Method and device for calibrating behavior prediction probability, storage medium and electronic equipment

Abstract

The embodiment of the invention provides a method and device for calibrating behavior prediction probabilities, a storage medium and electronic equipment, and the method comprises the steps: obtaining the behavior prediction probabilities of abnormal behaviors of N objects; based on the N behavior prediction probabilities and a first initial function, calibrating the N behavior prediction probabilities to obtain N first behavior probabilities; calculating gradients of the M first parameters by using the N first behavior probabilities, N real behavior tags corresponding to the N objects and a first objective function; iteratively updating the M first parameters based on the gradient of the M first parameters, and determining target group parameter values corresponding to the M first parameters under the condition that the M first parameters meet a preset condition; adjusting the preset function based on the target group parameter values corresponding to the M first parameters to obtain a second target function; and respectively calibrating the N behavior prediction probabilities by using a second target function to obtain N target behavior probabilities.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computers, and more specifically, to a method and device for calibrating behavior prediction probability, a storage medium, and an electronic device. Background Art

[0002] Score calibration is used to adjust the output of the model. In related technologies, the more common calibration schemes are Pratt scaling and binning calibration, but Pratt scaling requires the data to conform to a certain distribution function, that is, the normal distribution characteristics.

[0003] Bin calibration makes the predicted default probability close to the actual sample default probability by fitting the predicted bad percentage and the actual bad percentage of samples in each bin. However, this method may cause the results to be uncalibrated or destroy the order preservation of the data because there are no bad samples in the bucket or the distribution is not monotonic. Summary of the invention

[0004] The embodiments of the present application provide a method and device for calibrating behavior prediction probability, a storage medium, and an electronic device, so as to at least solve the problem in the related art that the prediction probability of abnormal behavior of a sample object cannot be accurately calibrated.

[0005] According to one embodiment of the present application, a method for calibrating behavior prediction probabilities is provided, including: obtaining behavior prediction probabilities of N objects having abnormal behaviors, wherein the behavior prediction probabilities are probability values ​​output by a target model based on the behavior data of the objects, and N is a positive integer greater than 0; based on the N behavior prediction probabilities and a first initial function, calibrating the N behavior prediction probabilities to obtain N first behavior probabilities, wherein the first initial function includes preset M first parameters, and M is a positive integer greater than 0, and the first initial function is used to calibrate the behavior prediction probabilities; using the N first behavior probabilities, the N real behavior parameters corresponding to the N objects, signature and a first objective function, calculate the gradients of the M first parameters, wherein the real behavior label is used to represent the real behavior of the object, and the first objective function includes the preset M first parameters; based on the gradients of the M first parameters, iteratively update the M first parameters, and when the M first parameters meet the preset conditions, determine the target group parameter values ​​corresponding to the M first parameters; adjust the preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second objective function, wherein the preset function is a monotonic function; use the second objective function to calibrate the N behavior prediction probabilities respectively to obtain N target behavior probabilities.

[0006] In an exemplary embodiment, before obtaining the behavior prediction probability of abnormal behavior of N objects, the method further includes: obtaining first information of the N objects respectively, wherein the first information includes behavior data and object information of the objects; inputting the first information of the N objects into a target model to obtain the behavior prediction probability of abnormal behavior of each object output by the target model; wherein the target model is a logistic regression model obtained by iteratively training an initial model using sample information of multiple groups of sample objects, the initial model includes a feature extraction sub-model and a classification sub-model, for each group of sample information, the feature extraction sub-model is used to convert the sample information into preset variables, and extract target variables that meet preset indicators from the preset variables, the classification sub-model is used to classify the behavior of the sample objects based on the target variables, the sample information includes sample behavior data, sample object information and real behavior data of the sample objects, and the preset indicators include weighted evidence indicators of the preset variables and stability indicators of the preset variables.

[0007] In an exemplary embodiment, the gradients of the M first parameters are calculated using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and the first objective function, including: performing partial derivative operations on the M first parameters respectively through the first objective function to obtain partial derivatives corresponding to the M first parameters; substituting the N first behavior probabilities and the N real behavior labels into the partial derivatives corresponding to the M first parameters to calculate the gradients corresponding to the M first parameters.

[0008] In an exemplary embodiment, based on the gradients of the M first parameters, the M first parameters are iteratively updated, and when the M first parameters meet the preset conditions, the target group parameter values ​​corresponding to the M first parameters are determined, including: when the gradients of the M first parameters do not meet the preset conditions, iterative calculation is performed on the gradients of the M first parameters, wherein, when the gradients of the M first parameters are subjected to the j-th round of iterative calculation, the steps performed include: obtaining M first target parameters, wherein the first target parameters are parameters obtained after the j-1-th round of updating; based on the update formulas of the M parameters and the M first target parameters, M second target parameters are calculated, wherein the M second target parameters are parameters obtained after the j-th round of iterative calculation; when the M second target parameters meet the preset conditions, the M second target parameters are determined as target group parameter values; wherein the j is a positive integer greater than 1.

[0009] In an exemplary embodiment, adjusting a preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second objective function includes: inputting the target group parameter values ​​into the M second parameters included in the preset function to obtain the second objective function, wherein the preset function is: The above A is the first subparameter of the M above second parameters, the above B is the second subparameter of the M above second parameters, the above first subparameter is used to represent the shape of the nonlinear relationship controlling the above preset function, and the above first subparameter is greater than 0, the above second subparameter is used to adjust the offset degree of the above preset function, the above i is used to represent the i-th object of the N above objects, and the above Y i It is used to represent the probability corresponding to the true behavior label of the above i-th object. The above p i Used to represent the predicted probability of the behavior of the i-th object mentioned above.

[0010] In an exemplary embodiment, the N behavior prediction probabilities are calibrated respectively using the second objective function to obtain N target behavior probabilities, including: for each of the behavior prediction probabilities of the objects, calibration is performed through the following steps: the probability corresponding to the true behavior label is input into the second objective function to calculate the calibration probability; the calibration probability is compared with the behavior prediction probability; when the difference between the calibration probability and the behavior prediction probability is less than or equal to a preset threshold, the behavior prediction probability is determined as the target behavior probability; when the difference between the calibration probability and the behavior prediction probability is greater than the preset threshold, the calibration probability is determined as the target behavior probability.

