Resource recommendation method and device, electronic equipment and medium

Abstract

The invention provides a resource recommendation method and device, electronic equipment and a medium, and relates to the field of data processing, in particular to the technical field of cloud. The specific implementation scheme is as follows: according to user feature data of a target user, resource feature data of candidate resources, current scene feature data, a pre-trained general scene model and a target standby scene model, obtaining a first proportion parameter of a candidate behavior operation probability generated by the target user on the candidate resources, the standby proportion parameter of the candidate behavior operation probability generated by the target user on the candidate resources on the target standby scene is obtained; determining a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter; and determining a recommendation result of the candidate resources according to the candidate behavior operation probability and the fusion proportion parameter corresponding to the candidate behavior operation probability. According to the scheme, the deviation caused by insufficient learning of the target standby scene in the general scene model is corrected, and the accuracy of user resource recommendation in different scenes is improved.

Description

Technical Field

[0001] The present disclosure relates to the field of information recommendation, in particular to the field of cloud technology, and specifically to a resource recommendation method, device, electronic device and medium. Background Art

[0002] Feed stream refers to an information stream that is continuously updated and presented to users. In the feed stream recommendation system, users have different preferences for resource types in different scenarios, such as the number of immersive refreshes, the time range of entry, and the network status. However, there is an uneven distribution of samples in different scenarios. The existing technical solution inputs all scene samples into a separate network model. Considering the sample distribution problem, there may be insufficient learning for scenes with fewer samples, and sufficient training cannot be performed. In actual use, it may be affected by the majority of samples, resulting in deviations in resource recommendation predictions, affecting the user experience. Summary of the invention

[0003] The present disclosure provides a method, device, electronic device and medium for resource recommendation.

[0004] According to one aspect of the present disclosure, a resource recommendation method is provided, comprising:

[0005] Inputting the user feature data of the target user, the resource feature data of the candidate resource and the current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource; wherein the general scenario model is obtained by training the network parameters by superimposing noise disturbance;

[0006] According to the user characteristic data of the target user, the resource characteristic data of the candidate resource, the current scenario characteristic data and the pre-trained target backup scenario model, a backup proportion parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario is obtained; wherein the target backup scenario model is obtained by training the network parameters by superimposing noise disturbance;

[0007] Determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter;

[0008] The recommendation result of the candidate resource is determined according to the candidate behavior operation probability and a fusion ratio parameter corresponding to the candidate behavior operation probability.

[0009] According to another aspect of the present disclosure, a method for training a general scene model is provided, comprising:

[0010] Determine the common scenario historical behavior data of sample users;

[0011] Inputting the general scenario historical behavior data into the general scenario model to obtain general initial network parameters;

[0012] Superimposing noise disturbance on the general initial network parameters to obtain a general scene model after disturbance;

[0013] Determine the general prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource output by the disturbed general scenario model;

[0014] recommending sample resources to the sample users by using the disturbed general scenario model, and determining general real behavior feedback data of the sample users on the sample resources;

[0015] The general scenario model is updated according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resource and the general real behavior feedback data to obtain a trained general scenario model.

[0016] According to another aspect of the present disclosure, a method for training a candidate backup scene model is provided, comprising:

[0017] Determine the common scenario historical behavior data of the sample users, and filter the candidate backup scenario historical behavior data from the common scenario historical behavior data according to the sample quantity ratio;

[0018] Inputting the candidate backup scenario historical behavior data into the candidate backup scenario model to obtain backup initial network parameters;

[0019] Superimposing noise disturbance on the backup initial network parameters to obtain a disturbed candidate backup scenario model;

[0020] Determine the standby prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource under the candidate standby scenario output by the disturbed candidate standby scenario model;

[0021] recommending sample resources to the sample user by using the disturbed candidate backup scenario model, and determining backup real behavior feedback data of the sample user on the sample resource in the candidate backup scenario;

[0022] The candidate backup scenario model is updated according to the backup prediction ratio of the probability of the sample user generating the candidate behavior operation on the sample resource in the candidate backup scenario and the backup real behavior feedback data to obtain a trained candidate backup scenario model.

[0023] According to another aspect of the present disclosure, there is provided a resource recommendation device, comprising:

[0024] A first proportion parameter determination module is used to input user feature data of a target user, resource feature data of a candidate resource, and current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user generates a candidate behavior operation on the candidate resource; wherein the general scenario model is obtained by training the network parameters by superimposing noise disturbances;

[0025] The standby ratio parameter determination module obtains the standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario according to the user characteristic data of the target user, the resource characteristic data of the candidate resource, the current scenario characteristic data and the pre-trained target standby scenario model; wherein the target standby scenario model is obtained by superimposing noise disturbance training on the network parameters;

[0026] A fusion proportion parameter determination module, used to determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter;

[0027] The resource recommendation module is used to determine the recommendation result of the candidate resource according to the candidate behavior operation probability and the fusion proportion parameter corresponding to the candidate behavior operation probability.

[0028] According to another aspect of the present disclosure, a training device for a general scenario model is provided, comprising:

[0029] A general sample data determination module is used to determine the general scenario historical behavior data of sample users;

[0030] A general initial parameter determination module, used for inputting the general scenario historical behavior data into the general scenario model to obtain general initial network parameters;

[0031] A first disturbance module, used for superimposing noise disturbance on the general initial network parameters to obtain a general scenario model after disturbance;

[0032] A first prediction module, used to determine the general prediction proportion of the probability of the sample user generating a candidate behavior operation on the sample resource output by the general scenario model after the disturbance;

[0033] A first feedback module, configured to recommend sample resources to the sample users by using the perturbed general scenario model, and determine general real behavior feedback data of the sample users on the sample resources;

[0034] The general scenario model updating module is used to update the general scenario model according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resources and the general real behavior feedback data to obtain a trained general scenario model.

[0035] According to another aspect of the present disclosure, a training device for a candidate backup scene model is provided, comprising:

[0036] A spare sample data determination module is used to determine the common scenario historical behavior data of the sample user, and to filter the candidate spare scenario historical behavior data from the common scenario historical behavior data according to the sample quantity ratio;

[0037] A standby initial parameter determination module, used for inputting the candidate standby scenario historical behavior data into the candidate standby scenario model to obtain standby initial network parameters;

[0038] A second disturbance module is used to superimpose noise disturbance on the backup initial network parameters to obtain a disturbed candidate backup scenario model;

[0039] A second prediction module is used to determine the standby prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource under the candidate standby scenario output by the disturbed candidate standby scenario model;

[0040] A second feedback module is used to recommend sample resources to the sample user by using the disturbed candidate backup scenario model, and determine the sample user's backup real behavior feedback data on the sample resource in the candidate backup scenario;

[0041] The backup scenario model updating module is used to update the candidate backup scenario model according to the backup prediction proportion of the sample user's candidate behavior operation probability on the sample resource in the candidate backup scenario and the backup real behavior feedback data to obtain a trained candidate backup scenario model.

[0042] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0043] at least one processor; and

[0044] a memory communicatively connected to the at least one processor; wherein,

[0045] 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 so that the at least one processor can execute the resource recommendation method, the training method of the general scene model, or the training method of the candidate backup scene model described in any embodiment of the present disclosure.

