Behavior prediction model generation method and device, electronic equipment and storage medium

Abstract

The invention relates to a behavior prediction model generation method and device, electronic equipment, a storage medium and a program product, and the method comprises the steps: constructing an energy model based on the model parameters of an initial behavior prediction model and an energy function, and calculating the gradient data of the energy model, the training data comprises click behavior sample data and click behavior labels corresponding to the click behavior sample data; sampling the random noise data for a target number of times based on the gradient data to obtain prior noise data; performing diffusion processing on the prior noise data through a diffusion model to obtain target sample data; and obtaining a trained target behavior prediction model based on the target sample data and the training data. By adopting the method, the diffusion model can obtain the sample data with relatively strong relevance with the click behavior sample data adopted by the training behavior prediction model, and the generalization ability of the model is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular, to a method, apparatus, electronic device, storage medium, and program product for generating a behavior prediction model. Background Art

[0002] With the continuous development of artificial intelligence technology, the scope of application of diffusion technology is also becoming wider and wider. In related technologies, a model can be trained using diffusion data obtained from a diffusion model; for example, generating samples through a diffusion model is generally performed by an unconstrained direct random sampling method. Although the above diffusion process can generate diverse samples, and even any possible sample can be generated through a sufficient number of sampling processes. However, the distribution of the generated samples cannot be estimated, and the correlation between the generated samples is weak, resulting in a poor generalization effect of the model obtained by training using such samples. Summary of the Invention

[0003] The present disclosure provides a method, apparatus, electronic device, and storage medium for generating a behavior prediction model to at least solve the problem of poor generalization effect of the model in related technologies. The technical solution of the present disclosure is as follows:

[0004] According to a first aspect of an embodiment of the present disclosure, a method for generating a behavior prediction model is provided, including:

[0005] Based on the model parameters of an initial behavior prediction model and an energy function, an energy model is constructed, and gradient data of the energy model is calculated. The initial behavior prediction model is trained based on training data, and the training data includes click behavior sample data and click behavior labels corresponding to each click behavior sample data. The click behavior sample data is used to represent recommended data in a recommendation scenario, and the click behavior label is used to represent an interaction operation of a target object on the recommended data;

[0006] Based on the gradient data, target - number sampling is performed on random noise data to obtain prior noise data, where the random noise data is click behavior data obtained by random sampling;

[0007] The prior noise data is diffusion - processed through a diffusion model to obtain target sample data;

[0008] Based on the target sample data and the training data, the initial behavior prediction model is trained to obtain a trained target behavior prediction model.

[0009] In one of the embodiments, the constructing an energy model based on the model parameters of an initial behavior prediction model and an energy function includes:

[0010] Construct an energy model corresponding to the initial behavior prediction model based on the network parameters of the initial behavior prediction model, click behavior sample data, click behavior labels corresponding to each click behavior sample data, and the relationships between variables characterized by the energy function.

[0011] In one embodiment, calculating the gradient data of the energy model includes:

[0012] Through the energy model, calculate the target distribution between the click behavior sample data and the click behavior label corresponding to the click behavior sample data;

[0013] Based on the target distribution, calculate the gradient data of the energy model.

[0014] In one embodiment, the sampling the random noise data for a target number of times based on the gradient data to obtain prior noise data includes:

[0015] Sample the random noise data for a target number of times based on the gradient data, and determine the noise data corresponding to the target number of times of sampling as the prior noise data.

[0016] In one embodiment, the method further includes:

[0017] For the i-th sampling in the target number of times, calculate the noise data corresponding to the i-th sampling based on the noise data corresponding to the (i - 1)-th sampling, sampling parameters, the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling, and standard parameters, where i is greater than or equal to 1 and i is a positive integer.

[0018] In one embodiment, the calculating the noise data corresponding to the i-th sampling based on the noise data corresponding to the (i - 1)-th sampling, sampling parameters, the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling, and standard parameters includes:

[0019] Adjust the noise data corresponding to the (i - 1)-th sampling and the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling based on the sampling parameters and the standard parameters to obtain an adjustment factor;

[0020] Adjust the noise data corresponding to the (i - 1)-th sampling through the adjustment factor to obtain the noise data corresponding to the i-th sampling.

[0021] In one embodiment, the diffusing the prior noise data through a diffusion model to obtain target sample data includes:

[0022] Determine the prior noise data as the sampling starting point of the diffusion model, and perform sampling for the number of time steps of the time step hyperparameter on the prior noise data based on the time step hyperparameter of the diffusion model to obtain target sample data.

[0023] In one embodiment, the target sample data includes a plurality of click behavior data; training the initial behavior prediction model based on the target sample data and the training data to obtain a trained target behavior prediction model includes:

[0024] Through the initial behavior prediction model, obtain the label data corresponding to each click behavior data included in the target sample data;

[0025] Based on each click behavior data included in the target sample data, the label data corresponding to each click behavior data, and the training data, train the initial behavior prediction model to obtain a trained target behavior prediction model.

[0026] According to a second aspect of the embodiments of the present disclosure, there is provided a generating device for a behavior prediction model, including:

[0027] A construction unit configured to execute constructing an energy model based on the model parameters of the initial behavior prediction model and the energy function, and calculating the gradient data of the energy model, where the initial behavior prediction model is trained based on training data, the training data includes click behavior sample data and click behavior labels corresponding to each click behavior sample data, the click behavior sample data is used to characterize the recommended data in the recommendation scenario, and the click behavior label is used to characterize the interaction operation of the target object on the recommended data;

[0028] A sampling unit configured to execute sampling the random noise data for a target number of times based on the gradient data to obtain prior noise data;

[0029] A diffusion unit configured to execute diffusing the prior noise data through a diffusion model to obtain target sample data;

[0030] A training unit configured to execute training the initial behavior prediction model based on the target sample data and the training data to obtain a trained target behavior prediction model.

