Training method, information push method and device for multi-business data prediction model

Abstract

The present disclosure is about a training method, information push method and device for a multi-business data prediction model, the method comprising: obtaining sample features of sample multimedia information and at least two sample business data; performing dimension increase processing on the screened sample features to obtain sample newly added dimension features; training a preset machine learning model according to the sample newly added dimension features and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; training a multi-task model according to the preset sample features, at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are sample features or sample newly added dimension features. The present disclosure improves the accuracy of the prediction results of the multi-business data prediction model and improves the conversion rate of the pushed information obtained according to the multi-business data prediction model.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a training method, an information push method and a device for a multi-business data prediction model. Background Art

[0002] In the relevant technology, feature crossover plays a vital role in the model training of the short video push system. It can reflect the interactive relationship between modeling features, better describe the interests of users and the characteristics of items, and thus achieve a more accurate push effect. In the short video scene, the interests of users and the popularity of videos change very quickly. Traditional push models are difficult to achieve personalized descriptions of each user or video, and the push accuracy is low.

[0003] Therefore, it is necessary to provide a training method, information push method and equipment for a multi-business data prediction model. During the model training process, a personalized model is constructed for each object and multimedia information, which solves the problem that feature cross-cutting in multimedia information push cannot achieve personalization, thereby improving the accuracy of the prediction results of the business data prediction model. Summary of the invention

[0004] The present disclosure provides a training method, information push method and device for a multi-service data prediction model, so as to at least solve the problem of low accuracy of prediction results of service data prediction models in related technologies. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, a training method for a multi-service data prediction model is provided, comprising:

[0006] Acquire sample features of sample multimedia information and at least two types of sample business data; the sample features include sample object features and sample multimedia information association features;

[0007] Performing dimension-increasing processing on the screening sample features to obtain new dimension features of the samples; the screening sample features are part or all of the sample features;

[0008] The preset machine learning model is trained according to the newly added dimension features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model;

[0009] A multi-task model is trained according to preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the sample newly added dimensional features.

[0010] In an exemplary embodiment, the multi-task model includes a gating network and at least two expert networks; the multi-task model is trained according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a multi-service data prediction model, including:

[0011] Training each expert network according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function for each expert network;

[0012] The gating network is trained according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data;

[0013] Determining model parameters of the multi-task model according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data;

[0014] Based on the multi-task model and the model parameters, the multi-service data prediction model is determined.

[0015] In an exemplary embodiment, determining the model parameters of the multi-task model according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data includes:

[0016] Calculate the product of the target cross-feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data to obtain the initial weight feature determination function of each expert network corresponding to each sample business data;

[0017] Calculating the sum of the initial weight feature determination functions of each expert network corresponding to each sample business data to obtain the target weight cross feature determination function corresponding to each sample business data;

[0018] The model parameters of the multi-task model are determined according to the target weight cross-feature determination function corresponding to each sample business data.

[0019] In an exemplary embodiment, the preset sample features are at least two, and the training of each expert network according to the preset sample features and the personalized parameter prediction function to obtain the target cross-feature determination function of each expert network includes:

[0020] Performing splicing processing on the at least two preset sample features to obtain a sample cross feature;

[0021] A preset number of cross-trainings are performed on each expert network according to the sample cross-features to obtain a target cross-feature determination function for each expert network.

[0022] In an exemplary embodiment, performing a preset number of cross-training on each expert network according to the sample cross-features to obtain a target cross-feature determination function for each expert network includes:

[0023] Performing a preset number of cross-training on each expert network according to the sample cross-features to obtain an initial cross-feature determination function for each expert network;

[0024] The initial cross-features of each expert network are subjected to low-rank decomposition processing, and the initial cross-feature determination function of each expert network is updated to the target cross-feature determination function.

[0025] In an exemplary embodiment, each expert network includes at least two cross layers, and training each expert network according to the preset sample features and the personalized parameter prediction function to obtain the target cross feature determination function of each expert network includes:

[0026] Determine a preset cross layer corresponding to the personalized parameter prediction function in each expert network;

[0027] At the preset cross layer of each expert network, each expert network is trained according to the preset sample features and the personalized parameter prediction function to obtain a target cross feature determination function of each expert network.

[0028] In an exemplary embodiment, the training of the gating network according to the preset sample features, the at least two sample service data, and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data includes:

[0029] The gated network is trained according to the preset sample features and the at least two sample business data to obtain an initial weight determination function of each expert network corresponding to each sample business data;

[0030] Based on the personalized parameter prediction function, the initial weight determination function of each expert network corresponding to each sample business data is updated to obtain the weight determination function of each expert network corresponding to each sample business data.

[0031] In an exemplary embodiment, the step of performing dimension increasing processing on the features of the screened samples to obtain features of the sample with newly added dimensions includes:

[0032] Obtaining dimension information of the preset machine learning model;

[0033] Based on the dimension information, the dimension of the screening sample feature is updated to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature;

[0034] Based on the sample update feature, the sample newly added dimension feature is determined; the preset sample feature is the sample feature, the sample update feature or the sample newly added dimension feature.

[0035] In an exemplary implementation, updating the dimension of the screening sample feature based on the dimension information to obtain the sample update feature includes:

[0036] Based on the dimensional information, determine the number of dimensions of the newly added dimensional features of the sample;

[0037] Based on the number of dimensions, the dimensions of the sample object features and / or the sample multimedia information associated features are updated to obtain the sample updated features.

[0038] According to a second aspect of an embodiment of the present disclosure, a multimedia information push method is provided, including:

[0039] Obtaining target features and target newly added dimension features of target multimedia information; the target features include target object features and target multimedia information association features;

[0040] Inputting the target newly added dimension features into a personalized parameter prediction model to obtain target personalized parameters;

[0041] Inputting the preset target feature and the target personalized parameter into a multi-service data prediction model to obtain at least two types of target service data; the preset target feature is the target feature or the target newly added dimension feature;

[0042] The target multimedia information is pushed to the target object according to the at least two target service data.

[0043] In an exemplary implementation, the pushing the target multimedia information to the target object according to the at least two target service data includes:

[0044] Determine the weight of each target business data;

[0045] Determining the comprehensive service data of the target multimedia information according to each target service data and the weight of each target service data;

[0046] If the comprehensive service data is greater than a preset threshold, the target multimedia information is pushed to the target object.

[0047] In an exemplary implementation, the number of the target multimedia information is at least two, and the step of determining the integrated service data of the target multimedia information includes:

[0048] Determine comprehensive business data for each target multimedia information;

[0049] The pushing the target multimedia information to the target object comprises:

[0050] sorting the at least two target multimedia information based on the comprehensive service data of each target multimedia information;

[0051] Determine the multimedia information to be pushed according to the sorting result;

[0052] Pushing the multimedia information to be pushed to the target object.

[0053] In an exemplary embodiment, the step of obtaining target features and target newly added dimension features of target multimedia information includes:

[0054] Acquiring target features of the target multimedia information;

[0055] Based on the target feature, determining a screening target feature; the screening sample feature is part or all of the target feature;

[0056] The screening target features are subjected to dimensionality increase processing to obtain target newly added dimensional features.

[0057] According to a third aspect of an embodiment of the present disclosure, a training device for a multi-service data prediction model is provided, comprising:

[0058] A sample data acquisition module is configured to acquire sample features of sample multimedia information and at least two types of sample business data; the sample features include sample object features and sample multimedia information association features;

[0059] The sample newly added dimension feature determination module is configured to perform dimension increase processing on the screened sample features to obtain the sample newly added dimension feature; the screened sample features are part or all of the sample features;

[0060] A personalized parameter prediction function determination module is configured to execute a preset machine learning model according to the newly added dimensional features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model;

[0061] The multi-business data prediction model determination module is configured to execute training of a multi-task model based on preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the newly added dimensional features of the samples.