[0046] 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 resource recommendation method, the general scenario model training method, or the candidate backup scenario model training method described in any embodiment of the present disclosure.

[0047] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which, when executed by a processor, implements the resource recommendation method, the general scenario model training method, or the candidate backup scenario model training method according to any embodiment of the present disclosure.

[0048] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0050] Figure 1 is a schematic diagram of a resource recommendation method according to an embodiment of the present disclosure;

[0051] Figure 2 is a schematic diagram of an architecture of a process for determining a recommendation result of a candidate resource according to an embodiment of the present disclosure;

[0052] Figure 3 is a schematic diagram of another resource recommendation method according to an embodiment of the present disclosure;

[0053] Figure 4 is a schematic diagram of another resource recommendation method according to an embodiment of the present disclosure;

[0054] Figure 5 is a schematic diagram of a training method for a general scene model according to an embodiment of the present disclosure;

[0055] Figure 6 is a schematic diagram of a method for training a candidate backup scene model according to an embodiment of the present disclosure;

[0056] Figure 7 is a structural schematic diagram of a resource recommendation device according to an embodiment of the present disclosure;

[0057] Figure 8 is a structural schematic diagram of a training device for a general scenario model according to an embodiment of the present disclosure;

[0058] Fig. 9 is a structural schematic diagram of a training device for a candidate backup scene model according to an embodiment of the present disclosure;

[0059] Fig.10 It is a block diagram of an electronic device used to implement the resource recommendation method, the training method of a general scenario model, or the training method of a candidate backup scenario model of the embodiment of the present disclosure. DETAILED DESCRIPTION

[0060] The following is a description of exemplary embodiments of the present disclosure 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 recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit 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.

[0061] Figure 1 is a schematic diagram of a resource recommendation method according to an embodiment of the present disclosure. This embodiment is applicable to the case of optimizing the resource recommendation method. The method can be executed by a resource recommendation device, which can be implemented by software and / or hardware and integrated in an electronic device. The electronic device involved in this embodiment can be a background server with computing capabilities, etc. Figure 1 , the method specifically includes the following:

[0062] S110, inputting the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource.

[0063] Among them, the target user refers to the user who has the demand for feed stream recommendation, and the user characteristic data includes information that can characterize the user's resource preference, such as the user characteristic data including the user's age, gender and the user's historical viewing behavior data; the candidate resource refers to the resource waiting to be recommended to the user, such as short videos, etc. The candidate resource can be preliminarily screened according to the historical viewing record of the target user, and the resource characteristic data of the candidate resource includes information that can characterize the resource type and content, such as the resource characteristic data including the resource duration and the resource type, etc.; the current scene characteristic data is used for the external attribute information of the user's current viewing feed stream, such as the current scene characteristic data including the viewing time, such as morning, noon or evening, whether it is the first viewing, such as the first brush or non-first brush, the network environment, such as WIFI or non-WIFI, and the current network transmission speed, the system of the user's viewing device, such as iOS or Android, etc., and the user's preference for the resource content of the viewing feed stream in different scenarios will be different. Among them, the user characteristic data is obtained on the basis that the user knows and agrees.

[0064] The candidate behavior operation probability refers to the probability that the target user will perform a certain behavior on the candidate resource. The candidate behavior operation probability may include the probabilities corresponding to multiple candidate behavior operations, such as the probability that the target user will click on the candidate resource, the probability of watching the whole video, the probability of liking the video, or the probability of frequency of the video. The candidate behavior operation probability is the prediction target of the target user's behavior when watching the candidate resource. It can also be said to be a number of target ranking influencing factors generated in the resource ranking stage. The candidate behavior operation probability can be determined by the user's historical viewing behavior, or by a pre-trained behavior prediction model. The behavior prediction model can be trained by the historical behavior sample set of the sample user. The method for determining the candidate behavior operation probability is not limited here.

[0065] The general scenario model is pre-trained based on the training sample set of sample users, wherein the training sample set includes data in all scenarios. The training input of the model is the user feature data of the sample user, the resource feature data of the historical resources, the corresponding historical scenario feature data, and the label ratio parameter of the probability of the sample user generating a candidate behavior operation on the historical resources, and the output is the predicted ratio parameter. The general scenario model learns the relationship between the user feature data of the sample user, the resource feature data of the historical resources, the corresponding historical scenario feature data, and the label ratio parameter of the probability of the sample user generating a candidate behavior operation on the historical resources, and predicts the first ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource through learning the user feature data of the target user, the resource feature data of the candidate resource, and the current scenario feature data according to the relationship.

[0066] The general scenario model is obtained by superimposing noise perturbations on the network parameters for training. That is, during the training process, the general scenario model superimposes noise on the network parameters of the model, and determines the training direction of the network parameters of the model according to the output results of the model after superimposing the noise.

[0067] The first proportion parameter is used to characterize the influence of the candidate behavior operation probability in recommending candidate resources to the target user. As the first proportion parameter increases, it means that the influence of the corresponding candidate behavior operation probability of the target user on the candidate resources on the recommendation result increases, that is, the corresponding candidate behavior operation probability increases the influence of the ranking result of the candidate resources.

[0068] Specifically, the user feature data of the target user, the resource feature data of the candidate resources, and the current scene feature data are obtained, and the above data are input into a general scene model. The model is trained based on the user feature data of the sample user, the resource feature data of the historical resources, and the corresponding historical scene feature data as input. The output of the model is the first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource.

[0069] Exemplarily, when the target user is in a wifi environment, on an iOS system, and just enters the video recommendation list, the features input into the general scenario model are [user feature 1, user feature 2..., resource feature 1, resource feature 2..., wifi environment, iOS, first time viewing, etc.], and the output of the model is the first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource.

[0070] S120, obtaining a standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario based on the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and a pre-trained target standby scenario model.

[0071] Among them, the target backup scenario model is obtained by superimposing noise disturbance training on the network parameters, that is, the target backup scenario model is obtained by superimposing noise on the network parameters of the model during the training process, and the training direction of the network parameters of the model is determined according to the output results of the model after superimposing noise.

[0072] The target backup scenario model is trained with the training sample set of sample users in the target backup scenario. Since the training sample set of the collected sample users has the problem of uneven distribution of samples in different scenarios, and since some scenarios have natural collection disadvantages, resulting in a small number of samples, in the general scenario model, there is a problem of insufficient learning for these unevenly distributed scenarios, which may be affected by the majority of samples in actual use, and thus there is an inaccurate prediction phenomenon for these scenarios. In this embodiment, separate training is performed for these unevenly distributed scenarios, and the samples of each scenario are used to guide the parameter evolution direction of each scenario, thereby improving the adequacy of learning for unevenly distributed scenarios and improving the prediction accuracy for unevenly distributed scenarios. The target backup scenario is an unevenly distributed scenario. When training the target backup scenario model, first, the training sample set of the target backup scenario is screened from the training sample set of all sample users, and then the network parameters are trained using these training sample sets. The training input of the target backup scenario model is the user feature data of the sample user in the target backup scenario, the resource feature data of the historical resources, the corresponding historical scenario feature data, and the label ratio parameter of the probability of the sample user generating a candidate behavior operation on the historical resources, and the output is the predicted ratio parameter. The target backup scenario model learns the relationship between the user feature data of the sample user in the target backup scenario, the resource feature data of the historical resources, the corresponding historical scenario feature data, and the label ratio parameter of the probability of the sample user generating a candidate behavior operation on the historical resources, so as to predict the backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resources in the target backup scenario through learning the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data according to the relationship.