[0031] According to a third aspect of the embodiments of the present disclosure, there is provided an electronic device, including:

[0032] A processor;

[0033] A memory for storing executable instructions of the processor;

[0034] Among them, the processor is configured to execute the instructions to implement the method for generating a behavior prediction model as described in any one of the above first aspects.

[0035] According to a fourth aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enabling the electronic device to execute the method for generating a behavior prediction model as described in any one of the above first aspects.

[0036] According to a fifth aspect of the embodiments of the present disclosure, there is provided a computer program product, when the instructions are executed by a processor of an electronic device, enabling the electronic device to execute the method for generating a behavior prediction model as described in any one of the above first aspects.

[0037] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0038] By adopting this method, based on the model parameters of the pre-trained behavior prediction model and the correlation relationships between the various variables characterized by the energy function, an energy model is constructed, and kinetic sampling is performed on the noise data randomly sampled at random sampling points through the gradient data of the energy model to obtain noise data that has incorporated the prior knowledge from the behavior prediction model, and this noise data is used as the sampling starting point of the diffusion model, which can make the sample data generated by the diffusion model through diffusion have a strong correlation with the click behavior sample data used for training the behavior prediction model, improving the generalization ability of the model. The diffusion model can diffuse meaningful and high-quality sample data based on the adapted noise data, also improving the prediction performance of the behavior prediction model, and further reducing the cost of generating training sample data, etc.

[0039] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure, and do not constitute an improper limitation to the present disclosure.

[0041] Figure 1 It is an application environment diagram of a method for generating a behavior prediction model shown according to an exemplary embodiment.

[0042] Figure 2 It is a flowchart of a method for generating a behavior prediction model shown according to an exemplary embodiment.

[0043] Figure 3It is a flowchart of the model training step in a method for generating a behavior prediction model shown according to an exemplary embodiment.

[0044] Figure 4 It is a flowchart of a method for generating a behavior prediction model shown according to another exemplary embodiment.

[0045] Figure 5 It is a block diagram of a device for generating a behavior prediction model shown according to an exemplary embodiment.

[0046] Figure 6 It is a block diagram of an electronic device shown according to an exemplary embodiment. Detailed implementation manners

[0047] To enable those of ordinary skill in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0048] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0049] It should also be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in the present disclosure are all information and data that have been authorized by the user or fully authorized by all parties.

[0050] Figure 1 It is a flowchart of a method for generating a behavior prediction model shown according to an exemplary embodiment. As Figure 1 shown, the method for generating a behavior prediction model is used in a terminal; it can be understood that this method can also be applied to a system including a terminal and a server and implemented through the interaction between the terminal and the server. Among them, the terminal can be but is not limited to various personal computers, laptop computers, smart phones, and tablet computers, and the server can be implemented by an independent server or a server cluster composed of multiple servers. In this exemplary embodiment, the method includes the following steps:

[0051] In step S110, an energy model is constructed based on the model parameters of the initial behavior prediction model and the energy function, and the gradient data of the energy model is calculated;

[0052] Among them, the initial behavior prediction model is trained based on training data. The training data includes click behavior sample data and click behavior labels corresponding to each click behavior sample data. The click behavior sample data is used to represent the recommended data in the recommendation scenario, and the click behavior label is used to represent the interaction operation of the target object on the recommended data. The energy function represents the correlation relationship between various variables in the energy model. The model parameters of the initial behavior prediction model can be the parameters of the network of this initial behavior prediction model, etc.; the gradient data of this energy model can be determined based on the click behavior sample data included in the training data and the click behavior labels corresponding to each click behavior sample data. The click behavior sample data can be exposure materials, such as video data, image data, or picture-text data, etc. The click behavior label corresponding to the click behavior sample data represents the probability value that the user will click on this exposure material, or can also include an identifier indicating that the user will click on this exposure material or an identifier indicating that the user will not click on this exposure material, etc.

[0053] Specifically, the terminal can process the click behavior sample data included in the training data, the click behavior labels corresponding to each click behavior sample data, and the model parameters of the initial behavior prediction model through the correlation relationship represented by the energy function, that is, input the above data into the energy function to obtain the energy model. The terminal can calculate the gradient data of this energy model based on the obtained energy model.

[0054] In step S120, the random noise data is sampled a target number of times based on the gradient data to obtain prior noise data;

[0055] Among them, the random noise data is click behavior data obtained by random sampling. The target number can be determined by a preset sampling method. The preset sampling can be the Langevin dynamics sampling method; the random noise data can be noise data obtained by randomly sampling from standard Gaussian distribution data. The random noise data x 0 is click behavior data, and this click behavior data can be an exposure material generated by random sampling. The random noise data can be initial noise.

[0056] Specifically, the terminal can perform processing that incorporates the prior indication of the initial behavior prediction model on the random noise data based on the calculated gradient data. Specifically, the terminal can perform sampling processing on the random noise data a target number of times based on the gradient data calculated by the energy model to obtain prior noise data that incorporates the prior knowledge of the initial behavior prediction model. Among them, the target number can be determined based on the actual application scenario.

[0057] Optionally, the terminal may perform Langevin dynamics sampling on the random noise data for a target number of times based on the gradient data, and determine the data obtained after the target number of samplings as the prior noise data.

[0058] In step S130, the prior noise data is processed by a diffusion model to obtain target sample data;

[0059] Among them, the diffusion model may be a preset diffusion model. The target sample data may include multiple click behavior data, for example, it may include multiple exposure material data for determining whether a user clicks, etc.

