Resource operation data prediction method, prediction model training method and device

Abstract

The present disclosure relates to a resource operation data prediction method, a prediction model training method, and an apparatus. The resource operation data prediction method includes: obtaining object feature information of a target object and resource feature information of a target multimedia resource; extracting the object feature information and the resource feature information based on a hidden layer of a resource operation prediction model to obtain first feature information; performing multi-scale information extraction on first input feature information based on a first multi-scale information extraction module of the resource operation prediction model to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information; and determining resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information. The present disclosure can improve the accuracy of resource operation prediction results.

Description

Technical Field

[0001] The present disclosure relates to the field of machine learning technology, and in particular to a resource operation data prediction method, a prediction model training method, and a device. Background Art

[0002] With the advancement of technology and the vigorous development of the multimedia industry, more and more people have become loyal users of multimedia resources. For massive multimedia resources, personalized distribution becomes increasingly important.

[0003] A widely used method in related technologies is to use machine learning to predict user operations on multimedia resources. This involves using a multi-layer perceptron for forward feature extraction, combined with regression or classification algorithms for target prediction. Ultimately, this resource operation prediction information is combined with strategies to rank multimedia resources, and the multimedia resources that are pushed to the user are determined based on the ranking results. However, these methods for predicting user resource operations in related technologies suffer from inaccurate predictions. Summary of the Invention

[0004] The present disclosure provides a resource operation data prediction method, prediction model training method, and device to at least address the problem of inaccurate user resource operation prediction in related technologies. The technical solutions of the present disclosure are as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a resource operation data prediction method is provided, comprising:

[0006] Acquiring object feature information of a target object and resource feature information of a target multimedia resource;

[0007] Extracting the object feature information and the resource feature information based on a hidden layer of a resource operation prediction model to obtain first feature information;

[0008] performing multi-scale information extraction on the first input feature information by a first multi-scale information extraction module based on the resource operation prediction model to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information;

[0009] Resource operation data of the target object on the target multimedia resource is determined based on the first multi-scale feature information and the first feature information.

[0010] In an exemplary embodiment, the resource operation prediction model further includes a multi-scale gating weight prediction module;

[0011] The method further comprises:

[0012] Performing weight prediction on feature information of each scale in the first multi-scale feature information based on the multi-scale gating weight prediction module to obtain scale gating weights corresponding to the feature information of each scale;

[0013] The determining, based on the first multi-scale feature information and the first feature information, resource operation data of the target object on the target multimedia resource includes:

[0014] Obtaining multi-scale weighted information based on the feature information of each scale and the scale gating weights corresponding to the feature information of each scale;

[0015] The resource operation data is determined based on the multi-scale weighted information and the first feature information.

[0016] In an exemplary embodiment, the resource operation prediction model further includes a second multi-scale information extraction module;

[0017] The method further comprises:

[0018] Performing multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information;

[0019] The first input feature information is determined based on the first feature information and the second multi-scale feature information.

[0020] In an exemplary embodiment, determining the resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information includes:

[0021] Calculating average scale feature information of feature information of each scale in the first multi-scale feature information;

[0022] performing information fusion on the average scale feature information and the first feature information to obtain target fused feature information;

[0023] The resource operation data is determined based on the target fusion feature information.

[0024] In an exemplary embodiment, the resource operation prediction model includes a plurality of hidden layers;

[0025] The determining the resource operation data based on the target fusion feature information includes:

[0026] Inputting the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extracting information from the target fusion feature information based on the next hidden layer to obtain second feature information;

[0027] Based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction is performed on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information;

[0028] Determining target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information;

[0029] Determining the next hidden layer as the current hidden layer;

[0030] Until there is no next hidden layer after the current hidden layer;

[0031] The resource operation data is determined based on the target fusion feature information corresponding to the current hidden layer.

[0032] In an exemplary embodiment, the resource operation data includes resource operation data for a plurality of candidate resources;

[0033] The method further comprises:

[0034] Determining, based on the resource operation data on the plurality of candidate resources, ranking information of each candidate resource in the push list;

[0035] Determining a target resource from the push list based on the sorting information;

[0036] Push the target resource to the target object.

[0037] According to a second aspect of an embodiment of the present disclosure, a resource operation prediction model training method is provided, comprising:

[0038] Acquire sample data; the sample data includes sample object feature information of the sample object, sample resource feature information of the sample multimedia resource, and sample resource operation tags;

[0039] Extracting the sample object feature information and the sample resource feature information based on the hidden layer of the resource operation prediction model to be trained to obtain first training feature information;

[0040] performing multi-scale information extraction on the first input training feature information based on the first multi-scale information extraction module of the resource operation prediction model to be trained to obtain first multi-scale training feature information; the first input training feature information is obtained by performing feature dimension conversion on the first training feature information;

[0041] Based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain a resource operation prediction model.

[0042] In an exemplary embodiment, the resource operation prediction model to be trained further includes a multi-scale gating weight prediction module;

[0043] The method further comprises:

[0044] Performing weight prediction on feature information of each scale in the first multi-scale training feature information based on the multi-scale gating weight prediction module to obtain scale gating training weights corresponding to the training feature information of each scale;

[0045] The training of the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain the resource operation prediction model includes:

[0046] Obtaining multi-scale weighted training information based on the training feature information of each scale and the scale gated training weights corresponding to the training feature information of each scale;

[0047] Based on the multi-scale weighted training information, the first training feature information, and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0048] In an exemplary embodiment, the resource operation prediction model to be trained further includes a second multi-scale information extraction module;

[0049] The method further comprises:

[0050] Performing multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information;

[0051] The first training feature information and the second multi-scale training feature information are fused to obtain the first input training feature information.

[0052] In an exemplary embodiment, the training of the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain the resource operation prediction model includes:

[0053] Calculating average scale training feature information of the training feature information of each scale in the first multi-scale training feature information;

[0054] performing information fusion on the average scale training feature information and the first training feature information to obtain target fused training feature information;

[0055] Based on the target fusion training feature information and the sample resource operation labels, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0056] In an exemplary embodiment, the resource operation prediction model to be trained includes a plurality of hidden layers;

[0057] The step of training the resource operation prediction model to be trained based on the target fusion training feature information and the sample resource operation label to obtain the resource operation prediction model includes:

[0058] Inputting the target fusion training feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extracting information from the target fusion training feature information based on the next hidden layer to obtain second training feature information;

[0059] performing multi-scale information extraction on the second input feature information based on the first multi-scale information extraction module corresponding to the next hidden layer to obtain third multi-scale training feature information; the second input training feature information is determined based on feature dimension conversion performed on the second training feature information;

[0060] Determining target fusion training feature information corresponding to the next hidden layer based on the third multi-scale training feature information and the second training feature information;

[0061] Determining the next hidden layer as the current hidden layer;

[0062] Until there is no next hidden layer after the current hidden layer;

[0063] Based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0064] In an exemplary embodiment, the training of the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain the resource operation prediction model includes:

[0065] Determining loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label;

[0066] Adjusting the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module;

[0067] The resource operation prediction model is obtained based on the adjusted hidden layer and the adjusted first multi-scale information extraction module.

[0068] According to a third aspect of an embodiment of the present disclosure, there is provided a resource operation data prediction device, comprising:

[0069] A feature information acquisition unit configured to acquire object feature information of a target object and resource feature information of a target multimedia resource;

[0070] A first feature information extraction unit is configured to extract the object feature information and the resource feature information based on a hidden layer of a resource operation prediction model to obtain first feature information;

[0071] a first multi-scale feature information extraction unit configured to execute a first multi-scale information extraction module based on the resource operation prediction model to perform multi-scale information extraction on the first input feature information to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information;

[0072] The resource operation data determining unit is configured to determine the resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information.

