Training method, resource recommendation method and device for click-through rate prediction model

Abstract

The present disclosure provides a training method, a resource recommendation method and an apparatus for a click-through rate prediction model, which relates to the field of artificial intelligence technology, in particular to the field of intelligent search and deep learning, and can be used in smart city scenarios. The specific implementation scheme is: controlling multiple sub-processes executed in parallel to filter out a first sample set from a candidate sample set and obtain a first parameter corresponding to the first sample set; summarizing the first parameters to obtain a first parameter set; training the target filter of the main process based on the sample status of each first sample set, the first parameter set and the screening action of each sub-process on the candidate sample set; filtering out a second sample set from the candidate sample set based on the target filter; training the target click-through rate prediction model of the main process based on the second sample set and the known positive sample set; and ending the training of the target click-through rate prediction model when the target click-through rate prediction model meets the training convergence conditions. The present disclosure can improve the learning efficiency of the model.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, in particular to the fields of intelligent search and deep learning, and can be used in smart city scenarios. Background Art

[0002] Recommendation systems are often found in online business systems such as search engines, e-commerce systems, and social applications. It's well known that if a recommendation system's recommendations match a user's interests, the user will experience a positive experience and be more likely to click on them. Click-through rate (CTR) prediction models are crucial in retrieving ads or recommended items of interest to users. Therefore, improving the CTR prediction capabilities of CTR prediction models is a worthy topic of research. Summary of the Invention

[0003] The present disclosure provides a training method for a click-through rate prediction model, a resource recommendation method, and a device.

[0004] According to one aspect of the present disclosure, a method for training a click-through rate prediction model is provided, comprising:

[0005] Controlling multiple sub-processes executed in parallel to select first sample sets from the candidate sample sets and obtain first parameters corresponding to the first sample sets;

[0006] Summarizing the first parameters corresponding to the first sample sets to obtain a first parameter set;

[0007] Training a target filter of the main process based on the sample status of each first sample set, the first parameter set, and the screening action of each sub-process on the candidate sample set;

[0008] Filtering a second sample set from the candidate sample set based on the target filter;

[0009] The target click rate prediction model of the main process is trained based on the second sample set and the known positive sample set;

[0010] When the target click-through rate prediction model meets the training convergence condition, the training of the target click-through rate prediction model is terminated.

[0011] According to another aspect of the present disclosure, a resource recommendation method is provided, comprising:

[0012] Responding to a recommendation request for a target user, obtaining candidate resources;

[0013] Input candidate resources into the click-through rate prediction model to obtain the click-through rate of target users on the candidate resources;

[0014] When it is determined based on the click rate of the candidate resource that the candidate resource is recommended to the target user, the candidate resource is sent to the target user.

[0015] According to another aspect of the present disclosure, a training device for a click-through rate prediction model is provided, comprising:

[0016] a parallel execution module, configured to control a plurality of sub-processes executed in parallel to respectively select first sample sets from the candidate sample sets and obtain first parameters corresponding to the first sample sets;

[0017] a reward determination module, configured to aggregate the first parameters corresponding to the first sample sets to obtain a first parameter set;

[0018] A first training module is configured to train a target filter of the main process based on the sample status of each first sample set, the first parameter set, and the screening action of each sub-process on the candidate sample set;

[0019] A screening module, configured to screen out a second sample set from the candidate sample set based on a target filter;

[0020] A second training module is used to train a target click rate prediction model of the main process based on the second sample set and the known positive sample set;

[0021] The end module is used to end the training of the target click-through rate prediction model when the target click-through rate prediction model meets the training convergence condition.

[0022] According to another aspect of the present disclosure, a resource recommendation device is provided, comprising:

[0023] An acquisition module, configured to acquire candidate resources in response to a recommendation request to a target user;

[0024] The prediction module is used to input candidate resources into the click-through rate prediction model to obtain the click-through rate of the candidate resources by target users;

[0025] The recommendation module is used to send the candidate resource to the target user when determining to recommend the candidate resource to the target user based on the click rate of the candidate resource.

[0026] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0027] at least one processor; and,

[0028] a memory communicatively connected to the at least one processor; wherein,

[0029] The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method of any embodiment of the present disclosure.

[0030] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method according to any embodiment of the present disclosure.

[0031] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program, which implements the method according to any embodiment of the present disclosure when executed by a processor.

[0032] The disclosed embodiments utilize a parallel learning process, using multiple subprocesses to simultaneously filter valid samples, enabling rapid optimization of the target filter. The optimized target filter then uses the valid samples to train the target click-through rate prediction model, minimizing ineffective learning and improving the learning efficiency of the target click-through rate prediction model.

[0033] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0035] Figure 1 is a schematic diagram of a joint training framework according to an embodiment of the present disclosure;

[0036] Figure 2 is a flowchart of a method for training a click-through rate prediction model according to an embodiment of the present disclosure;

[0037] FIG3( a ) is a schematic diagram of a parallel learning framework of a training method for a click-through rate prediction model according to an embodiment of the present disclosure;

[0038] FIG3( b ) is a schematic diagram of a parallel learning framework of a training method for a click-through rate prediction model according to another embodiment of the present disclosure;

[0039] Figure 4 is a flow chart illustrating a case where a target click-through rate prediction model does not meet a training convergence condition according to an embodiment of the present disclosure;

[0040] Figure 5 is a flowchart of a resource recommendation method according to an embodiment of the present disclosure;

[0041] Figure 6 1 is a structural diagram of a training device for a click-through rate prediction model according to an embodiment of the present disclosure;

[0042] Figure 7is a structural diagram of a resource recommendation device according to an embodiment of the present disclosure;

