Model training method, click rate determination method and related device

Abstract

Embodiments of the present disclosure provide a model training method, a click rate determination method and related equipment, which belong to the technical field of computers. The method comprises: obtaining a training sample set, the training sample set comprising training samples in multiple business scenarios and corresponding sample labels; inputting the training samples into a first prediction model and a second prediction model respectively, and migrating intermediate data of the first prediction model to the second prediction model to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; and adjusting model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample labels. The present disclosure can alleviate the problems of difficult model convergence and low prediction accuracy in scenarios with partial data sparsity, solve the problem of difficult maintenance of single-domain models, and solve the seesaw effect caused by modeling based on a mixed data set of multiple scenarios.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of computer technology, and more specifically, embodiments of the present disclosure relate to a model training method, a click rate determination method and related equipment. BACKGROUND

[0002] This section is intended to provide background or context to the embodiments of the disclosure recited in the claims. The description herein does not constitute admission that the prior art is prior art nor does it constitute an admission of any description in this section as prior art to an application.

[0003] With the popularity of the Internet, more and more users use Internet products such as web pages and APPs to obtain information. However, the Internet is full of a large amount of information, and how to quickly provide users with the information they need has become a major means to improve user satisfaction.

[0004] In related technologies, the prediction result of a click rate estimation model is usually used to recommend information to users. However, most of the existing technologies use interaction data of each scene to train a model in a single scene, which makes it difficult for the model to converge and reduces the prediction accuracy for scenes with sparse data. In addition, the number of models increases with the number of scenes, making it difficult to maintain the models and consuming a large amount of resources. The model training process based on a mixed data set of multiple scenes often has a seesaw effect, in which the performance of some scene models improves while the performance of some other scene models decreases. SUMMARY

[0005] Embodiments of the present disclosure provide a model training method, a click rate determination method and related equipment.

[0006] In a first aspect of the embodiments of the present disclosure, a model training method is provided, which includes: obtaining a training sample set, the training sample set including training samples in multiple business scenes and corresponding sample labels; inputting the training samples into a first prediction model and a second prediction model respectively, and migrating intermediate data of the first prediction model to the second prediction model to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; and adjusting model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample labels.

[0007] Optionally, the first prediction model includes scene models corresponding to multiple business scenes, and the training samples include sample scene features. The method further includes: determining, according to the sample scene features, a scene model corresponding to the training sample in the first prediction model; and inputting the training samples into the corresponding scene models and the second prediction model respectively.

[0008] Optionally, the obtaining the first training click rate comprises: based on first standardization processing, performing forward processing on the training sample by the first prediction model to obtain the first training click rate, wherein a standardization parameter of the first standardization processing is determined based on a training sample in a business scenario corresponding to the scenario model.

[0009] Optionally, the obtaining the second training click rate comprises: based on second standardization processing, performing forward processing on the intermediate data and the training sample by the second prediction model to obtain the second training click rate, wherein a standardization parameter of the second standardization processing is determined based on training samples in multiple business scenarios input into the second prediction model, and the training samples corresponding to the multiple business scenarios have feature correlations.

[0010] Optionally, the adjusting the model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample label comprises: determining a first loss function according to the first training click rate and the sample label; determining a second loss function according to the second training click rate and the sample label; determining a target loss function according to the first loss function and the second loss function; and adjusting the model parameters of the first prediction model and the second prediction model according to the target loss function.

[0011] Optionally, the migrating the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model by an adaptation network; and the adjusting the model parameters of the first prediction model and the second prediction model according to the target loss function comprises: adjusting the model parameters of the first prediction model, the second prediction model and the adaptation network by minimizing the target loss function and gradient back propagation.

[0012] In a second aspect of the embodiments of the present application, a click rate determination method is provided, which comprises: determining a target scenario model in a first prediction model according to a scenario feature of to-be-predicted data, wherein the first prediction model comprises scenario models for multiple business scenarios; inputting the to-be-predicted data into the target scenario model and a second prediction model respectively; migrating intermediate data of the target scenario model to the second prediction model to obtain a prediction output of the second prediction model; and determining a click rate of the to-be-predicted data according to the prediction output of the second prediction model.

[0013] Optionally, the migrating the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network, the adaptation network being configured to filter the migrated data based on an importance of the intermediate data.

[0014] Optionally, the target scenario model comprises at least one first hidden layer, and the second prediction model comprises at least one second hidden layer, and the obtaining the prediction output of the second prediction model comprises: performing first linear processing and first nonlinear conversion on input data layer by layer through each of the first hidden layers; performing second linear processing and second nonlinear conversion on the intermediate data migrated from each of the first hidden layers and the input data layer by layer through each of the second hidden layers; and performing nonlinear normalization processing on an output of the last one of the second hidden layers to obtain the prediction output.

[0015] Optionally, the target scenario model further comprises a first normalization layer, and the second prediction model further comprises a second normalization layer, and before the input data is input to the first hidden layer and the second hidden layer, the method further comprises: performing first normalization processing on the input data through the first normalization layer; and performing second normalization processing on the input data through the second normalization layer.

[0016] Optionally, the to-be-predicted data comprises user feature data, content feature data and association feature data, and before the to-be-predicted data is input to the target scenario model and the second prediction model respectively, the method further comprises: performing vector representation on the user feature data, the content feature data and the association feature data respectively by using a pre-trained embedding model, to correspondingly obtain a user feature vector, a content feature vector and an association feature vector; and the inputting the to-be-predicted data to the target scenario model and the second prediction model respectively comprises: inputting the user feature vector, the content feature vector and the association feature vector after being spliced to the target scenario model and the second prediction model respectively.

[0017] Optionally, the method further comprises: determining target recommendation information according to a click rate of the to-be-predicted data; and pushing the target recommendation information to a user.

[0018] In a third aspect of the embodiments of the present application, a model training apparatus is provided, and the apparatus comprises: an acquisition module configured to acquire a training sample set comprising training samples in multiple business scenarios and corresponding sample labels; a training module configured to input the training samples into a first prediction model and a second prediction model respectively, and migrate intermediate data of the first prediction model to the second prediction model, to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; and an adjustment module configured to adjust model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample labels.

[0019] Optionally, the first prediction model comprises scenario models corresponding to the multiple business scenarios, and the training samples comprise sample scenario features, and the apparatus further comprises a scenario model determination module configured to determine, in the first prediction model, a scenario model corresponding to the training sample according to the sample scenario features; and the training module is further configured to input the training sample into the corresponding scenario model and the second prediction model respectively.

[0020] Optionally, the training module is further configured to perform forward processing on the training sample by the first prediction model based on first standardization processing to obtain the first training click rate, and a standardization parameter of the first standardization processing is determined based on training samples in a business scenario corresponding to the scenario model.

[0021] Optionally, the training module is further configured to perform forward processing on the intermediate data and the training sample by the second prediction model based on second standardization processing to obtain the second training click rate, and a standardization parameter of the second standardization processing is determined based on training samples in multiple business scenarios input into the second prediction model, and the training samples in the multiple business scenarios have feature correlations.

[0022] Optionally, the adjustment module is further configured to determine a first loss function according to the first training click rate and the sample labels, to determine a second loss function according to the second training click rate and the sample labels, to determine a target loss function according to the first loss function and the second loss function, and to adjust the model parameters of the first prediction model and the second prediction model according to the target loss function.

[0023] Optionally, the training module is further configured to migrate the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network; and the adjusting module is further configured to adjust the model parameters of the first prediction model, the second prediction model and the adaptation network by minimizing the target loss function and gradient back propagation.

