A method for determining a multimedia resource, a method and device for training a prediction model

Abstract

The present disclosure relates to a multimedia resource determination method and a prediction model training method and device, and relates to the technical field of computers. Embodiments of the present disclosure at least solve the problem that the recommendation of multimedia resources to an account may be inaccurate in related technologies. The method comprises: obtaining an account reconstruction feature of a current account and a resource reconstruction feature of each candidate multimedia resource in a plurality of candidate multimedia resources; the account reconstruction feature is used to represent relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account; the resource reconstruction feature is used to represent relevant factors and irrelevant factors of the multimedia resource in the search result under the search condition affecting the interactive behavior of the account; according to the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource, a predicted interactive parameter of the current account is predicted, and a to-be-recommended multimedia resource for recommending to the current account is determined from the plurality of candidate multimedia resources.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to a multimedia resource determination method, a prediction model training method and device. BACKGROUND

[0002] In a multimedia resource recommendation scenario, an electronic device can determine a to-be-recommended multimedia resource for recommending to an account by using historical click behaviors of the account and a pre-trained prediction model. Specifically, the electronic device obtains account features of the account, resource features of historical multimedia resources clicked by the account in a historical time period, and resource features of each candidate multimedia resource, and inputs the account features, the resource features of the historical multimedia resources, and the resource features of each candidate multimedia resource into the prediction model to perform a prediction operation of a click rate of the account on each candidate multimedia video, and obtain a plurality of click rates. Further, the electronic device determines the to-be-recommended multimedia resource for recommending to the account according to the predicted plurality of click rates.

[0003] However, in the above multimedia resource determination method, there may be some confusing factors in the historical click behavior data, which cannot fully represent the real interests of the account. For example, there may be popularity bias or selection bias in the historical click behavior data, that is, the account performs a click operation only because the clicked multimedia resource is relatively popular, or the clicked multimedia resource in the multimedia resource recommended by the recommendation system to the account is relatively interesting to the account without other choices. In this way, based on the existence of the confusing factors in the historical click behavior data, the multimedia resource recommended to the account by using the above multimedia resource determination method is not accurate enough. SUMMARY

[0004] The present disclosure provides a multimedia resource determination method, a prediction model training method and device to at least solve the problem that the multimedia resource recommended to the account may not be accurate in the related art. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of the embodiments of the present disclosure, a method for determining a multimedia resource is provided, including: obtaining an account reconstruction feature of a current account and a resource reconstruction feature of each candidate multimedia resource in a plurality of candidate multimedia resources; the account reconstruction feature is used to represent relevant factors and irrelevant factors of a search behavior of the account affecting an interactive behavior of the account; the resource reconstruction feature is used to represent relevant factors and irrelevant factors of a search condition of the multimedia resource in a search result under the search condition affecting the interactive behavior of the account; predicting a predicted interactive parameter of the current account for each candidate multimedia resource according to the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource; and determining a to-be-recommended multimedia resource for recommending to the current account from the plurality of candidate multimedia resources according to the predicted interactive parameter of each candidate multimedia resource.

[0006] Optionally, the account reconstruction feature of the current account is obtained by: obtaining an account feature of the current account, a resource feature of a historical multimedia resource on which the current account has performed an interactive behavior in a historical time period, and a search condition feature of each historical multimedia resource; performing feature reconstruction processing on the resource feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource to obtain a resource reconstruction feature of each historical multimedia resource; and obtaining the account reconstruction feature of the current account according to the resource reconstruction feature of each historical multimedia resource and the account feature of the current account.

[0007] Optionally, the resource reconstruction feature of each historical multimedia resource is obtained by: determining a linear feature of each historical multimedia resource and a nonlinear feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource; the linear feature of each historical multimedia resource is used to represent relevant factors of each historical multimedia resource affecting the interactive behavior of the account under the search condition, and the nonlinear feature of each historical multimedia resource is used to represent irrelevant factors of each historical multimedia resource affecting the interactive behavior of the account under the search condition; and weighting the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource to obtain the resource reconstruction feature of each historical multimedia resource.

[0008] Optionally, the determining the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource comprises: obtaining a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource according to the search condition feature of each historical multimedia resource, the resource feature of each historical multimedia resource, and a preset regression operation; determining the fitting vector of each historical multimedia resource as the linear feature of each historical multimedia resource, and determining the residual vector of each historical multimedia resource as the nonlinear feature of each historical multimedia resource.

[0009] Optionally, the method further comprises: performing feature dimension conversion on the search condition feature of each historical multimedia resource to obtain converted search condition feature of each historical multimedia resource; the converted search condition feature of each historical multimedia resource has the same dimension as the resource feature of each historical multimedia resource; merging the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource to obtain merged feature of each historical multimedia resource; inputting the merged feature of each historical multimedia resource into a pre-trained first weight model to obtain the weight of the linear feature of each historical multimedia resource; the first weight model comprises a first parameter, and the first parameter is used to learn the converted search condition feature in the merged feature; inputting the merged feature of each historical multimedia resource into a pre-trained second weight model to obtain the weight of the nonlinear feature of each historical multimedia resource; the second weight model comprises a second parameter, and the second parameter is used to learn the resource feature in the merged feature.

[0010] Optionally, the obtaining the resource reconstruction feature of each candidate multimedia resource comprises: obtaining the resource feature of each candidate multimedia resource and the search condition feature of each candidate multimedia resource; determining the linear feature of each candidate multimedia resource and the nonlinear feature of each candidate multimedia resource according to the search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource; the linear feature of each candidate multimedia resource is used to represent relevant factors of each candidate multimedia resource affecting the account to perform the interactive behavior under the search condition, and the nonlinear feature of each candidate multimedia resource is used to represent non-relevant factors of each candidate multimedia resource affecting the account to perform the interactive behavior under the search condition; weighting the linear feature of each candidate multimedia resource and the nonlinear feature of each candidate multimedia resource to obtain the resource reconstruction feature of each candidate video.

[0011] Optionally, the determining of the linear feature and the nonlinear feature of each candidate multimedia resource based on the search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource comprises: obtaining a fitting vector and a residual vector of each candidate multimedia resource based on the search condition feature of each candidate multimedia resource, the resource feature of each candidate multimedia resource, and a preset regression operation; determining the fitting vector of each candidate multimedia resource as the linear feature of each candidate multimedia resource, and determining the residual vector of each candidate multimedia resource as the nonlinear feature of each candidate multimedia resource.

[0012] Optionally, the method further comprises: performing feature dimension conversion on the search condition feature of each candidate multimedia resource to obtain converted search condition features of each candidate multimedia resource; the converted search condition features of each candidate multimedia resource have the same dimension as the resource feature of each candidate multimedia resource; merging the converted search condition features of each candidate multimedia resource and the resource feature of each candidate multimedia resource to obtain merged features of each candidate multimedia resource; inputting the merged features of each candidate multimedia resource into a pre-trained first weight model to obtain weights of the linear feature of each candidate multimedia resource; the first weight model comprises first parameters, and the first parameters are used to learn the converted search condition features in the merged features; inputting the merged features of each candidate multimedia resource into a pre-trained second weight model to obtain weights of the nonlinear feature of each candidate multimedia resource; the second weight model comprises second parameters, and the second parameters are used to learn the resource feature in the merged features.

[0013] Optionally, the predicted interaction parameter of each candidate multimedia resource is obtained based on a pre-trained prediction model, and the method further comprises: obtaining account reconstruction features of a sample account, resource reconstruction features of a sample multimedia resource, and a sample interaction parameter of the sample multimedia resource; the sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period; training a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and taking the sample interaction parameter of the sample multimedia resource as a label to obtain the prediction model.

[0014] According to a second aspect of the embodiments of the present disclosure, a training method of a prediction model is provided. The method comprises: obtaining account reconstruction features of a sample account, resource reconstruction features of sample multimedia resources, and sample interaction parameters of the sample multimedia resources; the account reconstruction features are used to represent relevant factors and irrelevant factors of search behavior of the account affecting interaction behavior of the account; the sample multimedia resources are multimedia resources on which the sample account has performed interaction behavior in a sample time period; the resource reconstruction features are used to represent relevant factors and irrelevant factors of the multimedia resources in the search results affecting the account to perform interaction behavior under the search condition; the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resources are used as sample features, and the sample interaction parameters of the sample multimedia resources are used as labels, a preset neural network model is trained, and a prediction model is obtained; the prediction model is used to predict a predicted interaction parameter of the sample account to a candidate multimedia resource.

[0015] According to a third aspect of the embodiments of the present disclosure, a determination apparatus of a multimedia resource is provided. The apparatus comprises: an obtaining unit, a predicting unit, and a determining unit; the obtaining unit is configured to obtain account reconstruction features of a current account and resource reconstruction features of each candidate multimedia resource in a plurality of candidate multimedia resources; the account reconstruction features are used to represent relevant factors and irrelevant factors of search behavior of the account affecting interaction behavior of the account; the resource reconstruction features are used to represent relevant factors and irrelevant factors of the multimedia resources in the search results affecting the account to perform interaction behavior under the search condition; the predicting unit is configured to predict a predicted interaction parameter of the current account to each candidate multimedia resource according to the account reconstruction features of the current account and the resource reconstruction features of each candidate multimedia resource; and the determining unit is configured to determine a to-be-recommended multimedia resource for recommendation to the current account from the plurality of candidate multimedia resources according to the predicted interaction parameter of each candidate multimedia resource.

[0016] Optionally, the obtaining unit is specifically configured to: obtain account features of the current account, resource features of historical multimedia resources on which the current account has performed interaction behavior in a historical time period, and search condition features of each historical multimedia resource; perform feature reconstruction processing on the resource features of each historical multimedia resource according to the search condition features of each historical multimedia resource, to obtain resource reconstruction features of each historical multimedia resource; and obtain the account reconstruction features of the current account according to the resource reconstruction features of each historical multimedia resource and the account features of the current account.

[0017] Optionally, the acquisition unit is specifically configured to: determine linear features of each historical multimedia resource and nonlinear features of each historical multimedia resource according to the search condition features of each historical multimedia resource and the resource features of each historical multimedia resource; the linear features of each historical multimedia resource are used to represent relevant factors of each historical multimedia resource affecting the account to perform interactive behaviors under the search condition, and the nonlinear features of each historical multimedia resource are used to represent non-relevant factors of each historical multimedia resource affecting the account to perform interactive behaviors under the search condition; and weight the linear features of each historical multimedia resource and the nonlinear features of each historical multimedia resource to obtain resource reconstruction features of each historical multimedia resource.

[0018] Optionally, the acquisition unit is specifically configured to: obtain fitting vectors of each historical multimedia resource and residual vectors of each historical multimedia resource according to the search condition features of each historical multimedia resource, the resource features of each historical multimedia resource, and a preset regression operation; determine the fitting vectors of each historical multimedia resource as the linear features of each historical multimedia resource, and determine the residual vectors of each historical multimedia resource as the nonlinear features of each historical multimedia resource.

[0019] Optionally, the device further includes a processing unit; the processing unit is configured to perform feature dimension conversion on the search condition features of each historical multimedia resource to obtain converted search condition features of each historical multimedia resource; the converted search condition features of each historical multimedia resource have the same dimension as the resource features of each historical multimedia resource; the processing unit is further configured to combine the converted search condition features of each historical multimedia resource and the resource features of each historical multimedia resource to obtain combined features of each historical multimedia resource; the processing unit is further configured to input the combined features of each historical multimedia resource into a pre-trained first weight model to obtain weights of the linear features of each historical multimedia resource; the first weight model includes first parameters, and the first parameters are used to learn the converted search condition features in the combined features; and the processing unit is further configured to input the combined features of each historical multimedia resource into a pre-trained second weight model to obtain weights of the nonlinear features of each historical multimedia resource; the second weight model includes second parameters, and the second parameters are used to learn the resource features in the combined features.

