and numerical feature data in sample data in different ways to obtain two different multidimensional numerical feature data as training samples for the final risk control model. It also considers the correlation between feature data, avoids loss of feature information, improves the training effect of the final risk control model, and thus ensures the accuracy of subsequent model evaluation.

[0019] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0021] Figure 1 A flowchart illustrating a method for creating a risk control model according to an embodiment of the present invention is shown;

[0022] Figure 2 A flowchart illustrating another method for creating a risk control model provided by an embodiment of the present invention is shown;

[0023] Figure 3 This diagram illustrates the components of a risk control model creation device provided in an embodiment of the present invention.

[0024] Figure 4 This diagram illustrates the composition of another risk control model creation device provided in an embodiment of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0026] Risk control technology plays a crucial role in the field of credit risk control. Currently, the industry practice is to build machine learning models to assess the risk of individual or corporate users. For machine learning risk control models, the mainstream approaches currently used are based on three types of algorithms: Logistic Regression, Decision Tree, and Deep Learning. Since the associated data of individual or corporate users contains both categorical and numerical features, existing technologies generally employ one-hot encoding algorithms and mapping encoding algorithms (WoE encoding, bad rate encoding, etc.) to process the categorical feature data, converting it into numerical feature data. This numerical feature data is then used as training samples to build the risk control model. However, using one-hot encoding converts numerical feature data into high-dimensional sparse feature data, which is unsuitable for ensemble tree models and deep neural network models with better performance, easily leading to overfitting. Therefore, only simple logistic regression risk control models can be used, resulting in poor model training performance. Using mapping encoding algorithms results in the loss of some feature information, also leading to poor model training performance and affecting the accuracy of subsequent evaluations. To address this, this invention provides a method for creating a risk control model that solves the problem of poor model training performance when categorical feature data exists, thus ensuring the accuracy of subsequent evaluations. The specific execution steps are as follows: Figure 1 As shown, it includes:

[0027] 101. Obtain the sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data.

[0028] It should be noted that the target sample set is a collection of multiple target samples, including but not limited to individual users and enterprise users. This embodiment does not limit this. The sample data corresponding to the target samples includes, but is not limited to, basic personal information (gender, age, family situation, etc.), financial information (income, family assets, etc.), multi-loan information (the user's loan situation on multiple lending platforms), consumption information (e-commerce data, food delivery data, travel data, review data, etc.), historical platform performance (the user's historical performance on lending platforms, which can directly reflect the user's willingness to repay), event tracking information (aggregating features based on time, frequency, and behavioral rhythm at more granular points in the user's app), and external credit data. Among these, gender and place of origin are categorical feature data, that is, non-numerical data of a certain category. It is the result of classifying things, and the data is represented by categories and described in words. The consumption amount and travel mileage in the consumption information are numerical feature data, that is, observations measured on a numerical scale, and the results are expressed as specific numerical values. The sample data often contains both types of feature data. Therefore, in the process of obtaining sample data, categorical feature data and numerical feature data can be directly extracted and classified in order to perform subsequent steps 102.

[0029] 102. Process the categorical feature data and numerical feature data into the first multidimensional feature dataset.

