Model training method, information generation method, device, equipment, medium and product

Abstract

The embodiment of the invention discloses a model training method and device, an information generation method and device, equipment, a medium and a product. A specific embodiment of the method comprises the steps of determining monotonicity information corresponding to each sample feature; executing a training step: inputting the sample feature information subset of the forward monotone information into the first initial full-connection model to obtain a first output result; inputting the sample feature information subset of the negative monotone information into a second initial full-connection model to obtain a second output result; inputting the sample feature information subset without the monotone information into a third initial full-connection model to obtain a third output result; determining whether the training is finished; and determining the initial value influence information generation model as a value influence information generation model in response to determining that the training is ended. The implementation mode is related to artificial intelligence, the trained value influence information is utilized to generate the model, and the value influence information for the input data can be accurately generated.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technologies, and particularly to model training methods, information generation methods, apparatuses, devices, media, and products. Background Art

[0002] Currently, in various scenarios, the interpretable application of machine learning models is very extensive. When facing interpretable problems, the monotonicity between feature variables and target variables is often required. In this way, effective cause analysis can be performed on the model results. For the interpretable application of models, the commonly used method is to obtain an interpretable model by truncating the weight values of models such as linear regression and logistic regression.

[0003] However, the inventors have found that when using the above method, the following technical problems often exist:

[0004] There is no cross-processing between features in models such as linear regression and logistic regression, resulting in limited application of the obtained interpretable model and inaccurate prediction. In addition, for the interpretable adjustment of multi-layer neural network models, there are often problems such as complex network structures and difficulty in ensuring the monotonicity between different input features and target variables by simply constraining the weight values of each layer.

[0005] The above information disclosed in this background art section is only used to enhance the understanding of the background of the inventive concept, and thus, it may include information that does not form the prior art known to those of ordinary skill in the art in this country. Summary of the Invention

[0006] This summary of the disclosure is used to introduce concepts in a brief form, which will be described in detail in the following detailed implementation section. This summary of the disclosure is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to be used to limit the scope of the claimed technical solution.

[0007] Some embodiments of the present disclosure propose model training methods, information generation methods, apparatuses, devices, media, and products to solve the technical problems mentioned in the above background art section.

[0008] In a first aspect, some embodiments of the present disclosure provide a model training method, including: determining, according to a training sample set, monotonicity information corresponding to each sample feature in a sample feature set; for the training sample set, performing the following training steps: inputting a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is positive monotonic information, into a first initial fully connected model included in an initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to target training samples in the training sample set; inputting a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is negative monotonic information, into a second initial fully connected model included in the initial value impact information generation model to obtain a second output result; inputting a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is non-existent monotonic information, into a third initial fully connected model included in the initial value impact information generation model to obtain a third output result; determining whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result; in response to determining that the training is completed, determining the initial value impact information generation model as a value impact information generation model.

[0009] Optionally, the above method further includes: in response to determining that the training is not completed, updating model parameters of the initial value impact information generation model according to the first output result, the second output result, and the third output result to obtain an updated value impact information generation model, and removing the target training samples from the training sample set to obtain a post-removal training sample set; using the post-removal training sample set as the training sample set and the updated value impact information generation model as the initial value impact information generation model, and continuing to perform the above training steps.

[0010] Optionally, the determining whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result includes: inputting the first output result, the second output result, and the third output result into a fourth initial fully connected model included in the initial value impact information generation model to obtain a fourth output result; generating loss information for the fourth output result; determining whether the training of the initial value impact information generation model is completed according to the loss information.

[0011] Optionally, the generating loss information for the fourth output result includes: determining label loss information between the fourth output result and the corresponding sample label; generating model parameter value limit loss information according to each training parameter value included in the first initial fully connected model and each training parameter value included in the second initial fully connected model; generating the loss information according to the label loss information and the model parameter value limit loss information.

[0012] Optionally, the above method further includes: for each sample feature in the above sample feature set, performing the following first generation step: for each training sample in the above training sample set, performing the following second generation step: inputting the above training sample into the above value impact information generation model to generate value impact information; generating contribution information representing the contribution of the corresponding sample feature information to the above value impact information; generating a dot plot representing the obtained contribution information set and sample feature information set; according to the above dot plot, performing a monotonicity check on the monotonicity information corresponding to the above sample feature to obtain a check result.

[0013] Optionally, determining the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set includes: determining the feature value interval information corresponding to the above sample feature according to the above training sample set; performing interval information segmentation on the above feature value interval information to obtain an interval segmentation information set; determining the value impact information corresponding to each interval segmentation information in the above interval segmentation information set to obtain a value impact information set; determining the monotonicity information corresponding to the above sample feature according to the above value impact information set.

[0014] Optionally, each training parameter value included in the first fully connected model in the above value impact information generation model is greater than 0, the second fully connected model includes multiple fully connected layers connected in series, each training parameter value included in the second fully connected model is greater than 0, the model parameter value of the fourth fully connected model for the first output result is greater than 0, and the model parameter value of the fourth fully connected model for the second output result is less than 0.

[0015] Optionally, the training samples in the above training sample set are samples related to the value authorization risk control detection scenario.

[0016] Second aspect, some embodiments of the present disclosure provide a model training apparatus, including: a determination unit configured to determine monotonicity information corresponding to each sample feature in a sample feature set according to a training sample set; an execution unit configured to perform the following training steps for the training sample set: input a subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is positive monotonic information, into a first initial fully connected model included in an initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to a target training sample in the training sample set; input a subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is negative monotonic information, into a second initial fully connected model included in the initial value impact information generation model to obtain a second output result; input a subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is non-existent monotonic information, into a third initial fully connected model included in the initial value impact information generation model to obtain a third output result; determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result; in response to determining that the training is completed, determine the initial value impact information generation model as a value impact information generation model.

[0017] Optionally, the above apparatus further includes: in response to determining that the training is not completed, update model parameters of the initial value impact information generation model according to the first output result, the second output result, and the third output result to obtain an updated value impact information generation model, and remove the target training sample from the training sample set to obtain a post-removal training sample set; use the post-removal training sample set as the training sample set and the updated value impact information generation model as the initial value impact information generation model, and continue to perform the above training steps.