[0020] Optionally, the obtaining unit is specifically configured to: obtain resource features of each candidate multimedia resource and search condition features of each candidate multimedia resource; determine linear features of each candidate multimedia resource and nonlinear features of each candidate multimedia resource according to the search condition features of each candidate multimedia resource and the resource features of each candidate multimedia resource; the linear features of each candidate multimedia resource are used to represent relevant factors of each candidate multimedia resource affecting the account to perform the interactive behavior under the search condition, and the nonlinear features of each candidate multimedia resource are used to represent irrelevant factors of each candidate multimedia resource affecting the account to perform the interactive behavior under the search condition; and weight the linear features of each candidate multimedia resource and the nonlinear features of each candidate multimedia resource to obtain resource reconstruction features of each candidate video.

[0021] Optionally, the obtaining unit is specifically configured to: obtain fitting vectors of each candidate multimedia resource and residual vectors of each candidate multimedia resource according to the search condition features of each candidate multimedia resource, the resource features of each candidate multimedia resource, and a preset regression operation; determine the fitting vectors of each candidate multimedia resource as the linear features of each candidate multimedia resource, and determine the residual vectors of each candidate multimedia resource as the nonlinear features of each candidate multimedia resource.

[0022] Optionally, the device further includes a processing unit, which is configured to: perform feature dimension conversion on the search condition features of each candidate multimedia resource to obtain converted search condition features of each candidate multimedia resource; the converted search condition features of each candidate multimedia resource have the same dimension as the resource features of each candidate multimedia resource; combine the converted search condition features of each candidate multimedia resource and the resource features of each candidate multimedia resource to obtain combined features of each candidate multimedia resource; input the combined features of each candidate multimedia resource into a pre-trained first weight model to obtain weights of the linear features of each candidate multimedia resource; the first weight model includes first parameters used to learn the converted search condition features in the combined features; and input the combined features of each candidate multimedia resource into a pre-trained second weight model to obtain weights of the nonlinear features of each candidate multimedia resource; the second weight model includes second parameters used to learn the resource features in the combined features.

[0023] Optionally, the predicted interaction parameter of each candidate multimedia candidate resource is predicted based on a pre-trained prediction model, and the apparatus further comprises a training unit; the obtaining unit is further configured to obtain account reconstruction features of a sample account, resource reconstruction features of a sample multimedia resource, and a sample interaction parameter of the sample multimedia resource; the sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period; and the training unit is configured to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and taking the sample interaction parameter of the sample multimedia resource as a label, to obtain the prediction model.

[0024] According to a fourth aspect of the embodiments of the present disclosure, a training apparatus of a prediction model is provided, which comprises an obtaining unit and a training unit; the obtaining unit is configured to obtain account reconstruction features of a sample account, resource reconstruction features of a sample multimedia resource, and a sample interaction parameter of the sample multimedia resource; the account reconstruction features are used to represent relevant factors and irrelevant factors of a search behavior of an account affecting an interaction behavior of the account; the sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period; the resource reconstruction features are used to represent relevant factors and irrelevant factors of the multimedia resource in a search result under a search condition affecting an interaction behavior of an account; and the training unit is configured to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and taking the sample interaction parameter of the sample multimedia resource as a label, to obtain the prediction model; and the prediction model is used to predict a predicted interaction parameter of the sample account to a candidate multimedia resource.

[0025] According to a fifth aspect of the embodiments of the present disclosure, an electronic device is provided, which comprises a processor and a memory for storing instructions executable by the processor; and the processor is configured to execute the instructions to implement the method for determining a multimedia resource according to the first aspect and any possible design of the first aspect, or the method for training a prediction model according to the second aspect and any possible design of the second aspect.

[0026] According to a sixth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, which, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the method for determining a multimedia resource according to the first aspect and any possible design of the first aspect, or the method for training a prediction model according to the second aspect and any possible design of the second aspect.

[0027] According to a seventh aspect of the embodiments of the present disclosure, a computer program product is provided, which includes computer instructions, when the computer instructions are run on an electronic device, cause the electronic device to perform the method for determining a multimedia resource as provided by the first aspect and any possible design of the first aspect, or the method for training a prediction model as provided by the second aspect and any possible design of the second aspect.

[0028] The technical solutions provided by the present disclosure at least have the following beneficial effects: considering the causal relationship between the search behavior of an account and the interactive behavior of the account, using account reconstruction features to represent the behavior characteristics of the account, the relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account can make the structure of the determined account behavior characteristics more clear and more refined, thereby improving the accuracy of the account behavior characteristics. At the same time, considering the causal relationship between the search condition of the multimedia resource and the interactive behavior performed by the account, using resource reconstruction features to represent the relevant factors and irrelevant factors of the search condition of the multimedia resource affecting the interactive behavior performed by the account can make the results of representing the characteristics of the multimedia resource more clear and refined, thereby improving the accuracy of the characteristics of the multimedia resource. In this way, using the account reconstruction features and the resource reconstruction features as the basis for predicting the interactive parameters can maximize the distinction between the causal and non-causal parts of the account and the candidate multimedia resource, thereby improving the accuracy of the predicted interactive parameters, and further improving the accuracy of the determined multimedia resource to be recommended.

[0029] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present disclosure and, together with the specification, serve to explain the principles of the present disclosure, and do not constitute an undue limitation on the present disclosure.

[0031] Figure 1 is a method schematic diagram of a method for determining a multimedia resource according to an exemplary embodiment;

[0032] Figure 2 is a structural schematic diagram of a recommendation system according to an exemplary embodiment;

[0033] Figure 3 is one of the flow schematic diagrams of a method for determining a multimedia resource according to an exemplary embodiment;

[0034] Figure 4 is the second flow schematic diagram of a method for determining a multimedia resource according to an exemplary embodiment;

[0035] Figure 5 is a schematic diagram of determining account reconstruction features according to an example embodiment;

[0036] Figure 6 is a third flowchart of a method for determining multimedia resources according to an example embodiment;

[0037] Figure 7 is a fourth flowchart of a method for determining multimedia resources according to an example embodiment;

[0038] Figure 8 is a schematic diagram of performing feature reconstruction operations according to an example embodiment;

[0039] Figure 9 is a fifth flowchart of a method for determining multimedia resources according to an example embodiment;

[0040] Figure 10 is a flowchart of a method for training a prediction model according to an example embodiment;

[0041] Figure 11 is a structural schematic diagram of a device for determining multimedia resources according to an example embodiment;

[0042] Figure 12 is a structural schematic diagram of a device for training a prediction model according to an example embodiment;

[0043] Figure 13 is a structural schematic diagram of an electronic device according to an example embodiment. DETAILED DESCRIPTION

[0044] In order to make the ordinary person in the art better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings.

[0045] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following example embodiments does not represent all implementations consistent with the present disclosure. Rather, they are only examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0046] In addition, in the description of the embodiments of the disclosure, unless otherwise specified, " / " represents the meaning of or, for example, A / B can represent A or B. "And / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. In addition, in the description of the embodiments of the disclosure, "multiple" means two or more than two.

[0047] The invention principle of the multimedia resource determination method and the prediction model training method provided by the embodiments of the disclosure is introduced as follows:

[0048] In the related art, when recommending multimedia resources to an account, the sampling data used includes account features of the account, resource features of historical multimedia resources clicked by the account in a historical time period, and resource features of candidate multimedia resources. However, in actual scenarios, due to the existence of confusion factors in historical click behavior data, the resource features of historical multimedia resources cannot fully represent the real interests of the account, which may lead to that the predicted click rate cannot be fully based on the interests of the account. For example, as shown in FIG. 1, an electronic device inputs the account features of an account and the resource features of historical multimedia resources into a preset account portrait model, and obtains account portrait features representing the interests of the account. Further, the electronic device inputs the account portrait features and the resource features of candidate multimedia resources into a pre-trained prediction model, and obtains a predicted click rate of the candidate multimedia resources. Figure 1

[0049] In some embodiments using a prediction model, the related art also uses a preset self-attention model to construct the interests of the account according to the search behavior and the click behavior of the account through a self-attention mechanism in the self-attention model. However, in this embodiment, only the search behavior data and the click behavior data of the account are superimposed, and the confusion factors are not considered, and the above problems cannot be solved.

[0050] ​Embodiments of the present disclosure consider that there may be confounding factors in historical click behavior data, and therefore, in order to weaken the influence of confounding factors, a causal relationship between searching multimedia resources and performing interactive behaviors on multimedia resources by an account in a historical time period can be determined, and then relevant factors and irrelevant factors of the account performing interactive behaviors can be determined, wherein the relevant factors can be factors that reflect interactive behaviors on search results by searching behaviors in the historical time period, and the irrelevant factors can be factors that affect interactive behaviors on multimedia resources by the account in the historical time period due to the confounding factors. Further, the relevant factors and the irrelevant factors are decoupled and reconstructed into new features, which can greatly improve the influence relationship between the determined historical search behavior data of the account and interactive behaviors of the account, and then the accuracy of the predicted interactive parameters can be improved, and finally the accuracy of the recommended multimedia resources to the account can be greatly improved.

[0051] The multimedia resource determination method provided by the embodiments of the present disclosure can be applied to a recommendation system. Figure 2 A structural schematic diagram of the recommendation system is shown. As shown in the figure, Figure 2 The recommendation system 10 is used to solve the problem of inaccurate search result ranking in the related art. The recommendation system 10 includes a multimedia resource recommendation device (hereinafter referred to as a recommendation device for brevity) 11 and an electronic device 12. The recommendation device 11 is connected with the electronic device 12. The recommendation device 11 and the electronic device 12 can be connected in a wired manner or in a wireless manner, and the embodiments of the present disclosure do not limit this.

[0052] The recommendation device 11 can be used for data interaction with the electronic device 12, for example, the recommendation device 11 can obtain account information of an account, historical search behavior data (search text of searching behaviors in a historical time period) of the account, and historical click behavior data (resource information of multimedia resources clicked in a historical time period) of the account from the electronic device 12.

[0053] Meanwhile, the recommendation device 11 can also obtain search text and resource information of candidate multimedia resources from the electronic device 12.

[0054] The recommendation device 11 can also perform the method for determining a multimedia resource in the embodiments of the present disclosure, for example, according to the account information of the account and the historical search behavior data and the historical click behavior data of the account, determining the relevant factors and the irrelevant factors of the historical search behavior to the historical click behavior data, and further determining the account reconstruction features of the account according to the determined relevant factors and irrelevant factors. At the same time, the recommendation device 11 also determines the relevant factors and the irrelevant factors of the search text of the candidate multimedia resource to the resource information, and further determines the resource reconstruction features of the candidate multimedia resource according to the determined relevant factors and irrelevant factors.

[0055] Further, the recommendation device can also predict the interaction parameters of the account to each candidate multimedia resource based on the determined account reconstruction features, resource reconstruction features, and pre-trained prediction model, and further determine the multimedia resource for recommending to the account according to the predicted interaction parameters.

[0056] It should be noted that the interaction behavior involved in the embodiments of the present disclosure can specifically include the behaviors of the account clicking, playing, liking, following, and sharing the multimedia resource. The interaction parameter involved in the embodiments of the present disclosure can specifically include the click rate of the account to the multimedia resource, and can also include the like rate or following behavior of the account to the multimedia resource, or a parameter calculated based on the above click rate, like rate, and following behavior, which is used to reflect the preference degree of the account to the multimedia resource.

[0057] In another case, the recommendation device can also obtain the training sample of the prediction model based on the above method for determining the account reconstruction features and the resource reconstruction features, and train the prediction model according to the training sample.

[0058] It should be noted that the multimedia resource involved in the embodiments of the present disclosure can include video, audio, text, and the like, which is not specifically limited herein. Meanwhile, in the subsequent description of the embodiments of the present disclosure, video is taken as an example for subsequent description, and audio and text and other resources can be referred to the subsequent description.

[0059] The recommendation device 11 and the electronic device 12 can be independent devices, or can be integrated into the same device, which is not specifically limited herein.

