Training method of behavior prediction model, risk behavior prediction method and device

Abstract

The embodiments of the present specification describe a training method of a behavior prediction model, a risk behavior prediction method and device. According to the method of the embodiments, the type identification of a sample behavior event and the time information of the sample behavior event can be obtained when training the behavior prediction model. Then the sample behavior event is characterized in the continuous time domain, and then the behavior prediction model is trained according to the behavior event characterized in the continuous time domain, so as to optimize the prediction value of the type identification and the time output by the behavior prediction model. By characterizing the sample behavior event in the continuous time domain, the association between the behavior event and the time when it occurs is realized, so that the model can fully learn the regularity and periodicity characteristics presented by the behavior event and the time when it occurs, thereby improving the accuracy of risk behavior prediction.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of artificial intelligence, and more particularly to methods for training behavior prediction models, methods for predicting risky behavior, and apparatus. Background Technology

[0002] Data-driven risk control primarily assesses the risk of current events based on historical data. It compares information about the current event with information about historical events recorded in the risk control engine, determining the level of risk based on the degree of anomaly. This is currently the most widely used risk control method in industry.

[0003] However, user behavior events typically exhibit patterns or periodic variations over a continuous time domain, leading to lower accuracy in behavior prediction methods based on data at discrete time points. Summary of the Invention

[0004] This specification describes one or more embodiments of a training method for a behavior prediction model, a method and apparatus for predicting risky behavior, which can improve the accuracy of risky behavior prediction.

[0005] Based on the first aspect, training methods for behavior prediction models are provided, including:

[0006] Extract the type identifier of the sample behavior event and the corresponding time information from historical behavior data; each type of sample behavior event corresponds to a type identifier;

[0007] Based on the type identifier and time information of the sample behavioral events, the sample behavioral events are characterized in the continuous time domain to obtain behavioral event representations.

[0008] The behavior prediction model is trained based on the behavior event representation to optimize the type identifier and time prediction values ​​output by the behavior prediction model.

[0009] In one possible implementation, the step of representing the sample behavioral event in a continuous time domain based on the type identifier and time information of the sample behavioral event to obtain a behavioral event representation includes:

[0010] For each sample behavior event, execute:

[0011] Map the type identifier of the current sample behavior event to a preset space to obtain the mapping vector of the current sample behavior event;

[0012] Obtain the behavior event representation of the previous sample behavior event; wherein, when the current sample behavior event is the first sample behavior event, the behavior event representation of the previous sample behavior event is 0;

[0013] The behavioral event representation of the current sample behavior event is determined based on the mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavioral event representation of the previous sample behavior event.

[0014] In one possible implementation, determining the behavior event representation of the current sample behavior event based on the mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavior event representation of the previous sample behavior event includes:

[0015] The behavioral event representation of the j-th sample behavioral event is determined using the following formula:

[0016] h j =max{W y *y j +W t *t j +W h *h j-1 +b h ,0}

[0017] Among them, h j The behavioral event representation used to characterize the behavioral event of the j-th sample, y j The mapping vector t is used to characterize the behavioral event of the j-th sample. j h is used to characterize the time information corresponding to the behavior event of the j-th sample. j-1 W is used to characterize the behavioral event representation of the (j-1)th sample behavioral event. y The coefficients W used to characterize the linear transformation of behavioral events t The coefficients W are used to characterize the linear transformation of time information. h The coefficients, b, are used to characterize the linear transformation of the representation of the previous sample behavioral event. h The modifier used to characterize the behavioral event representation, max{…,0} is used to characterize the nonlinear activation function ReLU.

[0018] In one possible implementation, the behavior prediction model includes a first prediction probability distribution for predicting the type of behavioral event;

[0019] Training the behavior prediction model based on the behavior event representation includes:

[0020] Based on the behavioral event representation, a first prediction probability distribution is determined for predicting the type of behavioral event;

[0021] The model parameters in the first predicted probability distribution are optimized using the cross-entropy loss function.

[0022] In one possible implementation, determining a first prediction probability distribution for predicting the type of behavioral event based on the behavioral event representation includes:

[0023] The first prediction probability distribution for predicting the type of behavioral event is calculated using the following formula:

[0024]

[0025] Where P(k) is used to characterize the first prediction probability distribution, V k The parameter matrix h, whose type identifier is k, is used to characterize the neural network layer corresponding to the behavior prediction model. j The behavior event representation is used to characterize the behavior event of the j-th sample, K is used to characterize the number of types of sample behavior events, and b1 and b2 are both adjustment amounts used to adjust the predicted value.

[0026] In one possible implementation, the behavior prediction model includes a second prediction probability distribution that predicts the timing of the behavioral event.

[0027] Training the behavior prediction model based on the behavior event representation includes:

[0028] Based on the behavioral event characterization, a second prediction probability distribution is determined to predict the time of occurrence of the behavioral event.

[0029] The model parameters in the second predicted probability distribution are optimized using a logarithmic function.

[0030] In one possible implementation, the historical behavior data is t. i Data from a point in time prior;

[0031] The step of determining a second prediction probability distribution for predicting the time of occurrence of a behavioral event based on the behavioral event representation includes:

[0032] Based on the behavioral event representation, determine t i The first probability density function for a behavioral event occurring after time step;

[0033] Based on the first probability density function, determine the self-t i The second probability density function is defined as the function that indicates no behavioral events occur within a preset time interval starting from a given moment.

[0034] The second predicted probability distribution is determined based on the first probability density function and the second probability density function.

[0035] In one possible implementation, the step of determining t based on the behavioral event representation...i The first probability density function for a behavioral event occurring after a certain time includes:

[0036] The first probability density function is determined using the following formula:

[0037]

[0038] Where, λ * (t) is used to characterize the first probability density function, v t h is the model parameter for linearly transforming the representation of behavioral events. j The behavioral event representation used to characterize the behavioral event of the j-th sample, t j b is used to characterize the time when the j-th sample behavioral event occurs. t The correction is used to characterize the first probability density function. w1 and w2 are model parameters that transform the time quantity. p is used to characterize the hyperparameters of the behavior prediction model.