[0018] Optionally, the execution unit may be configured to: input the first output result, the second output result, and the third output result into a fourth initial fully connected model included in the initial value impact information generation model to obtain a fourth output result; generate loss information for the fourth output result; determine whether the training of the initial value impact information generation model is completed according to the loss information.

[0019] Optionally, the execution unit may be configured to: determine label loss information between the fourth output result and a corresponding sample label; generate model parameter value limit loss information according to each training parameter value included in the first initial fully connected model and each training parameter value included in the second initial fully connected model; generate the loss information according to the label loss information and the model parameter value limit loss information.

[0020] Optionally, the above device further includes: for each sample feature in the above sample feature set, performing the following first generation step: for each training sample in the above training sample set, performing the following second generation step: inputting the above training sample into the above value impact information generation model to generate value impact information; generating contribution information representing the contribution of the corresponding sample feature information to the above value impact information; generating a dot plot representing the obtained contribution information set and sample feature information set; according to the above dot plot, performing a monotonicity check on the monotonicity information corresponding to the above sample feature to obtain a check result.

[0021] Optionally, the determination unit may be configured to: determine the feature value interval information corresponding to the above sample feature according to the above training sample set; perform interval information segmentation on the above feature value interval information to obtain an interval segmentation information set; determine the value impact information corresponding to each interval segmentation information in the above interval segmentation information set to obtain a value impact information set; determine the monotonicity information corresponding to the above sample feature according to the above value impact information set.

[0022] Optionally, the training samples in the above training sample set are samples related to the value authorization risk control detection scenario.

[0023] In a third aspect, some embodiments of the present disclosure provide an information generation method, including: obtaining value impact associated data; inputting the above value impact associated data into a pre-trained value impact information generation model to generate value impact information, where the above value impact information generation model is generated based on a model training method.

[0024] In a fourth aspect, some embodiments of the present disclosure provide an information generation device, including: an obtaining unit configured to obtain value impact associated data; a generating unit configured to input the above value impact associated data into a pre-trained value impact information generation model to generate value impact information, where the above value impact information generation model is generated based on a model training method.

[0025] In a fifth aspect, some embodiments of the present disclosure provide an electronic device, including: one or more processors; a storage device having stored thereon one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method described in any implementation manner of the first aspect and the third aspect.

[0026] In a sixth aspect, some embodiments of the present disclosure provide a computer-readable medium having stored thereon a computer program, where the program when executed by a processor implements the method described in any implementation manner of the first aspect and the third aspect.

[0027] In a seventh aspect, some embodiments of the present disclosure provide a computer program product including a computer program which, when executed by a processor, implements the method described in any implementation manner of the first aspect and the third aspect above.

[0028] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the model training method of some embodiments of the present disclosure, a trained value impact information generation model can be utilized to accurately generate value impact information for input data. Specifically, the reason for the lack of precision in the relevant value impact information is as follows: In models such as linear regression and logistic regression, there is no cross-processing between features, resulting in limited application of the obtained interpretable model and inaccurate prediction problems. In addition, for the interpretable adjustment of multi-layer neural network models, there are often problems such as a relatively complex network structure and difficulty in ensuring the monotonicity between different input features and the target variable by simply constraining the weight values of each layer. Based on this, in the model training method of some embodiments of the present disclosure, first, according to the training sample set, the monotonicity information corresponding to each sample feature in the sample feature set can be accurately determined. Here, by determining the monotonicity information corresponding to each sample feature, the monotonicity information corresponding to the sample feature can be initially determined, so that when the training sample is input into the initial value impact information generation model later, the sample feature information in the training sample can be input into the corresponding fully connected model. Then, for the training sample set, the following training steps are performed: First step, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is positive monotonic information, into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result. Among them, the sample feature information set corresponds to the target training samples in the training sample set. Here, by inputting the subset of sample feature information with positive monotonic information into the first initial fully connected model, the positive monotonicity between the subset of sample features with positive monotonic information and the initial value impact information generation model is ensured, and the interpretability of the output result of the subsequent value impact information generation model is guaranteed. Second step, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is negative monotonic information, into the second initial fully connected model included in the initial value impact information generation model to obtain a second output result. Here, by inputting the subset of sample feature information with negative monotonic information into the second initial fully connected model, the negative monotonicity between the subset of sample features with negative monotonic information and the initial value impact information generation model is ensured, and the interpretability of the output result of the subsequent value impact information generation model is guaranteed. Third step, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is non-existent monotonic information, into the third initial fully connected model included in the initial value impact information generation model to obtain a third output result. Here, by inputting the subset of sample feature information with non-existent monotonic information into the third initial fully connected model, the sample feature content of the subset of sample feature information with non-existent monotonic information can be fully learned, and the model accuracy of the subsequent value impact information generation model is guaranteed. Fourth step, according to the first output result, the second output result, and the third output result, it can be accurately determined whether the training of the initial value impact information generation model is completed.In the fifth step, in response to determining the end of training, the initial value impact information generation model is determined as the value impact information generation model. In summary, by inputting the sample feature information corresponding to the monotonicity information into the corresponding fully-connected model, not only the monotonicity between some sample features in the sample feature set and the output variable is ensured, but also the output accuracy of the value impact information generation model is ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic and the elements and elements are not necessarily drawn to scale.

[0030] Figure 1 is a schematic diagram of an application scenario of a model training method according to some embodiments of the present disclosure;

[0031] Figure 2 is a flowchart of some embodiments of a model training method according to the present disclosure;

[0032] Figure 3 is a flowchart of some other embodiments of a model training method according to the present disclosure;

[0033] Figure 4 is a flowchart of some embodiments of an information generation method according to the present disclosure;

[0034] Figure 5 is a schematic structural diagram of some embodiments of a model training apparatus according to the present disclosure;

[0035] Figure 6 is a schematic structural diagram of some embodiments of an information generation apparatus according to the present disclosure;

[0036] Figure 7 is a schematic structural diagram of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not used to limit the protection scope of the present disclosure.