[0011] In an exemplary embodiment, after calibrating the N above-mentioned behavior prediction probabilities respectively using the above-mentioned second objective function to obtain N target behavior probabilities, the above-mentioned method also includes: calculating N target behavior probability ratios corresponding to the N above-mentioned target behavior probabilities, wherein the above-mentioned target behavior probability ratio is used to represent the ratio between the probability of the above-mentioned object occurring the above-mentioned abnormal behavior and the probability of not occurring the above-mentioned abnormal behavior; inputting the N above-mentioned target behavior probability ratios into the conversion function, and calculating the abnormal behavior scores of the N above-mentioned objects occurring the above-mentioned abnormal behaviors through the above-mentioned conversion function, wherein the above-mentioned conversion function is a function used to convert the above-mentioned target behavior probabilities into scores in a preset format.

[0012] According to another embodiment of the present application, a device for calibrating behavior prediction probabilities is provided, including: an acquisition module, used to acquire the behavior prediction probabilities of N objects having abnormal behaviors, wherein the above-mentioned behavior prediction probabilities are probability values ​​output by the target model based on the behavior data of the above-mentioned objects, and the above-mentioned N is a positive integer greater than 0; a first calibration module, used to calibrate the N above-mentioned behavior prediction probabilities based on the above-mentioned N behavior prediction probabilities and a first initial function to obtain N first behavior probabilities, wherein the above-mentioned first initial function includes preset M first parameters, and the above-mentioned M is a positive integer greater than 0, and the above-mentioned first initial function is used to calibrate the behavior prediction probabilities; a calculation module, used to use the N above-mentioned first behavior probabilities, the N real behaviors corresponding to the N above-mentioned objects A label and a first objective function are used to calculate the gradients of the M first parameters, wherein the real behavior label is used to represent the real behavior of the object, and the first objective function includes the preset M first parameters; a determination module is used to iteratively update the M first parameters based on the gradients of the M first parameters, and determine the target group parameter values ​​corresponding to the M first parameters when the M first parameters meet the preset conditions; an adjustment module is used to adjust the preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second objective function, wherein the preset function is a monotonic function; a second calibration module is used to use the second objective function to calibrate the N behavior prediction probabilities respectively to obtain N target behavior probabilities.

[0013] According to another embodiment of the present application, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0014] According to another embodiment of the present application, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0015] Through the present application, based on the behavior prediction probability of abnormal behavior occurring in N objects and its corresponding real behavior label and the first objective function, M first parameters are iteratively calculated, and finally the parameter values ​​of the M first parameters that meet the preset conditions are determined as the target group parameter values, and finally the second objective function is obtained, and the behavior prediction probability of abnormal behavior occurring in N objects is calibrated based on the second objective function to obtain the target behavior probability of abnormal behavior occurring in N objects, so that the calculated target behavior probability is closer to the real label of the object and is applicable to a wider range of object types. Therefore, the problem that the predicted probability of abnormal behavior occurring in sample objects cannot be accurately calibrated in the related art can be solved, thereby achieving the effect of accurately calibrating the predicted probability of abnormal behavior occurring in sample objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of a hardware environment of a method for calibrating behavior prediction probability according to an embodiment of the present application;

[0017] Figure 2 is a flow chart of a method for calibrating behavior prediction probability according to an embodiment of the present application;

[0018] Figure 3 is a structural block diagram of a device for calibrating behavior prediction probability according to an embodiment of the present application;

[0019] Figure 4 It is a flowchart of a method for calibrating behavior prediction probability in the financial field according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] The embodiments of the present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0021] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0022] The method embodiments provided in the embodiments of the present application can be executed in a server device or a similar computing device. Taking running on a server device as an example, Figure 1 Schematic diagram of the hardware environment of a method for calibrating behavior prediction probability according to an embodiment of the present application. Figure 1 As shown, the server device may include one or more ( Figure 1Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the server device may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above server device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0023] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the method for calibrating the prediction probability of behavior in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the server device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0024] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the server device. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0025] In this embodiment, a method for calibrating behavior prediction probability is provided. Figure 2 is a flow chart of a method for calibrating behavior prediction probability according to an embodiment of the present application, such as Figure 2 As shown, the process includes the following steps:

[0026] Step S202, obtaining the behavior prediction probability of abnormal behavior of N objects, wherein the behavior prediction probability is a probability value output by the target model based on the behavior data of the objects, and N is a positive integer greater than 0;

[0027] Step S204, based on the N above-mentioned behavior prediction probabilities and the first initial function, calibrate the N above-mentioned behavior prediction probabilities to obtain N first behavior probabilities, wherein the above-mentioned first initial function includes preset M first parameters, the above-mentioned M is a positive integer greater than 0, and the above-mentioned first initial function is used to calibrate the behavior prediction probabilities.

[0028] Step S206, calculating the gradients of the M first parameters using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and the first objective function, wherein the real behavior labels are used to represent the real behaviors of the objects, and the first objective function includes the preset M first parameters;

[0029] Step S208, iteratively updating the M first parameters based on the gradients of the M first parameters, and determining target group parameter values ​​corresponding to the M first parameters when the M first parameters meet a preset condition;

[0030] Step S210, adjusting the preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second target function, wherein the preset function is a monotonic function;

[0031] Step S212: calibrate the N behavior prediction probabilities respectively using the second objective function to obtain N target behavior probabilities.

[0032] Optionally, this embodiment can be applied in the field of gaming to predict the cheating behavior of players so as to further monitor it; or, it can be applied in the field of finance to calculate the probability of default of customers so as to further select their appropriate management strategies; or, it can be applied in the field of e-commerce to predict and calibrate the post-purchase return behavior of users so as to further make the recommended products more suitable for the actual needs of users, etc.