[0043] Figure 8 It is a block diagram of an electronic device used to implement the training method of the click rate prediction model or the resource recommendation method of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0044] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be appreciated by those skilled in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0045] The training samples of traditional click-through rate prediction models come from recommended samples displayed online. Among them, positive samples represent samples that have been clicked by users, while negative samples represent samples in recommended samples that have not been clicked by users (hereinafter referred to as unlabeled negative samples). However, in the existing training methods of click-through rate prediction models, all unclicked negative samples are directly regarded as unlabeled negative samples. However, unclicked samples are not necessarily samples that users are not interested in. Some samples that users are interested in may not be clicked by users due to display location or other reasons. This shows that there is noise in unlabeled negative samples. Since resources are massive, the number of clicked samples in these resources is always small. Therefore, directly treating unclicked samples as negative samples will result in large negative sample noise and truly effective feedback is too sparse. This results in the fact that the training samples of the click-through rate prediction model are not all valid samples. Since the click-through rate prediction model cannot be trained with valid samples, the actual performance of click-through rate prediction is reduced, and even the user experience is damaged.

[0046] In view of this, in order to improve the prediction function of the click-through rate prediction model, a training method for the click-through rate prediction model is provided in the embodiment of the present disclosure. The method proposes a joint training framework. Figure 1As shown in the figure, the framework includes a filter and a click-through rate prediction model. The filter is used to identify truly valid samples from a noisy sample set, and then the click-through rate prediction model is trained with the filtered sample set and the known labeled positive sample set. The reward is then determined based on the prediction results of each sample by the click-through rate prediction model. Based on the reward, the sample state of the sample set, and the action of the filter, reinforcement learning can be applied to the filter to optimize the filter. The optimization of the filter enables the filter to filter out truly valid negative samples and reduce sample noise. The above process can be called one round of training. In order to optimize the click-through rate prediction model, multiple rounds of training can be performed in a loop to improve the performance of the click-through rate prediction model.

[0047] In addition, if Figure 1 As shown, the training process of one round is long, and each operation needs to be performed serially. When multiple rounds of training are required to optimize the click-through rate prediction model, the entire training process takes a long time. In order to improve training efficiency, the present disclosure provides a parallel learning method to train the click-through rate prediction model.

[0048] The following describes the training method of the click-through rate prediction model provided by the embodiment of the present disclosure with reference to the accompanying drawings.

[0049] like Figure 2 FIG. 1 is a flow chart of a method for training a click-through rate prediction model according to an embodiment of the present disclosure, including the following contents:

[0050] S201: Control multiple sub-processes executed in parallel to select first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set. The first parameters can be understood as the aforementioned rewards.

[0051] S202: Summarize the first parameters corresponding to the first sample sets to obtain a first parameter set.

[0052] S203 : Based on the sample states of the first sample sets, the first parameter sets, and the screening actions of the sub-processes on the candidate sample sets, the target filter of the main process is trained.

[0053] S204: Filter out a second sample set from the candidate sample set based on the target filter.

[0054] S205 , training a target click rate prediction model of the main process based on the second sample set and the known positive sample set.

[0055] S206 , when the target click-through rate prediction model meets the training convergence condition, the training of the target click-through rate prediction model is terminated.

[0056] Figure 3(a) shows a schematic diagram of the parallel learning framework. This framework consists of a main process and multiple subprocesses. The main process includes a target filter and a target click-through rate prediction model. Multiple subprocesses work in parallel to filter valid samples from the candidate sample set and determine rewards to optimize the main process's target filter as quickly as possible. The optimized target filter then selects valid samples from the candidate sample set to optimize the target click-through rate prediction model.

[0057] In another embodiment, as shown in Figure 3(b), a sub-process includes a filter copy and a click-through rate prediction model copy. Each sub-process, based on the filter copy, selects a first sample set from a candidate sample set and inputs the first sample set into the click-through rate prediction model copy of that sub-process to obtain first parameters corresponding to the first sample set. The first sample set and first parameters of each sub-process are aggregated and used to train the target filter of the main process. The target click-through rate prediction model of the main process is then optimized based on the optimized target filter.

[0058] In summary, the parallel learning process based on multiple subprocesses can filter out valid samples through multiple subprocesses in parallel, optimizing the target filter as quickly as possible so that it can filter out valid samples. Based on this, model training is performed based on valid samples, enabling the target click-through rate prediction model to learn effectively, thereby minimizing ineffective learning processes, improving learning efficiency, and increasing model training speed.

[0059] In the disclosed embodiment, two candidate sample sets can be provided for training the target click-through rate prediction model. One is a candidate sample set consisting of unlabeled negative samples, through which valid negative samples can be screened out for learning. The other is a candidate sample set constructed from labeled positive samples and unlabeled samples. Through such a sample set, the filter can learn knowledge transfer and can not only screen out valid negative samples from the candidate sample set, but also learn to screen out valid positive samples from unlabeled samples for learning. The following describes the learning and training process for different candidate sample sets respectively.

[0060] 1) Unlabeled negative samples construct candidate sample sets

[0061] When a candidate sample set is constructed from unlabeled negative samples, each sub-process is used to filter out negative samples from the candidate sample set to obtain a first sample set.

[0062] In some embodiments, as shown in FIG3( b ), each sub-process has a corresponding filter copy and click-through rate prediction model copy. Controlling multiple sub-processes to be executed in parallel to filter out first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set can be implemented by performing the following operations for each sub-process:

[0063] The control sub-process filters out a first sample set from the candidate sample set when the filter copy of the sub-process is in exploration mode; the control sub-process inputs the first sample set and the known positive sample set into the click-through rate prediction model copy of the sub-process to obtain a first parameter corresponding to the first sample set.

