An interactive logistic regression modeling method to ensure reasonable weight coefficients

Abstract

The present invention provides an interactive logistic regression modeling method for ensuring reasonable weight coefficients, which belongs to the technical field of financial model training. The method specifically includes: taking a sub-model without abnormal credit data sources as a to-be-included model, determining the inclusion processing order of the to-be-included model according to the correlation between the to-be-included model and the credit data sources of different target sub-models and the deviation between the weight coefficient of the target sub-model and the weight coefficient interval, retraining the credit model based on the inclusion processing order and the abnormality processing strategy, determining the credit model reconstruction strategy based on the abnormal credit data source and the weight abnormality sub-model, and stopping the training until the training result meets the requirements or reaches the reconstruction strategy, thereby improving the efficiency of the training processing and reducing the influence of a single data source in the credit model.

Description

Technical Field

[0001] The present invention belongs to the technical field of financial model training, and in particular relates to an interactive logistic regression modeling method that ensures reasonable weight coefficients. Background Art

[0002] In credit business scenarios, model teams simultaneously build and launch numerous models to meet the needs of multi-angle analysis for various business scenarios. The explosive growth of model scores has made cross-validation model fusion no longer suitable, as it results in a dimensional explosion and poor interpretability of the fusion logic. For multi-model fusion, the current industry-wide approach is for strategy analysts to design and evaluate targets based on business scenarios and then integrate them by building logistic regression models.

[0003] However, the above technical solution has the following technical problems: the model divisions of different sub-models are often constructed based on multiple credit data sources. Therefore, when building a credit model based on multi-model fusion, the weight values ​​of different sub-models will be different. This may lead to a large number of sub-models constructed based on a certain data source and a large sum of weights. This makes the entire credit model more affected by a single data source, thereby affecting the stability of the overall credit model.

[0004] In order to solve the above technical problems, the present application provides an interactive logistic regression modeling method that ensures reasonable weight coefficients. Summary of the Invention

[0005] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0006] Specifically, the present application provides an interactive logistic regression modeling method that ensures reasonable weight coefficients, specifically including:

[0007] S1 determines the weight coefficient ranges of different target sub-models based on the overlap of the credit data sources referenced by the target sub-models;

[0008] S2 determines the weight coefficients of different target sub-models in the credit model through a logistic regression algorithm. When the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements, the target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model;

[0009] S3 determines the abnormality handling strategy for abnormal credit data sources based on the constituent data of the weight abnormality sub-model associated with different credit data sources and the deviation between the weight coefficient and the weight coefficient interval;

[0010] S4 takes the sub-model without abnormal credit data source as the model to be included, and determines the inclusion processing order of the model to be included according to the correlation between the model to be included and the credit data source of different target sub-models and the deviation between the weight coefficient of the target sub-model and the weight coefficient interval. The credit model is retrained according to the inclusion processing order and the abnormal processing strategy, and the credit model reconstruction strategy is determined based on the abnormal credit data source and the weight abnormal sub-model. Training is stopped until the training results meet the requirements or the reconstruction strategy is achieved.

[0011] The beneficial effects of the present invention are:

[0012] Based on the composition data of the weight abnormal sub-models associated with different credit data sources and the deviation between the weight coefficient and the weight coefficient interval, the abnormal processing strategy of the abnormal credit data source is determined, thereby avoiding the technical problem of low efficiency in the generation and processing of the credit model due to the large number of weight abnormal sub-models associated with the credit data source and the serious deviation, which may result in the failure to eliminate them. By specifically determining the elimination processing strategy, the dynamic update of the composition of the sub-models of the credit model is achieved, thereby improving the efficiency of the generation and processing.

