Multimedia content pushing method and device
Through the combination of event occurrence model and neural network model, extraction and training, the problem of poor multimedia content push is solved, more accurate multimedia content push is achieved, and conversion rate is improved.
Patent Information
- Application Number
- CN202310507983.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-06
- Publication Date
- 2024-11-08
AI Technical Summary
The prior art is difficult to accurately push multimedia content to different usage objects, resulting in poor pushing effect.
Through the training of the event generation model, the neural network model is used to extract the features of the sample object, determine the possibility of related events, and make the features nonlinearly independent according to the reference weight, thereby improving the conversion rate of multimedia content.
The push effect of multimedia content is improved, and the possibility of related events occurring in reference objects is accurately determined, and the conversion rate of push is improved.
Smart Images

Figure CN118921400A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of multimedia technology, and more particularly to a method and device for pushing multimedia content. Background Art
[0002] As the Internet develops, more and more applications emerge. Some applications have push functions that can push multimedia content to the users of the application. For example, some applications can push game videos, news information, etc. to the users. Generally, applications need to push different multimedia content to different users. Therefore, how to push multimedia content has become a problem that needs to be solved urgently. Summary of the invention
[0003] The present application provides a method and device for pushing multimedia content, which can be used to accurately determine the possibility of a reference object having a related event related to the multimedia content through an event occurrence model, which is conducive to improving the conversion rate of the multimedia content, thereby improving the pushing effect of the multimedia content. The technical solution includes the following contents.
[0004] In a first aspect, a method for training an event occurrence model is provided, the method comprising:
[0005] Acquire data of a plurality of sample objects and annotation information of each sample object, wherein the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs to the sample object;
[0006] Extracting features from the data of the plurality of sample objects by using a neural network model to obtain initial features of each sample object;
[0007] Determining prediction information of each sample object based on initial features of each sample object by using the neural network model, wherein the prediction information of the sample object represents the possibility of the related event occurring to the sample object;
[0008] Determining reference weights of the respective sample objects based on the initial features of the respective sample objects, wherein the reference weights of the sample objects are used to make the initial features of the sample objects tend to be nonlinearly independent;
[0009] Based on the labeling information, prediction information and reference weights of each sample object, the neural network model is trained to obtain an event occurrence model, and the event occurrence model is used to determine the possibility of the related event occurring in the reference object.
[0010] In a second aspect, a method for pushing multimedia content is provided, the method comprising:
[0011] Get the data of the reference object;
[0012] Determining initial features of the reference object based on the data of the reference object through an event occurrence model, wherein the event occurrence model is obtained by training according to the event occurrence model training method described in the first aspect;
[0013] Determining prediction information of the reference object based on initial features of the reference object through the event occurrence model, wherein the prediction information of the reference object represents the possibility of the occurrence of a relevant event related to the multimedia content by the reference object;
[0014] Based on the prediction information of the reference object, it is determined whether to push the multimedia content to the reference object.
[0015] In a third aspect, a training device for an event occurrence model is provided, the device comprising:
[0016] An acquisition module, used to acquire data of multiple sample objects and annotation information of each sample object, wherein the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs to the sample object;
[0017] A feature extraction module, used to extract features from the data of the plurality of sample objects through a neural network model to obtain initial features of each sample object;
[0018] A determination module, configured to determine prediction information of each sample object based on initial features of each sample object through the neural network model, wherein the prediction information of the sample object represents the possibility of the related event occurring to the sample object;
[0019] The determination module is further used to determine the reference weight of each sample object based on the initial features of each sample object, wherein the reference weight of the sample object is used to make the initial features of the sample object tend to be nonlinearly independent;
[0020] The training module is used to train the neural network model based on the labeling information, prediction information and reference weights of each sample object to obtain an event occurrence model, and the event occurrence model is used to determine the possibility of the related event occurring in the reference object.
[0021] In a possible implementation manner, the plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which the multimedia content is not pushed;
[0022] The determination module is used to determine the prediction information of each first object based on the first identification information and the initial features of the multiple first objects through the neural network model, and the first identification information represents the situation where the multimedia content is pushed; and to determine the prediction information of each second object based on the second identification information and the initial features of the multiple second objects through the neural network model, and the second identification information represents the situation where the multimedia content is not pushed.
[0023] In a possible implementation manner, the determination module is configured to perform Fourier transform on initial features of each sample object to obtain transformed features of each sample object; and determine a reference weight of each sample object based on the transformed features of each sample object.
[0024] In a possible implementation manner, the transformation feature of the sample object includes transformation sub-features of the sample object in multiple dimensions;
[0025] The determination module is used to determine, for any sample object, a first feature difference based on a transformation sub-feature of the any sample object in a first dimension and a transformation sub-feature of each sample object in the first dimension, where the first dimension is any dimension of the multiple dimensions; determine a second feature difference based on a transformation sub-feature of the any sample object in a second dimension and a transformation sub-feature of each sample object in the second dimension, where the second dimension is any dimension of the multiple dimensions except the first dimension; determine a transformation sub-feature correlation between the first dimension and the second dimension of any sample object based on the first feature difference and the second feature difference; and determine a reference weight of each sample object based on the transformation sub-feature correlation between each two dimensions of each sample object.
[0026] In a possible implementation, the determination module is used to determine an average transformation sub-feature of the first dimension based on the initial weights of the respective sample objects and the transformation sub-features of the respective sample objects in the first dimension; determine a weighted transformation sub-feature of any one of the sample objects in the first dimension based on the initial weight of any one of the sample objects and the transformation sub-feature of any one of the sample objects in the first dimension; and use a difference between the weighted transformation sub-feature of any one of the sample objects in the first dimension and the average transformation sub-feature of the first dimension as the first feature difference.
[0027] In a possible implementation, the training module is used to determine, for any sample object, a first sub-loss corresponding to the any sample object based on the labeling information of the any sample object and the prediction information of the any sample object; determine a weighted first sub-loss of the any sample object based on the first sub-loss corresponding to the any sample object and a reference weight of the any sample object; determine a first loss based on the weighted first sub-losses of the respective sample objects; and train the neural network model based on the first loss to obtain the event occurrence model.
[0028] In a possible implementation manner, the plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which the multimedia content is not pushed;
[0029] The training module is used to determine the first loss based on the labeling information, prediction information and reference weight of each sample object; determine the second loss based on the initial feature difference between the initial features of the multiple first objects and the initial features of the multiple second objects; and train the neural network model based on the first loss and the second loss to obtain the event occurrence model.
[0030] In one possible implementation, the training module is used to determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determine complexity information based on the parameters of the neural network model, wherein the complexity information characterizes the complexity of the neural network model; determine a third loss based on the complexity information and the number of sample objects; and train the neural network model based on the first loss and the third loss to obtain the event occurrence model.
[0031] In a possible implementation, the training module is used to determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determine a fourth loss based on the reference weight of each sample object and the number of sample objects; and train the neural network model based on the first loss and the fourth loss to obtain the event occurrence model.
[0032] In a fourth aspect, a multimedia content push device is provided, the device comprising:
[0033] An acquisition module, used to acquire data of a reference object;
[0034] a determination module, configured to determine the initial features of the reference object based on the data of the reference object through an event occurrence model, wherein the event occurrence model is obtained by training according to the event occurrence model training method described in the first aspect;
[0035] The determination module is further configured to determine prediction information of the reference object based on initial features of the reference object through the event occurrence model, wherein the prediction information of the reference object represents the possibility of the occurrence of a relevant event related to the multimedia content by the reference object;
[0036] The determination module is further configured to determine whether to push the multimedia content to the reference object based on the prediction information of the reference object.
[0037] In a possible implementation, the prediction information of the reference object includes first prediction information and second prediction information, the first prediction information representing the possibility of the relevant event occurring to the reference object when the multimedia content is pushed to the reference object, and the second prediction information representing the possibility of the relevant event occurring to the reference object when the multimedia content is not pushed to the reference object;
[0038] The determination module is used to push the multimedia content to the reference object if a first condition and a second condition are met, wherein the first condition is met when the first prediction information is greater than the second prediction information, and the second condition is met when the difference between the first prediction information and the second prediction information is greater than a threshold; if at least one of the first condition and the second condition is not met, determine not to push the multimedia content to the reference object.
[0039] In a fifth aspect, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor so that the electronic device implements the training method of the event occurrence model described in the first aspect or the method for pushing multimedia content described in the second aspect.
[0040] In the sixth aspect, a computer-readable storage medium is also provided, in which at least one computer program is stored, and the at least one computer program is loaded and executed by a processor so that the electronic device can implement the training method of the event occurrence model described in the first aspect or the method of pushing multimedia content described in the second aspect.
[0041] In the seventh aspect, a computer program or a computer program product is also provided, wherein at least one computer program is stored in the computer program or the computer program product, and the at least one computer program is loaded and executed by a processor so that the electronic device can implement the training method of the event occurrence model described in the first aspect or the method of pushing multimedia content described in the second aspect.
[0042] The technical solution provided by this application brings at least the following beneficial effects:
[0043] The technical solution provided by the present application extracts features from the data of the sample object through a neural network model to obtain the initial features of the sample object. On the one hand, the prediction information of the sample object is determined based on the initial features of the sample object through the neural network model, and the possibility of the occurrence of related events in the sample object is represented by the prediction information of the sample object, and whether the related events occur in the sample object is represented by the annotation information of the sample object. When the neural network model is trained through the annotation information and prediction information of the sample object, the prediction information output by the model can be made close to the annotation information, so that the event occurrence model can accurately determine the possibility of the occurrence of related events in the reference object. On the other hand, based on the initial features of the sample object, the reference weight of the sample object is determined, and the initial features of the sample object are made to be nonlinearly independent through the reference weight of the sample object. When the neural network model is trained through the reference weight of the sample object, the initial features determined by the model can be made to be close to nonlinear independence, so that the model has strong nonlinear mapping ability and generalization ability, so that the event occurrence model can more accurately determine the possibility of the occurrence of related events in the reference object. By determining the possibility of the occurrence of related events related to multimedia content in the reference object, it is beneficial to improve the conversion rate of multimedia content, thereby improving the push effect of multimedia content. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0045] Figure 1 It is a schematic diagram of an implementation environment of a method for training an event occurrence model or a method for pushing multimedia content provided in an embodiment of the present application;
[0046] Figure 2 It is a flow chart of a method for training an event occurrence model provided in an embodiment of the present application;
[0047] Figure 3 is a flowchart of a method for pushing multimedia content provided by an embodiment of the present application;
[0048] Figure 4 It is a schematic diagram of a cause-effect diagram provided in an embodiment of the present application;
[0049] Figure 5 It is a training framework diagram of an event occurrence model provided in an embodiment of the present application;
[0050] Figure 6 It is a structural schematic diagram of a training device for an event occurrence model provided in an embodiment of the present application;
[0051] Figure 7 It is a structural schematic diagram of a multimedia content push device provided in an embodiment of the present application;
[0052] Figure 8 is a schematic diagram of the structure of a terminal device provided in an embodiment of the present application;
[0053] Fig. 9 It is a structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION
[0054] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.
[0055] Figure 1 is a schematic diagram of an implementation environment of a training method for an event occurrence model or a method for pushing multimedia content provided in an embodiment of the present application, such as Figure 1 As shown, the implementation environment includes a terminal device 101 and a server 102. The event occurrence model training method or multimedia content push method in the embodiment of the present application can be executed by the terminal device 101, can also be executed by the server 102, or can be executed by the terminal device 101 and the server 102 together.