[0064] The child process's filter replica is in exploration mode, while the main process's target filter is in evaluation mode. In exploration mode, the filter replica explores the filtering action to filter samples based on an exploration strategy. This exploration strategy includes classic exploration methods such as epsilon-greedy, Gaussian noise exploration, and upper confidence bound (UCB), as well as reward-based and memory-based exploration methods.

[0065] In evaluation mode, the target filter in the main process does not adopt an exploration strategy and does not explore the screening actions. Instead, it uses the model prediction value of the target filter to screen samples.

[0066] As shown in Figure 3(b), each subprocess obtains the sample state corresponding to each sample in the candidate sample set and inputs this sample state into the filter replica. The filter replica is now in exploration mode and performs filtering based on the sample state. This filtering action is divided into selection and non-selection, with the selected samples forming the first sample set. By inputting the first sample set and the known positive sample set into the CTR prediction model replica, the first parameter can be obtained.

[0067] The sample states in the candidate sample set can be the sample input features. These input features can be determined based on specific business scenarios. Business scenarios can be divided into three categories: user information, sample information, and sample source.

[0068] Among them, user information includes but is not limited to basic characteristics of the user, such as the user's age, gender, zodiac sign, occupation, education level and life stage; location characteristics of the user, such as the user's birthplace, place of growth, place of work, home location, etc.; user behavior characteristics, such as the user's interest preference characteristics, activity characteristics, etc.

[0069] The sample information includes, but is not limited to, the sample title, sample quality, and sample type. Sample types include news, technology, history, etc.; sample types include pictures, videos, movies, novels, etc.

[0070] Sample sources can be categorized by the platform from which they originate, such as TikTok, Kuaishou, and related multimedia resource production platforms. Of course, sample sources can also be differentiated by different terminal devices.

[0071] In the disclosed embodiment, a subprocess contains copies of the filter and the click-through rate prediction model. Sample screening and reward acquisition can be completed in parallel by multiple subprocesses. This improves the efficiency of sample screening and reward determination, and consequently, the efficiency of training the target click-through rate prediction model.

[0072] Based on the content described above, the embodiment of the present disclosure optimizes the target filter based on a deep learning method. The deep learning method requires determining a reward. In the embodiment of the present disclosure, for each first sample set, a first parameter corresponding to the first sample set is obtained, which can be implemented as follows: based on the click-through rate prediction value of each target sample in the first sample set, the first sub-parameter of each target sample is determined; based on the model evaluation result of the target click-through rate prediction model, the second sub-parameter is determined; based on the first sub-parameter and the second sub-parameter of each sample, the first parameter corresponding to the first sample set is determined.

[0073] Therefore, the first sub-parameter reflects the click-through rate prediction of each target sample, and the second sub-parameter describes the overall prediction effect of the target click-through rate prediction model. The first parameter determined by the first sub-parameter and the second sub-parameter contains the reward information of a single sample and the overall information of the target click-through rate prediction model. By optimizing the target filter through the first parameter, the model parameters of the target filter can be adjusted to achieve a better effect as soon as possible, thereby improving the training efficiency.

[0074] Among them, when constructing a candidate sample set from unlabeled samples, the following method can be used

[0075] Determine the first sub-parameter for each target sample. This can be implemented as follows: if the target sample is determined to be a negative sample based on the click-through rate prediction value, the first sub-parameter of the target sample is determined to be a preset positive value; if the target sample is determined to be a positive sample based on the click-through rate prediction value, the first sub-parameter of the target sample is determined to be a preset negative value. Since the first sub-parameter represents the evaluation of each sample, it can be referred to as a single-step reward.

[0076] For example, if the default positive value is 1 and the default negative value is -1, then for the a-th target sample, if the filter copy selects this sample, it means that the filter copy considers it a negative sample. If the click-through rate prediction model copy also predicts this sample as a negative sample, then the filter copy and the click-through rate prediction model copy have the same judgment on this sample, and the first sub-parameter of this target sample is 1. Similarly, for the b-th sample, if the filter copy outputs the result that it is a negative sample, while the click-through rate prediction model copy outputs the result that it is a positive sample, then the two judgments are inconsistent, and the first sub-parameter of this sample is -1.

[0077] In the embodiment of the present disclosure, the first sub-parameter is determined based on whether the prediction result of the click-through rate prediction model is consistent with the action of the filter. The first sub-parameter thus determined can reflect the screening effect of the filter, and thus can better optimize the model parameters of the filter to improve the training effect of the target click-through rate prediction model.

[0078] Since the second sub-parameter is used to evaluate the target click-through rate prediction model, it can also be called the round reward. The Area Under Curve (AUC) metric can be used as the second sub-parameter in implementation. In addition to the aforementioned AUC metric, other evaluation metrics include model accuracy, mean absolute error (MAE), cross-validation, and F1-SCORE. It should be noted that different evaluation methods will result in different metrics.

[0079] After obtaining the first sub-parameter and the second sub-parameter, the first parameter corresponding to the first sample set is determined based on the first sub-parameter and the second sub-parameter of each sample. This can be implemented by performing a weighted summation of the second sub-parameter and the first sub-parameter of each sample to obtain the first parameter R, as shown in Expression (1).

[0080]

[0081] In formula (1), R i Represents the first sub-parameter of the i-th target sample, R e Represents the second sub-parameter, It represents the summation of the first sub-parameters of N target samples. Alpha represents the weight, which is a hyperparameter that can be adjusted based on the training effect.

[0082] In this disclosed embodiment, a weighted summation approach is used to obtain the first parameter, enabling the target filter to focus on learning a particular reward, or to learn a combination of both rewards. This learning approach helps better adjust the parameters of the target filter and improve its ability to screen samples.

