Recommendation model training method, device, computing equipment and medium

Abstract

The embodiments of the present disclosure provide a training method, apparatus, computing device and medium for a recommendation model. Sample feature information and sample recommendation scores of sample objects are obtained, and the first recommendation score and the second recommendation score of each sample object under each prediction target are determined respectively through the initial recommendation model, so as to train the initial recommendation model based on the sample recommendation score, the first recommendation score and the second recommendation score. The number of hidden layers included in the target task network does not exceed the first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed the second set threshold, so that the network parameters of the target task network are relatively small. On the basis of ensuring that there are fewer conflicts between multiple prediction targets through multiple initial task networks, the target task network can better learn the common knowledge between different task targets, thereby improving the accuracy of the trained network parameters and improving the ranking accuracy of the recommendation model.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of deep learning technology. More specifically, the embodiments of the present disclosure relate to a training method, apparatus, computing device, and medium for a recommendation model. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the present disclosure that are recited in the claims. No statement herein is admitted to be prior art by virtue of its inclusion in this section.

[0003] Recommendation algorithms, which can deeply explore user needs and provide targeted content recommendations, have been widely used in content recommendation processes. Generally speaking, recommendation algorithms are divided into two stages. The first stage is recall, which selects hundreds to thousands of candidate objects from the entire list of recommended objects. The second stage is ranking, which uses a ranking model to select a few to dozens of target objects from these hundreds to thousands of candidate objects and then rank these few to dozens of target objects.

[0004] In related technologies, the sorting model generally adopts a multi-gate mixture of experts (MMoE) model based on a deep neural network.

[0005] However, the MMoE model reduces conflicts between multiple task objectives by increasing network parameters. However, this increase in network parameters may lead to over-parameterization of the network, making the MMoE model unable to effectively learn the shared knowledge between different task objectives. This results in lower accuracy of the adjusted network parameters and, in turn, lower ranking accuracy. Summary of the Invention

[0006] In view of the problem that the model training effect in related technologies is poor, the accuracy of the network parameters obtained through training is low, and thus the ranking accuracy of the recommendation model is low, the embodiments of the present disclosure at least provide a training method, apparatus, computing device and medium for a recommendation model.

[0007] In a first aspect of the embodiments of the present disclosure, a method for training a recommendation model is provided, the method comprising:

[0008] Obtain sample feature information of multiple sample objects and sample recommendation scores corresponding to each sample object;

[0009] Inputting the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model respectively, and obtaining multiple sample first features output by each initial feature extraction network;

[0010] Based on the multiple sample first features output by each initial feature extraction network, a first recommendation score for each sample object under each prediction target is determined through the target task network of the initial recommendation model; and based on the multiple sample first features, a second recommendation score for each sample object under each prediction target is determined through the multiple initial task networks of the initial recommendation model; wherein the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold;

[0011] An initial recommendation model is trained based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object.

[0012] In one embodiment of the present disclosure, based on multiple sample first features output by multiple initial feature extraction networks, a target task network of an initial recommendation model is used to determine a first recommendation score for each sample object under each prediction target, including:

[0013] For any sample object, determine the mean feature of the first features of multiple samples corresponding to the sample object;

[0014] The mean feature is input into the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0015] In one embodiment of the present disclosure, the mean feature is input into the target task network to obtain the first recommendation score of the sample object under each prediction target, including:

[0016] The mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0017] In one embodiment of the present disclosure, each initial task network corresponds to an initial gating network;

[0018] Based on the first features of the multiple samples, the second recommendation score of each sample object under each prediction target is determined through the multiple initial task networks of the initial recommendation model, including:

[0019] For any sample object, multiple initial gating networks of the initial recommendation model are used to perform weighted summation on the first features of multiple samples to obtain the second features of the samples corresponding to each initial task network;

[0020] Input the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target;

[0021] Among them, the number of initial task networks matches the number of prediction targets.

[0022] In one embodiment of the present disclosure, the second features of each sample are input into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target, including:

[0023] The second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

[0024] In one embodiment of the present disclosure, obtaining sample feature information corresponding to a plurality of sample objects includes:

[0025] Based on the sample data corresponding to the multiple sample objects, obtaining multiple sample data features;

[0026] For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

[0027] In one embodiment of the present disclosure, the sample data feature includes at least one of a sample user feature, a sample object feature, and a sample scene feature;

[0028] Among them, the sample user characteristics are used to indicate the user identification of the sample user and / or the user behavior information of the sample user, the sample object characteristics are used to indicate the object identification of the sample object to be recommended and / or the attribute information of the sample object, and the sample scene characteristics are used to indicate the scene in which the sample user requests recommendation.

[0029] In one embodiment of the present disclosure, the sample feature information corresponding to each sample object is input into multiple initial feature extraction networks of the initial recommendation model respectively, and the first features of the samples output by each initial feature extraction network are obtained, including:

[0030] For any sample object, the sample feature information corresponding to the sample object is input into multiple initial feature extraction networks respectively, and the first sample features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks.

[0031] In one embodiment of the present disclosure, training an initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object includes:

[0032] Determining a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target;

[0033] An initial recommendation model is trained based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score.

[0034] In one embodiment of the present disclosure, after training the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object, the method further includes:

[0035] Based on the trained initial recommendation model, the target recommendation model is determined. The target recommendation model includes multiple feature extraction networks, multiple gating networks and multiple task networks.

[0036] In one embodiment of the present disclosure, after determining the target recommendation model based on the trained initial recommendation model, the method further includes:

[0037] Obtaining feature information of candidate objects;

[0038] Inputting the feature information into the target recommendation model, processing the feature information through multiple feature extraction networks of the target recommendation model to obtain multiple first features;

[0039] Inputting the plurality of first features into each gating network to obtain a second feature corresponding to each gating network;

[0040] Each second feature is input into the task network corresponding to the gating network, and the task network outputs the recommendation score of the candidate object under the corresponding prediction target, where one task network corresponds to one prediction target;

[0041] The target object is determined based on the recommendation scores of the candidate objects under each prediction target.