[0024] In a fourth aspect of the embodiments of the present application, a click rate determination apparatus is provided, which comprises: a scene model determination module configured to determine a target scene model in a first prediction model according to a scene feature of to-be-predicted data, the first prediction model comprising scene models for a plurality of business scenes; an input module configured to input the to-be-predicted data into the target scene model and a second prediction model respectively; a migration module configured to migrate intermediate data of the target scene model to the second prediction model to obtain a prediction output of the second prediction model; and a click rate determination module configured to determine a click rate of the to-be-predicted data according to the prediction output of the second prediction model.

[0025] Optionally, the migration module is further configured to migrate the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network, the adaptation network being used for screening the migrated data based on an importance of the intermediate data.

[0026] Optionally, the target scene model comprises at least one first hidden layer, the second prediction model comprises at least one second hidden layer, and the apparatus further comprises a prediction module configured to: perform first linear processing and first nonlinear conversion on input data layer by layer through each of the first hidden layers; perform second linear processing and second nonlinear conversion on the intermediate data migrated from each of the first hidden layers and the input data layer by layer through each of the second hidden layers; and perform nonlinear normalization processing on an output of the last one of the second hidden layers to obtain the prediction output.

[0027] Optionally, the target scene model further comprises a first normalization layer, the second prediction model further comprises a second normalization layer, and the apparatus further comprises a normalization module configured to, before the input data is input into the first hidden layers and the second hidden layers: perform first normalization processing on the input data through the first normalization layer; and perform second normalization processing on the input data through the second normalization layer.

[0028] Optionally, the to-be-predicted data comprises user feature data, content feature data and association feature data, and before the to-be-predicted data is input into the target scene model and the second prediction model respectively, the apparatus further comprises an embedding module configured to: perform vector representation on the user feature data, the content feature data and the association feature data respectively by using a pre-trained embedding model, to obtain a user feature vector, a content feature vector and an association feature vector correspondingly; and the input module is further configured to: input the user feature vector, the content feature vector and the association feature vector after splicing into the target scene model and the second prediction model respectively.

[0029] Optionally, the apparatus further comprises a pushing module configured to determine target recommendation information according to a click rate of the to-be-predicted data, and push the target recommendation information to a user.

[0030] In a fifth aspect of the embodiments of the present application, a storage medium is provided, which has stored thereon a program, and the program is executed by a processor to implement the method described in the above embodiments.

[0031] In a sixth aspect of the embodiments of the present application, an electronic device is provided, which comprises a processor and a memory, the memory stores executable instructions, and the processor is configured to invoke the executable instructions stored in the memory to execute the method described in the above embodiments.

[0032] According to the model training method provided by the embodiments of the present application, on the one hand, by inputting the training samples under multiple business scenes into the first prediction model and the second prediction model respectively, and adjusting the model parameters of the first prediction model and the second prediction model according to the first training click rate output by the first prediction model, the second training click rate output by the second prediction model and the sample label, the joint training of the first prediction model and the second prediction model is realized, which can alleviate the problems of model convergence difficulty and low estimation accuracy caused by sample sparsity in some business scenes. On the other hand, through the joint training of the model under multiple business scenes, the problems of difficult maintenance and large resource consumption of the multi-scene click rate estimation model are solved. Further, by migrating the intermediate data of the first prediction model to the second prediction model, the common features between different scenes are modeled, and at the same time, the characteristics within each business scene are trained by the first prediction model, which can avoid the seesaw effect of the model and fully utilize the data of all scenes to improve the prediction accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0033] The above and other objects, features and advantages of the exemplary embodiments of the present application will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0034] FIG. 1 Fig. 1 schematically shows one of the model training method flowcharts according to one embodiment of the present disclosure.

[0035] FIG. 2 Fig. 2 schematically shows another of the model training method flowcharts according to one embodiment of the present disclosure.

[0036] FIG. 3 Fig. 3 schematically shows the click rate determination method flowchart according to one embodiment of the present disclosure.

[0037] FIG. 4 Fig. 4 schematically shows the click rate prediction process of the trained model according to one embodiment of the present disclosure.

[0038] FIG. 5 Fig. 5 schematically shows the model training device structure block diagram according to one embodiment of the present disclosure.

[0039] FIG. 6 Fig. 6 schematically shows the click rate determination device structure block diagram according to one embodiment of the present disclosure.

[0040] FIG. 7 Fig. 7 schematically shows the structure diagram of an electronic device suitable for implementing the embodiments of the present disclosure.

[0041] In the drawings, the same or corresponding reference numbers indicate the same or corresponding parts. DETAILED DESCRIPTION

[0042] The principles and spirits of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present disclosure, and do not limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0043] Those skilled in the art know that the embodiments of the present disclosure can be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure can be embodied in the form of a complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0044] The data involved in the present disclosure can be data authorized by users or sufficiently authorized by all parties, and the collection, dissemination, use, etc. of the data comply with the requirements of relevant national laws and regulations, and the embodiments / examples of the present disclosure can be combined with each other.

[0045] According to an embodiment of the present disclosure, a model training method, a click rate determination method, and related equipment are provided. SUMMARY

[0047] Click-through rate: also known as Click-Through-Rate (CTR), is a commonly used term in Internet advertising, which refers to the click-through rate of network advertising (picture advertising / text advertising / keyword advertising / ranking advertising / video advertising, etc.), that is, the actual number of clicks of the advertisement (which can be the number of accesses to the target page) divided by the number of exposures of the advertisement.

[0048] Transfer learning: also known as Transfer Learning, is a machine learning method, which refers to using a model developed for task A as an initial point to develop a model for task B. Transfer learning can transfer learned powerful skills to related problems to save computing overhead.

[0049] Single-domain click-through rate estimation: also known as Single-Domain CTR Prediction, refers to using a model trained by user-object interaction data of a single domain to estimate the click rate.

[0050] Mixed multi-domain data training shared model: refers to mixing user-object interaction data of multiple domains as a data set, and training a shared model based on the mixed data set for CTR estimation of multiple domains.

[0051] In the related art, before the click rate estimation, a multi-path recall is usually used to recall a candidate set of items that the user may be interested in from a large number of item sets to improve the calculation efficiency. Then a CTR estimation model of a sorting layer is applied to estimate a CTR score for each item in the candidate set. Finally, a final recommendation list is obtained through reordering. As can be seen from the above process, the sorting layer is a core module in the recommendation system structure, and therefore the CTR estimation model in the sorting layer is crucial to the recommendation system.

[0052] In a large content platform (website / application), there are usually many scenarios (domains) that need to use CTR estimation to improve user satisfaction. Since different domains have overlapping users and objects, there is a certain commonality between different domains. Information sharing between different domains is beneficial to the learning of the CTR estimation model of each domain. However, different domains have different user behaviors and unique data distribution, and the model training process based on a mixed data set of different domains may lead to a seesaw effect of performance improvement in some domains and performance decline in some domains.

[0053] To solve the above problems, the application idea of the application is that by inputting training samples containing multiple scenes into a first prediction model and a second prediction model respectively, the characteristics in each business scene can be trained by the first prediction model, and the common features between different scenes can be trained by the second prediction model, thereby realizing the joint training of the scene model (the first prediction model) and the shared model (the second prediction model), and the problems of model convergence difficulty and low prediction accuracy caused by sparse scene data can be solved. By migrating the intermediate data of the first prediction model to the second prediction model and migrating the features related between scenes to the second prediction model, the seesaw effect caused by the equal treatment of scene data is avoided, thereby improving the model prediction accuracy of all scenes.