[0033] Optionally, the N objects in this embodiment do not limit their corresponding sample types, that is, the N object data do not need to conform to a certain distribution function, nor do they need to be binned.

[0034] Optionally, the abnormal behavior of the object in this embodiment corresponds to different behaviors in different fields. For example, in the gaming field, the abnormal behavior is the player's cheating behavior; in the financial field, the abnormal behavior is the customer's breach of contract behavior; in the e-commerce field, the abnormal behavior is the user's post-purchase return behavior.

[0035] Optionally, the real behavior label in this embodiment is a binary label indicating whether the object actually has abnormal behavior, which can reflect the real behavior status of the object. The real behavior label is determined based on subsequent observation and analysis by comparing with preset abnormal behavior standards, helping the model to distinguish between normal behavior and abnormal behavior.

[0036] For example, in the field of gaming, abnormal behavior refers to players using plug-ins or other unfair means to play games, so the real behavior label can be "using plug-ins" or "not using plug-ins". For example, if the monitoring system detects that a player uses plug-ins during the game, its real behavior label is marked as "1", indicating that the player has indeed performed abnormal behavior. On the contrary, if the player does not use plug-ins throughout the game, its label is marked as "0", indicating that no abnormal behavior has been performed; in the credit or payment services in the financial field, abnormal behavior refers to the customer's default or fraudulent behavior, so the real behavior label can be "default" or "non-default", "fraud" or "non-fraud". For example, if a customer fails to repay the loan on time within the repayment period, its real behavior label is The label is marked as "1", indicating a default; if the customer repays all loans on time, its label is marked as "0", indicating no default; in the field of e-commerce, abnormal behavior may refer to the behavior of users frequently returning goods after purchase. Therefore, the real behavior label can be "frequent returns" or "normal purchases". For example, if a user purchases a large number of goods within a certain period of time and then returns all of them, its real behavior label may be marked as "1", indicating abnormal post-purchase return behavior. If the user does not return the goods after purchasing them, or the return ratio is within a reasonable range, its label is marked as "0", indicating normal purchase behavior.

[0037] Optionally, in this embodiment, the numerical values ​​of the preset M first parameters can be set by technicians in this field based on experience, or can be calculated in the following way: using N behavior prediction probabilities and N real behavior labels corresponding to N objects, respectively calculate the parameter values ​​of the M first parameters in the preset function to obtain multiple groups of parameter values ​​corresponding to the M first parameters; then input each behavior prediction probability and each corresponding real behavior label into the preset function to obtain N first preset functions; arbitrarily combine the N first preset functions according to the number of M first parameters to obtain multiple groups of first functions; use each group of first functions to calculate the parameter values ​​of the M first parameters to obtain multiple groups of parameter values; then input the multiple groups of parameter values ​​into the third objective function respectively to obtain the loss value corresponding to each group of parameter values ​​calculated by the third objective function, wherein the third objective function includes variables corresponding to the M first parameters, and the third objective function is the loss function of the target model; compare the N loss values, and determine a group of parameter values ​​corresponding to the loss values ​​that meet the preset loss values ​​as the final preset M first parameters.

[0038] Optionally, the target behavior probability in this embodiment is the behavior prediction probability after calibration with the second objective function, and the gap value between the target behavior probability and the actual behavior label satisfies a preset value, that is, the calibrated prediction is closer to the actual behavior state.

[0039] Optionally, the preset function in this embodiment is a mathematical function used to associate the behavior prediction probability with the actual behavior label to solve or optimize M parameters. The preset function can be a function used to evaluate the prediction accuracy of the model, which is a probability calibration function, and its purpose is to make the prediction result closer to the actual behavior by adjusting the parameters.

[0040] Through the present application, based on the behavior prediction probability of abnormal behavior occurring in N objects and its corresponding real behavior label and the first objective function, M first parameters are iteratively calculated, and finally the parameter values ​​of the M first parameters that meet the preset conditions are determined as the target group parameter values, and finally the second objective function is obtained, and the behavior prediction probability of abnormal behavior occurring in N objects is calibrated based on the second objective function to obtain the target behavior probability of abnormal behavior occurring in N objects, so that the calculated target behavior probability is closer to the real label of the object and is applicable to a wider range of object types. Therefore, the problem that the predicted probability of abnormal behavior occurring in sample objects cannot be accurately calibrated in the related art can be solved, thereby achieving the effect of accurately calibrating the predicted probability of abnormal behavior occurring in sample objects.

[0041] In an exemplary embodiment, before obtaining the behavior prediction probability of abnormal behavior of N objects, the method further includes: obtaining first information of the N objects respectively, wherein the first information includes behavior data and object information of the objects; inputting the first information of the N objects into a target model to obtain the behavior prediction probability of abnormal behavior of each object output by the target model; wherein the target model is a logistic regression model obtained by iteratively training an initial model using sample information of multiple groups of sample objects, the initial model includes a feature extraction sub-model and a classification sub-model, for each group of sample information, the feature extraction sub-model is used to convert the sample information into preset variables, and extract target variables that meet preset indicators from the preset variables, the classification sub-model is used to classify the behavior of the sample objects based on the target variables, the sample information includes sample behavior data, sample object information and real behavior data of the sample objects, and the preset indicators include weighted evidence indicators of the preset variables and stability indicators of the preset variables.

[0042] Optionally, the first information in this embodiment is information with certain distinguishing power and high stability. The first information includes the behavior data and object information of the object. In the financial field, the behavior data of the object may include various activity records of the customer in a bank or financial institution, such as loan repayment records, credit card usage, bill payment history, transaction frequency, transaction amount, number of overdue payments, credit inquiry records, etc. The object information may include the customer's personal information (such as age, gender, occupation), income information, asset information, bank account information, credit report history, loan type and amount, etc.; in the e-commerce field, the behavior data of the object may include the user's shopping history, return history, shopping frequency, shopping amount, shopping time, purchased product types, shopping cart behavior (such as putting in the shopping cart, removing, closing The object information may include user registration information, historical shopping preferences, payment method preferences, delivery address information, user level, whether there is a violation record, etc. In the field of games, the behavior data of the object may include various behavior records of players in the game, such as game time, in-game purchase records, game score changes, interaction records with other players in the game (such as chatting, trading, teaming, etc.), frequency of actions in the game (such as using skills, attacking, moving, etc.), game device information (such as the type of hardware used, operating system, etc.), etc. The object information may include the player's account information, registration information, historical violation records, player level, the player's social network in the game (such as a friend list), etc.