[0039] And / or,

[0040] The step of determining from t based on the first probability density function i The second probability density function, which specifies that no behavioral events occur within a preset time interval starting from a given moment, includes:

[0041] The second probability density function is calculated using the following formula:

[0042]

[0043] Among them, S * (t) is used to characterize the second probability density function, λ * (t) is used to characterize the first probability density function;

[0044] And / or,

[0045] Determining the second predicted probability distribution based on the first probability density function and the second probability density function includes:

[0046] The second predicted probability distribution is calculated using the following formula:

[0047] f * (t)=λ * (t)·S * (t)

[0048] Among them, f * (t) is used to characterize the second prediction probability distribution, λ * (t) is used to characterize the first probability density function, S * (t) is used to characterize the second probability density function.

[0049] In one possible implementation, the sample behavioral event includes: user-generated transaction behavior; the transaction behavior includes at least one of shopping, transportation, dining, and entertainment.

[0050] According to the second aspect, a risk behavior prediction method is provided, including:

[0051] Get the Tth p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event;

[0052] The obtained type identifiers of at least one behavioral event and the corresponding time values ​​of each behavioral event are input into the behavior prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training method of the behavioral prediction model as described in any of the first aspects;

[0053] The historical behavior database is searched to determine if there is a data set that is consistent with both the predicted value of the type identifier and the predicted value of the time; wherein the historical behavior database stores at least one data set, and each data set includes a type identifier of a behavior event and a time value corresponding to the behavior event;

[0054] If not found, the predicted behavioral event is a risky behavioral event.

[0055] According to the third aspect, a training device for a behavior prediction model is provided, comprising: a sample data acquisition module, a representation module, and a training module;

[0056] The first acquisition module is configured to acquire the type identifier of the sample behavior event and the time information corresponding to the sample behavior event from historical behavior data; wherein, each type of sample behavior event corresponds to a type identifier;

[0057] The characterization module is configured to characterize the sample behavior event in a continuous time domain based on the type identifier and time information of the sample behavior event obtained by the first acquisition module, thereby obtaining a behavior event characterization.

[0058] The training module is configured to train the behavior prediction model based on the behavior event representation obtained by the representation module, so as to optimize the type identifier and time prediction value output by the behavior prediction model.

[0059] According to the fourth aspect, a risk behavior prediction device is provided, comprising: a second acquisition module, an input module, a search module, and a determination module;

[0060] The second acquisition module is configured to acquire the Tth... p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event;

[0061] The input module is configured to input the type identifier of the at least one behavioral event and the time value corresponding to each behavioral event obtained by the second acquisition module into the behavioral prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training device of the behavioral prediction model as described in the third aspect;

[0062] The search module is configured to search the historical behavior database for a data set that is consistent with the type identifier prediction value obtained by the input module and the time prediction value; wherein, the historical behavior database stores at least one data set, and each data set includes a type identifier of a behavior event and a time value corresponding to the behavior event;

[0063] The determining module is configured such that if the searching module does not find the event, the event to be predicted is a risky event.

[0064] According to a fifth aspect, a computing device is provided, comprising: a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of the first and second aspects described above.

[0065] According to the method and apparatus provided in the embodiments of this specification, when training a behavior prediction model, the type identifier of a sample behavior event and the corresponding time information are first obtained. Then, the sample behavior event can be represented in a continuous time domain based on the type identifier and time information. Finally, the behavior prediction model can be trained based on the obtained behavior event representation to optimize the predicted values ​​of the type identifier and time output by the behavior prediction model. Therefore, this application represents sample behavior events in a continuous time domain when training the model, realizing the correlation between behavior events and their occurrence time. Thus, the behavior prediction model trained based on the obtained behavior event representation can fully learn the regularity and periodicity of the behavior events and their occurrence time, thereby improving the accuracy of risk behavior prediction. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart illustrating a training method for a behavior prediction model provided in one embodiment of this specification;

[0068] Figure 2 This is a sequence diagram of a transaction behavior provided in one embodiment of this specification;

[0069] Figure 3 This is a flowchart illustrating a method for determining the representation of a behavioral event according to one embodiment of this specification;

[0070] Figure 4 This is a flowchart illustrating a method for training a behavior prediction model according to one embodiment of this specification;

[0071] Figure 5 This is a flowchart illustrating another method for training a behavior prediction model, provided in one embodiment of this specification.

[0072] Figure 6 This is a flowchart illustrating a method for determining a second predicted probability distribution according to one embodiment of this specification;

[0073] Figure 7 This is a flowchart of a risk behavior prediction method provided in one embodiment of this specification;

[0074] Figure 8 This is a schematic diagram of a training device for a behavior prediction model provided in one embodiment of this specification;

[0075] Figure 9 This is a schematic diagram of a risk behavior prediction device provided in one embodiment of this specification. Detailed Implementation

[0076] As mentioned earlier, the core idea of ​​prevention and control based on user historical behavior events is to compare the information of the current behavior event with the information of historical behavior events recorded in the risk control engine, and then determine the level of risk based on the degree of abnormality of the event.

[0077] Traditional risk control solutions typically pre-determine time intervals, converting time series into multiple equal-length time intervals. Events within each interval are then aggregated, and the resulting features are used for risk prediction. However, user behavior events often exhibit patterns and periodic variations over a continuous time domain. For example, users might buy breakfast at convenience stores or cafeterias between 7:00 and 9:00 AM, travel between 5:30 and 6:30 PM, or exercise or work out between 7:00 and 8:00 PM. Traditional solutions fail to adequately preserve time information and dynamic change models, losing the correlation between time and event information, thus leading to lower prediction accuracy.