[0060] Specifically, the terminal may input the prior noise data into the diffusion model, so that the diffusion model performs data diffusion processing based on the prior noise data, and determines the diffusion result output by the diffusion model as the target sample data. The specific diffusion process may be that the terminal uses the prior noise data as the sampling starting point of the diffusion model, performs a standard number of samplings within the diffusion model based on the sampling starting point, realizes the diffusion generation of click behavior data, and obtains the target sample data output by the diffusion model; the standard number may be the time step hyperparameter of the diffusion model.

[0061] That is to say, the terminal may input the prior noise data into the diffusion model, use the noise prior data as the sampling starting point in the diffusion model, and determine the number of samplings based on the time step hyperparameter of the diffusion model, that is, a standard number of times; perform a standard number of samplings based on the sampling starting point, and use the sampling result as the output result of the diffusion model, that is, use the output result of the diffusion model as the target sample data.

[0062] In step S140, the initial behavior prediction model is trained based on the target sample data and the training data to obtain a trained target behavior prediction model.

[0063] Specifically, the terminal may mix the target sample data, the click behavior sample data included in the training data, and the click behavior labels corresponding to each click behavior sample data to obtain the mixed sample data, and train the trained initial behavior prediction model in the above embodiment based on the mixed sample data. When the current initial behavior prediction model meets the preset training completion condition, a trained target behavior prediction model is obtained, and this target behavior prediction model can be used to predict the click probability of a user clicking on a certain exposure material.

[0064] The preset training completion condition may be that the calculated loss value has converged, or it may also be that the training duration reaches a preset duration threshold, or it may also be that the number of training iterations has reached a preset target number. The present disclosure does not limit the specific values of the preset duration threshold and the preset target number, and those skilled in the art can specifically determine them according to the actual application scenario.

[0065] Optionally, the mixed sample data includes sample exposure materials and sample click behavior labels corresponding to the sample exposure materials. The terminal can obtain the predicted click behavior labels corresponding to the sample exposure materials output by the initial behavior prediction model, and calculate the loss value based on the sample click behavior labels and the predicted click behavior labels. For example, the distance between the sample click behavior label and the predicted click behavior label can be calculated, and the loss value can be calculated through this distance, etc. In this way, the terminal can update the model parameters of each module included in the label synthesis model to be trained based on the loss value, and return to execute the training step until the loss value meets the preset training completion condition, and obtain the target behavior prediction model with training completed.

[0066] In the above method for generating the behavior prediction model, based on the model parameters of the pre-trained behavior prediction model and the correlation relationship between the variables characterized by the energy function, an energy model is constructed, and kinetic sampling is performed on the noise data randomly sampled at the random sampling points through the gradient data of the energy model, so as to obtain the noise data that has incorporated the prior knowledge from the behavior prediction model, and use this noise data as the sampling starting point of the diffusion model, which can make the sample data generated by the diffusion model through diffusion have a strong correlation with the click behavior sample data used for training the behavior prediction model, improve the generalization ability of the model, the diffusion model can diffuse meaningful and high-quality sample data based on the adapted noise data, also improve the prediction performance of the behavior prediction model, and further reduce the cost of generating training sample data, etc.

[0067] In an exemplary embodiment, the specific implementation process of the step "construct an energy model based on the model parameters of the initial behavior prediction model and the energy function" may include:

[0068] Construct an energy model corresponding to the initial behavior prediction model based on the network parameters of the initial behavior prediction model, the click behavior sample data, the click behavior labels corresponding to each click behavior sample data, and the relationship between the variables characterized by the energy function.

[0069] Among them, the click behavior labels corresponding to each click behavior sample data may be the sample labels corresponding to the click behavior sample data.

[0070] Specifically, the terminal can use the network parameters of the initial behavior prediction model, the click behavior sample data, the click behavior labels corresponding to the respective click behavior sample data, and the relationships between the various variables characterized by the energy function. The network parameters, the click behavior sample data, and the click behavior labels corresponding to the respective click behavior sample data are used as variables in the energy function to obtain the energy model corresponding to the initial behavior prediction model.

[0071] Based on the above solution, an energy model corresponding to the initial behavior prediction model can be constructed, which can provide a data basis for subsequently incorporating the initial noise data into the prior knowledge of the behavior prediction model.

[0072] In an exemplary embodiment, as Figure 2 shown, the specific implementation process of the step "calculating the gradient data of the energy model" may include:

[0073] Step 210, calculate the target distribution between the click behavior sample data and the click behavior label corresponding to the click behavior sample data through the energy model;

[0074] Among them, the second label of the click behavior label may be the distribution of the click category label.

[0075] Specifically, the terminal can calculate the target distribution between the click behavior sample data and the click behavior label corresponding to the click behavior sample data based on the energy model. Among them, the click behavior label corresponding to the click behavior sample data may be the sample label in the training data.

[0076] Step 220, calculate the gradient data of the energy model based on the target distribution.

[0077] Specifically, the terminal can perform a conversion process based on the first distribution and the second distribution to obtain the gradient data corresponding to the energy model. The terminal can obtain the gradient data corresponding to the energy model through the following formula.

[0078] In an example, the energy model corresponding to the initial behavior prediction model obtained by the terminal can be represented by the following formula:

[0079] E θ (x,y)=-f θ (x)[y]

[0080] The terminal can perform a conversion on the above energy model to obtain a converted energy model, which can be represented by the following formula:

[0081]

[0082] Based on this, the terminal can calculate the target distribution of the click behavior sample data and the click behavior label corresponding to the click behavior sample data based on the converted energy model. For example, the target distribution of the click behavior sample data and the click behavior label corresponding to the click behavior sample data can be calculated through the following formula:

[0083]

[0084] Z θ = ∫x exp(-E θ (x))

[0085] where x represents the click behavior sample data, y represents the click behavior label corresponding to the click behavior sample data, and pθ(x, y) represents the target distribution.