[0073] In an exemplary embodiment, the resource operation prediction model further includes a multi-scale gating weight prediction module;

[0074] The device further comprises:

[0075] a scale gating weight determining unit, configured to perform weight prediction on feature information of each scale in the first multi-scale feature information based on the multi-scale gating weight prediction module, to obtain a scale gating weight corresponding to the feature information of each scale;

[0076] The resource operation data determining unit includes:

[0077] a multi-scale weighted information unit configured to perform feature information based on each scale and scale gating weights corresponding to the feature information at each scale to obtain multi-scale weighted information;

[0078] The first determining unit is configured to determine the resource operation data based on the multi-scale weighted information and the first feature information.

[0079] In an exemplary embodiment, the resource operation prediction model further includes a second multi-scale information extraction module;

[0080] The device further comprises:

[0081] a second multi-scale feature information determining unit, configured to perform multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information;

[0082] The first input feature information determining unit is configured to determine the first input feature information based on the first feature information and the second multi-scale feature information.

[0083] In an exemplary embodiment, the resource operation data determining unit includes:

[0084] an average scale feature information determining unit, configured to calculate average scale feature information of feature information of each scale in the first multi-scale feature information;

[0085] a fused feature information determining unit, configured to perform information fusion on the average scale feature information and the first feature information to obtain target fused feature information;

[0086] The second determining unit is configured to determine the resource operation data based on the target fusion feature information.

[0087] In an exemplary embodiment, the resource operation prediction model includes a plurality of hidden layers;

[0088] The second determining unit includes:

[0089] The second feature information determining unit is configured to input the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extract information from the target fusion feature information based on the next hidden layer to obtain second feature information;

[0090] a third multi-scale feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information;

[0091] a third determining unit configured to determine target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information;

[0092] A first hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer;

[0093] A first hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer;

[0094] The fourth determining unit is configured to determine the resource operation data based on the target fusion feature information corresponding to the current hidden layer.

[0095] In an exemplary embodiment, the resource operation data includes resource operation data for a plurality of candidate resources;

[0096] The device further comprises:

[0097] a ranking information determining unit, configured to determine ranking information of each candidate resource in the push list based on the resource operation data on the plurality of candidate resources;

[0098] a target resource determining unit, configured to determine a target resource from the push list based on the sorting information;

[0099] The target resource pushing unit is configured to push the target resource to the target object.

[0100] According to a fourth aspect of an embodiment of the present disclosure, a resource operation prediction model training device is provided, comprising:

[0101] A sample data acquisition unit is configured to acquire sample data; the sample data includes sample object feature information of a sample object, sample resource feature information of a sample multimedia resource, and a sample resource operation tag;

[0102] A first training feature information determining unit is configured to extract the sample object feature information and the sample resource feature information based on a hidden layer of the resource operation prediction model to be trained to obtain first training feature information;

[0103] a first multi-scale training feature information extraction unit configured to perform multi-scale information extraction on first input training feature information based on a first multi-scale information extraction module based on the resource operation prediction model to be trained, to obtain first multi-scale training feature information; the first input training feature information is obtained by performing feature dimension conversion on the first training feature information;

[0104] The resource operation prediction model determination unit is configured to train the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain a resource operation prediction model.

[0105] In an exemplary embodiment, the resource operation prediction model to be trained further includes a multi-scale gating weight prediction module;

[0106] The device further comprises:

[0107] a scale-gated training weight determination unit, configured to perform weight prediction on feature information of each scale in the first multi-scale training feature information based on the multi-scale gating weight prediction module, to obtain scale-gated training weights corresponding to the training feature information of each scale;

[0108] The resource operation prediction model determination unit includes:

[0109] a multi-scale weighted training information determining unit configured to execute a scale-gated training weight based on the training feature information of each scale and the training feature information of each scale, to obtain multi-scale weighted training information;

[0110] The first training unit is configured to train the resource operation prediction model to be trained based on the multi-scale weighted training information, the first training feature information and the sample resource operation label to obtain the resource operation prediction model.

[0111] In an exemplary embodiment, the resource operation prediction model to be trained further includes a second multi-scale information extraction module;

[0112] The device further comprises:

[0113] a second multi-scale training feature information determining unit, configured to perform multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information;

[0114] The first input training feature information determining unit is configured to perform information fusion on the first training feature information and the second multi-scale training feature information to obtain the first input training feature information.

[0115] In an exemplary embodiment, the resource operation prediction model determination unit includes:

[0116] an average scale training feature information determining unit, configured to calculate average scale training feature information of the training feature information of each scale in the first multi-scale training feature information;

[0117] a fusion training feature information determining unit, configured to perform information fusion on the average scale training feature information and the first training feature information to obtain target fusion training feature information;

[0118] The second training unit is configured to train the resource operation prediction model to be trained based on the target fusion training feature information and the sample resource operation label to obtain the resource operation prediction model.

[0119] In an exemplary embodiment, the resource operation prediction model to be trained includes a plurality of hidden layers;

[0120] The second training unit includes:

[0121] a second training feature information determining unit configured to input the target fusion training feature information corresponding to the current hidden layer into a hidden layer next to the current hidden layer, and extract information from the target fusion training feature information based on the next hidden layer to obtain second training feature information;

[0122] a third multi-scale training feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale training feature information; the second input training feature information is determined based on feature dimension conversion performed on the second training feature information;

[0123] A second hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer;

[0124] A second hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer;

[0125] The fifth determining unit is configured to perform training on the resource operation prediction model to be trained based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label to obtain the resource operation prediction model.

[0126] In an exemplary embodiment, the resource operation prediction model determination unit includes:

[0127] a loss information determining unit, configured to determine loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label;

[0128] a parameter adjustment unit configured to adjust the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module;

[0129] The fourth determining unit is configured to execute the resource operation prediction model based on the adjusted hidden layer and the adjusted first multi-scale information extraction module.

[0130] According to the fifth aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the resource operation data prediction method or the resource operation prediction model training method as described above.

[0131] According to the sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When the instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the resource operation data prediction method or the resource operation prediction model training method as described above.

[0132] According to the seventh aspect of an embodiment of the present disclosure, a computer program product is provided, which includes a computer program stored in a readable storage medium, and at least one processor of a computer device reads and executes the computer program from the readable storage medium, so that the device executes the above-mentioned resource operation data prediction method or resource operation prediction model training method.

[0133] The technical solutions provided by the embodiments of the present disclosure bring at least the following beneficial effects:

[0134] The present disclosure extracts object feature information of a target object and resource feature information of a target multimedia resource through a hidden layer of a resource operation prediction model to obtain first feature information; performs multi-scale information extraction on the first input feature information based on a first multi-scale information extraction module to obtain first multi-scale feature information; and determines resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information. In the present disclosure, the first multi-scale information extraction module is used to perform multi-scale information extraction on the first input feature information, thereby obtaining feature information of multiple sampling scales corresponding to the first input feature information. The extracted feature information of multiple sampling scales can improve the feature representation capability of the target object and the target multimedia resource, and accordingly improve the feature extraction capability and information fitting capability of the resource operation prediction model, further improving the prediction accuracy of the resource operation prediction model.

[0135] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0136] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0137] Figure 1It is a schematic diagram of an implementation environment according to an exemplary embodiment.

[0138] Figure 2 The figure is a flow chart of a resource operation data prediction method according to an exemplary embodiment.

[0139] Figure 3 The present invention is a flowchart of a method for determining resource operation data based on multi-scale gating weights according to an exemplary embodiment.

[0140] Figure 4 The figure is a flow chart of a method for extracting multi-scale information from first feature information according to an exemplary embodiment.

[0141] Figure 5 The present invention is a flowchart of a method for determining resource operation data based on multi-scale feature information according to an exemplary embodiment.

[0142] Figure 6 The figure is a flow chart of another method for determining resource operation data according to an exemplary embodiment.

[0143] Figure 7 FIG. 4 is a schematic structural diagram of a first multi-scale information extraction module according to an exemplary embodiment.

[0144] Figure 8 FIG. 4 is a schematic structural diagram of a multi-scale gating weight prediction module according to an exemplary embodiment.