[0013] The reconstruction strategy of the credit model is determined based on the abnormal credit data source and the weight abnormal sub-model, thereby avoiding the technical problem of low efficiency of the credit model update processing caused by continued repeated training of the credit model when there are a large number of abnormal credit data sources or a large number of weight abnormal sub-models. Targeted use of credit data sources that do not have weight abnormal sub-models realizes the reconstruction of the credit model, laying the foundation for further improving the efficiency of the generation processing, while also reducing the influence of a single data source in the credit model, avoiding the technical problem of a large impact on the accuracy of the overall credit assessment results when data from a certain data source is missing.

[0014] A further technical solution is that the target sub-model is determined according to the random selection results of the training personnel, and specifically a preset number of target sub-models are selected.

[0015] A further technical solution is that the overlap of the credit data sources referenced by the target sub-model includes the number of overlaps of the credit data sources referenced by the target sub-model, wherein the referenced credit data source is the credit data source called by the target sub-model when performing credit risk assessment.

[0016] It is understandable that the credit data source includes the user's credit risk data, specifically including position, credit expenditure record, overdue record, age, marital status, housing data, mortgage data, vehicle data and car loan data.

[0017] A further technical solution is that the method for determining the weight coefficient interval of the target sub-model is:

[0018] Based on the overlap of the credit data sources referenced by the target sub-model, determine the overlap amount of the credit data sources referenced by the target sub-model with those of other target sub-models;

[0019] Determining a total number of overlaps of the target sub-model based on the number of overlaps in different referenced credit data sources;

[0020] The weight coefficient interval of the target sub-model is determined according to the total number of overlaps.

[0021] A further technical solution is that the method for determining the credit model reconstruction strategy is:

[0022] Based on the number of abnormal credit data sources of the credit model, determine the proportion of abnormal credit data sources in the credit data sources referenced by the sub-models of the credit model, and use this as the abnormal data source proportion;

[0023] Determine the number of weighted abnormal sub-models corresponding to different abnormal credit data sources based on the weighted abnormal sub-models corresponding to different abnormal credit data sources;

[0024] The reconstruction strategy of the credit model is determined based on the proportion of the abnormal data sources and the number of weighted abnormal sub-models corresponding to different abnormal credit data sources.

[0025] A further technical solution is to stop training until the training results meet the requirements or the reconstruction strategy is achieved, specifically including:

[0026] When the credit risk identification accuracy of the credit model in different customer groups is greater than the preset accuracy threshold, the training result is determined to meet the requirements;

[0027] If the credit risk identification accuracy of the credit model in different customer groups is not uniformly greater than the preset accuracy threshold, and the credit model reaches the conditions for rebuilding the credit model, the training process will be stopped and the credit model will be rebuilt.

[0028] Other features and advantages will be described in the following description. The objectives and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and drawings.

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the accompanying drawings.

[0031] Figure 1 It is a flowchart of an interactive logistic regression modeling method that ensures reasonable weight coefficients;

[0032] Figure 2 is a flow chart of a method for determining a weight coefficient interval of a target sub-model;

[0033] Figure 3 It is a flow chart of a method for determining whether a deviation between a weight coefficient and a weight coefficient interval does not meet the requirements;

[0034] Figure 4 is a flow chart of a method for determining an exception handling strategy for an abnormal credit data source;

[0035] Figure 5 This is a flowchart of a method for determining the order of inclusion processing of models to be included. DETAILED DESCRIPTION

[0036] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this specification without creative work should fall within the scope of protection of this specification.

[0037] Example 1

[0038] like Figure 1 As shown, the present application provides an interactive logistic regression modeling method that ensures reasonable weight coefficients, specifically including:

[0039] S1 determines the weight coefficient ranges of different target sub-models based on the overlap of the credit data sources referenced by the target sub-models;

[0040] Furthermore, the target sub-model is determined according to the random selection result of the training personnel, and is specifically selected according to a preset number of target sub-models.

[0041] Specifically, the overlap of the credit data sources referenced by the target sub-model includes the number of overlaps of the credit data sources referenced by the target sub-model, where the referenced credit data source is the credit data source called by the target sub-model when performing credit risk assessment.