[0086] The terminal can calculate the gradient data corresponding to the energy model based on the energy model and the target distribution. For example, the gradient data (which can be denoted as ) can be calculated through the following formula:

[0087]

[0088] Based on the above solution, the terminal can calculate the gradient data corresponding to the energy model based on the click behavior sample data used to train the initial behavior prediction model and the click behavior labels corresponding to each click behavior sample data, providing a data basis for subsequent kinetic sampling, also enhancing the correlation between the subsequently generated samples and the training data, and ensuring the quality of the generated sample data.

[0089] In an exemplary embodiment, the specific implementation process of the step "sampling the random noise data for a target number of times based on the gradient data to obtain prior noise data" may include:

[0090] Sampling the random noise data for a target number of times based on the gradient data, and determining the noise data corresponding to the target number of times of sampling as the prior noise data.

[0091] where the target number can be the number of sampling times determined based on the actual application scenario or the number of sampling times determined based on the sampling method.

[0092] Specifically, the initial random noise data is sampled a target number of times based on the calculated gradient data of the energy model. The specific sampling process can be as follows: The terminal performs the first sampling on the initial random noise data based on the initial gradient data to obtain the first sampling result, and the terminal can adjust the initial gradient data based on the first sampling result to obtain the gradient data corresponding to the first sampling. Based on this, the terminal can perform the second sampling, that is, perform the second sampling based on the first sampling result and the gradient data corresponding to the first sampling to obtain the second sampling result and the gradient data corresponding to the second sampling, until the sampling result of the target number of times is obtained. The terminal can determine the sampling result of the target number of times as the prior noise data.

[0093] Based on the above solution, the initial random noise data can be sampled multiple times using Langevin dynamics based on the gradient data. The prior knowledge of the initial behavior prediction model can be incorporated into the random noise data, which can improve the correlation between the prior noise data obtained after sampling and the training data used to train the initial behavior prediction model.

[0094] In an exemplary embodiment, the method further includes:

[0095] For the i-th sampling in the target number of times, based on the noise data corresponding to the (i - 1)-th sampling, the sampling parameter, the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling, and the standard parameter, calculate the noise data corresponding to the i-th sampling, where i is greater than or equal to 1 and i is a positive integer.

[0096] Among them, the standard parameter follows a standard distribution, and the standard distribution can be determined based on the sampling parameter. For example, the sampling parameter can be α / 2, and the standard parameter can be ∈.

[0097] Specifically, for the i-th sampling in the target number of times, the specific sampling process can be that the terminal can adjust the gradient data based on the sampling result of the (i - 1)-th time to obtain the gradient data corresponding to the (i - 1)-th sampling, and perform the i-th sampling based on the gradient data corresponding to the (i - 1)-th sampling, the sampling result of the (i - 1)-th sampling, the sampling parameter, and the standard parameter to obtain the sampling result of the i-th time.

[0098] Optionally, the terminal can calculate the sampling result of the (i + 1)-th time (denoted as x i+1 ) through the following formula:

[0099]

[0100] Among them, x 0 represents the random noise data, that is, the random noise data randomly sampled from the standard Gaussian distribution data. p0 (x) represents the distribution corresponding to the random noise data, and ∈ is the standard parameter.

[0101] Based on the above solution, multiple Langevin dynamics samplings can be performed on the initial random noise data based on the gradient data. The prior knowledge of the initial behavior prediction model can be incorporated into the random noise data, and the correlation between the prior noise data obtained after sampling and the training data used to train the initial behavior prediction model can be improved.

[0102] In an exemplary embodiment, the specific implementation process of the step "calculating the noise data corresponding to the i-th sampling based on the noise data corresponding to the (i - 1)-th sampling, the sampling parameter, the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling, and the standard parameter" may include:

[0103] Adjust the noise data corresponding to the (i - 1)-th sampling and the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling based on the sampling parameter and the standard parameter to obtain an adjustment factor;

[0104] Adjust the noise data corresponding to the (i - 1)-th sampling through the adjustment factor to obtain the noise data corresponding to the i-th sampling.

[0105] Among them, the sampling parameter and the standard parameter can be determined based on the actual scenario. The sampling parameter can be, for example, α / 2, and the standard parameter can be a parameter that follows a distribution, such as ∈.

[0106] Specifically, the terminal can calculate the derivative of the noise data corresponding to the (i - 1)-th sampling and the noise data obtained from the (i - 1)-th sampling, and perform a first adjustment on the derivative through the sampling parameter to obtain a first-adjusted derivative, and then perform a second adjustment on the first-adjusted derivative through the standard parameter, and determine the second-adjusted derivative as the adjustment factor. Based on this, the terminal can adjust the noise data corresponding to the (i - 1)-th sampling through the adjustment factor. For example, it can calculate the sum value between the noise data corresponding to the (i - 1)-th sampling and the adjustment factor, and determine the sum value as the noise data corresponding to the i-th sampling.

[0107] Based on the above solution, multiple Langevin dynamics samplings can be performed on the initial random noise data based on the gradient data. The prior knowledge of the initial behavior prediction model can be incorporated into the random noise data, and the correlation between the prior noise data obtained after sampling and the training data used to train the initial behavior prediction model can be improved.

[0108] In an exemplary embodiment, the specific implementation process of the step "performing diffusion processing on the prior noise data through a diffusion model to obtain target sample data" may include:

[0109] Determine the prior noise data as the sampling starting point of the diffusion model, and perform the number of sampling steps equal to the time step hyperparameter of the diffusion model on the prior noise data to obtain the target sample data.

[0110] Among them, the time step hyperparameter can be the time step hyperparameter in the diffusion model.