[0060] When the recommendation device 11 and the electronic device 12 are integrated into the same device, the communication mode between the recommendation device 11 and the electronic device 12 is the communication between the internal modules of the device. In this case, the communication process between the two is the same as the communication process between the recommendation device 11 and the electronic device 12 when they are independent of each other.

[0061] In the following embodiments provided by the present application, the present application is described by taking the example that the recommendation device 11 and the electronic device 12 are independently arranged.

[0062] In actual application, the multimedia resource determination method provided by the embodiments of the present application can be applied to the recommendation device or the electronic device. In the following, the multimedia resource determination method provided by the embodiments of the present application is described by taking the example that the multimedia resource determination method is applied to the electronic device.

[0063] As shown in Figure 3 The multimedia resource determination method provided by the embodiments of the present application includes the following S201-S203.

[0064] S201, the electronic device acquires the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource in the plurality of candidate multimedia resources.

[0065] The account reconstruction feature is used to represent the relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account. The resource reconstruction feature is used to represent the relevant factors and irrelevant factors of the multimedia resource in the search result under the search condition affecting the interactive behavior of the account.

[0066] As a possible implementation manner, for the account reconstruction feature of the current account, the electronic device acquires the account feature of the current account, the resource feature of the historical multimedia resource on which the current account has performed the interactive behavior in the historical time period, and the search condition feature of each historical multimedia resource, and determines the resource reconstruction feature of each historical multimedia resource according to the resource feature of each historical multimedia resource and the search condition feature of each historical multimedia resource.

[0067] Further, the electronic device inputs the account feature of the current account and the resource reconstruction feature of each historical multimedia resource into the preset account portrait model to obtain the account reconstruction feature of the current account.

[0068] It can be understood that the account reconstruction feature is the account portrait feature, which includes the influence factors of the historical interactive behavior on the account portrait feature. In the above influence factors, the part of the interactive behavior caused by the historical search behavior is the relevant factor, and the part of the interactive behavior caused by the characteristics of the multimedia resource itself is the irrelevant factor.

[0069] Meanwhile, for the resource reconstruction feature of each candidate multimedia resource, the electronic device acquires the resource feature of each candidate multimedia resource and the search condition feature of each candidate multimedia resource, and determines the resource reconstruction feature of each candidate multimedia resource according to the resource feature of each candidate multimedia resource and the search condition feature of each candidate multimedia resource.

[0070] It should be noted that the search condition feature involved in the embodiments of the present disclosure can be a search text feature.

[0071] It can be understood that the resource reconstruction feature includes relevant factors and irrelevant factors of the influence of the multimedia resource on the interaction parameter, wherein the relevant factors are the causal relationship between the search condition of the multimedia resource and the clicking of the multimedia resource, and the irrelevant factors are the non-causal relationship between the above-mentioned confusion factors and the clicking of the multimedia resource.

[0072] It should be noted that the current account can be an account of opening a preset client or application program, and the preset client or application program applies the above-mentioned recommendation system.

[0073] Exemplarily, after the current account logs in or the method above-mentioned preset client or application program, the electronic device obtains the account reconstruction feature of the current account, and simultaneously obtains the resource reconstruction feature of each candidate video in the 100 candidate videos.

[0074] The specific implementation of this step can refer to the subsequent description of the present disclosure, and will not be described here.

[0075] S202, the electronic device predicts the predicted interaction parameter of the current account to each candidate multimedia resource according to the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource.

[0076] The predicted interaction parameter is used to represent the probability of the account performing an interaction behavior on the candidate multimedia resource.

[0077] Exemplarily, the above-mentioned interaction behavior can include clicking, playing, liking, following, sharing and the like.

[0078] As a possible implementation manner, the electronic device inputs the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource into a pre-trained prediction model respectively, and takes the output result of the prediction model as the predicted interaction parameter of each candidate multimedia resource.

[0079] Exemplarily, in combination with the example in the above-mentioned S201, the electronic device inputs the account reconstruction feature of the account and the resource reconstruction feature of the candidate video 1 in the 100 candidate videos into the prediction model, and obtains the predicted interaction parameter of the candidate video 1. Subsequently, the electronic device inputs the account reconstruction feature of the account and the resource reconstruction feature of the candidate video 2 in the 100 candidate videos into the prediction model, and obtains the predicted interaction parameter of the candidate video 2. Similarly, in turn, until the predicted interaction parameters of the 100 candidate videos are obtained.

[0080] It should be noted that the prediction model is obtained by pre-training the electronic device according to the training sample and the label. The training sample includes the account reconstruction feature of the sample account and the resource reconstruction feature of the sample multimedia resource, and the label includes the sample interaction parameter of the sample multimedia resource.

[0081] The sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in the sample time period.

[0082] S203, the electronic device determines the to-be-recommended multimedia resource for recommending to the current account from the plurality of candidate multimedia resources according to the predicted interaction parameter of each candidate multimedia resource.

[0083] As a possible implementation manner, the electronic device can sort the predicted interaction parameters of the plurality of candidate multimedia resources, and select a preset number of candidate multimedia resources with the highest predicted interaction parameters as the to-be-recommended multimedia resources.

[0084] The specific implementation of this step can refer to the prior art, which will not be described here.

[0085] The technical solution provided by the present disclosure at least brings the following beneficial effects: considering the causal relationship between the search behavior of the account and the interaction behavior of the account, using the account reconstruction feature to represent the behavior characteristics of the account, the relevant factors and irrelevant factors of the search behavior of the account affecting the interaction behavior of the account can make the structure of the determined account behavior characteristics more clear and more refined, thereby improving the accuracy of the account behavior characteristics. At the same time, considering the causal relationship between the search condition of the multimedia resource and the interaction behavior performed by the account, using the resource reconstruction feature to represent the relevant factors and irrelevant factors of the search condition of the multimedia resource affecting the interaction behavior performed by the account can make the result of representing the characteristics of the multimedia resource more clear and refined, thereby improving the accuracy of the characteristics of the multimedia resource. In this way, using the account reconstruction feature and the resource reconstruction feature as the basis for predicting the interaction parameter can maximize the distinction between the causal part and the non-causal part of the account and the candidate multimedia resource, thereby improving the accuracy of the predicted interaction parameter, and further improving the accuracy of the determined to-be-recommended multimedia resource.

[0086] In one design, in order to obtain the account reconstruction feature of the current account, as shown in Figure 4 The S201 provided by the embodiment of the present disclosure specifically includes the following S2011-S2013:

[0087] S2011, the electronic device obtains the account feature of the current account, the resource feature of the historical multimedia resource on which the current account has performed an interaction behavior in the historical time period, and the search condition feature of each historical multimedia resource.

[0088] As a possible implementation manner, the electronic device obtains account information of the current account, and converts the account information of the current account into account features of the current account.

[0089] For example, the account information of the current account can include an identity of the current account, a location of the current account, a gender of the current account, an age of the current account, and the like. The electronic device can convert the account information of the current account into a digital string, and merge the digital string to obtain a vector, and use the vector as the account features of the current account. In this case, the account features of the current account can be represented as [6, 15, 23, 45, 67], where 6 represents the identity of the current account, 15 represents the age of the current account, 23 represents the identity of the region where the current account is located, 45 represents the gender of the current account, and 67 represents the registration duration of the current account. The present embodiment is not limited in this regard.

[0090] Meanwhile, for the resource features of the historical multimedia resources, the electronic device determines historical multimedia resources on which the current account has performed interaction behaviors in a historical time period, and obtains resource information of each historical multimedia resource. Further, the electronic device converts the resource information of each historical multimedia resource into a digital string, and merges the digital string to obtain a vector, and uses the vector as the resource features of the historical multimedia resources.

[0091] It should be noted that the historical time period can be any time period before the current account logs in or accesses a preset client or application.

[0092] For example, the resource information of the multimedia resource can include an identity of the multimedia resource, cover information, description information, subtitle information, an identity of an author, and the like. The electronic device can convert the resource information of the historical multimedia resource into a digital string, and merge the digital string to obtain the resource features of each historical multimedia resource. In this case, the resource features of each historical multimedia resource can be represented as [987, 654, 321, 012, 234], where 987 represents the identity of the historical multimedia resource, 654 represents the cover information of the historical multimedia resource, 321 represents the description information of the historical multimedia resource, 012 represents the subtitle information of the historical multimedia resource, and 234 represents the identity of the author of the historical multimedia resource. The present embodiment is not limited in this regard.

[0093] Further, for the search condition features of the historical multimedia resources, the electronic device obtains a search condition of each historical multimedia resource, and converts the search condition into a search condition feature.

[0094] It should be noted that the search condition can include a search text of the multimedia resource, a search channel, a search keyword, and the like.

[0095] For example, the search condition feature can be a matrix, where each row corresponds to a search condition of a historical multimedia resource. In this case, the search condition feature can be a 10*64 matrix.

[0096] S2012, the electronic device performs feature reconstruction processing on the resource feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource, to obtain a resource reconstruction feature of each historical multimedia resource.

[0097] As a possible implementation manner, the electronic device performs regression operation on the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource in the same feature space, to obtain a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource, respectively.

[0098] It can be understood that performing regression operation on the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource can determine the causal relationship between the search condition feature and the resource feature, and can determine the relevant factors and the irrelevant factors between the clicked multimedia resource and the search condition. The fitting vector is the relevant factor, and the residual vector is the irrelevant factor.

[0099] Further, the electronic device weights the fitting vector of each historical multimedia resource and the residual vector of each historical multimedia resource, to obtain the resource reconstruction feature of each historical multimedia resource.

[0100] The specific implementation manner of this step can be referred to the subsequent description of the embodiments of the present disclosure, which will not be described here.

[0101] S2013, the electronic device obtains an account reconstruction feature of the current account according to the resource reconstruction feature of each historical multimedia resource and the account feature of the current account.

[0102] As a possible implementation manner, the electronic device inputs the resource reconstruction feature of each historical multimedia resource and the account feature of the current account into a pre-trained account portrait model, and takes the output result of the account portrait model as the account reconstruction feature of the current account.

[0103] For example, if there are 50 historical multimedia resources, the electronic device inputs the account feature of the current account and the resource reconstruction feature of the 50 historical multimedia resources into the account portrait model, to obtain the account reconstruction feature of the current account.

[0104] In some embodiments, Figure 5 An example of determining the account reconstruction feature is shown, as Figure 5As shown, the electronic device obtains the account feature of the current account, the resource feature of the historical multimedia resources, and the search condition feature of each historical multimedia resource, and performs feature reconstruction processing on the resource feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource, to obtain the resource reconstruction feature of each historical multimedia resource. Further, the electronic device inputs the account feature of the current account and the resource reconstruction feature of each historical multimedia resource into a preset account portrait model, to obtain the account reconstruction feature of the current account.

[0105] The technical solutions provided by the present disclosure at least bring the following beneficial effects: the account reconstruction feature of the current account is determined by the resource reconstruction features of the plurality of historical multimedia resources and the account feature of the current account, so that the sampling data of the predicted interaction parameter can include relevant factors and irrelevant factors that affect the interaction behavior of the account.

[0106] In one design, in order to be able to determine the resource reconstruction feature of each historical multimedia resource, as Figure 6 As shown, S2012 provided by the embodiments of the present disclosure specifically includes the following S301-S302.

[0107] S301, the electronic device determines the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource.

[0108] The linear feature of each historical multimedia resource is used to represent the relevant factors of each historical multimedia resource affecting the account to perform the interaction behavior under the search condition, and the nonlinear feature of each historical multimedia resource is used to represent the irrelevant factors of each historical multimedia resource affecting the account to perform the interaction behavior under the search condition.

[0109] As a possible implementation manner, the electronic device performs a preset regression operation on the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource, to obtain the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource.

[0110] The specific description of this step can refer to the subsequent description of the embodiments of the present disclosure, which will not be described here.

[0111] S302, the electronic device weights the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource, to obtain the resource reconstruction feature of each historical multimedia resource.