[0038] In addition, it should be noted that for the convenience of description, only parts related to the relevant invention are shown in the drawings. Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0039] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence relationship of the functions executed by these devices, modules or units.

[0040] It should be noted that the modifications of "one" and "multiple" mentioned in the present disclosure are illustrative rather than restrictive. Those skilled in the art should understand that unless otherwise clearly specified in the context, it should be understood as "one or more".

[0041] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only for illustrative purposes, and are not used to limit the scope of these messages or information.

[0042] For operations such as the collection, storage, and use of the information (such as sample feature information) involved in the present disclosure, before performing the corresponding operations, relevant organizations or individuals shall fulfill obligations including conducting an information security impact assessment, fulfilling the obligation of informing the information subject, and obtaining the prior authorization and consent of the information subject.

[0043] The present disclosure will be described in detail below with reference to the drawings and in combination with embodiments.

[0044] Figure 1 It is a schematic diagram of an application scenario of a model training method according to some embodiments of the present disclosure.

[0045] In Figure 1In the application scenario, first, the electronic device 101 can determine the monotonicity information corresponding to each sample feature in the sample feature set 104 according to the training sample set 102. In this application scenario, the monotonicity information corresponding to the first sample feature is positive monotonicity information. The monotonicity information corresponding to the second sample feature is positive monotonicity information. Then, for the training sample set 102, the electronic device 101 can perform the following training steps: In the first step, the electronic device 101 can input the subset of sample feature information in the sample feature information set 104, whose corresponding monotonicity information is positive monotonicity information, into the first initial fully connected model 106 included in the initial value impact information generation model 105, and obtain the first output result 109. Among them, the sample feature information set 104 corresponds to the target training sample 103 in the training sample set 102. In this application scenario, the sample feature information set 104 may include: the first sample feature information 1041, the second sample feature information 1042, the third sample feature information 1043, and the fourth sample feature information 1044. The subset of sample feature information with positive monotonicity information includes: the first sample feature information 1041 and the second sample feature information 1042. In the second step, the electronic device 101 can input the subset of sample feature information in the sample feature information set 104, whose corresponding monotonicity information is negative monotonicity information, into the second initial fully connected model 107 included in the initial value impact information generation model 105, and obtain the second output result 110. In this application scenario, the subset of sample feature information with negative monotonicity information may include: the third sample feature information 1043. In the third step, the electronic device 101 can input the subset of sample feature information in the sample feature information set 104, whose corresponding monotonicity information is non-existent monotonicity information, into the third initial fully connected model 108 included in the initial value impact information generation model 105, and obtain the third output result 111. In this application scenario, the subset of sample feature information with non-existent monotonicity information includes: the fourth sample feature information 1044. In the fourth step, the electronic device 101 can determine whether the training of the initial value impact information generation model 105 is completed according to the first output result 109, the second output result 110, and the third output result 111. In the fifth step, in response to determining that the training is completed, the electronic device 101 can determine the initial value impact information generation model 105 as the value impact information generation model.

[0046] It should be noted that the above-mentioned electronic device 101 can be hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster composed of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is embodied as software, it can be installed in the above-mentioned hardware devices. It can be implemented as, for example, multiple software or software modules for providing distributed services, or as a single software or software module. No specific limitation is made here.

[0047] It should be understood that Figure 1 the number of electronic devices in Figure 1 is merely illustrative. According to actual needs, there can be any number of electronic devices.

[0048] Continuing to refer to Figure 2 , a flowchart 200 of some embodiments of a model training method according to the present disclosure is shown. The model training method includes the following steps:

[0049] Step 201, according to a training sample set, determine the monotonicity information corresponding to each sample feature in the sample feature set.

[0050] In some embodiments, the execution subject of the above model training method (such as Figure 1 the electronic device 101 shown) can determine the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set. Among them, each sample feature in the sample feature set can be a feature that subsequently affects the generation of value impact information. In practice, the sample feature set can include, but is not limited to, at least one of the following: value acquisition feature, negative order feature, user credit feature. In practice, the value impact information can be information that affects the value information. For an e-commerce scenario, the value impact information can be bad debt information. The value acquisition feature can be a user income feature. The bad debt information can characterize the bad debt situation of the user for a target bill. In practice, the bad debt information can include: information indicating whether it is a bad debt and the bad debt probability. The bad debt probability can be a value between 0 and 1. The negative order feature can be a user malicious order feature. The user credit feature can be a user credit score feature. The monotonicity information can characterize the monotonicity (i.e., the transformation relationship between the information) between the feature variable corresponding to the sample feature and the target output variable corresponding to the value impact information. In practice, the monotonicity information can include: positive monotonicity information, negative monotonicity information, and non-existent monotonic information. The positive monotonicity information can characterize the positive monotonicity between the feature variable corresponding to the sample feature and the target variable corresponding to the value impact information (i.e., as the value of the feature variable corresponding to the sample feature increases, the value of the corresponding target variable increases). The negative monotonicity information can characterize the negative monotonicity between the feature variable corresponding to the sample feature and the target variable corresponding to the value impact information (i.e., as the value of the feature variable corresponding to the sample feature increases, the value of the corresponding target variable decreases). The non-existent monotonic information can characterize the non-existence of a monotonic relationship between the feature variable corresponding to the sample feature and the target variable corresponding to the value impact information (i.e., there is no monotonic relationship between the transformation of the value of the feature variable corresponding to the sample feature and the transformation of the value of the target variable).

[0051] In some optional implementation manners of some embodiments, the above determining the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set may include the following steps:

[0052] First step, according to the above training sample set, determine the feature value interval information corresponding to the above sample features. Among them, the feature value interval information can characterize the value magnitudes of the sample features in the training sample set.

[0053] As an example, first, the above execution entity can determine the feature values of each training sample in the training sample set on the sample features, obtaining a feature value set. Then, according to the maximum value and the minimum value in the above feature value set, determine the feature value interval information corresponding to the sample features.