[0042] In one embodiment of the present disclosure, after determining the target object based on the recommendation scores of the candidate objects under each prediction target, the method further includes:

[0043] Sort the target objects in the set order.

[0044] In a second aspect of the embodiments of the present disclosure, a training device for a recommendation model is provided, the device comprising:

[0045] An acquisition module is used to obtain sample feature information of multiple sample objects and a sample recommendation score corresponding to each sample object;

[0046] A processing module, configured to input the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model, and obtain multiple sample first features output by each initial feature extraction network;

[0047] a determination module configured to determine, based on the multiple sample first features output by the multiple initial feature extraction networks, a first recommendation score for each sample object under each prediction target through a target task network of the initial recommendation model, and to determine, based on the multiple sample first features, a second recommendation score for each sample object under each prediction target through the multiple initial task networks of the initial recommendation model; wherein the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold;

[0048] The training module is used to train the initial recommendation model based on the sample recommendation score, the first recommendation score and the second recommendation score corresponding to each sample object.

[0049] In one embodiment of the present disclosure, the determination module, when used to determine the first recommendation score of each sample object under each prediction target through the target task network of the initial recommendation model based on the multiple sample first features output by the multiple initial feature extraction networks, includes a determination unit and a processing unit;

[0050] a determining unit, configured to determine, for any sample object, a mean feature of a plurality of first features of samples corresponding to the sample object;

[0051] The processing unit is used to input the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0052] In one embodiment of the present disclosure, the processing unit, when inputting the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target, is configured to:

[0053] The mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0054] In one embodiment of the present disclosure, each initial task network corresponds to an initial gating network;

[0055] a determination module, comprising a first processing unit and a second processing unit, when used to determine a second recommendation score for each sample object under each prediction target through multiple initial task networks of the initial recommendation model based on the multiple sample first features;

[0056] The first processing unit is configured to perform weighted summation of the first features of multiple samples for any sample object through multiple initial gating networks of the initial recommendation model to obtain the second features of the samples corresponding to each initial task network;

[0057] The second processing unit is used to input the second feature of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target;

[0058] Among them, the number of initial task networks matches the number of prediction targets.

[0059] In one embodiment of the present disclosure, the second processing unit, when inputting the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target, is used to:

[0060] The second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

[0061] In one embodiment of the present disclosure, the acquisition module, when used to acquire sample feature information corresponding to a plurality of sample objects, is configured to:

[0062] Based on the sample data corresponding to the multiple sample objects, obtaining multiple sample data features;

[0063] For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

[0064] In one embodiment of the present disclosure, the sample data feature includes at least one of a sample user feature, a sample object feature, and a sample scene feature;

[0065] Among them, the sample user characteristics are used to indicate the user identification of the sample user and / or the user behavior information of the sample user, the sample object characteristics are used to indicate the object identification of the sample object to be recommended and / or the attribute information of the sample object, and the sample scene characteristics are used to indicate the scene in which the sample user requests recommendation.

[0066] In one embodiment of the present disclosure, the processing module, when inputting sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model to obtain the first sample features output by each initial feature extraction network, is configured to:

[0067] For any sample object, the sample feature information corresponding to the sample object is input into multiple initial feature extraction networks respectively, and the first sample features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks.

[0068] In one embodiment of the present disclosure, a training module is used to train an initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object, including:

[0069] Determining a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target;

[0070] An initial recommendation model is trained based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score.

[0071] In one embodiment of the present disclosure, the determination module is further used to determine a target recommendation model based on the trained initial recommendation model, where the target recommendation model includes multiple feature extraction networks, multiple gating networks, and multiple task networks.

[0072] In one embodiment of the present disclosure, the acquisition module is further configured to acquire feature information of the candidate object;

[0073] The processing module is further configured to input the feature information into the target recommendation model, process the feature information through multiple feature extraction networks of the target recommendation model, and obtain multiple first features;

[0074] The processing module is further configured to input the plurality of first features into each gating network to obtain a second feature corresponding to each gating network;

[0075] The processing module is further configured to input each second feature into a task network corresponding to the gating network, and output a recommendation score of the candidate object under the corresponding prediction target through the task network, wherein one task network corresponds to one prediction target;

[0076] The determination module is further used to determine the target object based on the recommendation scores of the candidate objects under various prediction targets.

[0077] In one embodiment of the present disclosure, the apparatus further comprises:

[0078] The sorting module is used to sort the target objects according to the set order.

[0079] In a third aspect of the embodiments of the present disclosure, a computing device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the operations performed by the training method for the recommendation model provided in the first aspect and any embodiment of the first aspect are implemented.

[0080] In a fourth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, on which a program is stored, and the program is used by a processor to execute the operations performed by the training method of the recommendation model provided in the first aspect and any embodiment of the first aspect.

[0081] In a fifth aspect of the embodiments of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the operations performed by the training method for the recommendation model provided in the first aspect and any embodiment of the first aspect.