[0112] As a possible implementation manner, the electronic device weights the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource based on a preset weight, to obtain a resource reconstruction feature of each historical multimedia resource.

[0113] The technical solutions provided by the present disclosure at least bring the following beneficial effects: by weighting the linear feature and the nonlinear feature, the resource reconstruction feature can be determined, and since the linear feature is used to represent the relevant factors (causal part) and the nonlinear feature is used to represent the non-relevant factors (non-causal part), the relevant factors and the non-relevant factors can be ensured to be included in the resource reconstruction feature, thereby improving the accuracy of the determined multimedia resource to be recommended.

[0114] In some embodiments, as shown in Figure 7 To be able to determine the linear feature and the nonlinear feature of the historical multimedia resource, the S301 provided by the embodiments of the present disclosure specifically includes the following S3011-S3012.

[0115] S3011, the electronic device obtains a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource according to the search condition feature of each historical multimedia resource, the resource feature of each historical multimedia resource, and a preset regression operation.

[0116] As a possible implementation manner, the electronic device performs feature space conversion on the resource feature of each historical multimedia resource, to obtain the converted resource feature of each historical multimedia resource.

[0117] The feature space where the converted resource feature of each historical multimedia resource is located is linearly related to the feature space where the search condition feature of each historical multimedia resource is located.

[0118] It can be understood that the purpose of performing feature space conversion on the resource feature is to be able to perform a regression operation on the converted resource feature and the search condition feature subsequently.

[0119] For example, the electronic device can input the resource feature of the historical multimedia resource into a preset space conversion model, and determine the output result of the space conversion model as the converted resource feature of the historical multimedia resource.

[0120] The space conversion model can be a multilayer perceptron (MLP) model.

[0121] Further, the electronic device takes the search condition feature of each historical multimedia resource as an instrumental variable, takes the converted resource feature of each historical multimedia resource as a representation vector, and performs the above-mentioned preset regression operation on the search condition feature of each historical multimedia resource and the converted resource feature of each historical multimedia resource, to obtain a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource, respectively.

[0122] Specifically, the electronic device takes the search condition feature of each historical multimedia resource as an instrumental variable, takes the converted resource feature of each historical multimedia resource as a representation vector, and performs a two-stage least square regression (LSR) operation on the search condition feature of each historical multimedia resource and the converted resource feature of each historical multimedia resource, to obtain a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource, respectively.

[0123] The specific implementation of the LSR operation in this step can refer to the description in the prior art, and will not be described here.

[0124] It can be understood that the fitting vector is a relevant factor in the resource reconstruction feature, that is, corresponds to a relevant reason for causing the multimedia resource to be clicked. The residual vector is an irrelevant factor in the resource reconstruction feature, that is, corresponds to an irrelevant reason (confusion factor) for causing the multimedia resource to be reduced.

[0125] S3012, the electronic device determines the fitting vector of each historical multimedia resource as a linear feature of each historical multimedia resource, and determines the residual vector of each historical multimedia resource as a nonlinear feature of each historical multimedia resource.

[0126] In some embodiments, Figure 8 An illustrative diagram of performing a feature reconstruction operation is shown, as Figure 8 As shown, the electronic device inputs the resource feature of the historical multimedia resource into a resource feature input space conversion model to obtain a converted resource feature of the historical multimedia resource. Further, the electronic device takes the search condition feature of the historical multimedia resource as an instrumental variable, takes the converted resource feature of the historical multimedia resource as a representation vector, and performs an LSR operation on the search condition feature of the historical multimedia resource and the converted resource feature of the historical multimedia resource, to obtain a fitting vector of the historical multimedia resource and a residual vector of the historical multimedia resource, respectively. Finally, the electronic device weights the fitting vector of the historical multimedia resource and the residual vector of the historical multimedia resource to obtain a resource reconstruction feature of the historical multimedia resource.

[0127] The technical scheme provided by the present disclosure has at least the following beneficial effects: through a regression manner, a fitting vector (causal part) of multimedia resources being clicked due to account search behavior and a residual vector of multimedia resources being clicked by an account due to their own characteristics (non-causal part) can be determined based on the causal relationship between the account search behavior and the account interaction behavior, so that the fitting vector is determined as a linear feature and the residual vector is determined as a non-linear feature, which can make the determined linear feature and non-linear feature more accurate, thereby ensuring the accuracy of subsequent resource reconstruction features.

[0128] In one design, the weight of the fitting vector and the weight of the residual vector can also be determined by the device itself, as shown in Figure 9 The determination method of the multimedia resource provided by the embodiment of the present disclosure also includes the following S401-S404.

[0129] S401, the electronic device converts the search condition feature of each historical multimedia resource into a feature dimension to obtain the converted search condition feature of each historical multimedia resource.

[0130] The converted search condition feature of each historical multimedia resource has the same dimension as the resource feature of each historical multimedia resource.

[0131] For example, if the search condition feature of the historical multimedia resource is a 10*64 matrix and the resource feature of the historical multimedia resource is a 1*32 vector, the 10*64 search condition feature is converted into a 1*640 vector to achieve the same dimension as the resource feature.

[0132] S402, the electronic device combines the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource to obtain the combined feature of each historical multimedia resource.

[0133] As a possible implementation manner, the electronic device combines the converted search condition feature and the resource feature to obtain the combined feature of each historical multimedia resource.

[0134] For example, in the case of the converted search condition feature being a 1*640 vector and the resource feature being a 1*32 vector, the combined vector obtained by combination is a 1*672 vector.

[0135] S403, the electronic device inputs the combined feature of each historical multimedia resource into a pre-trained first weight model to obtain the weight of the linear feature of each historical multimedia resource.

[0136] The first weight model includes a first parameter, and the first parameter is used to learn the converted search condition feature in the combined feature.

[0137] As a possible implementation manner, the electronic device inputs the merged vector into the first weight model, and takes the output result of the first weight model as the weight of the linear feature of each historical multimedia resource.

[0138] Exemplarily, the first weight model can be a multilayer perceptron (MLP) model.

[0139] S404, the electronic device inputs the merged feature of each historical multimedia resource into the pre-trained second weight model to obtain the weight of the nonlinear feature of each historical multimedia resource.

[0140] The second weight model includes a second parameter, and the second parameter is used to learn the resource feature in the merged feature.

[0141] As a possible implementation manner, the electronic device inputs the merged vector into the second weight model, and takes the output result of the second weight model as the weight of the nonlinear feature of each historical multimedia resource.

[0142] Exemplarily, the second weight model can be a multilayer perceptron (MLP) model. In some embodiments, as shown in Figure 8 The search condition feature of each historical multimedia resource is converted in feature dimension to obtain the converted search condition feature of each historical multimedia resource, and the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource are merged to obtain the merged feature of each historical multimedia resource. Further, the electronic device respectively inputs the merged feature of each historical multimedia resource into the pre-trained first weight model and the second weight model to obtain the first weight corresponding to the fitting vector of each historical multimedia resource and the second weight corresponding to the residual vector of each historical multimedia resource.

[0143] The technical solutions provided by the present disclosure at least bring the following beneficial effects: the weight of the fitting vector and the weight of the residual vector are determined from the merged model through the preset weight model, which can make the determined weight more accurate, thereby making the subsequently determined resource reconstruction feature more accurate.

[0144] In one design, in order to obtain the resource reconstruction feature of each candidate multimedia resource, S201 provided by the embodiments of the present disclosure specifically further includes the following S2014-S2016:

[0145] S2014, the electronic device obtains the resource feature of each candidate multimedia resource and the search condition feature of each candidate multimedia resource.

[0146] The specific implementation of this step can refer to the specific description of the above S2011 of the embodiments of the present disclosure for acquiring the resource features of the historical multimedia resources and the search condition features of the historical multimedia, except that the execution objects are different, which will not be described here.

[0147] S2015, the electronic device respectively determines the linear feature of each candidate multimedia resource and the nonlinear feature of each candidate multimedia resource according to the search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource.

[0148] The linear feature of each candidate multimedia resource is used to represent the relevant factors of each candidate multimedia resource affecting the interactive behavior of the account under the search condition, and the nonlinear feature of each candidate multimedia resource is used to represent the non-relevant factors of each candidate multimedia resource affecting the interactive behavior of the account under the search condition.

[0149] The specific implementation of this step can refer to the specific description of the above S301 of the embodiments of the present disclosure, which will not be described here.

[0150] S2016, the electronic device weights the linear feature of each candidate multimedia resource and the nonlinear feature of each candidate multimedia resource to obtain the resource reconstruction feature of each candidate video.

[0151] The specific implementation of this step can refer to the specific description of the above S302 of the embodiments of the present disclosure for weighting the fitting vector of each historical multimedia resource and the residual vector of each historical multimedia resource, except that the execution objects are different, which will not be described here.

[0152] The technical solutions provided by the present disclosure at least bring the following beneficial effects: by weighting the linear feature and the nonlinear feature, the resource reconstruction feature can be determined, and since the linear feature is used to represent the relevant factors (causal part) and the nonlinear feature is used to represent the non-relevant factors (non-causal part), it can be ensured that the resource reconstruction feature includes the relevant factors and the non-relevant factors, thereby improving the accuracy of the determined multimedia resources to be recommended.

[0153] In one design, in order to be able to determine the linear feature and the nonlinear feature of each candidate multimedia resource, the above S2015 provided by the embodiments of the present disclosure specifically includes the following S2015a-S2015b.

[0154] S2015a, the electronic device respectively obtains the fitting vector of each candidate multimedia resource and the residual vector of each candidate multimedia resource according to the search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource and a preset regression operation.

[0155] The specific implementation of this step can refer to the specific description of determining the fitting vector and the residual vector of each historical multimedia resource in S3011 of the foregoing embodiment of the disclosure, except that the execution object is different, which will not be described here again.

[0156] S2015b, the electronic device determines the fitting vector of each candidate multimedia resource as the linear feature of each candidate multimedia resource, and determines the residual vector of each candidate multimedia resource as the nonlinear feature of each candidate multimedia resource.

[0157] The technical solution provided by the disclosure brings at least the following beneficial effects: through the regression method, the fitting vector (causal part) of the multimedia resource being clicked due to the account search behavior and the residual vector of the multimedia resource being clicked by the account due to its own characteristics (non-causal part) can be determined based on the causal relationship between the account search behavior and the account interaction behavior. In this way, the linear feature and the nonlinear feature determined by determining the fitting vector as the linear feature and the residual vector as the nonlinear feature can be more accurate, thereby ensuring the accuracy of the subsequent resource reconstruction feature.

[0158] In one design, the weight of the linear feature (fitting vector) and the weight of the nonlinear feature (residual vector) of each candidate multimedia resource can also be determined by the device itself. The determination method of the multimedia resource provided by the embodiment of the disclosure further includes the following S501-S504.

[0159] S501, the electronic device performs feature dimension conversion on the search condition feature of each candidate multimedia resource to obtain the converted search condition feature of each candidate multimedia resource.

[0160] The converted search condition feature of each candidate multimedia resource has the same dimension as the resource feature of each candidate multimedia resource.

[0161] The specific implementation of this step can refer to the specific description of performing feature dimension conversion on the search condition feature of each historical multimedia resource in S401 of the foregoing embodiment of the disclosure, except that the execution object is different, which will not be described here again.

[0162] S502, the electronic device combines the converted search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource to obtain the combined feature of each candidate multimedia resource.

[0163] The specific implementation of this step can refer to the specific description of combining the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource in S402 of the foregoing embodiment of the disclosure, except that the execution object is different, which will not be described here again.

[0164] S503, the electronic device inputs the merged feature of each candidate multimedia resource into the pre-trained first weight model to obtain the weight of the linear feature of each candidate multimedia resource.

[0165] The first weight model includes first parameters, and the first parameters are used to learn the converted search condition feature in the merged feature.

[0166] The specific implementation of this step can refer to the specific description of the step of inputting the merged feature of each historical multimedia resource into the pre-trained first weight model to obtain the weight of the linear feature of each historical multimedia resource in S403 of the embodiment of the present disclosure, and the difference is that the execution object is different, which will not be described here.