[0056] The terminal device 101 can be a smart phone, a game console, a desktop computer, a tablet computer, a laptop computer, a smart TV, a smart car device, an intelligent voice interaction device, a smart home appliance, etc. The server 102 can be a single server, or a server cluster consisting of multiple servers, or any one of a cloud computing platform and a virtualization center, which is not limited in the embodiments of the present application. The server 102 can be connected to the terminal device 101 through a wired network or a wireless network. The server 102 can have functions such as data processing, data storage, and data transmission and reception, which are not limited in the embodiments of the present application. The number of terminal devices 101 and servers 102 is not limited and can be one or more.
[0057] Each optional embodiment of the present application can be implemented based on artificial intelligence (AI) technology. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.
[0058] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, mechatronics and other technologies. Artificial intelligence software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, as well as machine learning / deep learning, autonomous driving, smart transportation and other major directions.
[0059] With the development of the Internet, the types, quantity, and functions of applications are increasing. Some applications can push multimedia content to users. Generally speaking, applications need to push different multimedia content to different users. Therefore, how to push multimedia content has become a technical problem that needs to be solved urgently.
[0060] The present application embodiment provides a method for training an event occurrence model, which can be applied in the above implementation environment, and can train an event occurrence model, and accurately determine the possibility of a reference object having a related event related to multimedia content through the event occurrence model, which is conducive to improving the conversion rate of multimedia content, thereby improving the push effect of multimedia content. Figure 2 Taking the flowchart of a training method for an event occurrence model provided in the embodiment of the present application as an example, for the convenience of description, the terminal device 101 or the server 102 that executes the training method for the event occurrence model in the embodiment of the present application is referred to as an electronic device, and the method can be executed by an electronic device. Figure 2 As shown, the method includes the following steps.
[0061] Step 201 : acquiring data of a plurality of sample objects and annotation information of each sample object, wherein the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs in the sample object.
[0062] In the embodiment of the present application, the sample object is a user of the first application. The first application has a multimedia content push function and can push multimedia content to the user of the first application. The multimedia content includes at least one of picture content, text content, audio content, etc.
[0063] There are multiple sample objects. For any sample object, the electronic device can obtain the data of the sample object. Optionally, the data of the sample object includes at least one of the basic information of the sample object, the first usage information and the second usage information. The basic information of the sample object is used to describe the attributes of the sample object itself. For example, the basic information of the sample object includes the age, height, gender, etc. of the sample object. The first usage information is the information of the sample object using the first application. Since the first application can push multimedia content to the sample object, the first usage information can include the number of multimedia content watched by the sample object, the duration of the sample object watching multimedia content, etc. The second usage information is the information of the sample object using the second application. For example, the second usage information includes the number of days the sample object uses the second application, the program account level of the sample object in the second application, and the number of times the sample object logs into the second application within a time period. The second application is an application related to multimedia content, and the second application and the first application are the same or different applications. For example, the first application is a video application, and multimedia content related to a game is pushed to the sample object through the video application. In this case, the second application is a game application corresponding to the game.
[0064] In the embodiment of the present application, the electronic device may also obtain the annotation information of any sample object, and characterize whether the sample object has a related event related to the multimedia content through the annotation information. Since the second application is related to the multimedia content, the related event is an event performed by the sample object on the second application. For example, if the second application is a game application, the related event is an event of logging into the game application, or an event of playing a game on the game application.
[0065] Optionally, if a relevant event occurs to the sample object, the labeling information of the sample object is the first data. If no relevant event occurs to the sample object, the labeling information of the sample object is the second data. The embodiment of the present application does not limit the first data and the second data. Exemplarily, the first data is 1 and the second data is 0.
[0066] Step 202: extract features from data of multiple sample objects using a neural network model to obtain initial features of each sample object.
[0067] The embodiments of the present application do not limit the structure, size, etc. of the neural network model. Exemplarily, the neural network model is an initial network model, and the initial network model includes a feature extraction part, a weight determination part, and an event prediction part. The functions of each part are described below and will not be repeated here. Alternatively, the neural network model is a model obtained by training the initial network model at least once according to an implementation principle similar to steps 201 to 205. In this case, the neural network model also includes a feature extraction part, a weight determination part, and an event prediction part, and the parameters of these three parts are different from the parameters of the three parts included in the initial network model.
[0068] The feature extraction part of the neural network model is used to extract features from the data of the sample object. That is, the data of any sample object is input into the feature extraction part of the neural network model, and the feature extraction part extracts features from the data of the sample object to obtain and output the initial features of the sample object.
[0069] In an exemplary embodiment, the feature extraction part is a deep neural network. The deep neural network includes at least one of a fully connected layer and an activation layer. The fully connected layer is at least one layer, and any fully connected layer includes multiple neurons. The activation layer is at least one layer, and any activation layer is an arbitrary activation function, and any activation layer is located before or after any fully connected layer. The fully connected layer is used for convolution processing, and the activation layer is used for activation processing.
[0070] For example, a deep neural network includes 3 fully connected layers and 2 activation layers, each fully connected layer includes 128 neurons, one activation layer is located after the first fully connected layer, and another activation layer is located after the second fully connected layer, and the activation layer is a linear rectifier unit (Linear Rectification Function, ReLU). When the feature extraction part is the aforementioned deep neural network, the data of the sample object is input into the feature extraction part, and after convolution processing through the fully connected layer and activation processing through the activation layer, the output of the last fully connected layer is obtained, that is, the initial features of the sample object are obtained.
[0071] Step 203: Determine prediction information of each sample object based on the initial features of each sample object through a neural network model. The prediction information of the sample object represents the possibility of a related event occurring in the sample object.
[0072] The event prediction part of the neural network model is used to predict the possibility of related events occurring in the sample object. That is, the input of the information prediction part includes the initial features of any sample object, and the input is mapped to output through the information prediction part. The output of the information prediction part is the prediction information of the sample object, that is, the possibility of related events occurring in the sample object. Among them, the prediction information of the sample object is 0 or a positive number, and the prediction information of the sample object is positively correlated with the possibility of related events occurring in the sample object, that is, the greater the prediction information of the sample object, the more likely it is that the related event will occur in the sample object.
[0073] Optionally, the prediction information of the sample object is a probability value greater than or equal to 0 and less than or equal to 1. In this case, the prediction information of the sample object is the probability of the sample object having a related event. Alternatively, the prediction information of the sample object is the first data or the second data. If the prediction information of the sample object is the first data, the possibility of the sample object having a related event is the first data; if the prediction information of the sample object is the second data, the possibility of the sample object having a related event is the second data.
[0074] For example, the first data mentioned above is 1 and the second data is 0. If the prediction information of the sample object is 1, the possibility of the sample object having the relevant event is 100%, that is, the sample object has the relevant event; if the prediction information of the sample object is 0, the possibility of the sample object having the relevant event is 0%, that is, the sample object has not the relevant event.
[0075] In a possible implementation, the plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed. In this case, step 203 includes steps 2031 to 2032 (not shown in the figure).
[0076] In an embodiment of the present application, multiple sample objects are divided into an experimental group and a control group according to information on whether the sample objects are pushed multimedia content. The experimental group includes multiple first objects, any of which is a sample object to which multimedia content is pushed. The control group includes multiple second objects, any of which is a sample object to which multimedia content is not pushed.
[0077] Step 2031, using a neural network model to determine prediction information of each first object based on the first identification information and initial features of multiple first objects, wherein the first identification information represents the situation of the pushed multimedia content. The first identification information is any number or any character, for example, the first identification information is 1.
[0078] Optionally, the event prediction part of the neural network model includes a first prediction network. The first prediction network is used to predict the possibility of a related event occurring to the first object in the experimental group. That is to say, for any first object in the experimental group, the initial features of the first object and the first identification information are spliced to obtain spliced information, and the spliced information is input into the first prediction network, and the prediction information of the first object is output through the first prediction network. Since the first identification information represents the situation where the multimedia content is pushed, and the multimedia content is pushed to the first object, the prediction information of the first object represents the possibility of a related event occurring to the first object when the multimedia content is pushed.
[0079] The embodiment of the present application does not limit the structure, size, etc. of the first prediction network. Exemplarily, the structure of the first prediction network is similar to that of the feature extraction part. The first prediction network includes at least one of a fully connected layer and an activation layer. The fully connected layer is at least one layer, and any fully connected layer includes multiple neurons. The activation layer is at least one layer, and any activation layer is an arbitrary activation function, and any activation layer is located before or after any fully connected layer. The fully connected layer is used for convolution processing, and the activation layer is used for activation processing.
[0080] For example, the first prediction network includes 3 fully connected layers and 2 activation layers, each fully connected layer includes 128 neurons, one activation layer is located after the first fully connected layer, and another activation layer is located after the second fully connected layer. If the prediction information of the first object is data greater than or equal to 0, for example, the prediction information of the first object is a probability value greater than or equal to 0 and less than or equal to 1, the activation layer is a ReLU activation function. If the prediction information of the first object is the first data or the second data, that is, the prediction information of the first object is a binary variable, the activation layer is an S-shaped generating curve, that is, a Sigmoid activation function.
[0081] When the first prediction network includes a fully connected layer and an activation layer, the concatenated information obtained by concatenating the initial features of the first object and the first identification information is input into the first prediction network, and after convolution processing is performed through the fully connected layer and activation processing is performed through the activation layer, the output of the last fully connected layer is obtained, that is, the prediction information of the first object is obtained.
[0082] By determining the prediction information of the first object through the first prediction network, it is possible to determine the possibility of the sample object to have a related event when the multimedia content is pushed. When the neural network model is subsequently trained based on the prediction information of the first object and the annotation information of the first object, the prediction information output by the first prediction network can be gradually approached to the annotation information, so that the first prediction network can accurately predict the possibility of any object to have a related event when the multimedia content is pushed, thereby improving the conversion rate of the multimedia content.
[0083] Step 2032: Determine prediction information of each second object based on the second identification information and initial features of the plurality of second objects through a neural network model, wherein the second identification information represents a situation where the multimedia content is not pushed. The second identification information is any number or any character different from the first identification information, for example, the second identification information is 0.
[0084] Optionally, the event prediction part of the neural network model includes a second prediction network. The second prediction network is used to predict the possibility of a related event occurring to the second object in the control group. That is to say, for any second object in the control group, the initial features of the second object and the second identification information are spliced to obtain spliced information, and the spliced information is input into the second prediction network, and the prediction information of the second object is output through the second prediction network. Since the second identification information represents the situation where the multimedia content is not pushed, and the multimedia content is not pushed to the second object, the prediction information of the second object represents the possibility of a related event occurring to the second object when the multimedia content is not pushed.
[0085] The embodiment of the present application does not limit the structure, size, etc. of the second prediction network. Exemplarily, the structure of the second prediction network is similar to that of the first prediction network. That is, the second prediction network includes at least one of a fully connected layer and an activation layer, the fully connected layer is at least one layer, any fully connected layer includes multiple neurons, the activation layer is at least one layer, any activation layer is an arbitrary activation function, and any activation layer is located before or after any fully connected layer. The fully connected layer is used for convolution processing, and the activation layer is used for activation processing.