[0042] It is understandable that the credit data source includes the user's credit risk data, specifically including position, credit expenditure record, overdue record, age, marital status, housing data, mortgage data, vehicle data and car loan data.

[0043] Specifically, such as Figure 2 As shown, the method for determining the weight coefficient interval of the target sub-model is:

[0044] Based on the overlap of the credit data sources referenced by the target sub-model, determine the overlap amount of the credit data sources referenced by the target sub-model with those of other target sub-models;

[0045] Determining a total number of overlaps of the target sub-model based on the number of overlaps in different referenced credit data sources;

[0046] The weight coefficient interval of the target sub-model is determined according to the total number of overlaps.

[0047] It can be understood that normalization processing is performed based on the total number of overlaps of different target sub-models to obtain recommended weight coefficients of different target sub-models. The weight coefficient range is determined based on the preset value with the recommended weight coefficient as the center. For example, if the recommended weight coefficient is 0.1, the weight coefficient range is 0.08-0.12.

[0048] Optionally, the method for determining the weight coefficient interval of the target sub-model is:

[0049] Based on the overlap of the credit data sources referenced by the target sub-model, determine the overlap amount of the credit data sources referenced by the target sub-model with those of other target sub-models;

[0050] Determining recommendation weight coefficients for different referenced credit data sources based on the number of overlaps among different referenced credit data sources;

[0051] The recommendation weight coefficient is obtained by normalizing the number of overlaps in different referenced credit data sources.

[0052] The sum of the recommended weight coefficients of different referenced credit data sources is normalized and combined with preset values ​​to determine the weight coefficient range of the target sub-model.

[0053] S2 determines the weight coefficients of different target sub-models in the credit model through a logistic regression algorithm. When the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements, the target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model;

[0054] Furthermore, the weight coefficients of different target sub-models in the credit model are determined by a logistic regression algorithm, specifically including:

[0055] Set the initial weight coefficients of different target sub-models, iteratively update the weights in the training set, adjust the parameters along the negative gradient direction of the loss function until the target number of iterations is reached or the loss function meets the requirements, and then determine the weight coefficients of different target sub-models.

[0056] Specifically, such as Figure 3 As shown, determining that the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements specifically includes:

[0057] The target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model;

[0058] Determining the number of weight anomaly sub-models in different credit data sources based on the credit data sources referenced by the weight anomaly sub-models;

[0059] According to the number of weight anomaly sub-models in different credit data sources, it is determined whether the deviation between the weight coefficient and the weight coefficient interval meets the requirements.

[0060] It should be noted that if there is a credit data source in which the number of weight abnormal sub-models is greater than the preset abnormal sub-model number threshold, it is determined that the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements.

[0061] In another embodiment, if the total number of weight abnormality sub-models of different credit data sources is more than 3 or the number of credit data sources with weight abnormality sub-models is more than 2, it is determined that the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements, and the specific threshold number needs to be dynamically adjusted according to the number of target sub-models.

[0062] It can be understood that when the deviation between the weight coefficient and the weight coefficient interval meets the requirements, the credit model will be used as the final credit model.

[0063] Optionally, determining that a deviation between the weight coefficient and the weight coefficient interval does not meet a requirement specifically includes:

[0064] The target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model;

[0065] Determining the number of weight anomaly sub-models in different credit data sources based on the credit data sources referenced by the weight anomaly sub-models;

[0066] Based on the number of credit data sources with weight anomaly sub-models, determine whether the deviation between the weight coefficient and the weight coefficient interval meets the requirements.