[0167] S504, the electronic device inputs the merged feature of each candidate multimedia resource into the pre-trained second weight model to obtain the weight of the nonlinear feature of each candidate multimedia resource.

[0168] The second weight model includes second parameters, and the second parameters are used to learn the resource feature in the merged feature.

[0169] The specific implementation of this step can refer to the specific description of the step of inputting the merged feature of each historical multimedia resource into the pre-trained second weight model to obtain the weight of the nonlinear feature of each historical multimedia resource in S404 of the embodiment of the present disclosure, and the difference is that the execution object is different, which will not be described here.

[0170] The technical solution provided by the present disclosure at least brings the following beneficial effects: the weight of the linear feature and the weight of the nonlinear feature are determined from the merged model by the pre-set weight model, which can make the determined weight more accurate, so that the subsequent determined resource reconstruction feature is more accurate.

[0171] In one design, since the predicted interaction parameter of each candidate multimedia candidate resource is obtained based on the pre-trained prediction model, in order to train the above-mentioned prediction model, the method for determining a multimedia resource provided by the embodiment of the present disclosure further includes the following S601-S602:

[0172] S601, the electronic device obtains the account reconstruction feature of the sample account, the resource reconstruction feature of the sample multimedia resource, and the sample interaction parameter of the sample multimedia resource.

[0173] The sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period.

[0174] It should be noted that the sample interaction parameter of the sample multimedia resource is a probability of the sample account performing an interaction behavior on the sample multimedia resource in a sample time period. The sample time period is a time period before the sample account logs in or accesses a preset client or application.

[0175] The specific implementation of the electronic device obtaining the account reconstruction feature of the sample account and the resource reconstruction feature of the sample multimedia resource in this step can refer to the specific description in S201, S2011-S2013 of the present disclosure, and the difference is that the processing objects are different, and details are not repeated here.

[0176] S602, the account reconstruction feature of the sample account and the resource reconstruction feature of the sample multimedia resource are taken as sample features, the sample interaction parameter of the sample multimedia resource is taken as a label, and a preset neural network model is trained to obtain a prediction model.

[0177] The specific implementation of this step can refer to the description in the prior art, and details are not repeated here.

[0178] The technical solution provided by the present disclosure at least brings the following beneficial effects: the prediction model obtained by training by this method can enable the prediction model to learn the relevant factors of the search behavior of the account on the interaction behavior of the account and the irrelevant factors in the interaction behavior of the account, and further enable the prediction model obtained by training to learn the causal part and the non-causal part of the account and the candidate multimedia resource, which can improve the accuracy of the predicted interaction parameter, thereby improving the accuracy of the determined multimedia to be recommended.

[0179] In actual application, the present application embodiment further provides a prediction model training method, which can be applied to a recommendation device or an electronic device. Hereinafter, the prediction model training method provided by the present application embodiment is described by taking the prediction model training method applied to an electronic device as an example.

[0180] As shown in FIG. 7, Figure 10 The prediction model training method provided by the present disclosure embodiment includes the following S701-S702.

[0181] S701, the electronic device obtains the account reconstruction feature of the sample account, the resource reconstruction feature of the sample multimedia resource, and the sample interaction parameter of the sample multimedia resource.

[0182] The account reconstruction feature is used to represent the relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account. The sample multimedia resource is a multimedia resource clicked by the sample account in a sample time period. The resource reconstruction feature is used to represent the relevant factors and irrelevant factors of the multimedia resource in the search result affecting the interactive behavior of the account under the search condition.

[0183] It should be noted that the sample interactive parameter of the sample multimedia resource is the probability of the sample account clicking the sample multimedia resource in the sample time period. The sample time period is a time period before the sample account logs in or accesses the preset client or application.

[0184] The specific implementation of the electronic device obtaining the account reconstruction feature of the sample account and the resource reconstruction feature of the sample multimedia resource in this step can refer to the specific description in the foregoing S201, S2011-S2013 of the present disclosure, and the difference is that the processing objects are different, and details are not described herein.

[0185] S702, the electronic device trains the preset neural network model by taking the account reconstruction feature of the sample account and the resource reconstruction feature of the sample multimedia resource as sample features and taking the sample interactive parameter of the sample multimedia resource as a label, to obtain a prediction model.

[0186] The prediction model is used to predict the predicted interactive parameter of the sample account to the candidate multimedia resource.

[0187] The specific implementation of this step can refer to the description in the prior art, and details are not described herein.

[0188] The technical solutions provided by the present disclosure at least bring the following beneficial effects: considering the causal relationship between the search behavior of the account and the interactive behavior of the account, the account reconstruction feature is used to represent the relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account in the behavior feature of the account, which can make the structure of the determined behavior feature of the account more clear and more refined, thereby improving the accuracy of the behavior feature of the account. At the same time, considering the causal relationship between the multimedia resource being searched and being clicked, the resource reconstruction feature is used to represent the relevant factors and irrelevant factors of the search condition of the multimedia resource affecting the interactive behavior of the account, which can make the result of representing the feature of the multimedia resource more clear and refined, thereby improving the accuracy of the feature of the multimedia resource. In this way, the prediction model trained can learn the causal part and the non-causal part of the account and the multimedia resource to the greatest extent, thereby improving the accuracy of the predicted interactive parameter, and thus improving the accuracy of the determined multimedia to be recommended.

[0189] In one design, to be able to obtain the account reconstruction feature of the sample account, the S701 provided by the embodiments of the present disclosure specifically includes the following S7011-S7013.

[0190] S7011, the electronic device obtains the account feature of the sample account, the resource feature of the historical sample multimedia resource on which the sample account has performed the interaction behavior in the historical sample time period, and the search condition feature of each historical sample multimedia resource.

[0191] The historical sample time period is a time period before the sample time period.

[0192] The specific implementation of this step can refer to the specific description of obtaining the resource feature of the historical multimedia resource and the search condition feature of the historical multimedia in S2011 of the embodiments of the present disclosure, and the difference is that the execution object is different, which will not be described here.

[0193] S7012, the electronic device performs feature reconstruction processing on the resource feature of each historical sample multimedia resource according to the search condition feature of each historical sample multimedia resource, to obtain the resource reconstruction feature of each historical sample video.

[0194] The specific implementation of this step can refer to the specific description of performing feature reconstruction processing on the resource feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource in S2012 of the embodiments of the present disclosure, and the difference is that the execution object is different, which will not be described here.

[0195] S7013, the electronic device obtains the account reconstruction feature of the sample account according to the historical resource reconstruction feature of each historical sample multimedia resource and the account feature of the sample account.

[0196] The specific implementation of this step can refer to the specific description of inputting the resource reconstruction feature of each historical multimedia resource and the account feature of the current account into the pre-trained account portrait model to obtain the account reconstruction feature of the current account in S2013 of the embodiments of the present disclosure, and the difference is that the execution object is different, which will not be described here.

[0197] The technical solutions provided by the present disclosure at least bring the following beneficial effects: the account reconstruction feature of the current account is determined by the resource reconstruction feature of the plurality of historical sample multimedia resources and the account feature of the current account, so that the sampling data of the predicted interaction parameter can include relevant factors and irrelevant factors that affect the interaction behavior of the account.

[0198] In one design, to be able to determine the resource reconstruction feature of each historical sample multimedia resource, the S7012 provided by the embodiments of the present disclosure specifically includes the following S801-S802.

[0199] S801, the electronic device determines linear features of each historical sample multimedia resource and nonlinear features of each historical sample multimedia resource according to the search condition features of each historical sample multimedia resource and the resource features of each historical sample multimedia resource.

[0200] The linear features of each historical sample multimedia resource are used to represent relevant factors of each historical sample multimedia resource affecting the account to perform interactive behaviors under the search condition, and the nonlinear features of each historical sample multimedia resource are used to represent irrelevant factors of each historical sample multimedia resource affecting the account to perform interactive behaviors under the search condition.

[0201] The specific implementation of this step can refer to the specific description of S301 in the above-mentioned embodiments of the present disclosure, except that the execution object is different, which will not be described here.

[0202] S802, the electronic device weights the linear features of each historical sample multimedia resource and the nonlinear features of each historical sample multimedia resource to obtain resource reconstruction features of each historical sample multimedia resource.

[0203] The specific implementation of this step can refer to the specific description of weighting the linear features of each historical multimedia resource and the nonlinear features of each historical multimedia resource in the above-mentioned S302 of the embodiments of the present disclosure, except that the execution object is different, which will not be described here.

[0204] The technical solutions provided by the present disclosure at least bring the following beneficial effects: through the regression method, the relevant factors (causal part) of the multimedia resource being clicked due to the account search behavior and the irrelevant factors (non-causal part) of the multimedia resource being clicked by the account due to its own characteristics can be determined based on the causal relationship between the account search behavior and the account interactive behavior, and the weighting of the linear features and the nonlinear features can ensure that the resource reconstruction features include the relevant factors and the irrelevant factors, thereby improving the accuracy of the determined multimedia resources to be recommended.

[0205] In one design, in order to determine the linear features and the nonlinear features of the historical sample multimedia resource, S801 provided by the embodiments of the present disclosure specifically includes S8011-S8012.

[0206] S8011, the electronic device obtains a fitting vector of each historical sample multimedia resource and a residual vector of each historical sample multimedia resource according to the search condition features of each historical sample multimedia resource, the resource features of each historical sample multimedia resource, and a preset regression operation.

[0207] The specific implementation of this step can refer to the specific description of S3011 in the foregoing embodiments of the present disclosure, except that the execution object is different, and thus no further description is provided herein.

[0208] S8012, the electronic device determines the fitting vector of each historical sample multimedia resource as the linear feature of each historical sample multimedia resource, and determines the residual vector of each historical sample multimedia resource as the nonlinear feature of each historical sample multimedia resource.

[0209] The technical solutions provided by the present disclosure at least bring the following beneficial effects: through the regression manner, the fitting vector (causal part) of the multimedia resource being clicked due to the account search behavior and the residual vector of the multimedia resource being clicked by the account due to its own characteristics (non-causal part) can be determined based on the causal relationship between the account search behavior and the account interaction behavior, so that the linear feature and the nonlinear feature determined by determining the fitting vector as the linear feature and the residual vector as the nonlinear feature can be more accurate, thereby ensuring the accuracy of the subsequent resource reconstruction feature.

[0210] In one design, the weight of the fitting vector and the weight of the residual vector can also be determined by the device itself, and the training method of the prediction model provided by the embodiments of the present disclosure further includes the following S901-S904.

[0211] S901, the electronic device performs feature dimension conversion on the search condition feature of each historical sample multimedia resource to obtain the converted search condition feature of each historical sample multimedia resource.

[0212] The converted search condition feature of each historical sample multimedia resource has the same dimension as the resource feature of each historical sample multimedia resource.

[0213] The specific implementation of this step can refer to the specific description of the feature dimension conversion on the search condition feature of each historical multimedia resource in the foregoing S401 of the embodiments of the present disclosure, except that the execution object is different, and thus no further description is provided herein.

[0214] S902, the electronic device combines the converted search condition feature of each historical sample multimedia resource and the resource feature of each historical sample multimedia resource to obtain the combined feature of each historical sample multimedia resource.

[0215] The specific implementation of this step can refer to the specific description of the combination of the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource in the foregoing S402 of the embodiments of the present disclosure, except that the execution object is different, and thus no further description is provided herein.

[0216] S903, the electronic device inputs the combined feature of each historical sample multimedia resource into the pre-trained first weight model to obtain the weight of the fitting vector of each historical sample multimedia resource.

[0217] The first weight model includes a first parameter, and the first parameter is used to learn the converted search condition feature in the combined feature.

[0218] The specific implementation of this step can refer to the specific description of the step of inputting the combined feature of each historical multimedia resource into the pre-trained first weight model to obtain the weight of the fitting vector of each historical multimedia resource in S403 of the foregoing embodiment of the disclosure, with the difference being that the execution object is different, and details are not repeated here.