[0030] The first multidimensional feature dataset consists of multiple first multidimensional feature data after feature association relationships have been established. In this embodiment, it should be noted that since categorical feature data cannot be directly used as training samples to train the model, it is necessary to encode and transform the categorical feature data into numerical feature data. Encoding methods include, but are not limited to, one-hot encoding algorithms and mapping encoding (Woe encoding, bad rate encoding, etc.). In this step, a one-hot encoding algorithm is used to process the categorical feature data to ensure the integrity of the feature information of the sample data. Specifically, the categorical feature data is one-hot encoded to obtain multiple high-dimensional sparse feature data. The high-dimensional sparse feature data and numerical feature data belonging to the same target sample are then concatenated to form a first concatenated dataset. The first concatenated dataset is used to train a first model and obtain a first output result set. The first model is a factorization machine (FM) model used to construct the feature association relationships of the first concatenated data, and the first output result set is used as the first multidimensional feature dataset. The first model uses the Factor Factorization Machine (FM) model, which is more suitable for high-dimensional sparse feature data. The training samples for this model are data concatenated from high-dimensional sparse feature data and numerical feature data. Adding numerical feature data can improve the performance of the Factor Factorization Machine (FM) model. After training, the model essentially constructs the correlation between the feature data in the first concatenated data through the Factor Factorization Machine (FM) model, that is, it constructs the cross features between feature data and increases the cross terms of feature data, thereby improving the performance of the model. This allows the Factor Factorization Machine (FM) model to output a more accurate first multi-dimensional numerical feature data, which can then be used as training samples to train a risk control model that can output the score corresponding to the target sample.

[0031] 103. Process categorical and numerical feature data into a second multidimensional feature dataset.

[0032] The second multidimensional feature dataset consists of multiple second multidimensional feature data after the high-dimensional feature mapping has been constructed. In this embodiment, it should be noted that, as seen in step 102, the categorical feature data needs to be encoded into numerical feature data. The encoding methods include, but are not limited to, one-hot encoding algorithms and mapping encoding (Woe encoding, bad rate encoding, etc.). In this step, the Woe encoding algorithm is used to process the categorical feature data, mapping it into one-dimensional numerical feature data, i.e., low-dimensional dense feature data, so that it can be used as training samples for various models. Specifically, the categorical feature data is mapped and encoded to obtain multiple low-dimensional dense feature data. The low-dimensional dense feature data and numerical feature data belonging to the same target sample are then concatenated to form a second concatenated dataset. The second concatenated dataset is used to train a second model, and a second output result set is obtained. The second model is a deep neural network (DNN) model used to map the second concatenated data into multidimensional numerical feature data, and the second output result set is used as the second multidimensional feature dataset. The second model uses a deep neural network (DNN) model. After training, it can map the one-dimensional numerical feature data (second concatenated data) obtained by concatenating low-dimensional dense feature data and numerical feature data into a second high-dimensional numerical feature data (second output result). This allows the second output result and the first output result to be used as data of the same level (both are multi-dimensional numerical feature data) as training samples to train a risk control model that can output the score corresponding to the target sample, thereby improving the training effect of the model.

[0033] 104. Train the third model using the first and second multidimensional feature datasets.

[0034] The third model is a logistic regression (LR) risk control model used to determine the scoring results of the target samples. It should be noted that, as shown in steps 102 and 103, the first multidimensional feature dataset consists of multiple first multidimensional feature data after feature associations have been established, while the second multidimensional feature dataset consists of multiple second multidimensional feature data after high-dimensional feature mappings have been established. Both datasets are outputs from the first model (factorization machine model) and the second model (deep neural network model) during model training. The first multidimensional feature data avoids omitting some feature information, ensuring feature integrity and the relationships between features. The second multidimensional feature data serves as a supplementary sample at the same level as the first multidimensional feature data, ensuring consistency as training samples for the third model. By fully utilizing the sample data to train the third model, the training effect of the final risk control model is improved, thereby ensuring the accuracy of subsequent model evaluations.

[0035] Based on the above Figure 1As can be seen from the implementation method, the risk control model creation method and apparatus provided by the present invention, when it is necessary to create a risk control model, firstly acquires the sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data. The categorical feature data and numerical feature data are processed into a first multidimensional feature dataset. At the same time, the first multidimensional feature dataset is composed of multiple first multidimensional feature data after the feature association relationship has been constructed. The categorical feature data and numerical feature data are processed into a second multidimensional feature dataset. The second multidimensional feature dataset is composed of multiple second multidimensional feature data after the feature high-dimensional mapping has been constructed. Finally, a third model is trained using the first multidimensional feature dataset and the second multidimensional feature dataset. The third model is a logistic regression risk control model used to determine the scoring result of the target sample. The risk control model creation scheme provided by this invention can process categorical and numerical feature data in sample data in different ways to obtain two different multidimensional numerical feature data as training samples for the final risk control model. It also considers the correlation between feature data, avoids loss of feature information, improves the training effect of the final risk control model, and thus ensures the accuracy of subsequent model evaluation.

