Click rate prediction method, device, equipment and computer program product

Abstract

The embodiment of the invention provides a click rate prediction method and device, equipment and a computer program product. The prediction method comprises the following steps: acquiring a click rate prediction request for a preset object; based on the click rate prediction request, determining a neural network model for realizing click rate prediction operation, the neural network model corresponding to model parameters; mapping the model parameters to a topological space, and performing Riemannian optimization on the model parameters to obtain optimized model parameters; updating the neural network model based on the optimized model parameters to obtain an optimized model; and performing a click rate prediction operation based on the optimized model to obtain a click rate prediction result corresponding to the preset object. In the embodiment of the invention, the optimization operation of the neural network model is realized in a Riemannian optimization mode, and then the click rate prediction operation is performed by using the optimized model, so that time and space resources required by the click rate prediction operation can be saved, and the prediction accuracy of the click rate can be ensured.

Description

Technical Field

[0001] The present invention relates to the field of network technology, and in particular to a method, device, equipment and computer program product for predicting click rate. Background Art

[0002] Click-through rate refers to the ratio of the number of times a certain content on a website page is clicked to the number of times it is displayed, that is, clicks / views. It is a percentage that can reflect the degree of attention paid to a certain content on a web page, and it can be used to measure the attractiveness of an advertisement. At present, the click-through rate is often obtained by statistically calculating the number of clicks and the number of displays, which not only takes a lot of time and resources, but also reduces the accuracy and efficiency of click-through rate prediction. Summary of the invention

[0003] The embodiments of the present invention provide a method, apparatus, device and computer program product for predicting click-through rate, which can save time and space resources required for click-through rate prediction operations and also ensure the prediction accuracy of click-through rate.

[0004] In a first aspect, an embodiment of the present invention provides a method for predicting a click rate, comprising:

[0005] Get click rate prediction request for preset objects;

[0006] Based on the click-through rate prediction request, determining a neural network model for implementing a click-through rate prediction operation, wherein the neural network model corresponds to model parameters;

[0007] Mapping the model parameters to the topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters;

[0008] Update the neural network model based on the optimized model parameters to obtain an optimized model;

[0009] A click rate prediction operation is performed based on the optimized model to obtain a click rate prediction result corresponding to the preset object.

[0010] In a second aspect, an embodiment of the present invention provides a click rate prediction device, comprising:

[0011] A first acquisition module, used to acquire a click rate prediction request for a preset object;

[0012] A first determination module is used to determine a neural network model for implementing a click-through rate prediction operation based on the click-through rate prediction request, wherein the neural network model corresponds to a model parameter;

[0013] A first processing module is used to map the model parameters to a topological space and perform Riemann optimization on the model parameters to obtain optimized model parameters;

[0014] The first processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model;

[0015] The first processing module is further used to perform a click-through rate prediction operation based on the optimized model to obtain a click-through rate prediction result corresponding to the preset object.

[0016] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the click-through rate prediction method in the above-mentioned first aspect.

[0017] In a fourth aspect, an embodiment of the present invention provides a computer storage medium for storing a computer program, wherein the computer program enables a computer to implement the click rate prediction method in the first aspect when executed.

[0018] In a fifth aspect, an embodiment of the present invention provides a computer program product, comprising: a computer-readable storage medium storing computer instructions, which, when executed by one or more processors, causes the one or more processors to execute the steps in the click-through rate prediction method shown in the first aspect above.

[0019] In a sixth aspect, an embodiment of the present invention provides a sample identification method, including:

[0020] Obtain samples to be identified;

[0021] Determining a neural network model for implementing a sample recognition operation, wherein the neural network model corresponds to model parameters;

[0022] Mapping the model parameters to the topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters;

[0023] Update the neural network model based on the optimized model parameters to obtain an optimized model;

[0024] The sample to be identified is processed based on the optimized model to obtain an identification result.

[0025] In a seventh aspect, an embodiment of the present invention provides a sample identification device, including:

[0026] The second acquisition module is used to acquire the sample to be identified;

[0027] A second determination module is used to determine a neural network model for implementing a sample recognition operation, wherein the neural network model corresponds to model parameters;

[0028] A second processing module is used to map the model parameters to the topological space and perform Riemann optimization on the model parameters to obtain optimized model parameters;

[0029] The second processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model;

[0030] The second processing module is further used to process the sample to be identified based on the optimized model to obtain an identification result.

[0031] In an eighth aspect, an embodiment of the present invention provides an electronic device, comprising: a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein when the one or more computer instructions are executed by the processor, the sample identification method in the above-mentioned sixth aspect is implemented.