[0078] In this approach, behavioral events are associated with the time in which they occur, and sample behavioral events are represented in a continuous time domain. Based on this, the behavioral prediction model trained can fully consider the regularity and periodicity of behavioral events and the time in which they occur, thereby improving the accuracy of risk behavior prediction.

[0079] like Figure 1 As shown in the embodiments of this specification, a method for training a behavior prediction model is provided, which may include the following steps:

[0080] Step 101: Obtain the type identifier of the sample behavior event and the time information corresponding to the sample behavior event from the historical behavior data; wherein, each type of sample behavior event corresponds to a type identifier;

[0081] Step 103: Based on the type identifier and time information of the sample behavioral events, represent the sample behavioral events in the continuous time domain to obtain the behavioral event representation;

[0082] Step 105: Train the behavior prediction model based on the behavior event representation to optimize the type identifier and time prediction values ​​output by the behavior prediction model.

[0083] When training a behavior prediction model, the first step is to obtain the type identifier of the sample behavior event and the corresponding time information. Then, based on the type identifier and time information, the sample behavior event can be represented in a continuous time domain. Finally, the behavior prediction model can be trained based on the obtained behavior event representation to optimize the predicted values ​​of the type identifier and time output by the behavior prediction model. Therefore, this application represents sample behavior events in a continuous time domain during model training, realizing the correlation between behavior events and their occurrence time. Thus, the behavior prediction model trained based on the obtained behavior event representation can fully learn the regularity and periodicity of the behavior events and their occurrence time, thereby improving the accuracy of risk behavior prediction.

[0084] The following describes the appendix in conjunction with specific embodiments. Figure 1 The steps in the process will be explained.

[0085] First, in step 101, the type identifier of the sample behavior event and the time information corresponding to the sample behavior event are obtained from the historical behavior data.

[0086] Historical behavioral data can be data that records a user's daily behavior. For example, sample behavioral events in this historical behavioral data can be transactional behaviors generated by the user, which can include shopping, transportation, dining, and entertainment. Figure 2 This is a sequence diagram of the user's transaction behavior at various times, which may include transportation, breakfast, lunch, afternoon tea, and shopping on Taobao.

[0087] Each type of sample behavior event can correspond to a type identifier. For example, a user's morning commute transaction corresponds to a type identifier A, a user's breakfast purchase transaction corresponds to a type identifier B, and a user's purchase on a shopping app transaction corresponds to a type identifier C, etc.

[0088] The type identifier corresponding to each type of sample behavior event can be used to encode each sample behavior event according to its type, so as to obtain the type identifier of each sample behavior event. For example, the code for "transportation" is "0001", "breakfast" is "0010", "lunch" is "0011", "shopping" is "0100", "online shopping" is "0101", and "exercise" is "0110", etc.

[0089] Then, in step 103, the sample behavioral events are characterized in the continuous time domain according to the type identifier and time information of the sample behavioral events to obtain the behavioral event representation.

[0090] After obtaining the type identifier and corresponding time of each sample behavioral event, consider representing each sample behavioral event in a continuous time domain. For example, ... Figure 3 As shown, step 103 can perform the following operations for each sample behavior event:

[0091] Step 301: Map the type identifier of the current sample behavior event to a preset space to obtain the mapping vector of the current sample behavior event;

[0092] Step 303: Obtain the behavior event representation of the previous sample behavior event; wherein, when the current sample behavior event is the first sample behavior event, the behavior event representation of the previous sample behavior event is 0;

[0093] Step 305: Determine the behavioral event representation of the current sample behavioral event based on the mapping vector of the current sample behavioral event, the time information corresponding to the current sample behavioral event, and the behavioral event representation of the previous sample behavioral event.

[0094] In this embodiment, when representing each sample behavior event in a continuous time domain, the type identifier of the current sample behavior event can first be mapped to a preset space. Then, the behavior event representation of the previous sample behavior event is obtained. Based on the obtained mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavior event representation of the previous sample behavior event, the behavior event representation of the current sample behavior event can be determined. This association between sample behavior events and time information, and the incorporation of the influence of the previous sample behavior event on the current sample behavior event, fully considers the representation of each sample behavior event in the time domain, thereby improving the accuracy of predictions when predicting possible future user behavior events.

[0095] Step 301 will be explained below.

[0096] In this step, after encoding the type identifiers for each sample behavior event, we consider mapping them to embedding vectors. This allows us to adjust parameters such as the mapping dimension coefficient based on the size of the vectors to be processed, thus mapping each type identifier to a preset spatial dimension. For example, in one possible implementation, the type identifiers can be mapped using the following formula:

[0097] y j =W em *z j +b em

[0098] Among them, y j The mapping vector W used to characterize the behavioral event of the j-th sample. em The mapping coefficient, z, can be determined based on the size of the space to be mapped. j b is the type identifier for the j-th sample behavior event. em This is the correction amount for the mapping.

[0099] Step 303 will be explained below.

[0100] The behavior event representation of the j-th sample behavior event within the recurrent layer needs to consider the behavior event representation of the (j-1)-th sample behavior event, therefore, it is necessary to obtain the behavior event representation of the previous sample behavior event. Of course, if the current sample behavior event is the first sample behavior event input into the recurrent layer for training the behavior prediction model, the behavior event representation of the previous sample behavior event is 0. That is, in this case, when determining the current sample behavior event, it is determined based on the mapping vector of the current sample behavior event and the corresponding time information.

[0101] Step 305 will be explained below.