[0043] Optionally, the target model in this embodiment is used to output the predicted probability of abnormal behavior of the object. In the financial field, the target model can be a credit scoring card model, which is mainly modeled using logistic regression. The output of the model is generally defined as the quality of credit business assets, such as whether there is a default. The main source of model input is the basic information of the enterprise corresponding to the credit business and the actual controller of the enterprise before the borrowing time, basic information of the bank account, deposit data, credit data and other multi-dimensional data details.

[0044] Optionally, the preset indicators in this embodiment include a discrimination index (Information Value, referred to as IV) of the preset variable and a stability index (Population Stability Index, referred to as PSI) of the preset variable, which are used to evaluate the impact of the variable on the prediction ability and stability of the model. For example, in the above-mentioned credit scoring card model, a higher IV value means that the discrimination of the indicator is higher, while a lower IV value means that the discrimination of the indicator is lower; the PSI value is calculated to monitor the stability of the model at different time points or groups, ensuring that the model remains effective for a long time, and can be adjusted in time to maintain the prediction accuracy even when the market environment changes or the characteristics of the customer group fluctuate.

[0045] Through the above steps, the behavior data and object information of N objects are collected respectively, and then the above first information is input into the target model to obtain the behavior prediction probability of abnormal behavior of each object. The target model is iteratively trained based on the feature extraction sub-model and the classification sub-model, which can more accurately extract key features from the object information and improve the accuracy of the probability of predicting abnormal behavior. The use of preset indicators ensures that the model training process can select more effective variables, further improving the generalization ability and prediction effect of the model.

[0046] In an exemplary embodiment, the gradients of the M first parameters are calculated using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and the first objective function, including: performing partial derivative operations on the M first parameters respectively through the first objective function to obtain partial derivatives corresponding to the M first parameters; substituting the N first behavior probabilities and the N real behavior labels into the partial derivatives corresponding to the M first parameters to calculate the gradients corresponding to the M first parameters.

[0047] Optionally, the partial derivative operation in this embodiment is performed according to the first objective function actually used. For example, when the first objective function is expressed as When Among them, A is the first sub-parameter of the M parameters, B is the second sub-parameter of the M parameters, the first sub-parameter is used to represent the shape of the nonlinear relationship of the control preset function, and the first sub-parameter is greater than 0, and the second sub-parameter is used to adjust the offset degree of the preset function, Y i represents the true behavior label of the i-th sample, Y i =1 is used to indicate the real behavior label of the object that has the preset abnormal behavior, y i Represents the calibrated probability of the i-th sample, and takes partial derivatives of A and B in the first objective function to obtain and Get the partial derivatives corresponding to parameters A and B.

[0048] Optionally, the real behavior label in this embodiment is marked with a numerical value. Through the above steps, the partial derivative operation is performed on the M first parameters in the first objective function to obtain the corresponding partial derivatives, and then the N first behavior probabilities and the real behavior labels are substituted to obtain the gradients corresponding to the M first parameters, so as to facilitate the subsequent optimal M first parameters.

[0049] In an exemplary embodiment, based on the gradients of the M first parameters, the M first parameters are iteratively updated, and when the M first parameters meet the preset conditions, the target group parameter values ​​corresponding to the M first parameters are determined, including: when the gradients of the M first parameters do not meet the preset conditions, iterative calculation is performed on the gradients of the M first parameters, wherein, when the gradients of the M first parameters are subjected to the j-th round of iterative calculation, the steps performed include: obtaining M first target parameters, wherein the first target parameters are parameters obtained after the j-1-th round of updating; based on the update formulas of the M parameters and the M first target parameters, M second target parameters are calculated, wherein the M second target parameters are parameters obtained after the j-th round of iterative calculation; when the M second target parameters meet the preset conditions, the M second target parameters are determined as target group parameter values; wherein the j is a positive integer greater than 1.

[0050] Optionally, the preset conditions in this embodiment are designed according to actual conditions, including but not limited to: the first objective function converges, the maximum number of iterations is reached, etc.

[0051] Optionally, in this embodiment, the update formula of M parameters is expressed as: new parameter = old parameter - learning rate × gradient of this parameter. For example, when the first objective function is expressed as When The update formulas for parameters A and B are and Among them, A is the first sub-parameter of the M parameters, B is the second sub-parameter of the M parameters, the first sub-parameter is used to represent the shape of the nonlinear relationship of the control preset function, and the first sub-parameter is greater than 0, and the second sub-parameter is used to adjust the offset degree of the preset function, Y i represents the true behavior label of the i-th sample, Y i =1 is used to indicate the real behavior label of the object that has the preset abnormal behavior, y i represents the calibrated probability of the i-th sample, η is used to represent the learning rate and control the step size of each update, and is the gradient of the first objective function with respect to A and B.

[0052] Optionally, the determination of the preset conditions in this embodiment is iterative and optimized, and needs to comprehensively consider business needs, data characteristics, model types and performance indicators, and is finally determined through experimental verification and adjustment. In practical applications, the preset conditions may need to be fine-tuned according to different business scenarios and data sets to ensure the best performance of the model under specific conditions. For example, in the return prediction in the e-commerce field, if the business focuses more on reducing prediction errors, the preset loss value in the preset conditions may need to be set at a level that ensures a lower mean absolute error (MAE) or mean square error (MSE) to optimize the prediction accuracy; and in the detection of game plug-in behavior, if the business goal is to increase the detection rate of plug-in behavior, the preset loss value included in the preset conditions may need to be set higher to allow the model to slightly sacrifice some accuracy while ensuring a higher recall rate.