[0062] In an exemplary embodiment, the multi-task model includes a gating network and at least two expert networks; the multi-service data prediction model determination module includes:

[0063] A target cross-feature determination function determination unit is configured to train each expert network according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network;

[0064] A weight determination function determination unit is configured to train the gating network according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data;

[0065] A model parameter determination unit, configured to determine the model parameters of the multi-task model according to a target cross-feature determination function of each expert network and a weight determination function of each expert network corresponding to each sample business data;

[0066] The multi-service data prediction model determination unit is configured to determine the multi-service data prediction model based on the multi-task model and the model parameters.

[0067] In an exemplary embodiment, the model parameter determination unit includes:

[0068] The initial weight feature determination function determination subunit is configured to calculate the product of the target cross feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data, so as to obtain the initial weight feature determination function of each expert network corresponding to each sample business data;

[0069] The target weight cross-feature determination function determination subunit is configured to calculate the sum of the initial weight feature determination functions of each expert network corresponding to each sample business data to obtain the target weight cross-feature determination function corresponding to each sample business data;

[0070] The model parameter determination subunit is configured to execute a target weight cross-feature determination function corresponding to each sample business data to determine the model parameters of the multi-task model.

[0071] In an exemplary embodiment, the number of the preset sample features is at least two, and the target cross-feature determination function determination unit includes:

[0072] A sample cross feature determination subunit is configured to perform a splicing process on the at least two preset sample features to obtain a sample cross feature;

[0073] The first function determination subunit is configured to perform a preset number of cross-training on each expert network according to the sample cross-features to obtain a target cross-feature determination function for each expert network.

[0074] In an exemplary embodiment, the first function determination subunit includes:

[0075] an initial cross-feature determination function determination subunit, configured to perform a preset number of cross-training operations on each expert network according to the sample cross-features, to obtain an initial cross-feature determination function of each expert network;

[0076] The function updating subunit is configured to perform low-rank decomposition processing on the initial cross-features of each expert network, and update the initial cross-feature determination function of each expert network to the target cross-feature determination function.

[0077] In an exemplary embodiment, each expert network includes at least two cross layers, and the target cross feature determination function determination unit includes:

[0078] A preset cross layer determination subunit is configured to determine the preset cross layer corresponding to the personalized parameter prediction function in each expert network;

[0079] The second function determination subunit is configured to execute at a preset cross layer of each expert network, train each expert network according to the preset sample features and the personalized parameter prediction function, and obtain a target cross feature determination function of each expert network.

[0080] In an exemplary embodiment, the weight determination function determination unit includes:

[0081] An initial weight determination function determination subunit is configured to train the gated network according to the preset sample features and the at least two sample business data to obtain an initial weight determination function of each expert network corresponding to each sample business data;

[0082] The weight determination function determination subunit is configured to execute, based on the personalized parameter prediction function, an update of the initial weight determination function of each expert network corresponding to each sample business data, to obtain the weight determination function of each expert network corresponding to each sample business data.

[0083] In an exemplary embodiment, the sample newly added dimension feature determination module includes:

[0084] A dimension information determining unit, configured to execute obtaining dimension information of the preset machine learning model;

[0085] A sample update feature determination unit is configured to update the dimension of the screening sample feature based on the dimension information to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature;

[0086] The sample newly added dimension feature determination unit is configured to determine the sample newly added dimension feature based on the sample update feature; the preset sample feature is the sample feature, the sample update feature or the sample newly added dimension feature.

[0087] In an exemplary embodiment, the sample update feature determination unit includes:

[0088] A dimension number determination subunit is configured to determine the number of dimensions of the newly added dimensional features of the sample based on the dimensional information;

[0089] The sample update feature determination subunit is configured to update the dimensions of the sample object feature and / or the sample multimedia information associated feature based on the number of dimensions to obtain the sample update feature.

[0090] According to a fourth aspect of an embodiment of the present disclosure, a multimedia information push device is provided, including:

[0091] A target feature acquisition module is configured to acquire target features and target newly added dimension features of target multimedia information; the target features include target object features and target multimedia information associated features;

[0092] A target personalized parameter determination module is configured to input the target newly added dimension feature into a personalized parameter prediction model to obtain a target personalized parameter;

[0093] The target business data determination module is configured to input the preset target feature and the target personalized parameter into the multi-business data prediction model to obtain at least two types of target business data; the preset target feature is the target feature or the target newly added dimension feature;

[0094] A target multimedia information pushing module, configured to push the target multimedia information to the target object according to the at least two target service data;

[0095] Wherein, the multi-service data prediction model is trained according to the above-mentioned multi-service data prediction model training method.

[0096] In an exemplary embodiment, the target multimedia information push module includes:

[0097] A weight determination unit, configured to determine the weight of each target business data;

[0098] a comprehensive service data determination unit, configured to determine the comprehensive service data of the target multimedia information according to each target service data and the weight of each target service data;

[0099] The target multimedia information pushing unit is configured to push the target multimedia information to the target object if the comprehensive service data is greater than a preset threshold.

[0100] In an exemplary embodiment, the number of target multimedia information is at least two, and the integrated service data determination unit includes:

[0101] The integrated service data determination subunit is configured to determine the integrated service data of each target multimedia information;

[0102] The target multimedia information pushing unit includes:

[0103] An information sorting subunit, configured to sort the at least two target multimedia information based on the comprehensive service data of each target multimedia information;

[0104] The multimedia information to be pushed determining subunit is configured to determine the multimedia information to be pushed according to the sorting result;

[0105] The information pushing subunit is configured to push the multimedia information to be pushed to the target object.

[0106] In an exemplary embodiment, the target feature acquisition module includes:

[0107] A target feature acquisition unit, configured to execute acquisition of target features of the target multimedia information;

[0108] A screening target feature determination unit is configured to determine a screening target feature based on the target feature; the screening sample feature is part or all of the target feature;

[0109] The feature processing unit is configured to perform dimensionality increase processing on the screening target features to obtain target newly added dimensional features.

[0110] According to a fifth aspect of an embodiment of the present disclosure, there is provided an electronic device, including:

[0111] processor;

[0112] a memory for storing instructions executable by the processor;

[0113] Wherein, the processor is configured to execute the instructions to implement the training method of the multi-service data prediction model or the multimedia information pushing method as described above.

[0114] According to the sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device can execute the training method of the multi-business data prediction model or the multimedia information push method as described above.

[0115] According to a seventh aspect of an embodiment of the present disclosure, a computer program product is provided, including a computer program, which, when executed by a processor, implements the training method of the multi-business data prediction model or the multimedia information push method as described above.

[0116] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0117] The present invention obtains sample features of sample multimedia information and at least two types of sample business data; performs dimensionality increase processing on the screened sample features to obtain sample newly added dimensional features; trains a preset machine learning model according to the sample newly added dimensional features and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; trains a multi-task model according to the preset sample features, the at least two types of sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; during the model training process, a personalized parameter prediction model is constructed for each sample multimedia information for feature crossing, which solves the problem that feature crossing in multimedia information push cannot achieve personalization, thereby improving the accuracy of the prediction results of the multi-business data prediction model and improving the conversion rate of the pushed information obtained according to the multi-business data prediction model.

[0118] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0120] Figure 1 It is an application environment diagram of a training method for a multi-business data prediction model according to an exemplary embodiment.

[0121] Figure 2 The present invention is a flowchart of a method for training a multi-service data prediction model according to an exemplary embodiment.

[0122] Figure 3 It is a flowchart of a method for performing dimensionality increase processing on the features of screened samples to obtain features of newly added dimensions of the samples according to an exemplary embodiment.

[0123] Figure 4 The present invention is a flowchart of a method for determining a multi-service data prediction model according to an exemplary embodiment.

[0124] Figure 5 The invention is a flowchart of a method for determining model parameters of a multi-task model according to an exemplary embodiment.

[0125] Figure 6 The figure is a flowchart of a multimedia information pushing method according to an exemplary embodiment.

[0126] Figure 7 FIG. 1 is a flowchart example 1 of a method for pushing target multimedia information to a target object according to an exemplary embodiment.

[0127] Figure 8 This is a second flowchart example of a method for pushing target multimedia information to a target object according to an exemplary embodiment.

[0128] Fig. 9 The present invention is a block diagram of a training device for a multi-service data prediction model according to an exemplary embodiment.