[0111] Specifically, in the diffusion model, the terminal can use the prior noise data as the sampling starting point for the sampling process of the diffusion model, and determine the number of sampling steps of the diffusion model based on the time step hyperparameter of the diffusion model. Perform sampling for the number of steps of the time step hyperparameter at the sampling starting point to obtain the sampling result, and use this sampling result as the target sample data output by the diffusion model.

[0112] Optionally, the time step hyperparameter can be T. The terminal can use this prior noise data as the sampling starting point for the sampling process in the diffusion model, and perform sampling for T steps, and use the obtained sampling result as the target sample data output by the diffusion model.

[0113] Based on the above solution, performing sample diffusion processing on the diffusion model based on the prior noise data can ensure that the generated target sample data can be constrained by the pre-trained initial behavior prediction model. There is a strong correlation between the target sample data and the training data used to train the initial behavior prediction model, and it can also enable the diffusion model to generate target sample data that better conforms to the user pattern, improving the quality of the obtained target sample data.

[0114] In an exemplary embodiment, the target sample data includes multiple click behavior data. Specifically, the multiple click behavior data included in the target sample data can be multiple exposure materials.

[0115] Correspondingly, as Figure 3 shown, the specific implementation process of the step "Based on the target sample data and the training data, train the initial behavior prediction model to obtain the trained target behavior prediction model" may include:

[0116] Step 310, through the initial behavior prediction model, obtain the label data corresponding to each click behavior data included in the target sample data;

[0117] Specifically, the terminal can input each click behavior data included in the target sample data into the initial behavior prediction model respectively, and obtain the label data corresponding to each click behavior data output by the initial behavior prediction model. This label data can be the probability that the initial behavior prediction model predicts that the user will click on this exposure material.

[0118] Step 320: Train the initial behavior prediction model based on each click behavior data included in the target sample data, the label data corresponding to each click behavior data, and the training data to obtain a trained target behavior prediction model.

[0119] Specifically, the terminal can mix each click behavior data included in the target sample data, the label data corresponding to each click behavior data, each click behavior sample data included in the training data, and the click behavior labels corresponding to each click behavior sample data to obtain mixed sample data. Then, based on the mixed sample data, train the trained initial behavior prediction model in the above embodiment. When the current initial behavior prediction model meets the preset training completion condition, obtain the trained target behavior prediction model, which can be used to predict the click probability of a user clicking on a certain exposure material.

[0120] Among them, the preset training completion condition can be that the calculated loss value has converged, or it can also be that the training duration reaches a preset duration threshold, or it can also be that the number of training iterations has reached a preset target number. The present disclosure does not limit the specific values of the preset duration threshold and the preset target number, and those skilled in the art can specifically determine them according to the actual application scenario.

[0121] Optionally, the mixed sample data includes a sample exposure material and a sample click behavior label corresponding to the sample exposure material. The terminal can obtain the predicted click behavior label corresponding to the sample exposure material output by the initial behavior prediction model, and calculate the loss value based on the sample click behavior label and the predicted click behavior label. For example, the distance between the sample click behavior label and the predicted click behavior label can be calculated, and the loss value can be calculated through this distance, etc. In this way, the terminal can update the model parameters of each module included in the label synthesis model to be trained based on the loss value, and return to execute the training step until the loss value meets the preset training completion condition to obtain the trained target behavior prediction model.

[0122] Based on the above solution, by adjusting the pre-trained initial behavior prediction model based on the mixed sample data, a prediction model with stronger generalization ability can be obtained, improving the performance of the behavior prediction model.

[0123] The following describes in detail the specific implementation process of the above method for generating a behavior prediction model in combination with a specific embodiment:

[0124] In the related art, a model can be trained with training data containing user click behaviors to obtain a behavior prediction model. When actually deployed to a production scenario, since the distribution of user samples to be predicted is somewhat offset from the distribution of user samples used by the model during training, the direct prediction ability of the behavior prediction model inevitably degrades. A diffusion model, as a generative model based on sample noise prediction, can generate samples that conform to the overall distribution of user click behavior samples. However, the diffusion model tends to generate samples highly correlated with the mean of its training samples. That is to say, there are significant differences between the samples directly generated by the diffusion model and the user click behavior samples used to actually train the behavior prediction model, as well as the user click behavior samples that the behavior prediction model in the production environment needs to predict. As a result, there is no guarantee for the generated samples to improve the generalization ability of the behavior prediction model.

[0125] In the process of generating target sample data by a diffusion model involved in a method for generating a behavior prediction model provided by the present disclosure, considering that user click behaviors are regular and evolutionarily continuous in time, the behavior prediction model can be guided on the initial noise prior during sampling, increasing the relevance of the sampled user click behavior samples to the samples used during training. Thereby, the generalization effect of the behavior prediction model is improved, and the probability of prediction errors on samples not seen during training can also be reduced. The usability of the model in the production environment is enhanced, ensuring the high quality and diversity of the samples generated by the diffusion model, and generating more diverse samples to assist the behavior prediction model in efficiently improving its generalization ability on production scenario samples not seen during training.

[0126] A method for generating a behavior prediction model provided by the present disclosure can estimate the possible distribution of the generated samples, ensuring the relevance of the generated samples to the samples and the model that actually need to be generalized. Moreover, it does not require a large number of samplings, improving the efficiency of enhancing the generalization ability of the behavior prediction model, achieving more comprehensive sample coverage, and being applicable to large-scale deployment in actual scenarios.

[0127] That is to say, a method for generating a behavior prediction model provided by the present disclosure can, during the process of the diffusion model sampling to generate user click behavior samples as training samples for the behavior prediction model to further improve its generalization ability, constrain the generated samples to both maintain the original generation quality and be adapted to the original training task objective (i.e., having a high correlation with the training data for training the initial behavior prediction model). That is, the generated samples have a good correlation with the original training samples, so that the generated samples can truly simulate the generalization target samples that the behavior prediction model has to face, namely the user click behavior samples with evolutionary continuity, and ultimately achieve an efficient improvement in the generalization ability of the behavior prediction model.