[0032] In a ninth aspect, an embodiment of the present invention provides a computer storage medium for storing a computer program, wherein the computer program enables a computer to implement the sample identification method in the sixth aspect when executed.

[0033] In the tenth aspect, an embodiment of the present invention provides a computer program product, comprising: a computer-readable storage medium storing computer instructions, which, when executed by one or more processors, causes the one or more processors to execute the steps in the sample identification method shown in the sixth aspect above.

[0034] In an eleventh aspect, an embodiment of the present invention provides an information recommendation method, comprising:

[0035] Obtaining an object to be recommended, wherein the object to be recommended has corresponding portrait information;

[0036] Determining a neural network model for implementing an advertisement recommendation operation, wherein the neural network model corresponds to model parameters;

[0037] Mapping the model parameters and the portrait information to a topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters;

[0038] Update the neural network model based on the optimized model parameters to obtain an optimized model;

[0039] Based on the optimized model and the portrait information, recommendation information matching the object to be recommended is determined.

[0040] In a twelfth aspect, an embodiment of the present invention provides an information recommendation device, including:

[0041] The third acquisition module is used to acquire the object to be recommended, and the object to be recommended has corresponding portrait information;

[0042] A third determination module is used to determine a neural network model for implementing an advertisement recommendation operation, wherein the neural network model corresponds to model parameters;

[0043] A third processing module is used to map the model parameters and the portrait information to a topological space, and perform Riemann optimization on the model parameters to obtain optimized model parameters;

[0044] The third processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model;

[0045] The third processing module is further used to determine recommendation information matching the object to be recommended based on the optimized model and the portrait information.

[0046] In the thirteenth aspect, an embodiment of the present invention provides an electronic device, comprising: a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the information recommendation method in the above-mentioned eleventh aspect.

[0047] In a fourteenth aspect, an embodiment of the present invention provides a computer storage medium for storing a computer program, wherein the computer program enables a computer to implement the information recommendation method in the eleventh aspect when executed.

[0048] In the fifteenth aspect, an embodiment of the present invention provides a computer program product, comprising: a computer-readable storage medium storing computer instructions, which, when the computer instructions are executed by one or more processors, causes the one or more processors to execute the steps of the information recommendation method in the above-mentioned eleventh aspect.

[0049] The click-through rate prediction method, apparatus, device and computer program product provided in the present embodiment obtains a click-through rate prediction request for a preset object, and then determines a neural network model for implementing a click-through rate prediction operation based on the click-through rate prediction request, and then maps the model parameters corresponding to the neural network model to the topological space, and performs Riemann optimization on the model parameters, so as to obtain the optimized model parameters; after obtaining the optimized model parameters, the neural network model can be updated based on the optimized model parameters to obtain the optimized model; then, the click-through rate prediction operation can be performed based on the optimized model to obtain a click-through rate prediction result corresponding to the preset object, which not only saves the time resources and space resources required for the click-through rate prediction operation, but also ensures the prediction accuracy and efficiency of the click-through rate, further improving the practicality of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 A schematic diagram of a scenario of a click rate prediction method provided by an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of a flow chart of a click rate prediction method provided by an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of a process for performing Riemann optimization on the model parameters to obtain optimized model parameters provided in an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the principle of determining an optimized model based on the optimized model parameters provided in an embodiment of the present invention;

[0055] Figure 5 A schematic diagram of a flow chart of a sample identification method provided by an embodiment of the present invention;

[0056] Figure 6 A flowchart of an information recommendation method provided by an embodiment of the present invention;

[0057] Figure 7 A schematic diagram of the structure of a click rate prediction device provided by an embodiment of the present invention;

[0058] Figure 8 For Figure 7A schematic diagram showing the structure of an electronic device corresponding to the click rate prediction device provided in the embodiment;

[0059] Fig. 9 A schematic diagram of the structure of a sample identification device provided by an embodiment of the present invention;

[0060] Fig.10 For Fig. 9 A schematic diagram of the structure of an electronic device corresponding to the sample identification device provided in the embodiment is shown;

[0061] Fig.11 A schematic diagram of the structure of an information recommendation device provided by an embodiment of the present invention;

[0062] Fig.12 For Fig.11 A schematic diagram of the structure of an electronic device corresponding to the information recommendation device provided in the embodiment is shown. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "an", and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two, but does not exclude the situation of including at least one.

[0065] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0066] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0067] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a product or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such a product or system. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the product or system including the elements.