[0054] The method and device of the embodiments of the present disclosure can be applied to at least one of electronic devices including but not limited to a server, a terminal, and the like, which can be configured to perform the method provided by the embodiments of the present disclosure. In other words, the model training method or the click rate determination method can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to a single server, a server cluster, a cloud server, or a cloud server cluster, and the like. The present disclosure takes the server as an example for description.

[0055] Exemplary methods

[0056] The preferred embodiments of the present disclosure are described below in conjunction with the accompanying drawings of the specification, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present disclosure, and are not used to limit the present disclosure, and the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0057] The model training method according to the exemplary embodiments of the present disclosure is described below with reference to FIG. 1 The model training method according to the exemplary embodiments of the present disclosure can include steps S110-S130.

[0058] In step S110, a training sample set is obtained, and the training sample set includes training samples in multiple business scenes and corresponding sample labels.

[0059] In step S120, the training samples are input into a first prediction model and a second prediction model respectively, and intermediate data of the first prediction model is migrated to the second prediction model, to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model.

[0060] In step S130, the model parameters of the first prediction model and the second prediction model are adjusted according to the first training click rate, the second training click rate, and the sample label.

[0061] The model training method provided by the embodiments of the present disclosure, on the one hand, by inputting the training samples in multiple business scenarios into the first prediction model and the second prediction model respectively, and adjusting the model parameters of the first prediction model and the second prediction model according to the first training click rate output by the first prediction model, the second training click rate output by the second prediction model and the sample label, the joint training of the first prediction model and the second prediction model is realized, which can alleviate the problems of model convergence difficulty and low estimation accuracy caused by sample sparsity in some business scenarios. On the other hand, through the joint training of the model in multiple business scenarios, the problems of difficult maintenance of the multi-scene click rate estimation model and large resource consumption are solved. Further, by migrating the intermediate data of the first prediction model to the second prediction model, the common features between different scenarios are modeled, and at the same time, the characteristics within each business scenario are trained through the first prediction model, which can avoid the seesaw effect of the model and fully utilize the data of all scenarios to improve the prediction accuracy of the model.

[0062] The above steps are described in detail below.

[0063] In step S110, a training sample set is obtained, and the training sample set includes training samples in multiple business scenarios and corresponding sample labels.

[0064] In the example embodiment, the recall layer (Matching) can be used to quickly reduce the massive candidate set to a scale of several thousand to several hundred, so that the items of interest to the user are quickly filtered out to form the training sample set.

[0065] In the example embodiment, the multiple business scenarios refer to multiple business scenarios in which the user interaction data of each scenario has an association relationship, for example, the business scenarios can be the business scenarios corresponding to various application platforms such as music platforms, news platforms and video platforms, or different business functions in one application platform, which are not limited in the example.

[0066] In the example embodiment, each training sample can include user interaction data, sample scene features, and sample labels. The user interaction data can be user interaction data within a specified time period, such as user interaction data within the last month or user interaction data within the same period in the last few years. The user interaction data refers to various data generated by the user's interaction behavior in the application platform, which can include user data and content data, etc. The user data can include user behavior sequence data (such as user's click, play, collect, purchase, browse, etc. related information of content) and user attribute data (such as user account / ID, occupation, etc. basic information), and the content data refers to related information of the user interaction object, such as object name, number, category, style, etc. information. The object can include goods, video files, audio files, text, etc. in the platform, which is not limited in the example. The sample scene features refer to features used to indicate the business scene to which the training sample belongs, and different business scenes can correspond to different values. The sample label refers to the click probability of the user on the recommended object, and the value of 1 indicates that the user clicks, and the value of 0 indicates that the user does not click.

[0067] In step S120, the training samples are input into the first prediction model and the second prediction model respectively, and the intermediate data of the first prediction model is migrated to the second prediction model to obtain the first training click rate output by the first prediction model and the second training click rate output by the second prediction model.

[0068] In the example embodiment, the first prediction model is used to learn the characteristics in each business scene, i.e. domain-specific towers, as shown in the following formula: FIG. 2 As shown in the formula, the first prediction model can include a plurality of parallel scene models (such as scene model 1, scene model 2, …, scene model N), and each scene model corresponds to learning the interaction characteristics in one business scene. Each scene model can be one of a feedforward neural network or various deep neural networks (DeepFM, Wide&Deep, Deep Interest Network DIN, etc.). For example, the scene model is a feedforward neural network (FNN), which is a kind of artificial neural network. The network structure is a one-way multi-layer structure, in which each layer contains a plurality of neurons, each neuron can receive data from the previous layer of neurons, process the data to generate new data and output to the next layer. The 0th layer is the input layer, the last layer is the output layer, and the other intermediate layers are called hidden layers. The hidden layer can be one layer or multiple layers (such as K layers).

[0069] In the example embodiment, the second prediction model is used to learn the common characteristics between each business scene, i.e. shared towers. The network structure of the second prediction model can be the same as that of a scene model, or it can be different, which is not limited in the example.

[0070] In the example embodiment, one-way data migration can be performed between the intermediate layers of the scenario models and the intermediate layers of the second prediction model to migrate the intermediate data of the scenario models to the second prediction model. The number of network layers of the scenario models and the second prediction model can be the same or different, which is not limited in the example. For example, the output data of each intermediate layer of the scenario model can be migrated to the corresponding intermediate layer of the second prediction model for data fusion, and then input to the next layer of the second prediction model, for example, as shown in FIG. 1B, the output of the first hidden layer 1 can be migrated to the input of the second hidden layer 2 through the adapter 1. The output data of multiple intermediate layers of the scenario model can also be migrated to the same layer of the second prediction model, or the output of one intermediate layer of the scenario model can be migrated to multiple layers of the second prediction model, which is not limited in the example. FIG. 2

[0071] For example, according to the sample scenario characteristics, the scenario model corresponding to the training sample can be determined in the first prediction model, and the training sample is input into the corresponding scenario model.

[0072] In the example embodiment, the first prediction model can include scenario models corresponding to multiple business scenarios. As shown in FIG. 1A, the first prediction model can include scenario models corresponding to multiple business scenarios. FIG. 2 For example, as shown in FIG. 1A, in the first prediction model, for the input training sample, according to the sample scenario characteristics of each training sample, the training sample is input into the corresponding scenario model to realize separate modeling of each scenario training data, and each scenario model learns the characteristics of the training sample corresponding to the scenario. The sample scenario characteristics can be automatically labeled by the system or manually labeled, which is not limited in the example.

[0073] For example, the first training click rate can be obtained by the following steps.

[0074] Based on the first standardization processing, the first prediction model is used to perform forward processing on the training sample to obtain the first training click rate.

[0075] In the example embodiment, the standardization parameter of the first standardization processing is determined based on the training sample in the business scenario corresponding to the scenario model. The first standardization processing can be batch standardization processing, and the standardization parameter can be a statistical quantity of the training sample in the scenario, such as mean, variance, etc. The first standardization processing of the scenario p can be represented as:

[0076]

[0077] In the formula, x represents the x (p) data after the first standardization processing, x (p) represents the training sample of the pth business scenario (domain), μ (p) and (σ (p) ) 2 ​The mean and variance of the training sample representing the pth business scenario (domain), a (p) and b (p) The training parameter of the pth business scenario (domain), respectively, the learnable parameter, and is a constant, used to prevent the denominator from being zero.

[0078] In the example embodiment, as shown in FIG. 2 , the first standardization processing can be performed in the first standardization layer and each hidden layer of the first prediction model, so as to keep the numerical value processed by the model within a certain range and reduce the calculation complexity. For example, in each hidden layer, the input data of the layer can be subjected to nonlinear transformation processing and then subjected to corresponding first standardization processing, such as using a PRelu activation function and corresponding first standardization processing.

[0079] Exemplarily, the second training click rate can be obtained through the following steps.