[0067] S3 determines the abnormality handling strategy for abnormal credit data sources based on the constituent data of the weight abnormality sub-model associated with different credit data sources and the deviation between the weight coefficient and the weight coefficient interval;

[0068] Specifically, such as Figure 4 As shown, the method for determining the abnormality handling strategy of the abnormal credit data source is:

[0069] Determine the associated weighted anomaly sub-model of the credit data source using the constituent data of the associated weighted anomaly sub-model of different credit data sources, and use it as the associated anomaly sub-model;

[0070] According to the deviation between the weight coefficients of different associated anomaly sub-models and the weight coefficient interval, the deviation amount from the adjacent endpoint of the weight coefficient is determined;

[0071] The exception handling strategy of the credit data source is determined by the deviation amounts of different associated exception sub-models.

[0072] It can be understood that when the number of associated abnormal sub-models whose deviation is greater than the preset deviation threshold is more than 3, the credit data source is determined to be an abnormal credit data source.

[0073] In another embodiment, determining the exception handling strategy of the credit data source based on the deviation amounts of different associated exception sub-models specifically includes:

[0074] When the proportion of the number of associated abnormal sub-models in the sub-models associated with the credit data source is greater than a preset proportion threshold, a first preset strategy is adopted to determine an abnormality handling strategy for the credit data source;

[0075] When the proportion of the number of sub-models associated with the credit data source that are associated with the abnormal sub-models is not above the preset proportion threshold but is within a preset range, a second preset strategy is adopted to determine an abnormality handling strategy for the credit data source;

[0076] When the proportion of the number of sub-models associated with the credit data source by the associated abnormal sub-models is not within a preset range, a third preset strategy is adopted to determine the abnormality handling strategy of the credit data source.

[0077] The first preset strategy is that after retraining, if there are abnormal sub-models associated with the credit data source in more than a target number of training processing times, all sub-models associated with the credit data source will be removed from the credit model;

[0078] The second preset strategy is that after retraining, when there are associated abnormal sub-models in more than the second target number of training processing times, all associated abnormal sub-models associated with the credit data source will be eliminated from the credit model. The third prediction strategy is that no abnormal processing is required.

[0079] In another embodiment, the method for determining the abnormality handling strategy of the abnormal credit data source is:

[0080] Using the constituent data of the weighted anomaly sub-models associated with different credit data sources, determine the weighted anomaly sub-model associated with the credit data source, and use it as the associated anomaly sub-model; and determine the deviation from the adjacent endpoint of the weight coefficient based on the deviation between the weight coefficients of the different associated anomaly sub-models and the weight coefficient interval;

[0081] It is understandable that in the above steps, if the number of associated abnormal sub-models is large or the number of associated abnormal sub-models in the associated sub-models accounts for a large proportion, that is, greater than the threshold, then the credit data source can be directly determined to be an abnormal credit data source, and the abnormality handling strategy for the credit data source can be determined by the first preset strategy;

[0082] It should also be noted that if the number of associated abnormal sub-models is small or the number of associated abnormal sub-models in the associated sub-models does not account for a large proportion, it is necessary to determine the deviation amounts of different associated abnormal sub-models. When the deviation amount is greater than the preset deviation amount and the number of associated abnormal sub-models does not meet the requirement, that is, when it is greater than the preset number threshold, the credit data source can be directly determined to be an abnormal credit data source, and the abnormality handling strategy for the credit data source is determined by the first preset strategy.

[0083] In addition, it can be understood that even if the number of associated abnormal sub-models with deviations greater than the preset deviation amount meets the requirement, it is necessary to further determine whether the average of the proportion of the number of associated abnormal sub-models in the associated sub-models and the proportion of the number of associated abnormal sub-models with deviations greater than the preset deviation amount in the associated sub-models is greater than the proportion threshold. If the average is greater than the proportion threshold, the credit data source can be directly determined to be an abnormal credit data source, and the abnormality handling strategy for the credit data source is determined by the first preset strategy;

[0084] When the number of associated abnormal sub-models whose deviation is greater than the preset deviation threshold is not more than 3, it is determined that the credit data source is not an abnormal credit data source.