[0128] A method for generating a behavior prediction model provided by the present disclosure is to adapt the initial random noise data to the prior diffusion model constraint sampling of the behavior prediction model, and then use the generated samples for the training of the behavior prediction model to enhance its generalization ability. For example, Figure 4 as shown, this step includes:

[0129] The terminal can collect user click behavior samples, perform behavior prediction model training, initial noise prior adaptation, and behavior sample diffusion generation to obtain click behavior generation samples, and fine-tune the behavior prediction model to obtain the target behavior prediction model.

[0130] Optionally, the step of behavior prediction model training can be to train the behavior prediction model to be trained according to the user click behavior samples to obtain an initial behavior prediction model. The predicted binary classification result output by this initial behavior prediction model represents the probability that the user clicks on the exposure material. The larger this value, the higher the probability that the surface model believes the user will perform a click behavior on the exposure material. The user click behavior samples can be click behavior labels indicating whether the user will click on a given exposure material, that is, the exposure material and the probability that the user corresponding to the exposure material will click on the exposure material.

[0131] Optionally, the step of initial noise prior adaptation, which is to adjust the initial noise of the diffusion model sampling with the prior provided by the behavior prediction model, can be: The diffusion model in the present disclosure is obtained through large-scale pre-training on a large number of user click behavior samples. In order to make the generated samples of the diffusion model have a stronger correlation with the sample prediction knowledge prior of the behavior prediction model, the present disclosure constructs the behavior prediction model as an energy model and incorporates the prior into the randomly sampled Gaussian noise. The behavior prediction model can be denoted as fθ(x), where θ is the model parameter of the behavior prediction model, x is the user click behavior sample, and the class label y ∈ {0, 1} represents the classification of the user click behavior. The corresponding constructed energy model is: The energy model corresponding to the initial behavior prediction model obtained by the terminal can be expressed by the following formula:

[0132] E θ (x,y) = -f θ (x)[y]

[0133] The terminal can transform the above energy model to obtain the transformed energy model, which can be expressed by the following formula:

[0134]

[0135] Based on this, the terminal can calculate the target distribution of the click behavior sample data and the click behavior label corresponding to the click behavior sample data based on the converted energy model. For example, the target distribution of the click behavior sample data and the click behavior label corresponding to the click behavior sample data can be calculated through the following formula:

[0136]

[0137] Z θ =∫xexp(-E θ (x))

[0138] where x represents the click behavior sample data, y represents the click behavior label corresponding to the click behavior sample data, and p θ (x,y) represents the target distribution.

[0139] The terminal can calculate the gradient data corresponding to the energy model based on the energy model and the target distribution. For example, the gradient data (which can be denoted as ) can be calculated through the following formula:

[0140]

[0141] The terminal can calculate the sampling result of the (i + 1)-th time (which can be denoted as x i+1 ) through the following formula:

[0142]

[0143] where, x 0 represents random noise data, that is, random noise data randomly sampled from standard Gaussian distribution data. p0(x) represents the distribution corresponding to the random noise data, and ∈ is a standard parameter.

[0144] Optionally, for the generation of behavior sample diffusion, starting from the initial noise incorporating the prior, the process of diffusion model sampling can specifically include: after the terminal performs prior adaptation on the initial noise x 0 from the behavior prediction model, the new initial noise obtained is used as the sampling starting point of the diffusion model, and the new initial noise is prior noise data; and behavior sample diffusion generation is performed through T-step sampling actions, where T is the time step hyperparameter of the diffusion model, and the sampling result is denoted as the target sample data output by the diffusion model. The behavior samples after sampling are subject to the prior constraint from the behavior prediction model and have a strong correlation with the samples used to train the behavior prediction model; the behavior samples generated by the diffusion model sampling process have high quality and are more in line with the user behavior pattern. After collecting the samples, a click behavior generation sample data set is obtained.

[0145] Optionally, the steps of fine-tuning the behavior prediction model to obtain a behavior prediction model with stronger generalization ability may include: obtaining target sample data pairs and the label data corresponding to the target sample data, mixing the target sample data pairs and the label data corresponding to the target sample data with user click behavior samples, and then fine-tuning the behavior prediction model in the same training manner as in the first step to obtain a target behavior prediction model, which can make the target behavior prediction model a prediction model with stronger generalization ability when deployed.

[0146] A method for generating a behavior prediction model provided by the present disclosure constrains the initial noise in the sampling process of the diffusion model, incorporates prior knowledge from the behavior prediction model in the early stage of the diffusion model sampling, makes the generated user click samples have a strong correlation with the user click behavior samples used for training the behavior prediction model, and promotes efficient generation beneficial to the model generalization performance. The adapted initial noise can still be used as the sampling starting point of the diffusion model, that is, the diffusion model can sample meaningful and high-quality samples based on the adapted initial noise, further promoting the performance improvement of the behavior prediction model.

[0147] It should be understood that although Figures 1-4 the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order restriction, and these steps can be executed in other orders. Moreover, Figures 1-4 at least a part of the steps in

[0148] may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0149] Figure 5 is a block diagram of a device 500 for generating a behavior prediction model shown according to an exemplary embodiment. Referring to Figure 5 , the device includes a construction unit 502, a sampling unit 504, a diffusion unit 506, and a training unit 508.

[0150] The construction unit 502 is configured to execute the model parameters and the energy function of the initial behavior prediction model, construct an energy model, and calculate the gradient data of the energy model. The initial behavior prediction model is trained based on training data, which includes click behavior sample data and click behavior labels corresponding to each click behavior sample data. The click behavior sample data is used to represent the recommended data in the recommendation scenario, and the click behavior label is used to represent the interaction operation of the target object on the recommended data.