[0053] Through the above steps, by judging whether the gradients of the M first parameters meet the preset conditions, and iteratively calculating and updating the first parameters, a set of parameter values ​​that meet the preset conditions are finally determined as the target group parameter values. By determining the target group parameter values ​​corresponding to the preset conditions, it is possible to ensure that the calibrated prediction probability is as close to the real behavior label as possible, improve the calibration effect of the model, help avoid overfitting during the model calibration process, and ensure the prediction ability of the model on unknown objects.

[0054] In an exemplary embodiment, adjusting a preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second objective function includes: inputting the target group parameter values ​​into the M second parameters included in the preset function to obtain the second objective function, wherein the preset function is: The above A is the first subparameter of the M above second parameters, the above B is the second subparameter of the M above parameters, the above first subparameter is used to represent the shape of the nonlinear relationship controlling the above preset function, and the above first subparameter is greater than 0, the above second subparameter is used to adjust the offset degree of the above preset function, the above i is used to represent the i-th object of the N above objects, and the above Y i It is used to represent the probability corresponding to the true behavior label of the above i-th object. The above p i Used to represent the predicted probability of the behavior of the i-th object mentioned above.

[0055] Optionally, in the financial field, the preset function in this embodiment is obtained by the following steps: Since the scorecard logistic regression model itself is a Perform regression and the predicted value is When the default probability prediction p is biased from the actual default probability of the sample, the fitting formula To obtain A and B, according to the formula Available Resolve

[0056] Specifically, the preset function can be derived based on the BETA distribution: let the default probability prediction value s obey the BETA distribution, then the probability density function is: Let p(s|+)=f(s, α0, β0), p(s|+)=f(s, α1, β1), then Among them, α=α0-α1, β=β1-β0, Let K = e c ,but, Then according to the Bayesian update Among them, the prior probability is a constant, then Resolve Let α=β=A, c=B to get the preset function. Among them, s represents the predicted value of default probability, which is the model's prediction of the probability of a certain object (such as a customer) defaulting in the future (i.e., abnormal behavior), ranging from 0 to 1; α and β are two parameters of the BETA distribution, which control the shape of the distribution. α and β are used to construct the probability density function of the predicted value of default probability s. α controls the positive probability characteristics of s, while β controls the reverse probability characteristics of s, i.e., the probability that s is close to 0; B(α, β) is the BETA function, defined as Where Γ is the gamma function; * indicates multiplication; likelihood ratio is the likelihood ratio, which is the ratio of the probability of s under the positive class (+, such as a defaulting customer) condition to the probability of s under the negative class (-, such as a normal customer) condition, which reflects the ratio of the probability that the object is a positive class to the probability of the negative class when the predicted probability s is given; α0, β0 are used to represent the parameters of the BETA distribution under the positive class condition, which are used to describe the probability distribution of s under the positive class (such as default); α1, β1 are used to represent the parameters of the BETA distribution under the negative class condition, which are used to describe the probability distribution of s under the negative class (such as normal repayment); α, β are simplified parameters, defined as α=α0-α1 and β=β1-β0, which are used to describe the relative difference between the two distributions in the likelihood ratio; K is a constant term, which is used to balance the differences in the likelihood ratio due to different α and β parameters; c is a logarithmic constant term; K=e c , which is used to adjust the likelihood ratio in the logarithmic space to make the optimization process more stable; y is used to represent the calibrated default probability prediction value, and the goal is to convert the original default probability prediction value s into a value y that is closer to the actual default rate through the preset function; A and B are parameters in the final preset function, which are used to control the shape and position of the calibration function and are the key adjustment parameters in the preset function.

[0057] Through the above steps, a preset function is obtained based on the existing function, and then the determined target group parameter value is input into the preset function, and the function parameters are adjusted to obtain the second objective function, so that the calibration process of the behavior prediction probability can be carried out according to the selected optimal parameter group, so as to realize customized calibration for a specific data set, and by optimizing the parameters, the second objective function can more accurately reflect the relationship between the behavior prediction probability and the true behavior label, thereby improving the accuracy of the prediction.

[0058] In an exemplary embodiment, the N behavior prediction probabilities are calibrated respectively using the second objective function to obtain N target behavior probabilities, including: for each of the behavior prediction probabilities of the objects, calibration is performed through the following steps: the probability corresponding to the true behavior label is input into the second objective function to calculate the calibration probability; the calibration probability is compared with the behavior prediction probability; when the difference between the calibration probability and the behavior prediction probability is less than or equal to a preset threshold, the behavior prediction probability is determined as the target behavior probability; when the difference between the calibration probability and the behavior prediction probability is greater than the preset threshold, the calibration probability is determined as the target behavior probability.

[0059] Optionally, in this embodiment, the calibration probability is compared with the original behavior prediction probability, and the difference between the two is calculated. If the difference is less than or equal to a preset threshold, it indicates that the original prediction probability is accurate enough and is directly used as the target behavior probability; conversely, if the difference is greater than the preset threshold, it indicates that the original prediction needs to be calibrated. At this time, the calibration probability is used as the target behavior probability to improve the accuracy of the prediction.

[0060] Through the above steps, the behavior prediction probability of each object is calibrated using the second objective function, and the calibration probability is calculated. Then, by comparing the calibration probability with the behavior prediction probability, the target behavior probability is determined according to the preset threshold. The difference between the calibration probability and the behavior prediction probability can be accurately judged to ensure that the calibrated target behavior probability is closer to the real behavior label. The use of the preset threshold avoids unnecessary over-calibration and maintains the stability and reliability of the final result.

[0061] In an exemplary embodiment, after calibrating the N above-mentioned behavior prediction probabilities respectively using the above-mentioned second objective function to obtain N target behavior probabilities, the above-mentioned method also includes: calculating N target behavior probability ratios corresponding to the N above-mentioned target behavior probabilities, wherein the above-mentioned target behavior probability ratio is used to represent the ratio between the probability of the above-mentioned object occurring the above-mentioned abnormal behavior and the probability of not occurring the above-mentioned abnormal behavior; inputting the N above-mentioned target behavior probability ratios into the conversion function, and calculating the abnormal behavior scores of the N above-mentioned objects occurring the above-mentioned abnormal behaviors through the above-mentioned conversion function, wherein the above-mentioned conversion function is a function used to convert the above-mentioned target behavior probabilities into scores in a preset format.