[0073] The backup ratio parameter is used to characterize the influence of the candidate behavior operation probability on the target backup scenario in recommending candidate resources to the target user. If the backup ratio parameter increases, it means that the influence of the corresponding candidate behavior operation probability of the target user on the candidate resources on the recommendation result in the target backup scenario increases, that is, the corresponding candidate behavior operation probability in the target backup scenario has an increased impact on the ranking result of the candidate resources.

[0074] Specifically, the user characteristic data of the target user, the resource characteristic data of the candidate resources, and the current scene characteristic data are obtained, and the above data are input into the target backup scene model. The model is trained based on the user characteristic data of the sample users in the target backup scene, the resource characteristic data of the historical resources, and the corresponding historical scene characteristic data as input. The output of the model is the backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scene.

[0075] Exemplarily, when the target user is in a wifi environment, an ios system, and has just entered the video recommendation list, the features input into the target backup scenario model are [user feature 1, user feature 2..., resource feature 1, resource feature 2..., wifi environment, ios, first time viewing, etc.], and the output of the model is the backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario. Optionally, since the target backup scenario model is an auxiliary scenario training model, its main function is to correct the general scenario model. Therefore, in order to reduce the time consumption of online recommendations, the input of the target backup scenario model can be simplified. The specific number and type of simplification can be determined based on the training effect. The input simplification content of the target backup scenario model is not restricted here.

[0076] Optionally, there may be at least two target backup scenario models corresponding to the target user. Since there may be multiple scenes with uneven distribution during training, it is necessary to train corresponding backup scenario models for multiple scenes with uneven distribution respectively. Therefore, the current scene feature data of the target user may correspond to multiple backup scenarios at the same time. In this case, it is necessary to input the user feature data of the target user, the resource feature data of the candidate resources and the current scene feature data into at least two target backup scenario models corresponding to successful matching respectively, and obtain the backup proportion parameters of the probability of at least two target users generating candidate behavior operations on the candidate resources in the target backup scenarios respectively.

[0077] S130: Determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter.

[0078] Since the first proportion parameter represents the influence of the target user on the probability of candidate behavior operations on candidate resources in all scenarios, the backup proportion parameter represents the influence of the target user on the probability of candidate behavior operations on candidate resources in the target backup scenario. In order to correct the deviation of insufficient learning of the target backup scenario in the first proportion parameter obtained by the general scenario model, the backup proportion parameter obtained by the target backup scenario model is applied back to the whole, thereby enhancing the modeling accuracy corresponding to the target backup scenario and optimizing the fusion effect of the target backup scenario.

[0079] Specifically, the result of directly multiplying the first proportion parameter and the spare proportion parameter can be used as the fusion proportion parameter, for example, w=wgeneral*wspare, where w represents the fusion proportion parameter, wgeneral represents the first proportion parameter, and wspare represents the spare proportion parameter. Optionally, there are multiple candidate behavior operation probabilities, namely qi, which represents the i-th candidate behavior operation probability, and each candidate behavior operation probability corresponds to a first proportion parameter wigeneral and at least one spare proportion parameter wispare. The number of spare proportion parameters is determined according to the number of target spare scenario models matched by the target user, and the fusion proportion parameter corresponding to the i-th candidate behavior operation probability is wi=wigeneral*wispare.

[0080] S140: Determine a recommendation result of the candidate resource according to the candidate behavior operation probability and a fusion ratio parameter corresponding to the candidate behavior operation probability.

[0081] Specifically, the weighted sum of the candidate behavior operation probability and the fusion ratio parameter corresponding to the candidate behavior operation probability can be used as the recommendation probability of the candidate resource, and the recommendation ranking result of the candidate resource can be determined based on the recommendation probability. Alternatively, the multiplication result of multiple candidate behavior operation probabilities and the fusion ratio parameter corresponding to the candidate behavior operation probabilities can be used as the recommendation probability of the candidate resource, for example, F=q1^w1*q2^w2*...*qn^wn, where F represents the recommendation probability of the candidate resource, qn represents the nth candidate behavior operation probability, and wn represents the fusion ratio parameter corresponding to the nth candidate behavior operation probability. By multiplying, it is convenient to add new factors and adjust the strength of some factors as the business direction and goals change. This factor represents the candidate behavior operation probability. Figure 2 The figure shows a schematic diagram of the process architecture for determining the recommendation results of candidate resources.

[0082] The solution of this embodiment obtains the backup proportion parameters of the probability of the target user generating candidate behavior operations on candidate resources in the target backup scenario through the target backup scenario model, and merges the backup proportion parameters with the general proportion parameters obtained by the general scenario model, so as to correct the deviation caused by insufficient learning of the target backup scenario in the first proportion parameters obtained by the general scenario model, improve the ability to analyze users' resource preferences in different scenarios, and improve the accuracy of recommending candidate resources to target users.

[0083] Figure 3 is a schematic diagram of another resource recommendation method according to an embodiment of the present disclosure. This embodiment is a further refinement of step S120 in the above technical solution, and further refines the input of the target backup scenario model. The technical solution in this embodiment can be combined with each optional solution in one or more of the above embodiments. Figure 3 As shown, the resource recommendation method includes the following:

[0084] S210: Input the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource.

[0085] S220 , screening the user characteristic data of the target user, the resource characteristic data of the candidate resources, and the current scenario characteristic data according to the target backup scenario corresponding to the target backup scenario model to obtain target characteristic data associated with the target backup scenario.

[0086] According to the target standby scenario corresponding to the target standby scenario model, relevant data with a high degree of relevance are selected from the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data, and the relevant data are input into the target standby scenario model. The degree of relevance between the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data and the target standby scenario can be obtained in advance through correlation analysis, and the method for determining the degree of relevance of the relevant data is not limited here.

[0087] S230: Input the target feature data into the target backup scenario model to obtain a backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario.

[0088] The filtered target feature data is input into the pre-trained target backup scenario model, and the output result is the backup ratio parameter of the target user's probability of generating candidate behavior operations on candidate resources in the target backup scenario.

[0089] Similarly, when training the target backup scenario model, the training input data also needs to be screened according to the screening method.

[0090] In another optional implementation of this embodiment, before S220, the following is further included:

[0091] The corresponding target backup scene model is determined from pre-trained candidate backup scene models according to the current scene feature data.

[0092] Among them, the candidate backup scene model is a model obtained by training the scene samples with uneven distribution problems determined according to the training sample set of the sample users. Specifically, during the initial training, statistics are performed according to the scenes based on the training sample sets of the sample users collected, and the sample size corresponding to each scene is determined respectively, and the candidate backup scenes with uneven distribution are determined according to the sample size. In another optional implementation of this embodiment, the proportion of the training sample size of the candidate backup scene in the total training sample size is less than the preset proportion threshold. That is, scenes with a small sample size are selected as candidate backup scenes. The corresponding candidate backup scene models are obtained by training according to the training sample sets corresponding to the candidate backup scenes. Separate training is performed for these unevenly distributed scenes, and the samples of each scene are used to guide the parameter evolution direction of each scene, thereby improving the adequacy of learning the unevenly distributed scenes and improving the prediction accuracy of the unevenly distributed scenes.