[0129] Fig.10 The figure is a block diagram of a multimedia information pushing device according to an exemplary embodiment.

[0130] Fig.11 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0131] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0132] 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 are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0133] 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 this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0134] See also Figure 1 , which is an application environment diagram of a training method for a multi-business data prediction model according to an exemplary embodiment. The application environment may include a server 01 and a client 02.

[0135] Specifically, in the embodiments of this specification, the server 01 may include an independently operated server, or a distributed server, or a server cluster composed of multiple servers, and may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Server 01 may include a network communication unit, a processor, a memory, and the like. Specifically, the server 01 can be used to obtain sample characteristics of sample multimedia information and at least two sample business data; perform dimensionality increase processing on the screened sample characteristics to obtain sample newly added dimensional characteristics; the screened sample characteristics are part or all of the sample characteristics; according to the sample newly added dimensional characteristics and the corresponding sample personalized parameters, a preset machine learning model is trained to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; according to the preset sample characteristics, the at least two sample business data and the personalized parameter prediction function, a multi-task model is trained to obtain a multi-business data prediction model; and based on the multi-business data prediction model, multi-business data is predicted for the target characteristics and the target newly added dimensional characteristics of the target multimedia information; and according to the prediction results, the target multimedia information is sent to the client 02.

[0136] Specifically, in the embodiments of this specification, the client 02 may include physical devices such as smart phones, desktop computers, tablet computers, laptop computers, digital assistants, smart wearable devices, and vehicle-mounted terminals, and may also include software running in physical devices, such as web pages provided by some service providers to users, or applications provided by these service providers to users. Specifically, the client 02 may receive and display the target multimedia information sent by the server 01.

[0137] Figure 2 is a flowchart of a method for training a multi-service data prediction model according to an exemplary embodiment. Figure 2 As shown, this method can be applied to Figure 1 The server 01 shown includes the following steps.

[0138] In step S201, sample features of sample multimedia information and at least two types of sample business data are obtained; the sample features include sample object features and sample multimedia information association features.

[0139] In some embodiments, multimedia information may include but is not limited to audio, short video, long video and other information; there is at least one sample feature, the sample object is the user corresponding to the sample multimedia information, and the sample objects are usually multiple; the sample object feature is the sample user feature, and the sample user feature may include the sample user ID (UserID) and the device ID (DeviceID) corresponding to the sample user, or other features that can characterize the sample user information, such as user interests, user age, user region and other features. The user feature may also be an average summed user behavior sequence, such as a video ID sequence that has been recently watched for a long time after the average value (sumpooling) processing.

[0140] In some embodiments, the sample multimedia information association features may include content features and context features of the sample multimedia information, wherein the content features may include the identification of the multimedia information, such as a video ID (PhotoID), and may also include features such as the release time of the multimedia information, information category (such as entertainment, food, fitness, parenting, etc.), and title. The context features may include recommendation association features of the sample multimedia information, and the recommendation association features may include, but are not limited to, features such as recommendation time and recommended page.

[0141] In step S203, the dimension-increasing processing is performed on the screening sample features to obtain sample newly added dimension features; the screening sample features are part or all of the sample features.

[0142] In some embodiments, all sample features can be processed with increased dimensions, and some features can be selected from the sample features for increased dimension processing; for example, corresponding dimensions can be added to video features and user features; you can also choose to add dimensions only to the video ID, or to add dimensions to multiple features at the same time; you can also choose to add dimensions only to the user ID, or to add dimensions to multiple features at the same time; you can also add dimensions to features used for feature cross-cutting. Adding corresponding dimensions means adding new dimensions while maintaining the original dimensions unchanged. The newly added dimensions may not participate in the calculation of the multi-business model. The newly added dimensions can be stored in a parameter server (PS) and used as parameters for the personalized model.

[0143] In some embodiments, Figure 3 As shown, the dimension-increasing process is performed on the features of the screened samples to obtain the features of the newly added dimensions of the samples, including:

[0144] In step S2031, dimension information of a preset machine learning model is obtained;

[0145] In some embodiments, the preset machine learning model may include but is not limited to a multilayer perceptron (MLP) model, a convolutional neural network (CNN) model, and a matrix multiplication model. The dimension information of the preset machine learning model is the dimension information of the model input features and output features.

[0146] For example, for feature crossover, an MLP model with both input and output dimensions of r, when r=8, the function h(*) corresponding to the model can be a three-layer MLP composed of (8*8+8)*3=216-dimensional personalized parameters. These parameters enable different users or videos to perform adaptive personalized feature crossover based on different features, so that the trained model can achieve better recommendation results. The personalized model can be added only to the last layer of feature crossover, or to each layer. For expert networks, h(*) is an MLP calculation with both input and output dimensions of M. Where M is the number of expert networks, for example, M=8, then h(*) can be a three-layer MLP composed of (8*8+8)*3=216-dimensional personalized parameters. These parameters enable different users or videos to perform adaptive feature crossover based on different expert networks, so that the trained model can achieve better recommendation results.

[0147] In step S2033, based on the dimension information, the dimension of the screening sample feature is updated to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature;

[0148] In some embodiments, the dimension of the sample feature can be updated according to the dimension information of the preset machine learning model to obtain the sample update feature, thereby increasing the dimension of the sample feature. Features of different dimensions correspond to different numbers of feature vectors.

[0149] In step S2035, based on the sample update feature, the sample newly added dimension feature is determined; the preset sample feature is the sample feature, the sample update feature or the sample newly added dimension feature.

[0150] In some embodiments, a first feature vector set can be obtained according to the dimension corresponding to the sample update feature; a second feature vector set can be obtained according to the dimension corresponding to the sample feature; the number of second feature vectors in the second feature vector set is greater than the number of first feature vectors in the first feature vector set; then the first feature vector set is filtered out from the first feature vector set, and the feature vector obtained is the sample newly added dimension feature. For example, the sample feature is a feature vector of two dimensions, X and Y; the sample update feature is a feature vector of three dimensions, X, Y, and Z; then the sample newly added dimension feature is the Z feature vector.

[0151] In the disclosed embodiment, the number of new dimensions of the feature can be determined according to the type of preset machine learning model, so as to obtain the new dimension features of the sample, thereby realizing adaptive personalized feature crossing based on different features, so that the trained model can achieve good recommendation effect and improve the conversion rate of recommended information.

[0152] In some embodiments, the updating of the dimension of the screening sample feature based on the dimension information to obtain the sample update feature includes:

[0153] Based on the dimensional information, determine the number of dimensions of the newly added dimensional features of the sample;

[0154] Based on the number of dimensions, the dimensions of the sample object features and / or the sample multimedia information associated features are updated to obtain the sample updated features.

[0155] In some embodiments, the dimensions of the sample object features and / or the sample multimedia information associated features are updated, that is, the dimensions of the corresponding features are increased; for example, corresponding dimensions may be increased for video features and user features; it may also be selected to increase the dimensions only for the video ID, or to increase the dimensions for multiple features at the same time; it may also be selected to increase the dimensions only for the user ID, or to increase the dimensions for multiple features at the same time; it may also be selected to increase the dimensions for the features used for feature crossover. Increasing the corresponding dimensions means adding new dimensions while maintaining the original dimensions unchanged.

[0156] In the disclosed embodiments, some or all sample features can be selected for dimensionality processing based on actual application scenarios, and a personalized parameter prediction model can be constructed based on the newly added dimensional features, thereby solving the problem that feature cross-talk cannot achieve personalization in multimedia information push.

[0157] In step S205, a preset machine learning model is trained according to the newly added dimensional features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model.

[0158] In some embodiments, the sample personalized parameters may include but are not limited to the time period and location information of the sample object viewing the multimedia information; it may also include the category of the multimedia information, the number of exposures, etc.; the newly added dimensional features of the sample and the corresponding sample personalized parameters may be used to train a preset machine learning model to obtain a personalized parameter prediction model and a corresponding personalized parameter prediction function.

[0159] In step S207, a multi-task model is trained according to preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the sample newly added dimensional features.