[0145] Figure 9 FIG. 4 is a schematic structural diagram of a second multi-scale information extraction module according to an exemplary embodiment.

[0146] Figure 10 The present invention is a flowchart of a resource operation prediction model training method according to an exemplary embodiment.

[0147] Figure 11 The present invention is a flowchart of a method for model training based on a multi-scale gating weight prediction module according to an exemplary embodiment.

[0148] Figure 12 This is a flowchart of a method for extracting information based on a second multi-scale information extraction module according to an exemplary embodiment.

[0149] Figure 13 The figure is a flow chart of another resource operation prediction model training method according to an exemplary embodiment.

[0150] Figure 14 The figure is a flow chart of another resource operation prediction model training method according to an exemplary embodiment.

[0151] Figure 15 The present invention is a flow chart showing a method for determining a resource operation prediction model based on network parameter adjustment according to an exemplary embodiment.

[0152] Figure 16 The figure is a flow chart of a resource pushing method according to an exemplary embodiment.

[0153] Figure 17 The figure is a schematic diagram of a click-through rate model and a conversion rate model with a gated network according to an exemplary embodiment.

[0154] Figure 18 The figure is a schematic diagram of a resource operation prediction device according to an exemplary embodiment.

[0155] Figure 19 The figure is a schematic diagram of a resource operation prediction model training device according to an exemplary embodiment.

[0156] Figure 20 The figure is a schematic diagram showing the structure of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0157] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0158] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0159] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0160] See also Figure 1 , which shows a schematic diagram of an implementation environment provided by an embodiment of the present disclosure, the implementation environment may include: at least one first terminal 110 and a second terminal 120, and the first terminal 110 and the second terminal 120 can communicate data through a network.

[0161] Specifically, based on the target object's object feature information, the second terminal 120 pushes multimedia resources that match the target object's object feature information to the first terminal 110 corresponding to the target object. Furthermore, the multimedia resources that match the target object's object feature information can be predicted based on a resource operation prediction model. Accordingly, the second terminal 120 can also train the resource operation prediction model to be trained based on sample data to obtain the resource operation prediction model.

[0162] The first terminal 110 can communicate with the second terminal 120 based on a browser / server (B / S) or client / server (C / S) model. The first terminal 110 may include a physical device such as a smartphone, tablet computer, laptop computer, digital assistant, smart wearable device, vehicle-mounted terminal, or server, and may also include software running on the physical device, such as an application. The operating system running on the first terminal 110 in the embodiments of the present disclosure may include, but is not limited to, Android, iOS, Linux, Windows, and the like.

[0163] The second terminal 120 can establish a communication connection with the first terminal 110 via wired or wireless communication. The second terminal 120 can include an independently operated server, a distributed server, or a server cluster consisting of multiple servers, where the server can be a cloud server.

[0164] To solve the problem of inaccurate prediction of user resource operations in related technologies, please refer to Figure 2 , which shows a resource operation data prediction method, the execution subject of this method can be Figure 1 The second terminal in; the method may specifically include:

[0165] S210. Obtain object feature information of the target object and resource feature information of the target multimedia resource.

[0166] In a specific embodiment, the target object's object feature information may include the target object's tag information, the target object's static information, and the target object's dynamic information, among others. The target object's tag information may be the target object's label classification information in the system, the target object's static information may include the target object's age, gender, place of residence, educational background, and other information, and the target object's dynamic information may include the target object's historically clicked multimedia resources and historically converted multimedia resources. The target multimedia resource's resource feature information includes the target multimedia resource's label classification information in the system, the target multimedia resource's author information, and the target multimedia resource's posterior statistical information.

[0167] S220. Extract the object feature information and the resource feature information based on the hidden layer of the resource operation prediction model to obtain first feature information.

[0168] In an optional embodiment, the resource operation prediction model may include multiple hidden layers, and one hidden layer may correspond to one multi-scale information extraction module, that is, the hidden layer and the multi-scale information extraction module may appear in pairs; the multi-scale information extraction module may also be a flexibly detachable module, so that when the multi-scale information extraction module is needed, the multi-scale information extraction module can be directly added.

[0169] In another optional embodiment, the resource operation prediction model may further include an input layer and an output layer. The input layer may be configured to receive raw feature information of a target object and raw feature information of a target multimedia resource, and process the raw feature information to obtain corresponding object feature information and resource feature information. The output layer may be configured to output specific resource operation data of the target object on the target multimedia resource.

[0170] In an optional embodiment, the object feature information and resource feature information input into the hidden layer can be obtained by feature processing based on the original feature information of the target object and the original feature information of the target multimedia resource. The object feature information and resource feature information are extracted through the hidden layer to obtain the first feature information.

[0171] S230. Perform multi-scale information extraction on the first input feature information based on the first multi-scale information extraction module of the resource operation prediction model to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information.

[0172] Specifically, the first multi-scale information extraction module can be used to extract information at multiple scales from the input information. Specifically, multiple pooling channels of different scales can be used to extract information from the first input feature information to obtain scale feature information corresponding to the pooling channel of each scale, and then obtain multi-scale feature information corresponding to the pooling channels of multiple scales.

[0173] In an optional embodiment, the first input feature information can be obtained based on feature dimension conversion of the first feature information; specifically, the first feature information can be sparse feature information, the first input feature information can be dense feature information, and the feature dimension of the first feature information can be greater than the first input feature information; thereby representing the first feature information of more dimensions by the first input feature information of fewer dimensions, which can improve the model's processing efficiency of object feature information and resource feature information.

[0174] In an optional embodiment, the feature dimension conversion operation may specifically be to perform element-wise information extraction, i.e., an inner product operation of a vector, thereby achieving feature dimension conversion of the first feature information and converting the first feature information into dense first input feature information.

[0175] S240. Determine resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information.

[0176] Both the first multi-scale feature information and the first feature information include feature information of the target object and the target multimedia resource, so that the resource operation data of the target object on the target multimedia resource can be determined based on the first multi-scale feature information and the first feature information; the resource operation of the target object on the target multimedia resource includes a click operation, a conversion operation, etc. of the target object on the target multimedia resource. Accordingly, the resource operation data can be used to characterize the possibility of the target object performing a click operation on the target multimedia resource, as well as the possibility of performing a conversion operation, etc.

[0177] In the present disclosure, a first multi-scale information extraction module is used to perform multi-scale information extraction on the first input feature information, thereby obtaining feature information of multiple sampling scales corresponding to the first input feature information. The extracted feature information of multiple sampling scales can improve the feature representation capability of the target object and the target multimedia resource, and correspondingly improve the feature extraction capability and information fitting capability of the resource operation prediction model, thereby further improving the prediction accuracy of the resource operation prediction model.

[0178] In an optional embodiment, the resource operation prediction model further includes a multi-scale gating weight prediction module; Figure 3 , which shows a method for determining resource operation data based on multi-scale gating weights, the method may include:

[0179] S310 . Perform weight prediction on feature information of each scale in the first multi-scale feature information based on the multi-scale gating weight prediction module to obtain scale gating weights corresponding to the feature information of each scale.

[0180] S320 . Obtain multi-scale weighted information based on the feature information at each scale and the scale gating weight corresponding to the feature information at each scale.

[0181] S330. Determine the resource operation data based on the multi-scale weighted information and the first feature information.

[0182] Specifically, the first multi-scale information extraction module may include multiple pooling channels, each having a different pooling scale. Thus, when information is extracted from the first input feature information through the multiple pooling channels, scale feature information corresponding to each pooling channel may be obtained. For different resource operation prediction targets, the degree of reliance on the information extracted by different pooling channels may vary. Therefore, based on the resource operation prediction model including the first multi-scale information extraction module, a multi-scale gating weight prediction module may be further added. The multi-scale gating weight prediction module can be used to predict the weights of each pooling channel. After the first multi-scale information extraction module outputs the multi-scale feature information extracted by each pooling channel from the first input feature information, the multi-scale gating weight prediction module can be used to weight the extracted information output by each pooling channel to obtain weighted multi-scale weighted information. The multi-scale weighted information is obtained by weighting the extracted information output by each pooling channel, and the scale gating weight corresponding to each pooling channel is associated with the prediction target. This makes the multi-scale weighted information more helpful in predicting the prediction target, thereby improving the accuracy of the prediction result.