[0068] In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0069] Definition of terms:

[0070] Riemann manifold: is a locally Euclidean topological space. Specifically, a Riemann manifold is a smooth manifold in which every point on the tangent space has a positive inner product. Commonly used Riemann manifolds are of the following types: Hyperbolic manifolds, Stiefel manifolds, Grassmann manifolds, and Symmetric Positive Definite manifolds.

[0071] A hyperbolic manifold is a Riemannian manifold with constant negative curvature c, which is defined as in, refers to the set of points on the hyperbolic manifold, X is the element in the set, c is the curvature of the Riemann manifold, R d is the field of real numbers, ||X|| is the norm of X.

[0072] Stiefel manifold St(p,d): represents a set of d×p matrices with orthogonal columns, that is, St(p,d)={X∈R d×p :X T X=I p}, where St(p,d) is the set of points on the Stiefel manifold, X is the element in the set, and R d×p It is the matrix information composed of d dimensions and p dimensions, where p <d,I p is the p-dimensional identity matrix and T is the transpose operation.

[0073] Grassmann manifold: a set of subspaces representing the extension of an orthogonal matrix [X] = {XR:X∈St(p,d),R∈O(p)}, where O(p) represents an orthogonal group, i.e. X∈R p×p , and R T R=RR T =I p .

[0074] Smooth positive definite manifold is a set of orthogonal symmetric matrices of d×d dimensions, that is is located in The point set, It means that v can be any data other than 0 in the real number domain.

[0075] Riemannian two-level optimization: is a hierarchical mathematical paradigm in which an outer optimization objective nests an inner optimization problem with matrix manifold constraints.

[0076] Riemann Meta-Learning Task: The goal of the Riemann Meta-Learning task is to learn a better initialization of the neural network parameters to help the neural network converge quickly and quickly converge to new tasks. The Riemann Meta-Learning task assumes that the neural network parameters are located on the Riemann manifold space.

[0077] Riemann hyperparameter optimization task: The goal of the Riemann hyperparameter optimization task is to learn better hyperparameters to help the neural network converge to a better solution. Hyperparameters are parameters that need to be manually set during model training, such as learning rate, regularization parameters, etc. Different hyperparameter combinations will have a significant impact on model performance, but finding a better hyperparameter combination usually requires a lot of experiments and searches.

[0078] Riemann meta-optimization task: The goal of the Riemann meta-optimization task is to learn an optimizer parameterized by a neural network. The optimizer learned in this data-driven way can capture the intrinsic geometric structure of the data and help the neural network parameters converge more quickly.

[0079] In order to facilitate understanding of the specific implementation process and implementation effect of the click rate prediction method, device, equipment and computer program product in this embodiment, the relevant technology is briefly described below:

[0080] In the application scenario of the network platform, the click-through rate refers to the ratio of the number of times a certain content on a website page is clicked to the number of times it is displayed, that is, clicks / views. It is a percentage that can reflect the degree of attention paid to a certain content on a web page, and it can be used to measure the attractiveness of an advertisement. At present, the click-through rate is often obtained by statistically calculating the number of clicks and the number of displays, which not only takes a lot of time and resources, but also reduces the accuracy and efficiency of click-through rate prediction.

[0081] In order to solve the problem that the estimation operation of click-through rate requires a lot of time and space resources, the relevant technology provides a method to realize the prediction operation of click-through rate through a neural network model. At this time, the prediction accuracy of the click-through rate is closely related to the reasoning accuracy of the neural network model. Therefore, the optimization operation of the neural network model is a technical hotspot that is focused on research.

[0082] At present, the optimization method of the neural network model can be modeled as a two-layer optimization problem, including: an outer optimization problem and an inner optimization problem, wherein the outer optimization problem is used to optimize the hyperparameters and initialization parameters of the neural network model; the inner optimization problem is used to optimize the model parameters (e.g., model weights) of the neural network model. The above two-layer optimization problem can be mapped to a Riemann meta-learning task (hereinafter referred to as a "meta-learning task") and a Riemann hyperparameter optimization task (hereinafter referred to as a "hyperparameter optimization task"). The mapped meta-learning task and hyperparameter optimization task play a vital role in the click-through rate prediction system, the search advertising recommendation system, or the sample recognition system. Among them, the meta-learning task can be used for tasks such as personalized recommendation and click-through rate prediction. As the environment of search advertising is constantly changing, new types of advertisements and user behavior patterns will continue to emerge, and traditional neural network models may require a large amount of data to adapt to these changes. The meta-learning task can learn from the experience of multiple related tasks, so that the model can adapt to the new environment more quickly when facing new tasks. In the search advertising recommendation system, due to the complexity of data and tasks, the hyperparameter optimization task becomes particularly important. Appropriate hyperparameters can significantly improve the performance of the neural network model and enhance the effectiveness of ad recommendations.