[0086] When the second prediction network includes a fully connected layer and an activation layer, the concatenated information obtained by concatenating the initial features of the second object and the second identification information is input into the second prediction network, and after convolution processing is performed through the fully connected layer and activation processing is performed through the activation layer, the output of the last fully connected layer is obtained, that is, the prediction information of the second object is obtained.
[0087] By determining the prediction information of the second object through the second prediction network, it is possible to determine the possibility of the sample object having a related event when the multimedia content is not pushed. When the neural network model is subsequently trained based on the prediction information of the second object and the annotation information of the second object, the prediction information output by the second prediction network can be gradually approached to the annotation information, so that the second prediction network can accurately predict the possibility of any object having a related event when the multimedia content is not pushed, thereby improving the conversion rate of the multimedia content.
[0088] Step 204: Determine a reference weight of each sample object based on the initial features of each sample object. The reference weight of the sample object is used to make the initial features of the sample object tend to be nonlinearly independent.
[0089] The weight determination part of the neural network model is used to determine the reference weight of the sample object. That is, the initial features of each sample object are input into the weight determination part, and the reference weight of each sample object is output through the weight determination part. Among them, the reference weight of the sample object is used to make the initial features of the sample object tend to be nonlinearly independent.
[0090] The initial features of the sample objects tend to be nonlinearly independent, which is equivalent to the initial features gradually tending to: the sub-features of any one dimension cannot be linearly represented by the sub-features of other dimensions, that is, the correlation between the sub-features of any two dimensions is weak, which can improve the stability of the initial features, thereby improving the generalization ability and accuracy of the model.
[0091] In a possible implementation, step 204 includes step A1 to step A2 (not shown in the figure).
[0092] Step A1, performing Fourier transform on the initial features of each sample object to obtain the transformed features of each sample object.
[0093] In the embodiment of the present application, the weight determination part of the neural network model includes a Fourier transform part, and the Fourier transform part is used to perform Fourier transform on the initial features of the sample object. In an exemplary embodiment, the Fourier transform part is a random Fourier features (RFF) transform function, and the initial features of any sample object are subjected to random Fourier transform through the RFF transform function to obtain the transformed features of the sample object. Optionally, the RFF transform function is shown in the following formula (1).
[0094]
[0095] in, represents the RFF transformation function. h represents the mapping method of the RFF transformation function. x represents the initial features of any sample object. Characterizes the transformation characteristics of the sample object, where cos represents cosine, ω is the parameter that obeys normal distribution, and φ is the parameter that obeys uniform distribution. N(0,1) represents normal distribution, Uniform(0,2π) represents standard distribution, and π is the circumference angle corresponding to the semicircle, that is, π = 180 degrees.
[0096] It can be understood that the initial features of the sample object are features of the original space, and the initial features of the sample object have the characteristics of nonlinear correlation. The transformed features of the sample object are features of the Hilbert space, and the transformed features of the sample object have the characteristics of linear correlation. By performing Fourier transform on the initial features of the sample object, it is achieved to map the nonlinearly correlated features in the original space to linearly correlated features in the Hilbert space, so that the transformed features of the sample object remain linearly correlated in the Hilbert space, which is conducive to the subsequent measurement of the correlation of the transformed features, so that the reference weight determined based on the correlation can make the initial features of the sample object tend to be nonlinearly independent, thereby improving the generalization ability of the event occurrence model.
[0097] Step A2: determining a reference weight of each sample object based on the transformation feature of each sample object.
[0098] In the embodiment of the present application, the reference weight of each sample object can be determined through the transformation features of each sample object, so that after the transformation features of any sample object are weighted based on the reference weight of the sample object, the correlation between the weighted transformation features of the sample object and the average value of the weighted transformation features of all sample objects is small. In general, the average value of the weighted transformation features of all sample objects can be called the mathematical expectation of the weighted transformation features, and the smaller the correlation between the weighted transformation features of the sample object and the mathematical expectation, the more independent the weighted transformation features of the sample object tend to be. Since the transformation features of the sample objects have the characteristic of linear correlation, the reference weights of the sample objects are used to make the transformation features of the sample objects tend to be linearly independent.
[0099] It can be understood that by performing Fourier transform on the initial features of the sample objects to obtain the transformed features of the sample objects, it is possible to map the initial features of the sample objects that are nonlinearly correlated in the original space to the transformed features of the sample objects that are linearly correlated in the Hilbert space. Therefore, the initial features of the sample objects have the characteristics of nonlinear correlation, based on which the reference weights of the sample objects are used to make the initial features of the sample objects tend to be nonlinearly independent.
[0100] In a possible implementation, the transformation feature of the sample object includes transformation sub-features of the sample object in multiple dimensions. In this case, step A2 includes steps A21 to A24 (not shown in the figure).
[0101] In the embodiment of the present application, the data of the sample object includes sub-data of the sample object in multiple dimensions, wherein the data of the sample object includes at least one of basic information of the sample object, first usage information and second usage information.
[0102] The basic information of the sample object includes sub-data of at least one dimension. For example, the basic information of the sample object includes the age, height, gender, etc. of the sample object, the age of the sample object is sub-data of one dimension, the height of the sample object is sub-data of another dimension, and the gender of the sample object is sub-data of yet another dimension.
[0103] Similarly, the first usage information includes sub-data of at least one dimension. For example, the first usage information includes the number of multimedia contents watched by the sample object, the duration of the multimedia contents watched by the sample object, etc. The number of multimedia contents watched by the sample object is sub-data of one dimension, and the duration of the multimedia contents watched by the sample object is sub-data of another dimension.
[0104] The second usage information includes sub-data of at least one dimension. For example, the second usage information includes the number of days the sample object uses the second application, the program account level of the sample object in the second application, and the number of times the sample object logs into the second application within a time period. Among them, the number of days the sample object uses the second application is sub-data of one dimension, the program account level of the sample object in the second application is sub-data of another dimension, and the number of times the sample object logs into the second application within a time period is sub-data of yet another dimension.
[0105] After extracting features from the data of the sample object to obtain the initial features of the sample object, the initial features of the sample object include initial sub-features of the sample object in each dimension, wherein the initial sub-features of the sample object in any dimension are obtained by extracting features from the sub-data of the sample object in the dimension. After mapping the initial features of the sample object to the transformed features of the sample object, the transformed features of the sample object include transformed sub-features of the sample object in each dimension, wherein the transformed sub-features of the sample object in any dimension are obtained by mapping the initial sub-features of the sample object in the dimension.
[0106] Step A21, for any sample object, based on the transformation sub-feature of any sample object in the first dimension and the transformation sub-features of each sample object in the first dimension, determine a first feature difference, where the first dimension is any dimension of multiple dimensions.
[0107] In the embodiment of the present application, any one of the multiple dimensions is used as the first dimension. The transformation sub-features of any sample object in the first dimension can be extracted from the transformation features of the sample object. The transformation sub-features of each sample object in the first dimension are averaged, weighted averaged, etc., and the obtained operation result is used as the average transformation sub-feature of the first dimension. The difference between the transformation sub-feature of the sample object in the first dimension and the average transformation sub-feature of the first dimension is used as the first feature difference, and the first feature difference reflects the difference between the transformation sub-feature of the sample object in the first dimension and the transformation sub-features of each sample object in the first dimension.
[0108] Optionally, step A21 includes steps A211 to A213 (not shown in the figure).
[0109] Step A211: determining the average transformed sub-feature of the first dimension based on the initial weight of each sample object and the transformed sub-feature of each sample object in the first dimension.
[0110] For any sample object, the set initial weight of the sample object can be obtained, or the initial weight of the sample object can be randomly generated. Based on the initial weight of the sample object, the transformation sub-feature of the sample object in the first dimension is weighted to obtain the weighted transformation sub-feature of the sample object in the first dimension. The weighted transformation sub-features of each sample object in the first dimension are averaged, and the average value obtained is the average transformation sub-feature of the first dimension.
[0111] Alternatively, the average transformed sub-feature of the first dimension can be expressed as: Where n is the number of sample objects, ∑ is the summation symbol, ω j is the initial weight of the jth sample object, u(A j ) is the transformed sub-feature of the j-th sample object in the first dimension A.
[0112] Step A212: determining a weighted transformed sub-feature of any sample object in the first dimension based on the initial weight of any sample object and the transformed sub-feature of any sample object in the first dimension.
[0113] In the embodiment of the present application, the transformation sub-feature of the sample object in the first dimension may be weighted based on the initial weight of the sample object to obtain the weighted transformation sub-feature of the sample object in the first dimension. Optionally, the weighted transformation sub-feature of the sample object in the first dimension may be expressed as: i u(A i ). Among them, ω i is the initial weight of the i-th sample object, u(A i ) is the transformed sub-feature of the i-th sample object in the first dimension A.
[0114] Step A213: taking the difference between the weighted transformed sub-feature of any sample object in the first dimension and the average transformed sub-feature of the first dimension as the first feature difference.
[0115] In the embodiment of the present application, the weighted transformation sub-feature of any sample object in the first dimension minus the average transformation sub-feature of the first dimension can be used as the first feature difference. Optionally, the first feature difference can be expressed as: Among them, the meaning of each parameter in the first characteristic difference has been described accordingly in step A211 and step A212, and will not be repeated here.
[0116] Step A22, determining a second feature difference based on the transformed sub-feature of any sample object in the second dimension and the transformed sub-features of each sample object in the second dimension, where the second dimension is any dimension of the multiple dimensions except the first dimension.
[0117] In the embodiment of the present application, any dimension other than the first dimension among the multiple dimensions is used as the second dimension. The transformation sub-feature of the sample object in the second dimension can be extracted from the transformation feature of any sample object. The transformation sub-features of each sample object in the second dimension are averaged, weighted averaged, etc., and the obtained operation result is used as the average transformation sub-feature of the second dimension. The difference between the transformation sub-feature of the sample object in the second dimension and the average transformation sub-feature of the second dimension is used as the second feature difference, and the second feature difference reflects the difference between the transformation sub-feature of the sample object in the second dimension and the transformation sub-feature of each sample object in the second dimension.
[0118] In a possible implementation, step A22 includes steps A221 to A223 (not shown in the figure).
[0119] Step A221: Determine the average transformed sub-feature of the second dimension based on the initial weight of each sample object and the transformed sub-feature of each sample object in the second dimension.
[0120] For any sample object, based on the initial weight of the sample object, the transformation sub-feature of the sample object in the second dimension is weighted to obtain the weighted transformation sub-feature of the sample object in the second dimension. The weighted transformation sub-features of each sample object in the second dimension are averaged, and the average value obtained is the average transformation sub-feature of the second dimension.
[0121] Alternatively, the average transformed sub-feature of the second dimension can be expressed as: Where n is the number of sample objects, ∑ is the summation symbol, ω j is the initial weight of the jth sample object, v(B j) is the transformed sub-feature of the j-th sample object in the second dimension B.
[0122] Step A222: Determine the weighted transformed sub-feature of any sample object in the second dimension based on the initial weight of any sample object and the transformed sub-feature of any sample object in the second dimension.
[0123] In the embodiment of the present application, the transformation sub-feature of the sample object in the second dimension may be weighted based on the initial weight of the sample object to obtain the weighted transformation sub-feature of the sample object in the second dimension. Optionally, the weighted transformation sub-feature of the sample object in the second dimension may be expressed as: i v(B i ). Among them, ω i is the initial weight of the i-th sample object, v(B i ) is the transformed sub-feature of the i-th sample object in the second dimension B.