[0219] S904, the electronic device inputs the combined feature of each historical sample multimedia resource into the pre-trained second weight model to obtain the weight of the residual vector of each historical sample multimedia resource.

[0220] The second weight model includes a second parameter, and the second parameter is used to learn the resource feature in the combined feature.

[0221] The specific implementation of this step can refer to the specific description of the step of inputting the combined feature of each historical multimedia resource into the pre-trained second weight model to obtain the weight of the residual vector of each historical multimedia resource in S404 of the foregoing embodiment of the disclosure, with the difference being that the execution object is different, and details are not repeated here.

[0222] The technical solution provided by the disclosure at least brings the following beneficial effects: the weight of the fitting vector and the weight of the residual vector are determined from the combined model through the pre-set weight model, which can make the determined weight more accurate, so that the subsequent determined resource reconstruction feature is more accurate.

[0223] In one design, in order to obtain the resource reconstruction feature of each sample multimedia resource, S701 provided by the embodiment of the disclosure, specifically further includes the following S7014-S7016:

[0224] S7014, the electronic device obtains the resource feature of the sample multimedia resource and the search condition feature of the sample multimedia resource.

[0225] The specific implementation of this step can refer to the specific description of the step of obtaining the resource feature of the historical multimedia resource and the search condition feature of the historical multimedia in S2011 of the foregoing embodiment of the disclosure, with the difference being that the execution object is different, and details are not repeated here.

[0226] S7015, the electronic device determines linear features of the sample multimedia resources and nonlinear features of the sample multimedia resources according to the search condition features of the sample multimedia resources and the resource features of the sample multimedia resources.

[0227] The linear features of the sample multimedia resources are used to represent relevant factors of the sample multimedia resources affecting the account to perform the interactive behavior under the search condition, and the nonlinear features of the sample multimedia resources are used to represent irrelevant factors of the sample multimedia resources affecting the account to perform the interactive behavior under the search condition.

[0228] The specific implementation of this step can refer to the specific description of S301 in the foregoing embodiments of the present disclosure, which will not be described here again.

[0229] S7016, the electronic device weights the fitting vectors of the sample multimedia resources and the residual vectors of the sample multimedia resources to obtain resource reconstruction features of the sample video.

[0230] The specific implementation of this step can refer to the specific description of the weighting of the fitting vectors of each historical multimedia resource and the residual vectors of each historical multimedia resource in S302 in the foregoing embodiments of the present disclosure, and the difference is that the execution objects are different, which will not be described here again.

[0231] The technical solutions provided by the present disclosure at least bring the following beneficial effects: by weighting the linear features and the nonlinear features, the resource reconstruction features can be determined, and since the linear features are used to represent the relevant factors (causal part) and the nonlinear features are used to represent the irrelevant factors (non-causal part), it can be ensured that the resource reconstruction features include the relevant factors and the irrelevant factors, thereby improving the accuracy of the determined multimedia resources to be recommended.

[0232] In one design, in order to determine the linear features and the nonlinear features of each sample multimedia resource, the foregoing S7015 provided by the embodiments of the present disclosure specifically includes the following S7015a-S7015b.

[0233] S7015a, the electronic device respectively obtains the fitting vectors of each sample multimedia resource and the residual vectors of each sample multimedia resource according to the search condition features of each sample multimedia resource, the resource features of each sample multimedia resource, and a preset regression operation.

[0234] The specific implementation of this step can refer to the specific description of the determination of the fitting vectors and the residual vectors of each historical multimedia resource in S3011 in the foregoing embodiments of the present disclosure, and the difference is that the execution objects are different, which will not be described here again.

[0235] S7015b, the electronic device determines the fitting vector of each sample multimedia resource as a linear feature of each sample multimedia resource, and determines the residual vector of each sample multimedia resource as a nonlinear feature of each sample multimedia resource.

[0236] The technical solutions provided by the present disclosure at least bring the following beneficial effects: through a regression manner, the fitting vector (causal part) of a multimedia resource being clicked due to an account search behavior and the residual vector of the multimedia resource being clicked by the account due to its own features (non-causal part) can be determined based on the causal relationship between the account search behavior and the account interaction behavior, so that the determined linear features and nonlinear features are more accurate by determining the fitting vector as a linear feature and the residual vector as a nonlinear feature, thereby ensuring the accuracy of subsequent resource reconstruction features.

[0237] In one design, the weight of the fitting vector and the weight of the residual vector can also be determined by the device itself, and the training method of the prediction model provided by the embodiments of the present disclosure further includes the following S1001-S1004.

[0238] S1001, the electronic device performs feature dimension conversion on the search condition features of the sample multimedia resource to obtain converted search condition features of the sample multimedia resource.

[0239] The converted search condition features of the sample multimedia resource have the same dimension as the resource features of the sample multimedia resource.

[0240] The specific implementation of this step can refer to the specific description of the feature dimension conversion on the search condition features of each historical multimedia resource in S401 of the embodiments of the present disclosure described above, and the difference is that the execution object is different, which will not be described here.

[0241] S1002, the electronic device combines the converted search condition features of the sample multimedia resource and the resource features of the sample multimedia resource to obtain combined features of the sample multimedia resource.

[0242] The specific implementation of this step can refer to the specific description of combining the converted search condition features of each historical multimedia resource and the resource features of each historical multimedia resource in S402 of the embodiments of the present disclosure described above, and the difference is that the execution object is different, which will not be described here.

[0243] S1003, the electronic device inputs the combined features of the sample multimedia resource into a pre-trained first weight model to obtain the weight of the fitting vector of the sample multimedia resource.

[0244] The first weight model includes a first parameter, and the first parameter is used to learn the converted search condition features in the combined features.

[0245] The specific implementation of this step can refer to the specific description of inputting the merged feature of each historical multimedia resource into the pre-trained first weight model to obtain the weight of the fitting vector of each historical multimedia resource in the foregoing S403 of the embodiments of the present disclosure, with the difference being that the execution objects are different, and details are not described herein again.

[0246] S1004, the electronic device inputs the merged feature of the sample multimedia resource into the pre-trained second weight model to obtain the weight of the residual vector of the sample multimedia resource.

[0247] The second weight model includes a second parameter, and the second parameter is used to learn the resource feature in the merged feature.

[0248] The specific implementation of this step can refer to the specific description of inputting the merged feature of each historical multimedia resource into the pre-trained second weight model to obtain the weight of the residual vector of each historical multimedia resource in the foregoing S404 of the embodiments of the present disclosure, with the difference being that the execution objects are different, and details are not described herein again.

[0249] The technical solutions provided by the present disclosure at least bring the following beneficial effects: determining the weight of the fitting vector and the weight of the residual vector from the merged model through the pre-set weight model can make the determined weight more accurate, so that the subsequent determined resource reconstruction feature is more accurate.

[0250] Figure 11 is a structural schematic diagram of a multimedia resource determination apparatus according to an exemplary embodiment. Referring to Figure 11 As shown in the figure, the multimedia resource determination apparatus 110 provided by the embodiments of the present disclosure can be applied to an electronic device, and is used to execute the multimedia resource determination method provided by the above embodiments. The determination apparatus 110 includes an acquisition unit 1101, a prediction unit 1102, and a determination unit 1103.

[0251] The acquisition unit 1101 is configured to acquire an account reconstruction feature of a current account and a resource reconstruction feature of each candidate multimedia resource in a plurality of candidate multimedia resources. The account reconstruction feature is used to represent relevant factors and irrelevant factors of the search behavior of the account affecting the interactive behavior of the account. The resource reconstruction feature is used to represent relevant factors and irrelevant factors of the multimedia resource in the search result under the search condition affecting the interactive behavior of the account.

[0252] The prediction unit 1102 is configured to predict a predicted interactive parameter of the current account for each candidate multimedia resource according to the account reconstruction feature of the current account and the resource reconstruction feature of each candidate multimedia resource.

[0253] The determining unit 1103 is configured to determine, from the plurality of candidate multimedia resources, a to-be-recommended multimedia resource for recommendation to the current account according to the predicted interaction parameter of each candidate multimedia resource.

[0254] Optionally, as shown in Figure 11 The obtaining unit 1101 is specifically configured to:

[0255] Obtain an account feature of the current account, a resource feature of a historical multimedia resource on which the current account has performed an interaction behavior in a historical time period, and a search condition feature of each historical multimedia resource.

[0256] Perform feature reconstruction processing on the resource feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource, to obtain a resource reconstruction feature of each historical multimedia resource.

[0257] Obtain an account reconstruction feature of the current account according to the resource reconstruction feature of each historical multimedia resource and the account feature of the current account.

[0258] Optionally, as shown in Figure 11 The obtaining unit 1101 is specifically configured to:

[0259] Determine a linear feature of each historical multimedia resource and a nonlinear feature of each historical multimedia resource according to the search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource. The linear feature of each historical multimedia resource is used to represent a relevant factor of each historical multimedia resource affecting the account to perform the interaction behavior under the search condition, and the nonlinear feature of each historical multimedia resource is used to represent a non-relevant factor of each historical multimedia resource affecting the account to perform the interaction behavior under the search condition.

[0260] Weight the linear feature of each historical multimedia resource and the nonlinear feature of each historical multimedia resource, to obtain the resource reconstruction feature of each historical multimedia resource.

[0261] Optionally, as shown in Figure 11 The obtaining unit 1101 is specifically configured to:

[0262] Obtain a fitting vector of each historical multimedia resource and a residual vector of each historical multimedia resource according to the search condition feature of each historical multimedia resource, the resource feature of each historical multimedia resource, and a preset regression operation.

[0263] Determine the fitting vector of each historical multimedia resource as the linear feature of each historical multimedia resource, and determine the residual vector of each historical multimedia resource as the nonlinear feature of each historical multimedia resource.

[0264] Optionally, as shown in Figure 11 The apparatus provided by the embodiment of the present disclosure further includes a processing unit 1104.

[0265] The processing unit 1104 is configured to perform feature dimension conversion on the search condition feature of each historical multimedia resource to obtain a converted search condition feature of each historical multimedia resource. The converted search condition feature of each historical multimedia resource has the same dimension as the resource feature of each historical multimedia resource.

[0266] The processing unit 1104 is further configured to merge the converted search condition feature of each historical multimedia resource and the resource feature of each historical multimedia resource to obtain a merged feature of each historical multimedia resource.

[0267] The processing unit 1104 is further configured to input the merged feature of each historical multimedia resource into a pre-trained first weight model to obtain a weight of a linear feature of each historical multimedia resource. The first weight model includes first parameters, and the first parameters are used to learn the converted search condition feature in the merged feature.

[0268] The processing unit 1104 is further configured to input the merged feature of each historical multimedia resource into a pre-trained second weight model to obtain a weight of a nonlinear feature of each historical multimedia resource. The second weight model includes second parameters, and the second parameters are used to learn the resource feature in the merged feature.

[0269] Optionally, as shown in Figure 11 The acquisition unit 1101 provided by the embodiment of the present disclosure is specifically configured to:

[0270] Acquire the resource feature of each candidate multimedia resource and the search condition feature of each candidate multimedia resource.

[0271] According to the search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource, respectively determine a linear feature of each candidate multimedia resource and a nonlinear feature of each candidate multimedia resource. The linear feature of each candidate multimedia resource is used to represent relevant factors of each candidate multimedia resource affecting the account to perform interactive behavior under the search condition, and the nonlinear feature of each candidate multimedia resource is used to represent irrelevant factors of each candidate multimedia resource affecting the account to perform interactive behavior under the search condition.

[0272] Weight the linear feature of each candidate multimedia resource and the nonlinear feature of each candidate multimedia resource to obtain a resource reconstruction feature of each candidate video.

[0273] Optionally, as shown in Figure 11As shown, the embodiment of the present disclosure provides an acquisition unit 1101, which is specifically used for:

[0274] According to the search condition feature of each candidate multimedia resource, the resource feature of each candidate multimedia resource, and the preset regression operation, a fitting vector of each candidate multimedia resource and a residual vector of each candidate multimedia resource are respectively obtained.