[0083] However, since data in practical applications often have obvious non-Euclidean characteristics, for example, commodity categories have a strong hierarchical structure. If Euclidean geometry is simply used to model data, the complex geometric characteristics of the data will be destroyed, resulting in serious data distortion and loss of semantic information. Therefore, Riemannian manifold space (such as hyperbolic manifold, Stiefel manifold, Grassmann manifold, smooth positive definite manifold, etc.) can be used to model data and construct corresponding manifold network models (such as hyperbolic neural networks).

[0084] When converting the two-layer optimization problem of the neural network model into a Riemann two-layer optimization problem (for example, outer optimization problem, inner optimization problem), since the outer optimization problem is coupled with the inner optimization problem, calculating the gradient of the outer optimization parameters requires storing, expanding, and differentiating the entire inner optimization process. Directly using automatic differentiation to solve the two-layer optimization problem will introduce higher computing pressure and space storage pressure. Therefore, how to solve the two-layer optimization problem is challenging.

[0085] In order to solve the problems of excessive time resource consumption and high space complexity in the prediction scenario of click rate in the related art, this embodiment provides a click rate prediction method, device, equipment and computer program product. For details, refer to the attached Figure 1As shown, the execution subject of the click-through rate prediction method can be a click-through rate prediction device 200, and the click-through rate prediction device 200 can be implemented as a local server, a server in the cloud, or a preset device. In the case where the click-through rate prediction device 200 is implemented as a server in the cloud, the click-through rate prediction method can be executed in the cloud, and several computing nodes (cloud servers) can be deployed in the cloud, and each computing node has computing, storage and other processing resources. In the cloud, multiple computing nodes can be organized to provide a certain service. Of course, a computing node can also provide one or more services. The way to provide the service in the cloud can be to provide a service interface to the outside, and the user calls the service interface to use the corresponding service. The service interface includes a software development kit (Software Development Kit, referred to as SDK), an application programming interface (Application Programming Interface, referred to as API), and the like.

[0086] The click rate prediction device 200 is connected to the client 100 in communication, wherein the client 100 is used for users to perform applications so as to trigger the prediction operation of the click rate. The above-mentioned client 100 can be any computing device with a certain click rate prediction capability. When implemented specifically, the client 100 can be a mobile phone, a personal computer PC, a tablet computer, a setting application, etc. In addition, the basic structure of the client 100 may include: at least one processor. The number of processors depends on the configuration and type of the client. The client 100 may also include a memory, which may be volatile, such as a random access memory (RAM), or non-volatile, such as a read-only memory (ROM), a flash memory, etc., or may include both types at the same time. The memory usually stores an operating system (OS), one or more application programs, and may also store program data, etc. In addition to the processing unit and the memory, the client 100 also includes some basic configurations, such as a network card chip, an IO bus, a display component, and some peripheral devices, etc. Optionally, some peripheral devices may include, for example, a keyboard, a mouse, an input pen, a printer, etc. Other peripheral devices are well known in the art and will not be described in detail here.

[0087] The click rate prediction device 200 refers to a device that can provide click rate prediction operations in a network virtual environment, and generally refers to a device that uses the network to perform information planning and click rate prediction operations. In physical implementation, the click rate prediction device 200 can be any device that can provide computing services and can perform corresponding click rate prediction operations, for example: it can be a cluster server, a conventional server, a cloud server, a cloud host, a virtual center, etc. The page building device 200 mainly includes a processor, a hard disk, a memory, a system bus, etc., which is similar to a general computer architecture.

[0088] In the above embodiment, the client 100 is connected to the click rate prediction device 200 through a network, and the network connection can be a wireless or wired network connection. If the client 100 can be connected to the click rate prediction device 200 through communication, the network standard of the mobile network can be any one of 2G (GSM), 2.5G (GPRS), 3G (WCDMA, TD-SCDMA, CDMA2000, UTMS), 4G (LTE), 4G+ (LTE+), WiMax, 5G, 6G, etc.

[0089] In the embodiment of the present application, the client 100 is used by the user to generate a click-through rate prediction request, wherein the click-through rate prediction request can be generated through a human-computer interaction operation or a voice interaction operation. After obtaining the click-through rate prediction request, the click-through rate prediction request can be sent to the click-through rate prediction device 200, so that the click-through rate prediction device 200 can perform a corresponding click-through rate prediction operation based on the click-through rate prediction request.