[0083] In some embodiments, the target filter of the main process is trained based on the sample states of each first sample set, the first parameter set, and the screening actions of each sub-process on the candidate sample set. Taking the training process as an example, when there are n sub-processes, where n is a positive integer, each sub-process can determine a first sample set and a corresponding first parameter, thereby obtaining n first sample sets and n first parameters. Therefore, the sample states of the n first sample sets are spliced ​​together to obtain a state set O, which can be expressed as O = {o1, o2, ..., o n}, concatenate n first parameters to obtain the first parameter set R Z , which can be expressed as R Z ={R1, R2, ..., R n}, the action set A can be obtained by splicing the screening actions of the candidate sample set by n sub-processes, which can be expressed as A={A1,A2,...,A n}, based on the state set O, the first parameter set R Z And the action set A is used to train the target filter of the main process, and the training method can use the Policy Gradient algorithm.

[0084] The Policy Gradient algorithm trains a filter based on reward results. This method increases the probability of actions with large rewards and decreases the probability of actions with small rewards. It also offers the advantage of round-based updates: after each round, the total reward is obtained, and the model parameters are automatically updated.

[0085] In each round of training, the target filter traverses all samples in the candidate sample set in evaluation mode to filter and obtain the second sample set.

[0086] When constructing a candidate sample set from unlabeled negative samples, the target filter acts as a negative sample filter. The second sample set now includes the filtered, valid negative samples. The target CTR prediction model of the main process is subsequently trained based on the second sample set and the known positive sample set. Training of the target CTR prediction model ends when the model meets the training convergence criteria.

[0087] In some embodiments, the training convergence condition for the target CTR prediction model may be whether the second sub-parameter is greater than a threshold. If the second sub-parameter is greater than the threshold, the training of the target CTR prediction model is terminated. The training convergence condition may also be whether the target CTR prediction model has reached a preset number of iterations. If the number of iterations meets the preset number, the training may also be terminated.

[0088] In addition, in the disclosed embodiment, after training the target filter of the main process, the model parameters of the target filter are synchronously updated to the filter copies of each child process. This facilitates the child process to use the appropriate filter copy to filter samples in the next round of training.

[0089] Similarly, after training the target click-through rate prediction model of the main process, the model parameters of the target click-through rate prediction model are synchronously updated to the click-through rate prediction model copies of each child process, so that the child process can be rewarded based on the appropriate click-through rate prediction model copy.

[0090] Therefore, by timely updating the filter copy and click-through rate prediction model copy of the child process, the model of the child process can be synchronized with the model of the main process, which is conducive to adjusting the parameters of the target click-through rate prediction model as soon as possible and improving training efficiency.

[0091] In some embodiments, when the target click-through rate prediction model does not meet the training convergence condition, the process returns to the step of controlling multiple sub-processes executed in parallel to screen out first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set. The specific process is as follows: Figure 4 As shown, until the target click-through rate prediction model meets the training convergence conditions.

[0092] S401 : Control multiple sub-processes executed in parallel to select first sample sets from candidate sample sets and obtain first parameters corresponding to the first sample sets.

[0093] S402: Summarize the first parameters corresponding to the first sample sets to obtain a first parameter set.

[0094] S403 : Based on the sample states of the first sample sets, the first parameter sets, and the screening actions of the sub-processes on the candidate sample sets, the target filter of the main process is trained.

[0095] S404: Filter out a second sample set from the candidate sample set based on the target filter.

[0096] S405 , training a target click rate prediction model of the main process based on the second sample set and the known positive sample set.

[0097] S406: Determine whether the click rate prediction model meets the convergence condition.

[0098] S407, when the click rate prediction model does not meet the convergence condition, the filter copy and the click rate prediction model copy in the child process are synchronously updated based on the filter and the click rate prediction model in the main process, and steps S401 to S406 are executed in a loop.

[0099] S408 : When the click-through rate prediction model satisfies the convergence condition, determine whether the click-through rate prediction model satisfies the convergence condition.

[0100] In the disclosed embodiment, the click-through rate prediction model and filter are updated in a loop-based manner, and sample screening is performed based on multiple sub-processes executed in parallel, which not only improves efficiency but also improves the accuracy of the model.

[0101] It should be noted that multiple sub-processes and the main process can be executed on one device or on multiple devices.

[0102] 2) Positive samples and unlabeled samples construct candidate sample sets

[0103] Among them, unlabeled samples may include unlabeled negative samples in recommended samples, and may also include samples of similar products. Among them, samples of similar products include, for example, samples recommended to users in the search box, similar samples recommended to users in the interface browsed by users, etc. Positive samples may include samples that users have clicked recently, such as in the past week. It should be noted that the purpose of constructing the candidate sample set from positive samples and unlabeled samples in the embodiments of the present disclosure is to expand the candidate sample set.

[0104] In the disclosed embodiment, the candidate sample set is expanded. At this time, the filter is not limited to being a negative sample filter, but a sample knowledge migrator, which migrates the valid sample knowledge of other products to the current product, so that the target filter can perform knowledge migration, that is, the target filter can not only filter out valid negative samples, but also filter out valid positive samples from unlabeled samples.

[0105] It can be understood that, when the candidate sample set includes positive samples and unlabeled samples, the first sample set includes the screened positive samples and the screened negative samples.

[0106] In some embodiments, each sub-process has a corresponding filter copy and click-through rate prediction model copy. Controlling multiple sub-processes executed in parallel to filter out first sample sets from candidate sample sets and obtain first parameters corresponding to each first sample set can be implemented by performing the following operations on each sub-process:

[0107] The control sub-process filters out a first sample set from the candidate sample set when the filter copy of the sub-process is in exploration mode; the control sub-process inputs the first sample set and the known positive sample set into the click-through rate prediction model copy of the sub-process to obtain a first parameter corresponding to the first sample set.