[0036] Furthermore, the preferred embodiment of the present invention is based on the above... Figure 1 Based on this, a detailed explanation of the risk control model creation process is provided, including the specific steps as follows: Figure 2 As shown, it includes:

[0037] 201. Obtain the sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data.

[0038] This step is consistent with the description of step 101 in the above method, so the same content will not be repeated here. It should be noted that, in order to ensure the balance of the target sample set and to make the trained model perform better, the target samples can be extracted before this step. The extracted samples can be drawn from historical samples in the credit reporting system of an authoritative institution, which includes, but is not limited to, major financial institutions, housing provident funds, social security and other institutions with complete data. Alternatively, historical samples can be extracted from the online credit reporting system. The credit reporting data of the online credit reporting system mainly comes from the Internet platform, and samples can be obtained by using Internet technology or interface cooperation. Specifically, before obtaining the sample data of each target sample in the target sample set, the method further includes: extracting a preset number of historical samples to form a historical sample set, determining the positive and negative sample ratio of the historical samples in the historical sample set, and judging whether the positive and negative sample ratio is within a preset threshold range. If so, the historical sample set is determined as the target sample set. The preset quantity can be customized. Positive samples are high-quality users with few or no negative records within a preset period, while negative samples are low-quality users with multiple negative records within the preset period. The balance of the target sample is ensured by checking whether the ratio of positive to negative samples in the historical sample set is within a preset threshold range. The preset threshold range is generally set to approximately 9 / 1, but the specific value can be customized; this embodiment does not impose a specific limitation. If the ratio of positive to negative samples is within the preset threshold range, the historical sample set is determined as the target sample set. If the ratio is not within the preset threshold range, the method further includes dynamically adjusting the number of positive and / or negative samples in the historical samples of the target sample set based on the ratio, so that the adjusted ratio of positive to negative samples in the historical samples of the target sample set is within the preset threshold range. Since the preset threshold range is already determined, positive samples, negative samples, or both can be added by comparing the ratio of positive to negative samples with the preset threshold range to ensure that the adjusted ratio of positive to negative samples in the historical samples of the target sample set is within the preset threshold range.

[0039] Furthermore, to determine the positive and negative sample ratio of historical samples in the historical sample set, the following steps are taken: First, obtain the correlation information of each historical sample in the historical sample set. This correlation information includes at least one of the following: basic attributes, behavioral performance, risk indicators, and credit information. Then, determine whether the correlation information conforms to business rules. These business rules are matching rules that determine if the historical sample corresponding to the correlation information is a positive sample. If it conforms, the historical sample corresponding to the correlation information is determined as a positive sample; otherwise, it is determined as a negative sample. The positive and negative sample ratio is determined based on the number of positive and negative samples in the historical sample set. The business rules can be user-defined, for example, using any one of the following criteria: basic attributes, behavioral performance, risk indicators, or credit information, to improve the accuracy of determining positive and negative samples and ensure the balance of the target sample.

[0040] 202. Perform one-hot encoding on the categorical feature data to obtain multiple high-dimensional sparse feature data.

[0041] This step is based on the description of step 102 in the above method, so the same content will not be repeated here. It should be noted that one-hot encoding is a process of converting categorical feature data into a form that is easy for machine learning algorithms to use, and it can ensure the integrity of the feature information of the sample data. Since different categorical feature data have different numbers of categories, after converting the categorical feature data, high-dimensional sparse feature data can be obtained, which is still essentially numerical feature data, so as to perform the subsequent step 203.

[0042] 203. The high-dimensional sparse feature data and numerical feature data belonging to the same target sample are concatenated sequentially to form the first concatenated dataset.