[0093] Specifically, the candidate backup scene corresponding to the candidate backup scene model is matched according to the current scene feature data to determine a successfully matched target backup scene and a corresponding target backup scene model. There may be at least one target backup scene model.

[0094] Exemplarily, the current scene feature data of the target user includes [wifi environment, ios, first refresh, etc.], where ios corresponds to a candidate backup scene, and the first refresh corresponds to a candidate backup scene. The corresponding target backup scene model is determined according to the current scene feature data as the ios target backup scene model and the first refresh target backup scene model.

[0095] In another optional implementation of this embodiment, the proportion of the training sample volume of the candidate backup scene corresponding to the candidate backup scene model in the total training sample volume is less than a preset proportion threshold.

[0096] S240: Determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter.

[0097] S250: Determine a recommendation result of the candidate resource according to the candidate behavior operation probability and a fusion ratio parameter corresponding to the candidate behavior operation probability.

[0098] The solution of this embodiment obtains target feature data of the target backup scene model by screening the input data, thereby streamlining the input of the target backup scene model, reducing the calculation time of the target backup scene model, and improving the efficiency of resource recommendation.

[0099] Figure 4is a schematic diagram of another resource recommendation method according to an embodiment of the present disclosure. This embodiment further refines S130 and S240 in the above technical solution, and further refines the determination process of the fusion ratio parameter. The technical solution in this embodiment can be combined with each optional solution in one or more of the above embodiments. Figure 4 As shown, the resource recommendation method includes the following:

[0100] S310: Input the user feature data of the target user, the resource feature data of the candidate resources, and the current scenario feature data into a pre-trained general scenario model to obtain a first ratio parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource.

[0101] S320, obtaining a standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario based on the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and a pre-trained target standby scenario model.

[0102] S330: Determine a scene weight parameter corresponding to the target backup scene.

[0103] Among them, the target backup scene is a scene with uneven sample size distribution. Due to the uneven sample size distribution of the target backup scene, the sample learning of the scene in the general scenario model will be insufficient. Therefore, it is necessary to correct the deviation caused by insufficient learning of the target backup scene. When correcting the deviation, it is necessary to determine the degree of deviation correction according to the unevenness of the target backup scene. The scene weight parameter corresponding to the target backup scene is used to characterize the degree of deviation correction of the target backup scene.

[0104] Specifically, the corresponding scenario weight parameter is determined according to the deviation correction degree of the target backup scenario. Exemplarily, the corresponding scenario weight parameter, such as the sample size ratio, can be determined according to the sample distribution parameter of the target backup scenario.

[0105] In another optional implementation of this embodiment, S330 includes:

[0106] The degree of deviation correction of the target backup scenario for the general scenario is determined based on the proportion of the training sample size of the target backup scenario in the total training sample size and the scenario importance parameter of the target backup scenario; the scenario weight parameter corresponding to the target backup scenario is determined based on the degree of deviation correction.

[0107] As the proportion of the training sample size of the target backup scenario in the total training sample size increases, the sufficiency of learning the target backup scenario in the general scenario model increases, and the degree of deviation correction of the target backup scenario for the general scenario decreases; the scenario importance parameter of the target backup scenario is used to standardize the importance of the target backup scenario to the final candidate resource prediction. As the scenario importance parameter of the target backup scenario increases, the importance of the final candidate resource prediction increases, and the degree of deviation correction of the target backup scenario for the general scenario increases. Therefore, the final degree of deviation correction is determined in combination with the proportion of the training sample size of the target backup scenario in the total training sample size and the scenario importance parameter of the target backup scenario. The specific determination method can be determined according to actual conditions. For example, a first mapping relationship between the sample size proportion interval and the degree of deviation correction is established in advance, and a second mapping relationship between the scene importance parameter and the degree of deviation correction is established. The first degree of deviation correction is determined according to the first mapping relationship, the second degree of deviation correction is determined according to the second mapping relationship, and the final degree of deviation correction is determined according to the first degree of deviation correction and the second degree of deviation correction.

[0108] Finally, the scene weight parameter is determined according to the degree of deviation correction. For example, the degree of deviation correction is directly used as the scene weight parameter, or a third mapping relationship between the degree of deviation correction and the scene weight parameter is established in advance, and the final scene weight parameter is determined according to the third mapping relationship.

[0109] The scenario weight parameters are determined by the sample size ratio and scenario importance parameters corresponding to the target backup scenario, thereby improving the accuracy of determining the scenario weight parameters of the target backup scenario and further improving the accuracy of deviation correction of the target backup scenario.

[0110] S340: Determine a target standby scene correction parameter according to a weighted value of a scene weight parameter and a standby ratio parameter.

[0111] The product of the scenario weight parameter corresponding to the target backup scenario and the corresponding backup proportion parameter is used as the target backup scenario correction parameter. The target backup scenario correction parameter represents the effect of the degree of scenario deviation correction on the probability of candidate behavior operation under the target backup scenario.

[0112] S350: Determine a fusion proportion parameter according to the target backup scene correction parameter and the first proportion parameter.

[0113] Since the first proportion parameter represents the influence of the target user on the probability of candidate behavior operations on candidate resources in all scenarios, the target backup scenario correction parameter represents the corrective influence of the target user on the probability of candidate behavior operations on candidate resources in the target backup scenario. In order to correct the insufficient learning deviation of the target backup scenario in the first proportion parameter obtained by the general scenario model, the target backup scenario correction parameter obtained by the target backup scenario model is applied back to the whole, thereby enhancing the modeling accuracy corresponding to the target backup scenario and optimizing the fusion effect of the target backup scenario.

[0114] Specifically, the result of directly multiplying the first proportion parameter and the target standby scene correction parameter can be used as the fusion proportion parameter, for example, w=wgeneral*wcorrected, where w represents the fusion proportion parameter, wgeneral represents the first proportion parameter, and wcorrected represents the target standby scene correction parameter. Optionally, there are multiple candidate behavior operation probabilities, namely qi, representing the i-th candidate behavior operation probability, each candidate behavior operation probability corresponds to a first proportion parameter wigeneral, and at least one target standby scene correction parameter wicorrected, the number of target standby scene correction parameters is determined according to the number of target standby scene models matched by the target user, then the fusion proportion parameter corresponding to the i-th candidate behavior operation probability is wi=wigeneral*wicorrected.

[0115] S360: Determine a recommendation result for the candidate resource according to the candidate behavior operation probability and a fusion ratio parameter corresponding to the candidate behavior operation probability.

[0116] The solution of this embodiment improves the accuracy of the standby ratio parameter obtained by the target standby scenario model through the scenario weight parameter, improves the accuracy of deviation correction of the target standby scenario, and further improves the accuracy of recommendation of candidate resources.