[0108] In some embodiments, when the candidate sample set includes positive samples and unlabeled samples, each sub-process obtains the sample state corresponding to each sample in the candidate sample set. The sample state can be the input feature of the corresponding sample. The sample state is input into the filter replica. The filter replica is now in exploration mode and performs a screening action based on the sample state. The screening action is divided into selection and non-selection. The selected samples constitute the first sample set. The first sample set and the known positive sample set are input into the click-through rate prediction model replica to obtain the first parameter.

[0109] Based on the content described above, the embodiment of the present disclosure optimizes the target filter based on a deep learning method. The deep learning method requires determining a reward. In the embodiment of the present disclosure, for each first sample set, a first parameter corresponding to the first sample set is obtained, which can be implemented as follows: based on the click-through rate prediction value of each target sample in the first sample set, the first sub-parameter of each target sample is determined; based on the model evaluation result of the target click-through rate prediction model, the second sub-parameter is determined; based on the first sub-parameter and the second sub-parameter of each sample, the first parameter corresponding to the first sample set is determined.

[0110] Wherein, when a candidate sample set is constructed from positive samples and unlabeled samples, the following method can be used to determine the first sub-parameter of each target sample. This can be implemented as follows: determining the first sub-parameter of the target sample based on a specified principle; wherein, when the target sample is a sample screened out from unlabeled samples, the specified principle is used to: when the first judgment result of the target sample based on the click-through rate prediction value is consistent with the second judgment result represented by the screening action, the first sub-parameter is a positive value; when the first judgment result and the second judgment result are inconsistent, the first sub-parameter is a negative value; the first judgment result and the second judgment result are used to indicate whether the target sample is a positive sample or a negative sample; when the target sample is a positive sample screened out from the positive samples, the specified principle is used to set the first sub-parameter of the target sample to a default value.

[0111] In some embodiments, the first sub-parameter R of the target sample is determined based on the following first sub-parameter determination formula that satisfies a specified criterion: i It can be shown as expression (2),

[0112]

[0113] Among them, clip() represents the truncation function, which means as well as The value of is limited to the range of (-1, 1), c i Represents the prediction result of the click-through rate prediction model, o i Represents the filter result of a filter.

[0114] Among them, the range of the first sub-parameter is limited based on the truncation function. When the output result of the click-through rate prediction model is inconsistent with the output result of the filter, the first sub-parameter is a negative number, indicating punishment; when the output result of the click-through rate prediction model is consistent with the output result of the filter, the first sub-parameter is a positive number, indicating reward.

[0115] In the embodiment of the present disclosure, the first sub-parameter is calculated using a truncation function, which can limit the reward to a range, making its distribution more concentrated, which is beneficial for subsequent calculations.

[0116] In the embodiment of the present disclosure, judging the judgment results of the filter and the judgment results of the click-through rate prediction model based on the specified principles can effectively determine whether the prediction results of the model are accurate, which helps to better adjust the parameters of the target filter and improve the target filter's ability to screen samples.

[0117] After obtaining the first sub-parameter and the second sub-parameter, the first parameter corresponding to the first sample set is determined based on the first sub-parameter and the second sub-parameter of each sample. The method for obtaining the first parameter is similar to the above, and will not be repeated here.

[0118] In some embodiments, the target filter of the main process is trained based on the sample states of each first sample set, the first parameter set, and the screening actions of each sub-process on the candidate sample set. Taking the training process as an example, when there are m sub-processes, where m is a positive integer, each sub-process can determine a first sample set and a corresponding first parameter, thereby obtaining m first sample sets and m first parameters. Thus, m first sample sets and m first parameters are obtained. Therefore, the sample states of the m first sample sets are spliced ​​together to obtain a state set O', which can be expressed as O'={o1, o2, ..., o m}, concatenate the m first parameters to obtain the first parameter set R Z ', can be expressed as R Z '={R1,R2,...,R m}, the action set A' can be obtained by splicing the screening actions of the candidate sample set by m sub-processes, which can be expressed as A'={A1,A2,...,A m}, based on the state set O', the first parameter set R Z' and action set A' are used to train the target filter of the main process, and the training method can use the Policy Gradient algorithm.

[0119] In each round of training, the target filter traverses all samples in the candidate sample set in evaluation mode to filter and obtain the second sample set.

[0120] In the case of constructing a candidate sample set of positive samples and unlabeled samples, the target filter acts as a sample knowledge migrator. At this time, the second sample set includes the filtered positive sample set and the filtered negative sample set.

[0121] Subsequently, the target click rate prediction model of the main process is trained based on the second sample set and the known positive sample set; when the target click rate prediction model meets the training convergence condition, the training of the target click rate prediction model is ended.

[0122] In some embodiments, the training convergence conditions of the target click-through rate prediction model are the same as those described above and are not described in detail here.

[0123] In addition, in the disclosed embodiment, after training the target filter of the main process, the model parameters of the target filter are synchronously updated to the filter copies of each child process. This facilitates the child process to use the appropriate filter copy to filter samples in the next round of training.

[0124] Similarly, after training the target click-through rate prediction model of the main process, the model parameters of the target click-through rate prediction model are synchronously updated to the click-through rate prediction model copies of each child process, so that the child process can be rewarded based on the appropriate click-through rate prediction model copy.

[0125] In some embodiments, when the target click-through rate prediction model does not meet the training convergence condition, the process returns to the step of controlling multiple sub-processes executed in parallel to screen out first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set. The specific process is as follows: Figure 4 As shown, until the target click-through rate prediction model meets the training convergence conditions.

