Resource Recommendation Model Training Method, Resource Information Recommendation Method and Device

Abstract

The present disclosure relates to a method for training a resource recommendation model, a method and apparatus for resource information recommendation. The method includes: inputting multiple weighted sample object resource information into a first model to be trained for recommendation recognition processing to determine first recommendation index information. Based on the first recommendation index information, update the preset weight information to obtain updated weight information corresponding to each of the multiple sample object resource information, and based on the updated weight information, determine target object resource information from the multiple sample object resource information. Based on the target object resource information, train a second model to be trained to obtain a resource recommendation model. This method can screen out target object resource information that is more important for the recommendation index information, thereby reducing the computational amount of feature processing, enhancing the learning of the model for the target object resource information during the training process, and improving the accuracy of resource recommendation.

Description

Technical Field

[0001] The present disclosure relates to the field of recommendation technologies, and in particular, to a method for training a resource recommendation model, a method for recommending resource information, and an apparatus. Background Art

[0002] With the rise of the Internet, the amount of information on the network has also shown an explosive growth. In order to provide a better user experience for users, it is necessary to mine a small amount of information from the vast amount of information and accurately push it to different users. In related technologies, when performing resource recommendation, a large amount of feature information needs to be processed, which increases the computational amount of resource recommendation and there is a problem that the feature information cannot be distinguished, resulting in a decrease in the accuracy of resource recommendation. Summary of the Invention

[0003] The present disclosure provides a method for training a resource recommendation model, a method for recommending resource information, and an apparatus, so as to at least solve the problems of large computational amount and low accuracy in resource recommendation in related technologies. The technical solutions of the present disclosure are as follows:

[0004] According to a first aspect of an embodiment of the present disclosure, a method for training a resource recommendation model is provided, and the method includes:

[0005] After performing weighted processing on multiple sample object resource information based on preset weight information, input it into a first model to be trained for recommendation recognition processing to determine first recommendation index information, where the sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object;

[0006] Based on the first recommendation index information, update the preset weight information to obtain updated weight information corresponding to each of the multiple sample object resource information, where the updated weight information is used to represent the importance degree of each of the multiple sample object resource information to the first recommendation index information;

[0007] Based on the updated weight information, determine target object resource information from the multiple sample object resource information;

[0008] Based on the target object resource information, train a second model to be trained to obtain a resource recommendation model, where the second model to be trained is obtained by adjusting the network structure in the trained first model to be trained.

[0009] As an optional embodiment, the updating the preset weight information based on the first recommendation index information to obtain updated weight information corresponding to each of the multiple sample object resource information includes:

[0010] Determine the first loss information based on the first recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource;

[0011] Update the preset weight information based on the first loss information to obtain the updated weight information.

[0012] As an optional embodiment, the updating the preset weight information based on the first loss information to obtain the updated weight information includes:

[0013] Update the model parameters in the first model to be trained based on the first loss information to obtain an initial feature screening model;

[0014] Update the preset weight information based on the initial feature screening model to obtain the updated weight information.

[0015] As an optional embodiment, the first recommendation metric information includes multiple recommendation metric information, and the determining the first loss information based on the first recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource includes:

[0016] Determine the recommendation loss information between each of the recommendation metric information and the preset recommendation annotation information;

[0017] Obtain the first loss information based on each of the recommendation loss information.

[0018] As an optional embodiment, the initial feature screening model includes metric identification layers corresponding to multiple recommendation metrics. After updating the model parameters in the first model to be trained based on the first loss information to obtain the initial feature screening model, the method further includes:

[0019] Determine the target recommendation metric among the multiple recommendation metrics;

[0020] Update the other metric identification layers in the metric identification layer except the target metric identification layer to feature fusion layers, where the target metric identification layer is the metric identification layer corresponding to the target recommendation metric;

[0021] Construct a second model to be trained based on the feature fusion layer and the target metric identification layer.

[0022] As an optional embodiment, the training the second model to be trained based on the target object resource information to obtain a resource recommendation model includes:

[0023] Input the target object resource information into the second model to be trained for recommendation recognition processing to determine the second recommendation metric information;

[0024] Obtain second loss information based on the second recommended metric information and the preset recommended annotation information corresponding to the sample multimedia resource;

[0025] Train the second model to be trained based on the second loss information to obtain the resource recommendation model.

[0026] As an optional embodiment, the second model to be trained includes a feature fusion layer and a target metric recognition layer. The target metric recognition layer includes a feature perception layer and a recommended metric determination layer. The feature perception layer includes multiple sequentially connected perception layers. Inputting the target object resource information into the second model to be trained for recommendation recognition processing and determining the second recommended metric information includes:

[0027] Input the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0028] When the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain the perception processing result corresponding to the first perception layer;

[0029] When the current perception layer is not the first perception layer, obtain the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0030] Input the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer;

[0031] Use the perception processing result output by the last perception layer as the perception feature information;

[0032] Input the fused feature information and the perception feature information into the recommended metric determination layer for recommended metric determination to obtain the second recommended metric information.

[0033] As an optional embodiment, determining the target object resource information from multiple pieces of sample object resource information based on the updated weight information includes:

[0034] Compare the updated weight information corresponding to each piece of sample object resource information with a preset weight threshold;

[0035] Use the sample object resource information with updated weight information greater than or equal to the preset weight threshold as the target object resource information.

[0036] According to a second aspect of the embodiments of the present disclosure, there is provided a resource information recommendation method, the method including:

[0037] Obtain multiple target object resource information from the object resource information corresponding to the target object. The multiple target object resource information are object resource information whose update weight information meets a preset condition. The update weight information represents the importance degree of the object resource information to the recommendation index information. The object resource information is determined based on the object information of the target object and the resource information of the multimedia resource to be recommended;

[0038] Input the target object resource information into the resource recommendation model obtained by the above resource recommendation model training method for recommendation recognition processing to determine the recommendation index information corresponding to the multimedia resource to be recommended;

[0039] Based on the recommendation index information, determine the target recommended resource from the multimedia resources to be recommended;

[0040] Recommend the target recommended resource to the target object.

[0041] As an optional embodiment, the resource recommendation model includes a feature fusion layer and a target index recognition layer. The target index recognition layer includes a feature perception layer and a recommendation index determination layer. The step of inputting the target object resource information into the resource recommendation model for recommendation recognition processing to determine the recommendation index information corresponding to the resource information to be recommended includes:

[0042] Input the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0043] Input the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information;

[0044] Input the fused feature information and the perceived feature information into the recommendation index determination layer for recommendation index determination to obtain the recommendation index information.

[0045] As an optional embodiment, the feature perception layer includes multiple sequentially connected perception layers. The step of inputting the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information includes:

[0046] When the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain the perception processing result corresponding to the first perception layer;

[0047] When the current perception layer is not the first perception layer, obtain the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0048] Input the target object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer;

[0049] Use the perception processing result output by the last perception layer as the perception feature information.

[0050] According to the third aspect of the embodiments of the present disclosure, a resource recommendation model training device is provided. The device includes:

[0051] A first recommendation recognition module, configured to perform weighted processing on multiple sample object resource information based on preset weight information, and then input it into a first model to be trained for recommendation recognition processing to determine first recommendation index information. The sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object;

[0052] A weight update module, configured to perform an update on the preset weight information based on the first recommendation index information to obtain updated weight information corresponding to each of the multiple sample object resource information. The updated weight information is used to represent the importance degree of each of the multiple sample object resource information to the first recommendation index information;

[0053] A target information acquisition module, configured to perform determination of target object resource information from the multiple sample object resource information based on the updated weight information;

[0054] A resource recommendation model training module, configured to perform training on a second model to be trained based on the target object resource information to obtain a resource recommendation model. The second model to be trained is obtained by adjusting the network structure in the trained first model to be trained.

[0055] As an optional embodiment, the weight update module includes:

[0056] A first loss information determination unit, configured to perform determination of first loss information based on the first recommendation index information and the preset recommendation annotation information corresponding to the sample multimedia resource;

[0057] A weight update unit, configured to perform an update on the preset weight information based on the first loss information to obtain the updated weight information.

[0058] As an optional embodiment, the weight update unit includes:

[0059] A model parameter update unit, configured to perform an update on the model parameters in the first model to be trained based on the first loss information to obtain an initial feature screening model;

[0060] An updated weight information acquisition unit, configured to perform an update on the preset weight information based on the initial feature screening model to obtain the updated weight information.

[0061] As an optional embodiment, the first recommendation metric information includes multiple pieces of recommendation metric information, and the first loss information determination unit includes:

[0062] A recommendation loss information determination unit, configured to perform a determination of the recommendation loss information between each piece of the recommendation metric information and the preset recommendation annotation information;

[0063] A first loss information acquisition unit, configured to perform an operation of obtaining the first loss information based on each of the recommendation loss information.