[0080] Based on the second standardization processing, the intermediate data and the training sample are subjected to forward processing by the second prediction model to obtain the second training click rate.

[0081] In the example embodiment, the standardization parameter of the second standardization processing is determined based on the training sample under the plurality of business scenarios input into the second prediction model, and the training samples corresponding to the plurality of business scenarios have feature correlation. The second standardization processing can be batch standardization processing, and for the shared tower, each training sample is equally important, so the standardization parameter of the second standardization processing is based on global statistics. The second standardization processing can be represented as:

[0082]

[0083] In the formula, x represents the data of x after the second standardization processing, x represents the input data, and a and b 2 respectively represent the mean and variance of all samples in the training sample set, a and b are learnable parameters of the shared tower, and is a constant, used to prevent the denominator from being zero.

[0084] In the example embodiment, similarly to the scenario model, as shown in FIG. 2 , the second standardization processing can be performed in the second standardization layer and each hidden layer of the second prediction model, such as using a PRelu activation function and second standardization processing.

[0085] In the above embodiments, batch normalization, also referred to as BN, is a method for making artificial neural network training faster and more stable by re-centering and re-scaling the hidden layer output. The premise of batch normalization is that the input features x are independent and identically distributed. For a multi-scene training sample set, each scene (domain) has a different data distribution, so batch normalization can be performed for each domain to avoid the problem of ignoring the statistical differences between different domains and affecting the final model performance when using global BN.

[0086] In step S130, the model parameters of the first prediction model and the second prediction model are adjusted according to the first training click rate, the second training click rate, and the sample label.

[0087] In the present example embodiment, the first prediction model can be trained based on the first training click rate (training output of the first prediction model) and the sample label. The second prediction model is trained based on the second training click rate (training output of the second prediction model) and the sample label.

[0088] Exemplarily, the model parameters of the first prediction model and the second prediction model can be adjusted according to the following steps.

[0089] According to the first training click rate and the sample label, a first loss function is determined; according to the second training click rate and the sample label, a second loss function is determined; according to the first loss function and the second loss function, a target loss function is determined; and according to the target loss function, the model parameters of the first prediction model and the second prediction model are adjusted.

[0090] In the present example embodiment, for the first prediction model (domain-specific tower), a loss function corresponding to the first prediction model (first loss function) can be determined based on the output of each scene model and the corresponding sample label. For the second prediction model (shared tower), a loss function of the second prediction model (second loss function) can be determined based on its output and the sample label. The first loss function and the second loss function can be set with a weight coefficient according to actual conditions to determine a target loss function of the entire model, and the model parameters of the first prediction model and the second prediction model are adjusted based on the global target loss function. The calculation of each loss function can use a commonly used loss function calculation method, such as a cross-entropy loss function.

[0091] For example, the calculation process of each loss function is as follows:

[0092]

[0093]

[0094]

[0095] where N denotes the total number of domains, denotes the final training loss denotes the second loss function, denotes the first loss function, denotes the target loss function, i denotes the sample serial number, M p denotes the total number of samples in the p-th domain, denotes the sample label, denotes the training output of the p-th domain corresponding scene model, denotes the training output of the second prediction model, λ1 and λ2 are hyperparameters used to control the intensity of the domain loss and the L2 regularization term, Θ denotes all parameters of the model, λ1 can take a value of about 0.5, and λ2 can take a value of about 0.001.

[0096] In some embodiments, as shown in FIG. 1, the intermediate data of the first prediction model can be migrated to the intermediate network layer of the second prediction model through an adaptation network, and the model parameters of the first prediction model, the second prediction model and the adaptation network are adjusted by minimizing the target loss function and gradient backpropagation. FIG. 2

[0097] In the present example embodiment, in the forward calculation process, the intermediate data of the first prediction model can be migrated to the intermediate network layer of the second prediction model through the adaptation network. In the gradient backpropagation process, the gradient propagation from the second prediction model to the first prediction model is stopped, i.e., the gradient propagation between the second prediction model and the first prediction model is not performed, such as the gradient stop in FIG. 2. In this way, the influence of the gradient backpropagation of the shared tower on the learning of the unique tower can be prevented, and the performance of the unique tower model can be ensured. FIG. 2

[0098] In the present example embodiment, the adaptation network can be various gating networks or knowledge migration networks, which are not limited in the present example. For example, the adaptation network can include a gating network in each migration route, and each gating network can include a migration threshold. When the intermediate data exceeds the migration threshold, the intermediate data is migrated. Through model training, the model parameters of the adaptation network are adjusted to enable the adaptation network to learn the migration characteristics of each domain, so as to accurately migrate the domain knowledge in the prediction process.

[0099] ​​The model training method of the present application models the inter-domain sharing and the intra-domain characteristics by constructing a domain-shared tower, referred to as a shared tower, and constructing a domain-specific tower, referred to as a specific tower, for each domain. Considering that each domain has different data distribution, a domain-specific standardization strategy (Domain-Specific Batch Normalization, DSBN) is adopted in the specific tower, i.e., the first standardization processing. Further, in order to inject the intra-domain characteristics to affect the final CTR prediction score, an adapter network is designed to inject the knowledge of the specific tower into the shared tower. And in order to prevent the shared tower gradient from affecting the learning of the specific tower, the gradient is stopped at the adapter.

[0100] The above is the training process of the click rate prediction model, which can be performed offline, and the trained model can be used for online click rate prediction.

[0101] The click rate determination method according to the exemplary embodiments of the present disclosure will be described below with reference to FIG. 3 which can be applied to various click rate real-time prediction scenarios and can include steps S310-S340.

[0102] In step S310, a target scene model is determined in a first prediction model according to the scene characteristics of the data to be predicted, and the first prediction model includes scene models for multiple business scenes.

[0103] In step S320, the data to be predicted is input into the target scene model and a second prediction model respectively.

[0104] In step S330, the intermediate data of the target scene model is migrated to the second prediction model to obtain the prediction output of the second prediction model.

[0105] In step S340, the click rate of the data to be predicted is determined according to the prediction output of the second prediction model.

[0106] In the click rate determination method provided by the present exemplary embodiments, the target scene model, the first prediction model and the second prediction model are all trained models, the inter-domain shared features can be extracted through the second prediction model, the characteristics within the scene can be extracted through the target scene model in the first prediction model, and the scene characteristics and shared features can be combined by migrating the intermediate data of the target scene model to the second prediction model, which can model the inter-domain commonality while preserving the intra-domain characteristics, fully utilizes the data of each domain, and can improve the CTR prediction accuracy of each domain.

[0107] The above steps S310-S340 will be described in detail below.

[0108] In step S310, a target scene model is determined in a first prediction model according to a scene feature of the data to be predicted, the first prediction model including scene models for a plurality of business scenes.

[0109] In the example embodiment, the data to be predicted can include user feature data (such as a recent user behavior sequence, user basic data such as a user ID, a user name, etc.), content feature data (related features of objects clicked by a user recently, such as related features of commodities, videos, audios, texts, etc.), and association feature data (such as context features or association object information of clicked objects). The scene feature can be determined according to a current prediction scene. The target scene model is a trained scene model corresponding to the scene feature.

[0110] In the example embodiment, the scene feature has a one-to-one correspondence with the scene model, and the corresponding scene model can be determined through the scene feature. For example, the scene feature can be represented by natural numbers 1, 2, …, N, and each natural number corresponds to a scene model. As shown in FIG. 1, the first prediction model can include scene model 1, …, scene model p, …, scene model N, p and N being natural numbers. FIG. 4

[0111] In step S320, the data to be predicted is input into the target scene model and a second prediction model respectively.