[0082] The present disclosure obtains sample feature information of multiple sample objects and sample recommendation scores corresponding to each sample object, thereby determining the first recommendation score and the second recommendation score of each sample object under each prediction target based on the multiple first sample features extracted by the multiple initial feature extraction networks through the target task network and multiple initial task networks of the initial recommendation model, so as to train the initial recommendation model based on the sample recommendation score, the first recommendation score and the second recommendation score. Wherein, the number of hidden layers included in the target task network does not exceed the first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed the second set threshold, so that the network parameters of the target task network are relatively small, while the network parameters of each initial task network are relatively large, so that on the basis of ensuring that there are fewer conflicts between multiple prediction targets through multiple initial task networks, the target task network can better learn the common knowledge between different task targets, thereby improving the accuracy of the trained network parameters and improving the ranking accuracy of the recommendation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example and not limitation, wherein:

[0084] Figure 1 is a flowchart of a training method for a recommendation model according to an exemplary embodiment of the present disclosure;

[0085] Figure 2 is a structural diagram of an initial recommendation model according to an exemplary embodiment of the present disclosure;

[0086] Figure 3 is a structural diagram of a target recommendation model according to an exemplary embodiment of the present disclosure;

[0087] Figure 4 is a flowchart of an object recommendation process according to an exemplary embodiment of the present disclosure;

[0088] Figure 5 is a block diagram of a training device for a recommendation model according to an exemplary embodiment of the present disclosure;

[0089] Figure 6is a schematic diagram of a computer-readable storage medium according to an exemplary embodiment of the present disclosure;

[0090] Figure 7 is a structural diagram of a computing device according to an exemplary embodiment of the present disclosure;

[0091] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. DETAILED DESCRIPTION

[0092] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. Rather, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0093] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0094] It should be understood herein that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0095] According to embodiments of the present disclosure, a method, apparatus, computing device, and medium for training a recommendation model are provided. The method can be executed by a computing device to train a recommendation model. The computing device can be a server, such as a single server, multiple servers, a server cluster, a cloud computing platform, or the like. This disclosure does not limit the type or number of computing devices.

[0096] This recommendation model can be used to recommend content items (such as news recommendations, advertising recommendations, novel recommendations, etc.), multimedia data recommendations (such as audio recommendations, video recommendations, etc.), product recommendations, etc. This disclosure does not limit the specific application scenarios of the recommendation model.

[0097] The following is a reference to the above-mentioned application scenarios. Figure 1 To describe the training method of the recommendation model provided according to the exemplary embodiment of the present disclosure. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0098] See also Figure 1 , Figure 1 FIG. 1 is a flowchart of a method for training a recommendation model according to an exemplary embodiment of the present disclosure, the method comprising:

[0099] S101: Obtain sample feature information of multiple sample objects and a sample recommendation score corresponding to each sample object.

[0100] The sample object may be a content item, multimedia data, commodity, etc. for which a sample recommendation score has been determined, and the sample feature information may be obtained based on the sample data features of the sample object.

[0101] S102: Inputting the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model respectively, and obtaining multiple sample first features output by each initial feature extraction network.

[0102] S103. Based on the multiple sample first features output by the multiple initial feature extraction networks, determine the first recommendation score of each sample object under each prediction target through the target task network of the initial recommendation model, and based on the multiple sample first features, determine the second recommendation score of each sample object under each prediction target through the multiple initial task networks of the initial recommendation model; wherein, the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold.

[0103] Among them, the prediction target can be used to indicate the user's interactive operations on the sample object. For example, the prediction target can be whether the user has finished reading the content item (or the length of time the user has read the content item), whether the user has clicked on the content item, whether the user has purchased the product, whether the user has clicked on the multimedia data, etc. The present disclosure does not limit the specific type of the prediction target.

[0104] Optionally, the first set threshold and the second set threshold may both be any positive integer values, and the present disclosure does not limit the specific values ​​of the first set threshold and the second set threshold.

[0105] S104: Train an initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object.

[0106] The present disclosure introduces a target task network into the initial recommendation model, and the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold, so that the network parameters of the target task network are relatively small, while the network parameters of each initial task network are relatively large. Therefore, on the basis of ensuring that there are fewer conflicts between multiple prediction targets through multiple initial task networks, the target task network can better learn the shared knowledge between different task targets, thereby improving the accuracy of the trained network parameters and improving the ranking accuracy of the recommendation model.

[0107] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0108] In some embodiments, for S101, when obtaining sample feature information corresponding to multiple sample objects, it can be achieved through the following steps:

[0109] S1011. Acquire multiple sample data features based on sample data corresponding to multiple sample objects.

[0110] The sample data may be user log data, and may include user data of the sample user and operation data of the sample user on the sample object (such as click operations, like operations, comment operations, and the time when each operation occurred). Sample data features include at least one of sample user features, sample object features, and sample scenario features. Optionally, the sample data may also include other types of data, and sample data features may also include other types of features, which are not limited in this disclosure.

[0111] In a possible implementation, sample user features may be acquired based on user data corresponding to the sample object, and sample object features and sample scene features may be acquired based on operation data of the sample user on the sample object.

[0112] It should be noted that the sample user characteristics are used to indicate the user identification (ID) of the sample user and / or the user behavior information of the sample user. The user identification of the sample user may be pre-assigned to each sample user, and each sample user has a different corresponding user identification, so that the user identification can serve as a unique identifier for the sample user. The user behavior information of the sample user can be used to indicate the behavioral preferences of the sample user. The user behavior information of the sample user can be determined based on the user's historical interactive operations, for example, the user behavior information of the sample user can be determined based on the user's historical video viewing history, or for another example, the user behavior information of the sample user can be determined based on the user's purchased products, and so on.

[0113] The sample object features are used to indicate the object identifier of the sample object and / or the attribute information of the sample object. The object identifier of the sample object may be pre-assigned to each sample object, and each sample object may have a different object identifier, so that the object identifier can serve as a unique identifier for the sample object. The attribute information of the sample object may be used to indicate the user's interactive operations on the sample object, and / or the attributes of the sample object itself. For example, taking a video as the sample object, the attribute information used to indicate the user's interactive operations on the sample object may include the number of historical video clicks, video viewing time, video playback progress, historical number of likes, historical number of comments, etc. The attribute information used to indicate the attributes of the sample object itself may include video length, video publisher, video type (such as suspense video, popular science video, funny video, etc.), etc.

[0114] The sample scenario feature is used to indicate the scenario in which the sample user requested the recommendation. For example, the sample scenario feature may include context information when the user requested the recommendation, such as time.