[0124] Step A223: taking the difference between the weighted transformation sub-feature of any sample object in the second dimension and the average transformation sub-feature of the second dimension as the second feature difference.
[0125] In the embodiment of the present application, the weighted transformation sub-feature of any sample object in the second dimension minus the average transformation sub-feature of the second dimension can be used as the second feature difference. Optionally, the second feature difference can be expressed as: Among them, the meaning of each parameter in the second characteristic difference has been described accordingly in step A221 and step A222, and will not be repeated here.
[0126] Step A23: determining the transformation sub-feature correlation between the first dimension and the second dimension of any sample object based on the first feature difference and the second feature difference.
[0127] In the embodiment of the present application, any calculation such as mean square error calculation and root mean square error calculation can be performed on the first feature difference and the second feature difference, and the calculation result obtained is used as the transformation sub-feature correlation between the first dimension and the second dimension of the sample object. Optionally, the transposed matrix of the first feature difference is determined, and the transposed matrix of the first feature difference and the second feature difference are multiplied, and the product obtained is used as the transformation sub-feature correlation between the first dimension and the second dimension of the sample object, that is, the transformation sub-feature correlation can be expressed as: Among them, T is the symbol of the transposed matrix, and the meanings of the remaining parameters are described in steps A211, A212, A221, and A222, which will not be repeated here.
[0128] Step A24: determining a reference weight of each sample object based on the correlation of the transformed sub-features of each sample object between every two dimensions.
[0129] In the embodiment of the present application, the correlation of the transformed sub-features between the first dimension and the second dimension of each sample object can be averaged, and the result of the operation is used as the average correlation of the transformed sub-features between the first dimension and the second dimension. Optionally, the average correlation of the transformed sub-features between the first dimension and the second dimension is determined according to the formula (2) shown below.
[0130]
[0131] in, is the average transformation sub-feature correlation between the first dimension A and the second dimension B. The meanings of the remaining parameters are described in steps A211, A212, A221, and A222, and will not be repeated here.
[0132] By the above method, the average transformation sub-feature correlation between every two dimensions can be determined. Based on the average transformation sub-feature correlation between every two dimensions, the reference weight of each sample object is determined. Optionally, the reference weight of each sample object is determined according to the formula (3) shown below.
[0133]
[0134] Among them, ω * Characterizes the reference weight of each sample object. Φ * Characterize the weighted transformation features of each sample object, where the weighted transformation features of any sample object are obtained by weighting the transformation features of the sample object based on the reference weight of the sample object. argmin is the value of the independent variable that minimizes the reference function, where the reference function is The independent variable is ω * and Φ * ∑ is the summation symbol, and m is the dimension of the transformed features of the sample object. is the average transformation sub-feature correlation between dimension i and second dimension j, which is equivalent to the formula (2)
[0135] In the embodiment of the present application, the above formula (2) and formula (3) are based on the implementation principle of the Hilbert-Schmidt Independence Criterion (HSIC) to determine the average transformation sub-feature correlation between two dimensions. The average transformation sub-feature correlation is also called the Hilbert covariance operator and the HSIC index. The Hilbert covariance operator is used to measure the correlation of the average transformation sub-features of the two dimensions in the Hilbert space. Based on this, in formula (3) is the Hilbert norm after Fourier transformation.
[0136] As mentioned above, the initial features of the sample object have the characteristics of nonlinear correlation, while the transformed features of the sample object obtained after Fourier transforming the initial features of the sample object have the characteristics of linear transformation. The HSIC index can intuitively measure the correlation between the average transformed sub-features of two dimensions in the linearly correlated transformed features, and the HSIC index can indirectly measure the correlation between the average transformed sub-features of two dimensions in the nonlinearly correlated initial features, so that the reference weight can be determined. The reference weight makes the correlation between the average transformed sub-features of two dimensions in the initial features as small as possible (that is, as close to 0 as possible). In other words, the reference weight can make the initial sub-features of two dimensions in the initial features approach nonlinear independence.
[0137] By using the reference weights to make the initial sub-features of each two dimensions in the initial features approach nonlinear independence, the initial features can be approached to nonlinear independence. Generally speaking, the closer the features are to nonlinear independence, the higher the stability of the features and the stronger the generalization ability of the model. Therefore, by training the neural network model with the reference weights to obtain the event occurrence model, the event occurrence model can output features with high stability and improve the generalization ability of the event occurrence model.
[0138] Step 205: Based on the labeling information, prediction information and reference weight of each sample object, the neural network model is trained to obtain an event occurrence model, which is used to determine the possibility of related events occurring in the reference object.
[0139] In an embodiment of the present application, the loss of the neural network model can be determined based on the labeling information, prediction information and reference weights of each sample object. The neural network model is trained based on the loss of the neural network model to obtain a trained neural network model. If the trained neural network model meets the training end condition, the trained neural network model is determined to be an event occurrence model. If the trained neural network model does not meet the training end condition, the trained neural network model is determined to be the neural network model for the next training, and the neural network model is trained for the next time in the manner of steps 201 to 205 until the trained neural network model meets the training end condition and the trained neural network model is determined to be an event occurrence model.
[0140] The embodiments of the present application do not limit the trained neural network model to satisfy the training end condition. Exemplarily, the trained neural network model satisfies the training end condition when: the number of training times corresponding to the trained neural network model reaches the number threshold, or the loss of the trained neural network model is within a set range.
[0141] The embodiment of the present application does not limit the method for determining the loss of the neural network model. Exemplarily, according to any one of implementation methods 1 to 4 as shown below, the loss of the neural network model is determined, and the neural network model is trained based on the loss of the neural network model.
[0142] In implementation mode 1, step 205 includes at least one item of steps B1 to B3 (not shown in the figure).
[0143] Step B1, for any sample object, based on the labeling information of any sample object and the prediction information of any sample object, determine the first sub-loss corresponding to any sample object, and based on the first sub-loss corresponding to any sample object and the reference weight of any sample object, determine the weighted first sub-loss of any sample object.
[0144] In an embodiment of the present application, any error calculation such as mean square error, absolute value error, relative entropy, etc. can be performed on the annotation information of any sample object and the prediction information of the sample object, and the error calculation result obtained is the first sub-loss corresponding to the sample object. Optionally, if the annotation information of the sample object is a discrete variable, the first sub-loss can be a cross entropy loss, and if the annotation information of the sample object is a continuous variable, the first sub-loss can be a mean square error loss. Then, based on the reference weight of the sample object, the first sub-loss corresponding to the sample object is weighted to obtain a weighted first sub-loss of the sample object.
[0145] Optionally, the weighted first sub-loss of the sample object can be expressed as: ω i *l h [(Φ(x i ),T i ),Y i ],ω i is the reference weight of the i-th sample object, l h [(Φ(x i ),T i ),Y i ] is the first sub-loss of the i-th sample object. Among them, l h [(Φ(x i ),T i ),Y i ]=L[h(Φ(x i ),T i ),Y i ], L is the sign of the first sub-loss, h(Φ(x i ),T i ) is the prediction information of the i-th sample object, Φ(x i ) represents the initial features of the i-th sample object, T iCharacterizes the sample identification information of the i-th sample object. If the sample object is pushed multimedia content, then T i is the first identification information of the i-th sample object. If the sample object is not pushed multimedia content, then T i Y is the second identification information of the i-th sample object. i Represents the annotation information of the i-th sample object.
[0146] Step B2: determining a first loss based on the weighted first sub-loss of each sample object.
[0147] In the embodiment of the present application, the weighted first sub-loss of each sample object can be summed up, averaged, etc., and the result of the operation is used as the first loss. Optionally, the first loss can be expressed as: Wherein, n represents the number of sample objects, and the meanings of the remaining parameters are described in step B1, which will not be repeated here.
[0148] Step B3: training the neural network model based on the first loss to obtain an event occurrence model.
[0149] In the embodiment of the present application, the first loss can be used as the loss of the neural network model, or the first loss can be weighted and the weighted result can be used as the loss of the neural network model. The neural network model is trained based on the loss of the neural network model to obtain an event occurrence model.
[0150] It can be understood that training the neural network model based on the first loss is to adjust the parameters of the neural network model based on the first loss so that the parameters of the neural network model are optimized in the direction of reducing the first loss. That is, h*=argminL[h(Φ(x i ),T i ),Y i ], where h * represents the adjusted parameters of the neural network model, argmin is the value of the independent variable that minimizes the reference function, where the reference function is L[h(Φ(x i ),T i ),Y i ], the independent variable is h, and the meanings of the remaining parameters are described in step B1, which will not be repeated here.
[0151] The neural network model includes a feature extraction part, a weight determination part and an event prediction part, and the event prediction part includes a first prediction network and a second prediction network. The parameters of the neural network model are optimized in the direction of reducing the first loss, which is equivalent to optimizing the parameters of the first prediction network and the parameters of the second prediction network in the direction of reducing the first loss.
[0152] In implementation mode 2, step 205 includes at least one item of steps C1 to C3 (not shown in the figure).
[0153] Step C1, determining a first loss based on the labeling information, prediction information and reference weight of each sample object. The implementation of step C1 can be seen in the description of steps B1 to B2, which will not be repeated here.
[0154] Step C2, determining a second loss based on the initial feature difference between the initial features of the plurality of first objects and the initial features of the plurality of second objects. The plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed. The contents of the first objects and the second objects can be seen in the description of step 203, which will not be repeated here.
[0155] In an embodiment of the present application, the first initial feature group includes initial features of each first object, and the second initial feature group includes initial features of each second object. The first initial feature group is subtracted from the second initial feature group to obtain an initial feature difference. The second loss is determined based on the initial feature difference according to the second loss function. The embodiment of the present application does not limit the second loss function. For example, the second loss function can be any one of a mean square error loss function, an absolute value error loss function, a relative entropy loss function, and the like.
[0156] Optionally, the second loss function is an integer probability metric (IPM) function. The IPM function is used to measure the distribution distance between two distributions, and the IPM function is shown in the following formula (4).
[0157] IPM G (p,q)=sup g∈G |∫ s g(s)[p(s)-q(s)]ds| Formula (4)
[0158] Among them, IPM G (p,g) is the IPM indicator, which can measure the distribution distance between distribution p(s) and distribution q(s). G represents a function group, and the function group includes at least one set function, wherein the function group includes function g(s). sup is the abbreviation of Supremum (i.e., the least upper bound). s is the independent variable of the function. The IPM function can be understood as: select function g(s) from the function group G, and when the distribution distance between distribution p(s) and distribution q(s) is measured according to function g(s), the distribution distance is the smallest (for example, close to 0). In other words, function g(s) can make distribution p(s) approach distribution q(s).
[0159] In the embodiment of the present application, when the second loss function is an IPM function, a reference function can be selected from a set function group, and the reference function is used to make the first initial feature group approach the second initial feature group. According to the reference function, the distribution distance between the first initial feature group and the second initial feature group is determined, wherein the distribution distance is the second loss.
[0160] Alternatively, the second loss can be expressed as: IPM G (p Φ(0) ,p Φ(1) ). Among them, p Φ(0) Characterize the second initial feature set, that is, p Φ(0) =Φ T=0 (X), T=0 indicates that no multimedia content has been pushed, and at this time, X indicates each second object. Φ(1) Characterize the first initial feature set, i.e., p Φ(1) =Φ T=1 (X), T=1 indicates that no multimedia content has been pushed, and at this time, X indicates each first object.
[0161] Step C3: training the neural network model based on the first loss and the second loss to obtain an event occurrence model.