[0064] As an optional embodiment, the initial feature screening model includes metric identification layers corresponding to multiple recommendation metrics, and the apparatus further includes:

[0065] A target metric determination module, configured to perform a determination of a target recommendation metric among the multiple recommendation metrics;

[0066] A model update module, configured to perform an update of other metric identification layers in the metric identification layer except the target metric identification layer into feature fusion layers, where the target metric identification layer is the metric identification layer corresponding to the target recommendation metric;

[0067] A second model to be trained construction module, configured to perform a construction of the second model to be trained based on the feature fusion layer and the target metric identification layer.

[0068] As an optional embodiment, the resource recommendation model training module includes:

[0069] A second recommendation metric determination unit, configured to perform a recommendation identification process by inputting the target object resource information into the second model to be trained to determine second recommendation metric information;

[0070] A second loss information determination unit, configured to perform an operation of obtaining second loss information based on the second recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource;

[0071] A model training unit, configured to perform a training of the second model to be trained based on the second loss information to obtain the resource recommendation model.

[0072] As an optional embodiment, the second model to be trained includes a feature fusion layer and a target metric recognition layer. The target metric recognition layer includes a feature perception layer and a recommended metric determination layer. The second recommended metric determination unit includes:

[0073] A feature fusion unit, configured to input the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0074] A first feature perception unit, configured to, when the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer;

[0075] A previous perception processing result acquisition unit, configured to, when the current perception layer is not the first perception layer, acquire a previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0076] A second feature perception unit, configured to input the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer;

[0077] A perceived feature information acquisition unit, configured to use the perception processing result output by the last perception layer as the perceived feature information;

[0078] A second recommended metric information acquisition unit, configured to input the fused feature information and the perceived feature information into the recommended metric determination layer for recommended metric determination to obtain the second recommended metric information.

[0079] As an optional embodiment, the target information acquisition module includes:

[0080] A comparison unit, configured to compare the updated weight information corresponding to each sample object resource information with a preset weight threshold;

[0081] A target information determination unit, configured to use the sample object resource information with updated weight information greater than or equal to the preset weight threshold as the target object resource information.

[0082] According to a fourth aspect of the embodiments of the present disclosure, there is provided a resource information recommendation device, the device includes:

[0083] An object resource acquisition module, configured to execute acquiring multiple target object resource information from the object resource information corresponding to the target object, where the multiple target object resource information are object resource information whose update weight information meets a preset condition, and the update weight information represents the importance degree of the object resource information to the recommendation index information, and the object resource information is determined based on the object information of the target object and the resource information of the multimedia resource to be recommended;

[0084] A recommendation recognition module, configured to execute inputting the target object resource information into a resource recommendation model obtained by the above resource recommendation model training method for recommendation recognition processing to determine the recommendation index information corresponding to the multimedia resource to be recommended;

[0085] A target recommended resource determination module, configured to execute determining a target recommended resource from the multimedia resources to be recommended based on the recommendation index information;

[0086] A recommendation module, configured to execute recommending the target recommended resource to the target object.

[0087] As an optional embodiment, the resource recommendation model includes a feature fusion layer and a target index recognition layer, the target index recognition layer includes a feature perception layer and a recommendation index determination layer, and the recommendation recognition module includes:

[0088] A feature fusion unit, configured to execute inputting the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0089] A feature perception unit, configured to execute inputting the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information;

[0090] A recommendation index determination unit, configured to execute inputting the fused feature information and the perceived feature information into the recommendation index determination layer for recommendation index determination to obtain the recommendation index information.

[0091] As an optional embodiment, the feature perception layer includes multiple sequentially connected perception layers, and the feature perception unit includes:

[0092] A first perception processing unit, configured to execute, when the current perception layer is the first perception layer, inputting the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer;

[0093] An upper perception processing result acquisition unit, configured to execute, when the current perception layer is not the first perception layer, acquiring an upper perception processing result corresponding to the previous perception layer of the current perception layer;

[0094] A second perception processing unit, configured to perform inputting the target object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer;

[0095] A perception feature information acquisition unit, configured to perform taking the perception processing result output by the last perception layer as the perception feature information.

[0096] According to a fifth aspect of the embodiments of the present disclosure, there is provided an electronic device, including:

[0097] A processor;

[0098] A memory for storing executable instructions of the processor;

[0099] Wherein, the processor is configured to execute the instructions to implement the above-mentioned resource recommendation model training method and the above-mentioned resource information recommendation method.

[0100] According to a sixth aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enabling the electronic device to execute the above-mentioned resource recommendation model training method and the above-mentioned resource information recommendation method.

[0101] According to a seventh aspect of the embodiments of the present disclosure, there is provided a computer program product, including a computer program, where when the computer program is executed by a processor, it implements the above-mentioned resource recommendation model training method and the above-mentioned resource information recommendation method.

[0102] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0103] Based on preset weight information, perform weighted processing on multiple sample object resource information corresponding to a sample object to obtain multiple weighted object resource information. Input the multiple weighted object resource information into a first model to be trained for recommendation recognition processing to determine first recommendation index information. Based on the first recommendation index information, update the preset weight information to obtain updated weight information corresponding to each of the multiple sample object resource information, and based on the updated weight information, determine target object resource information from the multiple sample object resource information. Based on the target object resource information, train a second model to be trained to obtain a resource recommendation model. This method can screen out target object resource information that is more important for the recommendation index information, thereby reducing the computational amount of feature processing, enhancing the learning of the model for the target object resource information during the training process, and improving the accuracy of resource recommendation

[0104] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0106] Figure 1 is a schematic diagram of an application scenario of a resource recommendation model training method shown according to an exemplary embodiment.

[0107] Figure 2 is a flowchart of a resource recommendation model training method shown according to an exemplary embodiment.

[0108] Figure 3 is a schematic diagram of a first model to be trained in a resource recommendation model training method shown according to an exemplary embodiment.

[0109] Figure 4 is a flowchart of obtaining updated weight information in a resource recommendation model training method shown according to an exemplary embodiment.

[0110] Figure 5 is a flowchart of constructing a second model to be trained in a resource recommendation model training method shown according to an exemplary embodiment.

[0111] Figure 6 is a flowchart of obtaining second recommendation metric information in a resource recommendation model training method shown according to an exemplary embodiment

[0112] Figure 7 is a flowchart of a resource information recommendation method shown according to an exemplary embodiment.

[0113] Figure 8 is a flowchart of determining recommendation metric information in a resource information recommendation method shown according to an exemplary embodiment.

[0114] Figure 9 is a schematic diagram of a resource recommendation model in a resource information recommendation method shown according to an exemplary embodiment.

[0115] Figure 10 is a schematic diagram of a feature crossing process in a resource information recommendation method shown according to an exemplary embodiment.

[0116] Figure 11 is a flowchart of obtaining perceptual feature information in a resource information recommendation method shown according to an exemplary embodiment.

[0117] Figure 12 It is a block diagram of a resource recommendation model training device shown according to an exemplary embodiment.

[0118] Figure 13 It is a block diagram of a resource information recommendation device shown according to an exemplary embodiment.

[0119] Figure 14 It is a block diagram of a server-side electronic device shown according to an exemplary embodiment. Detailed implementation manners

[0120] In order to enable those of ordinary skill 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.

[0121] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such used data may be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0122] 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 for display, data for analysis, etc.) involved in the present disclosure are all information and data that have been authorized by the user or fully authorized by all parties.

[0123] First, the following explanations are made for the related terms involved in the embodiments of the present application:

[0124] Click-Through-Rate (CTR): It refers to the click-through rate of multimedia resources, that is, the actual number of clicks on the multimedia resources divided by the display volume of the multimedia resources. Strictly speaking, the actual number of clicks can be the number of pages reaching the target page corresponding to the multimedia resources.

[0125] Conversion Rate (CVR): It is an index for measuring the recommendation effect of multimedia resources. Briefly speaking, it is the conversion rate from when a user clicks on a multimedia resource to becoming an effective activated or registered or even paying user. CVR = (conversion volume / click volume) * 100%.

[0126] The Sigmoid function is an S-shaped function commonly found in biology, also known as the S-shaped growth curve. In information science, due to its properties such as being monotonically increasing and its inverse function being monotonically increasing, the Sigmoid function is often used as an activation function in neural networks to map variables between 0 and 1.

[0127] Figure 1 It is a schematic diagram of an application scenario of a resource recommendation model training method shown according to an exemplary embodiment. As Figure 1 shown, this application scenario includes a client 110 and a server 120. The client 110 sends object information to the server 120. The server 120 obtains multiple sample object resource information corresponding to the sample object based on the object information and the resource information. The server 120 performs a weighting process on the sample object resource information based on the preset weight information to obtain multiple weighted object resource information, and inputs the multiple weighted object resource information into the first model to be trained for recommendation recognition processing to determine the first recommendation metric information. The server 120 updates the preset weight information based on the first recommendation metric information to obtain updated weight information corresponding to each of the multiple sample object resource information, and determines the target object resource information from the multiple sample object resource information based on the updated weight information. The server 120 trains the second model to be trained based on the target object resource information to obtain a resource recommendation model. The client 110 sends a resource recommendation request to the server 120. The server 120, in response to the resource recommendation request, determines the target recommended resource based on the resource recommendation model and sends the target recommended resource to the client 110.