[0062] Optionally, the conversion function in this embodiment is used to convert the probability into a score, so that the behavioral risks between different objects can be compared in the form of scores, and the score of the object's preset abnormal behavior = basic score + multiplier × the object's target behavior probability ratio, where the basic score is usually a reference score set according to business needs. For example, in credit scoring in the financial field, the basic score may be set to 680 points, representing a customer with average risk; the multiplier is a parameter that controls the range of score changes, which needs to be adjusted according to actual needs to ensure that the score can accurately reflect the difference in risk.

[0063] Optionally, in the financial field, the conversion function in this embodiment is The base score is set to 680 points when the good-bad ratio is 60:1. The determination of the multiplier may be based on the matching of the logarithmic probability and the actual score change. Here, it is set that the good-bad ratio doubles for every 30 points increase, that is, the multiplier is

[0064] Optionally, the performance of the conversion function in this embodiment is continuously monitored during the operation of the model. When the customer group, market environment or business needs change, the conversion function is adjusted or updated to ensure the accuracy and timeliness of the scoring.

[0065] Optionally, in this embodiment, by converting the target behavior probability into an abnormal behavior score, business decision makers can more intuitively understand the risk level of each object, thereby effectively performing risk management and resource allocation. For example, in credit approval, customers can be graded according to the abnormal behavior score to determine the loan amount and interest rate; on e-commerce platforms, abnormal behavior scores can be used to identify potential return risks and optimize inventory management; in game security systems, anti-cheating strategies can be dynamically adjusted based on abnormal behavior scores to improve player experience.

[0066] Through the above steps, based on N target behavior probabilities, the ratio of the probability of abnormal behavior to the probability of no abnormal behavior for each object is calculated, and then the target behavior probability ratio is converted into an abnormal behavior score through the conversion function. Converting the target behavior probability into an abnormal behavior score improves the interpretability and intuitiveness of the calibration results, making it easier for business personnel to understand and take action. The application of the conversion function ensures the consistency and standardization of the abnormal behavior score, so that the prediction results of different objects can be compared under the same standard, improving the efficiency and accuracy of the analysis.

[0067] The following is a specific example. For example, in the financial field, when predicting the probability of customer default, the calibration process for obtaining the probability of the object defaulting is as follows: Figure 3 As shown, the specific steps include:

[0068] Step S302: Obtain the sample default probability prediction value P output by the logistic regression model i :

[0069] The trained logistic regression model is used to analyze the multi-dimensional characteristics of each customer (sample), such as financial data, credit history, and transaction behavior. The logistic regression model outputs a value between 0 and 1, representing the probability of the customer defaulting in a certain period in the future, denoted as P. i , each customer corresponds to a predicted value of default probability, which constitutes the preliminary prediction result of the model, but may not be accurate enough due to the limitations of the model itself, uneven data distribution or sample bias, and needs further calibration.

[0070] Step S304: Fit the scoring calibration formula to obtain A and B:

[0071] The initial default probability prediction value P of all customers is obtained i After that, the scoring calibration formula is fitted to correct the output of the model to make it closer to the actual default rate. In this embodiment, the scoring calibration formula is based on the BETA distribution, and its formula form is: Among them, Y i is the predicted value of the default probability after calibration, A and B are the parameters to be solved, representing the linear combination of the logarithmic probability and the calibration formula. The specific method of fitting A and B may include minimizing the loss function between the predicted value and the actual value, such as the logarithmic loss function, and determining the optimal values ​​of A and B through an optimization algorithm (such as the gradient descent method or the quasi-Newton method). By adjusting A and B, the calibrated default probability Y i It is closer to the actual default probability while maintaining the ranking ability of the predicted values. That is, customers who originally have a high probability of default will still maintain a higher predicted value of default probability after calibration.

[0072] The steps for determining the optimal values ​​of A and B using the gradient descent method are as follows:

[0073] Step 1, initialize the initial values ​​of A and B;

[0074] Step 2: Calculate the calibrated probability y using the current A and B i ;

[0075] Step 3, calculating the gradients of A and B in the first objective function, wherein the first objective function is a function obtained by performing a logarithmic loss operation on the score calibration formula;

[0076] Step 4, update A and B;

[0077] Step 5: Repeat steps 2 to 4 until the first objective function converges or the maximum number of iterations is reached.

[0078] Step S306: Substitute P according to the calibration formula i Output calibrated default probability y i :

[0079] After obtaining A and B, the above scoring calibration formula is used to predict the initial default probability of each sample p i Calibrate and substitute into the formula to calculate the calibrated default probability Y i This process is equivalent to using the calibration formula to perform a nonlinear transformation on the model output to eliminate possible system deviations and ensure the reliability of the model prediction results. The default probability Y after calibration is i It will more accurately reflect the customer's actual default risk.

[0080] Step S308: y i Substitute the preset score conversion formula to find the score of each sample:

[0081] The calibrated default probability Y i Convert to a more intuitive scoring form that is easier for business decision makers to understand and apply. This conversion is usually based on a preset score conversion formula. For example, in the field of credit scoring, the following standardized formula can be used: The base score is 680, which represents the score value when the good-to-bad ratio is 60:1. By adjusting, the good-to-bad ratio doubles with every 30 points increase, and the calibrated default probability Y i Substituting into the formula, the credit score of each sample is calculated. This conversion method not only makes the default probability intuitively presented in the form of a score, but also retains the sorting ability of the model prediction, that is, customers with low credit scores are identified as high risk, and vice versa.

[0082] Through the above steps, the initial default probability prediction value output by logistic regression is calibrated and converted to a more accurate and intuitive credit score, which provides strong support for the credit approval, risk management and resource allocation of financial institutions. From obtaining the model prediction results, to fitting the score calibration formula, to calculating the calibrated default probability and converting the final score, each step ensures the accuracy, consistency and practicality of the model output.