[0102] After obtaining the mapping vector of the current sample behavior event, the time information of the current sample behavior event, and the behavior event representation of the previous sample behavior event, consider simultaneously using linear transformation and nonlinear activation functions to determine the behavior event representation of the current sample behavior event. For example, in one possible implementation, step 305 can use the following calculation formula to determine the behavior event representation of the j-th sample behavior event:

[0103] h j =max{W y *y j +W t *t j +W h *h j-1 +b h ,0}

[0104] Among them, h j The behavioral event representation used to characterize the behavioral event of the j-th sample, y j The mapping vector t is used to characterize the behavioral event of the j-th sample. j h is used to characterize the time information corresponding to the behavior event of the j-th sample. j-1 W is used to characterize the behavioral event representation of the (j-1)th sample behavioral event. y The coefficients W used to characterize the linear transformation of behavioral events t The coefficients W are used to characterize the linear transformation of time information. h The coefficients, b, are used to characterize the linear transformation of the representation of the previous sample behavioral event. h The modifier used to characterize the behavioral event representation, max{…,0} is used to characterize the nonlinear activation function ReLU.

[0105] From the above calculation formula, we can see that W y W t and W hIt is used for linear transformation and fully integrates type identification, event information, and the representation of the previous sample behavior event. Furthermore, max{…,0} represents the non-linear activation function ReLU, meaning that both linear and non-linear transformations are used to represent the current behavior event, thus ensuring that the resulting behavior event representation more comprehensively covers the feature information of the current sample behavior event. It should be noted that the ellipsis in max{…,0} refers to the term preceding the comma within the curly braces. For example, in the above formula for calculating the behavior event representation, the ellipsis represents W. y *y j +W t *t j +W h *h j-1 +b h However, if the current sample behavior event is the first sample behavior event in the recurrent layer of the input neural network, the ellipsis is W. y *y j +W t *t j +b h .

[0106] Further in step 105, a behavior prediction model is trained based on the behavior event representation to optimize the type identifier and time prediction values ​​output by the behavior prediction model.

[0107] In the embodiments of this specification, the behavior prediction model can output a prediction of type identifier and a prediction of time. For type identifier, the behavior prediction model gives a prediction based on a first prediction probability distribution that predicts the type of behavior event; while for time, the behavior prediction model gives a prediction based on a second prediction probability distribution that predicts the time when the behavior event occurs. These two cases will be explained below.

[0108] Scenario 1: The behavior prediction model includes a first prediction probability distribution that predicts the type of behavioral event;

[0109] like Figure 4 As shown, in this case, step 105 during the training of the behavior prediction model can be achieved through the following steps:

[0110] Step 401: Based on the behavioral event representation, determine the first prediction probability distribution for predicting the type of behavioral event;

[0111] Step 403: Optimize the model parameters in the first predicted probability distribution using the cross-entropy loss function.

[0112] In this embodiment, when training the behavior prediction model based on the behavior event representation, a first prediction probability distribution for predicting the type of behavior event is first determined based on the behavior event representation. Then, the model parameters in the first prediction probability distribution are optimized using the cross-entropy loss function. In this way, by continuously optimizing the first prediction probability distribution for predicting the type of behavior event, better model parameters are obtained, making the behavior prediction model more reliable in predicting the type of behavior event.

[0113] In step 401, when determining the first predicted probability distribution, the following formula can be used:

[0114]

[0115] Where P(k) is used to characterize the first prediction probability distribution, V k The parameter matrix h, whose type is identified as k, is used to characterize the neural network layer of the behavior prediction model. j The behavior event representation is used to characterize the behavior event of the j-th sample, K is used to characterize the number of types of sample behavior events, and b1 and b2 are both adjustment amounts used to adjust the predicted value.

[0116] Step 403 optimizes the model parameters in the first predicted probability distribution using the cross-entropy loss function with the goal of minimizing the cross-entropy loss function. Specifically, based on the cross-entropy loss function, in each iteration, the value of the loss function is used for backpropagation to update the model parameters in the first predicted probability distribution until the iteration stopping condition is met. The iteration stopping condition can be, for example, the loss function converging or the number of iterations reaching a preset threshold.

[0117] Scenario 2: The behavior prediction model includes a second prediction probability distribution that predicts the timing of the behavior event.

[0118] like Figure 5 As shown, in this case, step 105 during the training of the behavior prediction model can be achieved through the following steps:

[0119] Step 501: Based on the behavioral event representation, determine the second prediction probability distribution for predicting the time of occurrence of the behavioral event;

[0120] Step 503: Optimize the model parameters in the second predicted probability distribution using the logarithmic function.

[0121] In this embodiment, when training the behavior prediction model based on the behavior event representation, a second prediction probability distribution for predicting the time of occurrence of the behavior event is first determined based on the behavior event representation. Then, the model parameters in the second prediction probability distribution are optimized using a logarithmic function. Thus, by continuously optimizing the second prediction probability distribution for predicting the time of occurrence of the behavior event, better model parameters are obtained, thereby making the behavior prediction model more reliable in predicting the time of occurrence of the behavior event.

[0122] In this embodiment, since each behavioral event is characterized in a continuous time domain, a modeling approach based on time-series processes is considered to determine the second predicted probability distribution. The second predicted probability distribution describes the probability density of no behavioral events occurring within a certain time period, and the probability density function of behavioral events occurring after a certain time period, based on the time-series process. Therefore, the second predicted probability distribution can be further determined by determining the probability density function of no behavioral events occurring within a certain time period, and the probability density function of behavioral events occurring after a certain time period.

[0123] For example, such as Figure 6 As shown, when the historical behavior data is t i When dealing with data prior to time step, step 501 can determine the second prediction probability distribution through the following steps:

[0124] Step 601: Determine t based on the behavioral event representation. i The first probability density function for a behavioral event occurring after time step;

[0125] Step 603: Determine the value of t based on the first probability density function. i The second probability density function is defined as the function that indicates no behavioral events occur within a preset time interval starting from a given moment.

[0126] Step 605: Determine the second predicted probability distribution based on the first probability density function and the second probability density function.