[0128] In the embodiments of the present disclosure, the client 110 includes entity devices such as smartphones, desktop computers, tablet computers, laptop computers, digital assistants, smart wearable devices, etc., and may also include software running on the entity devices, such as application programs, etc. The operating systems running on the entity devices in the embodiments of the present application may include, but are not limited to, Android systems, IOS systems, linux, Unix, windows, etc. The client 110 includes a UI (User Interface) layer. The client 110 externally provides the collection of object information and the display of the target recommended resource through the UI layer. Additionally, the object information is sent to the server 120 based on the API (Application Programming Interface).

[0129] In an embodiment of the present disclosure, the server 120 may include an independently operating server, or a distributed server, or a server cluster composed of multiple servers. The server 120 may include a network communication unit, a processor, a memory, and so on. Specifically, the server 120 may train a first model to be trained, determine target object resource information, and train a second model to be trained to obtain a resource recommendation model. The server 120 performs a resource recommendation step for the target object based on the resource recommendation model.

[0130] Figure 2 is a flowchart of a resource recommendation model training method shown according to an exemplary embodiment. As Figure 2 shown, this method is used in a server and includes the following steps.

[0131] S210. Based on preset weight information, after weighting multiple sample object resource information, input it into the first model to be trained for recommendation recognition processing to determine first recommendation metric information. The sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object;

[0132] As an optional embodiment, the sample object resource information may include object feature information corresponding to the sample object and resource feature information corresponding to the sample multimedia resource. The object information of the sample object may be object attribute information, and the resource information of the sample multimedia resource may be resource attribute information. By performing feature extraction on the object attribute information and the resource attribute information, sample object resource information can be obtained. When performing feature extraction on the object attribute information and the resource attribute information, feature extraction may be performed through a preset feature extraction network, and converted into data input to the model. The preset feature extraction network may be implemented based on an embedding lookup function. Based on the embedding lookup function, feature processing is performed on the object attribute information and the resource attribute information to obtain sample object resource information that can be input into the first model to be trained. For example, feature extraction is performed on object attribute information such as user ID, user age, and historical click resource information to obtain object feature information, or feature extraction is performed on resource attribute information such as resource category and resource price to obtain resource feature information, and the object feature information and the resource feature information are used as sample object resource information. Among them, the sample multimedia resource may include different multimedia resources such as images, videos, advertisements, and goods.

[0133] As an optional embodiment, the preset weight information may be the weight information after activation processing. Based on a preset activation function, the preset weight information can be activated, so as to map the preset weight information within a preset interval. The activation function may be a sigmoid function, and the sigmoid function can map the preset weight information within the interval [0, 1], and this interval [0, 1] is the preset interval. For each sample object resource information, it can be first normalized, and then based on the preset weight information, each normalized sample object resource information is weighted to obtain multiple weighted object resource information. When normalizing, the batch normalization function (Batch Normalization) can be used to normalize the sample object resource information.

[0134] As an optional embodiment, the first model to be trained is a multi-task teacher model, which can obtain the features of tasks at a higher level than the target task. Inputting multiple weighted object resource information into the first model to be trained for recommendation recognition processing, the first recommendation index information can be determined, and the first recommendation index information can be the recommendation index information corresponding to multiple recommendation indexes. These multiple recommendation indexes correspond to the multi-tasks in the first model to be trained. For example, if the first model to be trained includes two tasks of CTR prediction and CVR prediction, the first recommendation index information includes the click-through rate index information corresponding to the CTR prediction task and the conversion rate index information corresponding to the CVR prediction task.

[0135] S220. Based on the first recommendation index information, update the preset weight information to obtain the updated weight information corresponding to each of the multiple sample object resource information. The updated weight information represents the importance of each of the multiple sample object resource information to the first recommendation index information;

[0136] As an optional embodiment, the updated weight information can represent the importance of each of the multiple sample object resource information to the first recommendation index information. The greater the updated weight of the sample object resource information, the greater its importance to the first recommendation index information, and the smaller the updated weight of the sample object resource information, the smaller its importance to the first recommendation index information. And based on the sample object resource information with greater importance, it is easier to determine the first recommendation index information.

[0137] As an optional embodiment, updating the preset weight information based on the first recommendation index information to obtain the updated weight information corresponding to each of the multiple sample object resource information includes:

[0138] Determine the first loss information based on the first recommendation index information and the preset recommendation annotation information corresponding to the sample multimedia resource;

[0139] Update the preset weight information based on the first loss information to obtain the updated weight information.

[0140] As an optional embodiment, the sample multimedia resources are multimedia resources for which a preset operation has been performed on the sample object and multimedia resources for which the preset operation has not been performed on the sample object. The multimedia resources for which the preset operation has been performed on the sample object are positive sample multimedia resources, and the multimedia resources for which the preset operation has not been performed on the sample object are negative sample multimedia resources. The preset operation can be a click operation or the like. The preset recommended annotation information corresponding to the sample multimedia resources is the positive sample annotation information and the negative sample annotation information. The positive sample annotation information indicates recommendation, and the negative sample annotation information indicates non-recommendation. The first recommended metric information includes two cases: indicating that the recommended metric is satisfied and indicating that the recommended metric is not satisfied. In the case where the recommended metric is satisfied, recommendation is performed, and in the case where the recommended metric is not satisfied, no recommendation is performed.

[0141] By calculating the difference between the first recommended metric information and the preset recommended annotation information corresponding to the sample multimedia resources, the first loss information can be obtained. The first loss information can be a cross-entropy loss function. Based on the first loss information, the preset weight information can be updated to obtain updated weight information. The value of the preset weight information can be set to an initial value of 0. After updating the preset weight information based on the updated weight information, the value of the obtained updated weight information can be a numerical value in the interval [0, 1]. In the interval [0, 1], the closer the updated weight information is to 0, the smaller the importance of the sample object resource information corresponding to the updated weight information to the recommended metric information. The closer the updated weight information is to 1, the greater the importance of the sample object resource information corresponding to the updated weight information to the recommended metric information.

[0142] Training the preset weight information with the first loss information to obtain updated weight information, where the updated weight information represents the importance of multiple sample object resource information to the recommended metric information respectively. Thus, based on the updated weight information, feature screening can be performed to obtain the target object resource information, thereby reducing the number of feature information, reducing the computational resource amount for the feature information, and improving the efficiency of feature information calculation.

[0143] As an optional embodiment, the first recommended metric information includes multiple recommended metric information. Based on the first recommended metric information and the preset recommended annotation information corresponding to the sample multimedia resources, determining the first loss information includes:

[0144] Determining the recommended loss information between each recommended metric information and the preset recommended annotation information;

[0145] Based on each recommended loss information, obtaining the first loss information.

[0146] As an alternative embodiment, the first recommended metric information may include multiple pieces of recommended metric information. In the case of including multiple pieces of recommended metric information, the recommended loss information between each piece of recommended metric information and the preset recommended characterization information may be determined, and the sum of each recommended loss information is calculated, and the obtained sum value is used as the first loss information. For example, the multiple pieces of recommended metric information include click-through rate metric information and conversion rate metric information. The click-through rate metric information may be CTR, and the conversion rate metric information may be CVR. Calculate the difference between the click-through rate metric information and the preset recommended annotation information to obtain the click-through rate loss information. Calculate the difference between the conversion rate metric information and the preset recommended annotation information to obtain the conversion rate loss information. Calculate the sum value between the click-through rate loss information and the conversion rate loss information to obtain the first loss information.

[0147] As an alternative embodiment, in the case where the first recommended metric information includes click-through rate metric information and conversion rate metric information, please refer to Figure 3 , such as Figure 3 shown in the schematic diagram of the first model to be trained of the multi-task structure. The first model to be trained may include a to-be-trained metric recognition layer corresponding to each of multiple recommended metrics. The to-be-trained metric recognition layer may include a to-be-trained click-through rate recognition layer and a to-be-trained conversion rate recognition layer, and a to-be-trained feature perception layer shared by the to-be-trained click-through rate recognition layer and the to-be-trained conversion rate recognition layer. The to-be-trained feature perception layer may be a single-layer perceptron structure, and both the to-be-trained click-through rate recognition layer and the to-be-trained conversion rate recognition layer may be multi-layer perceptron structures. After the sample object resource information is weighted, the weighted object resource information is obtained. The weighted object resource information is input into the to-be-trained feature perception layer of the first model to be trained for feature perception processing, that is, the weighted object resource information is weighted and summed again to obtain the training perception feature information. Then, the training perception feature information is input into the to-be-trained click-through rate recognition layer to determine the click-through rate metric information, and the training perception feature information is input into the to-be-trained conversion rate recognition layer to determine the conversion rate metric information. Calculate the cross-entropy loss function between the click-through rate metric information and the preset recommended annotation information to obtain the click-through rate loss information. Calculate the cross-entropy loss function between the conversion rate metric information and the preset recommended annotation information to obtain the conversion rate loss information. Calculate the sum value between the click-through rate loss information and the CVR loss information to obtain the first loss information.