[0126] Based on the same technical concept, the embodiment of the present disclosure also provides a resource recommendation method, which is applied to the click rate prediction model trained by the above method, such as Figure 5 Shown, including:

[0127] S501 : Responding to a recommendation request for a target user, obtaining candidate resources.

[0128] The recommendation request is the resource type that the target user is interested in, and the candidate resource is the sample set filtered out by the target filter based on the recommendation request.

[0129] S502: Input the candidate resources into a click-through rate prediction model to obtain the click-through rate of the candidate resources by the target users.

[0130] S503 : When it is determined based on the click rate of the candidate resource that the candidate resource is to be recommended to the target user, the candidate resource is sent to the target user.

[0131] In the disclosed embodiment, candidate resources recommended to target users based on a trained click-through rate prediction model can more accurately identify truly effective resources, thereby improving the actual performance of click-through rate prediction and improving the user experience.

[0132] Based on the same technical concept, the embodiment of the present disclosure also provides a training device for a click-through rate prediction model, such as Figure 6 Shown, including:

[0133] A parallel execution module 601 is used to control multiple parallel execution sub-processes to respectively select first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set;

[0134] A reward determination module 602 is configured to aggregate the first parameters corresponding to each first sample set to obtain a first parameter set;

[0135] A first training module 603 is configured to train a target filter of the main process based on the sample status of each first sample set, the first parameter set, and the screening action of each sub-process on the candidate sample set;

[0136] A screening module 604 is configured to screen out a second sample set from the candidate sample set based on a target filter;

[0137] A second training module 605 is configured to train a target click rate prediction model of the main process based on the second sample set and the known positive sample set;

[0138] Ending module 606 is used to end the training of the target click-through rate prediction model when the target click-through rate prediction model meets the training convergence condition.

[0139] In some embodiments, each sub-process has a corresponding filter copy and click-through rate prediction model copy, and the parallel execution module 601 is configured to perform the following operations on each sub-process:

[0140] Controlling the subprocess to filter out a first sample set from the candidate sample set when the filter copy of the subprocess is in an exploration mode;

[0141] The control sub-process inputs the first sample set and the known positive sample set into the click rate prediction model copy of the sub-process to obtain the first parameter corresponding to the first sample set.

[0142] In some embodiments, an update module is further included for:

[0143] After training the target filter of the main process, the model parameters of the target filter are updated synchronously to the filter copies of each child process; and

[0144] After the target click-through rate prediction model of the main process is trained, the model parameters of the target click-through rate prediction model are synchronously updated to the click-through rate prediction model copies of each sub-process.

[0145] In some embodiments, an iteration module is further included for:

[0146] If the target click-through rate prediction model does not meet the training convergence conditions, the process returns to the step of controlling multiple parallel sub-processes to screen out first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set until the target click-through rate prediction model meets the training convergence conditions.

[0147] In some embodiments, for each first sample set, a first parameter corresponding to the first sample set is obtained, and the parallel execution module 601 is configured to:

[0148] Determining a first sub-parameter of each target sample based on a predicted click-through rate value of each target sample in the first sample set;

[0149] Determine a second sub-parameter based on a model evaluation result of the target click-through rate prediction model;

[0150] A first parameter corresponding to the first sample set is determined based on the first sub-parameter and the second sub-parameter of each target sample.

[0151] In some embodiments, the candidate sample set includes positive samples and unlabeled samples; the first sample set includes screened positive samples and screened negative samples; the parallel execution module 601 is used to:

[0152] For each target sample, based on the click rate prediction value of the target sample, the first sub-parameter of the target sample is determined, including:

[0153] Determine the first sub-parameter of the target sample based on the specified principle;

[0154] Wherein, when the target sample is a sample screened from unlabeled samples, the specified principle is used: when the first determination result of the target sample based on the click-through rate prediction value is consistent with the second determination result represented by the screening action, the first sub-parameter is a positive value; when the first determination result and the second determination result are inconsistent, the first sub-parameter is a negative value; the first determination result and the second determination result are used to indicate whether the target sample is a positive sample or a negative sample;

[0155] In the case where the target sample is a positive sample screened out from the positive samples, the specified principle is used to set the first sub-parameter of the target sample to a default value.

[0156] In some embodiments, the first sub-parameter of the target sample is determined based on the following first sub-parameter determination formula that satisfies specified criteria:

[0157]

[0158] Among them, clip() represents the truncation function, which means as well as The value of is limited to the range of (-1, 1), c i Represents the prediction result of the click-through rate prediction model, o i Represents the filter result of a filter.

[0159] In some embodiments, the candidate sample set includes unlabeled negative samples; the first sample set includes the screened negative samples;

[0160] For each target sample, based on the click rate prediction value of the target sample, the first sub-parameter of the target sample is determined, and the module 601 is executed in parallel, for:

[0161] In a case where the target sample is determined to be a negative sample based on the click rate prediction value, determining the first sub-parameter of the target sample to be a preset positive value;

[0162] In the case where the target sample is determined to be a positive sample based on the click rate prediction value, the first sub-parameter of the target sample is determined to be a preset negative value.

[0163] In some embodiments, based on the first sub-parameter and the second sub-parameter of each target sample, the first parameter corresponding to the first sample set is determined, and the parallel execution module 601 is configured to:

[0164] The second sub-parameter and the accumulated sum of the first sub-parameters of each target sample are weighted summed to obtain the first parameter.