[0160] In some embodiments, the sample business data may include but is not limited to the data of the sample object's operation behavior on the sample multimedia information (such as likes, attention, viewing time, etc.). The sample business data for the multi-task model is multiple, that is, the model corresponds to the prediction of multiple business indicators;

[0161] Among them, the multi-task model (Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts, MMOE) can improve the learning efficiency and quality of each task by learning the connections and differences between different tasks. The framework of multi-task learning widely adopts the Shared-Bottom structure; the idea of ​​Shared-Bottom is that multiple targets share a set of shared bottom hidden layers (layers) at the bottom, and different sub-networks (Towers) are built on top of them based on different targets. The advantage of this is that the bottom layer is reused, which reduces the amount of calculation and prevents overfitting.

[0162] In some embodiments, the sample feature and the sample newly added dimension feature are both feature vectors corresponding to the same feature. For example, the sample feature is a feature vector of two dimensions, X and Y; the sample newly added dimension feature is a Z feature vector.

[0163] In some embodiments, the original sample features, the sample update features, or the sample newly added dimension features may be combined with a personalized parameter prediction function to train a multi-task model, thereby obtaining a multi-business data prediction model.

[0164] In some embodiments, the multi-task model MMOE includes a gating network and at least two expert networks, wherein the number of expert networks can generally be determined based on multiple experimental results, for example, can be 4-8.

[0165] In some embodiments, the multi-task model includes a gating network and at least two expert networks; Figure 4 As shown, the multi-task model is trained according to the preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model, including:

[0166] In step S2071, each expert network is trained according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network;

[0167] In some embodiments, the preset sample features are at least two, and the training of each expert network according to the preset sample features and the personalized parameter prediction function to obtain the target cross-feature determination function of each expert network includes:

[0168] Performing splicing processing on the at least two preset sample features to obtain a sample cross feature;

[0169] In some embodiments, during the multi-task model training process, it is usually necessary to splice multiple preset sample features to obtain sample cross-features; multiple preset sample features can be feature vectorized (embedding) to obtain multiple feature vectors, and then the multiple feature vectors can be spliced ​​to obtain sample cross-features.

[0170] A preset number of cross-trainings are performed on each expert network according to the sample cross-features and the personalized parameter prediction function to obtain a target cross-feature determination function for each expert network.

[0171] In some embodiments, a multi-layer feature cross network can be constructed based on the cross layer (CrossLayer) of low rank expansion and used as an expert network, that is, each expert network includes a multi-layer feature cross network; the input features of each expert network are sample cross features, and the training focus of each expert network is different. For example, different expert networks can input features of different time periods for training, one focusing on information recommendation in the morning time, and the other focusing on information recommendation in the afternoon time.

[0172] In some embodiments, the calculation formula of the L+1th layer output feature in each expert network is as follows:

[0173] x l+1 =x0⊙(W l x l +b l )+x l

[0174] Among them, ⊙ refers to the inner product, and the features in the formula are all feature vectors; x0 is the input sample cross feature, x l and x l+1 are the outputs of the Lth and L+1th layer feature crossover, L = 0, 1, ...; x0, x l ,x l+1 ,b l are both d-dimensional vectors. l is a d*d dimensional vector, which is related to b l These are all training parameters. During the model training process, the parameters can be adjusted. After the model training is completed, the fixed values ​​of the parameters can be obtained.

[0175] In the disclosed embodiment, multiple preset sample features may be concatenated and then input into various expert networks for training in combination with personalized parameter prediction functions; thereby cross-training of personalized features may be achieved, thereby improving the prediction accuracy of the model.

[0176] In some embodiments, performing a preset number of cross-training on each expert network according to the sample cross-features and the personalized parameter prediction function to obtain a target cross-feature determination function for each expert network includes:

[0177] A preset number of cross-trainings are performed on each expert network according to the sample cross-features and the personalized parameter prediction function to obtain an initial cross-feature determination function of each expert network.

[0178] In the embodiment of the present disclosure, the crossover process is performed for a preset number of times (N times) to obtain a d-dimensional feature crossover output x N , the value of N can be set according to the actual situation, for example, it can be set to 3. The initial cross-feature determination function y of each expert network N as follows:

[0179] y N =x N W N

[0180] Among them, W N is a d*o dimensional vector, y N is an o-dimensional vector.

[0181] The initial cross-features of each expert network are subjected to low-rank decomposition processing, and the initial cross-feature determination function of each expert network is updated to the target cross-feature determination function.

[0182] In some embodiments, in order to reduce the complexity of the model calculation process, the initial cross-features of each expert network may be subjected to low-rank decomposition processing, and the formula is as follows:

[0183] W l =U l V l T

[0184] Among them, U l ,V l Both are d*r dimensional matrices, r< <d。

[0185] The update calculation formula for the L+1th layer output features in each expert network is as follows:

[0186] x l+1 =x0⊙(U l V l T x l +b l )+x l

[0187] And can be based on the updated x l+1 The calculation formula is used to get the updated x N , and then get the updated y N , and the updated y N As the target cross-feature determination function.

[0188] Assuming there are M expert networks and K types of sample business data, based on MMOE, the above feature crossover method will be repeated M times with the same input and different model parameters to obtain M o-dimensional output vectors, which are the output results of the expert network.

[0189] In the disclosed embodiment, the initial cross-features of each expert network may be subjected to low-rank decomposition processing, thereby reducing the complexity of the model calculation process and improving the training speed of the multi-business data prediction model.

[0190] In some embodiments, each expert network includes at least two cross layers, and training each expert network according to the preset sample features and the personalized parameter prediction function to obtain the target cross feature determination function of each expert network includes:

[0191] Determine a preset cross layer corresponding to the personalized parameter prediction function in each expert network.

[0192] In some embodiments, the preset cross-layer may be any one, multiple or all layers of the expert network; for example, the personalized parameter prediction function may be added to the last layer of feature cross-layer or to each layer.

[0193] At the preset cross layer of each expert network, each expert network is trained according to the preset sample features and the personalized parameter prediction function to obtain a target cross feature determination function of each expert network.

[0194] In some embodiments, if the preset machine learning model is MLP, the update calculation formula of the L+1th layer output feature in each expert network can be adjusted in the determined preset cross layer:

[0195] x l+1 =x0⊙(U l h(V l T x l )+b l )+x l

[0196] Where h is an MLP calculation with input and output dimensions r. For example, if r = 8, h can be a three-layer MLP composed of (8*8+8)*3=216 dimensional personalized parameters. These parameters enable different users or videos to perform adaptive personalized feature crossover based on different features, achieving better recommendation results.

[0197] If the preset machine learning model is CNN, the formula for the output features of the L+1th layer in each expert network is adjusted as follows:

[0198] x l+1 =x0⊙(U l c(V l T x l )+b l )+x l

[0199] Where c is the personalized parameter prediction function corresponding to the CNN model whose input channel and output channel are both r.

[0200] If the preset machine learning model is a matrix multiplication model, the formula for the output features of the L+1th layer in each expert network is adjusted as follows:

[0201] x l+1 =x0⊙(U l WV l T x l +b l )+x l

[0202] Among them, W is an r*r matrix, which ensures that the dimension of the parameters remains unchanged after matrix multiplication.

[0203] In the disclosed embodiment, the preset cross-layer corresponding to the personalized parameter prediction function can be determined according to actual conditions, so as to implement cross-training of personalized features in the preset cross-layer of the expert network and improve the prediction accuracy of the multi-business data prediction model.

[0204] In step S2073, the gated network is trained according to the preset sample features, the at least two sample service data, and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data;

[0205] In the disclosed embodiment, the gating network is composed of MLP, and the gating network can be trained according to a variety of sample business data. The preset sample features are combined with the personalized parameter prediction function to train the gating network to obtain an output result; and when the output result matches the various sample business data, the weight determination function of each expert network corresponding to each sample business data is determined; thereby obtaining the weight of each expert network relative to any sample business data.

[0206] In some embodiments, the training of the gating network according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain the weight determination function of each expert network corresponding to each sample service data includes:

[0207] The gated network is trained according to the preset sample features and the at least two sample business data to obtain initial weight determination functions of each expert network corresponding to each sample business data.