[0148] In the first model to be trained, set the to-be-trained metric recognition layers corresponding to multiple recommendation metrics, so that the first model to be trained is a multi-task model, and feature learning and model training can be performed through other recommendation metrics other than the target recommendation metric. Therefore, the target object resource information obtained based on the first model to be trained can have knowledge corresponding to other recommendation metrics, thereby improving the accuracy of resource recommendation on the premise of ensuring the computational amount and recommendation efficiency of the resource recommendation model deployed online.

[0149] As an optional embodiment, please refer to Figure 4 , based on the first loss information, update the preset weight information to obtain the updated weight information, including:

[0150] S410. Based on the first loss information, update the model parameters in the first model to be trained to obtain an initial feature screening model;

[0151] S420. Based on the initial feature screening model, determine the weight update parameters corresponding to the preset weight information;

[0152] S430. Based on the weight update parameters, update the preset weight information to obtain the updated weight information.

[0153] As an optional embodiment, the first model to be trained includes a multi-layer neural network. Based on the first loss information, perform backpropagation on the first model to be trained to update the model parameters in the first model to be trained, and obtain the first model to be trained with updated model parameters. Based on the preset weight information, perform weighted processing on multiple sample object resource information and then input it into the first model to be trained with updated model parameters for recommendation recognition processing to obtain updated first recommendation metric information, and based on the updated first recommendation metric information, update the first loss information. Repeat the steps from updating the model parameters in the first model to be trained based on the first loss information to updating the first loss information based on the updated first recommendation metric information. When the model parameters meet the preset parameter convergence condition, an initial feature screening model can be obtained.

[0154] Determine the loss information corresponding to the initial feature screening model, perform backpropagation based on the loss information corresponding to the initial feature screening model, and determine the weight update parameters corresponding to the preset weight information. The weight update parameter can be the gradient information of the preset weight information. Based on the weight update parameter, update the preset weight information to obtain the updated weight information.

[0155] As an optional embodiment, the first model to be trained includes a click-through rate recognition layer to be trained, a conversion rate recognition layer to be trained, and a feature perception layer shared by the click-through rate recognition layer to be trained and the conversion rate recognition layer to be trained. The feature perception layer can be a single-layer perceptron structure, and both the click-through rate recognition layer to be trained and the conversion rate recognition layer to be trained can be multi-layer perceptron structures. Based on the first loss information, backpropagation is performed on the first model to be trained to update the model parameters corresponding to the click-through rate recognition layer to be trained, the conversion rate recognition layer to be trained, and the feature perception layer in the first model to be trained. When the model parameters converge, an initial feature screening model can be obtained. Then, based on the initial feature screening model, the weight update parameters corresponding to the preset weight information are determined. The weight update parameters can be the gradient information of the preset weight information. Based on the weight update parameters, the preset weight information is updated to obtain updated weight information.

[0156] Inputting the first loss information into the first model to be trained for backpropagation can train the preset weight information to obtain updated weight information. Based on the backpropagation method, the interference in updating the preset weight information can be reduced, thereby improving the accuracy of the updated weight information. In subsequent steps, the target object resource information is determined based on the updated weight information, improving the effectiveness of determining the target object resource information.

[0157] As an optional embodiment, please refer to Figure 5 , such as Figure 5 shown, the initial feature screening model includes index recognition layers corresponding to multiple recommendation metrics. After obtaining the initial feature screening model by updating the model parameters in the first model to be trained based on the first loss information, the method further includes:

[0158] S510. Determine the target recommendation metric among multiple recommendation metrics;

[0159] S520. Update the other index recognition layers in the index recognition layer except the target index recognition layer to feature fusion layers, where the target index recognition layer is the index recognition layer corresponding to the target recommendation metric;

[0160] S530. Construct a second model to be trained based on the feature fusion layer and the target index recognition layer.

[0161] As an alternative embodiment, based on the first loss information, the first model to be trained is trained to obtain an initial feature screening model, that is, the first loss information is input into the first model to be trained for backpropagation, and the first model parameter update information of each layer of neural network in the first model to be trained is calculated. The first model parameter update information can be the gradient information corresponding to the model parameters of each layer of neural network in the first model to be trained. Based on the first model parameter update information, the model parameters of each layer of neural network in the first model to be trained are updated, and an initial feature screening model can be obtained. The initial feature screening model includes an index identification layer corresponding to multiple recommendation metrics, where the model parameters in the index identification layer have been trained based on the first loss information, and the multiple recommendation metrics can be two or more recommendation metrics. For example, when the recommendation metrics include the CTR recommendation metric and the CVR recommendation metric, the index identification layer can include a click-through rate identification layer corresponding to the CTR recommendation metric and a conversion rate identification layer corresponding to the CVR recommendation metric.

[0162] The target recommendation metric is the recommendation metric applied when recommending multimedia resources online. When the recommendation metrics include the CTR recommendation metric and the CVR recommendation metric, the CTR recommendation metric can be used as the target recommendation metric. Then, the click-through rate identification layer corresponding to the CTR recommendation metric is the target metric identification layer, and the conversion rate identification layer corresponding to the CVR recommendation metric is the other metric identification layer. The conversion rate identification layer corresponding to the CVR recommendation metric can be updated to a feature fusion layer. Based on the feature fusion layer and the target metric identification layer, a second model to be trained can be constructed. The target metric identification layer in the second model to be trained retains the model parameters that have been trained in the initial feature screening model.

[0163] Modify part of the network structure in the already trained initial feature screening model to obtain the second model to be trained, so that the already trained part of the model parameters in the initial feature screening model can be directly utilized, reducing the number of training rounds of the second model to be trained, thereby improving the training efficiency of the second model to be trained.

[0164] S230. Based on the updated weight information, determine the target object resource information from multiple sample object resource information;

[0165] As an alternative embodiment, determining the target object resource information from multiple sample object resource information based on the updated weight information includes:

[0166] Compare the updated weight information corresponding to each sample object resource information with a preset weight threshold;

[0167] Take the sample object resource information with the updated weight information greater than or equal to the preset weight threshold as the target object resource information.

[0168] As an optional embodiment, based on the updated weight information, the advantageous feature information that is more important for determining the recommended metric information can be filtered out, that is, the target object resource information. The target object resource information is the sample object resource information for which the updated weight information meets the preset conditions. If the updated weight information is greater than the preset weight threshold, the sample object resource information with the updated weight information greater than the preset weight threshold can be used as the target object resource information. If the updated weight information is less than or equal to the preset weight threshold, the sample object resource information corresponding to the updated weight information is not selected as the target object resource information. The preset weight threshold can be 0.5. The number of the target object resource information is less than that of the sample object resource information.

[0169] The updated weight information being greater than the preset weight threshold indicates that the sample object resource information corresponding to the updated weight information is the advantageous feature information with the feature discrimination degree meeting the preset conditions. The sample object resource information corresponding to the updated weight information is more important for the recommended metric information compared with other sample object resource information, and it is easier to obtain the recommended metric information based on the sample object resource information corresponding to the updated weight information. Other sample object resource information refers to the sample object resource information other than the sample object resource information corresponding to the updated weight information greater than the preset weight threshold among the sample object resource information.

[0170] Based on the target object resource information, the recommended metric information can be calculated more accurately. For example, according to the feature of the user's stay duration on the resource details page, the user's CVR recommendation metric can be determined more accurately, and then this feature can be the target object resource information.

[0171] Determining the sample object resource information with the updated weight information greater than the preset weight threshold as the target object resource information can filter out the feature information with high feature discrimination degree and great importance for the recommended metric information, thereby improving the accuracy of determining the target object resource information.

[0172] S240. Based on the target object resource information, train the second model to be trained to obtain a resource recommendation model. The second model to be trained is obtained by adjusting the network structure in the already trained first model to be trained.

[0173] As an optional embodiment, the second model to be trained is a single-task student model. Based on the target object resource information, train the second model to be trained, and the second model to be trained can learn the features corresponding to other tasks except the target task in the target object resource information. Therefore, although the resource recommendation model obtained after training the second model to be trained is a single-task model, it can adopt a feature recognition process similar to that of the trained first model to be trained to identify the target object resource information from the input object resource information.

[0174] As an optional embodiment, training the second model to be trained based on the target object resource information to obtain a resource recommendation model includes:

[0175] Input the target object resource information into the second model to be trained for recommendation recognition processing to determine the second recommendation metric information;

[0176] Based on the second recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource, obtain the second loss information;

[0177] Based on the second loss information, train the second model to be trained to obtain a resource recommendation model.