[0083] It should be noted that, through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the above methods of each embodiment of the present application.

[0084] In this embodiment, a device for calibrating the prediction probability of behavior is also provided, and the device is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived.

[0085] Figure 4 is a structural block diagram of a device for calibrating behavior prediction probability according to an embodiment of the present application, such as Figure 4 As shown, the device comprises:

[0086] An acquisition module 4002 is used to acquire the behavior prediction probability of abnormal behavior of N objects, wherein the behavior prediction probability is a probability value output by the target model based on the behavior data of the objects, and N is a positive integer greater than 0;

[0087] A first calibration module 4004 is used to calibrate the N behavior prediction probabilities based on the N behavior prediction probabilities and a first initial function to obtain N first behavior probabilities, wherein the first initial function includes M preset first parameters, M is a positive integer greater than 0, and the first initial function is used to calibrate the behavior prediction probabilities;

[0088] A calculation module 4006, configured to calculate the gradients of the M first parameters using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and a first objective function, wherein the real behavior labels are used to represent the real behaviors of the objects, and the first objective function includes the preset M first parameters;

[0089] A determination module 4008 is used to iteratively update the M first parameters based on the gradients of the M first parameters, and determine the target group parameter values ​​corresponding to the M first parameters when the M first parameters meet the preset conditions; an adjustment module 4010 is used to adjust the preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second target function, wherein the preset function is a monotonic function;

[0090] The second calibration module 4012 is used to calibrate the N behavior prediction probabilities respectively using the second objective function to obtain N target behavior probabilities.

[0091] In an exemplary embodiment, the acquisition module 4002 further includes: a first acquisition unit, for respectively acquiring the first information of the N objects, wherein the first information includes the behavior data and object information of the objects; a first input unit, for inputting the first information of the N objects into the target model to obtain the behavior prediction probability of abnormal behavior of each object output by the target model; wherein the target model is a logistic regression model obtained by iteratively training the initial model using sample information of multiple groups of sample objects, the initial model includes a feature extraction sub-model and a classification sub-model, for each group of sample information, the feature extraction sub-model is used to convert the sample information into preset variables, and extract target variables that meet preset indicators from the preset variables, the classification sub-model is used to classify the behavior of the sample objects based on the target variables, the sample information includes sample behavior data, sample object information and real behavior data of the sample objects, and the preset indicators include weighted evidence indicators of the preset variables and stability indicators of the preset variables.

[0092] In an exemplary embodiment, the above-mentioned calculation module 4006 includes: a first execution unit, used to perform partial derivative operations on the M above-mentioned first parameters through a first objective function, and obtain partial derivatives corresponding to the M above-mentioned first parameters; a first calculation unit, used to substitute the N above-mentioned first behavior probabilities and N true behavior labels into the partial derivatives corresponding to the M above-mentioned first parameters, and calculate the gradients corresponding to the M above-mentioned first parameters.

[0093] In an exemplary embodiment, the determination module 4008 includes: a second calculation unit, for performing iterative calculation on the gradients of the M first parameters when the gradients of the M first parameters do not meet the preset conditions, wherein when the gradients of the M first parameters are subjected to the j-th round of iterative calculation, the steps performed include: obtaining M first target parameters, wherein the first target parameters are parameters obtained after the j-1-th round of updating; based on the update formulas of the M parameters and the M first target parameters, calculating M second target parameters, wherein the M second target parameters are parameters obtained after the j-th round of iterative calculation; when the M second target parameters meet the preset conditions, determining the M second target parameters as target group parameter values; wherein j is a positive integer greater than 1.

[0094] In an exemplary embodiment, the adjustment module 4010 includes: a second input unit, configured to input the target group parameter value into the M second parameters included in the preset function to obtain a second target function, wherein the preset function is: The above A is the first subparameter of the M above second parameters, the above B is the second subparameter of the M above second parameters, the above first subparameter is used to represent the shape of the nonlinear relationship controlling the above preset function, and the above first subparameter is greater than 0, the above second subparameter is used to adjust the offset degree of the above preset function, the above i is used to represent the i-th object of the N above objects, and the above Y i It is used to represent the probability corresponding to the true behavior label of the above i-th object. The above p i Used to represent the predicted probability of the behavior of the i-th object mentioned above.

[0095] In an exemplary embodiment, the second calibration module 4012 includes: a third calculation unit, used to input the probability corresponding to the true behavior label into the second objective function to calculate the calibration probability; a first comparison unit, used to compare the calibration probability with the behavior prediction probability; a first determination unit, used to determine the behavior prediction probability as the target behavior probability when the difference between the calibration probability and the behavior prediction probability is less than or equal to a preset threshold; a second determination unit, used to determine the calibration probability as the target behavior probability when the difference between the calibration probability and the behavior prediction probability is greater than the preset threshold.

[0096] In an exemplary embodiment, the second calibration module 4012 further includes: a fourth calculation unit, used to calculate N target behavior probability ratios corresponding to the N target behavior probabilities, wherein the target behavior probability ratios are used to represent the ratio between the probability that the object undergoes the abnormal behavior and the probability that the abnormal behavior does not occur; a fifth calculation unit, used to input the N target behavior probability ratios into a conversion function, and calculate the abnormal behavior scores of the N objects undergoing the abnormal behavior through the conversion function, wherein the conversion function is a function for converting the target behavior probabilities into scores in a preset format.

[0097] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0098] An embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0099] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0100] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0101] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0102] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0103] Obviously, those skilled in the art should understand that the above modules or steps of the present application can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present application is not limited to any specific combination of hardware and software.