[0112] In the example embodiment, the target scene model is used to extract the intra-domain characteristics of the data to be predicted, and the second prediction model is used to extract the common characteristics of the data to be predicted between different domains, and the second prediction model is a trained model.

[0113] In some embodiments, the data to be predicted can be vectorized before being input into the model. For example, a pre-trained embedding model can be used to vectorize the user feature data, the content feature data, and the association feature data respectively, to obtain a user feature vector, a content feature vector, and an association feature vector.

[0114] In the example embodiment, the pre-trained embedding model can be implemented through a representation embedding layer, i.e., embedding, which can map high-dimensional sparse features to low-dimensional dense features. By using a low-dimensional, dense, and continuous vector to represent an object, the embedding vector can represent some features of the object, and the similarity or relationship between objects can be reflected by the distance between vectors.

[0115] Then, the user feature vector, the content feature vector, and the association feature vector are spliced and input into the target scene model and the second prediction model respectively.

[0116] ​In the example embodiment, the user feature vector, the content feature vector and the association feature vector can be vector spliced in a specified order, and the spliced vector is input as input data into the target scenario model and the second prediction model respectively.

[0117] Exemplarily, the to-be-predicted data after embedding representation can be represented as:

[0118] x1={E(u1),E(u2),…,E(u2);E(v1),E(v2),…,E(v m );E(c1),E(c2),…,E(c k )}

[0119] Wherein, (u1, u2, …, u2) represents user features, (v1, v2, …, v m ) represents object features, and (c1, c2, …, c k ) represents context features, E(·) represents a pre-trained embedding model, i.e., a representation embedding layer, and n, m, k are natural numbers.

[0120] In step S330, the intermediate data of the target scenario model is migrated to the second prediction model to obtain a prediction output of the second prediction model.

[0121] In the example embodiment, the intermediate data of the first prediction model can be migrated to the intermediate network layer of the second prediction model through an adaptation network to migrate the intra-domain characteristics extracted by the target scenario model to the second prediction model. The adaptation network is a trained network, and through the trained adaptation network (such as FIG. 4 As shown, adapter 1, …, adapter K-1), the migration data can be filtered based on the importance of the intermediate data. That is, the important data in the domain can be migrated to the second prediction model for click rate prediction through the adaptation network.

[0122] For example, as shown in FIG. 4 The target scenario model can include a first normalization layer, at least one first hidden layer (such as first hidden layer 1, first hidden layer 2, …, first hidden layer K), and the second prediction model can include a second normalization layer, at least one second hidden layer (such as second hidden layer 1, second hidden layer 2, …, second hidden layer K). The processing process of the click rate prediction model on the input data can include the following steps.

[0123] The input data is first normalized by the first normalization layer; and the input data is second normalized by the second normalization layer.

[0124] In the example embodiment, the first and second standardization processes are batch standardization processes, and the input data can be batch-standardized by the model parameters of the standardization process trained by the first and second prediction models, respectively, to control the data scale. The specific processing process can refer to the model training method described above.

[0125] The input data is subjected to first linear processing and first non-linear conversion by each first hidden layer layer by layer.

[0126] In the example embodiment, each first hidden layer can include linear mapping and non-linear conversion of data, for example, as shown in FIG. 4 The target scene model can include K layers of hidden layers, the PRelu activation function and the first standardization process strategy are used from the first to the (K-1) layer, the Sigmoid activation function is used for the output of the last layer, and the final output is the CTR estimation result of the specific domain p, and the specific expression is as follows:

[0127]

[0128] In the formula, and represent the training weights and biases of the target scene model (domain p), represents the output of the first hidden layer of the first layer of the target scene model, represents the final CTR estimation score of the target scene model (domain p).

[0129] The intermediate data migrated from each first hidden layer and the input data are subjected to second linear processing and second non-linear conversion by each second hidden layer layer by layer.

[0130] In the example embodiment, the intermediate data migrated from each first hidden layer can include the output of one or more first hidden layers, such as one or more of In the formula above. For example, as shown in FIG. 4 The target scene model and the second prediction model can have the same number of hidden layers, and the output of each first hidden layer is migrated to the corresponding second hidden layer of the second prediction model through an adapter network, so that the output of the migrated data of the previous first hidden layer and the output of the second hidden layer are jointly input into the next second hidden layer for processing. For example, the output of the first hidden layer of the target scene model is migrated to the same position of the second prediction model through the adapter 1. Each migration path corresponds to an adapter to achieve selective data migration.

[0131] The output of the last second hidden layer is subjected to non-linear normalization processing to obtain the prediction output.

[0132] In the example embodiment, the nonlinear normalization processing is used to normalize map the hidden layer output to the output dimension to obtain the click probability of the to-be-predicted data, for example, the nonlinear normalization processing can be performed by using a Sigmoid activation function.

[0133] For example, as shown in FIG. 4 The shared tower and the unique tower structure are basically the same, and also have K layers of the second hidden layer, the first K-1 layers use the PRelu activation function and the second standardization processing, the last layer output uses the Sigmoid activation, and the final output is the CTR estimation result, and the calculation expression is as follows:

[0134]

[0135]

[0136] In the formula, W1, W2, …, W K and b1, b2, …, b K respectively represent the training weights and biases of the second prediction model (shared tower), g(·) represents an adapter function, and respectively represent parameters in the adapter function for controlling the domain p knowledge transfer, ⊙ represents element multiplication operation, SG represents a stop gradient backpropagation function, z1 represents the output of the first layer hidden layer of the second prediction model, σ(·) represents a gating function, represents the final prediction output.

[0137] In the above embodiment, the features of the target scenario model (unique tower) are transferred to the second prediction model (shared tower) through the adaptation network (adapter 1, adapter 2, …, adapter K-1), the inter-domain commonality is extracted while the intra-domain characteristics are retained, and the CTR estimation accuracy of all domains is improved.

[0138] In step S340, the click rate of the to-be-predicted data is determined according to the prediction output of the second prediction model.

[0139] In the example embodiment, the prediction output of the second prediction model can be a probability vector, and the probability value in the probability vector is taken as the click probability of the object.

[0140] In some embodiments, the method further comprises: determining target recommendation information according to the click rate of the to-be-predicted data; and pushing the target recommendation information to the user.

[0141] In the example embodiment, the object (goods, video, audio, text, etc.) with a higher click rate can be taken as the target recommendation information, for example, the object with a click rate exceeding a click rate threshold can be taken as the target recommendation information, or the objects can be sorted according to the click rate to form a recommendation list and pushed to the user, and the example is not limited in this regard.

[0142] The following introduces a specific process of a click rate determination method of an embodiment of the present application.

[0143] In a first step, a training sample set is obtained, which can include training samples in multiple business scenarios and corresponding sample labels.

[0144] In a second step, the training samples are input into a first prediction model and a second prediction model respectively, and intermediate data of the first prediction model is migrated to the second prediction model, to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model.

[0145] In a third step, model parameters of the first prediction model and the second prediction model are adjusted according to the first training click rate, the second training click rate and the sample labels, to obtain trained first and second prediction models.

[0146] In a fourth step, a target scenario model is determined in the trained first prediction model according to a scenario feature of to-be-predicted data.

[0147] In a fifth step, the to-be-predicted data are input into the target scenario model and the trained second prediction model respectively.

[0148] In a sixth step, intermediate data of the target scenario model is migrated to the trained second prediction model, to obtain a prediction output of the second prediction model.

[0149] In a seventh step, a click rate of the to-be-predicted data is determined according to the prediction output of the second prediction model.

[0150] In an eighth step, target recommendation information is determined according to the click rate of the to-be-predicted data, and the target recommendation information is pushed to a user.