[0115] S1012: For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

[0116] In a possible implementation, for any sample object, multiple sample data features corresponding to the sample object may be concatenated into a vector, and the concatenated vector is used as sample feature information (or input feature) of the sample object.

[0117] Taking the multiple sample data features corresponding to the sample object including sample user features, sample object features and sample scene features as an example, when obtaining the sample feature information of the sample object, the sample user features, sample object features and sample scene features can be spliced ​​into a vector, and the spliced ​​vector can be used as the sample feature information of the sample object.

[0118] After obtaining the sample feature information of the sample object, the obtained sample feature information can be processed through the initial recommendation model.

[0119] Among them, the initial recommendation model can be an MMoE model with the target task network added, that is, the initial recommendation model can include multiple initial feature extraction networks (feature extraction networks can also be called expert networks), multiple initial gating networks, multiple initial task networks and a target task network, each initial task network corresponds to an initial gating network, so that the number of initial gating networks and initial task networks included in the initial recommendation model is consistent.

[0120] See also Figure 2 , Figure 2is a structural diagram of an initial recommendation model according to an exemplary embodiment of the present disclosure. Figure 2 As shown in the figure, the initial recommendation model can include three initial feature extraction networks (expert network 1, expert network 2 and expert network 3), two initial gating networks (gating network A and gating network B), two initial task networks (task network A and task network B) and one target task network (or self-attached task network).

[0121] It should be noted that the above Figure 2 The structure of the initial recommendation model shown is only an example of the structure of the initial recommendation model and does not constitute a limitation on the structure of the initial recommendation model. That is, in more possible implementations, the initial recommendation model may include more or fewer initial feature extraction networks, initial gating networks, and initial task networks, and it is only necessary to ensure that the number of initial gating networks and initial task networks is consistent.

[0122] Each initial feature extraction network may include one or more hidden layers. Therefore, for S102, when the sample feature information corresponding to each sample object is input into the multiple initial feature extraction networks of the initial recommendation model respectively, and multiple sample first features output by each initial feature extraction network are obtained, it can be achieved by the following method:

[0123] For any sample object, the sample feature information corresponding to the sample object is input into multiple initial feature extraction networks respectively, and the sample first features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks to obtain multiple sample first features. The number of sample first features output is consistent with the number of initial feature extraction networks. For example, if there are k initial feature extraction networks in the initial recommendation model, k sample first features can be obtained, and the sample first features can be recorded as e i , i=1,2,3,…,k, where k is any positive integer value.

[0124] by Figure 2 As an example, the initial recommendation model shown in Figure 2 The initial recommendation model shown in FIG includes three initial feature extraction networks. Figure 2 The initial recommendation model shown can obtain the first features of the three samples.

[0125] It should be noted that after obtaining the first feature of the sample, the recommendation score can be predicted based on the first feature of the sample.

[0126] In some embodiments, for S103, when determining the first recommendation score of each sample object under each prediction target based on the multiple sample first features output by the multiple initial feature extraction networks and the target task network of the initial recommendation model, the following steps may be included:

[0127] S1031-1. For any sample object, determine the mean feature of multiple sample first features corresponding to the sample object.

[0128] Among them, each sample first feature can be represented in the form of a vector. In a possible implementation method, the mean vector of multiple sample first features corresponding to the object can be determined, and the determined mean vector can be used as the mean feature corresponding to the sample object.

[0129] For example, the mean feature corresponding to the sample object can be determined by the following formula (1):

[0130]

[0131] Among them, f AVG represents the mean feature, k represents the total number of the first features of the sample (that is, the number of initial feature extraction networks), e i Represents the first feature of each sample.

[0132] S1032-1. Input the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0133] In one possible implementation, the mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0134] It should be noted that the number of predicted targets is consistent with the number of initial task networks. Figure 2 As an example, the initial recommendation model shown in FIG. 1 includes two initial task networks, task network A and task network B. Figure 2 The target task network in the initial recommendation model shown can obtain the first recommendation scores of the sample object under two prediction targets (that is, prediction target A and prediction target B), thereby obtaining two first recommendation scores.

[0135] Among them, the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold, and the target task network includes fewer network parameters, thereby reducing the impact of competition between network parameters corresponding to multiple prediction targets during training, so that the shared structure in the initial recommendation model (that is, the initial feature extraction network) can better learn more effective features.

[0136] The above only describes the process of determining the first recommendation score. However, in S103, based on the first features of the multiple samples and using the multiple initial task networks of the initial recommendation model, determining the second recommendation score of each sample object under each prediction target may include the following steps:

[0137] S1031-2. For any sample object, perform weighted summation on multiple sample first features through multiple initial gating networks of the initial recommendation model to obtain the sample second features corresponding to each initial task network.

[0138] Each initial gating network may include one or more hidden layers. Therefore, S1031-2 may be implemented as follows:

[0139] For any sample object, the sample feature information of the sample object is respectively input into multiple initial gating networks of the initial recommendation model, and the sample feature information is processed by the hidden layer included in the initial gating network to obtain multiple weighted features corresponding to the sample feature information. Then, based on the multiple weighted features, the multiple sample first features are weighted and summed to obtain the sample second features corresponding to each initial task network.

[0140] The number of weighted features is consistent with the number of initial gating networks. If the initial recommendation model includes T initial gating networks, then T weighted features can be obtained through these T initial gating networks. The obtained weighted features can be recorded as g j , j=1,2,3,…,T.

[0141] Each weighted feature can be a multidimensional vector, and the dimension of the weighted feature is consistent with the number of initial feature extraction networks. For example, if there are k initial feature extraction networks in the initial recommendation model, the weighted feature can be a k-dimensional vector, denoted as g j =[g j1 ,g j2 ,…,g jk ].