[0043] It should be noted that since the high-dimensional sparse feature data is obtained by converting the categorical feature data through one-hot encoding, it is not the complete data of the target sample. Therefore, in order to ensure the model's performance, the original numerical feature data can be concatenated with the high-dimensional sparse feature data to make full use of the sample data, thereby ensuring the effectiveness of the model training in the subsequent step 204.

[0044] 204. Train the first model using the first concatenated dataset and obtain the first output result set.

[0045] The first model is a Factorization Machine (FM) model used to construct the feature association relationships of the first concatenated data. It should be noted that since the first concatenated data contains the high-dimensional sparse feature data mentioned in step 203, a Factorization Machine (FM) model, capable of establishing association relationships between features in the first concatenated data, is chosen for better processing. The advantages of this model are that it can perform reasonable parameter estimation in very sparse data, its complexity is linear, and its optimization effect is good. Furthermore, it is a general model, applicable to any situation where features have real values; that is, it has linear computational complexity even in highly sparse data scenarios. Therefore, after training on the first concatenated dataset, a more accurate multi-dimensional numerical feature dataset, i.e., the first output result set, can be output as one of the training samples for the subsequent risk control model.

[0046] 205. Use the first output result set as the first multidimensional feature dataset.

[0047] In this step, since the output of the first model consists of multidimensional numerical feature data, the set of multidimensional numerical feature data corresponding to the sample data of the target sample can be used as the first multidimensional feature dataset, so as to serve as one of the training samples for the subsequent risk control model.

[0048] 206. Map and encode the categorical feature data separately to obtain multiple low-dimensional dense feature data.

[0049] This step is based on the description of step 103 in the above method, so the same content will not be repeated here. It should be noted that mapping encoding is one of the mapping encoding algorithms. It is also a process of converting categorical feature data into a form that is easy for machine learning algorithms to use. It encodes categorical feature data into one-dimensional numerical feature data so that subsequent step 207 can be performed.

[0050] 207. The low-dimensional dense feature data and numerical feature data belonging to the same target sample are sequentially concatenated to form the second concatenated dataset.

[0051] It should be noted that since the low-dimensional dense feature data is obtained by mapping and encoding categorical feature data, it is not the complete data of the target sample. Therefore, in order to ensure the model's performance, the original numerical feature data can be concatenated with the low-dimensional dense feature data to make full use of the sample data, thereby ensuring the effectiveness of model training in the subsequent step 208. However, the mapping and encoding method has drawbacks, which can easily cause the loss of some feature information. But it is essentially numerical feature data and can be used as training samples for model training in the subsequent step 208.

[0052] 208. Train the second model using the second concatenated dataset and obtain the second output result set.

[0053] The second model is a deep neural network (DNN) model used to map the second concatenated data into multidimensional numerical feature data. Since both the low-dimensional dense feature data and the original numerical feature data are one-dimensional, a deep neural network (DNN) model can be used for training to map the second concatenated dataset into multidimensional feature data at the same level as the first output set. This ensures that the second output set becomes multidimensional numerical feature data at the same level as the first output set, thereby guaranteeing the consistency of training samples for subsequent training of the risk control model.

[0054] 209. Use the second output result set as the second multidimensional feature dataset.

[0055] In this step, since the output of the second model is also multidimensional numerical feature data, the set of multidimensional numerical feature data corresponding to the sample data of the target sample can be used as the second multidimensional feature dataset, so as to serve as one of the training samples for the subsequent risk control model.

[0056] 210. Train the third model using the first and second multidimensional feature datasets.

[0057] The third model is a logistic regression risk control model used to determine the scoring results of the target sample. This step is related to the description of step 104 in the above method, so the same content will not be repeated here.