[0208] In some embodiments, in a multi-task scenario, multiple gating networks can be used; for each task, a separate gating network can be assigned; preset sample features can be input into the gating network for training, and the output layer is an M-dimensional normalized exponential function (softmax); for K types of sample business data, each gating network outputs the weight determination function of each expert network corresponding to its corresponding sample business data Thus, the multi-expert network output result of each sample business data can be obtained, that is, the initial weight determination function of each expert network corresponding to each sample business data, and the calculation formula is as follows:

[0209]

[0210] Among them, f k (x0) is the output result corresponding to the k-th sample business data, which is o-dimensional. is the softmax output of the i-th dimension of the k-th MLP, f i (x0) is the output of the i-th expert network, i.e., y above N x0 can use the original feature dimensions, or only use the newly added extra dimensions for feature crossover, or use the original dimensions plus the newly added dimensions. For example, if there are 10 original 32-dimensional features, you can use these 320 dimensions for feature crossover, or you can add 8 dimensions to each feature and only use the newly added 80 dimensions for feature crossover.

[0211] Based on the personalized parameter prediction function, the initial weight determination function of each expert network corresponding to each sample business data is updated to obtain the weight determination function of each expert network corresponding to each sample business data.

[0212] In some embodiments, if the preset machine learning model is MLP, the weight determination function of each expert network corresponding to each sample business data can be determined based on the personalized parameter prediction function, and the calculation formula is as follows:

[0213]

[0214] Where h is an MLP calculation with both input and output dimensions of M. Where M is the number of expert networks, for example, M = 8, then h can be a three-layer MLP composed of (8*8+8)*3=216 dimensional personalized parameters. These parameters enable different users or videos to perform adaptive feature crossover based on different expert networks, achieving better recommendation results.

[0215] If the preset machine learning model is CNN, based on the personalized parameter prediction function, the weight determination function of each expert network corresponding to each sample business data is determined. The calculation formula is as follows:

[0216]

[0217] Where c is a CNN model with M input channels and M output channels.

[0218] If the preset machine learning model is a matrix multiplication model, based on the personalized parameter prediction function, the weight determination function of each expert network corresponding to each sample business data is determined. The calculation formula is as follows:

[0219]

[0220] Where W is an M*M matrix, which ensures that the dimension of the parameters remains unchanged after matrix multiplication.

[0221] In the disclosed embodiment, the gated network can be trained according to the preset sample features and the at least two sample business data to obtain the initial weight determination function of each expert network corresponding to each sample business data; then, according to the personalized parameter prediction function, the initial weight determination function of each expert network corresponding to each sample business data is updated to obtain the weight determination function of each expert network corresponding to each sample business data, which can improve the accuracy of the weight of each expert network corresponding to each sample business data, thereby improving the prediction accuracy of the multi-business data prediction model.

[0222] In step S2075, the model parameters of the multi-task model are determined according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data.

[0223] In some embodiments, Figure 5 As shown, the model parameters of the multi-task model are determined according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data, including:

[0224] In step S20751, the product of the target cross-feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data is calculated to obtain the initial weight feature determination function of each expert network corresponding to each sample business data;

[0225] In some embodiments, the product of the target cross-feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data may be used as the initial weight feature determination function of each expert network corresponding to each sample business data.

[0226] In step S20753, the sum of the initial weight feature determination functions of the expert networks corresponding to each type of sample business data is calculated to obtain the target weight cross feature determination function corresponding to each type of sample business data;

[0227] In some embodiments, the sum of the initial weight feature determination functions of the expert networks corresponding to each type of sample business data may be determined as the target weight cross-feature determination function corresponding to each type of sample business data.

[0228] In step S20755, the model parameters of the multi-task model are determined according to the target weight cross-feature determination function corresponding to each sample business data.

[0229] In some embodiments, the target weight cross-feature determination function corresponding to each sample business data may be determined as a model parameter of the multi-task model.

[0230] In some embodiments, the obtained M f k (x0) is spliced ​​to the input side of the M multi-task estimation tower (MLP) to achieve multi-task feature crossover.

[0231] In the disclosed embodiment, model training can be performed based on the target cross-feature determination function and the weight determination function of each expert network corresponding to various sample business data, so as to quickly and accurately determine the model prediction parameters of various sample business data, thereby constructing a multi-business data prediction model.

[0232] In step S2077, the multi-service data prediction model is determined based on the multi-task model and the model parameters.

[0233] In the disclosed embodiment, the expert network and the gated network can be trained separately to accurately obtain the target cross-feature determination function that combines personalized parameter features and the weight determination function of each expert network corresponding to each sample business data, thereby improving the prediction accuracy of the multi-business data prediction model.

[0234] The present invention obtains sample features of sample multimedia information and at least two types of sample business data; performs dimensionality increase processing on the screened sample features to obtain sample newly added dimensional features; trains a preset machine learning model according to the sample newly added dimensional features and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; trains a multi-task model according to the preset sample features, the at least two types of sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; during the model training process, a personalized parameter prediction model is constructed for each sample multimedia information for feature crossing, which solves the problem that feature crossing in multimedia information push cannot achieve personalization, thereby improving the accuracy of the prediction results of the multi-business data prediction model and improving the conversion rate of the pushed information obtained according to the multi-business data prediction model.

[0235] Figure 6 is a flowchart of a multimedia information push method according to an exemplary embodiment; Figure 6 As shown, this method can be applied to Figure 1 The server 01 shown includes the following steps.

[0236] In step S601, the target features and target newly added dimension features of the target multimedia information are obtained; the target features include target object features and target multimedia information associated features. The target multimedia information and the sample multimedia information are information of the same field and the same type; the target features and the sample features are features of the same type, the target newly added dimension features and the sample newly added dimension features are features of the same type; the target object features and the sample object features are features of the same type, and the target multimedia information associated features and the sample multimedia information associated features are features of the same type. The target object may be a terminal to be recommended.

[0237] In some embodiments, the step of obtaining target features and target newly added dimension features of target multimedia information includes:

[0238] Acquiring target features of the target multimedia information;

[0239] Based on the target feature, determining a screening target feature; the screening sample feature is part or all of the target feature;

[0240] In the disclosed embodiment, the screening target feature corresponds to the screening sample feature and is the same type of feature.

[0241] The screening target features are subjected to dimensionality increase processing to obtain target newly added dimensional features.

[0242] In the disclosed embodiment, the dimension increasing process of screening target features is consistent with the dimension increasing process of screening sample features.

[0243] In the disclosed embodiment, the screening target features after dimensionality increase processing can be determined from the target features to obtain the target newly added dimension features, so as to facilitate obtaining the target personalized parameters, and use the target personalized parameters to input the model for data prediction, thereby improving the prediction accuracy of the multi-business data prediction model.

[0244] In step S603, the target newly added dimension features are input into a personalized parameter prediction model to obtain target personalized parameters.

[0245] In the disclosed embodiment, the personalized parameter prediction model is obtained by training based on the above-mentioned preset machine learning model. The multi-service data prediction model is obtained by training according to the above-mentioned multi-service data prediction model training method.

[0246] In step S605, the preset target feature and the target personalized parameter are input into a multi-business data prediction model to obtain at least two types of target business data; the preset target feature is the target feature or the target newly added dimension feature.

[0247] In the disclosed embodiment, the target business data may represent the operation behavior score of the target object on the target multimedia information; for example, the like behavior score of a user on a short video may be predicted to determine whether the user will like the short video.

[0248] In step S607, the target multimedia information is pushed to the target object according to the at least two target service data.

[0249] In some embodiments, Figure 7 As shown, the pushing the target multimedia information to the target object according to the at least two target service data includes:

[0250] In step S6071, the weight of each target business data is determined;

[0251] In some embodiments, the weight of each target business data can be determined according to the actual application scenario and business objectives. For example, in the short video recommendation scenario, if the business objective is viewing time, a higher weight is set for viewing time, and lower weights are set for other target business data.