[0183] In one embodiment, see Figure 4 , which shows a method for extracting multi-scale information from first feature information, the method may include:

[0184] S410. Perform multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information.

[0185] S420. Determine the first input feature information based on the first feature information and the second multi-scale feature information.

[0186] exist Figure 2 In the method shown, a first multi-scale information extraction module is used to perform multi-scale information extraction on the first input feature information, thereby achieving the extraction of rich feature information from the first input feature information; in a specific embodiment, multi-scale information extraction can also be performed on the first feature information; Figure 2 In the first multi-scale information extraction module, the feature information extracted has been highly abstracted, resulting in the loss of feature information. Figure 2 On the basis of the above, a second multi-scale information extraction module is added again. The second multi-scale information extraction module can realize information extraction of multiple scales on the first feature information, which can improve the diversity and richness of the extracted feature information, and thus improve the feature representation ability of the extracted feature information for the target object and target multimedia resource; further, the input information of a multi-scale information extraction module is determined by the first feature information and the second multi-scale feature information, thereby improving the richness of the first input feature information.

[0187] In one embodiment, see Figure 5 , which shows a method for determining resource operation data based on multi-scale feature information, the method may include:

[0188] S510: Calculate average scale feature information of feature information at each scale in the first multi-scale feature information.

[0189] S520. Perform information fusion on the average scale feature information and the first feature information to obtain fused feature information.

[0190] S530. Determine the resource operation data based on the target fusion feature information.

[0191] The first-scale feature information contains information extracted from multiple scales of the first input feature information. This information can be averaged to obtain the corresponding average-scale feature information. This averaging operation smoothes the output data, preventing overly abrupt output information from a particular pooling channel and improving the stability of the output data. Furthermore, the integration of the average-scale feature information with the first-scale feature information enhances the feature expression capability of the feature data used to determine resource operation data, further improving the accuracy of prediction results.

[0192] In an optional embodiment, the resource operation prediction model includes multiple hidden layers; accordingly, see Figure 6 , which shows a method for determining resource operation data based on multiple hidden layers, the method may include:

[0193] S610. Input the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extract information from the target fusion feature information based on the next hidden layer to obtain second feature information.

[0194] S620. Based on the first multi-scale information extraction module corresponding to the next hidden layer, perform multi-scale information extraction on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information.

[0195] S630. Determine target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information.

[0196] S640. Determine the next hidden layer as the current hidden layer.

[0197] S650. Determine whether there is a next hidden layer after the current hidden layer; if so, execute step S610; if not, execute step S660.

[0198] S660. Determine the resource operation data based on the target fusion feature information corresponding to the current hidden layer.

[0199] In a specific embodiment, the resource operation prediction model may include multiple hidden layers, each hidden layer may correspond to a first multi-scale information extraction module, and the output information of the first multi-scale information extraction module corresponding to the current hidden layer and the first feature information of the current hidden layer may be used as input information for the next hidden layer.

[0200] Each hidden layer and the first multi-scale information extraction module corresponding to each hidden layer can further abstractly extract the input feature information based on the target fusion feature information obtained by the previous hidden layer and the first multi-scale information extraction module corresponding to the previous hidden layer, thereby extracting feature information through multiple hidden layers and the first multi-scale information extraction modules corresponding to the multiple hidden layers, which can improve the feature expression ability of the extracted information, improve the accuracy of feature extraction, and thus improve the accuracy of the prediction results of the prediction model.

[0201] In the specific implementation process, for the above Figure 2 The first multi-scale information extraction module shown in Figure 3 The multi-scale gating weight prediction module shown in Figure 4 The second multi-scale information extraction module shown in can be regarded as part of the multi-scale gating module. The first multi-scale information extraction module, the multi-scale gating weight prediction module and the second multi-scale information extraction module can be used in combination. The specific combinations may include: the first multi-scale information extraction module + the multi-scale gating weight prediction module, the first multi-scale information extraction module + the second multi-scale information extraction module, the first multi-scale information extraction module + the multi-scale gating weight prediction module + the second multi-scale information extraction module, etc. The superposition of multiple modules can make the feature expression ability of the extracted information stronger, and further improve the accuracy of the model prediction results.

[0202] In one specific embodiment, see Figure 7 , which shows the structural diagram of the first multi-scale information extraction module, Figure 7 It can be seen that the first multi-scale information extraction module specifically includes pooling channels with scales of 1, 2, 4, 8, and 16. These pooling channels are used to extract the first input feature information respectively, and then the sigmoid activation function is used to activate the feature information extracted by multiple pooling channels. The average information of the feature information extracted by each pooling channel is then calculated to obtain the output information of the current multi-scale gating module.

[0203] In one specific embodiment, see Figure 8 , which shows the structural diagram of the multi-scale gating weight prediction module. Figure 8 It can be seen that by weighting the output information of different pooling channels, the multi-scale weighted information corresponding to the different pooling channels is obtained, and the average calculation is further performed based on the multi-scale weighted information, and the output information of the current multi-scale gating module is determined based on the average calculation result.

[0204] In one specific embodiment, see Figure 9 , which shows the structural diagram of the second multi-scale information extraction module, Figure 9 It can be seen that the second multi-scale information extraction module performs a multi-scale information extraction operation on the first feature information to obtain corresponding second multi-scale feature information; thus, the input information of the first multi-scale information extraction module is determined based on the fusion result of the first feature information and the second multi-scale feature information.

[0205] In one example, the resource operation data includes resource operation data for multiple candidate resources; accordingly, based on the resource operation data for the multiple candidate resources, the ranking information of each candidate resource in the push list can be determined; based on the ranking information, a target resource is determined from the push list; and the target resource is pushed to the target object. Thus, based on the resource operation data predicted by the resource operation prediction model described above, the corresponding target resource is pushed to the target object, thereby achieving adaptability between the target resource and the target object and improving the user experience.

[0206] In one embodiment, see Figure 10 , which shows a resource operation prediction model training method, which may include:

[0207] S1010. Obtain sample data; the sample data includes sample object feature information of a sample object, sample resource feature information of a sample multimedia resource, and a sample resource operation tag.

[0208] Specifically, sample data includes positive and negative samples. For resource click prediction models, positive samples are multimedia resources that subjects clicked on, while negative samples are multimedia resources that subjects were shown but did not click on. For resource conversion prediction models, positive samples are multimedia resources that subjects clicked on and subsequently converted to, while negative samples are multimedia resources that subjects clicked but did not convert to. The sample resource operation label corresponding to positive samples can be 1, while the sample resource operation label corresponding to negative samples can be 0.

[0209] S1020. Extract the sample object feature information and the sample resource feature information based on the hidden layer of the resource operation prediction model to be trained to obtain first training feature information.

[0210] S030. Based on the first multi-scale information extraction module of the resource operation prediction model to be trained, multi-scale information extraction is performed on the first input training feature information to obtain first multi-scale training feature information; the first input training feature information is obtained by performing feature dimension conversion on the first training feature information.

[0211] S1040. Based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain a resource operation prediction model.

[0212] In the present disclosure, a first multi-scale information extraction module is used to perform multi-scale information extraction on the first input training feature information, thereby obtaining feature information of multiple sampling scales corresponding to the first input training feature information. The extracted feature information of multiple sampling scales can improve the feature representation capability of the target object and the target multimedia resource, and correspondingly improve the feature extraction capability and information fitting capability of the resource operation prediction model, further improving the prediction accuracy of the resource operation prediction model.

[0213] In an alternative embodiment, see Figure 11 , which shows a method for model training based on a multi-scale gating weight prediction module, which may include:

[0214] S1110. Perform weight prediction on feature information of each scale in the first multi-scale training feature information based on the multi-scale gating weight prediction module to obtain scale gating training weights corresponding to the training feature information of each scale.