[0178] As an optional embodiment, inputting the target object resource information into the second model to be trained for recommendation recognition processing can determine the second recommendation metric information. The second recommendation metric information is the recommendation metric information corresponding to the target recommendation metric. For example, when the target recommendation metric is the CTR recommendation metric, the second recommendation metric information is the click-through rate metric information corresponding to the CTR recommendation metric. When inputting the target object resource information into the second model to be trained, since the target object resource information is

[0179] As an optional embodiment, calculating the difference between the second recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource can obtain the second loss information, and the second loss information can be a cross-entropy loss function. When the target recommendation metric is the CTR recommendation metric, the second loss information is the CTR loss information. Input the second loss information into the second model to be trained for backpropagation, and calculate the second model parameter update information for each layer of neural network in the second model to be trained. The second model parameter update information can be the gradient information corresponding to the model parameters of each layer of neural network in the second model to be trained. Based on the second model parameter update information, update the model parameters of each layer of neural network in the second model to be trained to obtain a resource recommendation model.

[0180] Based on the target object resource information, training the second model to be trained can train a resource recommendation model based on the features that are more important for the recommendation metric information, thereby improving the training effectiveness of the second model to be trained.

[0181] As an optional embodiment, please refer to Figure 6 , the second model to be trained includes a feature fusion layer and a target metric recognition layer, the target metric recognition layer includes a feature perception layer and a recommendation metric determination layer, the feature perception layer includes multiple sequentially connected perception layers, and inputting the target object resource information into the second model to be trained for recommendation recognition processing to determine the second recommendation metric information includes:

[0182] S610. Input the target object resource information into the feature fusion layer for feature fusion to obtain the fused feature information;

[0183] S620. When the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain the perception processing result corresponding to the first perception layer;

[0184] S630. When the current perception layer is not the first perception layer, obtain the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0185] S640. Input the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer;

[0186] S650. Use the perception processing result output by the last perception layer as the perception feature information;

[0187] S660. Input the fused feature information and the perception feature information into the recommendation metric determination layer for recommendation metric determination to obtain the second recommendation metric information.

[0188] As an optional embodiment, in the feature fusion layer of the second model to be trained, the target object resource information can be first subjected to feature cross-processing to obtain cross-feature information. Then, the cross-feature information is summed to obtain the fused feature information. The processing methods of feature cross can include inner product, cross product, bilinear transformation, etc. The cross-feature information is the second-order feature information of the target object resource information.

[0189] As an optional embodiment, the feature perception layer of the second model to be trained may include multiple sequentially connected perception layers. In the feature perception layer of the second model to be trained, the target object resource information can be input into the first perception layer for feature perception processing to obtain a first perception processing result, that is, the perception processing result corresponding to the first perception layer. The first perception processing result and the target object resource information are input into the second perception layer for feature perception processing, and a second perception processing result, that is, the perception processing result corresponding to the second perception layer, can be obtained. The second perception processing result and the target object resource information are input into the third perception layer for feature perception processing, and a third perception processing result, that is, the perception processing result corresponding to the third perception layer, can be obtained. And so on. When the current perception layer is not the first perception layer, the output of the previous perception layer of the current perception layer and the target object resource information are input into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer. The output of the previous perception layer of the current perception layer is the previous perception processing result. Until when the current perception layer is the last perception layer, the output of the penultimate perception layer and the target object resource information are input into the last perception layer for feature perception processing to obtain the perception processing result output by the last perception layer, and the perception processing result output by the last perception layer can be used as the perception feature information.

[0190] As an optional embodiment, the fused feature information and the perception feature information are input into the recommendation metric determination layer of the second model to be trained, the fused feature information and the perception feature information are summed, and the sum result is subjected to feature normalization processing to obtain the second recommendation metric information.

[0191] Inputting the target object resource information into the second model to be trained for recommendation metric recognition, performing feature fusion processing on the target object resource information based on the feature fusion layer, and performing feature perception processing on the target object resource information based on multiple perception layers can enhance the second model to be trained's learning of features. Then, determining the second recommendation metric information based on the fused feature information and the perception feature information can improve the accuracy of the second recommendation metric information.

[0192] The embodiments of the present disclosure provide a resource information recommendation method. Please refer to Figure 7 , and the method includes:

[0193] S710. Obtain multiple target object resource information from the object resource information corresponding to the target object. The multiple target object resource information are object resource information whose updated weight information meets a preset condition. The updated weight information represents the importance degree of the object resource information to the recommendation metric information. The object resource information is determined based on the object information of the target object and the resource information of the multimedia resource to be recommended corresponding to the target object;

[0194] Input the target object resource information into the resource recommendation model obtained by training the second model to be trained above for recommendation recognition processing, and determine the recommendation index information corresponding to the multimedia resource to be recommended;

[0195] Based on the recommendation index information, determine the target recommended resource from the multimedia resources to be recommended;

[0196] Recommend the target recommended resource to the target object.

[0197] As an optional embodiment, the target object is a user who needs to recommend multimedia resources, and the object resource information corresponding to the target object is determined based on the object attribute information of the target object and the resource attribute information of the multimedia resource to be recommended. The object attribute information may include user ID, user age, historical click resource information, etc., and the resource attribute information may include resource category, resource price, etc. The object resource information may be feature information extracted by a feature extraction network, and the target object resource information is the object resource information whose updated weight information meets the preset conditions and is the feature information screened by the initial feature screening model.

[0198] During the process of offline model training, when training the first model to be trained to obtain the initial feature screening model, based on the first loss information, perform a backpropagation operation on the first model to be trained, and the weight update parameters corresponding to the preset weight information can be obtained. Based on the weight update parameters, the preset weight information can be updated to obtain the updated weight information corresponding to each sample object resource information. When the updated weight information meets the preset conditions, the sample object resource information corresponding to the updated weight information that meets the preset conditions can be determined as the target object resource information. Based on the target object resource information obtained during the training process, the object resource information that matches the target object resource information can be screened out from the object resource information to obtain the target object resource information during the model application process. This target object resource information is the advantageous feature information that is more important for determining the recommendation index information.

[0199] Input the target object resource information into the resource recommendation model obtained by training the second model to be trained above for recommendation recognition processing, and the recommendation index information corresponding to the multimedia resource to be recommended can be determined. Sort the multimedia resources to be recommended according to the size of the recommendation index information to obtain a sequence of resources to be recommended. Use the first preset number of multimedia resources to be recommended in the sequence of resources to be recommended as the target recommended resources, or use the multimedia resources to be recommended in the sequence of resources to be recommended whose recommendation index information is greater than or equal to the preset recommendation threshold as the target recommended resources.

[0200] In response to a resource recommendation request sent by a target object, send the target recommended resource to the target object or, based on preset resource recommendation information, send the target recommended resource to the target object.

[0201] Input the target object's resource information into a resource recommendation model to identify recommendation metrics, and determine the target recommended resource based on the output recommendation metric information. Recommending the target recommended resource to the target object can enable the resource recommendation model to obtain knowledge of other metric identification layers in the initial feature screening model based on the target object's resource information when identifying the target recommendation metrics, thereby reducing resource consumption and improving the efficiency of resource recommendation.

[0202] As an optional embodiment, please refer to Figure 8 , the resource recommendation model includes a feature fusion layer and a target metric identification layer. The target metric identification layer includes a feature perception layer and a recommendation metric determination layer. Inputting the target object's resource information into the resource recommendation model for recommendation identification processing and determining the recommendation metric information corresponding to the resource information to be recommended includes:

[0203] S810. Input the target object's resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0204] S820. Input the target object's resource information into the feature perception layer for feature perception processing to obtain perceived feature information;

[0205] S830. Input the fused feature information and the perceived feature information into the recommendation metric determination layer for recommendation metric determination to obtain recommendation metric information.

[0206] As an optional embodiment, please refer to Figure 9 , such as Figure 9 shown in the schematic diagram of the resource recommendation model with a single-task structure. The resource recommendation model includes a feature fusion layer and a target metric identification layer. The target metric identification layer includes a feature perception layer and a recommendation metric determination layer. The target metric identification layer is the metric identification layer corresponding to the target recommendation metric, and the target metric identification layer can be an MLP structure. Inputting the target object's resource information into the feature fusion layer of the resource recommendation model for feature fusion can obtain fused feature information. Inputting the target object's resource information into the feature perception layer of the resource recommendation model for feature perception processing can obtain perceived feature information, and then inputting the fused feature information and the perceived feature information into the recommendation metric determination layer of the resource recommendation model for recommendation metric determination can obtain recommendation metric information.

[0207] In the feature fusion layer of the resource recommendation model, the feature cross - processing can be first performed on the target object resource information to obtain cross - feature information. Then, the cross - feature information is summed to obtain the fusion feature information. The processing methods of feature cross - processing can include inner product, cross product, bilinear transformation, etc. The cross - feature information is the second - order feature information of the target object resource information.

[0208] Please refer to Figure 10 , such as Figure 10 shown in the schematic diagram of the inner product of features in the feature fusion layer. As Figure 10 shown, the target object resource information is n feature information, including e1, e2, e3, ……, en. The inner product is performed pairwise on the target object resource information, and the results obtained from the inner product calculation can include c(1,2), c(1,3), ……, c(1,n), c(2,3), ……, c(2,n), ……, c(n - 1,n). The results obtained from the inner product calculation are used as the fusion feature information. When the target object resource information is n feature information, the feature fusion layer performs inner product calculation on pairwise feature information among the n feature information, and n(n - 1) / 2 cross - feature information can be obtained. Based on the following formula, the cross - feature information can be summed to obtain the fusion feature information.