[0151] The specific details of the above embodiments have been described in detail in the foregoing model training method and click rate determination method, and thus will not be described here again.

[0152] Traditional CTR estimation models are generally for a single domain, however, in large content platforms (websites / applications), there are usually many scenarios (domains) that need to use CTR estimation to improve user experience. Since different domains have overlapping users and objects, there is a certain commonality between different domains. Single-domain prediction does not utilize the commonality between domains, resulting in low prediction accuracy and a large number of models that are difficult to maintain. Different domains have different user behaviors and data distributions, which can easily lead to a seesaw effect.

[0153] To solve the above problems, one aspect of the present disclosure models the inter-domain commonality through a second prediction model, and the inter-domain commonality can alleviate the problem of inaccurate CTR estimation in some sparse domains to some extent. On the other hand, the first prediction model models the intra-domain characteristics of multiple scenarios, which can estimate the CTR of multiple scenarios and solve the problems of difficult maintenance, large resource consumption and large human resource investment. Further, by migrating the intermediate data of the first prediction model to the second prediction model, the intra-domain characteristics are retained while modeling the inter-domain commonality, improving the CTR estimation accuracy of all domains, fully utilizing the data of each domain, and solving the seesaw effect.

[0154] The present application can be applied to a recommendation system in various fields, which can discover the interest points of users, predict the information that users may need, and push the information to users in priority according to the behavior characteristics and other data of users, thereby improving the efficiency of users in obtaining information. The present application can be applied to a recommendation system serving multiple products (such as multiple application programs or clients), which can improve the prediction accuracy in multiple fields and multiple scenarios, reduce the difficulty of system maintenance, and save resources.

[0155] Exemplary apparatus

[0156] It should be noted that the model training method and the click rate determination method provided by the embodiments of the present disclosure can be executed by a corresponding device. Next, reference is made to FIG. 5 The model training device of the exemplary embodiments of the present disclosure is described first.

[0157] FIG. 5 A block diagram of a model training device according to one embodiment of the present application is schematically shown.

[0158] Referring to FIG. 5 The model training device 500 according to one embodiment of the present application is shown in FIG. 5. The device 500 can include an acquisition module 510 configured to acquire a training sample set, the training sample set including training samples and corresponding sample labels in multiple business scenarios; a training module 520 configured to input the training samples into a first prediction model and a second prediction model respectively, and migrate the intermediate data of the first prediction model to the second prediction model, to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; and an adjustment module 530 configured to adjust the model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample labels.

[0159] In some embodiments of the present disclosure, based on the foregoing scheme, the first prediction model comprises a plurality of scene models corresponding to a plurality of business scenarios, and the training sample comprises a sample scene feature. The device 500 can further comprise a scene model determination module configured to determine, according to the sample scene feature, a scene model corresponding to the training sample in the first prediction model; and the training module 520 can be further configured to input the training sample into the corresponding scene model and the second prediction model respectively.

[0160] In some embodiments of the present disclosure, based on the foregoing scheme, the training module 520 can be further configured to perform forward processing on the training sample by the first prediction model based on first standardization processing to obtain a first training click rate, wherein a standardization parameter of the first standardization processing is determined based on the training sample in the business scenario corresponding to the scene model.

[0161] In some embodiments of the present disclosure, based on the foregoing scheme, the training module 520 can be further configured to perform forward processing on the intermediate data and the training sample by the second prediction model based on second standardization processing to obtain a second training click rate, wherein a standardization parameter of the second standardization processing is determined based on the training sample in a plurality of business scenarios input into the second prediction model, and the training samples corresponding to the plurality of business scenarios have feature correlations.

[0162] In some embodiments of the present disclosure, based on the foregoing scheme, the adjustment module 530 can be further configured to determine a first loss function according to the first training click rate and the sample label; determine a second loss function according to the second training click rate and the sample label; determine a target loss function according to the first loss function and the second loss function; and adjust the model parameters of the first prediction model and the second prediction model according to the target loss function.

[0163] In some embodiments of the present disclosure, based on the foregoing scheme, the training module 520 can be further configured to migrate the intermediate data of the first prediction model to an intermediate network layer of the second prediction model by an adaptation network; and the adjustment module 530 can be further configured to adjust the model parameters of the first prediction model, the second prediction model and the adaptation network by minimizing the target loss function and gradient back propagation.

[0164] Reference is made to FIG. 6As shown, the click rate determination apparatus 600 according to an embodiment of the present application can include a scene model determination module 610, which can be configured to determine a target scene model in a first prediction model according to a scene feature of to-be-predicted data, the first prediction model including scene models for a plurality of business scenes; an input module 620, which can be configured to input the to-be-predicted data into the target scene model and a second prediction model respectively; a migration module 630, which can be configured to migrate intermediate data of the target scene model to the second prediction model to obtain a prediction output of the second prediction model; and a click rate determination module 640, which can be configured to determine a click rate of the to-be-predicted data according to the prediction output of the second prediction model.

[0165] In some embodiments of the present disclosure, based on the foregoing scheme, the migration module 630 can be further configured to migrate the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network, the adaptation network being used for screening the migrated data based on an importance of the intermediate data.

[0166] In some embodiments of the present disclosure, based on the foregoing scheme, the target scene model includes at least one first hidden layer, the second prediction model includes at least one second hidden layer, and the apparatus 600 further includes a prediction module, which can be configured to: perform first linear processing and first nonlinear conversion on the input data layer by layer through each first hidden layer; perform second linear processing and second nonlinear conversion on the intermediate data migrated from each first hidden layer and the input data layer by layer through each second hidden layer; and perform nonlinear normalization processing on an output of the last second hidden layer to obtain the prediction output.

[0167] In some embodiments of the present disclosure, based on the foregoing scheme, the target scene model further includes a first normalization layer, the second prediction model further includes a second normalization layer, and the apparatus 600 further includes a normalization module, which can be configured to, before the input data is input into the first hidden layer and the second hidden layer: perform first normalization processing on the input data through the first normalization layer; and perform second normalization processing on the input data through the second normalization layer.

[0168] In some embodiments of the present disclosure, based on the foregoing scheme, the to-be-predicted data includes user feature data, content feature data, and association feature data, and before the to-be-predicted data is input into the target scene model and the second prediction model respectively, the apparatus 600 further includes an embedding module, which can be configured to: utilize a pre-trained embedding model to perform vector representation on the user feature data, the content feature data, and the association feature data respectively, to correspondingly obtain a user feature vector, a content feature vector, and an association feature vector. The input module 620 can be further configured to input the user feature vector, the content feature vector, and the association feature vector after splicing to the target scene model and the second prediction model respectively.

[0169] In some embodiments of the present disclosure, based on the foregoing scheme, the device 600 further comprises a pushing module, which can be configured to determine target recommendation information according to the click rate of the data to be predicted, and push the target recommendation information to the user.

[0170] The specific details of the modules or units in the above model training device and click rate determination device have been described in detail in the corresponding model training method and click rate determination method, and thus will not be described here again.

[0171] Exemplary media

[0172] After introducing the method of the exemplary embodiments of the present application, next, the medium of the exemplary embodiments of the present application is described.

[0173] In some possible embodiments, various aspects of the present application can also be implemented as a storage medium having program codes stored thereon, which, when executed by a processor of a device, are used to implement the steps in the model training method according to various exemplary embodiments of the present application described in the above “Exemplary Method” section of the present specification.

[0174] Specifically, the processor of the device, when executing the program codes, is used to implement the following steps:

[0175] obtain a training sample set, the training sample set comprising training samples in multiple business scenarios and corresponding sample labels; input the training samples into a first prediction model and a second prediction model respectively, and migrate intermediate data of the first prediction model to the second prediction model to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; and adjust model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate, and the sample labels.