[0054] Second step, perform interval information segmentation on the above feature value interval information, obtaining an interval segmentation information set. Among them, the number of each interval segmentation information included in the interval segmentation information set can be preset. For example, the number can be 20.

[0055] As an example, first, the above execution entity can sort the feature value set in descending order of feature values, obtaining a feature value sequence. Then, perform segmentation processing on the feature value sequence at intervals of a target number of feature values, obtaining multiple feature value subsequences. Finally, use the feature value subsequences as interval segmentation information, and use the multiple feature value subsequences as the interval segmentation information set.

[0056] Third step, determine the value impact information corresponding to each interval segmentation information in the above interval segmentation information set, obtaining a value impact information set.

[0057] As an example, for each interval segmentation information, first, determine the subset of training samples corresponding to the interval segmentation information. Then, determine the subset of value impact information corresponding to the subset of training samples. According to the subset of value impact information, determine the bad debt rate for the interval segmentation information as the value impact information.

[0058] Fourth step, according to the above value impact information set, determine the monotonicity information corresponding to the above sample features.

[0059] As an example, first, the above execution entity can sort the interval segmentation information set in ascending order of numerical values, obtaining an interval segmentation information sequence. Then, sort the value impact information set according to the corresponding relationship between the interval segmentation information and the value impact information, obtaining a value impact information sequence. Finally, according to the magnitudes of each interval segmentation information in the interval segmentation information sequence and the magnitudes of each value impact information in the value impact information sequence, determine the monotonicity information corresponding to the sample features.

[0060] As another example, the above execution entity can perform corresponding combination of the value impact information set and the interval segmentation information sequence to obtain an information group. Then, generate a trend chart for the information group. Finally, according to the trend chart, determine the monotonicity information corresponding to the sample features.

[0061] In some alternative implementations of some embodiments, the training samples in the above training sample set are samples related to the value authorization risk control detection scenario. In practice, for the financial field, the value authorization risk control detection scenario may be a risk control detection scenario for credit operations. For example, credit operations may include, but are not limited to, at least one of the following: loan operations, bill financing operations, financial leasing operations, and trade financing operations.

[0062] Step 202, for the training sample set, perform the following training steps:

[0063] Step 2021, input the subset of sample feature information in the sample feature information set corresponding to the positive monotonicity information into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result.

[0064] In some embodiments, the above execution subject may input the subset of sample feature information in the sample feature information set corresponding to the positive monotonicity information into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result. Among them, the initial value impact information generation model may be a value impact information generation model whose training has not ended. The value impact information generation model may be a neural network model for generating value impact information. The first initial fully connected model may be a first fully connected model whose training has not ended. The first fully connected model may include at least one fully connected layer. In practice, the first fully connected model includes: at least one fully connected layer connected in parallel. The first output result may represent the sample feature semantic information corresponding to the subset of sample feature information with positive monotonicity information. The first output result may be a result in matrix form. Among them, the sample feature information set corresponds to the target training samples in the training sample set.

[0065] In practice, the target training samples may be training samples randomly selected from the training sample set. The sample feature information set may be a feature information set corresponding to the sample feature set included in the target training samples.

[0066] Step 2022, input the subset of sample feature information in the sample feature information set corresponding to the negative monotonicity information into the second initial fully connected model included in the initial value impact information generation model to obtain a second output result.

[0067] In some embodiments, the above-mentioned execution entity may input a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is negative monotonic information, into the second initial fully-connected model included in the initial value impact information generation model to obtain a second output result. The second initial fully-connected model may be a second fully-connected model whose training has not ended. The second fully-connected model may include at least one fully-connected layer. In practice, the second fully-connected model includes: at least one fully-connected layer connected in parallel. The second output result may represent the sample feature semantic information corresponding to the subset of sample feature information with negative monotonic information. The second output result may be in the form of a matrix.

[0068] Step 2023: Input a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is non-existent monotonic information, into the third initial fully-connected model included in the initial value impact information generation model to obtain a third output result.

[0069] In some embodiments, the above-mentioned execution entity may input a subset of sample feature information in the sample feature information set, where the corresponding monotonicity information is non-existent monotonic information, into the third initial fully-connected model included in the initial value impact information generation model to obtain a third output result. The third initial fully-connected model may be a third fully-connected model whose training has not ended. The third fully-connected model may include at least one fully-connected layer. In practice, the third fully-connected model includes: at least one fully-connected layer connected in parallel. The third output result may represent the sample feature semantic information corresponding to the subset of sample feature information with non-existent monotonic information. The third output result may be in the form of a matrix.

[0070] Step 2024: Determine whether the initial value impact information generation model has finished training according to the first output result, the second output result, and the third output result.

[0071] In some embodiments, the above-mentioned execution entity may determine whether the initial value impact information generation model has finished training according to the first output result, the second output result, and the third output result.

[0072] As an example, first, the above-mentioned execution entity may splice the first output result, the second output result, and the third output result to obtain a spliced result. Then, input the spliced result into the activation function layer to generate model loss information. Next, in response to determining that the model loss information is less than the target value, determine that the initial value impact information generation model has finished training. In response to determining that the model loss information is greater than or equal to the target value, determine that the initial value impact information generation model has not finished training.

[0073] In some alternative implementations of some embodiments, determining whether the initial value impact information generation model has finished training based on the first output result, the second output result, and the third output result may include the following steps:

[0074] First, input the first output result, the second output result, and the third output result into the fourth initial fully connected model included in the initial value impact information generation model to obtain a fourth output result. Among them, the fourth initial fully connected model may be the fourth fully connected model when the model has not finished training. The fourth fully connected model may include at least one fully connected layer. In practice, the fourth fully connected model includes: at least one fully connected layer connected in parallel.

[0075] Second, generate loss information for the fourth output result.

[0076] As an example, the execution subject may input the fourth output result into a target loss function to generate the loss information. Among them, the target loss function may be a pre-set loss function. For example, the target loss function may be a cross-entropy loss function.

[0077] Third, determine whether the initial value impact information generation model has finished training according to the loss information.