[0085] The associated sub-models other than the weight abnormal sub-model are regarded as associated sub-models, and the associated sub-models whose deviation from the adjacent endpoints of the adjacent weight coefficient interval is within a preset range are regarded as critical sub-models;

[0086] It is also understandable that, in the above steps, if there is no critical sub-model, then the second preset strategy can be directly determined to determine the exception handling strategy for the credit data source.

[0087] In addition, it should be noted that in the above steps, if a critical sub-model exists and the ratio of the number of critical sub-models to the number of associated abnormal sub-models is greater than 0.6, then the credit data source can be directly determined to be an abnormal credit data source, and the abnormality handling strategy for the credit data source can be determined using the first preset strategy;

[0088] It should also be noted that if the ratio of the number of critical sub-models to the number of associated abnormal sub-models is greater than 0.6, for example, less than a certain threshold, the credit data source can be directly determined to be an abnormal credit data source, and the abnormality handling strategy for the credit data source can be determined by the second preset strategy;

[0089] The abnormality handling strategy of the credit data source is determined by the deviation amount of different associated abnormal sub-models and the number of critical sub-models, and in combination with the number of sub-models of the credit data source.

[0090] In a possible embodiment, based on the deviation amounts of different associated abnormal sub-models, preset deviation weight coefficients corresponding to the deviation amounts of different associated abnormal sub-models are determined, and the impact value is determined according to the sum of the preset deviation weight coefficients of different associated abnormal sub-models and the average of the proportion of the number of critical sub-models and associated abnormal sub-models in the number of sub-models in the credit model.

[0091] When the impact value does not meet the requirements, that is, is greater than the preset impact threshold, the exception handling strategy of the credit data source is the first preset strategy, and in other cases it is the second preset strategy.

[0092] S4 takes the sub-model without abnormal credit data source as the model to be included, and determines the inclusion processing order of the model to be included according to the correlation between the model to be included and the credit data source of different target sub-models and the deviation between the weight coefficient of the target sub-model and the weight coefficient interval. The credit model is retrained according to the inclusion processing order and the abnormal processing strategy, and the credit model reconstruction strategy is determined based on the abnormal credit data source and the weight abnormal sub-model. Training is stopped until the training results meet the requirements or the reconstruction strategy is achieved.

[0093] Specifically, such as Figure 5 As shown, the method for determining the order of inclusion processing of the models to be included is:

[0094] Determine the overlap between the credit data sources referenced by the model to be included and the different target sub-models based on the association between the model to be included and the credit data sources of the different target sub-models;

[0095] Determine the target sub-model of the credit data source with the overlap using the overlap situation, and use it as the data source overlap model;

[0096] The order of inclusion of the models to be included is determined based on the deviation between the weight coefficients of the overlapping models of different data sources and the weight coefficient intervals.

[0097] It is understandable that the order of inclusion in the model is determined based on the deviation between the weight coefficients and weight coefficient intervals of the credit data sources of different data source overlap models, specifically including:

[0098] The order of inclusion of the models to be included is determined based on the number of data source overlap models whose weight coefficients are not within the weight coefficient interval, from small to large.

[0099] Furthermore, the method for determining the credit model reconstruction strategy is:

[0100] Based on the number of abnormal credit data sources of the credit model, determine the proportion of abnormal credit data sources in the credit data sources referenced by the sub-models of the credit model, and use this as the abnormal data source proportion;

[0101] Determine the number of weighted abnormal sub-models corresponding to different abnormal credit data sources based on the weighted abnormal sub-models corresponding to different abnormal credit data sources;

[0102] The reconstruction strategy of the credit model is determined based on the proportion of the abnormal data sources and the number of weighted abnormal sub-models corresponding to different abnormal credit data sources.