[0117] Figure 5 is a schematic diagram of a method for training a general scene model according to an embodiment of the present disclosure. This embodiment is a further refinement of the training process of the general scene model in the above technical solution. The technical solution in this embodiment can be combined with each optional solution in one or more of the above embodiments. Figure 5 As shown, the training method of the general scene model includes the following:

[0118] S410: Determine the common scenario historical behavior data of the sample user.

[0119] The sample users refer to multiple users with resource recommendation history records collected. The general scenario historical behavior data is determined based on the resource recommendation history records of the sample users. The general scenario historical behavior data includes historical behavior data in all scenarios. The general scenario historical behavior data includes user feature data of the sample users in each scenario, resource feature data of the recommended resources, and corresponding historical scenario feature data. In addition, the general scenario historical behavior data also includes corresponding tag general proportion parameters, which can be determined according to a heuristic algorithm.

[0120] S420: Input the common scenario historical behavior data into the common scenario model to obtain common initial network parameters.

[0121] The general scenario historical behavior data is input into the general scenario model for preliminary training to obtain the initial network parameters of the general scenario model. Specifically, the general scenario model can adopt a DNN network structure.

[0122] S430, superimposing noise disturbance on the general initial network parameters to obtain a general scenario model after disturbance.

[0123] Specifically, the general initial network parameters are superimposed with noise to perturb, and the general scenario model after perturbation is obtained according to the initial network parameters with superimposed noise. Exemplarily, the weight parameters and deviations of the DNN are superimposed with perturbation noise through the ES evolution strategy, and the value of the perturbation noise can be a fixed value, or formed according to a preset fixed distribution.

[0124] In another optional implementation of this embodiment, superimposing noise disturbance on the common initial network parameters includes:

[0125] Determining initial disturbance parameters according to a preset fixed distribution;

[0126] Scaling the initial disturbance parameters according to the order of magnitude of the initial network parameters to obtain the target disturbance parameters;

[0127] Superimpose the target perturbation parameters on the initial network parameters.

[0128] The preset fixed distribution may be a Gaussian distribution, etc., and is not limited thereto. The initial disturbance parameters for each superposition are determined by the preset fixed distribution, and the initial disturbance parameters are scaled according to the order of magnitude of the initial network parameter object to be superimposed, and then superimposed with the initial network parameters after the scaling is completed, and the general scenario model after the disturbance is determined according to the superimposed network parameters.

[0129] When the general scenario model is initialized to build network parameters, some features may have poor feature importance and discrimination in some scenarios during network training due to uneven sample distribution. For example, in some features, the proportion of samples from the first scenario is significantly higher than that of samples from the second scenario, resulting in extremely small weights in some of the final initialized parameters.

[0130] The initial perturbation parameters generated each time during the training process obey a fixed distribution, and the initial perturbation parameter values ​​are within a certain range. When an initial network parameter with an extremely small value is superimposed on a larger initial perturbation parameter, the impact of the initial network parameter on the predicted value will be greatly increased, resulting in fluctuations and difficulty in convergence.

[0131] Therefore, in the process of training the general scenario model, the order of magnitude of the initial network parameters is determined, and then the initial perturbation parameters are determined according to the preset fixed distribution. The initial perturbation parameters are scaled according to the order of magnitude of the initial network parameters to ensure that the initial perturbation parameters are at the same order of magnitude as the initial network parameters of the superimposed object. This avoids the problem of large training fluctuations and difficulty in network convergence due to uneven distribution of some scene feature samples during network training, thereby improving network training efficiency and model accuracy.

[0132] S440, determining the general prediction proportion of the probability of the sample user generating a candidate behavior operation on the sample resource output by the general scenario model after disturbance.

[0133] The perturbed general scenario model is used to make sample resource recommendation predictions for sample users and obtain the general prediction ratio.

[0134] Specifically, the user feature data of the sample user, the resource feature data of the sample resource, and the current scenario feature data are determined and input into the general scenario model after disturbance, and the model output is the general predicted proportion of the probability of the sample user generating a candidate behavior operation on the sample resource.

[0135] S450: Use the disturbed general scenario model to recommend sample resources to sample users, and determine the general real behavior feedback data of the sample users on the sample resources.

[0136] The perturbed general scenario model is deployed online to recommend sample resources to sample users. Based on the real-time feedback from online sample users, the general real behavior feedback data of sample users on sample resources is determined.

[0137] S460. Update the general scenario model according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resources and the general real behavior feedback data to obtain a trained general scenario model.

[0138] The general real behavior feedback data of sample users on sample resources can reflect the impact of superimposed disturbance parameters on the model prediction results. Therefore, the general scenario model is updated according to the degree of influence, and the superimposed disturbance is performed again according to the updated general scenario model. S340 to S360 are repeated until the general prediction ratio of the probability of candidate behavior operations generated by sample users on sample resources matches the general real behavior feedback data, and it is determined that the training of the general scenario model is completed.

[0139] Exemplarily, the general initial network parameters can be perturbed to various degrees at the same time to obtain multiple general scenario models after perturbation, determine the general predicted proportion of the probability of sample users generating candidate behavioral operations on sample resources output by each general scenario model after perturbation, and use each general scenario model after perturbation to recommend sample resources to sample users, determine the general real behavior feedback data of sample users on sample resources, determine the update parameters of the general scenario model based on the general predicted proportion of the probability of sample users generating candidate behavioral operations on sample resources corresponding to different degrees of perturbation and the general real behavior feedback data, and obtain a trained general scenario model.

[0140] The solution of this embodiment trains the general scenario model through the general scenario historical behavior data of sample users to achieve general resource recommendations for users in multiple scenarios; and in the model training process, by superimposing noise perturbations on the network parameters, combined with real behavior feedback data, the training accuracy and prediction accuracy of the final model are improved.

[0141] Figure 6 Schematic diagram of a training method for candidate backup scene models according to an embodiment of the present disclosure. This embodiment further refines the specific training process of the candidate backup scene models in the above technical solution. The technical solution in this embodiment can be combined with each optional solution in one or more of the above embodiments. Figure 6 As shown, the training method of the candidate backup scene model includes the following:

[0142] S510: Determine the common scenario historical behavior data of the sample users, and filter out candidate backup scenario historical behavior data from the common scenario historical behavior data according to the proportion of the sample quantity.

[0143] The historical behavior data of general scenarios are classified according to the scenarios, and the historical behavior data and sample quantity under each type of scenario are obtained. The proportion of sample quantity corresponding to each type of scenario is determined, and the scenarios are sorted in ascending order according to the proportion of sample quantity. A preset number of scenes are selected from the first place as candidate scenarios. The specific value of the preset number can be determined based on the recommendation accuracy and is not limited here.

[0144] According to the candidate backup scenarios, the matching historical behavior data is screened from the general historical behavior data as the candidate backup scenario historical behavior data. Exemplarily, the candidate backup scenarios include the iOS system scenario and the first-swipe scenario, and the historical behavior data of the sample user using the iOS system in the general scenario historical behavior data is used as the iOS scenario historical behavior data; similarly, the historical behavior data of the sample user in the general scenario historical behavior data belonging to the first-swipe is used as the first-swipe scenario historical behavior data.

[0145] S520: Input the candidate backup scenario historical behavior data into the candidate backup scenario model to obtain backup initial network parameters.