[0162] In an embodiment of the present application, the first loss and the second loss may be summed, weighted summed, or the like, and the result of the operation obtained is the loss of the neural network model. The neural network model is trained using the loss of the neural network model to obtain an event occurrence model.
[0163] It can be understood that the second loss can measure the initial feature distance between the first object and the second object. Training the neural network model based on the second loss is to adjust the parameters of the neural network model based on the second loss so that the parameters of the neural network model are optimized in the direction of reducing the second loss, that is, the parameters of the neural network model are optimized in the direction of reducing the initial feature distance between the first object and the second object. By optimizing the initial features of the sample object, it is beneficial to optimize the reference weight of the sample object and the weighted transformation features of the sample object, so that the initial features of the sample object tend to be nonlinearly independent. That is, ω * ,Φ * =argminIPM[Φ T=0 (X),Φ Φ=1 (X)], where ω * The reference weight that characterizes the sample object, Φ * Characterize the weighted transformation features of the sample object, argmin is the value of the independent variable that minimizes the reference function, where the reference function is IPM[Φ T=0 (X),Φ T=1 (X)], the independent variable is ω *,Φ * , IPM[Φ T=0 (X),Φ T=1 (X)] characterizes the second loss.
[0164] In implementation mode 3, step 205 includes at least one item of steps D1 to D4 (not shown in the figure).
[0165] Step D1, determining a first loss based on the labeling information, prediction information and reference weight of each sample object. The implementation of step D1 can be seen in the description of steps B1 to B2, which will not be repeated here.
[0166] Step D2, determining complexity information based on the parameters of the neural network model, wherein the complexity information represents the complexity of the neural network model.
[0167] In an embodiment of the present application, at least one of the number of parameters and the number of calculations can be determined based on the parameters of the neural network model. Complexity information is determined based on at least one of the number of parameters and the number of calculations to measure the complexity of the neural network model through the complexity information. For example, the complexity information can be the number of parameters or the number of calculations, or the complexity information is obtained by weighting the number of parameters and the number of calculations.
[0168] Step D3, determining a third loss based on the complexity information and the number of sample objects.
[0169] The embodiment of the present application does not limit the determination method of the third loss. Exemplarily, the ratio between the complexity information and the number of sample objects is the third loss, or the third loss is determined based on the complexity information, the number of sample objects and the set weight. Optionally, the third loss is: Among them, λ h is to set the weight, R(h) is the complexity information, and n is the number of sample objects.
[0170] Step D4, based on the first loss and the third loss, the neural network model is trained to obtain an event occurrence model.
[0171] In an embodiment of the present application, the first loss and the third loss may be summed, weighted summed, or the like, and the result of the operation obtained is the loss of the neural network model. The neural network model is trained using the loss of the neural network model to obtain an event occurrence model.
[0172] In the embodiment of the present application, the third loss is positively correlated with the complexity of the neural network model and negatively correlated with the number of sample objects. The neural network model is trained by the third loss so that the parameters of the neural network model are optimized in the direction of reducing the third loss, which is conducive to reducing the complexity of the model and improving the computing and processing capabilities of the model.
[0173] In implementation mode 4, step 205 includes at least one item of steps E1 to E3 (not shown in the figure).
[0174] Step E1, determining a first loss based on the labeling information, prediction information and reference weight of each sample object. The implementation of step E1 can be seen in the description of step B1 to step B2, which will not be repeated here.
[0175] Step E2: determining a fourth loss based on the reference weight of each sample object and the number of sample objects.
[0176] In the embodiment of the present application, the reference weight of each sample object is subjected to norm calculation to obtain a calculation result, and the calculation result is used to describe the statistical information of the reference weight of each sample object. When the 0-norm calculation is performed, the calculation result is used to describe the number of non-zero reference weights in the reference weight of each sample object. When the 1-norm calculation is performed, the calculation result is used to describe the sum of the absolute values of the reference weights of each sample object. When the 2-norm calculation is performed, the calculation result is used to describe the 1 / 2 power of the sum of the squares of the reference weights of each sample object.
[0177] Next, based on the calculation result and the number of sample objects, a fourth loss is determined. The embodiment of the present application does not limit the calculation method of the fourth loss. For example, the fourth loss is the ratio between the calculation result and the number of sample objects, or the fourth loss is obtained by weighting the ratio between the calculation result and the number of sample objects. Optionally, the fourth loss is: Among them, λ ω To set the weight, ‖ω‖2 is the calculation result, ω is the reference weight set, the reference weight set includes the reference weight of each sample object, ‖‖2 is the 2-norm sign, and n is the number of sample objects.
[0178] Step E3: training the neural network model based on the first loss and the fourth loss to obtain an event occurrence model.
[0179] In an embodiment of the present application, the first loss and the fourth loss may be summed, weighted summed, or the like, and the result of the operation obtained is the loss of the neural network model. The neural network model is trained using the loss of the neural network model to obtain an event occurrence model.
[0180] In the embodiment of the present application, the fourth loss is positively correlated with the statistical information of the reference weights of each sample object, and negatively correlated with the number of sample objects. The neural network model is trained by the fourth loss, so that the parameters of the neural network model are optimized in the direction of reducing the fourth loss, so that the reference weights of each sample object gradually tend to be the same, so that the initial features gradually tend to be nonlinearly independent, and the model's computing processing ability, generalization ability and accuracy are improved.
[0181] It can be understood that the loss of the neural network model is determined based on at least one of the first loss, the second loss, the third loss and the fourth loss. Optionally, the loss of the neural network model is as shown in formula (5) below.
[0182]
[0183] Among them, L(h,Φ,ω) represents the loss of the neural network model, h is the parameter of the neural network model, Φ represents the initial features, and ω represents the reference weight. For the first loss, the meaning of each parameter can be found in the description of implementation method 1 and will not be repeated here. The meaning of each parameter can be found in the description of implementation mode 3, and will not be repeated here. G (p Φ(0) ,p Φ(1) ) is the second loss. The meaning of each parameter can be found in the description of implementation method 2 and will not be repeated here. α is the weight of the second loss. This is the fourth loss. The meaning of each parameter can be found in the description of implementation method 4 and will not be repeated here.
[0184] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards in the relevant regions. For example, the data of sample objects and the annotation information of sample objects involved in this application are all obtained with full authorization.
[0185] The above method extracts features from the data of the sample object through a neural network model to obtain the initial features of the sample object. On the one hand, the prediction information of the sample object is determined based on the initial features of the sample object through the neural network model, and the possibility of the occurrence of related events in the sample object is represented by the prediction information of the sample object, and whether the related events occur in the sample object is represented by the annotation information of the sample object. When the neural network model is trained through the annotation information and prediction information of the sample object, the prediction information output by the model can be made close to the annotation information, so that the event occurrence model can accurately determine the possibility of the occurrence of related events in the reference object. On the other hand, based on the initial features of the sample object, the reference weight of the sample object is determined, and the initial features of the sample object are made to be nonlinearly independent through the reference weight of the sample object. When the neural network model is trained through the reference weight of the sample object, the initial features determined by the model can be made to be close to nonlinear independence, so that the model has strong nonlinear mapping ability and generalization ability, so that the event occurrence model can more accurately determine the possibility of the occurrence of related events in the reference object. By determining the possibility of the occurrence of related events related to multimedia content in the reference object, it is beneficial to improve the conversion rate of multimedia content, thereby improving the push effect of multimedia content.
[0186] The embodiment of the present application provides a method for pushing multimedia content. The method can be applied to the above implementation environment. The possibility of a reference object having a related event related to the multimedia content can be accurately determined through an event occurrence model, which is conducive to improving the conversion rate of the multimedia content, thereby improving the pushing effect of the multimedia content. Figure 3 Taking the flowchart of a multimedia content push method provided by the embodiment of the present application as an example, for the convenience of description, the terminal device 101 or the server 102 that executes the multimedia content push method in the embodiment of the present application is referred to as an electronic device, and the method can be executed by an electronic device. Figure 3 As shown, the method includes the following steps.
[0187] Step 301, obtaining data of a reference object.
[0188] The implementation of step 301 is similar to the content of the "sample object data" in step 201, and can be seen in the description of step 201, which will not be repeated here.
[0189] Step 302: Determine the initial features of the reference object based on the data of the reference object through the event occurrence model. Figure 2 The event occurrence model is trained using a related training method.
[0190] The implementation of step 302 is similar to the content of “initial features of the sample object” in step 202 , and can be seen in the description of step 202 , which will not be repeated here.
[0191] Step 303 : Determine prediction information of the reference object based on the initial features of the reference object through an event occurrence model, where the prediction information of the reference object represents the possibility of the reference object having a relevant event related to the multimedia content.
[0192] The implementation of step 303 is similar to the content of "prediction information of sample object" in step 203, and can be seen in the description of step 203, which will not be repeated here.
[0193] Step 304: determine whether to push multimedia content to the reference object based on the prediction information of the reference object.
[0194] Since the prediction information of the reference object can reflect the possibility of the reference object undergoing relevant events related to the multimedia content, and the conversion rate of the multimedia content is positively correlated with the possibility of the reference object undergoing relevant events, that is, the higher the possibility of the reference object undergoing relevant events, the higher the conversion rate of the multimedia content, therefore, the prediction information of the reference object can also reflect the conversion rate of the multimedia content.
[0195] If the prediction information of the reference object is greater than the reference threshold, it is determined to push the multimedia content to the reference object, so as to make the reference object have relevant events as much as possible and improve the conversion rate of the multimedia content. If the prediction information of the reference object is less than the reference threshold, it is determined not to push the multimedia content to the reference object, so as to avoid blindly pushing the multimedia content and reduce the push cost of the multimedia content.
[0196] It is understandable that if the prediction information of the reference object is equal to the reference threshold, the multimedia content may be pushed to the reference object, or the multimedia content may not be pushed to the reference object, which is not limited here.
[0197] The embodiment of the present application does not limit the method for determining the reference threshold. Exemplarily, the reference threshold is a numerical value determined based on experience. Alternatively, feature extraction is performed on the data of each sample object through an event occurrence model to obtain sample features of each sample object. Based on the sample features of each sample object, the event occurrence model determines the sample information of each sample object, and the sample information of the sample object represents the possibility of a related event occurring in the sample object. Among them, each sample object includes a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed. By performing statistical processing on the sample information of each first object and the sample information of each second object, a critical value of the sample information is obtained, and the critical value is the reference threshold. Among them, the sample features of the sample object and the initial features of the sample object are determined in a similar manner, as can be seen in the description of step 202, and the sample information of the sample object and the predicted information of the sample object are determined in a similar manner, as can be seen in the description of step 203, which will not be repeated here.
[0198] In a possible implementation, the prediction information of the reference object includes first prediction information and second prediction information, the first prediction information represents the possibility of the reference object having a related event when the reference object is pushed multimedia content, and the second prediction information represents the possibility of the reference object having a related event when the reference object is not pushed multimedia content. In this case, step 304 includes steps 3041 to 3042 (not shown in the figure).
[0199] In the embodiment of the present application, the event occurrence model includes a first prediction network and a second prediction network. In the process of training the event occurrence model, the first identification information and the initial features of the first object are input into the first prediction network, and the prediction information of the first object is output through the first prediction network. Since the first identification information represents the situation of the pushed multimedia content, the first prediction network can continuously learn the possibility of the first object having a related event when the first object is pushed multimedia content.