[0103] It is understandable that, based on the proportion of abnormal data sources and the number of weighted abnormal sub-models corresponding to different abnormal credit data sources, the credit model reconstruction strategy is determined, specifically including:

[0104] When the proportion of abnormal data sources is greater than the preset abnormal data source proportion threshold, it means that the number of abnormal data sources is relatively large. Therefore, on this basis, if the number of abnormal data sources eliminated in the credit model is greater than the preset number of data sources, and the remaining abnormal data sources still have weight abnormality sub-models, the credit model will be rebuilt;

[0105] When the proportion of abnormal data sources is not greater than the preset abnormal data source proportion threshold, it means that the number of abnormal data sources at this time is small. Therefore, on this basis, it is also necessary to determine whether the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is greater than the preset abnormal sub-model number threshold. If so, it means that the number of sub-models with abnormal weight coefficients is large. Therefore, on this basis, if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is still greater than the preset abnormal sub-model number threshold after retraining for more than 3 times, the credit model will be reconstructed; and if there are still weighted abnormal sub-models after retraining for more than 10 times, the credit model will be reconstructed.

[0106] However, if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is not greater than the preset abnormal sub-model number threshold, it is necessary to determine the retraining number threshold based on the proportion of the number of weighted abnormal sub-models. When the number of retraining times is greater than the retraining number threshold and there are still weighted abnormal sub-models, the credit model will be reconstructed.

[0107] The retraining number threshold is determined based on a preset number threshold corresponding to the proportion of the number of weight abnormal sub-models, or can be determined based on the ratio of a preset value to the proportion of the number of weight abnormal sub-models. The larger the proportion of the number of weight abnormal sub-models, the smaller the retraining number threshold is, and its value range is between 15 and 20 times.

[0108] In another possible embodiment, the method for determining the credit model reconstruction strategy is:

[0109] S41 determines, based on the number of abnormal credit data sources of the credit model, a proportion of the abnormal credit data sources in the credit data sources referenced by the sub-models of the credit model, and uses the proportion as the abnormal data source proportion;

[0110] It is understandable that in the above steps, it is necessary to determine whether the number of abnormal credit data sources and the proportion of abnormal data sources meet the requirements, that is, whether they are greater than the preset threshold. When any of them does not meet the requirements, it means that the number of abnormal credit data sources in the credit model is large. Therefore, on this basis, in order to improve the training processing efficiency of the credit model, if the number of abnormal data sources eliminated in the credit model is greater than the preset number of data sources, and the remaining abnormal data sources still have weight abnormality sub-models, the credit model will be rebuilt;

[0111] It should also be noted that in the above steps, when the number of abnormal credit data sources is within the preset number of data sources and the proportion of abnormal data sources is less than 0.1, it means that the number of abnormal credit data sources is small. Therefore, in order to avoid excessive repetition of building the credit model, only when the number of retraining times is greater than the retraining number threshold and there is still a weight abnormality sub-model, the credit model will be rebuilt. In other cases, proceed to the next step.

[0112] S42 determines the number of weighted abnormal sub-models corresponding to different abnormal credit data sources according to the weighted abnormal sub-models corresponding to different abnormal credit data sources, and determines the attention data source in the abnormal credit data source based on the number of weighted abnormal sub-models;

[0113] It should be noted that, in one embodiment, the attention data source is an abnormal credit data source whose number of weighted abnormal sub-models is more than 5.

[0114] It can be understood that in the above steps, it is necessary to determine whether the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is greater than the preset abnormal sub-model number threshold. If so, it means that the number of sub-models with abnormal weight coefficients is large. Therefore, on this basis, if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is still greater than the preset abnormal sub-model number threshold after retraining for more than 3 times, the credit model will be reconstructed; and if there are still abnormal weighted sub-models after retraining for more than 10 times, the credit model will be reconstructed.