[0146] The candidate backup scenario historical behavior data is input into the candidate backup scenario model for preliminary training to obtain the backup initial network parameters of the candidate backup scenario model. Specifically, the candidate backup scenario model can adopt a DNN network structure.

[0147] S530 , superimposing noise disturbance on the backup initial network parameters to obtain a disturbed candidate backup scenario model.

[0148] Specifically, the backup initial network parameters are superimposed with noise to perturb, and the candidate backup scenario model after perturbation is obtained according to the backup initial network parameters with superimposed noise. Exemplarily, the weight parameters and deviations of the DNN are superimposed with perturbation noise through the ES evolution strategy, and the value of the perturbation noise can be a fixed value, or formed according to a preset fixed distribution.

[0149] S540: Determine the backup prediction ratio of the probability of the sample user generating the candidate behavior operation on the sample resource output by the disturbed candidate backup scenario model in the candidate backup scenario.

[0150] The disturbed candidate backup scenario model is used to make sample resource recommendation predictions for sample users and obtain the backup prediction ratio.

[0151] Specifically, the user feature data of the sample user, the resource feature data of the sample resource, and the current scenario feature data are determined and input into the disturbed candidate backup scenario model to obtain the model output as the backup prediction ratio of the probability that the sample user will generate a candidate behavior operation on the sample resource.

[0152] S550: Use the disturbed candidate backup scenario model to recommend sample resources to sample users, and determine the sample users' backup real behavior feedback data on the sample resources in the candidate backup scenarios.

[0153] The disturbed candidate backup scenario model is deployed online, and sample resources are actually recommended to sample users in the candidate backup scenarios. Based on the real-time feedback from online sample users, the sample users' real backup behavior feedback data for sample resources is determined.

[0154] S560, updating the candidate backup scenario model according to the backup prediction ratio of the probability of the sample user generating the candidate behavior operation on the sample resource in the candidate backup scenario and the backup real behavior feedback data, to obtain a trained candidate backup scenario model.

[0155] The sample user's backup real behavior feedback data on the sample resource in the candidate backup scenario can reflect the impact of the superimposed disturbance parameters on the model prediction results. Therefore, the candidate backup scenario model is updated according to the degree of influence, and the superimposed disturbance is performed again according to the updated candidate backup scenario model. S440 to S460 are repeated until the sample user's backup prediction ratio of the probability of generating candidate behavior operations on the sample resource matches the backup real behavior feedback data, and it is determined that the candidate backup scenario model training is completed.

[0156] Exemplarily, the backup initial network parameters can be simultaneously perturbed to various degrees to obtain multiple disturbed candidate backup scenario models, determine the backup prediction ratio of the probability of sample users generating candidate behavior operations on sample resources output by each disturbed candidate backup scenario model, and use each disturbed candidate backup scenario model to recommend sample resources to sample users on the candidate backup scenarios, determine the backup real behavior feedback data of sample users on sample resources on the candidate backup scenarios, determine the update parameters of the candidate backup scenario model based on the backup prediction ratio of the probability of sample users generating candidate behavior operations on sample resources corresponding to different degrees of disturbance and the backup real behavior feedback data, and obtain a trained candidate backup scenario model.

[0157] The solution of this embodiment trains the candidate backup scenario model through the historical behavior data of the candidate backup scenarios of sample users, so as to recommend resources for a single scenario to the user, so as to make up for the deviation caused by the insufficient learning of the unevenly distributed single scenario in the general scenario model; and in the process of model training, the network parameters are superimposed with noise disturbance, combined with the real behavior feedback data to improve the training accuracy and prediction accuracy of the final model.

[0158] Figure 7 is a schematic diagram of the structure of a resource recommendation device according to an embodiment of the present disclosure, and the device can execute the resource recommendation method involved in any embodiment of the present disclosure; Figure 7 The resource recommendation device 600 includes: a first proportion parameter determination module 610, a standby proportion parameter determination module 620, a fusion proportion parameter determination module 630 and a resource recommendation module 640.

[0159] A first proportion parameter determination module 610 is used to input the user characteristic data of the target user, the resource characteristic data of the candidate resource, and the current scene characteristic data into a pre-trained general scene model to obtain a first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource; wherein the general scene model is obtained by training the network parameters by superimposing noise disturbance;

[0160] The standby ratio parameter determination module 620 is used to obtain the standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario according to the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and the pre-trained target standby scenario model; wherein the target standby scenario model is obtained by superimposing noise disturbance training on the network parameters;

[0161] A fusion ratio parameter determination module 630, configured to determine a fusion ratio parameter according to the first ratio parameter and the standby ratio parameter;

[0162] The resource recommendation module 640 is used to determine the recommendation result of the candidate resource according to the candidate behavior operation probability and the fusion ratio parameter corresponding to the candidate behavior operation probability.

[0163] The solution of this embodiment obtains the backup proportion parameters of the probability of the target user generating candidate behavior operations on candidate resources in the target backup scenario through the target backup scenario model, and merges the backup proportion parameters with the general proportion parameters obtained by the general scenario model, so as to correct the deviation caused by insufficient learning of the target backup scenario in the first proportion parameters obtained by the general scenario model, improve the ability to analyze users' resource preferences in different scenarios, and improve the accuracy of recommending candidate resources to target users.

[0164] In an optional implementation of this embodiment, the standby ratio parameter determination module is specifically used to:

[0165] A feature data screening unit, configured to screen the user feature data of the target user, the resource feature data of the candidate resources and the current scene feature data according to the target backup scene corresponding to the target backup scene model, to obtain target feature data associated with the target backup scene;

[0166] The standby ratio parameter determination unit is used to input the target feature data into the target standby scenario model to obtain the standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario.

[0167] In an optional implementation of this embodiment, the device further includes a target backup scenario model determination module, which is used to obtain the backup proportion parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario based on the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and the pre-trained target backup scenario model.

[0168] The corresponding target backup scene model is determined from pre-trained candidate backup scene models according to the current scene feature data.

[0169] In an optional implementation of this embodiment, the proportion of the training sample volume of the candidate backup scene corresponding to the candidate backup scene model in the total training sample volume is less than a preset proportion threshold. In an optional implementation of this embodiment, the fusion proportion parameter determination module includes:

[0170] A scene weight parameter determination unit, used to determine a scene weight parameter corresponding to a target backup scene;

[0171] A scene correction parameter determination unit, used to determine a target standby scene correction parameter according to a weighted value of the scene weight parameter and the standby proportion parameter;

[0172] A fusion ratio parameter determination unit is used to determine a fusion ratio parameter according to the target backup scene correction parameter and the first ratio parameter.

[0173] In an optional implementation of this embodiment, the scene weight parameter determination unit is specifically configured to:

[0174] Determine the degree of deviation correction of the target backup scenario to the general scenario according to the sample size ratio of the training sample size of the target backup scenario in the total training sample size and the scenario importance parameter of the target backup scenario;

[0175] The scenario weight parameter corresponding to the target backup scenario is determined according to the deviation correction degree.

[0176] The resource recommendation device can execute the resource recommendation method provided by any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, please refer to the resource recommendation method provided by any embodiment of the present disclosure.