[0200] Based on this, the first identification information and the initial features of the reference object are input into the first prediction network of the event occurrence model, and the first prediction information of the reference object is output through the first prediction network, and the first prediction information can characterize the possibility of the reference object having a related event when the reference object is pushed multimedia content. It can be understood that the method for determining the first prediction information is similar to the method for determining the prediction information of the first object, which can be seen in the description of step 2031 and will not be repeated here.
[0201] Similarly, in the process of training the event occurrence model, the second identification information and the initial features of the second object are input into the second prediction network, and the prediction information of the second object is output through the second prediction network. Since the second identification information represents the situation where the multimedia content is not pushed, the second prediction network can continuously learn the possibility of the second object having a related event when the multimedia content is not pushed to the second object.
[0202] Based on this, the second identification information and the initial features of the reference object are input into the second prediction network of the event occurrence model, and the second prediction information of the reference object is output through the second prediction network, and the second prediction information can characterize the possibility of the reference object having a related event when the multimedia content is not pushed to the reference object. It can be understood that the method for determining the second prediction information is similar to the method for determining the prediction information of the second object, which can be seen in the description of step 2032 and will not be repeated here.
[0203] Step 3041: If the first condition and the second condition are met, the multimedia content is pushed to the reference object. The first condition is met when the first prediction information is greater than the second prediction information, and the second condition is met when the difference between the first prediction information and the second prediction information is greater than a threshold.
[0204] If the first prediction information is greater than the second prediction information, it means that the reference object is more likely to have a related event when the multimedia content is pushed than the reference object when the multimedia content is not pushed. When the multimedia content is pushed to the reference object, the multimedia content is more likely to be converted. In addition, if the difference between the first prediction information and the second prediction information is greater than a threshold, it means that when the multimedia content is pushed to the reference object, the conversion rate of the multimedia content is significantly improved. Therefore, when the first prediction information is greater than the second prediction information, and the difference between the first prediction information and the second prediction information is greater than a threshold, it is determined to push the multimedia content to the reference object.
[0205] The embodiment of the present application does not limit the method for determining the threshold. Exemplarily, the threshold is a numerical value determined based on experience. Alternatively, feature extraction is performed on the data of each sample object through an event occurrence model to obtain sample features of each sample object. Based on the sample features of each sample object, the event occurrence model determines the sample information of each sample object, and the sample information of the sample object represents the possibility of a related event occurring in the sample object. Among them, each sample object includes a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed. By performing statistical processing on the sample information of each first object, a range of sample information (recorded as a first range) is obtained; by performing statistical processing on the sample information of each second object, a range of sample information (recorded as a second range) is obtained. The threshold is determined based on the first range and the second range. Among them, the sample features of the sample object and the initial features of the sample object are determined in a similar manner, as can be seen in the description of step 202, and the sample information of the sample object and the predicted information of the sample object are determined in a similar manner, as can be seen in the description of step 203, which will not be repeated here.
[0206] Step 3042: if at least one of the first condition and the second condition is not satisfied, it is determined not to push the multimedia content to the reference object.
[0207] If the first prediction information is less than the second prediction information, it means that the possibility of the relevant event occurring when the reference object is not pushed multimedia content is higher than the possibility of the relevant event occurring when the reference object is pushed multimedia content. In other words, when multimedia content is not pushed to the reference object, the multimedia content is more likely to be converted. In addition, if the difference between the first prediction information and the second prediction information is less than a threshold value, it means that even if the multimedia content is pushed to the reference object, the conversion rate of the multimedia content will not be significantly improved. Therefore, when the first prediction information is less than or equal to the second prediction information, and / or the difference between the first prediction information and the second prediction information is less than or equal to a threshold value, it is determined not to push the multimedia content to the reference object.
[0208] It can be understood that the first prediction information being equal to the second prediction information and the difference between the first prediction information and the second prediction information being equal to the threshold are two special cases. In one possible implementation, if the first prediction information is greater than or equal to the second prediction information and the difference between the first prediction information and the second prediction information is greater than or equal to the threshold, it is determined to push the multimedia content to the reference object.
[0209] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards in the relevant regions. For example, the data of the reference objects involved in this application are all obtained with full authorization.
[0210] The event occurrence model in the above method can accurately determine the possibility of the occurrence of relevant events in the reference object. By determining whether to push multimedia content to the reference object based on the possibility of the occurrence of relevant events related to the multimedia content in the reference object, it is beneficial to improve the conversion rate of multimedia content, thereby improving the push effect of multimedia content.
[0211] The above describes the training method of the event occurrence model or the method of pushing multimedia content in the embodiment of the present application from the perspective of method steps. The following is a systematic and comprehensive description. Among them, the training method of the event occurrence model and the method of pushing multimedia content are both applicable to the scenario of causal relationship. In the scenario of causal relationship, there are three types of variables, which are confounding variables, intervention variables and outcome variables.
[0212] Confounder (X) refers to a variable that affects both the intervention variable and the outcome variable. For example, Figure 4 The causal diagram shown can reflect the causal relationship between the variables, and the arrow from A to B indicates that A affects B. Based on this, it can be seen that the confounding variable will affect both the intervention variable and the outcome variable. In the embodiment of the present application, the confounding variable is the object data, which can be the data of the sample object mentioned above, or the data of the reference object mentioned above.
[0213] Treatment (T) refers to a variable that is affected by the confounding variable and affects the outcome variable. Figure 4 In the causal diagram shown, the intervening variable is affected by the confounding variable, and the intervening variable will also affect the outcome variable. In the embodiment of the present application, the intervening variable is whether the object is pushed multimedia content. In other words, the "situation of being pushed multimedia content" and "situation of not being pushed multimedia content" mentioned above are both intervening variables.
[0214] The outcome variable (Outcome, Y) is a variable that is affected by both confounding variables and intervening variables. Figure 4In the causal diagram shown, the outcome variable is affected by the confounding variable and the intervention variable. In the embodiment of the present application, the outcome variable is whether the subject has a related event related to the multimedia content. In general, if the intervention variable is "the subject is pushed multimedia content", the intervention variable can be recorded as "T = 1". At this time, the outcome variable affected by the confounding variable and the intervention variable is "ITE = Y (T = 1, X = x)", and if the intervention variable is "the subject is not pushed multimedia content", the intervention variable can be recorded as "T = 0". At this time, the outcome variable affected by the confounding variable and the intervention variable is "ITE = Y (T = 0, X = x)". Among them, x represents the object data, and ITE represents the individual causal effect (Individual Treatment Effect, ITE).
[0215] Based on the above causal relationship scenario, the present application embodiment provides a training framework diagram of an event occurrence model, such as Figure 5 As shown. In the embodiment of the present application, the neural network model can be trained to obtain an event occurrence model. The neural network model includes a feature extraction part, a weight determination part and an event prediction part, and the event occurrence model also includes a feature extraction part, a weight determination part and an event prediction part. Each part included in the event occurrence model and the part included in the neural network model have different parameters. For example, the feature extraction part included in the event occurrence model and the feature extraction part included in the neural network model have different parameters.
[0216] In the embodiment of the present application, object data can be obtained, and the object data corresponds to the data of the sample object mentioned above. The object data is input into the feature extraction part included in the neural network model, and the feature extraction part extracts the features of the object data to obtain initial features, and the initial features correspond to the initial features of the sample object mentioned above. The content of feature extraction can be seen in the relevant description of step 202, which will not be repeated here.
[0217] On the one hand, the initial features are input into the event prediction part included in the neural network model, and the first identification information or the second identification information is input into the event prediction part. Among them, the first identification information represents the situation of the multimedia content being pushed. If the initial features and the first identification information are input into the event prediction part, the prediction information is determined through the first prediction network, and the prediction information represents the possibility of the occurrence of related events when the multimedia content is pushed. The second identification information represents the situation of the multimedia content not being pushed. If the initial features and the second identification information are input into the event prediction part, the prediction information is determined through the second prediction network, and the prediction information represents the possibility of the occurrence of related events when the multimedia content is not pushed. Among them, the method for determining the prediction information can be seen in the relevant description of step 203, which will not be repeated here.
[0218] On the other hand, the initial features are input into the weight determination part included in the neural network model, and the initial features are Fourier transformed through the Fourier transform part in the weight determination part to obtain transformation features, which correspond to the transformation features of the sample object mentioned above. Then, by performing correlation calculation on the transformation features, a reference weight is obtained and output, which corresponds to the reference weight of the sample object mentioned above. The content of Fourier transform can be seen in the relevant description of step A1, and the content of correlation calculation can be seen in the relevant description of step A2, which will not be repeated here.
[0219] Next, the annotation information may be obtained, and the annotation information indicates whether the relevant event occurs. The first loss may be determined based on the prediction information, the annotation information, and the reference weight. The determination method of the first loss may refer to the relevant description of implementation method 1, and will not be repeated here.
[0220] It is also possible to extract the initial features corresponding to the first identification information, and to extract the initial features corresponding to the second identification information. The distribution distance between the initial features corresponding to the first identification information and the initial features corresponding to the second identification information is determined by the initial features corresponding to the first identification information and the initial features corresponding to the second identification information, and the distribution distance is the second loss. The initial features corresponding to the first identification information are the initial features of the first object mentioned above, and the initial features corresponding to the second identification information are the initial features of the second object mentioned above. The method for determining the second loss can be found in the relevant description of implementation method 2, which will not be repeated here.
[0221] Based on the first loss and the second loss, the neural network model is trained to obtain an event occurrence model. The training method can be seen in the relevant description of step 205, which will not be repeated here.
[0222] After obtaining the event occurrence model, object data can be obtained, and the object data corresponds to the data of the reference object mentioned above. The object data is input into the feature extraction part included in the event occurrence model, and the feature extraction part extracts the features of the object data to obtain initial features, which correspond to the initial features of the reference object mentioned above. The content of feature extraction can be seen in the relevant description of step 202, and the implementation principles of the two are similar, which will not be repeated here.
[0223] Next, the initial features are input into the event prediction part included in the event occurrence model. In addition, the first identification information and the second identification information are also input into the event prediction part. Among them, the first identification information represents the situation of the multimedia content being pushed. The first prediction information is determined based on the initial features and the first identification information through the first prediction network of the event prediction part. The first prediction information represents the possibility of the occurrence of related events when the multimedia content is pushed. The second identification information represents the situation of the multimedia content not being pushed. The second prediction information is determined based on the initial features and the second identification information through the second prediction network of the event prediction part. The second prediction information represents the possibility of the occurrence of related events when the multimedia content is not pushed. Among them, the determination method of the first prediction information and the second prediction information can be seen in the relevant description of step 203. The implementation principles of the two are similar and will not be repeated here.
[0224] If the first prediction information is greater than the second prediction information, and the difference between the first prediction information and the second prediction information is greater than a threshold, the multimedia content is pushed; if the first prediction information is not greater than the second prediction information, and / or the difference between the first prediction information and the second prediction information is not greater than a threshold, the multimedia content is not pushed. This part of the content can be seen in the description of step 3041 to step 3042, and will not be repeated here.