[0078] As an example, in response to determining that the loss information is less than the target value, it is determined that the initial value impact information generation model has finished training. In response to determining that the loss information is greater than or equal to the target value, it is determined that the initial value impact information generation model has not finished training.

[0079] Optionally, generating the loss information for the fourth output result may include the following steps:

[0080] First, determine the label loss information between the fourth output result and the corresponding sample label. Among them, the sample label may be the label in the training sample. That is, the training sample includes: a sample label and a sample feature information set. The sample label may be pre-annotated value impact information. In practice, the sample label may be the actual value impact information corresponding to the user.

[0081] As an example, first, the execution subject may convert the sample label into a label vector. Then, input the fourth output result and the label vector into the cross-entropy loss function to generate cross-entropy loss information as the label loss information.

[0082] Step 2: Generate model parameter value limit loss information according to each training parameter value included in the first initial fully-connected model and each training parameter value included in the second initial fully-connected model. Among them, the model parameter value limit loss information may be the parameter value loss information of each model parameter in the first initial fully-connected model and the second initial fully-connected model.

[0083] As an example, the above execution entity inputs each training parameter value included in the first initial fully-connected model and each training parameter value included in the second initial fully-connected model into the training parameter value constraint loss function to generate model parameter value limit loss information. Among them, the training parameter value constraint loss function constrains that each training parameter value included in the first fully-connected model in the value impact information generation model is greater than 0. The second fully-connected model includes multiple fully-connected layers connected in series. Each training parameter value included in the last fully-connected layer of the second fully-connected model is greater than 0. Each training parameter value included in the second fully-connected model is greater than 0. In practice, the training parameter value constraint loss function may be a function obtained by performing weighted summation processing on each training parameter value based on the Linear rectification function.

[0084] As another example, the above execution entity inputs each training parameter value included in the first initial fully-connected model and each training parameter value included in the second initial fully-connected model into the training parameter value transformation constraint loss function to generate model parameter value limit loss information. Among them, the training parameter value transformation constraint loss function constrains the parameter transformation situation of each training parameter corresponding to each training parameter value. Specifically, the training parameter value transformation constraint loss function may be the mean squared error loss function.

[0085] Step 3: Generate the above loss information according to the above label loss information and the above model parameter value limit loss information.

[0086] As an example, first, the above execution entity may perform information weighted summation processing on the label loss information and the model parameter value limit loss information to obtain the weighted summation processing loss information as the loss information.

[0087] Step 2025: In response to determining that the training is over, determine the initial value impact information generation model as the value impact information generation model.

[0088] In some embodiments, the above execution entity may, in response to determining that the training is over, determine the initial value impact information generation model as the value impact information generation model. Among them, the value impact information generation model may be a neural network model after the model parameters are updated.

[0089] In some alternative implementations of some embodiments, each training parameter value included in the first fully-connected model in the above value impact information generation model is greater than 0. The second fully-connected model includes multiple fully-connected layers connected in series. Each training parameter value included in the second fully-connected model is greater than 0, and the model parameter values of the fourth fully-connected model for the first output result are greater than 0, and the model parameter values of the fourth fully-connected model for the second output result are less than 0.

[0090] As an example, the first fully-connected model includes 4 convolutional layers. That is, each training parameter value of the 4 convolutional layers is greater than 0. The second fully-connected model includes: 5 fully-connected layers connected in series. That is, each training parameter value of the 5 fully-connected layers connected in series is greater than 0. The fourth fully-connected layer includes 3 fully-connected layers connected in series. The parameter size of the model parameter values of the fourth fully-connected layer for the first output result is greater than 0. The parameter size of the model parameter values of the fourth fully-connected layer for the second output result is less than 0.

[0091] In some alternative implementations of some embodiments, after step 202, the steps further include:

[0092] First step, in response to determining that the training is not over, update the model parameters of the initial value impact information generation model according to the first output result, the second output result, and the third output result to obtain an updated value impact information generation model, and remove the target training sample from the training sample set to obtain a post-removal training sample set.

[0093] As an example, first, the above execution entity can generate model loss information according to the first output result, the second output result, and the third output result. Then, according to the above model loss information, use the backpropagation method to update the various model parameters in the initial value impact information generation model to obtain an updated value impact information generation model.

[0094] Second step, use the post-removal training sample set as the training sample set and the updated value impact information generation model as the initial value impact information generation model, and continue to execute the above training steps.

[0095] The above-mentioned various embodiments of the present disclosure have the following beneficial effects: Through the model training method of some embodiments of the present disclosure, a trained value impact information generation model can be utilized to accurately generate value impact information for input data. Specifically, the reason for the inaccuracy of the relevant value impact information is that there is no cross-processing between features in models such as linear regression and logistic regression, resulting in limited application of the obtained interpretable model and problems with inaccurate prediction. In addition, for the interpretable adjustment of multi-layer neural network models, there are often problems such as a relatively complex network structure and difficulty in ensuring the monotonicity between different input features and the target variable by simply constraining the weight values of each layer. Based on this, the model training method of some embodiments of the present disclosure, first, according to the training sample set, the monotonicity information corresponding to each sample feature in the sample feature set can be accurately determined. Here, by determining the monotonicity information corresponding to each sample feature, the monotonicity information corresponding to the sample feature can be initially determined, so that when the training sample is input into the initial value impact information generation model later, the sample feature information in the training sample can be input into the corresponding fully connected model. Then, for the training sample set, the following training steps are performed: First step, the subset of sample feature information in the sample feature information set corresponding to the positive monotonicity information is input into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result. Among them, the sample feature information set corresponds to the target training samples in the training sample set. Here, by inputting the subset of sample feature information with positive monotonicity information into the first initial fully connected model, the positive monotonicity between the subset of sample features with positive monotonicity information and the initial value impact information generation model is ensured, and the interpretability of the output result of the subsequent value impact information generation model is ensured. Second step, the subset of sample feature information in the sample feature information set corresponding to the negative monotonicity information is input into the second initial fully connected model included in the initial value impact information generation model to obtain a second output result. Here, by inputting the subset of sample feature information with negative monotonicity information into the second initial fully connected model, the negative monotonicity between the subset of sample features with negative monotonicity information and the initial value impact information generation model is ensured, and the interpretability of the output result of the subsequent value impact information generation model is ensured. Third step, the subset of sample feature information in the sample feature information set corresponding to the non-existent monotonicity information is input into the third initial fully connected model included in the initial value impact information generation model to obtain a third output result. Here, by inputting the subset of sample feature information with non-existent monotonicity information into the third initial fully connected model, the sample feature content of the subset of sample feature information with non-existent monotonicity information can be fully learned, and the model accuracy of the subsequent value impact information generation model is ensured. Fourth step, according to the first output result, the second output result and the third output result, it can be accurately determined whether the training of the initial value impact information generation model is completed.In the fifth step, in response to determining that the training is completed, the initial value impact information generation model is determined as the value impact information generation model. In summary, by inputting the sample feature information corresponding to the monotonicity information into the corresponding fully connected model, not only the monotonicity between some sample features in the sample feature set and the output variable is ensured, but also the output accuracy of the value impact information generation model is ensured.