[0090] The click-through rate prediction device 200 is used to obtain a click-through rate prediction request for a preset object through the client 100, and then determine a neural network model for implementing the click-through rate prediction operation based on the click-through rate prediction request, wherein the neural network model may correspond to model parameters, and the model parameters may include model weight parameters of the neural network model; in order to achieve optimization operations on the neural network model, the obtained model parameters may be mapped to a Riemann manifold, and the model parameters may be Riemann optimized to obtain optimized model parameters; after obtaining the optimized model parameters, the neural network model may be updated based on the optimized model parameters to obtain the optimized model; then, the click-through rate prediction operation may be performed based on the optimized model, and a click-through rate prediction result corresponding to the preset object may be obtained, thereby completing the click-through rate prediction operation.

[0091] In this embodiment, by obtaining a click-through rate prediction request for a preset object, a neural network model for implementing a click-through rate prediction operation can be determined based on the click-through rate prediction request, and then the model parameters corresponding to the neural network model are mapped to the Riemann manifold, and the model parameters are Riemann optimized to obtain the optimized model parameters; after obtaining the optimized model parameters, the neural network model can be updated based on the optimized model parameters to obtain the optimized model; then, the click-through rate prediction operation can be performed based on the optimized model, thereby effectively realizing the click-through rate prediction operation using the optimized model, which not only saves the time and space resources required for the click-through rate prediction operation, but also ensures the accuracy and efficiency of the click-through rate prediction, further improving the practicality of the method.

[0092] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and the features in the embodiments can be combined with each other. In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0093] Figure 2 A flowchart of a method for predicting click rate provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, this embodiment provides a method for predicting click rate, the execution subject of the method is a prediction device for click rate, specifically, the prediction device for click rate can be implemented as software, or a combination of software and hardware, when the prediction device for click rate is implemented as hardware, it can be various electronic devices that can implement click rate prediction operations, including but not limited to personal computers, servers, databases, etc., when the prediction device for click rate is implemented as software, it can be installed in the electronic devices listed above. Based on the above-mentioned prediction device for click rate, the prediction operation of click rate can be implemented, specifically, the prediction method for click rate can include:

[0094] Step S201: obtaining a click rate prediction request for a preset object.

[0095] Step S202: Based on the click-through rate prediction request, determine a neural network model for implementing the click-through rate prediction operation, and the neural network model corresponds to model parameters.

[0096] Step S203: Map the model parameters to the topological space, and perform Riemann optimization on the model parameters to obtain optimized model parameters.

[0097] Step S204: updating the neural network model based on the optimized model parameters to obtain an optimized model.

[0098] Step S205: Perform a click-through rate prediction operation based on the optimized model to obtain a click-through rate prediction result corresponding to a preset object.

[0099] The specific implementation process and effects of each of the above steps are described in detail below:

[0100] Step S201: obtaining a click rate prediction request for a preset object.

[0101] When a user has a click-through rate prediction demand for a preset object, the click-through rate prediction device can obtain a click-through rate prediction request for the preset object, where the preset object may include any one of the following objects: goods, pictures, articles, videos, etc. In different application scenarios, different click-through rate prediction requests can be generated for different preset objects.

[0102] Specifically, this embodiment does not limit the specific implementation method for obtaining the click-through rate prediction request. In some instances, the click-through rate prediction request for the preset object can be obtained through human-computer interaction operations. At this time, obtaining the click-through rate prediction request for the preset object may include: displaying the human-computer interaction interface corresponding to the preset object; determining the execution operation input by the user in the human-computer interaction interface; and obtaining the click-through rate prediction request for the preset object based on the execution operation.

[0103] In other instances, the click-through rate prediction request for the preset object can be obtained not only through human-computer interaction operations, but also through a preset device. In this case, obtaining the click-through rate prediction request for the preset object may include: determining a preset device (for example, a client) that is communicatively connected to the click-through rate prediction device, wherein the preset device stores or generates the click-through rate prediction request for the preset object; actively or passively obtaining the click-through rate prediction request for the preset object through the preset device, thereby also ensuring the accuracy and reliability of obtaining the click-through rate prediction request.

[0104] Step S202: Based on the click-through rate prediction request, determine a neural network model for implementing the click-through rate prediction operation, and the neural network model corresponds to model parameters.

[0105] In order to efficiently implement the click-through rate prediction operation, a pre-trained neural network model can be used to implement the click-through rate prediction operation. At this time, the accuracy of the click-through rate prediction is closely related to the model reasoning accuracy of the neural network model. In order to ensure the accuracy and reliability of the click-through rate prediction operation, the neural network model used to implement the click-through rate prediction operation can be optimized. Specifically, after obtaining the click-through rate prediction request, the neural network model used to implement the click-through rate prediction operation can be determined based on the click-through rate prediction request.