[0215] S1120. Obtain multi-scale weighted training information based on the training feature information of each scale and the scale-gated training weights corresponding to the training feature information of each scale.

[0216] S1130. Based on the multi-scale weighted training information, the first training feature information, and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0217] Specifically, the first multi-scale information extraction module may include multiple pooling channels, each having a different pooling scale. Thus, when information is extracted from the first input training feature information through the multiple pooling channels, scale feature information corresponding to each pooling channel may be obtained. For different resource operation prediction targets, the degree of reliance on the information extracted by different pooling channels may vary. Therefore, based on the resource operation prediction model including the first multi-scale information extraction module, a multi-scale gating weight prediction module may be further added. The multi-scale gating weight prediction module can be used to predict the weights of each pooling channel. After the first multi-scale information extraction module outputs the multi-scale feature information extracted by each pooling channel from the first input training feature information, the multi-scale gating weight prediction module can be used to weight the extracted information output by each pooling channel to obtain weighted multi-scale weighted training information. The multi-scale weighted training information is obtained by weighting the extracted information output by each pooling channel. The scale gating training weight corresponding to each pooling channel is associated with the prediction target, making the multi-scale weighted training information more helpful in predicting the prediction target, thereby improving the accuracy of the prediction result.

[0218] In one embodiment, see Figure 12 , which shows a method for extracting information based on the second multi-scale information extraction module, the method may include:

[0219] S1210. Perform multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information.

[0220] S1220. Perform information fusion on the first training feature information and the second multi-scale training feature information to obtain the first input training feature information.

[0221] exist Figure 11 In the first multi-scale information extraction module, the feature information extracted has been highly abstracted, resulting in the loss of feature information. Figure 11 On the basis of the above, a second multi-scale information extraction module is added again. The second multi-scale information extraction module can realize information extraction of multiple scales on the first feature training information, which can improve the diversity and richness of the extracted feature information, and thus improve the feature representation ability of the extracted feature information for sample objects and sample multimedia resources; further, the input information of a multi-scale information extraction module is determined by the first training feature information and the second multi-scale training feature information, thereby improving the richness of the first input training feature information.

[0222] In an alternative embodiment, see Figure 13, which shows another resource operation prediction model training method, which may include:

[0223] S1310. Calculate the average scale training feature information of the training feature information at each scale in the first multi-scale training feature information.

[0224] S1320. Perform information fusion on the average scale training feature information and the first training feature information to obtain target fused training feature information.

[0225] S1330. Based on the target fusion training feature information and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0226] The averaging operation smoothes the output data, preventing abrupt outputs from any pooling channel and improving the stability of the output data. Furthermore, the integration of average-scale training feature information with the first training feature information enhances the feature representation of resource operation data, further improving the accuracy of model training.

[0227] In one embodiment, see Figure 14 , which shows a resource operation prediction model training method based on multiple hidden layers, the method may include:

[0228] S1410. Input the target fusion training feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extract information from the target fusion training feature information based on the next hidden layer to obtain second training feature information.

[0229] S1420. Based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction is performed on the second input feature information to obtain third multi-scale training feature information; the second input training feature information is obtained by performing feature dimension conversion on the second training feature information.

[0230] S1430. Based on the third multi-scale training feature information and the second training feature information, determine the target fusion training feature information corresponding to the next hidden layer.

[0231] S1440. Determine the next hidden layer as the current hidden layer.

[0232] S1450. Determine whether there is a next hidden layer after the current hidden layer; if so, execute step S1410; if not, execute step S1460.

[0233] S1460. Based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

[0234] In a specific embodiment, the resource operation prediction model to be trained may include multiple hidden layers, each hidden layer may correspond to a first multi-scale information extraction module, and the output information of the first multi-scale information extraction module corresponding to the current hidden layer and the first feature information of the current hidden layer may be used as input information for the next hidden layer.

[0235] Each hidden layer and the first multi-scale information extraction module corresponding to each hidden layer can further abstractly extract the input feature information based on the target fusion feature information obtained by the previous hidden layer and the first multi-scale information extraction module corresponding to the previous hidden layer, thereby performing information extraction through multiple hidden layers and the first multi-scale information extraction modules corresponding to the multiple hidden layers, which can improve the feature expression ability of the extracted information, improve the accuracy of feature extraction, and thus improve the accuracy of the prediction results.

[0236] In an alternative embodiment, see Figure 15 , which shows a method for determining a resource operation prediction model based on network parameter adjustment, the method may include:

[0237] S1510. Determine loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label.

[0238] S1520. Adjust the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module.

[0239] S1530. Based on the adjusted hidden layer and the adjusted first multi-scale information extraction module, obtain the resource operation prediction model.

[0240] In the specific implementation process, the above Figure 10 The first multi-scale information extraction module shown in Figure 11 The multi-scale gating weight prediction module shown in Figure 12The second multiscale information extraction module shown in FIG can be considered part of the multiscale gating module. The first multiscale information extraction module, the multiscale gating weight prediction module, and the second multiscale information extraction module can be used in combination. Specific combinations may include: a first multiscale information extraction module + a multiscale gating weight prediction module, a first multiscale information extraction module + a second multiscale information extraction module, a first multiscale information extraction module + a multiscale gating weight prediction module + a second multiscale information extraction module, and so on. Accordingly, during the training process, resource operation prediction models containing different modules can be trained separately, thereby obtaining resource operation prediction models containing the first multiscale information extraction module + a multiscale gating weight prediction module, a first multiscale information extraction module + a second multiscale information extraction module, and a first multiscale information extraction module + a multiscale gating weight prediction module + a second multiscale information extraction module. The superposition of multiple modules can enhance the feature expression capability of the extracted information, further improving the fitting capability of the trained resource operation prediction model, and thereby enhancing the accuracy of model predictions.

[0241] In an alternative embodiment, see Figure 16 , which shows a resource push method, which may include:

[0242] S1610. Obtain sample data; the sample data includes sample object feature information of a sample object, sample resource feature information of a sample multimedia resource, and a sample resource operation tag.

[0243] S1620. Extract the sample object feature information and the sample resource feature information based on the hidden layer of the resource operation prediction model to be trained to obtain first training feature information.

[0244] S1630. Based on the first multi-scale information extraction module of the resource operation prediction model to be trained, multi-scale information extraction is performed on the first input training feature information to obtain first multi-scale training feature information; the first input training feature information is obtained by performing feature dimension conversion on the first training feature information.

[0245] S1640. Based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain a resource operation prediction model.

[0246] S1650. Extract the object feature information of the target object and the resource feature information of the target multimedia resource based on the hidden layer of the resource operation prediction model to obtain first feature information.

[0247] S1660. Perform multi-scale information extraction on the first input feature information based on the first multi-scale information extraction module of the resource operation prediction model to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information.

[0248] S1670. Determine resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information.

[0249] S1680. Based on the resource operation data on the multiple candidate resources, determine the ranking information of each candidate resource in the push list.

[0250] S1690. Based on the sorting information, determine the target resource from the push list, and push the target resource to the target object.

[0251] The specific details of the training process of the resource operation prediction model and the prediction process of the resource operation data of the target object can be found in the above embodiments of the present disclosure, and will not be repeated here.

[0252] The following is a specific example to illustrate the specific implementation process of this disclosure. Figure 17 , which shows a click-through rate model and a conversion rate model with a gating network. The input and output of the model may specifically include:

[0253] 1. Object feature design: including object marking, static and dynamic behavior, etc.

[0254] 2. Video feature design: including video tags, video author, and video posterior statistical information;

[0255] 3. Design of positive and negative samples: We designed click-through rate and conversion rate models. For the click-through rate model, positive samples are videos that subjects clicked on, while negative samples are videos that subjects were shown but did not click on. For the conversion rate model, positive samples are videos that subjects clicked and then converted, while negative samples are videos that subjects clicked but did not convert.