[0096] Further referring to Figure 3 , a flow 300 of some other embodiments of the model training method according to the present disclosure is shown. The model training method includes the following steps:

[0097] Step 301, according to the training sample set, determine the monotonicity information corresponding to each sample feature in the sample feature set.

[0098] Step 302, for the training sample set, perform the following training steps:

[0099] Step 3021, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is positive monotonic information, into the first initial fully connected model included in the initial value impact information generation model, and obtain a first output result.

[0100] Step 3022, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is negative monotonic information, into the second initial fully connected model included in the initial value impact information generation model, and obtain a second output result.

[0101] Step 3023, input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is non-existent monotonic information, into the third initial fully connected model included in the initial value impact information generation model, and obtain a third output result.

[0102] Step 3024, determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result.

[0103] Step 3025, in response to determining that the training is completed, determine the initial value impact information generation model as the value impact information generation model.

[0104] In some embodiments, for the specific implementation of steps 301-302 and the technical effects brought thereby, reference can be made to Figure 2 Steps 201-202 in the corresponding embodiments, which will not be elaborated here.

[0105] Step 303, for each sample feature in the above sample feature set, perform the following first generation step:

[0106] Step 3031: For each training sample in the above training sample set, perform the following second generation steps:

[0107] Step 30311: Input the above training sample into the above value impact information generation model to generate value impact information.

[0108] In some embodiments, the execution entity (such as Figure 1 the electronic device 101 shown) may input the above training sample into the above value impact information generation model to generate value impact information.

[0109] Step 30312: Generate contribution information representing the contribution of the corresponding sample feature information to the above value impact information.

[0110] In some embodiments, the above execution entity may generate contribution information representing the contribution of the corresponding sample feature information to the above value impact information. Among them, the contribution information can represent the influence degree of the sample feature information on the value impact information. That is, the contribution information determines the information size of the value impact information to a certain extent.

[0111] As an example, the above execution entity may use the SHAP analysis method to generate collinear information of the corresponding sample feature information with respect to the above value impact information.

[0112] Step 3032: Generate a scatter plot representing the obtained contribution information set and sample feature information set.

[0113] In some embodiments, the above execution entity may generate a scatter plot representing the obtained contribution information set and sample feature information set.

[0114] Step 3033: According to the above scatter plot, perform a monotonicity check on the monotonicity information corresponding to the above sample features to obtain a check result.

[0115] In some embodiments, the above execution entity may perform a monotonicity check on the monotonicity information corresponding to the above sample features according to the above scatter plot to obtain a check result. Among them, the check result includes: a result indicating that the monotonicity information corresponding to the sample features is correct, and a result indicating that the monotonicity information corresponding to the sample features is incorrect.

[0116] As an example, first, the above execution entity may perform a monotonicity check on the monotonicity information corresponding to the above sample features by analyzing the graph transformation trend in the scatter plot to obtain a check result.

[0117] From Figure 3 it can be seen that compared with the description of some embodiments corresponding to Figure 2 Figure 3The process 300 of the model training method in some corresponding embodiments ensures the accuracy of the output result of the value impact information generation model by verifying the monotonicity of each sample feature to avoid errors in the subsequent input of feature information into the fully connected model.

[0118] Continuing to refer to Figure 4 FIG. 400 shows a process of some embodiments of an information generation method according to the present disclosure. The information generation method includes the following steps:

[0119] Step 401, obtaining value impact associated data.

[0120] In some embodiments, the execution subject of the above information generation method (e.g., Figure 1 the electronic device 101 shown) can obtain the value impact associated data by wired or wireless means. Among them, the value impact associated data can be the model input data to be used for subsequent value impact information prediction. Specifically, the value impact associated data may include: a set of feature information corresponding to the feature set. Each feature included in the feature set is the same as each sample feature included in the sample feature set. In practice, for the financial field, the value impact associated data may be credit associated data. That is, the credit associated data may be data associated with the credit operation.

[0121] Step 402, inputting the above value impact associated data into a pre-trained value impact information generation model to generate value impact information.

[0122] In some embodiments, the above execution subject can input the above value impact associated data into a pre-trained value impact information generation model to generate value impact information. Among them, the above value impact information generation model is generated based on the model training method.

[0123] The above embodiments of the present disclosure have the following beneficial effects: The information generation method according to some embodiments of the present disclosure can accurately generate value impact information for the value impact associated data.

[0124] Further referring to Figure 5 As an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a model training apparatus. These apparatus embodiments correspond to Figure 2 the method embodiments shown, and the model training apparatus can be specifically applied to various electronic devices.

[0125] As shown in Figure 5As shown in the figure, a model training device 500 includes: a determination unit 501 and an execution unit 502. Among them, the determination unit 501 is configured to determine the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set; the execution unit 502 is configured to perform the following training steps for the training sample set: input the subset of sample feature information in the sample feature information set corresponding to the positive monotonicity information into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to the target training samples in the training sample set; input the subset of sample feature information in the sample feature information set corresponding to the negative monotonicity information into the second initial fully connected model included in the initial value impact information generation model to obtain a second output result; input the subset of sample feature information in the sample feature information set corresponding to the non-existent monotonicity information into the third initial fully connected model included in the initial value impact information generation model to obtain a third output result; determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result and the third output result; in response to determining that the training is completed, determine the initial value impact information generation model as the value impact information generation model.