[0106] For a neural network model, it may correspond to model parameters. In some scenarios, the model parameters may be implemented as model internal parameters, which may specifically include model weights corresponding to the neural network model; in some other scenarios, the model parameters may be implemented as model external parameters, which may include initialization parameters and hyperparameters corresponding to the neural network model that need to be pre-configured by the user; in other scenarios, the model parameters may be implemented as model internal parameters and model external parameters, and, for a neural network model, different neural network models may correspond to different model parameters.

[0107] In addition, this embodiment does not limit the method for determining the neural network model. In some instances, the neural network model can be determined by a preset mapping relationship. In this case, based on the click-through rate prediction request, determining the neural network model used to implement the click-through rate prediction operation may include: based on the click-through rate prediction request, determining the application function identifier corresponding to the click-through rate prediction operation; determining the mapping relationship used to determine the neural network model, the mapping relationship including the correspondence between each application function and the corresponding neural network model; based on the mapping relationship and the function identifier, the neural network model used to implement the click-through rate prediction operation can be determined, thereby effectively ensuring the accuracy and reliability of the determination of the neural network model.

[0108] Step S203: Map the model parameters to the topological space, and perform Riemann optimization on the model parameters to obtain optimized model parameters.

[0109] After obtaining the model parameters, in order to optimize the neural network model, the model parameters can be mapped to the topological space. In some instances, the topological space can be implemented as a topological space corresponding to any manifold, such as Riemann manifold, symplectic manifold, contact manifold, etc. For the sake of convenience, the topological space corresponding to the Riemann manifold is used as an example for the specific implementation method. At this time, the model parameters can be directly mapped to the Riemann manifold. Specifically, the model parameters can be mapped to the Riemann manifold through a preset recovery operation function and / or an orthogonal projection operation function. The Riemann manifold can include any one of the following: hyperbolic manifolds, Stiefel manifolds, Grassmann manifolds, smooth positive definite manifolds, etc.

[0110] After mapping the model parameter X to the Riemann manifold, in order to optimize the neural network model, the model parameter X can be Riemann optimized to obtain the optimized model parameter X *In some examples, the Riemann optimization operation can be implemented by a Riemann function. In this case, performing Riemann optimization on the model parameter X may include: constructing a Riemann function for performing an optimization operation on the model parameter X; performing an optimization operation on the model parameter X using the Riemann function, so as to obtain the optimized model parameter X. * , which effectively ensures the optimization of the model parameters X * The accuracy and reliability of acquisition.

[0111] In other instances, in order to ensure the quality and effect of optimizing the model parameters, in the application scenario of click-through rate prediction, not only can the model parameters be mapped to the Riemann manifold, but the object features of the preset object can also be obtained, and then the object features and the model parameters can be mapped to the Riemann manifold together, that is, the model parameters can be Riemann optimized in combination with the corresponding object features, so as to obtain more accurate and reliable optimized model parameters that match the click-through rate prediction request, thereby further ensuring the accuracy of determining the optimized model parameters.

[0112] It should be noted that since the model parameters can be implemented as at least one of the following: model intrinsic parameters and model extrinsic parameters; when performing Riemann optimization operations on the model parameters, different optimization operations can be implemented according to the needs of the application scenario, for example: only the model internal parameters can be Riemann optimized to obtain the optimized model internal parameters; or, the model external parameters can be Riemann optimized to obtain the optimized model external parameters; or, the model internal parameters and model external parameters can be Riemann optimized to obtain the optimized model internal parameters and the optimized model external parameters, thereby effectively achieving the flexibility and reliability of Riemann optimization of the model parameters.

[0113] Step S204: updating the neural network model based on the optimized model parameters to obtain an optimized model.

[0114] After obtaining the optimized model parameters, the neural network model can be updated based on the optimized model parameters. In some instances, the optimized model parameters can be directly used to replace the model parameters in the neural network model, so that the optimized model can be stably obtained. Compared with the neural network model corresponding to the model parameters, the reasoning accuracy of the optimized model is higher than the reasoning accuracy of the neural network model; or, the reasoning efficiency of the optimized model is higher than the reasoning efficiency of the neural network model.

[0115] Step S205: Perform a click-through rate prediction operation based on the optimized model to obtain a click-through rate prediction result corresponding to a preset object.

[0116] After obtaining the optimized model, the click-through rate prediction operation can be performed based on the optimized model. Specifically, the click-through rate prediction operation based on the optimized model may include: obtaining click-related data of the preset object; inputting the preset object and click-related data into the optimized model for analysis and processing, so as to obtain the click-through rate prediction result output by the optimized model, thereby completing the click-through rate prediction operation.