[0256] 4. Design of Neural Networks: See Figure 17 The neural network is mainly divided into two parts: click-through rate network and conversion rate network. At the same time, a gating network is added to the click-through rate network to extract the hidden layer structured information of the neural network.

[0257] 5. Gating Module - Feature Pyramid ( Figure 7The first multi-scale information extraction module in the

[15] : performs multi-scale information extraction on the gated hidden layer output. First, element-wise information extraction is performed on the hidden layer output to obtain a dense feature layer. Then, average pooling operations with scales of 1, 2, 4, 8, 16, and 32 are used to extract information from the dense feature layer. Then, the sigmoid activation function is used to activate the multi-scale extracted information. Finally, the gated outputs of different scales are averaged to obtain the importance bias of the feature pyramid gating module on the hidden layer high-dimensional abstract features.

[0258] 6. Gate control module - gate-reweight module ( Figure 8 The Multi-Scale Gated Weight Prediction Module in the Feature Pyramid Module performs channel-by-channel importance assessments based on the importance bias extracted at multiple scales in the Feature Pyramid Module. Different estimated targets have a dependency bias on the importance extracted at multiple scales. By designing the Se-gig module, we can determine the varying degrees of dependency of estimated targets at different scales.

[0259] 7. Gate control module - input calibration module Gsig ( Figure 9 The second multi-scale information extraction module in the feature pyramid module applies multi-scale feature biases to the input hidden layer features of the feature pyramid module. Because the features extracted by the feature pyramid module are already highly abstracted, there is information loss. By adding the Gsig module, the importance of the scales that the feature pyramid module relies on is increased, while the reliance on less important scales is reduced.

[0260] 8. Forward learning of the network: For the object side, the click model and conversion model are a multi-layer fully connected neural network. The input is the object side features and multimedia resource side features, which are transformed through multiple layers of nonlinear transformations, such as Figure 17 As shown, the hidden layer vectors A1 and A2 are output respectively. A1 and A2 represent the top vector of the click rate and the top vector of the conversion rate model respectively.

[0261] 9. Loss function: The hidden vectors A1 and A2 are converted into estimated probability values ​​through the sigmoid function. The sigmoid function calculation formula is as follows:

[0262] σ(a)=1 / (1+exp(-a)) (1)

[0263] The sigmoid function is in the range of (0, 1), thus obtaining the probability values ​​of click-through rate and conversion rate. Then, based on the sample labels (positive and negative samples of click-through rate, positive and negative samples of conversion rate), the logloss loss function is used. The loss function is as follows:

[0264] l t (w t )=-y tlog p t -(1-y t )log(1-p t ) (2)

[0265] Among them, x t ∈R d , the model parameter is w t , the estimated probability is p t =σ(w t ·x t ), σ is the sigmoid function, y t ∈{0,1} is the sample label.

[0266] 10. Network Backward Learning: Stochastic gradient descent is used to minimize the logloss loss function, calculate the gradient of the loss function, and ultimately update the network parameters layer by layer. For click-through rate prediction, the loss value is calculated based on the loss function formula for the samples composed of click-through rate estimates, and the related gradients are calculated. Backpropagation is from top to bottom, first updating the parameters of the top layer A1, then updating the parameters of the next top layer downwards, and finally updating the input feature embedding. For the conversion rate model, the loss value is also calculated based on the loss function formula for the samples composed of conversion rate estimates, and the related gradients are calculated, and the model parameters are updated from top to bottom.

[0267] 11. Online Prediction: After a neural network has fully learned from samples, it can accurately rank online requests. For example, when a user sends a video request, the click-through rate model estimates the user's click-through rate for a particular video, while the conversion rate model estimates the user's purchase rate for the products featured in the video. The click-through rate ranges from 0 to 1, with a higher value indicating a greater likelihood that the model believes the user will click on the video.

[0268] This disclosure significantly improves the accuracy of neural networks in video recommendation ranking. By constructing a priori relationships between neuron combinations, the multilayer perceptron is equipped with the ability to extract structured information, significantly enhancing the model's fitting capabilities and making it suitable for different prediction targets. Specifically, the pyramid feature extractor effectively improves the model's feature extraction capabilities while significantly reducing the computational resources of the multi-scale gating network. The gated bias network fits the biases applied to the pyramid feature extractor's gates by different prediction targets. The input bias network fits the biases applied to the pyramid feature inputs by different prediction targets.

[0269] Figure 18 A resource operation prediction device according to an exemplary embodiment includes:

[0270] The feature information acquisition unit 1810 is configured to acquire object feature information of a target object and resource feature information of a target multimedia resource;

[0271] A first feature information extraction unit 1820 is configured to extract the object feature information and the resource feature information based on the hidden layer of the resource operation prediction model to obtain first feature information;

[0272] A first multi-scale feature information extraction unit 1830 is configured to execute a first multi-scale information extraction module based on the resource operation prediction model to perform multi-scale information extraction on the first input feature information to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information;

[0273] The resource operation data determining unit 1840 is configured to determine the resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information.

[0274] In an exemplary embodiment, the resource operation prediction model further includes a multi-scale gating weight prediction module;

[0275] The device further comprises:

[0276] a scale gating weight determining unit, configured to perform weight prediction on feature information of each scale in the first multi-scale feature information based on the multi-scale gating weight prediction module, to obtain a scale gating weight corresponding to the feature information of each scale;

[0277] The resource operation data determining unit 1840 includes:

[0278] a multi-scale weighted information unit configured to perform feature information based on each scale and scale gating weights corresponding to the feature information at each scale to obtain multi-scale weighted information;

[0279] The first determining unit is configured to determine the resource operation data based on the multi-scale weighted information and the first feature information.

[0280] In an exemplary embodiment, the resource operation prediction model further includes a second multi-scale information extraction module;

[0281] The device further comprises:

[0282] a second multi-scale feature information determining unit, configured to perform multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information;

[0283] The first input feature information determining unit is configured to determine the first input feature information based on the first feature information and the second multi-scale feature information.

[0284] In an exemplary embodiment, the resource operation data determining unit 1840 includes:

[0285] an average scale feature information determining unit, configured to calculate average scale feature information of feature information of each scale in the first multi-scale feature information;

[0286] a fused feature information determining unit, configured to perform information fusion on the average scale feature information and the first feature information to obtain target fused feature information;

[0287] The second determining unit is configured to determine the resource operation data based on the target fusion feature information.

[0288] In an exemplary embodiment, the resource operation prediction model includes a plurality of hidden layers;

[0289] The second determining unit includes:

[0290] The second feature information determining unit is configured to input the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extract information from the target fusion feature information based on the next hidden layer to obtain second feature information;

[0291] a third multi-scale feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information;

[0292] a third determining unit configured to determine target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information;

[0293] A first hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer;

[0294] A first hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer;

[0295] The fourth determining unit is configured to determine the resource operation data based on the target fusion feature information corresponding to the current hidden layer.

[0296] In an exemplary embodiment, the resource operation data includes resource operation data for a plurality of candidate resources;

[0297] The device further comprises:

[0298] a ranking information determining unit, configured to determine ranking information of each candidate resource in the push list based on the resource operation data on the plurality of candidate resources;

[0299] a target resource determining unit, configured to determine a target resource from the push list based on the sorting information;

[0300] The target resource pushing unit is configured to push the target resource to the target object.

[0301] Figure 19 A resource operation prediction model training device according to an exemplary embodiment includes:

[0302] The sample data acquisition unit 1910 is configured to acquire sample data; the sample data includes sample object feature information of a sample object, sample resource feature information of a sample multimedia resource, and a sample resource operation tag;

[0303] A first training feature information determining unit 1920 is configured to extract the sample object feature information and the sample resource feature information based on the hidden layer of the resource operation prediction model to be trained to obtain first training feature information;

[0304] A first multi-scale training feature information extraction unit 1930 is configured to perform multi-scale information extraction on first input training feature information based on a first multi-scale information extraction module based on the resource operation prediction model to be trained, to obtain first multi-scale training feature information; the first input training feature information is obtained by performing feature dimension conversion on the first training feature information;

[0305] The resource operation prediction model determination unit 1940 is configured to train the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain a resource operation prediction model.