[0115] It can also be understood that if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is not greater than the preset abnormal sub-model number threshold, it is also necessary to determine whether there is an attention data source. When there is an attention data source, if the number of attention data sources is greater than the preset attention data source number threshold, it means that the number of sub-models with abnormal weight coefficients is large. Therefore, on this basis, if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is still greater than the preset abnormal sub-model number threshold after retraining for more than 3 times, the credit model will be reconstructed; and if the weighted abnormal sub-model still exists after retraining for more than 10 times, the credit model will be reconstructed, and in other cases, proceed to step S43.

[0116] S43 determines the reconstruction strategy of the credit model based on the proportion of the abnormal data sources, the proportion of the number of weighted abnormal sub-models, and the composition data of the attention data source.

[0117] It should be noted that the weighted outlier value of the credit model is determined by taking the average of the proportion of abnormal data sources, the proportion of the number of weighted outlier sub-models, and the proportion of attention data sources in the data sources referenced by the credit model.

[0118] It can be understood that when the weight anomaly value of the credit model is greater than the preset anomaly threshold, if the number of abnormal data sources eliminated in the credit model is greater than the preset number of data sources, and the remaining abnormal data sources still have weight anomaly sub-models, the credit model will be reconstructed. In other cases, it is necessary to determine whether the sum of the number of weight anomaly sub-models corresponding to different abnormal credit data sources is greater than the preset abnormal sub-model number threshold. If so, it means that there are many sub-models with abnormal weight coefficients. Therefore, on this basis, if the sum of the number of weight anomaly sub-models corresponding to different abnormal credit data sources is still greater than the preset abnormal sub-model number threshold after retraining for more than 3 times, the credit model will be reconstructed; and if the weight anomaly sub-model still exists after retraining for more than 10 times, the credit model will be reconstructed.

[0119] However, if the sum of the number of weighted abnormal sub-models corresponding to different abnormal credit data sources is not greater than the preset abnormal sub-model number threshold, it is necessary to determine the retraining number threshold based on the proportion of the number of weighted abnormal sub-models. When the number of retraining times is greater than the retraining number threshold and there are still weighted abnormal sub-models, the credit model will be reconstructed.

[0120] It should be noted that the credit model is reconstructed, specifically including:

[0121] The credit model is reconstructed based on the credit data source that does not have a weight abnormality sub-model in the credit model and the sub-model corresponding to the credit data source that does not have a weight abnormality sub-model.

[0122] Furthermore, training is stopped until the training results meet the requirements or the strategy is rebuilt, including:

[0123] When the credit risk identification accuracy of the credit model in different customer groups is greater than the preset accuracy threshold, the training result is determined to meet the requirements;

[0124] If the credit risk identification accuracy of the credit model in different customer groups is not uniformly greater than the preset accuracy threshold, and the credit model reaches the conditions for rebuilding the credit model, the training process will be stopped and the credit model will be rebuilt.

[0125] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, refer to the descriptions of the method embodiments.

[0126] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0127] The foregoing description is merely one or more embodiments of this specification and is not intended to limit this specification. It will be apparent to those skilled in the art that various modifications and variations may be made to one or more embodiments of this specification. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of one or more embodiments of this specification are intended to be within the scope of the claims of this specification.