[0117] The click-through rate prediction method provided in this embodiment obtains a click-through rate prediction request for a preset object, and then determines a neural network model for implementing a click-through rate prediction operation based on the click-through rate prediction request, and then maps the model parameters corresponding to the neural network model to the topological space, and performs Riemann optimization on the model parameters, so as to obtain the optimized model parameters; after obtaining the optimized model parameters, the neural network model can be updated based on the optimized model parameters to obtain the optimized model; then, a click-through rate prediction operation can be performed based on the optimized model to obtain a click-through rate prediction result corresponding to the preset object, which not only saves the time resources and space resources required for the click-through rate prediction operation, but also ensures the prediction accuracy and efficiency of the click-through rate, further improving the practicality of the method.

[0118] Figure 3 The present invention provides a flow chart of performing Riemann optimization on model parameters to obtain optimized model parameters; based on the above embodiment, refer to the attached Figure 3 As shown, for the model parameters, not only can the Riemann optimization operation of the model parameters be implemented by the Riemann function, but also the Riemann optimization operation can be implemented by constructing a loss function corresponding to the neural network model. At this time, the Riemann optimization of the model parameters in this embodiment to obtain the optimized model parameters may include:

[0119] Step S301: Construct a loss function corresponding to the neural network model.

[0120] In order to optimize the neural network model, a loss function corresponding to the neural network model can be constructed first, wherein when the model parameters are implemented as model internal parameters, the constructed loss function can be a loss function used to perform inner layer optimization operations on the neural network model. Specifically, this embodiment does not limit the specific construction method of the loss function. In some instances, the loss function can be determined by analyzing and processing the neural network model. At this time, constructing the loss function corresponding to the neural network model may include: obtaining historical click data and historical click-through rate results corresponding to the historical object; analyzing and processing the historical click data using the neural network model to obtain predicted click information corresponding to the historical click data; and constructing a loss function corresponding to the neural network model based on the predicted click information and the historical click-through rate results. This effectively ensures the accuracy and reliability of the construction of the loss function.

[0121] In other instances, the loss function can not only be determined by analyzing and processing the neural network model, but also the model parameters can be mapped into the Riemann space to construct a loss function corresponding to the neural network model. At this time, constructing the loss function corresponding to the model parameters can include: mapping the model parameters into the Riemann space to obtain mapping parameters corresponding to the model parameters; based on the model parameters and the mapping parameters, constructing a loss function corresponding to the neural network model.

[0122] In order to stably implement the Riemann optimization operation on the model parameters, the model parameters can be mapped to the Riemann space. Specifically, the recovery operation function in the Riemann manifold can be used to map the model parameters to the Riemann space, so that the mapping parameters corresponding to the model parameters can be obtained. For example, when the model parameter is X, since the optimization operation of the neural network model requires multiple iterative operations on the model parameters, the update function of the model parameter X can be defined as f θ (X), and then we can use the recycling operation function Γ in the Riemann manifold X To map the model parameters to the Riemann space, we obtain the mapping parameters Γ corresponding to the model parameters X (f θ (X)).

[0123] After obtaining the mapping parameters, the model parameters and the mapping parameters can be analyzed and processed to construct a loss function corresponding to the neural network model, thereby effectively ensuring the accuracy and reliability of the loss function. Specifically, the loss function can be implemented as: in, is the loss function corresponding to the neural network model, X is the model parameter, Γ X (f θ (X)) is the mapping parameter corresponding to the model parameter, Γ X() is the recycling operation function in the Riemann manifold, and “||||” is the norm.

[0124] Step S302: Perform Riemann optimization on the loss function to obtain the minimum value of the loss function.

[0125] After obtaining the loss function, the loss function can be subjected to Riemann optimization to obtain the minimum value of the loss function. In some instances, since the loss function is a function related to the model parameters, for example, when the loss function is When , different X values ​​can correspond to different loss function values In the process of optimizing the neural network model, the model parameters will be continuously iterated and updated, so that multiple different loss function values ​​corresponding to each model parameter can be obtained, and then the minimum value of the loss function can be determined among the multiple different loss function values. This effectively ensures the accuracy and reliability of determining the minimum value of the loss function.

[0126] In other instances, the Riemann optimization operation of the loss function can be determined not only by different model parameters, but also by the gradient of the model parameters. In this case, performing Riemann optimization on the loss function to obtain the minimum value of the loss function may include: determining the gradient of the model parameters based on the loss function; determining the inner iteration step size for Riemann optimization of the loss function; iterating the loss function based on the gradient of the model parameters and the inner iteration step size and using the Riemann gradient descent method to obtain the minimum value of the loss function.