[0209] y 1 = ∑c i

[0210] where y1 is the output of the feature fusion layer, that is, the fusion feature information, ci is the cross - feature information, and i is a value between 1 and n(n - 1) / 2.

[0211] When performing feature perception processing on the target object resource information in the feature perception layer of the resource recommendation model, the weighted sum of the input target object resource information is calculated to obtain the perception feature information. Based on the following formula, the target object resource information can be summed to obtain the perception feature information.

[0212] y 2 = w mlp e i

[0213] where y2 is the output of the feature perception layer, that is, the perception feature information, ei is the target object resource information, and i is a value between 1 and n.

[0214] In the recommendation metric determination layer, a feature mapping step can be performed on the sum value obtained by adding the perceived feature information and the fused feature information, to obtain the value corresponding to the sum value of the perceived feature information and the fused feature information within a preset interval, that is, the recommendation metric information is obtained. When performing the feature mapping step, a preset activation function can be used, and this activation function can be sigmoid. The recommendation metric information can be calculated based on the following formula.

[0215] y out =σ(y 1 ,y 2 )

[0216] Among them, yout is the output of the recommendation metric determination layer, that is, the recommendation metric information, σ() represents the activation function, y1 is the fused feature information, and y2 is the perceived feature information.

[0217] By performing feature fusion in the way of feature crossing to obtain the fused feature information, the dimension of the features can be increased. At the same time, by using a multi-layer perceptron to obtain the perceived feature information, the resource recommendation model's learning of the features can be enhanced. Then, based on the fused feature information and the perceived feature information to determine the recommendation metric information, the effectiveness of determining the recommendation metric information can be improved.

[0218] As an optional embodiment, please refer to Figure 11 , the feature perception layer includes multiple sequentially connected perception layers. Inputting the target object resource information into the feature perception layer for feature perception processing, the obtained perceived feature information includes:

[0219] S1110. When the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain the perception processing result corresponding to the first perception layer;

[0220] S1120. When the current perception layer is not the first perception layer, obtain the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0221] S1130. Input the target object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer;

[0222] S1140. Use the perception processing result output by the last perception layer as the perceived feature information.

[0223] As an alternative embodiment, the feature perception layer of the resource recommendation model may include multiple sequentially connected perception layers. The target object resource information is input into the first perception layer for feature perception processing to obtain a first perception processing result, that is, the perception processing result corresponding to the first perception layer. The first perception processing result and the target object resource information are input into the second perception layer for feature perception processing to obtain a second perception processing result, that is, the perception processing result corresponding to the second perception layer. The second perception processing result and the target object resource information are input into the third perception layer for feature perception processing to obtain a third perception processing result, that is, the perception processing result corresponding to the third perception layer. And so on. When the current perception layer is not the first perception layer, the output of the previous perception layer of the current perception layer and the target object resource information are input into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer. The output of the previous perception layer of the current perception layer is the previous perception processing result. Until when the current perception layer is the last perception layer, the output of the penultimate perception layer and the target object resource information are input into the last perception layer for feature perception processing to obtain the perception processing result output by the last perception layer. The perception processing result output by the last perception layer can be used as the perception feature information.

[0224] As an alternative embodiment, please refer to Figure 9 , such as Figure 9 shown in the schematic diagram of the resource recommendation model with a single-task structure. In the resource recommendation model as shown in Figure 9 , based on the concat operation, the target object resource information is connected to each perception layer in the feature perception layer, so that the target object resource information can be input into each perception layer.

[0225] Connecting the target object resource information to each perception layer in the feature perception layer can enhance the learning of the resource recommendation model for the target object resource information. The target object resource information is information whose importance to the recommendation index information meets the preset conditions. Therefore, the effectiveness of generating recommendation index information in the resource recommendation model can be improved, thereby improving the accuracy of resource recommendation.

[0226] As an optional embodiment, the above-mentioned resource recommendation model training method and resource information recommendation method can be applied to various resource recommendation scenarios such as the advertising business field and the video business field, and the target object is the user. In a specific resource recommendation scenario, during model training, the server obtains multiple sample object resource information corresponding to the sample object, and based on the preset weight information, performs weighted processing on the multiple sample object resource information to obtain multiple weighted object resource information. The server constructs a first model to be trained with multiple tasks, inputs the weighted object resource information into the first model to be trained for recommendation recognition processing, and obtains first recommendation index information, where the first recommendation index information includes the recommendation index information corresponding to each task. Based on the first recommendation index information, the server can update the preset weight information to obtain the updated weight information corresponding to each of the multiple sample object resource information, and based on the updated weight information, determine the target object resource information from the sample object resource information, where the target object resource information is the object resource information whose importance to the recommendation index information meets the preset conditions.

[0227] The server can replace other index recognition layers except the target index recognition layer in the initially obtained feature screening model after training the first model to be trained with a feature fusion layer to obtain a second model to be trained. The server trains the second model to be trained based on the target object resource information to obtain a resource recommendation model.

[0228] During model application, only the resource recommendation model is deployed online, and the initial feature screening model is not deployed online. When the server receives a resource recommendation request sent by the user or obtains resource recommendation information, it can determine the target object resource information from the object resource information corresponding to the user, input the target object resource information into the resource recommendation model obtained by the above model training method for recommendation recognition processing, and determine the recommendation index information corresponding to the multimedia resource to be recommended. The server obtains the multimedia resources to be recommended whose recommendation index information meets the preset conditions as the target resource information, and recommends these target resource information to the user.

[0229] An embodiment of the present disclosure provides a method for training a resource recommendation model. The method includes: based on preset weight information, performing weighted processing on multiple sample object resource information corresponding to a sample object to obtain multiple weighted object resource information. Inputting the multiple weighted object resource information into a first model to be trained for recommendation recognition processing to determine first recommendation index information. Based on the first recommendation index information, updating the preset weight information to obtain updated weight information corresponding to each of the multiple sample object resource information, and based on the updated weight information, determining target object resource information from the multiple sample object resource information. Based on the target object resource information, training a second model to be trained to obtain a resource recommendation model. This method can screen out target object resource information that is more important for the recommendation index information, thereby reducing the computational complexity of feature processing, enhancing the learning of the model for the target object resource information during the training process, and improving the accuracy of resource recommendation.

[0230] Figure 12 is a block diagram of a resource recommendation model training device shown according to an exemplary embodiment. Refer to Figure 12 , the device includes:

[0231] A first recommendation recognition module 1210, configured to perform weighted processing on multiple sample object resource information based on preset weight information and then input it into a first model to be trained for recommendation recognition processing to determine first recommendation index information, where the sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object;

[0232] A weight update module 1220, configured to perform updating on the preset weight information based on the first recommendation index information to obtain updated weight information corresponding to each of the multiple sample object resource information, and the updated weight information is used to represent the importance degree of each of the multiple sample object resource information to the first recommendation index information;

[0233] A target information acquisition module 1230, configured to perform determining target object resource information from the multiple sample object resource information based on the updated weight information;

[0234] A resource recommendation model training module 1240, configured to perform training on a second model to be trained based on the target object resource information to obtain a resource recommendation model, where the second model to be trained is obtained by adjusting the network structure in the trained first model to be trained.

[0235] As an optional embodiment, the weight update module 1220 includes:

[0236] A first loss information determination unit, configured to perform determining first loss information based on the first recommendation index information and preset recommendation annotation information corresponding to the sample multimedia resource;

[0237] A weight update unit, configured to perform an update on preset weight information based on first loss information to obtain updated weight information.

[0238] As an optional embodiment, the weight update unit includes:

[0239] A model parameter update unit, configured to perform an update on model parameters in a first model to be trained based on first loss information to obtain an initial feature screening model;

[0240] An updated weight information acquisition unit, configured to perform an update on preset weight information based on the initial feature screening model to obtain updated weight information.

[0241] As an optional embodiment, the first recommendation metric information includes multiple types of recommendation metric information, and the first loss information determination unit includes:

[0242] A recommendation loss information determination unit, configured to determine recommendation loss information between each type of recommendation metric information and preset recommendation annotation information;

[0243] A first loss information acquisition unit, configured to obtain first loss information based on each piece of recommendation loss information.

[0244] As an optional embodiment, the initial feature screening model includes metric identification layers corresponding to multiple types of recommendation metrics, and the apparatus further includes:

[0245] A target metric determination module, configured to determine a target recommendation metric among multiple types of recommendation metrics;

[0246] A model update module, configured to update other metric identification layers except the target metric identification layer in the metric identification layer into feature fusion layers, where the target metric identification layer is the metric identification layer corresponding to the target recommendation metric;

[0247] A second model to be trained construction module, configured to construct a second model to be trained based on the feature fusion layer and the target metric identification layer.