Claims

1. An interactive logistic regression modeling method that ensures reasonable weight coefficients, characterized in that: Specifically include: Determine the weight coefficient ranges of different target sub-models based on the overlap of credit data sources referenced by the target sub-models; Determine the weight coefficients of different target sub-models in the credit model using a logistic regression algorithm. When the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements, the target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model; Determine the abnormality handling strategy for abnormal credit data sources based on the constituent data of the weight abnormality sub-model associated with different credit data sources and the deviation between the weight coefficient and the weight coefficient interval; The sub-models without abnormal credit data sources are selected as the models to be included. The order of inclusion of the models to be included is determined based on the correlation between the models to be included and the credit data sources of different target sub-models, as well as the deviation between the weight coefficient of the target sub-model and the weight coefficient interval. The credit model is retrained based on the inclusion order and the abnormality handling strategy. The credit model reconstruction strategy is determined based on the abnormal credit data source and the weight abnormality sub-model. Training is stopped until the training results meet the requirements or the reconstruction strategy is achieved. The method for determining the abnormality handling strategy of the abnormal credit data source is as follows: Determine the associated weighted anomaly sub-model of the credit data source using the constituent data of the associated weighted anomaly sub-model of different credit data sources, and use it as the associated anomaly sub-model; According to the deviation between the weight coefficients of different associated anomaly sub-models and the weight coefficient interval, the deviation amount from the adjacent endpoint of the weight coefficient is determined; Determining an exception handling strategy for the credit data source based on the deviation amounts of different associated exception sub-models; The method for determining the credit model reconstruction strategy is: Based on the number of abnormal credit data sources of the credit model, determine the proportion of abnormal credit data sources in the credit data sources referenced by the sub-models of the credit model, and use this as the abnormal data source proportion; Determine the number of weighted abnormal sub-models corresponding to different abnormal credit data sources based on the weighted abnormal sub-models corresponding to different abnormal credit data sources; Determining a reconstruction strategy for the credit model based on the proportion of abnormal data sources and the number of weighted abnormal sub-models corresponding to different abnormal credit data sources; Training stops until the training results meet the requirements or the strategy is rebuilt, including: When the credit risk identification accuracy of the credit model in different customer groups is greater than the preset accuracy threshold, the training result is determined to meet the requirements; If the credit risk identification accuracy of the credit model in different customer groups is not uniformly greater than the preset accuracy threshold, and the credit model reaches the conditions for rebuilding the credit model, the training process will be stopped and the credit model will be rebuilt.

2. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The target sub-model is determined based on the random selection results of the trainers.

3. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The overlap of the credit data sources referenced by the target sub-model includes the overlap number of the credit data sources referenced by the target sub-model.

4. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The credit data source includes the user's credit risk data, including position, credit expenditure record, overdue record, age, marital status, housing data, mortgage data, vehicle data and car loan data.

5. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The method for determining the weight coefficient interval of the target sub-model is: Based on the overlap of the credit data sources referenced by the target sub-model, determine the overlap amount of the credit data sources referenced by the target sub-model with those of other target sub-models; Determining a total number of overlaps of the target sub-model based on the number of overlaps in different referenced credit data sources; The weight coefficient interval of the target sub-model is determined according to the total number of overlaps.

6. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The weight coefficients of different target sub-models in the credit model are determined by the logistic regression algorithm, including: Set the initial weight coefficients of different target sub-models, iteratively update the weights in the training set, adjust the parameters along the negative gradient direction of the loss function until the target number of iterations is reached or the loss function meets the requirements, and then determine the weight coefficients of different target sub-models.

7. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: Determining whether the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements specifically includes: The target sub-model whose weight coefficient is not within the weight coefficient interval is regarded as a weight abnormal sub-model; Determining the number of weight anomaly sub-models in different credit data sources based on the credit data sources referenced by the weight anomaly sub-models; According to the number of weight anomaly sub-models in different credit data sources, it is determined whether the deviation between the weight coefficient and the weight coefficient interval meets the requirements.

8. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 7, characterized in that: If there is a credit data source in which the number of weight abnormal sub-models is greater than the preset abnormal sub-model number threshold, it is determined that the deviation between the weight coefficient and the weight coefficient interval does not meet the requirements.

9. The interactive logistic regression modeling method for ensuring reasonable weight coefficients according to claim 1, characterized in that: The method for determining the order of inclusion processing of the models to be included is: Determine the overlap between the credit data sources referenced by the model to be included and the different target sub-models based on the association between the model to be included and the credit data sources of the different target sub-models; Determine the target sub-model of the credit data source with the overlap using the overlap situation, and use it as the data source overlap model; The order of inclusion of the models to be included is determined based on the deviation between the weight coefficients of the overlapping models of different data sources and the weight coefficient intervals.

Images