[0142] Taking the process of weighted summation of multiple sample first features through any initial gating network as an example, the weighted features corresponding to the initial gating network can be multiplied by the multiple sample first features respectively, and the multiple multiplication results can be summed to obtain the sample second features processed by the initial gating network, which can be used as the sample second features to be input into the initial task network corresponding to the initial gating network.

[0143] For example, when performing weighted summation of the first features of multiple samples through the j-th initial gating network, it can be achieved by the following formula (2):

[0144]

[0145] Among them, fj represents the second feature of the sample obtained after processing by the jth initial gating network, g ji represents the i-th dimension element of the j-th weighted feature, e i Represents the first feature of the i-th sample.

[0146] S1032-2. Input the second feature of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target.

[0147] The number of initial task networks matches the number of prediction targets, and one initial task network corresponds to one prediction target. Each initial task network can include one or more hidden layers.

[0148] In one possible implementation, the second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

[0149] It should be noted that after obtaining the first recommendation score and the second recommendation score, the subsequent model training process can be continued.

[0150] In some embodiments, for S104, when training the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object, the following steps may be included:

[0151] S1041 : Determine a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target.

[0152] In a possible implementation, a weighted sum of the first recommendation score and the second recommendation score of each sample object under each prediction target may be performed to obtain a prediction recommendation score corresponding to each sample object.

[0153] S1042: Train an initial recommendation model based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score.

[0154] In one possible implementation, the parameters of the initial recommendation model may be adjusted based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score of each sample object until a training cutoff condition is met, thereby obtaining a trained initial recommendation model.

[0155] It should be noted that the Adam optimizer can be used to adjust the parameters of the initial recommendation model based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score of each sample object. Optionally, other types of optimizers can also be used, which is not limited in this disclosure.

[0156] The loss function may be a result obtained by weighted summing a first loss function indicating the difference between the first recommendation score of the sample object under each prediction target and the sample recommendation score, and a second loss function indicating the difference between the second recommendation score of the sample object under each prediction target and the sample recommendation score.

[0157] Among them, the types of the first loss function and the second loss function can be determined based on the type of prediction target. For example, when the prediction target is whether the reading time is greater than the set time threshold, whether the reading percentage is greater than the set percentage, whether the reading task is completed, or whether there is an interactive behavior task, cross entropy can be used as the first loss function and the second loss function. When the prediction target is to predict the specific value of the reading time, square loss can be used as the first loss function and the second loss function.

[0158] It should be noted that S102 to S104 can be an iterative execution process, that is, the sample feature information corresponding to the first sample object can be input into the initial recommendation model to obtain the predicted recommendation score corresponding to the first sample object, and then the parameters of the initial recommendation model can be adjusted based on the loss function indicating the difference between the sample recommendation score and the predicted recommendation score of the first sample object to obtain the recommendation model after the first parameter adjustment; then the sample feature information corresponding to the second sample object can be input into the recommendation model after the first parameter adjustment to obtain the predicted recommendation score corresponding to the second sample object, and then the parameters of the recommendation model after the first parameter adjustment can be continued to be adjusted based on the loss function indicating the difference between the sample recommendation score and the predicted recommendation score of the second sample object to obtain the recommendation model after the second parameter adjustment, and so on, until the training cutoff condition is met, and the trained initial recommendation model can be obtained.

[0159] Among them, the training cutoff condition can be that the number of iterations reaches a set number, the loss function sets a condition, etc. The present disclosure does not limit which specific condition is used as the training cutoff condition.

[0160] In some embodiments, after the training of the initial recommendation model is completed, a target recommendation model for object recommendation can be obtained based on the trained initial recommendation model.

[0161] In one possible implementation, a target recommendation model can be determined based on a trained initial recommendation model. The target recommendation model includes multiple feature extraction networks, multiple gating networks, and multiple task networks. In other words, the target recommendation model can be composed of multiple feature extraction networks, multiple gating networks, and multiple task networks in the trained initial recommendation model.

[0162] Among them, the structure of the target recommendation model can be found in Figure 3 , Figure 3 is a structural diagram of a target recommendation model according to an exemplary embodiment of the present disclosure. Figure 3 As shown in the figure, the target recommendation model can include 3 feature extraction networks (expert network 1, expert network 2 and expert network 3), 2 gating networks (gating network A and gating network B) and 2 task networks (task network A and task network B).

[0163] It should be noted that after obtaining the target recommendation model, the target recommendation model can be used to implement the online object recommendation process.

[0164] When the computing device is a server, after the server has trained and obtained a target recommendation model, it can directly use the target recommendation model to implement online recommendations. Optionally, the server can also deploy the target recommendation model to a terminal device, which can then implement online recommendations using the target recommendation model. This disclosure does not limit whether online recommendations are implemented by the server or the terminal device.

[0165] In some embodiments, the process of implementing the online object recommendation process through the target recommendation model can be referred to Figure 4 , Figure 4 FIG. 1 is a flowchart of an object recommendation process according to an exemplary embodiment of the present disclosure. The process may include the following steps:

[0166] S401: Acquire feature information of candidate objects.

[0167] It should be noted that, for the introduction of candidate objects and feature information, please refer to S101 and the embodiments corresponding to S101, which will not be repeated here.

[0168] S402: Input the feature information into the target recommendation model, and process the feature information through multiple feature extraction networks of the target recommendation model to obtain multiple first features.

[0169] It should be noted that, for the introduction of S402, reference can be made to S102 and the corresponding embodiments of S102, which will not be repeated here.

[0170] S403: Input the multiple first features into each gated network to obtain a second feature corresponding to each gated network.

[0171] It should be noted that for the introduction of S403, please refer to S1031-2, which will not be repeated here.

[0172] S404: Input each second feature into the task network corresponding to the gating network, and output the recommendation score of the candidate object under the corresponding prediction target through the task network, where one task network corresponds to one prediction target.

[0173] It should be noted that for the introduction of S404, please refer to S1032-2, which will not be repeated here.