[0058] Based on the above Figure 1As can be seen from the implementation, the risk control model creation method and apparatus provided by this invention, when creating a risk control model, first acquires sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data. By performing one-hot encoding on the categorical feature data respectively, multiple high-dimensional sparse feature data are obtained. The high-dimensional sparse feature data and numerical feature data belonging to the same target sample are then concatenated to form a first concatenated dataset. The first concatenated dataset is used to train a first model and obtain a first output result set. The first output result set is used as the first multi-dimensional feature dataset. This method avoids the loss of some feature information, ensures the integrity of features and the correlation between features, and enables the processing of sample data. To fully utilize the data and ensure the effectiveness of subsequent model training, categorical feature data is mapped and encoded separately to obtain multiple low-dimensional dense feature data. Low-dimensional dense feature data and numerical feature data belonging to the same target sample are then concatenated to form a second concatenated dataset. This second concatenated dataset is used to train the second model and obtain a second output result set, which is then used as the second multidimensional feature dataset. This second output result set serves as a supplementary sample at the same level as the first multidimensional feature data, ensuring consistency as training samples for the third model. Finally, the first and second multidimensional feature datasets are used to train the third model, thus fully utilizing the sample data to train the third model, improving the training effect of the final risk control model, and ensuring the accuracy of subsequent model evaluation.

[0059] Furthermore, as a response to the above Figure 1-2 The implementation of the method embodiment shown in this invention provides a risk control model creation device. This device is used to solve the problem of poor model training performance when categorical feature data exists, thereby ensuring the accuracy of subsequent model evaluation. The embodiment of this device corresponds to the aforementioned method embodiment. For ease of reading, this embodiment will not repeat the details of the aforementioned method embodiment, but it should be understood that the device in this embodiment can implement all the contents of the aforementioned method embodiment. Specifically, as shown... Figure 3 As shown, the device includes:

[0060] The acquisition unit 31 is used to acquire sample data of each target sample in the target sample set, wherein the sample data includes categorical feature data and numerical feature data.

[0061] The first processing unit 32 is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit 31 into a first multidimensional feature dataset, wherein the first multidimensional feature dataset is composed of multiple first multidimensional feature data after feature association has been constructed.

[0062] The second processing unit 33 is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit 31 into a second multidimensional feature dataset. The second multidimensional feature dataset is composed of multiple second multidimensional feature data after the feature high-dimensional mapping has been constructed.

[0063] Training unit 34 is used to train a third model using the first multidimensional feature dataset obtained by the first processing unit 32 and the second multidimensional feature dataset obtained by the second processing unit 33. The third model is a logistic regression risk control model used to predict the scoring results of the target sample.

[0064] Furthermore, such as Figure 4 As shown, the first processing unit 32 includes:

[0065] The first encoding module 321 is used to perform one-hot encoding on the categorical feature data respectively to obtain multiple high-dimensional sparse feature data.

[0066] The first splicing module 322 is used to sequentially splice the high-dimensional sparse feature data and the numerical feature data obtained by the first encoding module 321 belonging to the same target sample to form a first spliced ​​dataset.

[0067] The first training module 323 is used to train a first model using the first splicing dataset obtained by the first splicing module 322 and obtain a first output result set. The first model is a factorization machine model used to construct the feature association relationship of the first splicing data.

[0068] The first determining module 324 is used to take the first output result set obtained by the first training module 323 as the first multidimensional feature dataset.

[0069] Furthermore, such as Figure 4 As shown, the second processing unit 33 includes:

[0070] The second encoding module 331 is used to map and encode the categorical feature data respectively to obtain multiple low-dimensional dense feature data.

[0071] The second splicing module 332 is used to sequentially splice the low-dimensional dense feature data and the numerical feature data obtained by the second encoding module 331 belonging to the same target sample to form a second spliced ​​dataset.

[0072] The second training module 333 is used to train the second model using the second splicing dataset obtained by the second splicing module 332 and obtain the second output result set. The second model is a deep neural network model used to map the second splicing data into multidimensional numerical feature data.

[0073] The second determining module 334 is used to take the second output result set obtained by the second training module 333 as the second multidimensional feature dataset.