[0248] As an optional embodiment, the resource recommendation model training module 1240 includes:

[0249] A second recommendation metric determination unit, configured to input target object resource information into the second model to be trained for recommendation identification processing to determine second recommendation metric information;

[0250] A second loss information determination unit, configured to obtain second loss information based on the second recommendation metric information and preset recommendation annotation information corresponding to the sample multimedia resource;

[0251] The model training unit is configured to train a second model to be trained based on second loss information to obtain a resource recommendation model.

[0252] As an optional embodiment, the second model to be trained includes a feature fusion layer and a target metric identification layer, and the target metric identification layer includes a feature perception layer and a recommendation metric determination layer. The second recommendation metric determination unit includes:

[0253] The feature fusion unit is configured to input target object resource information into the feature fusion layer for feature fusion to obtain fused feature information;

[0254] The first feature perception unit is configured to, when the current perception layer is the first perception layer, input target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer;

[0255] The previous perception processing result acquisition unit is configured to, when the current perception layer is not the first perception layer, acquire the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0256] The second feature perception unit is configured to input object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer;

[0257] The perceived feature information acquisition unit is configured to use the perception processing result output by the last perception layer as the perceived feature information;

[0258] The second recommendation metric information acquisition unit is configured to input the fused feature information and the perceived feature information into the recommendation metric determination layer for recommendation metric determination to obtain second recommendation metric information.

[0259] As an optional embodiment, the target information acquisition module 1230 includes:

[0260] The comparison unit is configured to compare the updated weight information corresponding to each sample object resource information with a preset weight threshold;

[0261] The target information determination unit is configured to use the sample object resource information with updated weight information greater than or equal to the preset weight threshold as the target object resource information.

[0262] Regarding the devices in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0263] Figure 13It is a block diagram of a resource information recommendation device shown according to an exemplary embodiment. Referring to Figure 11 , the device includes:

[0264] An object resource acquisition module 1310, configured to obtain multiple target object resource information from the object resource information corresponding to the target object. The multiple target object resource information is object resource information whose updated weight information meets a preset condition. The updated weight information represents the importance degree of the object resource information to the recommendation index information. The object resource information is determined based on the object information of the target object and the resource information of the multimedia resource to be recommended corresponding to the target object;

[0265] A recommendation recognition module 1320, configured to perform recommendation recognition processing by inputting the target object resource information into a resource recommendation model obtained during the training process of the above-mentioned second model to be trained, and determine the recommendation index information corresponding to the multimedia resource to be recommended;

[0266] A target recommendation resource determination module 1330, configured to determine a target recommendation resource from the multimedia resources to be recommended based on the recommendation index information;

[0267] A recommendation module 1340, configured to recommend the target recommendation resource to the target object.

[0268] As an optional embodiment, the resource recommendation model includes a feature fusion layer and a target index recognition layer. The target index recognition layer includes a feature perception layer and a recommendation index determination layer. The recommendation recognition module includes:

[0269] A feature fusion unit, configured to perform feature fusion by inputting the target object resource information into the feature fusion layer to obtain fused feature information;

[0270] A feature perception unit, configured to perform feature perception processing by inputting the target object resource information into the feature perception layer to obtain perceived feature information;

[0271] A recommendation index determination unit, configured to perform recommendation index determination by inputting the fused feature information and the perceived feature information into the recommendation index determination layer to obtain recommendation index information.

[0272] As an optional embodiment, the feature perception layer includes multiple sequentially connected perception layers. The feature perception unit includes:

[0273] A first perception processing unit, configured to perform feature perception processing by inputting the target object resource information into the first perception layer when the current perception layer is the first perception layer, and obtain a perception processing result corresponding to the first perception layer;

[0274] The previous perception processing result acquisition unit is configured to, when the current perception layer is not the first perception layer, acquire the previous perception processing result corresponding to the previous perception layer of the current perception layer;

[0275] The second perception processing unit is configured to perform inputting the target object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer;

[0276] The perception feature information acquisition unit is configured to perform taking the perception processing result output by the last perception layer as the perception feature information.

[0277] Regarding the device in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0278] Figure 14 is a block diagram of an electronic device for resource recommendation model training or resource information recommendation shown according to an exemplary embodiment. The electronic device may be a server, and its internal structure diagram may be as Figure 14 shown. The electronic device includes a processor, a memory, and a network interface connected through a system bus. Among them, 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 through a network connection. When the computer program is executed by the processor, it implements a method for resource recommendation model training or resource information recommendation.

[0279] Those skilled in the art can understand that Figure 14 the structure shown in

[0280] is only a block diagram of some structures related to the solution of the present disclosure, and does not constitute a limitation on the electronic device to which the solution of the present disclosure is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0281] In an exemplary embodiment, a computer program product is further provided, including a computer program which, when executed by a processor, implements the resource recommendation model training method and the resource information recommendation method as described above.

[0282] Those skilled in the art will readily conceive of other embodiments of the present disclosure 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, which follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not disclosed herein. The specification and examples are only to be considered as exemplary, and the true scope and spirit of the present disclosure are pointed out by the following claims.

[0283] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure is only limited by the appended claims.

Claims

1. A method for training a resource recommendation model, characterized in that, the method includes: Based on preset weight information, after weighting multiple sample object resource information, input it into a first model to be trained for recommendation recognition processing to determine first recommendation index information. The sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object; Based on the first recommendation index information, update the preset weight information to obtain updated weight information corresponding to each of the multiple sample object resource information. The updated weight information is used to characterize the importance of each of the multiple sample object resource information to the first recommendation index information; Based on the updated weight information, determine target object resource information from the multiple sample object resource information; Based on the target object resource information, train a second model to be trained to obtain a resource recommendation model. The second model to be trained is obtained by adjusting the network structure in the trained first model to be trained; The first model to be trained is used to update model parameters based on first loss information to obtain an initial feature screening model. The first loss information is determined based on the first recommendation index information and preset recommendation annotation information corresponding to the sample multimedia resource; The second model to be trained is constructed based on the initial feature screening model. The second model to be trained includes a feature fusion layer and a target index recognition layer. The target index recognition layer includes a feature perception layer and a recommendation index determination layer. The feature perception layer includes multiple sequentially connected perception layers. The feature fusion layer is used to perform feature fusion on the target object resource information to obtain fused feature information. The perception layer is used to perform feature perception processing on the target object resource information and the previous perception processing result corresponding to the previous perception layer to obtain perception feature information. The recommendation index determination layer is used to determine a recommendation index according to the fused feature information and the perception feature information.

2. The resource recommendation model training method according to claim 1, characterized in that, the updating the preset weight information based on the first recommendation index information to obtain updated weight information corresponding to each of the multiple sample object resource information includes: Determine first loss information based on the first recommendation index information and preset recommendation annotation information corresponding to the sample multimedia resource; Based on the first loss information, update the preset weight information to obtain the updated weight information.

3. The resource recommendation model training method according to claim 2, characterized in that, the updating the preset weight information based on the first loss information to obtain the updated weight information includes: Based on the first loss information, update the model parameters in the first model to be trained to obtain an initial feature screening model; Based on the initial feature screening model, determine weight update parameters corresponding to the preset weight information; Based on the weight update parameters, update the preset weight information to obtain the updated weight information.

4. The method for training a resource recommendation model according to any one of claims 2 or 3, wherein, the first recommendation metric information includes multiple pieces of recommendation metric information, and determining the first loss information based on the first recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource includes: determining the recommendation loss information between each piece of the recommendation metric information and the preset recommendation annotation information; obtaining the first loss information based on each piece of the recommendation loss information.

5. The method for training a resource recommendation model according to claim 3, wherein, the initial feature screening model includes metric identification layers corresponding to multiple recommendation metrics. After updating the model parameters in the first model to be trained based on the first loss information to obtain the initial feature screening model, the method further includes: determining a target recommendation metric among the multiple recommendation metrics; updating the other metric identification layers in the metric identification layers except the target metric identification layer to feature fusion layers, where the target metric identification layer is the metric identification layer corresponding to the target recommendation metric; constructing the second model to be trained based on the feature fusion layers and the target metric identification layer.

6. The method for training a resource recommendation model according to claim 1, wherein, training the second model to be trained based on the target object resource information to obtain a resource recommendation model includes: inputting the target object resource information into the second model to be trained for recommendation recognition processing to determine second recommendation metric information; obtaining second loss information based on the second recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource; training the second model to be trained based on the second loss information to obtain the resource recommendation model.

7. The method for training a resource recommendation model according to claim 6, wherein, inputting the target object resource information into the second model to be trained for recommendation recognition processing to determine second recommendation metric information includes: inputting the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information; when the current perception layer is the first perception layer, inputting the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer; when the current perception layer is not the first perception layer, obtaining the previous perception processing result corresponding to the previous perception layer of the current perception layer; inputting the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer; using the perception processing result output by the last perception layer as perception feature information; inputting the fused feature information and the perception feature information into the recommendation metric determination layer for recommendation metric determination to obtain the second recommendation metric information.

8. The method for training a resource recommendation model according to claim 1, wherein, Determining the target object resource information from multiple pieces of the sample object resource information based on the updated weight information includes: Comparing the updated weight information corresponding to each piece of the sample object resource information with a preset weight threshold; Taking the sample object resource information with the updated weight information greater than or equal to the preset weight threshold as the target object resource information.