[0177] Figure 8 is a schematic diagram of a general scene model training device according to an embodiment of the present disclosure, and the device can execute the general scene model training method involved in any embodiment of the present disclosure; Figure 8The training device 700 of the general scenario model includes: a general sample data determination module 710, a general initial parameter determination module 720, a first disturbance module 730, a first prediction module 740, a first feedback module 750 and a general scenario model update module 760.

[0178] The general sample data determination unit module 710 is used to determine the general scenario historical behavior data of the sample user;

[0179] A general initial parameter determination module unit 720 is used to input the general scenario historical behavior data into a general scenario model to obtain general initial network parameters;

[0180] A first disturbance module unit 730 is used to perturb the general initial network parameters and superimpose noise disturbance to obtain a general scenario model after disturbance;

[0181] The first prediction module unit 740 is used to determine the general prediction proportion of the probability of the sample user generating a candidate behavior operation on the sample resource output by the disturbed general scenario model;

[0182] A first feedback module unit 750 is used to recommend sample resources to the sample users by using the perturbed general scenario model, and determine general real behavior feedback data of the sample users on the sample resources;

[0183] The general scenario model updating module unit 760 is used to update the general scenario model according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resource and the general real behavior feedback data to obtain a trained general scenario model.

[0184] The solution of this embodiment trains the general scenario model through the general scenario historical behavior data of sample users to achieve general resource recommendations for users in multiple scenarios; and in the model training process, by superimposing noise perturbations on the network parameters, combined with real behavior feedback data, the training accuracy and prediction accuracy of the final model are improved.

[0185] In an optional implementation of this embodiment, the first disturbance unit module is specifically used to:

[0186] Determining initial disturbance parameters according to a preset fixed distribution;

[0187] Scaling the initial disturbance parameter according to the order of magnitude of the initial network parameter to obtain a target disturbance parameter;

[0188] The target disturbance parameter is superimposed on the initial network parameter.

[0189] The training device of the general scene model can execute the training method of the general scene model provided in any embodiment of the present disclosure, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the training method of the general scene model provided in any embodiment of the present disclosure.

[0190] Fig. 9 is a schematic diagram of a structure of a training device for a candidate backup scene model according to an embodiment of the present disclosure, and the device can execute a training method for a candidate backup scene model involved in any embodiment of the present disclosure; Fig. 9 The training device 800 for the candidate backup scene model includes: a backup sample data determination unit module 810, a backup initial parameter determination module 820, a second disturbance module 830, a second prediction module 840, a second feedback module 850 and a backup scene model update module 860.

[0191] A spare sample data determination module 810 is used to determine the common scenario historical behavior data of the sample user, and to filter the candidate spare scenario historical behavior data from the common scenario historical behavior data according to the sample quantity ratio;

[0192] A standby initial parameter determination module 820 is used to input the candidate standby scenario historical behavior data into the candidate standby scenario model to obtain standby initial network parameters;

[0193] A second disturbance module 830 is used to perturb the backup initial network parameters to superimpose noise disturbance to obtain a disturbed candidate backup scenario model;

[0194] The second prediction module 840 is used to determine the standby prediction proportion of the probability of the sample user generating the candidate behavior operation on the sample resource under the candidate standby scenario output by the disturbed candidate standby scenario model;

[0195] A second feedback module 850 is used to recommend sample resources to the sample user by using the disturbed candidate backup scenario model, and determine the sample user's backup real behavior feedback data on the sample resource in the candidate backup scenario;

[0196] The backup scenario model updating module 860 is used to update the candidate backup scenario model according to the backup prediction ratio of the sample user's probability of generating a candidate behavior operation on the sample resource in the candidate backup scenario and the backup real behavior feedback data to obtain a trained candidate backup scenario model.

[0197] The solution of this embodiment trains the candidate backup scenario model through the historical behavior data of the candidate backup scenarios of sample users, so as to recommend resources for a single scenario to the user, so as to make up for the deviation caused by the insufficient learning of the unevenly distributed single scenario in the general scenario model; and in the process of model training, the network parameters are superimposed with noise disturbance, combined with the real behavior feedback data to improve the training accuracy and prediction accuracy of the final model.

[0198] The training device for the candidate backup scene model can execute the training method for the candidate backup scene model provided in any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, please refer to the training method for the candidate backup scene model provided in any embodiment of the present disclosure.

[0199] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0200] 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.

[0201] Fig.10 A schematic block diagram of an example electronic device 600 that can be used to implement an embodiment 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 processing, cellular phones, smart phones, 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 required herein.

[0202] like Fig.10 As shown, the device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the device 900 can also be stored. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0203] A number of components in the device 900 are connected to the I / O interface 905, including: an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0204] The computing unit 901 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 901 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 running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 901 performs the various methods and processes described above, such as a resource recommendation method, a training method for a general scenario model, or a training method for a candidate standby scenario model. For example, in some embodiments, the resource recommendation method, the training method for a general scenario model, or the training method for a candidate standby scenario model may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 908. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the resource recommendation method, the training method for a general scenario model, or the training method for a candidate standby scenario model described above may be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to execute a resource recommendation method, a general scenario model training method, or a candidate backup scenario model training method in any other appropriate manner (eg, by means of firmware).

[0205] Various implementations of the systems and techniques described above 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), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including 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.

[0206] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0207] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may 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 may 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.

[0208] 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).

[0209] The systems and techniques described herein may be implemented in a computing system that includes backend 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 frontend components (e.g., a user computer with 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 backend components, middleware components, or frontend components. The components of the system may 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), a blockchain network, and the Internet.

[0210] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, a server of a distributed system, or a server combined with a blockchain.

[0211] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0212] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A resource recommendation method, comprising: Inputting the user feature data of the target user, the resource feature data of the candidate resource and the current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user will generate a candidate behavior operation on the candidate resource; wherein the general scenario model is obtained by training the network parameters by superimposing noise disturbance; According to the user characteristic data of the target user, the resource characteristic data of the candidate resource, the current scenario characteristic data and the pre-trained target backup scenario model, a backup proportion parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario is obtained; wherein the target backup scenario model is obtained by training the network parameters by superimposing noise disturbance; Determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter; The recommendation result of the candidate resource is determined according to the candidate behavior operation probability and a fusion ratio parameter corresponding to the candidate behavior operation probability.

2. The method according to claim 1, wherein: According to the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and the pre-trained target backup scenario model, a backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario is obtained, including: Screening the user characteristic data of the target user, the resource characteristic data of the candidate resources, and the current scenario characteristic data according to the target backup scenario corresponding to the target backup scenario model to obtain target characteristic data associated with the target backup scenario; The target feature data is input into the target backup scenario model to obtain a backup ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario.

3. The method according to claim 1 or 2, wherein: Before obtaining the standby proportion parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario according to the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and the pre-trained target standby scenario model, the method further includes: A corresponding target backup scene model is determined from pre-trained candidate backup scene models according to the current scene feature data.

4. The method according to claim 3, wherein: The proportion of the training sample volume of the candidate backup scenario corresponding to the candidate backup scenario model in the total training sample volume is less than a preset proportion threshold.

5. The method according to claim 1, wherein: Determining a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter includes: Determine the scenario weight parameters corresponding to the target backup scenario; Determining a target standby scene correction parameter according to a weighted value of the scene weight parameter and the standby proportion parameter; A fusion proportion parameter is determined according to the target backup scene correction parameter and the first proportion parameter.