[0151] The sampling unit 504 is configured to perform target - number sampling on the random noise data based on the gradient data to obtain prior noise data.

[0152] The diffusion unit 506 is configured to perform diffusion processing on the prior noise data through a diffusion model to obtain target sample data.

[0153] The training unit 508 is configured to perform training on the initial behavior prediction model based on the target sample data and the training data to obtain a trained target behavior prediction model.

[0154] In one embodiment, the construction unit is specifically configured to:

[0155] Based on the network parameters of the initial behavior prediction model, the click behavior sample data, the click behavior labels corresponding to each click behavior sample data, and the relationship between the variables represented by the energy function, construct the energy model corresponding to the initial behavior prediction model.

[0156] In one embodiment, the construction unit is further specifically configured to:

[0157] Through the energy model, calculate the target distribution between the click behavior sample data and the click behavior label corresponding to the click behavior sample data.

[0158] Based on the target distribution, calculate the gradient data of the energy model.

[0159] In one embodiment, the sampling unit is specifically configured to:

[0160] Perform target - number sampling on the random noise data based on the gradient data, and determine the noise data corresponding to the target - number sampling as the prior noise data.

[0161] In one embodiment, the device further includes:

[0162] A first computing unit, configured to perform the i-th sampling among a target number of times, and calculate the noise data corresponding to the i-th sampling based on the noise data corresponding to the (i - 1)-th sampling, sampling parameters, the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling, and standard parameters, where i is greater than or equal to 1 and i is a positive integer.

[0163] In one embodiment, the first computing unit is specifically configured to:

[0164] Adjust the noise data corresponding to the (i - 1)-th sampling and the gradient data corresponding to the noise data obtained from the (i - 1)-th sampling based on the sampling parameters and the standard parameters to obtain an adjustment factor;

[0165] Adjust the noise data corresponding to the (i - 1)-th sampling through the adjustment factor to obtain the noise data corresponding to the i-th sampling.

[0166] In one embodiment, the diffusion unit is specifically configured to:

[0167] Determine the prior noise data as the sampling starting point of the diffusion model, and perform sampling for the number of time steps of the time step hyperparameter of the diffusion model on the prior noise data to obtain target sample data.

[0168] In one embodiment, the target sample data includes multiple click behavior data; the training unit is specifically configured to:

[0169] Obtain the label data corresponding to each click behavior data included in the target sample data through the initial behavior prediction model;

[0170] Train the initial behavior prediction model based on each click behavior data included in the target sample data, the label data corresponding to each click behavior data, and the training data to obtain a trained target behavior prediction model.

[0171] Regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0172] Figure 6 It is a block diagram of an electronic device 600 for a method of generating a behavior prediction model shown according to an exemplary embodiment. For example, the electronic device 600 can be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0173] Refer to Figure 6, the electronic device 600 may include one or more of the following components: a processing component 602, a memory 604, a power supply component 606, a multimedia component 608, an audio component 610, an input / output (I / O) interface 612, a sensor component 614, and a communication component 616.

[0174] The processing component 602 generally controls the overall operation of the electronic device 600, such as operations associated with display, telephone calls, data communication, camera operations, and recording operations. The processing component 602 may include one or more processors 620 to execute instructions to complete all or part of the steps of the above-described methods. In addition, the processing component 602 may include one or more modules to facilitate the interaction between the processing component 602 and other components. For example, the processing component 602 may include a multimedia module to facilitate the interaction between the multimedia component 608 and the processing component 602.

[0175] The memory 604 is configured to store various types of data to support the operation of the electronic device 600. Examples of such data include instructions for any application or method operating on the electronic device 600, contact data, phone book data, messages, pictures, videos, etc. The memory 604 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disks, optical disks, or graphene memory.

[0176] The power supply component 606 provides power to various components of the electronic device 600. The power supply component 606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device 600.

[0177] The multimedia component 608 includes a screen that provides an output interface between the electronic device 600 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, the multimedia component 608 includes a front camera and / or a rear camera. When the electronic device 600 is in an operating mode, such as a shooting mode or a video mode, the front camera and / or the rear camera can receive external multimedia data. Each of the front camera and the rear camera can be a fixed optical lens system or have a focal length and optical zoom capabilities.

[0178] The audio component 610 is configured to output and / or input audio signals. For example, the audio component 610 includes a microphone (MIC) that is configured to receive external audio signals when the electronic device 600 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals can be further stored in the memory 604 or transmitted via the communication component 616. In some embodiments, the audio component 610 further includes a speaker for outputting audio signals.

[0179] The I / O interface 612 provides an interface between the processing component 602 and a peripheral interface module, which can be a keyboard, a click wheel, buttons, etc. These buttons can include but are not limited to: a home button, a volume button, a power button, and a lock button.

[0180] The sensor component 614 includes one or more sensors for providing an assessment of the various aspects of the state of the electronic device 600. For example, the sensor component 614 can detect the on / off state of the electronic device 600, the relative positioning of components, such as the display and the keypad of the electronic device 600. The sensor component 614 can also detect a change in the position of the electronic device 600 or components of the electronic device 600, the presence or absence of user contact with the electronic device 600, the orientation or acceleration / deceleration of the device 600, and a change in the temperature of the electronic device 600. The sensor component 614 can include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor component 614 can also include a light sensor, such as a CMOS or a CCD image sensor, for use in imaging applications. In some embodiments, the sensor component 614 can further include an acceleration sensor, a gyro sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0181] The communication component 616 is configured to facilitate communication between the electronic device 600 and other devices in a wired or wireless manner. The electronic device 600 can access a communication standard-based wireless network, such as WiFi, a carrier network (such as 2G, 3G, 4G, or 5G), or a combination thereof. In an exemplary embodiment, the communication component 616 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 616 further includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra-Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0182] In an exemplary embodiment, the electronic device 600 can be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components for performing the above method.