[0165] Based on the same technical concept, the embodiment of the present disclosure also provides a resource recommendation device, which is applied to the click rate prediction model trained by the above device, such as Figure 7 Shown, including:

[0166] The acquisition module 701 is used to acquire candidate resources in response to a recommendation request to a target user;

[0167] Confirmation module 702, used to input candidate resources into the click-through rate prediction model to obtain the click-through rate of the candidate resources by target users;

[0168] The sending module 703 is configured to send the candidate resource to the target user when it is determined based on the click rate of the candidate resource that the candidate resource is recommended to the target user.

[0169] For the description of specific functions and examples of each module and submodule of the device in the embodiment of the present disclosure, please refer to the relevant description of the corresponding steps in the above method embodiment, which will not be repeated here.

[0170] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0171] In the technical solutions disclosed herein, the acquisition, storage, and application of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0172] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0173] like Figure 8 As shown, the device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. Various programs and data required for the operation of the device 800 can also be stored in the RAM 803. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0174] Various components in device 800 are connected to I / O interface 805, including an input unit 806, such as a keyboard, mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, optical disk, etc.; and a communication unit 809, such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows device 800 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0175] The computing unit 801 can be various general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the training method of the click-through rate prediction model or the resource recommendation method. For example, in some embodiments, the training method of the click-through rate prediction model or the resource recommendation method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the training method of the click-through rate prediction model or the resource recommendation method described above can be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to execute the training method of the click-through rate prediction model or the resource recommendation method in any other appropriate manner (for example, by means of firmware).

[0176] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0177] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0178] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0179] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0180] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0181] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain. In the embodiments of the present disclosure, a server may be used to execute the protein generation method.

[0182] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0183] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A training method for a click-through rate prediction model, comprising: Controlling multiple sub-processes executed in parallel to filter out first sample sets from the candidate sample set based on their respective filter copies and obtaining first parameters corresponding to each first sample set; the first parameters include a reward for a single sample; Summarizing the first parameters corresponding to the first sample sets to obtain a first parameter set; Based on the sample states of each first sample set, the first parameter set, and the screening actions of each child process on the candidate sample set, a target filter of the main process is trained; wherein, for any two screening actions with different reward sizes, the target filter is trained by increasing the probability of the screening action with the larger reward and decreasing the probability of the screening action with the smaller reward; wherein the target filter is used to update the filter copy in each child process; Filtering a second sample set from the candidate sample set based on the target filter; Training the target click rate prediction model of the main process based on the second sample set and the known positive sample set; When the target click-through rate prediction model meets the training convergence condition, the training of the target click-through rate prediction model is terminated.

2. The method according to claim 1, wherein Each sub-process has a corresponding copy of the click-through rate prediction model. The sub-processes that control the multiple parallel executions respectively select first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set, including: Perform the following operations for each child process: controlling the subprocess to filter out a first sample set from the candidate sample set when the filter copy of the subprocess is in an exploration mode; The sub-process is controlled to input the first sample set and the known positive sample set into the click rate prediction model copy of the sub-process to obtain a first parameter corresponding to the first sample set.

3. The method according to claim 1 or 2, further comprising: After training the target filter of the main process, the model parameters of the target filter are synchronously updated to the filter copies of each sub-process; as well as, After the target click-through rate prediction model of the main process is trained, the model parameters of the target click-through rate prediction model are synchronously updated to the click-through rate prediction model copies of each sub-process.

4. The method according to any one of claims 1 to 3, further comprising: If the target click-through rate prediction model does not meet the training convergence condition, return to the step of controlling multiple parallel execution sub-processes to filter out first sample sets from the candidate sample sets and obtain first parameters corresponding to each first sample set until the target click-through rate prediction model meets the training convergence condition.

5. The method according to any one of claims 1 to 4, wherein For each first sample set, obtaining a first parameter corresponding to the first sample set includes: Determining a first sub-parameter of each target sample based on the click-through rate prediction value of each target sample in the first sample set; the first sub-parameter reflects the click-through rate prediction of the target sample; Determining a second sub-parameter based on a model evaluation result of the target click-through rate prediction model; the second sub-parameter is used to evaluate a prediction effect of the target click-through rate prediction model; Based on the first sub-parameter and the second sub-parameter of each target sample, a first parameter corresponding to the first sample set is determined.

6. The method according to claim 5, wherein: The candidate sample set includes positive samples and unlabeled samples; the first sample set includes screened positive samples and screened negative samples; For each target sample, based on the click rate prediction value of the target sample, determining a first sub-parameter of the target sample includes: Determining a first sub-parameter of the target sample based on a specified principle; Wherein, when the target sample is a sample screened out from unlabeled samples, the specified principle is used to: when a first determination result of the target sample based on the click-through rate prediction value is consistent with a second determination result represented by a screening action, the first sub-parameter is a positive value; when the first determination result and the second determination result are inconsistent, the first sub-parameter is a negative value; the first determination result and the second determination result are used to indicate whether the target sample is a positive sample or a negative sample; In a case where the target sample is a positive sample screened out from positive samples, the specified principle is used to set the first sub-parameter of the target sample to a default value.

7. The method according to claim 6, wherein the first sub-parameter expression that satisfies the specified principle comprises: Among them, clip() represents the truncation function, which means as well as The value of is limited to the range of (-1, 1), c i Represents the prediction result of the click-through rate prediction model, o i Represents the filter result of a filter.

8. The method according to claim 5, wherein The candidate sample set includes unlabeled negative samples; the first sample set includes screened negative samples; For each target sample, based on the click rate prediction value of the target sample, determining a first sub-parameter of the target sample includes: In a case where the target sample is determined to be a negative sample based on the click rate prediction value, determining the first sub-parameter of the target sample to be a preset positive value; In a case where the target sample is determined to be a positive sample based on the click rate prediction value, the first sub-parameter of the target sample is determined to be a preset negative value.