[0127] In this embodiment, when determining the second prediction probability distribution for predicting the time of occurrence of a behavioral event, t can first be determined based on the behavioral event representation. i The first probability density function is used to determine the probability of an event occurring after time t. Then, the probability density function is used to determine the probability of the event occurring after time t. i The second probability density function is defined as the probability density function of the first probability density function and the second probability density function, where no behavioral events occur within a preset time period starting from time t. Thus, the joint probability density of t represents the probability density of t. i The probability density function of no behavioral events occurring within time t, and the probability of a behavioral event occurring at time t in the future, is the second prediction probability distribution.

[0128] Step 601 will be explained below.

[0129] The core of a time series point process is its conditional intensity function λ. * (t), which is the first probability density function. Given an infinitesimal time window [t, t+dt], based on historical events before time t, we have: H t ={t i ,z i |t i ≤t}, where z i Represents time t i The type of behavioral event that has occurred is identified. Therefore, the probability of a future behavioral event can be formally defined as: λ * (t)dt=P(eventtin[t,t+dt]|H t )=E(dN(t)|H t ), where E(dN(t)|H t ) indicates based on historical event H t The expected number of events occurring within a time window [t, t+dt], where N(t) represents the number of events occurring before time t. If we assume that two events will not occur at the same time, i.e., dN(t) ∈ {0, 1}, then omitting the given condition, we can obtain λ. * (t), the main difference between different point process models is λ. * (t) different forms.

[0130] In this application, the first probability density function is determined using the following formula:

[0131]

[0132] Where, λ * (t) is used to characterize the first probability density function, v t h is the model parameter for linearly transforming the representation of behavioral events. j The behavioral event representation used to characterize the behavioral event of the j-th sample, t j b is used to characterize the time when the j-th sample behavioral event occurs. t The correction is used to characterize the first probability density function. w1 and w2 are model parameters that transform the time quantity. p is used to characterize the hyperparameters of the behavior prediction model.

[0133] The conditional strength function defined in the traditional way is λ * (t)=exp(v t ·h j +w t (tt j )+b t However, in the further calculation of the second predicted probability distribution, the model parameters w that are transformed by the time amount...t This can occur in division operations, which can easily cause numerical calculation problems, such as when w... t When the value is near 0, it is very easy to cause a division-by-zero exception. This application considers utilizing... and Replace w t By introducing an exponential function, the subsequent division calculations will not cause numerical calculation problems.

[0134] Furthermore, since the range of values ​​for w1 and w2 is (-∞, +∞), therefore Greater than 0 Less than 0. And in exp(v t ·h j +w t (tt j )+b t In ), w t The range of values ​​for is (-∞, +∞). To compensate for... and The insufficiency of two component degrees of freedom will reduce the two functions and A linear combination of these is used as the first probability density function.

[0135] It should be noted that p is a hyperparameter of the behavior prediction model, which can be determined through empirical or experimental values.

[0136] Therefore, it can be seen that the first probability density function determined by this scheme can avoid w in the calculation. t The division calculation can avoid w t The calculation anomalies caused by division can improve the accuracy of predicting the timing of events.

[0137] Step 603 will be explained below.

[0138] In this step, the self-t is determined according to the first probability density function. i When no behavioral events occur within a preset time interval starting from a given moment, the second probability density function can be determined using the following formula:

[0139]

[0140] Among them, S * (t) is used to characterize the second probability density function, λ * (t) is used to characterize the first probability density function;

[0141] Step 605 will be explained below.

[0142] When the description t is obtained iThe first probability density function after time t, where an action event occurs, and the probability density function of time t. i After determining the second probability density function, which indicates that no behavioral events occur within a preset time period starting from time t, the joint probability density function of the first and second probability density functions is considered. This joint probability density function represents the probability density of no behavioral events occurring within a certain time period, but with a behavioral event occurring at time t in the future; this is the second predicted probability distribution. In one possible implementation, step 605 can determine the second predicted probability distribution using the following formula:

[0143] f * (t)=λ * (t)·S * (t)

[0144] Among them, f * (t) is used to characterize the second prediction probability distribution, λ * (t) is used to characterize the first probability density function, S * (t) is used to characterize the second probability density function.

[0145] Thus, the relationship between behavioral events and the timing of their occurrence can be described through the second prediction probability distribution, thereby enabling the prediction of when behavioral events will occur.

[0146] like Figure 7 As shown in the embodiments of this specification, a method for predicting risky behavior is also provided, which may include the following steps:

[0147] Step 701: Obtain the Tth... p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event;

[0148] Step 703: Input the type identifier of at least one behavioral event and the corresponding time value of each behavioral event into the behavior prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training method of the behavioral prediction model as described in any embodiment of this specification;

[0149] Step 705: Search the historical behavior database for a data set that is consistent with both the type identifier prediction value and the time prediction value; wherein, the historical behavior database stores at least one data set, and each data set includes a type identifier for a behavior event and a time value corresponding to that behavior event;

[0150] Step 707: If not found, the behavioral event to be predicted is a risky behavioral event.

[0151] In this embodiment, when predicting risky behavior, the Tth digit can be obtained. p The type identifier and corresponding time value of the behavioral event prior to time are input into the behavior prediction model to obtain T. p The predicted type and time values ​​of the behavioral event to be predicted after a given time point are calculated. Since the type and timing of behavioral events often exhibit periodicity and regularity, a search can be conducted in a historical database to determine if a set of data exists that matches both the predicted type and time values. If a set is found, it indicates that the user's behavior falls within the scope of normal daily behavior and is not risky. If no set is found, it indicates that the behavioral event to be predicted is a risky event.

[0152] It's easy to understand that the historical behavior database should store users' non-risky daily behaviors. The historical behavior database should include several data sets, each containing a type identifier for a behavioral event and a corresponding time value.