[0174] S405: Determine the target object based on the recommendation scores of the candidate objects under each prediction target.

[0175] In one possible implementation, the recommendation scores of the candidate objects under various prediction targets may be weighted and summed to obtain a final recommendation score, thereby determining the candidate objects whose final recommendation scores are greater than a set score threshold as the target objects to be recommended.

[0176] The set score threshold may be any integer value, and the present disclosure does not limit the specific value of the set score threshold.

[0177] In some embodiments, after the target objects to be recommended are determined, the target objects may be sorted according to a set order.

[0178] For example, the target objects can be sorted in descending order of recommendation scores, so that the target objects can be displayed based on the sorting results, so that users can preferentially see target objects with higher recommendation scores, thereby improving recommendation accuracy and thus improving user experience.

[0179] After introducing the training method of the recommendation model according to the exemplary embodiment of the present disclosure, the structure of the training apparatus for the recommendation model according to the exemplary embodiment of the present disclosure and the computing device for implementing the training method of the recommendation model will be described next.

[0180] See also Figure 5 , Figure 51 is a block diagram of a training device for a recommendation model according to an exemplary embodiment of the present disclosure, the device comprising:

[0181] An acquisition module 501 is configured to acquire sample feature information of multiple sample objects and a sample recommendation score corresponding to each sample object;

[0182] Processing module 502, configured to input sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model, and obtain multiple sample first features output by each initial feature extraction network;

[0183] Determination module 503 is configured to determine, based on the multiple sample first features output by the multiple initial feature extraction networks, a first recommendation score for each sample object under each prediction target through a target task network of the initial recommendation model, and to determine, based on the multiple sample first features, a second recommendation score for each sample object under each prediction target through the multiple initial task networks of the initial recommendation model; wherein the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold;

[0184] The training module 504 is configured to train an initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object.

[0185] In one embodiment of the present disclosure, the determination module 503, when used to determine the first recommendation score of each sample object under each prediction target based on the multiple sample first features output by the multiple initial feature extraction networks and through the target task network of the initial recommendation model, includes a determination unit and a processing unit;

[0186] a determining unit, configured to determine, for any sample object, a mean feature of a plurality of first features of samples corresponding to the sample object;

[0187] The processing unit is used to input the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0188] In one embodiment of the present disclosure, the processing unit, when inputting the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target, is configured to:

[0189] The mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

[0190] In one embodiment of the present disclosure, each initial task network corresponds to an initial gating network;

[0191] The determination module 503 is configured to determine, based on the plurality of first features of the samples, a second recommendation score for each sample object under each prediction target through the plurality of initial task networks of the initial recommendation model, and includes a first processing unit and a second processing unit;

[0192] The first processing unit is configured to perform weighted summation of the first features of multiple samples for any sample object through multiple initial gating networks of the initial recommendation model to obtain the second features of the samples corresponding to each initial task network;

[0193] The second processing unit is used to input the second feature of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target;

[0194] Among them, the number of initial task networks matches the number of prediction targets.

[0195] In one embodiment of the present disclosure, the second processing unit, when inputting the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target, is used to:

[0196] The second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

[0197] In one embodiment of the present disclosure, the acquisition module 501, when used to acquire sample feature information corresponding to a plurality of sample objects, is used to:

[0198] Based on the sample data corresponding to the multiple sample objects, obtaining multiple sample data features;

[0199] For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

[0200] In one embodiment of the present disclosure, the sample data feature includes at least one of a sample user feature, a sample object feature, and a sample scene feature;

[0201] Among them, the sample user characteristics are used to indicate the user identification of the sample user and / or the user behavior information of the sample user, the sample object characteristics are used to indicate the object identification of the sample object to be recommended and / or the attribute information of the sample object, and the sample scene characteristics are used to indicate the scene in which the sample user requests recommendation.

[0202] In one embodiment of the present disclosure, the processing module 502, when inputting the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model and obtaining the sample first features output by each initial feature extraction network, is configured to:

[0203] For any sample object, the sample feature information corresponding to the sample object is input into multiple initial feature extraction networks respectively, and the first sample features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks.

[0204] In one embodiment of the present disclosure, the training module 504, when used to train the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object, is used to:

[0205] Determining a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target;

[0206] An initial recommendation model is trained based on a loss function indicating the difference between the sample recommendation score and the predicted recommendation score.

[0207] In one embodiment of the present disclosure, the determination module 503 is further configured to determine a target recommendation model based on the trained initial recommendation model, where the target recommendation model includes multiple feature extraction networks, multiple gating networks, and multiple task networks.

[0208] In one embodiment of the present disclosure, the acquisition module 501 is further configured to acquire feature information of a candidate object;

[0209] The processing module 502 is further configured to input the feature information into the target recommendation model, and process the feature information through multiple feature extraction networks of the target recommendation model to obtain multiple first features;

[0210] The processing module 502 is further configured to input the plurality of first features into each gating network to obtain a second feature corresponding to each gating network;

[0211] The processing module 502 is further configured to input each second feature into a task network corresponding to the gating network, and output a recommendation score of the candidate object under the corresponding prediction target through the task network, wherein one task network corresponds to one prediction target;

[0212] The determination module 503 is further configured to determine a target object based on the recommendation scores of the candidate objects under various prediction targets.

[0213] In one embodiment of the present disclosure, the apparatus further comprises:

[0214] The sorting module is used to sort the target objects according to the set order.

[0215] It should be noted that although several modules or units of the training device for the recommendation model are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided to be embodied by multiple modules or units.

[0216] The embodiment of the present disclosure also provides a computer-readable storage medium. Figure 6 is a schematic diagram of a computer-readable storage medium according to an exemplary embodiment of the present disclosure, such as Figure 6 As shown, the storage medium stores a computer program 601, which, when executed by a processor, can execute the training method of the recommendation model provided by any embodiment of the present disclosure.