[0176] In other possible embodiments, when the program codes are executed by the processor of the device, they are used to implement the steps in the click rate determination method according to various exemplary embodiments of the present application described in the above “Exemplary Method” section of the present specification.

[0177] Specifically, the processor of the device, when executing the program codes, is used to implement the following steps:

[0178] determine a target scenario model in the first prediction model according to the scenario characteristics of the data to be predicted, the first prediction model comprising scenario models for multiple business scenarios; input the data to be predicted into the target scenario model and the second prediction model respectively; migrate intermediate data of the target scenario model to the second prediction model to obtain a prediction output of the second prediction model; and determine the click rate of the data to be predicted according to the prediction output of the second prediction model.

[0179] The above is a schematic solution of the computer readable storage medium of the embodiment. It should be noted that the technical solution of the storage medium belongs to the same concept as the technical solutions of the model training method and the click rate determination method, and the details of the technical solution of the storage medium that are not described in detail can be seen from the description of the technical solution of the model training method and the click rate determination method.

[0180] It should be noted that the storage medium described above can be a readable storage medium. The readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0181] The program code contained in the readable storage medium can be transmitted by any suitable medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0182] The program code for performing the operations of the present application can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C++, etc., and a conventional procedural programming language such as "C" language or similar programming languages. The program code can be executed entirely on the user electronic device, partially on the user electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In the case of a remote electronic device, the remote electronic device can be connected to the user electronic device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external electronic device (for example, connected through the Internet by using an Internet service provider).

[0183] Exemplary electronic devices

[0184] After introducing the method, medium and device of the exemplary embodiments of the present disclosure, next, an electronic device according to another exemplary embodiment of the present disclosure is introduced.

[0185] Those skilled in the art can understand that each aspect of the present application can be implemented as a system, a method or a program product. Therefore, each aspect of the present application can be embodied in a form of entirely hardware, entirely software (including firmware, microcode, etc.), or a combination of hardware and software, which can be collectively referred to as "circuitry", "module" or "system".

[0186] The electronic device 700 according to this embodiment of the present application will be described below with reference to FIG. 7 FIG. 7 The electronic device 700 is merely an example and should not impose any limitation on the function and scope of use of the embodiments of the present application.

[0187] As shown in FIG. 7 The electronic device 700 is in the form of a general electronic device. The components of the electronic device 700 can include, but are not limited to, the at least one processing unit 710 described above, the at least one storage unit 720 described above, and a bus 730 connecting different system components, including the storage unit 720 and the processing unit 710.

[0188] The storage unit stores program codes which can be executed by the processing unit 710, so that the processing unit 710 performs the steps according to various exemplary embodiments of the present application described in the "Exemplary Method" section of the present specification.

[0189] The storage unit 720 can include a readable medium in the form of a volatile storage unit, such as a random access memory (RAM) 7201 and / or a cache memory 7202, and can further include a read-only memory (ROM) 7203.

[0190] The storage unit 720 can further include a program / utility 7204 having a set of program modules 7205, including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination thereof can include implementation of a network environment.

[0191] The bus 730 can represent one or more of several types of bus structures, including a storage unit bus or storage unit controller, a peripheral bus, a graphics acceleration port, a processing unit bus, or a local bus using any of a variety of bus architectures.

[0192] ​The electronic device 700 can also communicate with one or more external devices such as a keyboard or a pointing device, through I / O interface 750 and display unit 740. In general, the electronic device 700 will communicate with one or more other devices, such as a router, a modem, or a network adapter, through network adapter 760 and bus 730. It should be appreciated that the network adapter 760 can also be used to enable connection to other types of networks and remote computing resources, such as a mobile telephone network, a local area network (LAN), a wide area network (WAN), and / or the Internet. It will also be appreciated that the electronic device 700 might include other modules that might be used in connection with the various embodiments of the disclosure, such as security measures, an email client application, an instant messaging application, and the like.

[0193] From the above description of embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solutions according to the embodiments of the disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash disk, a mobile hard disk, or the like) or a network, and includes a number of instructions to enable an electronic device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to perform the methods according to the embodiments of the disclosure.

[0194] The above is a schematic solution of the electronic device 700 of the embodiment. It should be noted that the technical solution of the electronic device 700 belongs to the same concept as the technical solutions of the model training method and the click rate determination method, and the technical solution of the electronic device that is not described in detail can be understood with reference to the description of the technical solution of the call not connected state determination method.

[0195] It should be noted that although several modules or sub-modules of the call not connected state determination apparatus are mentioned in the above detailed description, such a division is merely exemplary and is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into a plurality of modules or units.

[0196] Furthermore, although the operations of the method of the present application are described in a particular, sequential order, this order is not meant to be a limitation and is not intended to imply that there is an absolute requirement that the operations be performed in the order described. On the contrary, many of the operations can be performed in any order, added to, or omitted from the operations described. Additionally or alternatively, certain steps can be performed in parallel with one another.

[0197] While the principles and spirit of the application have been described with reference to several specific embodiments, it is to be understood that the application is not limited to the specific embodiments disclosed and that modifications to, and the use of, each of the various aspects are intended to be included as is within the spirit and scope of the application. The application is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A model training method, characterized in that, The method comprises: obtaining a training sample set comprising training samples in multiple business scenarios and corresponding sample labels; the training sample comprises user interaction data, sample scene features, user interaction data refers to data generated by user interaction behavior in the application platform, and sample label refers to the click probability of the user on the recommended object; inputting the training sample into the first prediction model and the second prediction model respectively, and migrating the intermediate data of the first prediction model to the second prediction model to obtain the first training click rate output by the first prediction model and the second training click rate output by the second prediction model; the first prediction model is used to learn the characteristics in the business scenario, and the second prediction model is used to learn the common characteristics between the business scenarios; adjusting the model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample label; wherein migrating the intermediate data of the first prediction model to the second prediction model comprises migrating the intermediate data of the first prediction model to the intermediate network layer of the second prediction model, and inputting the data after fusion into the next layer in the second prediction model.

2. The method of claim 1, wherein, The first prediction model comprises multiple scene models corresponding to the business scenarios, and the training sample comprises sample scene features, and the method further comprises: determining the scene model corresponding to the training sample in the first prediction model according to the sample scene features; inputting the training sample into the corresponding scene model and the second prediction model respectively. The first training click rate is obtained in the following way:

3. The method of claim 2, wherein, based on the first standardization processing, the first prediction model is used to forward process the training sample to obtain the first training click rate, and the standardization parameter of the first standardization processing is determined based on the training sample in the business scenario corresponding to the scene model. The second training click rate is obtained in the following way:

4. The method of claim 3, wherein, based on the second standardization processing, the second prediction model is used to forward process the intermediate data and the training sample to obtain the second training click rate, and the standardization parameter of the second standardization processing is determined based on the training sample in multiple business scenarios input into the second prediction model, and the training samples corresponding to the multiple business scenarios have feature correlation. The adjustment of the model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample label comprises:

5. The method of claim 1, wherein, determining a first loss function according to the first training click rate and the sample label; determining a second loss function according to the second training click rate and the sample label; determining a target loss function according to the first loss function and the second loss function; adjusting the model parameters of the first prediction model and the second prediction model according to the target loss function. The migration of the intermediate data of the first prediction model to the second prediction model comprises:

6. The method of claim 5, wherein, ​ migrating intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network; the adjusting the model parameters of the first prediction model and the second prediction model according to the target loss function comprises: the model parameters of the first prediction model, the second prediction model and the adaptation network are adjusted through minimizing the target loss function and gradient back propagation.