[0126] In some optional implementation manners of some embodiments, the above device 500 further includes: an update unit and a first step execution unit (not shown in the figure). Among them, the above update unit can be configured to: in response to determining that the training is not completed, update the model parameters of the initial value impact information generation model according to the first output result, the second output result and the third output result to obtain an updated value impact information generation model, and remove the target training samples from the training sample set to obtain a post-removal training sample set. The first step execution unit can be configured to: use the post-removal training sample set as the training sample set and the updated value impact information generation model as the initial value impact information generation model, and continue to perform the above training steps.

[0127] In some optional implementation manners of some embodiments, the execution unit 502 can be further configured to: input the above first output result, the above second output result and the above third output result into the fourth initial fully connected model included in the initial value impact information generation model to obtain a fourth output result; generate loss information for the above fourth output result; determine whether the training of the above initial value impact information generation model is completed according to the above loss information.

[0128] In some alternative implementations of some embodiments, the execution unit 502 may be further configured to: determine the label loss information between the fourth output result and the corresponding sample label; generate model parameter value limit loss information according to each training parameter value included in the first initial fully connected model and each training parameter value included in the second initial fully connected model; and generate the loss information according to the label loss information and the model parameter value limit loss information.

[0129] In some alternative implementations of some embodiments, the apparatus 500 further includes: a second step execution unit (not shown in the figure). The second step execution unit may be configured to: for each sample feature in the sample feature set, perform the following first generation step: for each training sample in the training sample set, perform the following second generation step: input the training sample into the value impact information generation model to generate value impact information; generate contribution information representing the contribution of the corresponding sample feature information to the value impact information; generate a dot plot representing the obtained contribution information set and the sample feature information set; and perform a monotonicity check on the monotonicity information corresponding to the sample feature according to the dot plot to obtain a check result.

[0130] In some alternative implementations of some embodiments, the determination unit 501 may be further configured to: determine the feature value range information corresponding to the sample feature according to the training sample set; perform interval information segmentation on the feature value range information to obtain an interval segmentation information set; determine the value impact information corresponding to each interval segmentation information in the interval segmentation information set to obtain a value impact information set; and determine the monotonicity information corresponding to the sample feature according to the value impact information set.

[0131] In some alternative implementations of some embodiments, each training parameter value included in the first fully connected model in the value impact information generation model is greater than 0, the second fully connected model includes multiple fully connected layers connected in series, each training parameter value included in the second fully connected model is greater than 0, the model parameter value of the fourth fully connected model for the first output result is greater than 0, and the model parameter value of the fourth fully connected model for the second output result is less than 0.

[0132] In some alternative implementations of some embodiments, the training samples in the training sample set are samples related to the value authorization risk control detection scenario.

[0133] It can be understood that the various units described in the model training apparatus 500 are related to the reference Figure 2corresponds to each step in the described method. Thus, the operations, features, and beneficial effects described above for the method also apply to the model training device 500 and the units included therein, and will not be elaborated here.

[0134] Further referring to Figure 6 , as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of an information generation device, and these device embodiments correspond to Figure 4 the method embodiments shown, and the information generation device can be specifically applied to various electronic devices.

[0135] As Figure 6 shown, an information generation device 600 includes: an acquisition unit 601 and a generation unit 602. Among them, the acquisition unit 601 is configured to acquire value impact association data; the generation unit 602 is configured to input the above value impact association data into a pre-trained value impact information generation model to generate value impact information, where the above value impact information generation model is generated based on a model training method.

[0136] It can be understood that the units described in the information generation device 600 correspond to each step in the method described with reference to Figure 4 . Thus, the operations, features, and beneficial effects described above for the method also apply to the information generation device 600 and the units included therein, and will not be elaborated here.

[0137] Next, referring to Figure 7 , which shows a schematic structural diagram of an electronic device (such as the electronic device 101 in Figure 1 ) 700 suitable for implementing some embodiments of the present disclosure. Figure 7 The electronic device shown is only an example and should not impose any limitations on the functions and usage scopes of the embodiments of the present disclosure.

[0138] As Figure 7 shown, the electronic device 700 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in the read-only memory 702 or a program loaded from the storage device 708 into the random access memory 703. In the random access memory 703, various programs and data required for the operation of the electronic device 700 are also stored. The processing device 701, the read-only memory 702, and the random access memory 703 are connected to each other through a bus 704. The input / output interface 705 is also connected to the bus 704.

[0139] Typically, the following devices can be connected to the input / output interface 705: an input device 706 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 709. The communication device 709 can allow the electronic device 700 to communicate with other devices wirelessly or wireline to exchange data. Although Figure 7 the electronic device 700 with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices can be implemented or had. Figure 7 Each block shown in

[0140]

[0141] Specifically, according to some embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of the present disclosure include a computer program product that includes a computer program carried on a computer-readable medium, and the computer program includes program codes for performing the methods shown in the flowcharts. In such some embodiments, the computer program can be downloaded and installed from a network through the communication device 709, or installed from the storage device 708, or installed from the read-only memory 702. When the computer program is executed by the processing device 701, the above functions defined in the methods of some embodiments of the present disclosure are performed.It should be noted that, in some embodiments of the present disclosure, the above-mentioned computer-readable medium may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In some embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of the present disclosure, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium, and the computer-readable signal medium may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0142] In some embodiments, the client and the server may communicate using any currently known or future-developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with digital data communication in any form or medium (e.g., a communication network). Examples of communication networks include local area networks ("LANs"), wide area networks ("WANs"), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0143] The above computer-readable medium may be included in the above electronic device; or it may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs. When the above one or more programs are executed by the electronic device, the electronic device is caused to: determine the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set; for the training sample set, perform the following training steps: input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is positive monotonic information, into the first initial fully-connected model included in the initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to the target training samples in the training sample set; input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is negative monotonic information, into the second initial fully-connected model included in the initial value impact information generation model to obtain a second output result; input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is non-existent monotonic information, into the third initial fully-connected model included in the initial value impact information generation model to obtain a third output result; determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result and the third output result; in response to determining that the training is completed, determine the initial value impact information generation model as the value impact information generation model. Obtain value impact association data; input the above value impact association data into the pre-trained value impact information generation model to generate value impact information, where the above value impact information generation model is generated based on the model training method.