[0217] The present disclosure also provides a computing device, which may include a memory and a processor. The memory is used to store computer instructions that can be executed on the processor, and the processor is used to implement the training method of the recommendation model provided by any embodiment of the present disclosure when executing the computer instructions. Figure 7 , Figure 7 7 is a structural diagram of a computing device according to an exemplary embodiment of the present disclosure. The computing device 700 may include but is not limited to: a processor 710, a memory 720, and a bus 730 connecting different system components (including the memory 720 and the processor 710).

[0218] The memory 720 stores computer instructions that can be executed by the processor 710, so that the processor 710 can execute the training method of the recommendation model provided by any embodiment of the present disclosure. The memory 720 may include a random access memory unit RAM 721, a cache memory unit 722 and / or a read-only memory unit ROM 723. The memory 720 may also include: a program tool 725 having a set of program modules 724, the program modules 724 including but not limited to: an operating system, one or more application programs, other program modules and program data, and one or more combinations of these program modules may include the implementation of a network environment.

[0219] The bus 730 may include, for example, a data bus, an address bus, and a control bus. The computing device 700 may also communicate with an external device 750 via an I / O interface 740. The external device 750 may be, for example, a keyboard, a Bluetooth device, etc. The computing device 700 may also communicate with one or more networks via a network adapter 760. For example, the network may be a local area network, a wide area network, a public network, etc. Figure 7 As shown, the network adapter 760 can also communicate with other modules of the computing device 700 via the bus 730 .

[0220] The embodiments of the present disclosure further provide a computer program product, which includes a computer program. When the program is executed by the processor 710 of the computing device 700, it can implement the training method of the recommendation model provided by any embodiment of the present disclosure.

[0221] Furthermore, although the operations of the disclosed method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0222] Although the spirit and principles of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features in these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the appended claims.

Claims

1. A training method for a recommendation model, characterized in that: The method comprises: Obtaining sample feature information of multiple sample objects and a sample recommendation score corresponding to each sample object; wherein the sample objects include at least multimedia data and commodities, and the sample feature information is determined based at least on sample users' operation data on the sample objects, wherein the operation data includes at least click operations, like operations, comment operations, and the time when each operation occurs; Inputting the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model respectively, and obtaining multiple sample first features output by each initial feature extraction network; Based on the multiple sample first features output by the multiple initial feature extraction networks, a first recommendation score for each sample object under each prediction target is determined through the target task network of the initial recommendation model, and based on the multiple sample first features, a second recommendation score for each sample object under each prediction target is determined through the multiple initial task networks of the initial recommendation model; wherein the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold; the target task network is used to learn shared knowledge between different prediction targets; The initial recommendation model is trained based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object.

2. The method according to claim 1, characterized in that The determining, based on the multiple sample first features output by the multiple initial feature extraction networks, of a first recommendation score for each sample object under each prediction target through the target task network of the initial recommendation model includes: For any sample object, determining the mean feature of the first features of multiple samples corresponding to the sample object; The mean feature is input into the target task network to obtain the first recommendation score of the sample object under each prediction target.

3. The method according to claim 2, characterized in that Inputting the mean feature into the target task network to obtain a first recommendation score of the sample object under each prediction target includes: The mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

4. The method according to claim 1, wherein Each initial task network corresponds to an initial gating network; The determining, based on the multiple sample first features and using the multiple initial task networks of the initial recommendation model, a second recommendation score for each sample object under each prediction target includes: For any sample object, a weighted sum of the first features of the multiple samples is performed through the multiple initial gating networks of the initial recommendation model to obtain the second features of the samples corresponding to each initial task network; Input the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target; The number of the initial task networks matches the number of the predicted targets.

5. The method according to claim 4, characterized in that Inputting the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target includes: The second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

6. The method according to claim 1, characterized in that The obtaining of sample feature information corresponding to the plurality of sample objects includes: Acquire multiple sample data features based on the sample data corresponding to the multiple sample objects; For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

7. The method according to claim 6, characterized in that The sample data feature includes at least one of a sample user feature, a sample object feature, and a sample scene feature; Among them, the sample user characteristics are used to indicate the user identification of the sample user and / or the user behavior information of the sample user, the sample object characteristics are used to indicate the object identification of the sample object to be recommended and / or the attribute information of the sample object, and the sample scene characteristics are used to indicate the scene in which the sample user requests recommendation.

8. The method according to claim 1, characterized in that The step of inputting the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model to obtain the first features of the samples output by each initial feature extraction network includes: For any sample object, the sample feature information corresponding to the sample object is input into the multiple initial feature extraction networks respectively, and the first sample features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks.

9. The method according to claim 1, characterized in that The training of the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object includes: Determining a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target; The initial recommendation model is trained based on a loss function indicating a difference between the sample recommendation score and the predicted recommendation score.

10. The method according to claim 1, characterized in that After training the initial recommendation model based on the sample recommendation scores, the first recommendation scores, and the second recommendation scores corresponding to the respective sample objects, the method further includes: Based on the trained initial recommendation model, a target recommendation model is determined, where the target recommendation model includes multiple feature extraction networks, multiple gating networks, and multiple task networks.

11. The method according to claim 10, characterized in that After determining the target recommendation model based on the trained initial recommendation model, the method further includes: Obtaining feature information of candidate objects; Inputting the feature information into the target recommendation model, and processing the feature information through multiple feature extraction networks of the target recommendation model to obtain multiple first features; Inputting the plurality of first features into each gating network to obtain a second feature corresponding to each gating network; Each second feature is input into the task network corresponding to the gating network, and the task network outputs the recommendation score of the candidate object under the corresponding prediction target, where one task network corresponds to one prediction target; The target object is determined based on the recommendation scores of the candidate objects under various prediction targets.