[0306] In an exemplary embodiment, the resource operation prediction model to be trained further includes a multi-scale gating weight prediction module;

[0307] The device further comprises:

[0308] a scale-gated training weight determination unit, configured to perform weight prediction on feature information of each scale in the first multi-scale training feature information based on the multi-scale gating weight prediction module, to obtain scale-gated training weights corresponding to the training feature information of each scale;

[0309] The resource operation prediction model determination unit includes:

[0310] a multi-scale weighted training information determining unit configured to execute a scale-gated training weight based on the training feature information of each scale and the training feature information of each scale, to obtain multi-scale weighted training information;

[0311] The first training unit is configured to train the resource operation prediction model to be trained based on the multi-scale weighted training information, the first training feature information and the sample resource operation label to obtain the resource operation prediction model.

[0312] In an exemplary embodiment, the resource operation prediction model to be trained further includes a second multi-scale information extraction module;

[0313] The device further comprises:

[0314] a second multi-scale training feature information determining unit, configured to perform multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information;

[0315] The first input training feature information determining unit is configured to perform information fusion on the first training feature information and the second multi-scale training feature information to obtain the first input training feature information.

[0316] In an exemplary embodiment, the resource operation prediction model determination unit includes:

[0317] an average scale training feature information determining unit, configured to calculate average scale training feature information of the training feature information of each scale in the first multi-scale training feature information;

[0318] a fusion training feature information determining unit, configured to perform information fusion on the average scale training feature information and the first training feature information to obtain target fusion training feature information;

[0319] The second training unit is configured to train the resource operation prediction model to be trained based on the target fusion training feature information and the sample resource operation label to obtain the resource operation prediction model.

[0320] In an exemplary embodiment, the resource operation prediction model to be trained includes a plurality of hidden layers;

[0321] The second training unit includes:

[0322] a second training feature information determining unit configured to input the target fusion training feature information corresponding to the current hidden layer into a hidden layer next to the current hidden layer, and extract information from the target fusion training feature information based on the next hidden layer to obtain second training feature information;

[0323] a third multi-scale training feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale training feature information; the second input training feature information is determined based on feature dimension conversion performed on the second training feature information;

[0324] A second hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer;

[0325] A second hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer;

[0326] The fifth determining unit is configured to perform training on the resource operation prediction model to be trained based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label to obtain the resource operation prediction model.

[0327] In an exemplary embodiment, the resource operation prediction model determination unit includes:

[0328] a loss information determining unit, configured to determine loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label;

[0329] a parameter adjustment unit configured to adjust the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module;

[0330] The fourth determining unit is configured to execute the resource operation prediction model based on the adjusted hidden layer and the adjusted first multi-scale information extraction module.

[0331] Regarding the apparatus in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0332] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided. Optionally, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a tape, a floppy disk, an optical data storage device, etc.; when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform any of the methods described above.

[0333] In an exemplary embodiment, a computer program product is also provided, which includes a computer program stored in a readable storage medium, and at least one processor of a computer device reads and executes the computer program from the readable storage medium, so that the device performs any of the above methods.

[0334] Figure 20 This is a block diagram of an electronic device for a resource operation data prediction method or a resource operation prediction model training method according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as shown in FIG. Figure 19 As shown. The electronic device includes a processor, a memory, and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a resource operation data prediction method or a resource operation prediction model training method is implemented.

[0335] Those skilled in the art will understand that Figure 20 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0336] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the following claims.

[0337] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A resource operation data prediction method, characterized in that: include: Acquiring object feature information of a target object and resource feature information of a target multimedia resource; Extracting the object feature information and the resource feature information based on a hidden layer of a resource operation prediction model to obtain first feature information; performing multi-scale information extraction on the first input feature information by a first multi-scale information extraction module based on the resource operation prediction model to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information; determining resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information; The determining, based on the first multi-scale feature information and the first feature information, resource operation data of the target object on the target multimedia resource includes: Obtaining multi-scale weighted information based on feature information of each scale in the first multi-scale feature information and scale gating weights corresponding to the feature information of each scale; determining the resource operation data based on the multi-scale weighted information and the first feature information; or, Calculating average scale feature information of feature information of each scale in the first multi-scale feature information; fusing the average scale feature information and the first feature information to obtain target fused feature information; and determining the resource operation data based on the target fused feature information.

2. The method according to claim 1, characterized in that The resource operation prediction model also includes a multi-scale gating weight prediction module; The method further comprises: Based on the multi-scale gating weight prediction module, weight prediction is performed on feature information of each scale in the first multi-scale feature information to obtain scale gating weights corresponding to the feature information of each scale.

3. The method according to claim 1, characterized in that The resource operation prediction model also includes a second multi-scale information extraction module; The method further comprises: Performing multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information; The first input feature information is determined based on the first feature information and the second multi-scale feature information.

4. The method according to claim 1, wherein The resource operation prediction model includes multiple hidden layers; The determining the resource operation data based on the target fusion feature information includes: Inputting the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extracting information from the target fusion feature information based on the next hidden layer to obtain second feature information; Based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction is performed on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information; Determining target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information; Determining the next hidden layer as the current hidden layer; Until there is no next hidden layer after the current hidden layer; The resource operation data is determined based on the target fusion feature information corresponding to the current hidden layer.

5. The method according to claim 1, wherein The resource operation data includes resource operation data for a plurality of candidate resources; The method further comprises: Determining, based on the resource operation data on the plurality of candidate resources, ranking information of each candidate resource in the push list; Based on the ranking information, determining a target resource from the push list; Push the target resource to the target object.

6. A resource operation prediction model training method, characterized in that: include: Acquire sample data; the sample data includes sample object feature information of the sample object, sample resource feature information of the sample multimedia resource, and sample resource operation tags; Extracting the sample object feature information and the sample resource feature information based on the hidden layer of the resource operation prediction model to be trained to obtain first training feature information; Performing multi-scale information extraction on the first input training feature information based on the first multi-scale information extraction module of the resource operation prediction model to be trained to obtain first multi-scale training feature information; The first input training feature information is obtained by performing feature dimension conversion on the first training feature information; The method further comprises: training the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain a resource operation prediction model; and training the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain a resource operation prediction model, including: Based on the training feature information of each scale in the first multi-scale training feature information and the scale gated training weights corresponding to the training feature information of each scale, multi-scale weighted training information is obtained; based on the multi-scale weighted training information, the first training feature information and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model; or, Calculate the average scale training feature information of the training feature information of each scale in the first multi-scale training feature information; perform information fusion on the average scale training feature information and the first training feature information to obtain target fused training feature information; and train the resource operation prediction model to be trained based on the target fused training feature information and the sample resource operation label to obtain the resource operation prediction model.

7. The method according to claim 6, characterized in that The resource operation prediction model to be trained also includes a multi-scale gating weight prediction module; The method further comprises: Based on the multi-scale gating weight prediction module, weight prediction is performed on the feature information of each scale in the first multi-scale training feature information to obtain the scale gating training weight corresponding to the training feature information of each scale.

8. The method according to claim 6, characterized in that The resource operation prediction model to be trained also includes a second multi-scale information extraction module; The method further comprises: Performing multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information; The first training feature information and the second multi-scale training feature information are fused to obtain the first input training feature information.

9. The method according to claim 6, characterized in that The resource operation prediction model to be trained includes multiple hidden layers; The step of training the resource operation prediction model to be trained based on the target fusion training feature information and the sample resource operation label to obtain the resource operation prediction model includes: Inputting the target fusion training feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extracting information from the target fusion training feature information based on the next hidden layer to obtain second training feature information; performing multi-scale information extraction on the second input feature information based on the first multi-scale information extraction module corresponding to the next hidden layer to obtain third multi-scale training feature information; the second input training feature information is determined based on feature dimension conversion performed on the second training feature information; Determining target fusion training feature information corresponding to the next hidden layer based on the third multi-scale training feature information and the second training feature information; Determining the next hidden layer as the current hidden layer; Until there is no next hidden layer after the current hidden layer; Based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label, the resource operation prediction model to be trained is trained to obtain the resource operation prediction model.