6. The method according to claim 5, wherein: Determine the scenario weight parameters corresponding to the target backup scenario, including: Determine the degree of deviation correction of the target backup scenario to the general scenario according to the sample size ratio of the training sample size of the target backup scenario in the total training sample size and the scenario importance parameter of the target backup scenario; The scenario weight parameter corresponding to the target backup scenario is determined according to the deviation correction degree.

7. A method for training a general scene model, comprising: Determine the common scenario historical behavior data of sample users; Inputting the general scenario historical behavior data into the general scenario model to obtain general initial network parameters; Superimposing noise disturbance on the general initial network parameters to obtain a general scene model after disturbance; Determine the general prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource output by the disturbed general scenario model; recommending sample resources to the sample users by using the disturbed general scenario model, and determining general real behavior feedback data of the sample users on the sample resources; The general scenario model is updated according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resource and the general real behavior feedback data to obtain a trained general scenario model.

8. The method according to claim 7, wherein: Superimposing noise disturbance on the general initial network parameters includes: Determining initial disturbance parameters according to a preset fixed distribution; Scaling the initial disturbance parameter according to the order of magnitude of the initial network parameter to obtain a target disturbance parameter; The target disturbance parameter is superimposed on the initial network parameter.

9. A method for training a candidate backup scene model, comprising: Determine the common scenario historical behavior data of the sample users, and filter the candidate backup scenario historical behavior data from the common scenario historical behavior data according to the sample quantity ratio; Inputting the candidate backup scenario historical behavior data into the candidate backup scenario model to obtain backup initial network parameters; Superimposing noise disturbance on the backup initial network parameters to obtain a disturbed candidate backup scenario model; Determine the standby prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource under the candidate standby scenario output by the disturbed candidate standby scenario model; recommending sample resources to the sample user by using the disturbed candidate backup scenario model, and determining backup real behavior feedback data of the sample user on the sample resource in the candidate backup scenario; The candidate backup scenario model is updated according to the backup prediction ratio of the probability of the sample user generating the candidate behavior operation on the sample resource in the candidate backup scenario and the backup real behavior feedback data to obtain a trained candidate backup scenario model.

10. A resource recommendation device, comprising: A first proportion parameter determination module is used to input user feature data of a target user, resource feature data of a candidate resource, and current scenario feature data into a pre-trained general scenario model to obtain a first proportion parameter of the probability that the target user generates a candidate behavior operation on the candidate resource; wherein the general scenario model is obtained by training the network parameters by superimposing noise disturbances; A standby ratio parameter determination module is configured to obtain a standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in a target standby scenario based on the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data, and a pre-trained target standby scenario model; wherein the target standby scenario model is obtained by training the network parameters with superimposed noise disturbance; A fusion proportion parameter determination module, used to determine a fusion proportion parameter according to the first proportion parameter and the standby proportion parameter; The resource recommendation module is used to determine the recommendation result of the candidate resource according to the candidate behavior operation probability and the fusion proportion parameter corresponding to the candidate behavior operation probability.

11. The device according to claim 10, wherein: The standby ratio parameter determination module is specifically used for: A feature data screening unit, configured to screen the user feature data of the target user, the resource feature data of the candidate resources and the current scene feature data according to the target backup scene corresponding to the target backup scene model, to obtain target feature data associated with the target backup scene; The standby ratio parameter determination unit is used to input the target feature data into the target standby scenario model to obtain the standby ratio parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target standby scenario.

12. The device according to claim 10 or 11, wherein: The device also includes a target backup scenario model determination module, which is used to obtain the backup proportion parameter of the probability of the target user generating a candidate behavior operation on the candidate resource in the target backup scenario based on the user feature data of the target user, the resource feature data of the candidate resource, the current scenario feature data and the pre-trained target backup scenario model, A corresponding target backup scene model is determined from pre-trained candidate backup scene models according to the current scene feature data.

13. The device according to claim 12, wherein: The proportion of the training sample volume of the candidate backup scenario corresponding to the candidate backup scenario model in the total training sample volume is less than a preset proportion threshold.

14. The device according to claim 10, wherein: The fusion ratio parameter determination module includes: A scene weight parameter determination unit, used to determine a scene weight parameter corresponding to a target backup scene; A scene correction parameter determination unit, used to determine a target standby scene correction parameter according to a weighted value of the scene weight parameter and the standby proportion parameter; A fusion ratio parameter determination unit is used to determine a fusion ratio parameter according to the target backup scene correction parameter and the first ratio parameter.

15. The device according to claim 14, wherein: The scene weight parameter determination unit is specifically used for: Determine the degree of deviation correction of the target backup scenario to the general scenario according to the sample size ratio of the training sample size of the target backup scenario in the total training sample size and the scenario importance parameter of the target backup scenario; The scenario weight parameter corresponding to the target backup scenario is determined according to the deviation correction degree.

16. A training device for a general scene model, comprising: A general sample data determination module is used to determine the general scenario historical behavior data of sample users; A general initial parameter determination module, used for inputting the general scenario historical behavior data into the general scenario model to obtain general initial network parameters; A first disturbance module is used to superimpose noise disturbance on the general initial network parameters to obtain a general scenario model after disturbance; A first prediction module, used to determine the general prediction proportion of the probability of the sample user generating a candidate behavior operation on the sample resource output by the general scenario model after the disturbance; A first feedback module, configured to recommend sample resources to the sample users by using the perturbed general scenario model, and determine general real behavior feedback data of the sample users on the sample resources; The general scenario model updating module is used to update the general scenario model according to the general predicted proportion of the probability of the sample user generating candidate behavior operations on the sample resources and the general real behavior feedback data to obtain a trained general scenario model.

17. The device according to claim 16, wherein: The first disturbance module is specifically used for: Determining initial disturbance parameters according to a preset fixed distribution; Scaling the initial disturbance parameter according to the order of magnitude of the initial network parameter to obtain a target disturbance parameter; The target disturbance parameter is superimposed on the initial network parameter.

18. A training device for a candidate backup scene model, comprising: A spare sample data determination module is used to determine the common scenario historical behavior data of the sample user, and to filter the candidate spare scenario historical behavior data from the common scenario historical behavior data according to the sample quantity ratio; A standby initial parameter determination module, used for inputting the candidate standby scenario historical behavior data into the candidate standby scenario model to obtain standby initial network parameters; A second disturbance module is used to superimpose noise disturbance on the backup initial network parameters to obtain a disturbed candidate backup scenario model; A second prediction module is used to determine the standby prediction ratio of the probability of the sample user generating a candidate behavior operation on the sample resource under the candidate standby scenario output by the disturbed candidate standby scenario model; A second feedback module is used to recommend sample resources to the sample user by using the disturbed candidate backup scenario model, and determine the sample user's backup real behavior feedback data on the sample resource in the candidate backup scenario; The backup scenario model updating module is used to update the candidate backup scenario model according to the backup prediction proportion of the sample user's candidate behavior operation probability on the sample resource in the candidate backup scenario and the backup real behavior feedback data to obtain a trained candidate backup scenario model.

19. 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 9.

20. 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 1-9.

21. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images