[0153] It is worth noting that when inputting the type identifier of a behavioral event and its corresponding time value into the behavioral prediction model to obtain the predicted type identifier value, the type identifier with the highest probability value can be taken as the predicted type identifier value for the behavioral event to be predicted. However, for the prediction of the time of occurrence of a behavioral event, due to the second prediction probability distribution f... * (t) is a probability density function over continuous time. Outputting the value at the moment of maximum probability density is not necessarily reasonable; therefore, we can consider taking its expectation as the probability density function for t. P+1 The prediction of time can be further estimated using importance sampling. For example, the expectation of time t can be calculated using the following process. P+1 Estimate at any time:

[0154]

[0155]

[0156]

[0157] in, That is, for t P+1 The estimated value at time t, p(t) is an exponential distribution c·exp(-c(tt)). P ), c is the conditional intensity function defined in the traditional way, where t P The value at time t is c = exp(v t ·h j +w t (tt P )+b t ) = exp(vt ·h j +b t Then, randomly generate N random numbers {r1, r2, ..., r...}. N}, then we have,

[0158]

[0159] Therefore, based on the above prediction and estimation method, even when the first probability density function is complex or the numerical solution is not easy to calculate, the timing of the occurrence of behavioral events can be quickly predicted through random sampling algorithms.

[0160] like Figure 8 As shown in the embodiments of this specification, a training device for a behavior prediction model is also provided, including: a first acquisition module 801, a representation module 802, and a training module 803;

[0161] The first acquisition module 801 is configured to acquire the type identifier of the sample behavior event and the time information corresponding to the sample behavior event from historical behavior data; wherein, each type of sample behavior event corresponds to a type identifier;

[0162] The characterization module 802 is configured to characterize the sample behavior event in a continuous time domain based on the type identifier and time information of the sample behavior event obtained by the first acquisition module 801, thereby obtaining a behavior event characterization.

[0163] Training module 803 is configured to train a behavior prediction model based on the behavior event representation obtained by representation module 802, so as to optimize the type identifier and time prediction value output by the behavior prediction model.

[0164] In one possible implementation, when the representation module 802 represents the sample behavior event in the continuous time domain according to the type identifier and time information of the sample behavior event to obtain the behavior event representation, it is configured to perform the following operation for each sample behavior event:

[0165] Map the type identifier of the current sample behavior event to a preset space to obtain the mapping vector of the current sample behavior event;

[0166] Obtain the behavior event representation of the previous sample behavior event; where, when the current sample behavior event is the first sample behavior event, the behavior event representation of the previous sample behavior event is 0;

[0167] The behavioral event representation of the current sample behavioral event is determined based on the mapping vector of the current sample behavioral event, the time information corresponding to the current sample behavioral event, and the behavioral event representation of the previous sample behavioral event.

[0168] In one possible implementation, when the representation module 802 determines the behavior event representation of the current sample behavior event based on the mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavior event representation of the previous sample behavior event, it is configured to use the following calculation formula to determine the behavior event representation of the j-th sample behavior event:

[0169] h j =max{W y *y j +W t *t j +W h *h j-1 +b h ,0}

[0170] Among them, h j The behavioral event representation used to characterize the behavioral event of the j-th sample, y j The mapping vector t is used to characterize the behavioral event of the j-th sample. j h is used to characterize the time information corresponding to the behavior event of the j-th sample. j-1 W is used to characterize the behavioral event representation of the (j-1)th sample behavioral event. y The coefficients W used to characterize the linear transformation of behavioral events t The coefficients W are used to characterize the linear transformation of time information. h The coefficients, b, are used to characterize the linear transformation of the representation of the previous sample behavioral event. h The modifier used to characterize the behavioral event representation, max{…,0} is used to characterize the nonlinear activation function ReLU.

[0171] In one possible implementation, when the behavior prediction model includes a first prediction probability distribution for predicting the type of behavior event, the training module 803 is configured to perform the following operations when training the behavior prediction model based on the behavior event representation:

[0172] Based on the behavioral event representation, determine the first prediction probability distribution for predicting the type of behavioral event;

[0173] The model parameters in the first predicted probability distribution are optimized using the cross-entropy loss function.

[0174] In one possible implementation, when determining the first prediction probability distribution for predicting the type of a behavioral event based on the behavioral event representation, the training module 803 is configured to calculate the first prediction probability distribution for predicting the type of a behavioral event using the following formula:

[0175]

[0176] Where P(k) is used to characterize the first prediction probability distribution, V k The parameter matrix h, whose type is identified as k, is used to characterize the neural network layer of the behavior prediction model. j The behavior event representation is used to characterize the behavior event of the j-th sample, K is used to characterize the number of types of sample behavior events, and b1 and b2 are both adjustment amounts used to adjust the predicted value.

[0177] In one possible implementation, when the behavior prediction model includes a second prediction probability distribution that predicts the time of occurrence of a behavioral event, the training module 803 is configured to perform the following operations when training the behavior prediction model based on the behavioral event representation:

[0178] Based on the behavioral event representation, a second prediction probability distribution is determined to predict the time of occurrence of the behavioral event.

[0179] The model parameters in the second predicted probability distribution are optimized using a logarithmic function.

[0180] In one possible implementation, the historical behavior data is t. i Data from a point in time prior;

[0181] When training module 803 determines the second prediction probability distribution for predicting the time of occurrence of a behavioral event based on the behavioral event representation, it is configured to perform the following operations:

[0182] Based on the behavioral event representation, determine t i The first probability density function for a behavioral event occurring after time step;

[0183] Based on the first probability density function, determine the self-t i The second probability density function is defined as the function that indicates no behavioral events occur within a preset time interval starting from a given moment.

[0184] The second predicted probability distribution is determined based on the first probability density function and the second probability density function.