10. The method according to claim 6, characterized in that The training of the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain the resource operation prediction model includes: Determining loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label; Adjusting the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module; The resource operation prediction model is obtained based on the adjusted hidden layer and the adjusted first multi-scale information extraction module.

11. A resource operation data prediction device, characterized in that: include: A feature information acquisition unit configured to acquire object feature information of a target object and resource feature information of a target multimedia resource; A first feature information extraction unit is configured to extract the object feature information and the resource feature information based on a hidden layer of a resource operation prediction model to obtain first feature information; a first multi-scale feature information extraction unit configured to execute a first multi-scale information extraction module based on the resource operation prediction model to perform multi-scale information extraction on the first input feature information to obtain first multi-scale feature information; the first input feature information is obtained by performing feature dimension conversion on the first feature information; a resource operation data determining unit, configured to determine resource operation data of the target object on the target multimedia resource based on the first multi-scale feature information and the first feature information; The resource operation data determining unit includes: a multi-scale weighted information unit configured to obtain multi-scale weighted information based on feature information of each scale in the first multi-scale feature information and scale gating weights corresponding to the feature information of each scale; a first determining unit configured to determine the resource operation data based on the multi-scale weighted information and the first feature information; or, The resource operation data determination unit includes: an average scale feature information determination unit, configured to calculate average scale feature information of feature information of each scale in the first multi-scale feature information; a fusion feature information determination unit, configured to perform information fusion on the average scale feature information and the first feature information to obtain target fusion feature information; and a second determination unit, configured to determine the resource operation data based on the target fusion feature information.

12. The device according to claim 11, characterized in that The resource operation prediction model also includes a multi-scale gating weight prediction module; The device further comprises: The scale gating weight determining unit is configured to perform weight prediction on feature information of each scale in the first multi-scale feature information based on the multi-scale gating weight prediction module to obtain a scale gating weight corresponding to the feature information of each scale.

13. The device according to claim 11, characterized in that The resource operation prediction model also includes a second multi-scale information extraction module; The device further comprises: a second multi-scale feature information determining unit, configured to perform multi-scale information extraction on the first feature information based on the second multi-scale information extraction module to obtain second multi-scale feature information; The first input feature information determining unit is configured to determine the first input feature information based on the first feature information and the second multi-scale feature information.

14. The device according to claim 11, characterized in that The resource operation prediction model includes multiple hidden layers; The second determining unit includes: The second feature information determining unit is configured to input the target fusion feature information corresponding to the current hidden layer into the next hidden layer of the current hidden layer, and extract information from the target fusion feature information based on the next hidden layer to obtain second feature information; a third multi-scale feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale feature information; the second input feature information is obtained by performing feature dimension conversion on the second feature information; a third determining unit configured to determine target fusion feature information corresponding to the next hidden layer based on the third multi-scale feature information and the second feature information; A first hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer; A first hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer; The fourth determining unit is configured to determine the resource operation data based on the target fusion feature information corresponding to the current hidden layer.

15. The device according to claim 11, characterized in that The resource operation data includes resource operation data for a plurality of candidate resources; The device further comprises: a ranking information determining unit, configured to determine ranking information of each candidate resource in the push list based on the resource operation data on the plurality of candidate resources; a target resource determining unit, configured to determine a target resource from the push list based on the sorting information; The target resource pushing unit is configured to push the target resource to the target object.

16. A resource operation prediction model training device, characterized in that: include: A sample data acquisition unit is configured to acquire sample data; the sample data includes sample object feature information of a sample object, sample resource feature information of a sample multimedia resource, and a sample resource operation tag; A first training feature information determining unit is configured to extract the sample object feature information and the sample resource feature information based on a hidden layer of the resource operation prediction model to be trained to obtain first training feature information; A first multi-scale training feature information extraction unit is configured to perform multi-scale information extraction on the first input training feature information based on the first multi-scale information extraction module based on the resource operation prediction model to be trained, to obtain first multi-scale training feature information; The first input training feature information is obtained by performing feature dimension conversion on the first training feature information; a resource operation prediction model determination unit, configured to train the resource operation prediction model to be trained based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label to obtain a resource operation prediction model; The resource operation prediction model determination unit includes: A multi-scale weighted training information determination unit is configured to execute training feature information of each scale in the first multi-scale training feature information, and scale gated training weights corresponding to the training feature information of each scale, to obtain multi-scale weighted training information; a first training unit is configured to execute training of the resource operation prediction model to be trained based on the multi-scale weighted training information, the first training feature information, and the sample resource operation label, to obtain the resource operation prediction model; or, The resource operation prediction model determination unit includes: The average scale training feature information determination unit is configured to calculate the average scale training feature information of the training feature information of each scale in the first multi-scale training feature information; the fusion training feature information determination unit is configured to perform information fusion on the average scale training feature information and the first training feature information to obtain target fusion training feature information; the second training unit is configured to train the resource operation prediction model to be trained based on the target fusion training feature information and the sample resource operation label to obtain the resource operation prediction model.

17. The device according to claim 16, characterized in that The resource operation prediction model to be trained also includes a multi-scale gating weight prediction module; The device further comprises: The scale gated training weight determination unit is configured to perform weight prediction on the feature information of each scale in the first multi-scale training feature information based on the multi-scale gated weight prediction module, and obtain the scale gated training weight corresponding to the training feature information of each scale.

18. The device according to claim 16, characterized in that The resource operation prediction model to be trained also includes a second multi-scale information extraction module; The device further comprises: a second multi-scale training feature information determining unit, configured to perform multi-scale information extraction on the first training feature information based on the second multi-scale information extraction module to obtain second multi-scale training feature information; The first input training feature information determining unit is configured to perform information fusion on the first training feature information and the second multi-scale training feature information to obtain the first input training feature information.

19. The device according to claim 16, characterized in that The resource operation prediction model to be trained includes multiple hidden layers; The second training unit includes: a second training feature information determining unit configured to input the target fusion training feature information corresponding to the current hidden layer into a hidden layer next to the current hidden layer, and extract information from the target fusion training feature information based on the next hidden layer to obtain second training feature information; a third multi-scale training feature information determining unit configured to execute, based on the first multi-scale information extraction module corresponding to the next hidden layer, multi-scale information extraction on the second input feature information to obtain third multi-scale training feature information; the second input training feature information is determined based on feature dimension conversion performed on the second training feature information; A second hidden layer determining unit is configured to determine the next hidden layer as a current hidden layer; A second hidden layer determination unit is configured to execute until there is no next hidden layer after the current hidden layer; The fifth determining unit is configured to perform training on the resource operation prediction model to be trained based on the target fusion training feature information corresponding to the current hidden layer and the sample resource operation label to obtain the resource operation prediction model.

20. The device according to claim 16, wherein The resource operation prediction model determination unit includes: a loss information determining unit, configured to determine loss information based on the first multi-scale training feature information, the first training feature information, and the sample resource operation label; a parameter adjustment unit configured to adjust the network parameters of the hidden layer and the network parameters of the first multi-scale information extraction module based on the loss information to obtain an adjusted hidden layer and an adjusted first multi-scale information extraction module; The fourth determining unit is configured to execute the resource operation prediction model based on the adjusted hidden layer and the adjusted first multi-scale information extraction module.

21. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the resource operation data prediction method according to any one of claims 1 to 5, or the resource operation prediction model training method according to any one of claims 6 to 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 enabled to execute the resource operation data prediction method as described in any one of claims 1 to 5, or the resource operation prediction model training method as described in any one of claims 6 to 10.

23. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting resource operation data according to any one of claims 1 to 5 or the method for training a resource operation prediction model according to any one of claims 6 to 10 is implemented.

Images