9. A resource information recommendation method, characterized in that, the method includes: Obtaining multiple target object resource information from the object resource information corresponding to a target object, where the multiple target object resource information are object resource information whose updated weight information meets a preset condition, the updated weight information characterizes the importance degree of the object resource information to the recommendation index information, and the object resource information is determined based on the object information of the target object and the resource information of the multimedia resource to be recommended; Inputting the target object resource information into a resource recommendation model trained by the resource recommendation model training method according to any one of claims 1-8 for recommendation recognition processing to determine the recommendation index information corresponding to the multimedia resource to be recommended; Determining a target recommended resource from the multimedia resources to be recommended based on the recommendation index information; Recommending the target recommended resource to the target object.

10. The resource information recommendation method according to claim 9, characterized in that, the resource recommendation model includes a feature fusion layer and a target index recognition layer, the target index recognition layer includes a feature perception layer and a recommendation index determination layer, and inputting the target object resource information into the resource recommendation model for recommendation recognition processing to determine the recommendation index information corresponding to the multimedia resource to be recommended includes: Inputting the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information; Inputting the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information; Inputting the fused feature information and the perceived feature information into the recommendation index determination layer for recommendation index determination to obtain the recommendation index information.

11. The resource information recommendation method according to claim 10, characterized in that, the feature perception layer includes multiple sequentially connected perception layers, and inputting the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information includes: When the current perception layer is the first perception layer, inputting the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer; When the current perception layer is not the first perception layer, obtaining the previous perception processing result corresponding to the previous perception layer of the current perception layer; Inputting the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer; Taking the perception processing result output by the last perception layer as the perceived feature information.

12. A resource recommendation model training device, characterized in that, the device includes: The first recommendation recognition module is configured to perform weighted processing on multiple sample object resource information based on preset weight information, and then input it into the first model to be trained for recommendation recognition processing to determine the first recommendation index information. The sample object resource information is determined based on the object information of the sample object and the resource information of the sample multimedia resource corresponding to the sample object; The weight update module is configured to perform an update on the preset weight information based on the first recommendation index information to obtain updated weight information corresponding to each of the multiple sample object resource information. The updated weight information is used to characterize the importance of each of the multiple sample object resource information to the first recommendation index information; The target information acquisition module is configured to determine target object resource information from the multiple sample object resource information based on the updated weight information; The resource recommendation model training module is configured to perform training on the second model to be trained based on the target object resource information to obtain a resource recommendation model. The second model to be trained is obtained by adjusting the network structure in the already trained first model to be trained; The first model to be trained is used to update model parameters based on first loss information to obtain an initial feature screening model. The first loss information is determined based on the first recommendation index information and the preset recommendation annotation information corresponding to the sample multimedia resource; The second model to be trained is constructed based on the initial feature screening model. The second model to be trained includes a feature fusion layer and a target index recognition layer. The target index recognition layer includes a feature perception layer and a recommendation index determination layer. The feature perception layer includes multiple sequentially connected perception layers. The feature fusion layer is used to perform feature fusion on the target object resource information to obtain fused feature information. The perception layer is used to perform feature perception processing on the object resource information and the previous perception processing result corresponding to the previous perception layer to obtain perception feature information. The recommendation index determination layer is used to determine a recommendation index based on the fused feature information and the perception feature information.

13. The resource recommendation model training device according to claim 12, wherein, the weight update module includes: The first loss information determination unit is configured to determine first loss information based on the first recommendation index information and the preset recommendation annotation information corresponding to the sample multimedia resource; The weight update unit is configured to perform an update on the preset weight information based on the first loss information to obtain the updated weight information.

14. The resource recommendation model training device according to claim 13, wherein, the weight update unit includes: The model parameter update unit is configured to perform an update on the model parameters in the first model to be trained based on the first loss information to obtain an initial feature screening model; The updated weight information acquisition unit is configured to perform an update on the preset weight information based on the initial feature screening model to obtain the updated weight information.

15. The resource recommendation model training device according to any one of claims 13 or 14, characterized in that, the first recommendation metric information includes a variety of recommendation metric information, and the first loss information determination unit includes: a recommendation loss information determination unit configured to determine the recommendation loss information between each of the recommendation metric information and the preset recommendation annotation information; a first loss information acquisition unit configured to obtain the first loss information based on each of the recommendation loss information.

16. The resource recommendation model training device according to claim 14, characterized in that, the initial feature screening model includes metric identification layers corresponding to a variety of recommendation metrics, and the device further includes: a target metric determination module configured to determine a target recommendation metric among the variety of recommendation metrics; a model update module configured to update other metric identification layers in the metric identification layer except the target metric identification layer to feature fusion layers, where the target metric identification layer is the metric identification layer corresponding to the target recommendation metric; a second model to be trained construction module configured to construct the second model to be trained based on the feature fusion layer and the target metric identification layer.

17. The resource recommendation model training device according to claim 12, characterized in that, the resource recommendation model training module includes: a second recommendation metric determination unit configured to input the target object resource information into the second model to be trained for recommendation identification processing to determine second recommendation metric information; a second loss information determination unit configured to obtain second loss information based on the second recommendation metric information and the preset recommendation annotation information corresponding to the sample multimedia resource; a model training unit configured to train the second model to be trained based on the second loss information to obtain the resource recommendation model.

18. The resource recommendation model training device according to claim 17, characterized in that, the second recommendation metric determination unit includes: a feature fusion unit configured to input the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information; a first feature perception unit configured to, when the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer; a previous perception processing result acquisition unit configured to, when the current perception layer is not the first perception layer, acquire a previous perception processing result corresponding to the previous perception layer of the current perception layer; a second feature perception unit configured to input the object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain a perception processing result corresponding to the current perception layer; a perception feature information acquisition unit configured to use the perception processing result output by the last perception layer as perception feature information; The second recommended metric information acquisition unit is configured to input the fused feature information and the perceived feature information into the recommended metric determination layer to determine the recommended metric, and obtain the second recommended metric information.

19. The resource recommendation model training device according to claim 12, wherein, the target information acquisition module includes: a comparison unit configured to compare the updated weight information corresponding to each sample object resource information with a preset weight threshold; a target information determination unit configured to use the sample object resource information with the updated weight information greater than or equal to the preset weight threshold as the target object resource information.

20. A resource information recommendation device, wherein, the device includes: an object resource acquisition module configured to obtain a plurality of target object resource information from the object resource information corresponding to the target object, the plurality of target object resource information being object resource information whose updated weight information meets a preset condition, the updated weight information representing the importance degree of the object resource information to the recommended metric information, and the object resource information being determined based on the object information of the target object and the resource information of the multimedia resource to be recommended; a recommendation recognition module configured to input the target object resource information into a resource recommendation model trained by the resource recommendation model training method according to any one of claims 1-8 for recommendation recognition processing to determine the recommended metric information corresponding to the multimedia resource to be recommended; a target recommended resource determination module configured to determine a target recommended resource from the multimedia resources to be recommended based on the recommended metric information; a recommendation module configured to recommend the target recommended resource to the target object.

21. The resource information recommendation device according to claim 20, wherein, the resource recommendation model includes a feature fusion layer and a target metric recognition layer, the target metric recognition layer includes a feature perception layer and a recommended metric determination layer, and the recommendation recognition module includes: a feature fusion unit configured to input the target object resource information into the feature fusion layer for feature fusion to obtain fused feature information; a feature perception unit configured to input the target object resource information into the feature perception layer for feature perception processing to obtain perceived feature information; a recommended metric determination unit configured to input the fused feature information and the perceived feature information into the recommended metric determination layer to determine the recommended metric, and obtain the recommended metric information.

22. The resource information recommendation device according to claim 21, wherein, the feature perception layer includes a plurality of sequentially connected perception layers, and the feature perception unit includes: a first perception processing unit configured to, when the current perception layer is the first perception layer, input the target object resource information into the first perception layer for feature perception processing to obtain a perception processing result corresponding to the first perception layer; The previous perception processing result acquisition unit is configured to, when the current perception layer is not the first perception layer, acquire the previous perception processing result corresponding to the previous perception layer of the current perception layer; The second perception processing unit is configured to perform inputting the target object resource information and the previous perception processing result into the current perception layer for feature perception processing to obtain the perception processing result corresponding to the current perception layer; The perception feature information acquisition unit is configured to perform using the perception processing result output by the last perception layer as the perception feature information.

23. An electronic device Characterized in that It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to execute the instructions to implement the resource recommendation model training method according to any one of claims 1 to 8 and the resource information recommendation method according to any one of claims 9 to 11.

24. A computer-readable storage medium Characterized in that When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the resource recommendation model training method according to any one of claims 1 to 8 and the resource information recommendation method according to any one of claims 9 to 11.

25. A computer program product comprising a computer program Characterized in that When the computer program is executed by a processor, it implements the resource recommendation model training method according to any one of claims 1 to 8 and the resource information recommendation method according to any one of claims 9 to 11.

Images