[0185] In one possible implementation, the training module 803 determines t based on the behavioral event representation. i When determining the first probability density function for a behavioral event occurring after a certain time, the configuration is such that the first probability density function is determined using the following formula:

[0186]

[0187] Where, λ * (t) is used to characterize the first probability density function, v t h is the model parameter for linearly transforming the representation of behavioral events. j The behavioral event representation used to characterize the behavioral event of the j-th sample, tj b is used to characterize the time when the j-th sample behavioral event occurs. t The correction is used to characterize the first probability density function. w1 and w2 are model parameters that transform the time quantity. p is used to characterize the hyperparameters of the behavior prediction model.

[0188] In one possible implementation, the training module 803 determines the self-t based on the first probability density function. i When no behavioral events occur within a preset time interval starting from a given moment, the second probability density function is configured to be calculated using the following formula:

[0189]

[0190] Among them, S * (t) is used to characterize the second probability density function, λ * (t) is used to characterize the first probability density function;

[0191] In one possible implementation, when determining the second predicted probability distribution based on the first probability density function and the second probability density function, the training module 803 is configured to calculate the second predicted probability distribution using the following formula:

[0192] f * (t)=λ * (t)·S * (t)

[0193] Among them, f * (t) is used to characterize the second prediction probability distribution, λ * (t) is used to characterize the first probability density function, S * (t) is used to characterize the second probability density function.

[0194] In one possible implementation, the sample behavioral events acquired by the first acquisition module 801 include: user-generated transaction behaviors; the transaction behaviors include at least one of shopping, transportation, dining, and entertainment.

[0195] like Figure 9 As shown, this specification also provides a risk behavior prediction device, including: a second acquisition module 901, an input module 902, a search module 903, and a determination module 904;

[0196] The second acquisition module 901 is configured to acquire the Tth... p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event;

[0197] Input module 902 is configured to input the type identifier of at least one behavioral event and the time value corresponding to each behavioral event obtained by the second acquisition module 901 into the behavioral prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training device of the behavioral prediction model as described in any of the above embodiments;

[0198] The lookup module 903 is configured to search the historical behavior database for a data set that is consistent with the type identifier prediction value obtained by the input module 902 and the time prediction value; wherein, the historical behavior database stores at least one data set, and each data set includes a type identifier of a behavior event and a time value corresponding to the behavior event;

[0199] The determination module 904 is configured such that if the search module 903 does not find the event, the event to be predicted is a risky event.

[0200] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods in any of the embodiments of the specification.

[0201] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any of the embodiments in this specification.

[0202] It is understood that the structures illustrated in the embodiments of this specification do not constitute a specific limitation on the training device for the behavior prediction model and the risk behavior prediction device. In other embodiments of this specification, the training device for the behavior prediction model and the risk behavior prediction device may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0203] The information interaction and execution process between the various units in the above-mentioned device are based on the same concept as the method embodiments in this specification, and the specific details can be found in the descriptions in the method embodiments in this specification, so they will not be repeated here.

[0204] Those skilled in the art will recognize that, in one or more of the examples above, the functions described herein can be implemented using hardware, software, widgets, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0205] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects described in this specification. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. Training methods for behavior prediction models, including: Extract the type identifier of the sample behavior event and the corresponding time information from historical behavior data; each type of sample behavior event corresponds to a type identifier; Based on the type identifier and time information of the sample behavioral events, the sample behavioral events are characterized in the continuous time domain to obtain behavioral event representations. The behavior prediction model is trained based on the behavior event representation to optimize the type identifier and time prediction values ​​output by the behavior prediction model; The behavior prediction model includes a second prediction probability distribution for predicting the time of occurrence of behavioral events; The step of training the behavior prediction model based on the behavior event representation includes: determining a second prediction probability distribution for predicting the time of occurrence of the behavior event based on the behavior event representation; and optimizing the model parameters in the second prediction probability distribution using a logarithmic function. The historical behavior data includes t i Data from a point in time prior; The step of determining the second prediction probability distribution for predicting the time of occurrence of a behavioral event based on the behavioral event representation includes: determining t based on the behavioral event representation. i The first probability density function of a behavioral event occurring after time t; based on the first probability density function, determine the probability density function of time t. i A second probability density function is defined as follows: no behavioral events occur within a preset time interval starting from a given time. The second predicted probability distribution is determined based on the first probability density function and the second probability density function. Wherein, determining the second predicted probability distribution based on the first probability density function and the second probability density function includes: The second predicted probability distribution is calculated using the following formula: f * (t)=λ * (t)·S * (t) Among them, f * (t) is used to characterize the second prediction probability distribution, λ * (t) is used to characterize the first probability density function, S * (t) is used to characterize the second probability density function.

2. The method according to claim 1, wherein, The step of representing the sample behavioral event in a continuous time domain based on the type identifier and time information of the sample behavioral event to obtain the behavioral event representation includes: For each sample behavior event, execute: Map the type identifier of the current sample behavior event to a preset space to obtain the mapping vector of the current sample behavior event; Obtain the behavior event representation of the previous sample behavior event; wherein, when the current sample behavior event is the first sample behavior event, the behavior event representation of the previous sample behavior event is 0; The behavioral event representation of the current sample behavior event is determined based on the mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavioral event representation of the previous sample behavior event.

3. The method according to claim 2, wherein, The step of determining the behavior event representation of the current sample behavior event based on the mapping vector of the current sample behavior event, the time information corresponding to the current sample behavior event, and the behavior event representation of the previous sample behavior event includes: The behavioral event representation of the j-th sample behavioral event is determined using the following formula: h j =max{W y *y j +W t *t j +W h *h j-1 +b h ,0} Among them, h j The behavioral event representation used to characterize the behavioral event of the j-th sample, y j The mapping vector t is used to characterize the behavioral event of the j-th sample. j h is used to characterize the time information corresponding to the behavior event of the j-th sample. j-1 W is used to characterize the behavioral event representation of the (j-1)th sample behavioral event. y The coefficients W used to characterize the linear transformation of behavioral events t The coefficients W are used to characterize the linear transformation of time information. h The coefficients, b, are used to characterize the linear transformation of the representation of the previous sample behavioral event. h The modifier used to characterize the behavioral event representation, max{…,0} is used to characterize the nonlinear activation function ReLU.