[0225] Figure 6 FIG. 1 is a schematic diagram of a structure of a training device for an event occurrence model provided in an embodiment of the present application. Figure 6 As shown, the device comprises:
[0226] An acquisition module 601 is used to acquire data of multiple sample objects and annotation information of each sample object, where the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs in the sample object;
[0227] A feature extraction module 602 is used to extract features from data of multiple sample objects through a neural network model to obtain initial features of each sample object;
[0228] A determination module 603 is used to determine prediction information of each sample object based on the initial features of each sample object through a neural network model, where the prediction information of the sample object represents the possibility of a related event occurring in the sample object;
[0229] The determination module 603 is further used to determine the reference weight of each sample object based on the initial features of each sample object, and the reference weight of the sample object is used to make the initial features of the sample object tend to be nonlinearly independent;
[0230] The training module 604 is used to train the neural network model based on the labeling information, prediction information and reference weight of each sample object to obtain an event occurrence model, and the event occurrence model is used to determine the possibility of related events occurring in the reference object.
[0231] In a possible implementation, the plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed;
[0232] The determination module 603 is used to determine the prediction information of each first object based on the first identification information and the initial features of multiple first objects through a neural network model, and the first identification information represents the situation where the multimedia content is pushed; and to determine the prediction information of each second object based on the second identification information and the initial features of multiple second objects through a neural network model, and the second identification information represents the situation where the multimedia content is not pushed.
[0233] In a possible implementation, the determination module 603 is configured to perform Fourier transform on the initial features of each sample object to obtain a transformed feature of each sample object; and determine a reference weight of each sample object based on the transformed feature of each sample object.
[0234] In a possible implementation, the transformation feature of the sample object includes transformation sub-features of the sample object in multiple dimensions;
[0235] The determination module 603 is used to determine, for any sample object, a first feature difference based on the transformation sub-features of any sample object in a first dimension and the transformation sub-features of each sample object in the first dimension, where the first dimension is any dimension of multiple dimensions; determine a second feature difference based on the transformation sub-features of any sample object in a second dimension and the transformation sub-features of each sample object in the second dimension, where the second dimension is any dimension of multiple dimensions except the first dimension; determine a correlation between the transformation sub-features of any sample object in the first dimension and the second dimension based on the first feature difference and the second feature difference; and determine a reference weight of each sample object based on the correlation between the transformation sub-features of each sample object between every two dimensions.
[0236] In a possible implementation, the determination module 603 is used to determine the average transformation sub-feature of the first dimension based on the initial weights of the respective sample objects and the transformation sub-features of the respective sample objects in the first dimension; determine the weighted transformation sub-feature of any sample object in the first dimension based on the initial weight of any sample object and the transformation sub-feature of any sample object in the first dimension; and use the difference between the weighted transformation sub-feature of any sample object in the first dimension and the average transformation sub-feature of the first dimension as the first feature difference.
[0237] In one possible implementation, the training module 604 is used to determine, for any sample object, a first sub-loss corresponding to any sample object based on the labeling information of any sample object and the prediction information of any sample object; determine a weighted first sub-loss of any sample object based on the first sub-loss corresponding to any sample object and a reference weight of any sample object; determine a first loss based on the weighted first sub-losses of each sample object; and train a neural network model based on the first loss to obtain an event occurrence model.
[0238] In a possible implementation, the plurality of sample objects include a plurality of first objects to which multimedia content is pushed and a plurality of second objects to which multimedia content is not pushed;
[0239] The training module 604 is used to determine the first loss based on the labeling information, prediction information and reference weight of each sample object; determine the second loss based on the initial feature difference between the initial features of multiple first objects and the initial features of multiple second objects; and train the neural network model based on the first loss and the second loss to obtain an event occurrence model.
[0240] In one possible implementation, the training module 604 is used to determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determine complexity information based on the parameters of the neural network model, the complexity information characterizing the complexity of the neural network model; determine a third loss based on the complexity information and the number of sample objects; and train the neural network model based on the first loss and the third loss to obtain an event occurrence model.
[0241] In one possible implementation, the training module 604 is used to determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determine a fourth loss based on the reference weight of each sample object and the number of sample objects; and train a neural network model based on the first loss and the fourth loss to obtain an event occurrence model.
[0242] The above-mentioned device extracts features from the data of the sample object through a neural network model to obtain the initial features of the sample object. On the one hand, the prediction information of the sample object is determined based on the initial features of the sample object through the neural network model, and the possibility of the occurrence of related events of the sample object is represented by the prediction information of the sample object, and whether the related events occur in the sample object is represented by the annotation information of the sample object. When the neural network model is trained through the annotation information and prediction information of the sample object, the prediction information output by the model can be made close to the annotation information, so that the event occurrence model can accurately determine the possibility of the occurrence of related events of the reference object. On the other hand, based on the initial features of the sample object, the reference weight of the sample object is determined, and the initial features of the sample object are made to be nonlinearly independent through the reference weight of the sample object. When the neural network model is trained through the reference weight of the sample object, the initial features determined by the model can be made to be close to nonlinearly independent, so that the model has strong nonlinear mapping ability and generalization ability, so that the event occurrence model can more accurately determine the possibility of the occurrence of related events of the reference object. By determining the possibility of the occurrence of related events related to multimedia content in the reference object, it is beneficial to improve the conversion rate of multimedia content, thereby improving the push effect of multimedia content.
[0243] It should be understood that the above Figure 6 When the device provided realizes its functions, only the division of the above-mentioned functional modules is used as an example for illustration. In practical applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.
[0244] Figure 7 FIG. 1 is a schematic diagram of a structure of a multimedia content push device provided in an embodiment of the present application. Figure 7 As shown, the device comprises:
[0245] An acquisition module 701 is used to acquire data of a reference object;
[0246] The determination module 702 is used to determine the initial features of the reference object based on the data of the reference object through the event occurrence model. The event occurrence model is based on Figure 2 The event occurrence model is trained by a related training method;
[0247] The determination module 702 is further configured to determine prediction information of the reference object based on the initial features of the reference object through the event occurrence model, wherein the prediction information of the reference object represents the possibility of the reference object having a relevant event related to the multimedia content;
[0248] The determination module 702 is further configured to determine whether to push multimedia content to the reference object based on the prediction information of the reference object.
[0249] In a possible implementation, the prediction information of the reference object includes first prediction information and second prediction information, the first prediction information represents the possibility of the reference object having a related event when the multimedia content is pushed to the reference object, and the second prediction information represents the possibility of the reference object having a related event when the multimedia content is not pushed to the reference object;
[0250] Determination module 702 is used to push multimedia content to the reference object if a first condition and a second condition are met, the first condition being met when the first prediction information is greater than the second prediction information, and the second condition being met when the difference between the first prediction information and the second prediction information is greater than a threshold; if at least one of the first condition and the second condition is not met, determine not to push the multimedia content to the reference object.
[0251] The event occurrence model in the above device can accurately determine the possibility of the occurrence of relevant events in the reference object. By determining whether to push multimedia content to the reference object based on the possibility of the occurrence of relevant events related to the multimedia content in the reference object, it is beneficial to improve the conversion rate of multimedia content, thereby improving the push effect of multimedia content.
[0252] It should be understood that the above Figure 7 When the device provided realizes its functions, only the division of the above-mentioned functional modules is used as an example for illustration. In practical applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.
[0253] Figure 8 The structure block diagram of a terminal device 800 provided by an exemplary embodiment of the present application is shown. The terminal device 800 includes: a processor 801 and a memory 802 .
[0254] The processor 801 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 801 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 801 may also include a main processor and a coprocessor. The main processor is a processor for processing data in an awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 801 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 801 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.
[0255] The memory 802 may include one or more computer-readable storage media, which may be non-transitory. The memory 802 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 802 is used to store at least one computer program, which is used to be executed by the processor 801 to implement the training method of the event occurrence model or the push method of multimedia content provided in the method embodiment of the present application.
[0256] In some embodiments, the terminal device 800 may further optionally include: a peripheral device interface 803 and at least one peripheral device. The processor 801, the memory 802 and the peripheral device interface 803 may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral device interface 803 via a bus, a signal line or a circuit board. Specifically, the peripheral device includes: at least one of a radio frequency circuit 804, a display screen 805, a camera assembly 806, an audio circuit 807 and a power supply 808.
[0257] The peripheral device interface 803 may be used to connect at least one peripheral device related to I / O (Input / Output) to the processor 801 and the memory 802. In some embodiments, the processor 801, the memory 802, and the peripheral device interface 803 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 801, the memory 802, and the peripheral device interface 803 may be implemented on a separate chip or circuit board, which is not limited in this embodiment.
[0258] The radio frequency circuit 804 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 804 communicates with communication networks and other communication devices through electromagnetic signals. The radio frequency circuit 804 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. Optionally, the radio frequency circuit 804 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and the like. The radio frequency circuit 804 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: the World Wide Web, a metropolitan area network, an intranet, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency circuit 804 may also include circuits related to NFC (Near Field Communication), which is not limited in this application.
[0259] The display screen 805 is used to display the UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 805 is a touch display screen, the display screen 805 also has the ability to collect touch signals on the surface or above the surface of the display screen 805. The touch signal can be input to the processor 801 as a control signal for processing. At this time, the display screen 805 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, the display screen 805 can be one, set on the front panel of the terminal device 800; in other embodiments, the display screen 805 can be at least two, respectively set on different surfaces of the terminal device 800 or in a folding design; in other embodiments, the display screen 805 can be a flexible display screen, set on the curved surface or folding surface of the terminal device 800. Even, the display screen 805 can also be set to a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 805 can be made of materials such as LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode, organic light-emitting diode).
[0260] The camera assembly 806 is used to capture images or videos. Optionally, the camera assembly 806 includes a front camera and a rear camera. Typically, the front camera is arranged on the front panel of the terminal, and the rear camera is arranged on the back of the terminal. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize panoramic shooting and VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 806 may also include a flash. The flash can be a monochrome temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.
[0261] The audio circuit 807 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals and input them into the processor 801 for processing, or input them into the radio frequency circuit 804 to achieve voice communication. For the purpose of stereo acquisition or noise reduction, there may be multiple microphones, which are respectively arranged at different parts of the terminal device 800. The microphone may also be an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 801 or the radio frequency circuit 804 into sound waves. The speaker may be a traditional film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 807 may also include a headphone jack.
[0262] The power supply 808 is used to power various components in the terminal device 800. The power supply 808 can be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 808 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery charged through a wired line, and a wireless rechargeable battery is a battery charged through a wireless coil. The rechargeable battery can also be used to support fast charging technology.
[0263] In some embodiments, the terminal device 800 further includes one or more sensors 809 , including but not limited to: an acceleration sensor 811 , a gyroscope sensor 812 , a pressure sensor 813 , an optical sensor 814 , and a proximity sensor 815 .
[0264] The acceleration sensor 811 can detect the magnitude of acceleration on the three coordinate axes of the coordinate system established by the terminal device 800. For example, the acceleration sensor 811 can be used to detect the components of gravity acceleration on the three coordinate axes. The processor 801 can control the display screen 805 to display the user interface in a horizontal view or a vertical view according to the gravity acceleration signal collected by the acceleration sensor 811. The acceleration sensor 811 can also be used for collecting game or user motion data.
[0265] The gyroscope sensor 812 can detect the body direction and rotation angle of the terminal device 800, and the gyroscope sensor 812 can cooperate with the acceleration sensor 811 to collect the user's 3D actions on the terminal device 800. The processor 801 can implement the following functions based on the data collected by the gyroscope sensor 812: motion sensing (such as changing the UI according to the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.