[0252] In step S6073, the comprehensive service data of the target multimedia information is determined according to each target service data and the weight of each target service data;

[0253] In the disclosed embodiment, the comprehensive business data of the target multimedia information can be calculated according to the linear combination (ensemble sort) formula; the product of each target business data and the corresponding weight can be calculated to obtain the initial comprehensive business data corresponding to each target business data; and then the sum of the initial comprehensive business data corresponding to various target business data can be calculated to obtain the comprehensive business data of the target multimedia information.

[0254] In some embodiments, the number of the target multimedia information is at least two, and the step of determining the integrated service data of the target multimedia information includes:

[0255] Determine the comprehensive business data of each target multimedia information.

[0256] In step S6075, if the comprehensive service data is greater than a preset threshold, the target multimedia information is pushed to the target object.

[0257] In some embodiments, the preset threshold can be set according to actual conditions. When the comprehensive business data is less than or equal to the preset threshold, the target multimedia information is not pushed to the target object, thereby avoiding wasting traffic and reducing the user experience of the target object.

[0258] In the disclosed embodiment, the weight of each target business data can be set according to actual conditions, and the comprehensive business data of the target multimedia information can be obtained based on each target business data and the corresponding weight, thereby improving the accuracy of the comprehensive business data of the target multimedia information and the accuracy of information recommendation.

[0259] In some embodiments, Figure 8 As shown, the pushing of the target multimedia information to the target object includes:

[0260] In step S60751, the at least two target multimedia information are sorted based on the comprehensive service data of each target multimedia information;

[0261] In step S60753, the multimedia information to be pushed is determined according to the sorting result;

[0262] In step S60755, the multimedia information to be pushed is pushed to the target object.

[0263] In some embodiments, multiple target multimedia information may be sorted from high to low according to the corresponding integrated business data size, and a preset number of target multimedia information at the top of the sort may be determined as the multimedia information to be pushed; the preset number may be one or more.

[0264] In the disclosed embodiment, the multimedia information to be pushed can be determined based on the comprehensive business data sorting results of multiple target multimedia information; thereby, a preset number of multimedia information can be recommended to the target object according to actual conditions, thereby increasing the number of recommended information and improving user experience.

[0265] The present disclosure proposes to construct a separate personalized small model for each new user and new video for feature crossover, so that each user and video can personalize and select different features for feature crossover, solving the problem that feature crossover in short video recommendation cannot be personalized; the present disclosure proposes to combine MMOE to realize multi-objective modeling of feature crossover, solving the problem that feature crossover in short video recommendation cannot realize multi-task estimation.

[0266] The present invention also proposes to construct a personalized small model for each new user and new video, and adaptively select different expert networks for multi-objective modeling, so as to ensure that different expert networks learn different information, thereby solving the problem that the MMoe model cannot realize personalized multi-objective modeling and easily learns convergent information; the present invention can not only automatically perform adaptive high-order feature combination for different users and videos, but also automatically perform adaptive multi-task modeling.

[0267] In the embodiment of the present disclosure, the newly added dimension features of the target can be input into the personalized parameter prediction model to obtain the target personalized parameters; then the preset target features and the target personalized parameters can be input into the multi-business data prediction model to obtain at least two types of target business data; thereby realizing information recommendation by integrating personalized features, thereby improving the accuracy and conversion rate of recommended information.

[0268] Fig. 9 1 is a block diagram of a training device for a multi-service data prediction model according to an exemplary embodiment. Fig. 9 , the device comprises:

[0269] The sample data acquisition module 910 is configured to acquire sample features of sample multimedia information and at least two types of sample business data; the sample features include sample object features and sample multimedia information association features;

[0270] The sample newly added dimension feature determination module 920 is configured to perform dimension increase processing on the screened sample features to obtain the sample newly added dimension feature; the screened sample feature is part or all of the sample feature;

[0271] The personalized parameter prediction function determination module 930 is configured to execute a preset machine learning model according to the newly added dimension features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model;

[0272] The multi-business data prediction model determination module 940 is configured to execute training of a multi-task model based on preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the sample newly added dimensional features.

[0273] In some embodiments, the multi-task model includes a gating network and at least two expert networks; the multi-service data prediction model determination module includes:

[0274] A target cross-feature determination function determination unit is configured to train each expert network according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network;

[0275] A weight determination function determination unit is configured to train the gating network according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data;

[0276] A model parameter determination unit, configured to determine the model parameters of the multi-task model according to a target cross-feature determination function of each expert network and a weight determination function of each expert network corresponding to each sample business data;

[0277] The multi-service data prediction model determination unit is configured to determine the multi-service data prediction model based on the multi-task model and the model parameters.

[0278] In some embodiments, the model parameter determination unit includes:

[0279] The initial weight feature determination function determination subunit is configured to calculate the product of the target cross feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data, so as to obtain the initial weight feature determination function of each expert network corresponding to each sample business data;

[0280] The target weight cross-feature determination function determination subunit is configured to calculate the sum of the initial weight feature determination functions of each expert network corresponding to each sample business data to obtain the target weight cross-feature determination function corresponding to each sample business data;

[0281] The model parameter determination subunit is configured to execute a target weight cross-feature determination function corresponding to each sample business data to determine the model parameters of the multi-task model.

[0282] In some embodiments, the number of the preset sample features is at least two, and the target cross-feature determination function determination unit includes:

[0283] A sample cross feature determination subunit is configured to perform a splicing process on the at least two preset sample features to obtain a sample cross feature;

[0284] The first function determination subunit is configured to perform a preset number of cross-training on each expert network according to the sample cross-features to obtain a target cross-feature determination function for each expert network.

[0285] In some embodiments, the first function determination subunit includes:

[0286] an initial cross-feature determination function determination subunit, configured to perform a preset number of cross-training operations on each expert network according to the sample cross-features, to obtain an initial cross-feature determination function of each expert network;

[0287] The function updating subunit is configured to perform low-rank decomposition processing on the initial cross-features of each expert network, and update the initial cross-feature determination function of each expert network to the target cross-feature determination function.

[0288] In some embodiments, each expert network includes at least two cross layers, and the target cross feature determination function determination unit includes:

[0289] A preset cross layer determination subunit is configured to determine the preset cross layer corresponding to the personalized parameter prediction function in each expert network;

[0290] The second function determination subunit is configured to execute at a preset cross layer of each expert network, train each expert network according to the preset sample features and the personalized parameter prediction function, and obtain a target cross feature determination function of each expert network.

[0291] In some embodiments, the weight determination function determination unit includes:

[0292] An initial weight determination function determination subunit is configured to train the gated network according to the preset sample features and the at least two sample business data to obtain an initial weight determination function of each expert network corresponding to each sample business data;

[0293] The weight determination function determination subunit is configured to execute, based on the personalized parameter prediction function, an update of the initial weight determination function of each expert network corresponding to each sample business data, to obtain the weight determination function of each expert network corresponding to each sample business data.

[0294] In some embodiments, the sample newly added dimension feature determination module includes:

[0295] A dimension information determining unit, configured to execute obtaining dimension information of the preset machine learning model;

[0296] A sample update feature determination unit is configured to update the dimension of the screening sample feature based on the dimension information to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature;

[0297] The sample newly added dimension feature determination unit is configured to determine the sample newly added dimension feature based on the sample update feature; the preset sample feature is the sample feature, the sample update feature or the sample newly added dimension feature.

[0298] In some embodiments, the sample update feature determination unit includes:

[0299] A dimension number determination subunit is configured to determine the number of dimensions of the newly added dimensional features of the sample based on the dimensional information;

[0300] The sample update feature determination subunit is configured to update the dimensions of the sample object feature and / or the sample multimedia information associated feature based on the number of dimensions to obtain the sample update feature.