7. A click rate determination method, characterized by, The method comprises: determining a target scene model in the first prediction model according to a scene feature of to-be-predicted data, the first prediction model comprising scene models for a plurality of business scenes; inputting the to-be-predicted data into the target scene model and a second prediction model respectively; the first prediction model is used to learn characteristics within a business scene, and the second prediction model is used to learn common characteristics between business scenes; migrating intermediate data of the target scene model to the second prediction model to obtain a prediction output of the second prediction model; determining a click rate of the to-be-predicted data according to the prediction output of the second prediction model; wherein the migrating the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model, and inputting the data into a next layer in the second prediction model after data fusion.

8. The method of claim 7, wherein, The migrating the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network, the adaptation network being used to filter the migrated data based on an importance of the intermediate data.

9. The method of claim 7, wherein, The target scene model comprises at least one first hidden layer, and the second prediction model comprises at least one second hidden layer, and the obtaining the prediction output of the second prediction model comprises: performing first linear processing and first nonlinear conversion on input data layer by layer through each first hidden layer; performing second linear processing and second nonlinear conversion on the input data and intermediate data migrated from each first hidden layer layer by layer through each second hidden layer; performing nonlinear normalization processing on an output of a last second hidden layer to obtain the prediction output.

10. The method of claim 9, wherein, The target scene model further comprises a first standardization layer, and the second prediction model further comprises a second standardization layer, and the method further comprises: performing first standardization processing on the input data through the first standardization layer; performing second standardization processing on the input data through the second standardization layer.

11. The method of claim 7, wherein, The to-be-predicted data comprises user feature data, content feature data and association feature data, and before the to-be-predicted data is input into the target scene model and the second prediction model respectively, the method further comprises: using a pre-trained embedding model to perform vector representation on the user feature data, the content feature data and the association feature data respectively, to correspondingly obtain a user feature vector, a content feature vector and an association feature vector; The inputting the to-be-predicted data into the target scene model and the second prediction model respectively comprises: After the user feature vector, the content feature vector and the association feature vector are spliced, the spliced result is input into the target scene model and the second prediction model respectively.

12. The method according to any one of claims 7-11, characterized in that, The method further comprises: determining target recommendation information according to the click rate of the to-be-predicted data, and pushing the target recommendation information to the user.

13. A model training apparatus, comprising: The device comprises: The obtaining module is configured to obtain a training sample set, the training sample set comprising training samples in multiple business scenarios and corresponding sample labels; the training sample comprises user interaction data and sample scene features, the user interaction data being data generated by user interaction behavior in an application platform, and the sample label being a click probability of a user on a recommended object; The training module is configured to input the training sample into a first prediction model and a second prediction model respectively, and migrate intermediate data of the first prediction model to the second prediction model, to obtain a first training click rate output by the first prediction model and a second training click rate output by the second prediction model; the first prediction model is used to learn characteristics in a business scenario, and the second prediction model is used to learn common characteristics between business scenarios; The adjusting module is configured to adjust model parameters of the first prediction model and the second prediction model according to the first training click rate, the second training click rate and the sample label. The intermediate data of the first prediction model is migrated to an intermediate network layer of the second prediction model, and after data fusion, the intermediate data is input into a next layer in the second prediction model.

14. The apparatus of claim 13, wherein, The first prediction model comprises scene models corresponding to multiple business scenarios, and the training sample comprises sample scene features, and the device further comprises: A scene model determination module determines a scene model corresponding to the training sample in the first prediction model according to the sample scene features. The training module is further configured to: input the training sample into the corresponding scene model and the second prediction model respectively.

15. The apparatus of claim 14, wherein, The training module is further configured to: perform forward processing on the training sample through the first prediction model based on first standardization processing, to obtain the first training click rate, a standardization parameter of the first standardization processing being determined based on training samples in a business scenario corresponding to the scene model.

16. The apparatus of claim 15, wherein, The training module is further configured to: perform forward processing on the intermediate data and the training sample through the second prediction model based on second standardization processing, to obtain the second training click rate, a standardization parameter of the second standardization processing being determined based on training samples in multiple business scenarios input into the second prediction model, the training samples in the multiple business scenarios having feature association.

17. The apparatus of claim 13, wherein, The adjusting module is further configured to: determine a first loss function according to the first training click rate and the sample label; determine a second loss function according to the second training click rate and the sample label; and determine a target loss function according to the first loss function and the second loss function; adjust model parameters of the first prediction model and the second prediction model according to the target loss function.

18. The apparatus of claim 17, wherein, The migration of the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network; The adjustment module is further configured to: adjust model parameters of the first prediction model, the second prediction model and the adaptation network by minimizing the target loss function and gradient back propagation.

19. A click rate determining apparatus characterized by comprising: The device comprises: a scenario model determination module configured to determine a target scenario model in a first prediction model according to a scenario feature of to-be-predicted data, the first prediction model comprising scenario models for a plurality of business scenarios; an input module configured to input the to-be-predicted data into the target scenario model and a second prediction model respectively, the first prediction model being used to learn characteristics within a business scenario, and the second prediction model being used to learn common characteristics between business scenarios; a migration module configured to migrate intermediate data of the target scenario model to the second prediction model to obtain a prediction output of the second prediction model; a click rate determination module configured to determine a click rate of the to-be-predicted data according to the prediction output of the second prediction model. The migration of the intermediate data of the first prediction model to the second prediction model comprises: migrating the intermediate data of the first prediction model to an intermediate network layer of the second prediction model, and inputting the data into a next layer in the second prediction model after data fusion.

20. The apparatus of claim 19, wherein, The migration module is further configured to: migrate the intermediate data of the first prediction model to an intermediate network layer of the second prediction model through an adaptation network, the adaptation network being used to filter the migrated data based on an importance of the intermediate data.

21. The apparatus of claim 19, wherein, The target scenario model comprises at least one first hidden layer, the second prediction model comprises at least one second hidden layer, and the device further comprises a prediction module configured to: perform first linear processing and first nonlinear conversion on input data layer by layer through each of the first hidden layers; perform second linear processing and second nonlinear conversion on the intermediate data migrated from each of the first hidden layers and the input data layer by layer through each of the second hidden layers; perform nonlinear normalization processing on an output of a last layer of the second hidden layers to obtain the prediction output.

22. The apparatus of claim 21, wherein, The target scenario model further comprises a first normalization layer, the second prediction model further comprises a second normalization layer, and the device further comprises a normalization module configured to: perform first normalization processing on the input data through the first normalization layer before the input data is input into the first hidden layers and the second hidden layers; perform second normalization processing on the input data through the second normalization layer.

23. The apparatus of claim 19, wherein, The to-be-predicted data includes user feature data, content feature data, and association feature data. Before the to-be-predicted data is input into the target scene model and the second prediction model respectively, the apparatus further includes an embedding module configured to: vectorize the user feature data, the content feature data, and the association feature data respectively by using a pre-trained embedding model, to obtain a user feature vector, a content feature vector, and an association feature vector respectively; the input module is further configured to: input the user feature vector, the content feature vector, and the association feature vector into the target scene model and the second prediction model respectively after splicing.

24. The apparatus of any of claims 19-23, wherein, The apparatus further includes: a push module configured to determine target recommendation information according to a click rate of the to-be-predicted data, and push the target recommendation information to a user.

25. An electronic device, comprising: a processor and a memory, the memory storing executable instructions, and the processor being configured to invoke the executable instructions stored in the memory to perform the method according to any one of claims 1 to 12.

26. A storage medium having a program stored thereon, the program being executed by a processor to implement the method according to any one of claims 1 to 12.

Images