[0183] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 604 including instructions, and the above instructions can be executed by a processor 620 of the electronic device 600 to complete the above method. For example, the computer-readable storage medium can be a ROM, Random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0184] In an exemplary embodiment, a computer program product is also provided, and the computer program product includes instructions that can be executed by a processor 620 of the electronic device 600 to complete the above method.

[0185] It should be noted that the above-mentioned device, electronic device, computer-readable storage medium, computer program product, etc. may also include other implementation manners according to the description of the method embodiments. The specific implementation manners can refer to the description of the relevant method embodiments and will not be elaborated here one by one.

[0186] Those skilled in the art will readily think of other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present disclosure are pointed out by the claims.

[0187] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure is only limited by the appended claims.

Claims

1. A method for generating a behavior prediction model, characterized in that: include: Based on the model parameters and energy function of the initial behavior prediction model, an energy model is constructed, and the gradient data of the energy model is calculated, wherein the initial behavior prediction model is obtained by training based on training data, wherein the training data includes click behavior sample data and click behavior labels corresponding to each click behavior sample data, wherein the click behavior sample data is used to characterize the recommended data in the recommendation scenario, and the click behavior labels are used to characterize the interactive operation of the target object on the recommended data; Based on the gradient data, random noise data is sampled a target number of times to obtain prior noise data, where the random noise data is click behavior data obtained by random sampling; Performing diffusion processing on the prior noise data through a diffusion model to obtain target sample data; Based on the target sample data and the training data, the initial behavior prediction model is trained to obtain a trained target behavior prediction model.

2. The method for generating a behavior prediction model according to claim 1, characterized in that: The energy model is constructed based on the model parameters and energy function of the initial behavior prediction model, including: Based on the network parameters of the initial behavior prediction model, the click behavior sample data, the click behavior labels corresponding to each of the click behavior sample data, and the relationship between the variables represented by the energy function, an energy model corresponding to the initial behavior prediction model is constructed.

3. The method for generating a behavior prediction model according to claim 2, characterized in that: The step of calculating the gradient data of the energy model comprises: Calculating target distribution between the click behavior sample data and the click behavior labels corresponding to the click behavior sample data through an energy model; Based on the target distribution, the gradient data of the energy model is calculated.

4. The method for generating a behavior prediction model according to claim 1, characterized in that: The step of sampling random noise data a target number of times based on the gradient data to obtain prior noise data includes: The random noise data is sampled a target number of times based on the gradient data, and the noise data corresponding to the target number of samplings is determined as the priori noise data.

5. The method for generating a behavior prediction model according to claim 4, characterized in that: The method further comprises: For the i-th sampling in the target number of times, based on the noise data corresponding to the i-1-th sampling, the sampling parameters, the gradient data corresponding to the noise data obtained by the i-1-th sampling, and the standard parameters, the noise data corresponding to the i-th sampling is calculated, where i is greater than or equal to 1 and i is a positive integer.

6. The method for generating a behavior prediction model according to claim 5, characterized in that: The calculating of the noise data corresponding to the i-th sampling based on the noise data corresponding to the i-th sampling, the sampling parameters, the gradient data corresponding to the noise data obtained by the i-th sampling, and the standard parameters includes: Based on the sampling parameters and the standard parameters, the noise data corresponding to the i-1th sampling and the gradient data corresponding to the noise data obtained by the i-1th sampling are adjusted to obtain an adjustment factor; The noise data corresponding to the i-1th sampling is adjusted by the adjustment factor to obtain the noise data corresponding to the i-th sampling.

7. The method for generating a behavior prediction model according to claim 1, characterized in that: The step of performing diffusion processing on the prior noise data through a diffusion model to obtain target sample data includes: The priori noise data is determined as the sampling starting point of the diffusion model, and based on the time step hyperparameter of the diffusion model, the priori noise data is sampled for the time step hyperparameter times to obtain target sample data.

8. The method for generating a behavior prediction model according to claim 1, characterized in that: The target sample data includes a plurality of click behavior data; the initial behavior prediction model is trained based on the target sample data and the training data to obtain a trained target behavior prediction model, including: Obtaining label data corresponding to each click behavior data included in the target sample data through the initial behavior prediction model; Based on the click behavior data contained in the target sample data, the label data corresponding to each click behavior data, and the training data, the initial behavior prediction model is trained to obtain a trained target behavior prediction model.

9. A device for generating a behavior prediction model, characterized in that: include: a construction unit configured to execute model parameters and an energy function based on an initial behavior prediction model, construct an energy model, and calculate gradient data of the energy model, wherein the initial behavior prediction model is obtained by training based on training data, wherein the training data includes click behavior sample data and click behavior labels corresponding to each click behavior sample data, wherein the click behavior sample data is used to characterize recommendation data in a recommendation scenario, and the click behavior labels are used to characterize an interactive operation of a target object on the recommendation data; A sampling unit is configured to perform a target number of samplings on random noise data based on the gradient data to obtain prior noise data; A diffusion unit is configured to perform diffusion processing on the prior noise data through a diffusion model to obtain target sample data; The training unit is configured to train the initial behavior prediction model based on the target sample data and the training data to obtain a trained target behavior prediction model.

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the method for generating a behavior prediction model as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method for generating a behavior prediction model as described in any one of claims 1 to 8.

12. A computer program product, comprising instructions, characterized in that: When the instruction is executed by a processor of an electronic device, the electronic device is enabled to execute the method for generating a behavior prediction model as described in any one of claims 1 to 8.

Images