[0144] Computer program code for performing the operations of some embodiments of the present disclosure may be written in one or more programming languages or combinations thereof. The above programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, executed as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (for example, by using an Internet service provider to connect through the Internet).

[0145] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a portion of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur in a different order than that noted in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or by a combination of dedicated hardware and computer instructions.

[0146] The units described in some embodiments of the present disclosure can be implemented in software or in hardware. The described units can also be provided in a processor. For example, it can be described as: a processor includes a determination unit and an execution unit. Among them, the names of these units do not constitute a limitation on the unit itself in some cases. For example, the determination unit can also be described as "the unit that determines the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set".

[0147] The functions described above herein can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on Chip (SOCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0148] Some embodiments of the present disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the above model training methods or information generation methods.

[0149] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A model training method, comprising: Determine the monotonicity information corresponding to each sample feature in the sample feature set according to the training sample set; For the training sample set, perform the following training steps: Input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is positive monotonic information, into the first initial fully-connected model included in the initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to the target training samples in the training sample set; Input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is negative monotonic information, into the second initial fully-connected model included in the initial value impact information generation model to obtain a second output result; Input the subset of sample feature information in the sample feature information set, whose corresponding monotonicity information is non-existent monotonic information, into the third initial fully-connected model included in the initial value impact information generation model to obtain a third output result; Determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result; In response to determining that the training is completed, determine the initial value impact information generation model as the value impact information generation model.

2. The method according to claim 1, wherein, The method further includes: In response to determining that the training is not completed, update the model parameters of the initial value impact information generation model according to the first output result, the second output result, and the third output result to obtain an updated value impact information generation model, and remove the target training samples from the training sample set to obtain a post-removal training sample set; Use the post-removal training sample set as the training sample set and the updated value impact information generation model as the initial value impact information generation model, and continue to execute the training steps.

3. The method according to claim 1, wherein, The determining whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result includes: Input the first output result, the second output result, and the third output result into the fourth initial fully-connected model included in the initial value impact information generation model to obtain a fourth output result; Generate loss information for the fourth output result; Determine whether the training of the initial value impact information generation model is completed according to the loss information.

4. The method according to claim 3, wherein, The generating loss information for the fourth output result includes: Determine the label loss information between the fourth output result and the corresponding sample label; Generate model parameter value limit loss information according to each training parameter value included in the first initial fully-connected model and each training parameter value included in the second initial fully-connected model; Generate the loss information according to the label loss information and the model parameter value limit loss information.

5. The method according to claim 1, wherein, The method further includes: For each sample feature in the sample feature set, perform the following first generation step: For each training sample in the training sample set, perform the following second generation step: Input the training sample into the value impact information generation model to generate value impact information; Generate contribution information representing the contribution of the corresponding sample feature information to the value impact information; Generate a dot plot representing the obtained contribution information set and sample feature information set; According to the dot plot, perform a monotonicity check on the monotonicity information corresponding to the sample feature to obtain a check result.

6. The method according to claim 1, wherein, The determining, according to the training sample set, the monotonicity information corresponding to each sample feature in the sample feature set includes: According to the training sample set, determine the feature value interval information corresponding to the sample feature; Perform interval information segmentation on the feature value interval information to obtain an interval segmentation information set; Determine the value impact information corresponding to each interval segmentation information in the interval segmentation information set to obtain a value impact information set; According to the value impact information set, determine the monotonicity information corresponding to the sample feature.

7. The method according to claim 3, wherein, Each training parameter value included in the first fully connected model in the value impact information generation model is greater than 0, the second fully connected model includes multiple fully connected layers connected in series, each training parameter value included in the second fully connected model is greater than 0, the model parameter values of the fourth fully connected model for the first output result are greater than 0, and the model parameter values of the fourth fully connected model for the second output result are less than 0.

8. The method according to claim 1, wherein, The training samples in the training sample set are samples related to the value authorization risk control detection scenario.

9. An information generation method, comprising: Obtain value impact associated data; Input the value impact associated data into a pre-trained value impact information generation model to generate value impact information, where the value impact information generation model is generated based on the method described in any one of claims 1-8.

10. A model training device, comprising: A determination unit configured to determine, according to the training sample set, the monotonicity information corresponding to each sample feature in the sample feature set; An execution unit configured to perform the following training steps for the training sample set: input a subset of the sample feature information in the sample feature information set corresponding to the monotonicity information as positive monotonic information into the first initial fully connected model included in the initial value impact information generation model to obtain a first output result, where the sample feature information set corresponds to the target training samples in the training sample set; input a subset of the sample feature information in the sample feature information set corresponding to the monotonicity information as negative monotonic information into the second initial fully connected model included in the initial value impact information generation model to obtain a second output result; input a subset of the sample feature information in the sample feature information set corresponding to the monotonicity information as non-existent monotonic information into the third initial fully connected model included in the initial value impact information generation model to obtain a third output result; determine whether the training of the initial value impact information generation model is completed according to the first output result, the second output result, and the third output result; In response to determining that the training is completed, determine the initial value impact information generation model as the value impact information generation model.

11. An information generation device, comprising: An acquisition unit configured to acquire value impact associated data; A generation unit configured to input the value impact associated data into a pre-trained value impact information generation model to generate value impact information, where the value impact information generation model is generated based on the method described in any one of claims 1-8.

12. An electronic device, comprising: One or more processors; A storage device on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-9.

13. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, the method according to any one of claims 1-9 is implemented.

14. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-9.

Images