4. The method according to claim 1, wherein, The behavior prediction model includes a first prediction probability distribution for predicting the type of behavioral event; Training the behavior prediction model based on the behavior event representation includes: Based on the behavioral event representation, a first prediction probability distribution is determined for predicting the type of behavioral event; The model parameters in the first predicted probability distribution are optimized using the cross-entropy loss function.

5. The method according to claim 4, wherein, The step of determining a first prediction probability distribution for predicting the type of behavioral event based on the behavioral event representation includes: The first prediction probability distribution for predicting the type of behavioral event is calculated using the following formula: Where P(k) is used to characterize the first prediction probability distribution, V k The parameter matrix h, whose type identifier is k, is used to characterize the neural network layer corresponding to the behavior prediction model. j The behavior event representation is used to characterize the behavior event of the j-th sample, K is used to characterize the number of types of sample behavior events, and b1 and b2 are both adjustment amounts used to adjust the predicted value.

6. The method according to claim 1, wherein, The determination of t based on the behavioral event representation i The first probability density function for a behavioral event occurring after a certain time includes: The first probability density function is determined using the following formula: Where, λ * (t) is used to characterize the first probability density function, v t h is the model parameter for linearly transforming the representation of behavioral events. j The behavioral event representation used to characterize the behavioral event of the j-th sample, t j b is used to characterize the time when the j-th sample behavioral event occurs. t The correction is used to characterize the first probability density function. w1 and w2 are model parameters that transform the time quantity. p is used to characterize the hyperparameters of the behavior prediction model. And / or, The step of determining from t based on the first probability density function i The second probability density function, which specifies that no behavioral events occur within a preset time interval starting from a given moment, includes: The second probability density function is calculated using the following formula: Among them, S * (t) is used to characterize the second probability density function, λ * (t) is used to characterize the first probability density function.

7. The method according to any one of claims 1 to 6, wherein, The sample behavioral events include: user-generated transaction behaviors; the transaction behaviors include at least one of shopping, transportation, dining, and entertainment.

8. Risk behavior prediction methods, including: Get the Tth p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event; The obtained type identifiers of at least one behavioral event and the corresponding time values ​​of each behavioral event are input into the behavior prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training method of the behavioral prediction model as described in any one of claims 1 to 7; The historical behavior database is searched to determine if there is a data set that is consistent with both the predicted value of the type identifier and the predicted value of the time; wherein the historical behavior database stores at least one data set, and each data set includes a type identifier of a behavior event and a time value corresponding to the behavior event; If not found, the predicted behavioral event is a risky behavioral event.

9. A training device for a behavior prediction model, comprising: The first module consists of an acquisition module, a representation module, and a training module; The first acquisition module is configured to acquire the type identifier of the sample behavior event and the time information corresponding to the sample behavior event from historical behavior data; wherein, each type of sample behavior event corresponds to a type identifier; The characterization module is configured to characterize the sample behavior event in a continuous time domain based on the type identifier and time information of the sample behavior event obtained by the first acquisition module, thereby obtaining a behavior event characterization. The training module is configured to train the behavior prediction model based on the behavior event representation obtained by the representation module, so as to optimize the type identifier and time prediction value output by the behavior prediction model. The behavior prediction model includes a second prediction probability distribution that predicts the timing of behavioral events. When training a behavior prediction model based on behavior event representations, the training module is configured to perform the following operations: determine a second prediction probability distribution for predicting the time of occurrence of a behavior event based on the behavior event representation; and optimize the model parameters in the second prediction probability distribution using a logarithmic function. Historical behavioral data includes t i Data from a point in time prior; When the training module determines the second prediction probability distribution for predicting the time of occurrence of a behavioral event based on the behavioral event representation, it is configured to perform the following operation: determine t based on the behavioral event representation. i The first probability density function of a behavioral event occurring after time t; based on the first probability density function, determine the probability density function of time t. i A second probability density function is defined as follows: no behavioral events occur within a preset time interval starting from a given time. A second predicted probability distribution is determined based on the first and second probability density functions. Specifically, when determining the second predicted probability distribution based on the first probability density function and the second probability density function, the training module is configured to calculate the second predicted probability distribution using the following formula: f * (t)=λ * (t)·S * (t) Among them, f * (t) is used to characterize the second prediction probability distribution, λ * (t) is used to characterize the first probability density function, S * (t) is used to characterize the second probability density function.

10. A risk behavior prediction device, comprising: The second module includes an acquisition module, an input module, a search module, and a confirmation module. The second acquisition module is configured to acquire the Tth... p The type identifier of at least one behavioral event prior to the current time and the corresponding time value of each behavioral event; The input module is configured to input the type identifier of the at least one behavioral event and the time value corresponding to each behavioral event obtained by the second acquisition module into the behavioral prediction model to obtain the Tth... p The predicted value of the type identifier of a behavioral event to be predicted after a certain time, and the predicted value of the time corresponding to the behavioral event to be predicted; wherein, the behavioral prediction model is trained using the training device of the behavioral prediction model as described in claim 9; The search module is configured to search the historical behavior database for a data set that is consistent with the type identifier prediction value obtained by the input module and the time prediction value; wherein, the historical behavior database stores at least one data set, and each data set includes a type identifier of a behavior event and a time value corresponding to the behavior event; The determining module is configured such that if the searching module does not find the event, the event to be predicted is a risky event.

11. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-8.

Images