[0275] The fitting vector of each candidate multimedia resource is determined as a linear feature of each candidate multimedia resource, and the residual vector of each candidate multimedia resource is determined as a nonlinear feature of each candidate multimedia resource.

[0276] Optionally, as shown in Figure 11 As shown, the embodiment of the present disclosure provides a device further comprising a processing unit 1104. The processing unit 1104 is used for:

[0277] The search condition feature of each candidate multimedia resource is converted in feature dimension to obtain a converted search condition feature of each candidate multimedia resource. The converted search condition feature of each candidate multimedia resource has the same dimension as the resource feature of each candidate multimedia resource.

[0278] The converted search condition feature of each candidate multimedia resource and the resource feature of each candidate multimedia resource are merged to obtain a merged feature of each candidate multimedia resource.

[0279] The merged feature of each candidate multimedia resource is input into a pre-trained first weight model to obtain a weight of the linear feature of each candidate multimedia resource. The first weight model comprises a first parameter, and the first parameter is used for learning the converted search condition feature in the merged feature.

[0280] The merged feature of each candidate multimedia resource is input into a pre-trained second weight model to obtain a weight of the nonlinear feature of each candidate multimedia resource. The second weight model comprises a second parameter, and the second parameter is used for learning the resource feature in the merged feature.

[0281] Optionally, as shown in Figure 11 As shown, the embodiment of the present disclosure provides that the predicted interaction parameter of each candidate multimedia candidate resource is predicted based on a pre-trained prediction model. The device further comprises a training unit 1105.

[0282] The acquisition unit 1101 is further used for acquiring an account reconstruction feature of a sample account, a resource reconstruction feature of a sample multimedia resource, and a sample interaction parameter of the sample multimedia resource. The sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period.

[0283] The training unit 1105 is configured to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and taking the sample interaction parameters of the sample multimedia resource as labels, to obtain a prediction model.

[0284] Figure 12 FIG. 12 is a structural schematic diagram of a training device of a prediction model according to an example embodiment. Referring to FIG. 12, the training device 120 of the prediction model provided by the example embodiment can be used in the electronic device described above, and is specifically configured to perform the training method of the prediction model provided by the example embodiment. The training device 120 of the prediction model includes an obtaining unit 1201 and a training unit 1202. Figure 12

[0285] The obtaining unit 1201 is configured to obtain the account reconstruction features of the sample account, the resource reconstruction features of the sample multimedia resource, and the sample interaction parameters of the sample multimedia resource. The account reconstruction features are used to represent the relevant factors and irrelevant factors of the search behavior of the account affecting the interaction behavior of the account. The sample multimedia resource is a multimedia resource on which the sample account has performed an interaction behavior in a sample time period. The resource reconstruction features are used to represent the relevant factors and irrelevant factors of the multimedia resource in the search result affecting the interaction behavior of the account under the search condition.

[0286] The training unit 1202 is configured to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and taking the sample interaction parameters of the sample multimedia resource as labels, to obtain a prediction model. The prediction model is used to predict the predicted interaction parameters of the sample account on the candidate multimedia resource.

[0287] As to the device in the above example embodiments, the specific manners in which various modules perform operations have been described in detail in the example embodiments related to the method, and will not be described in detail here.

[0288] Figure 13 FIG. 13 is a structural schematic diagram of an electronic device provided by the example embodiment. As shown in FIG. 13, the electronic device 130 can include at least one processor 1301 and a memory 1303 configured to store processor-executable instructions. The processor 1301 is configured to execute the instructions in the memory 1303 to implement the determination method of the multimedia resource in the example embodiments. Figure 13

[0289] In addition, the electronic device 130 can further include a communication bus 1302 and at least one communication interface 1304.

[0290] ​​The processor 1301 can be a central processing unit (CPU), a micro-processing unit, an ASIC, or one or more integrated circuits used to control program execution of the present disclosure.

[0291] The communication bus 1302 can include a path for transmitting information between the above components.

[0292] The communication interface 1304, using any transceiver-like device, is used to communicate with other devices or communication networks, such as an Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc.

[0293] The memory 1303 can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but not limited to. The memory can exist independently, connected to the processing unit through a bus. The memory can also be integrated with the processing unit.

[0294] The memory 1303 is configured to store instructions for executing the present disclosure, and the processor 1301 is configured to control the execution of the instructions. The processor 1301 is configured to execute the instructions stored in the memory 1303, thereby realizing the functions in the method of the present disclosure.

[0295] As an example, in combination with Figure 11 , the functions realized by the acquisition unit 1101, the prediction unit 1102, the determination unit 1103, the processing unit 1104, and the training unit 1105 in the electronic device are the same as the functions of the processor 1301 in Figure 13 .

[0296] In a specific implementation, as an embodiment, the processor 1301 can include one or more CPUs, such as Figure 13CPU0 and CPU1 in FIG. 1.

[0297] In a specific implementation, as an example, the electronic device 130 can include multiple processors, such as the processor 1301 and the processor 1307 in FIG. 1. Each of the processors can be a single-CPU processor or a multi-CPU processor. The processor herein can refer to one or more devices, circuits, and / or processing cores for processing data (for example, computer program instructions). Figure 13

[0298] In a specific implementation, as an example, the electronic device 130 can further include an output device 1305 and an input device 1306. The output device 1305 communicates with the processor 1301 and can display information in various ways. For example, the output device 1305 can be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector, etc. The input device 1306 communicates with the processor 1301 and can accept input of an account in various ways. For example, the input device 1306 can be a mouse, a keyboard, a touch screen device, or a sensor device, etc.

[0299] Those skilled in the art can understand that the structure shown in FIG. 1 does not constitute a limitation on the electronic device 130, and can include more or fewer components than shown, or combine certain components, or adopt a different arrangement of components. Figure 13

[0300] Meanwhile, another hardware structure of the electronic device provided by the embodiments of the present disclosure can also refer to the description of the electronic device in the above Figure 13 embodiments, which will not be described here. The difference is that the processor included in the electronic device is used to execute the steps in the training method of the prediction model performed by the electronic device in the above embodiments.

[0301] In addition, the present disclosure also provides a computer readable storage medium, when the instructions in the computer readable storage medium are executed by the processor of the electronic device, the electronic device can execute the determination method of the multimedia resource and the training method of the prediction model provided by the above embodiments.

[0302] In addition, the present disclosure also provides a computer program product, including computer instructions, when the computer instructions run on the electronic device, the electronic device executes the determination method of the multimedia resource and the training method of the prediction model provided by the above embodiments. ​​

[0303] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. It is intended that the disclosure be construed as including any variations, uses, or adaptations of the specific embodiments following, including equivalents thereof, which are within the scope of the disclosure and including such as come within the general scope of the following claims. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the disclosure being indicated by the following claims.

Claims

1. A method for determining multimedia resources, characterized in that, include: Obtain the account reconstruction features of the current account, and the resource reconstruction features of each candidate multimedia resource among multiple candidate multimedia resources; The account reconstruction feature is used to characterize the relevant and irrelevant factors that influence an account's search behavior on its interactive behavior; the resource reconstruction feature is used to characterize the relevant and irrelevant factors that influence an account's interactive behavior based on the multimedia resources in the search results under the search conditions. Based on the account reconstruction features of the current account and the resource reconstruction features of each candidate multimedia resource, predict the predicted interaction parameters of the current account for each candidate multimedia resource; Based on the predicted interaction parameters of each candidate multimedia resource, determine the multimedia resources to be recommended to the current account from the plurality of candidate multimedia resources; This includes obtaining the resource reconstruction features of each candidate multimedia resource from multiple candidate multimedia resources, including: For each candidate multimedia resource, a linear characteristic and a nonlinear characteristic of the candidate multimedia resource are determined; the linear characteristic of the candidate multimedia resource is used to characterize the relevant factors that affect the account's interactive behavior under search conditions, and the nonlinear characteristic of the candidate multimedia resource is used to characterize the irrelevant factors that affect the account's interactive behavior under search conditions. The resource reconstruction characteristics of the candidate multimedia resources are determined based on their linear and nonlinear characteristics.

2. The method for determining multimedia resources according to claim 1, characterized in that, The process of obtaining the account reconstruction features of the current account includes: Obtain the account characteristics of the current account, the resource characteristics of the historical multimedia resources in which the current account has performed interactive behaviors within a historical time period, and the search condition characteristics of each historical multimedia resource; Based on the search condition features of each historical multimedia resource, feature reconstruction processing is performed on the resource features of each historical multimedia resource to obtain the resource reconstruction features of each historical multimedia resource; The account reconstruction features of the current account are obtained based on the resource reconstruction features of each historical multimedia resource and the account features of the current account.

3. The method for determining multimedia resources according to claim 2, characterized in that, The step of performing feature reconstruction processing on the resource features of each historical multimedia resource based on the search condition features of each historical multimedia resource to obtain the resource reconstruction features of each historical multimedia resource includes: Based on the search condition characteristics and resource characteristics of each historical multimedia resource, the linear characteristics and nonlinear characteristics of each historical multimedia resource are determined respectively. The linear characteristics of each historical multimedia resource are used to characterize the relevant factors that affect the account's interactive behavior under search conditions, and the nonlinear characteristics of each historical multimedia resource are used to characterize the irrelevant factors that affect the account's interactive behavior under search conditions. By weighting the linear and nonlinear characteristics of each historical multimedia resource, the resource reconstruction characteristics of each historical multimedia resource are obtained.

4. The method for determining multimedia resources according to claim 3, characterized in that, The step of determining the linear characteristics and nonlinear characteristics of each historical multimedia resource based on the search condition characteristics and resource characteristics of each historical multimedia resource includes: Based on the search condition features of each historical multimedia resource, the resource features of each historical multimedia resource, and the preset regression operation, the fitting vector and the residual vector of each historical multimedia resource are obtained respectively. The fitted vector of each historical multimedia resource is determined as the linear feature of each historical multimedia resource, and the residual vector of each historical multimedia resource is determined as the nonlinear feature of each historical multimedia resource.

5. The method for determining multimedia resources according to claim 3, characterized in that, The method further includes: The search condition features of each historical multimedia resource are transformed by feature dimension to obtain the transformed search condition features of each historical multimedia resource; the transformed search condition features of each historical multimedia resource have the same dimension as the resource features of each historical multimedia resource. The search condition features after conversion of each historical multimedia resource and the resource features of each historical multimedia resource are combined to obtain the combined features of each historical multimedia resource; The merged features of each historical multimedia resource are input into a pre-trained first weight model to obtain the weights of the linear features of each historical multimedia resource; the first weight model includes a first parameter, which is used to learn the transformed search condition features in the merged features; The merged features of each historical multimedia resource are input into a pre-trained second weight model to obtain the weights of the nonlinear features of each historical multimedia resource; the second weight model includes a second parameter, which is used to learn the resource features in the merged features.

6. The method for determining multimedia resources according to claim 1, characterized in that, Determining the linear characteristics and nonlinear characteristics of the candidate multimedia resources includes: Obtain the resource characteristics of each candidate multimedia resource, and the search condition characteristics of each candidate multimedia resource; Based on the search condition features and resource features of each candidate multimedia resource, the linear features and nonlinear features of each candidate multimedia resource are determined respectively. The linear features of each candidate multimedia resource are used to characterize the relevant factors that affect the account's interactive behavior under search conditions, and the nonlinear features of each candidate multimedia resource are used to characterize the irrelevant factors that affect the account's interactive behavior under search conditions. The step of determining the resource reconstruction characteristics of the candidate multimedia resources based on their linear and nonlinear characteristics includes: The linear and nonlinear features of each candidate multimedia resource are weighted to obtain the resource reconstruction features of each candidate video.