9. The method according to claim 5, wherein: The determining, based on the first sub-parameter and the second sub-parameter of each target sample, a first parameter corresponding to the first sample set includes: The first parameter is obtained by performing weighted sum processing on the accumulated sum of the second sub-parameter and the first sub-parameters of each target sample.

10. A resource recommendation method, using a click-through rate prediction model trained by the method according to any one of claims 1 to 9, comprising: Responding to a recommendation request for a target user, obtaining candidate resources; Inputting the candidate resource into the click-through rate prediction model to obtain the click-through rate of the candidate resource by the target user; In a case where it is determined based on the click rate of the candidate resource that the candidate resource is recommended to the target user, the candidate resource is sent to the target user.

11. A training device for a click-through rate prediction model, comprising: A parallel execution module is configured to control multiple sub-processes to be executed in parallel, each sub-process having a corresponding filter copy to filter out first sample sets from the candidate sample set and obtain first parameters corresponding to each first sample set; the first parameters include a reward for a single sample; a reward determination module, configured to aggregate the first parameters corresponding to the first sample sets to obtain a first parameter set; a first training module, configured to train a target filter of the main process based on the sample states of each first sample set, the first parameter set, and the screening actions of each sub-process on the candidate sample set; wherein, for any two screening actions with different reward sizes, the target filter is trained by increasing the probability of the screening action with the larger reward and decreasing the probability of the screening action with the smaller reward; wherein the target filter is used to update the filter copy in each sub-process; a screening module, configured to screen out a second sample set from the candidate sample set based on the target filter; A second training module is used to train the target click rate prediction model of the main process based on the second sample set and the known positive sample set; The ending module is used to end the training of the target click-through rate prediction model when the target click-through rate prediction model meets the training convergence condition.

12. The device according to claim 11, wherein Each sub-process has a corresponding copy of the click-through rate prediction model. The parallel execution module is used to perform the following operations on each sub-process: controlling the subprocess to filter out a first sample set from the candidate sample set when the filter copy of the subprocess is in an exploration mode; The sub-process is controlled to input the first sample set and the known positive sample set into the click rate prediction model copy of the sub-process to obtain a first parameter corresponding to the first sample set.

13. The apparatus according to claim 11 or 12, further comprising: An update module is used to synchronously update the model parameters of the target filter of the main process to the filter copies of each sub-process after training the target filter of the main process; And, after the target click-through rate prediction model of the main process is trained, the model parameters of the target click-through rate prediction model are synchronously updated to the click-through rate prediction model copies of each sub-process.

14. The apparatus according to any one of claims 11 to 13, further comprising: The iterative module is used to return to the step of controlling multiple parallel execution sub-processes to screen out first sample sets from candidate sample sets and obtain first parameters corresponding to each first sample set when the target click-through rate prediction model does not meet the training convergence condition, until the target click-through rate prediction model meets the training convergence condition.

15. The device according to any one of claims 11 to 14, wherein: For each first sample set, a first parameter corresponding to the first sample set is obtained, and the parallel execution module is configured to: Determining a first sub-parameter of each target sample based on the click-through rate prediction value of each target sample in the first sample set; the first sub-parameter reflects the click-through rate prediction of the target sample; Determining a second sub-parameter based on a model evaluation result of the target click-through rate prediction model; the second sub-parameter is used to evaluate a prediction effect of the target click-through rate prediction model; Based on the first sub-parameter and the second sub-parameter of each target sample, a first parameter corresponding to the first sample set is determined.

16. The device according to claim 15, wherein The candidate sample set includes positive samples and unlabeled samples; the first sample set includes screened positive samples and screened negative samples; the parallel execution module is used to determine, for each target sample, the first sub-parameter of the target sample based on the following method: Determining a first sub-parameter of the target sample based on a specified principle; Wherein, when the target sample is a sample screened out from unlabeled samples, the specified principle is used to: when a first determination result of the target sample based on the click-through rate prediction value is consistent with a second determination result represented by a screening action, the first sub-parameter is a positive value; when the first determination result and the second determination result are inconsistent, the first sub-parameter is a negative value; the first determination result and the second determination result are used to indicate whether the target sample is a positive sample or a negative sample; In a case where the target sample is a positive sample screened out from positive samples, the specified principle is used to set the first sub-parameter of the target sample to a default value.

17. The apparatus according to claim 16, wherein the first sub-parameter expression satisfying the specified principle comprises: Among them, clip() represents the truncation function, which means as well as The value of is limited to the range of (-1, 1), c i Represents the prediction result of the click-through rate prediction model, o i Represents the filter result of a filter.

18. The device according to claim 15, wherein The candidate sample set includes unlabeled negative samples; the first sample set includes screened negative samples; the parallel execution module is used to: In a case where the target sample is determined to be a negative sample based on the click rate prediction value, determining the first sub-parameter of the target sample to be a preset positive value; In a case where the target sample is determined to be a positive sample based on the click rate prediction value, the first sub-parameter of the target sample is determined to be a preset negative value.

19. The device according to claim 15, wherein The parallel execution module is used to: The first parameter is obtained by performing weighted sum processing on the accumulated sum of the second sub-parameter and the first sub-parameters of each target sample.

20. A resource recommendation device, using a click-through rate prediction model trained by the device according to any one of claims 11 to 19, comprising: An acquisition module, configured to acquire candidate resources in response to a recommendation request to a target user; A prediction module, configured to input the candidate resource into the click-through rate prediction model to obtain the click-through rate of the candidate resource by the target user; The recommendation module is configured to send the candidate resource to the target user when it is determined based on the click rate of the candidate resource that the candidate resource is recommended to the target user.

21. An electronic device comprising: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 10.

22. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-10.

23. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 10.

Images