[0301] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0302] Fig.10 is a block diagram of a multimedia information push device according to an exemplary embodiment. Fig.10 , the device comprises:

[0303] The target feature acquisition module 1010 is configured to execute acquisition of target features and target newly added dimension features of target multimedia information; the target features include target object features and target multimedia information associated features;

[0304] The target personalized parameter determination module 1020 is configured to input the target newly added dimension feature into the personalized parameter prediction model to obtain the target personalized parameter;

[0305] The target business data determination module 1030 is configured to input a preset target feature and the target personalized parameter into a multi-business data prediction model to obtain at least two types of target business data; the preset target feature is the target feature or the target newly added dimension feature;

[0306] The target multimedia information pushing module 1040 is configured to push the target multimedia information to the target object according to the at least two target service data.

[0307] In some embodiments, the multi-service data prediction model is trained according to the multi-service data prediction model training method described above.

[0308] In some embodiments, the target multimedia information push module includes:

[0309] A weight determination unit, configured to determine the weight of each target business data;

[0310] a comprehensive service data determination unit, configured to determine the comprehensive service data of the target multimedia information according to each target service data and the weight of each target service data;

[0311] The target multimedia information pushing unit is configured to push the target multimedia information to the target object if the comprehensive service data is greater than a preset threshold.

[0312] In some embodiments, the number of the target multimedia information is at least two, and the integrated service data determination unit includes:

[0313] The integrated service data determination subunit is configured to determine the integrated service data of each target multimedia information;

[0314] The target multimedia information pushing unit includes:

[0315] An information sorting subunit, configured to sort the at least two target multimedia information based on the comprehensive service data of each target multimedia information;

[0316] The multimedia information to be pushed determining subunit is configured to determine the multimedia information to be pushed according to the sorting result;

[0317] The information pushing subunit is configured to push the multimedia information to be pushed to the target object.

[0318] In some embodiments, the target feature acquisition module includes:

[0319] A target feature acquisition unit, configured to execute acquisition of target features of the target multimedia information;

[0320] A screening target feature determination unit is configured to determine a screening target feature based on the target feature; the screening sample feature is part or all of the target feature;

[0321] The feature processing unit is configured to perform dimensionality increase processing on the screening target features to obtain target newly added dimensional features.

[0322] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0323] In an exemplary embodiment, an electronic device is also provided, including a processor; a memory for storing instructions executable by the processor; wherein, when the processor is configured to execute the instructions stored in the memory, it implements the training method of the multi-service data prediction model or the multimedia information push method provided in any of the above embodiments.

[0324] The electronic device may be a terminal, a server or a similar computing device. For example, the electronic device is a server. Fig.11 is a block diagram of an electronic device according to an exemplary embodiment. Fig.11 As shown, the server 1100 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 1110 (the processor 1110 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 1130 for storing data, and one or more storage media 1120 (such as one or more mass storage devices) for storing application programs 1123 or data 1122. Among them, the memory 1130 and the storage medium 1120 can be short-term storage or permanent storage. The program stored in the storage medium 1120 may include one or more modules, each of which may include a series of instruction operations on the server. Furthermore, the central processing unit 1110 can be configured to communicate with the storage medium 1120 and execute a series of instruction operations in the storage medium 1120 on the server 1100. The server 1100 may also include one or more power supplies 1160, one or more wired or wireless network interfaces 1150, one or more input and output interfaces 1140, and / or one or more operating systems 1121, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

[0325] The input / output interface 1140 may be used to receive or send data via a network. The specific example of the network may include a wireless network provided by a communication provider of the server 1100. In one example, the input / output interface 1140 includes a network adapter (Network Interface Controller, NIC), which may be connected to other network devices via a base station so as to communicate with the Internet. In one example, the input / output interface 1140 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0326] It can be understood by those skilled in the art that Fig.11 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig.11 More or fewer components as shown, or with Fig.11 Different configurations are shown.

[0327] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as a memory 1130 including instructions, and the instructions can be executed by the processor 1110 of the device 1100 to complete the above method. Alternatively, the computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0328] In an exemplary embodiment, a computer program product is also provided, including a computer program, which, when executed by a processor, implements the training method of the multi-service data prediction model or the multimedia information pushing method provided in any of the above embodiments.

[0329] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0330] Those skilled in the art will readily appreciate 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 that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0331] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A training method for a multi-service data prediction model, characterized in that: include: Acquire sample features of sample multimedia information and at least two types of sample business data; the sample features include sample object features and sample multimedia information association features; Performing dimension-increasing processing on the screening sample features to obtain new dimension features of the samples; the screening sample features are part or all of the sample features; The preset machine learning model is trained according to the newly added dimension features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; A multi-task model is trained according to preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the sample newly added dimensional features.

2. The method according to claim 1, characterized in that The multi-task model includes a gated network and at least two expert networks; the multi-task model is trained according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a multi-service data prediction model, including: Training each expert network according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function for each expert network; The gating network is trained according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data; Determining the model parameters of the multi-task model according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data; Based on the multi-task model and the model parameters, the multi-service data prediction model is determined.

3. The method according to claim 2, characterized in that Determining the model parameters of the multi-task model according to the target cross-feature determination function of each expert network and the weight determination function of each expert network corresponding to each sample business data includes: Calculate the product of the target cross-feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data to obtain the initial weight feature determination function of each expert network corresponding to each sample business data; Calculating the sum of the initial weight feature determination functions of each expert network corresponding to each sample business data to obtain the target weight cross feature determination function corresponding to each sample business data; The model parameters of the multi-task model are determined according to the target weight cross-feature determination function corresponding to each sample business data.

4. The method according to claim 2, characterized in that: The preset sample features are at least two, and each expert network is trained according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network, including: Performing splicing processing on the at least two preset sample features to obtain a sample cross feature; A preset number of cross-trainings are performed on each expert network according to the sample cross-features and the personalized parameter prediction function to obtain a target cross-feature determination function for each expert network.

5. The method according to claim 4, characterized in that The cross-training of each expert network for a preset number of times according to the sample cross-features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network includes: Performing a preset number of cross-training on each expert network according to the sample cross-features and the personalized parameter prediction function to obtain an initial cross-feature determination function for each expert network; The initial cross-features of each expert network are subjected to low-rank decomposition processing, and the initial cross-feature determination function of each expert network is updated to the target cross-feature determination function.

6. The method according to claim 2, characterized in that Each expert network includes at least two cross layers, and the training of each expert network according to the preset sample features and the personalized parameter prediction function to obtain the target cross feature determination function of each expert network includes: Determine a preset cross layer corresponding to the personalized parameter prediction function in each expert network; At the preset cross layer of each expert network, each expert network is trained according to the preset sample features and the personalized parameter prediction function to obtain a target cross feature determination function of each expert network.

7. The method according to claim 2, characterized in that The step of training the gating network according to the preset sample features, the at least two sample service data, and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data includes: The gated network is trained according to the preset sample features and the at least two sample business data to obtain an initial weight determination function of each expert network corresponding to each sample business data; Based on the personalized parameter prediction function, the initial weight determination function of each expert network corresponding to each sample business data is updated to obtain the weight determination function of each expert network corresponding to each sample business data.

8. The method according to claim 1, characterized in that The dimension-increasing process is performed on the features of the screened samples to obtain the features of the newly added dimensions of the samples, including: Obtaining dimension information of the preset machine learning model; Based on the dimension information, the dimension of the screening sample feature is updated to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature; Based on the sample update feature, the sample newly added dimension feature is determined; the preset sample feature is the sample feature or the sample newly added dimension feature.

9. The method according to claim 8, characterized in that The updating of the dimension of the screening sample feature based on the dimension information to obtain the sample update feature includes: Based on the dimensional information, determining the number of dimensions of the newly added dimensional features of the sample; Based on the number of dimensions, the dimensions of the sample object features and / or the sample multimedia information associated features are updated to obtain the sample updated features.

10. A multimedia information push method, characterized in that: The method comprises: Obtaining target features and target newly added dimension features of target multimedia information; the target features include target object features and target multimedia information association features; Inputting the target newly added dimension features into a personalized parameter prediction model to obtain target personalized parameters; Inputting the preset target feature and the target personalized parameter into a multi-service data prediction model to obtain at least two types of target service data; the preset target feature is the target feature or the target newly added dimension feature; The target multimedia information is pushed to the target object according to the at least two target service data.