12. The method according to claim 11, characterized in that After determining the target object based on the recommendation scores of the candidate objects under each prediction target, the method further includes: Sort the target objects according to a set order.

13. A training device for a recommendation model, characterized in that: The device comprises: An acquisition module, configured to acquire sample feature information of a plurality of sample objects and a sample recommendation score corresponding to each sample object; wherein the sample objects include at least multimedia data and commodities, and the sample feature information is determined based at least on sample users' operation data on the sample objects, wherein the operation data includes at least click operations, like operations, comment operations, and the time when each operation occurs; A processing module, configured to input the sample feature information corresponding to each sample object into multiple initial feature extraction networks of the initial recommendation model, and obtain multiple sample first features output by each initial feature extraction network; a determination module configured to determine, based on the multiple sample first features output by the multiple initial feature extraction networks, a first recommendation score for each sample object under each prediction target through the target task network of the initial recommendation model, and to determine, based on the multiple sample first features, a second recommendation score for each sample object under each prediction target through the multiple initial task networks of the initial recommendation model; wherein the number of hidden layers included in the target task network does not exceed a first set threshold, and / or the number of neurons included in each hidden layer in the target task network does not exceed a second set threshold; and the target task network is configured to learn shared knowledge between different prediction targets; A training module is configured to train the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object.

14. The device according to claim 13, characterized in that The determination module, when used to determine the first recommendation score of each sample object under each prediction target based on the multiple sample first features output by the multiple initial feature extraction networks and through the target task network of the initial recommendation model, includes a determination unit and a processing unit; The determining unit is configured to determine, for any sample object, a mean feature of a plurality of sample first features corresponding to the sample object; The processing unit is configured to input the mean feature into the target task network to obtain a first recommendation score for the sample object under each prediction target.

15. The device according to claim 14, characterized in that The processing unit, when used to input the mean feature into the target task network to obtain the first recommendation score of the sample object under each prediction target, is used to: The mean feature is input into the target task network, and the mean feature is processed by the hidden layer included in the target task network to obtain the first recommendation score of the sample object under each prediction target.

16. The device according to claim 13, characterized in that Each initial task network corresponds to an initial gating network; The determining module, when used to determine the second recommendation score of each sample object under each prediction target based on the multiple sample first features through the multiple initial task networks of the initial recommendation model, includes a first processing unit and a second processing unit; The first processing unit is configured to perform weighted summation of the plurality of sample first features for any sample object through the plurality of initial gating networks of the initial recommendation model to obtain the sample second features corresponding to the respective initial task networks; The second processing unit is configured to input the second feature of each sample into the corresponding initial task network to obtain a second recommendation score for the sample object under each prediction target; The number of the initial task networks matches the number of the predicted targets.

17. The device according to claim 16, characterized in that The second processing unit, when used to input the second features of each sample into the corresponding initial task network to obtain the second recommendation score of the sample object under each prediction target, is used to: The second features of each sample are respectively input into the corresponding initial task network, and the corresponding second features of the sample are processed through the hidden layer included in each initial task network to obtain the second recommendation score of the sample object under each prediction target.

18. The device according to claim 13, characterized in that The acquisition module, when used to acquire sample feature information corresponding to a plurality of sample objects, is used to: Acquire multiple sample data features based on the sample data corresponding to the multiple sample objects; For any sample object, multiple sample data features corresponding to the sample object are spliced ​​to obtain sample feature information of the sample object.

19. The device according to claim 18, characterized in that The sample data feature includes at least one of a sample user feature, a sample object feature, and a sample scene feature; Among them, the sample user characteristics are used to indicate the user identification of the sample user and / or the user behavior information of the sample user, the sample object characteristics are used to indicate the object identification of the sample object to be recommended and / or the attribute information of the sample object, and the sample scene characteristics are used to indicate the scene in which the sample user requests recommendation.

20. The device according to claim 13, wherein The processing module is configured to input the sample feature information corresponding to each sample object into the multiple initial feature extraction networks of the initial recommendation model to obtain the first sample features output by each initial feature extraction network, and is configured to: For any sample object, the sample feature information corresponding to the sample object is input into the multiple initial feature extraction networks respectively, and the first sample features corresponding to each initial feature extraction network are output through the hidden layers included in the multiple initial feature extraction networks.

21. The device according to claim 13, characterized in that The training module is configured to train the initial recommendation model based on the sample recommendation score, the first recommendation score, and the second recommendation score corresponding to each sample object, including: Determining a predicted recommendation score corresponding to each sample object based on the first recommendation score of each sample object under each prediction target and the second recommendation score of each sample object under each prediction target; The initial recommendation model is trained based on a loss function indicating a difference between the sample recommendation score and the predicted recommendation score.

22. The device according to claim 13, characterized in that The determination module is further used to determine a target recommendation model based on the trained initial recommendation model, where the target recommendation model includes multiple feature extraction networks, multiple gating networks, and multiple task networks.

23. The device according to claim 22, characterized in that The acquisition module is further used to obtain feature information of the candidate object; The processing module is further configured to input the feature information into the target recommendation model, and process the feature information through multiple feature extraction networks of the target recommendation model to obtain multiple first features; The processing module is further configured to input the plurality of first features into each gating network to obtain a second feature corresponding to each gating network; The processing module is further configured to input each second feature into a task network corresponding to the gating network, and output a recommendation score of the candidate object under the corresponding prediction target through the task network, wherein one task network corresponds to one prediction target; The determination module is further configured to determine a target object based on the recommendation scores of the candidate objects under various prediction targets.

24. The device according to claim 23, characterized in that The device further comprises: The sorting module is used to sort the target objects according to a set order.

25. A computing device, characterized in that The computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the operations performed by the training method for the recommendation model according to any one of claims 1 to 12 are implemented.

26. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, and the processor executes the operations performed by the training method for the recommendation model according to any one of claims 1 to 12.

Images