7. The method for determining multimedia resources according to claim 6, characterized in that, The step of determining the linear characteristics and nonlinear characteristics of each candidate multimedia resource based on the search condition characteristics and resource characteristics of each candidate multimedia resource includes: Based on the search condition features of each candidate multimedia resource, the resource features of each candidate multimedia resource, and the preset regression operation, the fitting vector and the residual vector of each candidate multimedia resource are obtained respectively. The fitted vector of each candidate multimedia resource is determined as the linear feature of each candidate multimedia resource, and the residual vector of each candidate multimedia resource is determined as the nonlinear feature of each candidate multimedia resource.

8. The method for determining multimedia resources according to claim 6, characterized in that, The method further includes: The search condition features of each candidate multimedia resource are transformed in terms of feature dimension to obtain the transformed search condition features of each candidate multimedia resource; the transformed search condition features of each candidate multimedia resource have the same dimension as the resource features of each candidate multimedia resource. The search condition features after conversion of each candidate multimedia resource and the resource features of each candidate multimedia resource are combined to obtain the combined features of each candidate multimedia resource; The merged features of each candidate multimedia resource are input into a pre-trained first weight model to obtain the weights of the linear features of each candidate multimedia resource; the first weight model includes a first parameter, which is used to learn the transformed search condition features in the merged features; The merged features of each candidate multimedia resource are input into a pre-trained second weight model to obtain the weights of the nonlinear features of each candidate multimedia resource; the second weight model includes a second parameter, which is used to learn the resource features in the merged features.

9. The method for determining multimedia resources according to any one of claims 1-8, characterized in that, The predicted interaction parameters for each candidate multimedia resource are obtained based on a pre-trained prediction model, and the method further includes: The account reconstruction features of the sample account, the resource reconstruction features of the sample multimedia resources, and the sample interaction parameters of the sample multimedia resources are obtained; the sample multimedia resources are the multimedia resources that the sample account has interacted with during the sample time period. The account reconstruction features of the sample accounts and the resource reconstruction features of the sample multimedia resources are used as sample features, and the sample interaction parameters of the sample multimedia resources are used as labels. A preset neural network model is trained to obtain the prediction model.

10. A method for training a prediction model, characterized in that, include: The system obtains the account reconstruction features of sample accounts, the resource reconstruction features of sample multimedia resources, and the sample interaction parameters of the sample multimedia resources; the account reconstruction features are used to characterize the relevant and irrelevant factors that influence the account's search behavior on the account's interaction behavior. The sample multimedia resources are the multimedia resources in which the sample account has performed interactive behaviors within the sample time period; the resource reconstruction features are used to characterize the relevant and irrelevant factors that influence the account's interactive behaviors based on the multimedia resources in the corresponding search results under search conditions. The account reconstruction features of the sample accounts and the resource reconstruction features of the sample multimedia resources are used as sample features, and the sample interaction parameters of the sample multimedia resources are used as labels. A preset neural network model is trained to obtain the prediction model; the prediction model is used to predict the predicted interaction parameters of the sample accounts to the candidate multimedia resources. Among them, the resource reconstruction features of the sample multimedia resources are obtained, including: The linear and nonlinear characteristics of the sample multimedia resources are determined; the linear characteristics of the sample multimedia resources are used to characterize the relevant factors that influence the account's interactive behavior under search conditions, and the nonlinear characteristics of the sample multimedia resources are used to characterize the irrelevant factors that influence the account's interactive behavior under search conditions. The resource reconstruction characteristics of the sample multimedia resources are determined based on the linear and nonlinear characteristics of the sample multimedia resources.

11. A device for determining multimedia resources, characterized in that, It includes an acquisition unit, a prediction unit, and a determination unit; The acquisition unit is used to acquire the account reconstruction features of the current account and the resource reconstruction features of each candidate multimedia resource among multiple candidate multimedia resources. The account reconstruction feature is used to characterize the relevant and irrelevant factors that influence an account's search behavior on its interactive behavior; the resource reconstruction feature is used to characterize the relevant and irrelevant factors that influence an account's interactive behavior based on the multimedia resources in the search results under the search conditions. The prediction unit is configured to predict the predicted interaction parameters of the current account for each candidate multimedia resource based on the account reconstruction features of the current account and the resource reconstruction features of each candidate multimedia resource. The determining unit is configured to determine, from the plurality of candidate multimedia resources, a multimedia resource to be recommended to the current account based on the predicted interaction parameters of each candidate multimedia resource. The acquisition unit is specifically used for: For each candidate multimedia resource, a linear characteristic and a nonlinear characteristic of the candidate multimedia resource are determined; the linear characteristic of the candidate multimedia resource is used to characterize the relevant factors that affect the account's interactive behavior under search conditions, and the nonlinear characteristic of the candidate multimedia resource is used to characterize the irrelevant factors that affect the account's interactive behavior under search conditions. The resource reconstruction characteristics of the candidate multimedia resources are determined based on their linear and nonlinear characteristics.

12. The multimedia resource determination device according to claim 11, characterized in that, The acquisition unit is specifically used for: Obtain the account characteristics of the current account, the resource characteristics of the historical multimedia resources in which the current account has performed interactive behaviors within a historical time period, and the search condition characteristics of each historical multimedia resource; Based on the search condition features of each historical multimedia resource, feature reconstruction processing is performed on the resource features of each historical multimedia resource to obtain the resource reconstruction features of each historical multimedia resource; The account reconstruction features of the current account are obtained based on the resource reconstruction features of each historical multimedia resource and the account features of the current account.

13. The multimedia resource determination device according to claim 12, characterized in that, The acquisition unit is specifically used for: Based on the search condition characteristics and resource characteristics of each historical multimedia resource, the linear characteristics and nonlinear characteristics of each historical multimedia resource are determined respectively. The linear characteristics of each historical multimedia resource are used to characterize the relevant factors that influence the account's interactive behavior under search conditions, and the nonlinear characteristics of each historical multimedia resource are used to characterize the irrelevant factors that influence the account's interactive behavior under search conditions. By weighting the linear and nonlinear characteristics of each historical multimedia resource, the resource reconstruction characteristics of each historical multimedia resource are obtained.

14. The multimedia resource determination device according to claim 13, characterized in that, The acquisition unit is specifically used for: Based on the search condition features of each historical multimedia resource, the resource features of each historical multimedia resource, and the preset regression operation, the fitting vector and the residual vector of each historical multimedia resource are obtained respectively. The fitted vector of each historical multimedia resource is determined as the linear feature of each historical multimedia resource, and the residual vector of each historical multimedia resource is determined as the nonlinear feature of each historical multimedia resource.

15. The multimedia resource determination device according to claim 13, characterized in that, The device also includes a processing unit; The processing unit is used to transform the feature dimension of the search condition features of each historical multimedia resource to obtain the transformed search condition features of each historical multimedia resource; the transformed search condition features of each historical multimedia resource have the same dimension as the resource features of each historical multimedia resource. The processing unit is further configured to merge the search condition features after conversion of each historical multimedia resource and the resource features of each historical multimedia resource to obtain the merged features of each historical multimedia resource. The processing unit is further configured to input the merged features of each historical multimedia resource into a pre-trained first weight model to obtain the weights of the linear features of each historical multimedia resource; the first weight model includes a first parameter, which is used to learn the transformed search condition features in the merged features; The processing unit is further configured to input the merged features of each historical multimedia resource into a pre-trained second weight model to obtain the weights of the nonlinear features of each historical multimedia resource; the second weight model includes a second parameter, which is used to learn the resource features in the merged features.

16. The multimedia resource determination device according to claim 11, characterized in that, The acquisition unit is specifically used for: Obtain the resource characteristics of each candidate multimedia resource, and the search condition characteristics of each candidate multimedia resource; Based on the search condition features and resource features of each candidate multimedia resource, the linear features and nonlinear features of each candidate multimedia resource are determined respectively. The linear characteristics of each candidate multimedia resource are used to characterize the relevant factors that influence the account's interactive behavior under search conditions, and the nonlinear characteristics of each candidate multimedia resource are used to characterize the irrelevant factors that influence the account's interactive behavior under search conditions. The linear and nonlinear features of each candidate multimedia resource are weighted to obtain the resource reconstruction features of each candidate video.

17. The multimedia resource determination device according to claim 16, characterized in that, The acquisition unit is specifically used for: Based on the search condition features of each candidate multimedia resource, the resource features of each candidate multimedia resource, and the preset regression operation, the fitting vector and the residual vector of each candidate multimedia resource are obtained respectively. The fitted vector of each candidate multimedia resource is determined as the linear feature of each candidate multimedia resource, and the residual vector of each candidate multimedia resource is determined as the nonlinear feature of each candidate multimedia resource.

18. The multimedia resource determination device according to claim 16, characterized in that, The device further includes a processing unit; the processing unit is used for: The search condition features of each candidate multimedia resource are transformed in terms of feature dimension to obtain the transformed search condition features of each candidate multimedia resource; the transformed search condition features of each candidate multimedia resource have the same dimension as the resource features of each candidate multimedia resource. The search condition features after conversion of each candidate multimedia resource and the resource features of each candidate multimedia resource are combined to obtain the combined features of each candidate multimedia resource; The merged features of each candidate multimedia resource are input into the pre-trained first weight model to obtain the weights of the linear features of each candidate multimedia resource. The first weight model includes a first parameter, which is used to learn the transformed search condition features in the merged features; The merged features of each candidate multimedia resource are input into a pre-trained second weight model to obtain the weights of the nonlinear features of each candidate multimedia resource. The second weighting model includes a second parameter, which is used to learn resource features in the merged features.

19. The apparatus for determining multimedia resources according to any one of claims 11-18, characterized in that, The predicted interaction parameters of each candidate multimedia resource are obtained based on a pre-trained prediction model, and the device also includes a training unit. The acquisition unit is further configured to acquire the account reconstruction features of the sample account, the resource reconstruction features of the sample multimedia resources, and the sample interaction parameters of the sample multimedia resources; the sample multimedia resources are multimedia resources that the sample account has performed interactive behaviors in the sample time period. The training unit is used to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resources as sample features and the sample interaction parameters of the sample multimedia resources as labels, so as to obtain the prediction model.

20. A training device for a prediction model, characterized in that, Includes acquisition units and training units; The acquisition unit is used to acquire the account reconstruction features of the sample account, the resource reconstruction features of the sample multimedia resources, and the sample interaction parameters of the sample multimedia resources; the account reconstruction features are used to characterize the relevant and irrelevant factors that influence the account's search behavior on the account's interaction behavior. The sample multimedia resources are the multimedia resources in which the sample account has performed interactive behaviors within the sample time period; the resource reconstruction features are used to characterize the relevant and irrelevant factors that influence the account's interactive behaviors based on the multimedia resources in the corresponding search results under search conditions. The training unit is used to train a preset neural network model by taking the account reconstruction features of the sample account and the resource reconstruction features of the sample multimedia resource as sample features and the sample interaction parameters of the sample multimedia resource as labels, to obtain the prediction model; the prediction model is used to predict the predicted interaction parameters of the sample account to the candidate multimedia resource. The acquisition unit is specifically used for: Determine the linear characteristics and nonlinear characteristics of the sample multimedia resources; The linear characteristics of the sample multimedia resources are used to characterize the relevant factors that influence the account's interactive behavior under search conditions, and the nonlinear characteristics of the sample multimedia resources are used to characterize the irrelevant factors that influence the account's interactive behavior under search conditions. The resource reconstruction characteristics of the sample multimedia resources are determined based on the linear and nonlinear characteristics of the sample multimedia resources.

21. An electronic device, characterized in that, include: A processor and a memory for storing instructions executable by the processor; wherein the processor is configured to execute instructions to implement the method for determining multimedia resources according to any one of claims 1-9, or the method for training a prediction model according to claim 10.

22. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the method for determining multimedia resources as described in any one of claims 1-9, or the method for training the prediction model as described in claim 10.

23. A computer program product, comprising instructions, characterized in that, The computer program product includes computer instructions that, when executed on an electronic device, cause the electronic device to perform the method for determining multimedia resources as described in any one of claims 1-9, or the method for training the prediction model as described in claim 10.

Images