11. The method according to claim 10, characterized in that The pushing the target multimedia information to the target object according to the at least two target service data includes: Determine the weight of each target business data; Determining the comprehensive service data of the target multimedia information according to each target service data and the weight of each target service data; If the comprehensive service data is greater than a preset threshold, the target multimedia information is pushed to the target object.

12. The method according to claim 11, characterized in that The target multimedia information is at least two, and the determining of the comprehensive service data of the target multimedia information includes: Determine comprehensive business data for each target multimedia information; The pushing the target multimedia information to the target object comprises: sorting the at least two target multimedia information based on the comprehensive service data of each target multimedia information; Determine the multimedia information to be pushed according to the sorting result; Pushing the multimedia information to be pushed to the target object.

13. The method according to claim 10, characterized in that The step of obtaining target features and target newly added dimension features of target multimedia information includes: Acquiring target features of the target multimedia information; Based on the target feature, determining a screening target feature; the screening target feature is part or all of the target feature; The screening target features are subjected to dimensionality increase processing to obtain target newly added dimensional features.

14. A training device for a multi-service data prediction model, characterized in that: include: A sample data acquisition module is configured to acquire sample features of sample multimedia information and at least two types of sample business data; the sample features include sample object features and sample multimedia information association features; The sample newly added dimension feature determination module is configured to perform dimension increase processing on the screened sample features to obtain the sample newly added dimension feature; the screened sample features are part or all of the sample features; A personalized parameter prediction function determination module is configured to execute a preset machine learning model according to the newly added dimensional features of the sample and the corresponding sample personalized parameters to obtain a personalized parameter prediction function corresponding to the personalized parameter prediction model; The multi-business data prediction model determination module is configured to execute training of a multi-task model based on preset sample features, the at least two sample business data and the personalized parameter prediction function to obtain a multi-business data prediction model; the preset sample features are the sample features or the newly added dimensional features of the samples.

15. The device according to claim 14, characterized in that The multi-task model includes a gating network and at least two expert networks; the multi-service data prediction model determination module includes: A target cross-feature determination function determination unit is configured to train each expert network according to the preset sample features and the personalized parameter prediction function to obtain a target cross-feature determination function of each expert network; A weight determination function determination unit is configured to train the gating network according to the preset sample features, the at least two sample service data and the personalized parameter prediction function to obtain a weight determination function of each expert network corresponding to each sample service data; A model parameter determination unit, configured to determine the model parameters of the multi-task model according to a target cross-feature determination function of each expert network and a weight determination function of each expert network corresponding to each sample business data; The multi-service data prediction model determination unit is configured to determine the multi-service data prediction model based on the multi-task model and the model parameters.

16. The device according to claim 15, characterized in that The model parameter determination unit comprises: The initial weight feature determination function determination subunit is configured to calculate the product of the target cross feature determination function and the weight determination function corresponding to each expert network corresponding to each sample business data, so as to obtain the initial weight feature determination function of each expert network corresponding to each sample business data; The target weight cross-feature determination function determination subunit is configured to calculate the sum of the initial weight feature determination functions of each expert network corresponding to each sample business data to obtain the target weight cross-feature determination function corresponding to each sample business data; The model parameter determination subunit is configured to execute a target weight cross-feature determination function corresponding to each sample business data to determine the model parameters of the multi-task model.

17. The device according to claim 15, characterized in that The preset sample features are at least two, and the target cross-feature determination function determination unit includes: A sample cross feature determination subunit is configured to perform a splicing process on the at least two preset sample features to obtain a sample cross feature; The first function determination subunit is configured to perform a preset number of cross-training on each expert network according to the sample cross-features to obtain a target cross-feature determination function for each expert network.

18. The device according to claim 17, characterized in that The first function determination subunit comprises: an initial cross-feature determination function determination subunit, configured to perform a preset number of cross-trainings on each expert network according to the sample cross-features, to obtain an initial cross-feature determination function of each expert network; The function updating subunit is configured to perform low-rank decomposition processing on the initial cross-features of each expert network, and update the initial cross-feature determination function of each expert network to the target cross-feature determination function.

19. The device according to claim 15, characterized in that Each expert network includes at least two cross layers, and the target cross feature determination function determination unit includes: A preset cross layer determination subunit is configured to determine the preset cross layer corresponding to the personalized parameter prediction function in each expert network; The second function determination subunit is configured to execute at a preset cross layer of each expert network, train each expert network according to the preset sample features and the personalized parameter prediction function, and obtain a target cross feature determination function of each expert network.

20. The device according to claim 15, characterized in that The weight determination function determination unit comprises: An initial weight determination function determination subunit is configured to train the gated network according to the preset sample features and the at least two sample business data to obtain an initial weight determination function of each expert network corresponding to each sample business data; The weight determination function determination subunit is configured to execute, based on the personalized parameter prediction function, an update of the initial weight determination function of each expert network corresponding to each sample business data, to obtain the weight determination function of each expert network corresponding to each sample business data.

21. The device according to claim 14, characterized in that The sample newly added dimension feature determination module includes: A dimension information determining unit, configured to execute obtaining dimension information of the preset machine learning model; A sample update feature determination unit is configured to update the dimension of the screening sample feature based on the dimension information to obtain a sample update feature; the dimension of the sample update feature is greater than the dimension of the sample feature; The sample newly added dimension feature determination unit is configured to determine the sample newly added dimension feature based on the sample update feature; the preset sample feature is the sample feature, the sample update feature or the sample newly added dimension feature.

22. The device according to claim 21, characterized in that The sample update feature determination unit includes: A dimension number determination subunit is configured to determine the number of dimensions of the newly added dimensional features of the sample based on the dimensional information; The sample update feature determination subunit is configured to update the dimensions of the sample object feature and / or the sample multimedia information associated feature based on the number of dimensions to obtain the sample update feature.

23. A multimedia information push device, characterized in that: The device comprises: A target feature acquisition module is configured to acquire target features and target newly added dimension features of target multimedia information; the target features include target object features and target multimedia information associated features; A target personalized parameter determination module is configured to input the target newly added dimension feature into a personalized parameter prediction model to obtain a target personalized parameter; The target business data determination module is configured to input the preset target feature and the target personalized parameter into the multi-business data prediction model to obtain at least two types of target business data; the preset target feature is the target feature or the target newly added dimension feature; The target multimedia information pushing module is configured to push the target multimedia information to the target object according to the at least two target service data.

24. The device according to claim 23, characterized in that The target multimedia information push module includes: A weight determination unit, configured to determine the weight of each target business data; a comprehensive service data determination unit, configured to determine the comprehensive service data of the target multimedia information according to each target service data and the weight of each target service data; The target multimedia information pushing unit is configured to push the target multimedia information to the target object if the comprehensive service data is greater than a preset threshold.

25. The device according to claim 24, characterized in that The target multimedia information is at least two, and the integrated service data determination unit includes: The integrated service data determination subunit is configured to determine the integrated service data of each target multimedia information; The target multimedia information pushing unit includes: An information sorting subunit, configured to sort the at least two target multimedia information based on the comprehensive service data of each target multimedia information; The multimedia information to be pushed determining subunit is configured to determine the multimedia information to be pushed according to the sorting result; The information pushing subunit is configured to push the multimedia information to be pushed to the target object.

26. The device according to claim 24, characterized in that The target feature acquisition module comprises: A target feature acquisition unit, configured to execute acquisition of target features of the target multimedia information; A screening target feature determination unit is configured to determine a screening target feature based on the target feature; the screening target feature is part or all of the target feature; The feature processing unit is configured to perform dimensionality increase processing on the screening target features to obtain target newly added dimensional features.

27. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the training method of the multi-service data prediction model as described in any one of claims 1-9, or the multimedia information push method as described in any one of claims 10-13.

28. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by an electronic device processor, the electronic device is enabled to execute the training method of a multi-service data prediction model as described in any one of claims 1-9, or the multimedia information push method as described in any one of claims 10-13.

29. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the training method of the multi-service data prediction model as described in any one of claims 1-9, or the multimedia information push method as described in any one of claims 10-13.

Images