[0104] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for calibrating behavior prediction probability, characterized in that: include: Obtaining a behavior prediction probability of abnormal behavior of N objects, wherein the behavior prediction probability is a probability value output by a target model based on the behavior data of the object, and N is a positive integer greater than 0; Based on the N behavior prediction probabilities and the first initial function, calibrating the N behavior prediction probabilities to obtain N first behavior probabilities, wherein the first initial function includes preset M first parameters, M is a positive integer greater than 0, and the first initial function is used to calibrate the behavior prediction probabilities; Calculating the gradients of the M first parameters using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and a first objective function, wherein the real behavior labels are used to represent the real behaviors of the objects, and the first objective function includes the preset M first parameters; Iteratively update the M first parameters based on the gradients of the M first parameters, and determine the target group parameter values ​​corresponding to the M first parameters when the M first parameters meet a preset condition; Adjusting a preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second target function, wherein the preset function is a monotonic function; The N behavior prediction probabilities are calibrated respectively using the second objective function to obtain N target behavior probabilities.

2. The method according to claim 1, characterized in that Before obtaining the behavior prediction probability of abnormal behavior of N objects, the method further includes: Acquire first information of N objects respectively, wherein the first information includes behavior data and object information of the objects; Inputting the first information of the N objects into the target model, and obtaining the behavior prediction probability of abnormal behavior of each object outputted by the target model; Among them, the target model is a logistic regression model obtained by iteratively training the initial model using sample information of multiple groups of sample objects. The initial model includes a feature extraction sub-model and a classification sub-model. For each group of the sample information, the feature extraction sub-model is used to convert the sample information into preset variables, and extract the target variables that meet the preset indicators from the preset variables. The classification sub-model is used to classify the behavior of the sample objects based on the target variables. The sample information includes sample behavior data of the sample objects, sample object information and real behavior data of the sample objects. The preset indicators include weighted evidence indicators of the preset variables and stability indicators of the preset variables.

3. The method according to claim 1, characterized in that Calculating the gradients of the M first parameters by using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and the first objective function, includes: By using a first objective function, performing partial derivative operations on the M first parameters respectively, to obtain partial derivatives corresponding to the M first parameters; Substitute the N first behavior probabilities and the N true behavior labels into the partial derivatives corresponding to the M first parameters, and calculate the gradients corresponding to the M first parameters.

4. The method according to claim 1, characterized in that: Iteratively updating the M first parameters based on the gradients of the M first parameters, and determining target group parameter values ​​corresponding to the M first parameters when the M first parameters meet a preset condition, includes: When the gradients of the M first parameters do not satisfy the preset condition, iterative calculation is performed on the gradients of the M first parameters, wherein when the j-th round of iterative calculation is performed on the gradients of the M first parameters, the steps performed include: Obtain M first target parameters, where the first target parameters are parameters obtained after the j-1th round of updating; Based on the update formulas of the M parameters and the M first target parameters, M second target parameters are calculated, wherein the M second target parameters are parameters obtained after the j-th round of iterative calculation; In a case where the M second target parameters meet the preset condition, determining the M second target parameters as target group parameter values; Wherein, j is a positive integer greater than 1.

5. The method according to claim 1, characterized in that Adjusting a preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second target function includes: The target group parameter value is input into the M second parameters included in the preset function to obtain a second target function, wherein the preset function is: A is the first sub-parameter among the M second parameters, B is the second sub-parameter among the M second parameters, the first sub-parameter is used to represent the shape of the nonlinear relationship that controls the preset function, and the first sub-parameter is greater than 0, the second sub-parameter is used to adjust the offset degree of the preset function, i is used to represent the i-th object among the N objects, Yi is used to represent the probability corresponding to the true behavior label of the i-th object, and pi is used to represent the behavior prediction probability of the i-th object.

6. The method according to claim 1, characterized in that The N behavior prediction probabilities are calibrated respectively by using the second objective function to obtain N target behavior probabilities, including: The predicted probability of the behavior of each object is calibrated by the following steps: Inputting the probability corresponding to the true behavior label into the second objective function to calculate a calibration probability; comparing the calibrated probability with the behavior prediction probability; When the difference between the calibration probability and the behavior prediction probability is less than or equal to a preset threshold, determining the behavior prediction probability as the target behavior probability; When the difference between the calibration probability and the behavior prediction probability is greater than the preset threshold, the calibration probability is determined as the target behavior probability.

7. The method according to claim 1, characterized in that After the N behavior prediction probabilities are calibrated respectively by using the second objective function to obtain N target behavior probabilities, the method further includes: Calculating N target behavior probability ratios corresponding to the N target behavior probabilities, wherein the target behavior probability ratio is used to represent the ratio between the probability that the object has the abnormal behavior and the probability that the object has not the abnormal behavior; The N target behavior probability ratios are input into a conversion function, and the abnormal behavior scores of the N objects that have the abnormal behaviors are calculated by the conversion function, wherein the conversion function is a function for converting the target behavior probability into a score in a preset format.

8. A device for calibrating behavior prediction probability, characterized in that: include: An acquisition module, used to acquire the behavior prediction probability of abnormal behavior of N objects, wherein the behavior prediction probability is a probability value output by a target model based on the behavior data of the object, and N is a positive integer greater than 0; A first calibration module, configured to calibrate the N behavior prediction probabilities based on the N behavior prediction probabilities and a first initial function to obtain N first behavior probabilities, wherein the first initial function includes preset M first parameters, M is a positive integer greater than 0, and the first initial function is used to calibrate the behavior prediction probabilities; a calculation module, configured to calculate the gradients of the M first parameters using the N first behavior probabilities, the N real behavior labels corresponding to the N objects, and a first objective function, wherein the real behavior labels are used to represent the real behaviors of the objects, and the first objective function includes the preset M first parameters; a determination module, configured to iteratively update the M first parameters based on the gradients of the M first parameters, and determine target group parameter values ​​corresponding to the M first parameters when the M first parameters meet a preset condition; An adjustment module, configured to adjust a preset function based on the target group parameter values ​​corresponding to the M first parameters to obtain a second target function, wherein the preset function is a monotonic function; The second calibration module is used to calibrate the N behavior prediction probabilities respectively by using the second objective function to obtain N target behavior probabilities.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program implements the steps of the method described in any one of claims 1 to 7 when executed by a processor.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method described in any one of claims 1 to 7 are implemented.

Images