[0074] Furthermore, such as Figure 4 As shown, before acquiring the sample data of each target sample in the target sample set, the device further includes:

[0075] Extraction unit 35 is used to extract a preset number of historical samples to form a historical sample set;

[0076] The first determining unit 36 ​​is used to determine the ratio of positive to negative samples in the historical sample set obtained by the extraction unit 35.

[0077] Judgment unit 37 is used to determine whether the ratio of positive and negative samples obtained by the first determination unit 36 ​​is within a preset threshold range;

[0078] The second determining unit 38 is used to determine the historical sample set as the target sample set if the judging unit 37 determines that the ratio of positive to negative samples is not within a preset threshold range.

[0079] Furthermore, such as Figure 4 As shown, the first determining unit 36 ​​includes:

[0080] The acquisition module 361 is used to acquire the association information of each historical sample in the historical sample set, and the association information includes at least basic attributes, behavioral performance, risk indicators and credit information.

[0081] The judgment module 362 is used to sequentially judge whether the association information obtained by the acquisition module 361 conforms to the business rules, wherein the business rules are the matching rules for determining that the historical sample corresponding to the association information is a positive sample;

[0082] The third determining module 363 is used to determine that the historical sample corresponding to the associated information is a positive sample if the judging module 362 determines that the associated information conforms to the business rules.

[0083] The fourth determining module 364 is used to determine that the historical sample corresponding to the associated information is a negative sample if the judging module 362 determines that the associated information does not conform to the business rules.

[0084] The fifth determining module 365 is used to determine the positive-negative sample ratio based on the number of positive samples obtained by the third determining module 363 and the number of negative samples obtained by the fourth determining module 364 of the historical samples in the historical sample set.

[0085] Furthermore, such as Figure 4As shown, if it is determined that the ratio of positive to negative samples is not within a preset threshold range, the device further includes:

[0086] The adjustment unit 39 is configured to dynamically adjust the number of positive samples and / or the number of negative samples in the historical samples of the target sample set based on the positive-negative sample ratio, so that the positive-negative sample ratio of the historical samples in the adjusted target sample set is within the preset threshold range. Furthermore, this embodiment of the invention also provides a storage medium for storing a computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the above-described... Figure 1-2 The method for creating the risk control model described in the document.

[0087] Furthermore, embodiments of the present invention also provide a processor for running a program, wherein the program executes the above-described... Figure 1-2 The method for creating the risk control model described in the document.

[0088] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the related descriptions in other embodiments.

[0089] It is understood that the related features in the above methods and apparatus can be referenced interchangeably. Furthermore, the terms "first," "second," etc., in the above embodiments are used to distinguish between embodiments and do not represent the superiority or inferiority of any particular embodiment.

[0090] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0091] The algorithms and displays provided herein are not inherently associated with any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0092] In addition, the memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0093] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0098] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0099] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0100] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0101] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for creating a risk control model, characterized in that, include: Obtain sample data for each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data. The sample data includes personal basic information, financial information, multiple account information, consumption information, historical platform performance, data tracking information, and external credit data. The categorical feature data and the numerical feature data are processed into a first multidimensional feature dataset, which is composed of multiple first multidimensional feature data after feature association has been established; The categorical feature data and the numerical feature data are processed into a second multidimensional feature dataset, which is composed of multiple second multidimensional feature data after the high-dimensional feature mapping has been constructed. A third model is trained using the first multidimensional feature dataset and the second multidimensional feature dataset. The third model is a logistic regression risk control model used to determine the scoring results of the target sample.

2. The method according to claim 1, characterized in that, The categorical feature data and the numerical feature data are processed into a first multidimensional feature dataset. This first multidimensional feature dataset consists of multiple first multidimensional feature data sets for which feature relationships have already been established, including: The categorical feature data are individually encoded to obtain multiple high-dimensional sparse feature data. The high-dimensional sparse feature data and the numerical feature data belonging to the same target sample are sequentially concatenated to form a first concatenated dataset; A first model is trained using the first spliced ​​dataset, and a first output result set is obtained. The first model is a factorization machine model used to construct the feature association relationship of the first spliced ​​data. Use the first output result set as the first multidimensional feature dataset.