[0266] The pressure sensor 813 can be set on the side frame of the terminal device 800 and / or the lower layer of the display screen 805. When the pressure sensor 813 is set on the side frame of the terminal device 800, it can detect the user's holding signal of the terminal device 800, and the processor 801 performs left and right hand recognition or shortcut operation according to the holding signal collected by the pressure sensor 813. When the pressure sensor 813 is set on the lower layer of the display screen 805, the processor 801 controls the operability controls on the UI interface according to the user's pressure operation on the display screen 805. The operability controls include at least one of a button control, a scroll bar control, an icon control, and a menu control.
[0267] The optical sensor 814 is used to collect the ambient light intensity. In one embodiment, the processor 801 can control the display brightness of the display screen 805 according to the ambient light intensity collected by the optical sensor 814. Specifically, when the ambient light intensity is high, the display brightness of the display screen 805 is increased; when the ambient light intensity is low, the display brightness of the display screen 805 is reduced. In another embodiment, the processor 801 can also dynamically adjust the shooting parameters of the camera assembly 806 according to the ambient light intensity collected by the optical sensor 814.
[0268] The proximity sensor 815, also called a distance sensor, is usually arranged on the front panel of the terminal device 800. The proximity sensor 815 is used to collect the distance between the user and the front of the terminal device 800. In one embodiment, when the proximity sensor 815 detects that the distance between the user and the front of the terminal device 800 is gradually decreasing, the processor 801 controls the display screen 805 to switch from the screen-on state to the screen-off state; when the proximity sensor 815 detects that the distance between the user and the front of the terminal device 800 is gradually increasing, the processor 801 controls the display screen 805 to switch from the screen-off state to the screen-on state.
[0269] Those skilled in the art will understand that Figure 8 The structure shown in the figure does not constitute a limitation on the terminal device 800, and may include more or fewer components than shown in the figure, or combine certain components, or adopt a different component arrangement.
[0270] Fig. 9This is a schematic diagram of the structure of the server provided in the embodiment of the present application. The server 900 may have relatively large differences due to different configurations or performances, and may include one or more processors 901 and one or more memories 902, wherein the one or more memories 902 store at least one computer program, and the at least one computer program is loaded and executed by the one or more processors 901 to implement the event occurrence model training method or multimedia content push method provided in the above-mentioned various method embodiments. Exemplarily, the processor 901 is a CPU. Of course, the server 900 may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server 900 may also include other components for implementing device functions, which will not be repeated here.
[0271] In an exemplary embodiment, a computer-readable storage medium is also provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to enable an electronic device to implement any of the above-mentioned event occurrence model training methods or multimedia content push methods.
[0272] Optionally, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.
[0273] In an exemplary embodiment, a computer program or a computer program product is also provided, in which at least one computer program is stored, and the at least one computer program is loaded and executed by a processor to enable an electronic device to implement any of the above-mentioned event occurrence model training methods or multimedia content push methods.
[0274] It should be understood that the "plurality" mentioned in this article refers to two or more. "And / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.
[0275] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.
[0276] The above description is only an exemplary embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for training an event occurrence model, characterized in that: The method comprises: Acquire data of a plurality of sample objects and annotation information of each sample object, wherein the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs to the sample object; Extracting features from the data of the plurality of sample objects by using a neural network model to obtain initial features of each sample object; Determining prediction information of each sample object based on initial features of each sample object by using the neural network model, wherein the prediction information of the sample object represents the possibility of the related event occurring to the sample object; Based on the initial features of the respective sample objects, determining reference weights of the respective sample objects, wherein the reference weights of the sample objects are used to make the initial features of the sample objects tend to be nonlinearly independent; Based on the labeling information, prediction information and reference weights of each sample object, the neural network model is trained to obtain an event occurrence model, and the event occurrence model is used to determine the possibility of the related event occurring in the reference object.
2. The method according to claim 1, characterized in that The plurality of sample objects include a plurality of first objects to which the multimedia content is pushed and a plurality of second objects to which the multimedia content is not pushed; Determining prediction information of each sample object based on the initial features of each sample object by using the neural network model includes: Determining, by the neural network model, prediction information of each of the first objects based on first identification information and initial features of the plurality of first objects, wherein the first identification information represents a situation in which the multimedia content is pushed; The prediction information of each second object is determined by the neural network model based on the second identification information and the initial features of the plurality of second objects, wherein the second identification information represents a situation where the multimedia content is not pushed.
3. The method according to claim 1, characterized in that The determining, based on the initial features of the respective sample objects, the reference weights of the respective sample objects comprises: Performing Fourier transform on the initial features of each sample object to obtain transformed features of each sample object; Based on the transformation features of the respective sample objects, a reference weight of the respective sample objects is determined.
4. The method according to claim 3, characterized in that The transformation feature of the sample object includes transformation sub-features of the sample object in multiple dimensions; The determining, based on the transformation features of the respective sample objects, the reference weights of the respective sample objects comprises: For any sample object, determining a first feature difference based on a transformation sub-feature of the any sample object in a first dimension and a transformation sub-feature of each sample object in the first dimension, where the first dimension is any dimension of the multiple dimensions; determining a second feature difference based on the transformation sub-feature of any one of the sample objects in a second dimension and the transformation sub-features of each of the sample objects in the second dimension, where the second dimension is any one of the multiple dimensions except the first dimension; Determining, based on the first feature difference and the second feature difference, a transformation sub-feature correlation between the first dimension and the second dimension of any sample object; Based on the transformation sub-feature correlation between each two dimensions of each sample object, a reference weight of each sample object is determined.
5. The method according to claim 4, characterized in that The determining the first feature difference based on the transformation sub-feature of any one of the sample objects in the first dimension and the transformation sub-features of each of the sample objects in the first dimension includes: determining an average transformed sub-feature of the first dimension based on the initial weights of the respective sample objects and the transformed sub-features of the respective sample objects in the first dimension; Determining a weighted transformation sub-feature of any one of the sample objects in the first dimension based on an initial weight of any one of the sample objects and a transformation sub-feature of any one of the sample objects in the first dimension; The difference between the weighted transformed sub-feature of any sample object in the first dimension and the average transformed sub-feature of the first dimension is used as the first feature difference.
6. The method according to claim 1, characterized in that The step of training the neural network model based on the labeling information, prediction information and reference weight of each sample object to obtain an event occurrence model includes: For any sample object, determine a first sub-loss corresponding to the any sample object based on the labeling information of the any sample object and the prediction information of the any sample object, and determine a weighted first sub-loss of the any sample object based on the first sub-loss corresponding to the any sample object and a reference weight of the any sample object; Determine a first loss based on the weighted first sub-losses of the respective sample objects; The neural network model is trained based on the first loss to obtain the event occurrence model.
7. The method according to claim 1, characterized in that The plurality of sample objects include a plurality of first objects to which the multimedia content is pushed and a plurality of second objects to which the multimedia content is not pushed; The step of training the neural network model based on the labeling information, prediction information and reference weight of each sample object to obtain an event occurrence model includes: Determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determining a second loss based on an initial feature difference between the initial features of the plurality of first objects and the initial features of the plurality of second objects; Based on the first loss and the second loss, the neural network model is trained to obtain the event occurrence model.
8. The method according to claim 1, characterized in that The step of training the neural network model based on the labeling information, prediction information and reference weight of each sample object to obtain an event occurrence model includes: Determine a first loss based on the labeling information, prediction information and reference weight of each sample object; Determining complexity information based on parameters of the neural network model, wherein the complexity information represents the complexity of the neural network model; Determining a third loss based on the complexity information and the number of the sample objects; Based on the first loss and the third loss, the neural network model is trained to obtain the event occurrence model.
9. The method according to claim 1, characterized in that: The step of training the neural network model based on the labeling information, prediction information and reference weight of each sample object to obtain an event occurrence model includes: Determine a first loss based on the labeling information, prediction information and reference weight of each sample object; determining a fourth loss based on the reference weights of the respective sample objects and the number of the sample objects; Based on the first loss and the fourth loss, the neural network model is trained to obtain the event occurrence model.
10. A method for pushing multimedia content, characterized in that: The method comprises: Get the data of the reference object; Determining the initial features of the reference object based on the data of the reference object through an event occurrence model, wherein the event occurrence model is obtained by training according to the event occurrence model training method according to any one of claims 1 to 9; Determining prediction information of the reference object based on initial features of the reference object through the event occurrence model, wherein the prediction information of the reference object represents the possibility of the occurrence of a relevant event related to the multimedia content by the reference object; Based on the prediction information of the reference object, it is determined whether to push the multimedia content to the reference object.
11. The method according to claim 10, characterized in that The prediction information of the reference object includes first prediction information and second prediction information, the first prediction information representing the possibility of the relevant event occurring to the reference object when the multimedia content is pushed to the reference object, and the second prediction information representing the possibility of the relevant event occurring to the reference object when the multimedia content is not pushed to the reference object; The determining whether to push the multimedia content to the reference object based on the prediction information of the reference object includes: If a first condition and a second condition are met, pushing the multimedia content to the reference object, the first condition being met when the first prediction information is greater than the second prediction information, and the second condition being met when the difference between the first prediction information and the second prediction information is greater than a threshold; If at least one of the first condition and the second condition is not satisfied, it is determined not to push the multimedia content to the reference object.
12. A training device for an event occurrence model, characterized in that: The device comprises: An acquisition module, used to acquire data of multiple sample objects and annotation information of each sample object, wherein the annotation information of the sample object indicates whether a relevant event related to the multimedia content occurs to the sample object; A feature extraction module, used to extract features from the data of the plurality of sample objects through a neural network model to obtain initial features of each sample object; A determination module, configured to determine prediction information of each sample object based on initial features of each sample object through the neural network model, wherein the prediction information of the sample object represents the possibility of the related event occurring to the sample object; The determination module is further used to determine the reference weight of each sample object based on the initial features of each sample object, wherein the reference weight of the sample object is used to make the initial features of the sample object tend to be nonlinearly independent; The training module is used to train the neural network model based on the labeling information, prediction information and reference weights of each sample object to obtain an event occurrence model, and the event occurrence model is used to determine the possibility of the related event occurring in the reference object.
13. A multimedia content push device, characterized in that: The device comprises: An acquisition module, used to acquire data of a reference object; a determination module, configured to determine the initial features of the reference object based on the data of the reference object through an event occurrence model, wherein the event occurrence model is obtained by training according to the event occurrence model training method according to any one of claims 1 to 9; The determination module is further configured to determine prediction information of the reference object based on initial features of the reference object through the event occurrence model, wherein the prediction information of the reference object represents the possibility of the occurrence of a relevant event related to the multimedia content by the reference object; The determination module is further configured to determine whether to push the multimedia content to the reference object based on the prediction information of the reference object.
14. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor so that the electronic device implements the training method of the event occurrence model as described in any one of claims 1 to 9 or implements the method for pushing multimedia content as described in any one of claims 10 to 11.
15. A computer-readable storage medium, characterized in that: At least one computer program is stored in the computer-readable storage medium, and the at least one computer program is loaded and executed by the processor to enable the electronic device to implement the training method of the event occurrence model as described in any one of claims 1 to 9 or the method for pushing multimedia content as described in any one of claims 10 to 11.
16. A computer program product, characterized in that The computer program product stores at least one computer program, and the at least one computer program is loaded and executed by a processor so that the electronic device implements the training method of the event occurrence model as described in any one of claims 1 to 9 or implements the method for pushing multimedia content as described in any one of claims 10 to 11.