3. The method according to claim 1, characterized in that, The categorical feature data and the numerical feature data are processed into a second multidimensional feature dataset. This second multidimensional feature dataset consists of multiple second multidimensional feature data sets after a high-dimensional feature mapping has been constructed, including: The categorical feature data are mapped and encoded respectively to obtain multiple low-dimensional dense feature data; The low-dimensional dense feature data and the numerical feature data belonging to the same target sample are sequentially concatenated to form a second concatenated dataset; The second model is trained using the second spliced ​​dataset, and a second output result set is obtained. The second model is a deep neural network model used to map the second spliced ​​data into multidimensional numerical feature data. Use the second output result set as the second multidimensional feature dataset.

4. The method according to claim 1, characterized in that, Before acquiring the sample data of each target sample in the target sample set, the method further includes: Extract a predetermined number of historical samples to form a historical sample set; Determine the ratio of positive to negative samples in the historical sample set; Determine whether the ratio of positive to negative samples is within a preset threshold range; If so, then the historical sample set is determined to be the target sample set.

5. The method according to claim 4, characterized in that, Determining the ratio of positive to negative samples in the historical sample set includes: Obtain the association information of each historical sample in the historical sample set, wherein the association information includes at least one of the following: basic attributes, behavioral performance, risk indicators, and credit information; The association information is sequentially determined to conform to the business rules, wherein the business rules are the matching rules that determine the historical sample corresponding to the association information as a positive sample; If the conditions are met, the historical sample corresponding to the associated information is determined to be a positive sample; If the condition is not met, the historical sample corresponding to the associated information is determined to be a negative sample. The positive-to-negative sample ratio is determined based on the number of positive and negative samples in the historical sample set.

6. The method according to any one of claims 4-5, characterized in that, If it is determined that the ratio of positive to negative samples is not within a preset threshold range, the method further includes: The number of positive samples and / or negative samples in the historical samples of the target sample set are dynamically adjusted based on the positive-negative sample ratio, so that the positive-negative sample ratio of the historical samples in the adjusted target sample set is within the preset threshold range.

7. A device for creating a risk control model, characterized in that, include: The acquisition unit is used to acquire sample data of each target sample in the target sample set. The sample data includes categorical feature data and numerical feature data. The sample data includes personal basic information, financial information, multiple account information, consumption information, historical platform performance, data tracking information and external credit data. The first processing unit is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit into a first multidimensional feature dataset, wherein the first multidimensional feature dataset is composed of multiple first multidimensional feature data after feature association has been constructed. The second processing unit is used to process the categorical feature data and the numerical feature data obtained by the acquisition unit into a second multidimensional feature dataset. The second multidimensional feature dataset is composed of multiple second multidimensional feature data after the feature high-dimensional mapping has been constructed. The training unit is used to train a third model using the first multidimensional feature dataset obtained by the first processing unit and the second multidimensional feature dataset obtained by the second processing unit. The third model is a logistic regression risk control model used to predict the scoring results of the target sample.

8. The apparatus according to claim 7, characterized in that, The first processing unit includes: The first encoding module is used to perform one-hot encoding on the categorical feature data respectively to obtain multiple high-dimensional sparse feature data. The first splicing module is used to sequentially splice the high-dimensional sparse feature data and the numerical feature data belonging to the same target sample to form a first spliced ​​dataset. The first training module is used to train a first model using the first concatenated dataset and obtain a first output result set. The first model is a factorization machine model used to construct the feature association relationship of the first concatenated data. The first determining module is used to take the first output result set as the first multidimensional feature dataset.

9. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device where the storage medium is located to perform the method for creating a risk control model as described in any one of claims 1 to 6.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the method for creating a risk control model as described in any one of claims 1 to 6.

Images