[0127] After obtaining the loss function, the loss function can be analyzed and processed to determine the gradient of the model parameters. For example, when the loss function is When , the gradient of the model parameter X can be determined based on the above loss function In order to accurately obtain the minimum value of the loss function, the inner iteration step size T of the Riemann optimization operation on the loss function can be determined. In some instances, the inner iteration step size T can be determined through human-computer interaction, or the inner iteration step size T can be a pre-configured parameter. After determining the gradient of the model parameters and the inner iteration step size of the Riemann optimization operation on the loss function, the Riemann gradient descent method, the gradient of the model parameters and the inner iteration step size can be used to iterate the loss function, so that the minimum value of the loss function can be stably obtained.

[0128] In some other instances, in order to stably implement Riemann optimization operations on the neural network model, the model parameters can be iteratively updated in the process of iteratively processing the loss function based on the gradient of the model parameters and the inner iteration step size and using the gradient descent method. At this time, the method in this embodiment may also include: mapping the gradient of the model parameters to the tangent space to obtain the mapping gradient; based on the mapping gradient, determining the iterative model parameters corresponding to the model parameters; and updating the model parameters to the iterative model parameters.

[0129] Specifically, after obtaining the gradient of the model parameters Afterwards, the gradient of the model parameters can be Mapping to the tangent space, the gradient of the model parameters can be obtained by the orthogonal projection operation function π() in the Riemann manifold. Mapped to the tangent space, the mapping gradient can be obtained After getting the mapping gradient Afterwards, we can map the gradient Perform analytical processing to determine the iterative model parameters corresponding to the model parameters, for example: when the mapping gradient is When , the determined iterative model parameters can be Among them, ξ can be the preset iteration step size, Γ X () is a recycling operation function on the Riemann manifold, which is used to map the iterative model parameters to the topological space (specifically, it can be implemented as a Riemann manifold), so that the iterative model parameters located on the Riemann manifold can be obtained.

[0130] After the iterative model parameters corresponding to the model parameters are determined by the above formula, the model parameters can be updated to the iterative model parameters. Specifically, when the model parameters are X, the iterative model parameters are When , the model parameters can be updated to the iterative model parameters, that is, This effectively ensures that the model parameters can be continuously iteratively updated as the loss function is iteratively processed, thus ensuring the quality and effect of the Riemann optimization operation on the neural network model.

[0131] In some other instances, in order to accurately implement the iterative update operation of the model parameters, in the process of iteratively processing the loss function based on the gradient of the model parameters and the inner iteration step and using the gradient descent method, not only can the gradient of the model parameters be mapped to the tangent space to obtain the mapping gradient; the object features of the preset object can also be mapped to the tangent space, and then the model parameters and object features mapped on the tangent space can be combined to obtain the mapping gradient corresponding to the model parameters, which effectively ensures the accurate and reliable determination of the mapping gradient, and then the iterative model parameters corresponding to the model parameters can be determined based on the mapping gradient; then the model parameters can be updated to the iterative model parameters, thereby ensuring the flexibility and reliability of the update operation of the model parameters.

[0132] Step S303: Determine the target parameter corresponding to the minimum value as the optimized model parameter.

[0133] Since the loss function is a function related to the model parameters, after obtaining the minimum value of the loss function, the target parameter corresponding to the minimum value can be determined based on the minimum value of the loss function, and then the target parameter corresponding to the minimum value can be determined as the optimized model parameter. When Get the minimum value The corresponding target parameter is determined as the optimized model parameter X * , This effectively ensures the accuracy and reliability of determining the optimized model parameters.

[0134] In this embodiment, by constructing a loss function corresponding to the neural network model, the loss function is Riemann optimized to obtain the minimum value of the loss function, and then the target parameters corresponding to the minimum value can be determined as the optimized model parameters, thereby effectively ensuring the accuracy and reliability of the determination of the optimized model parameters. When the click-through rate prediction operation is performed based on the optimized model parameters, the accuracy of the click-through rate prediction operation can be improved.

[0135] Figure 4 The schematic diagram of the principle of determining the optimized model based on the optimized model parameters provided in the embodiment of the present invention; on the basis of any one of the above embodiments, refer to the attached Figure 4 As shown, when optimizing the neural network model, not only the model internal parameters of the neural network model can be optimized and updated, but also the model external parameters of the neural network model can be optimized and updated. At this time, based on the optimized model parameters, determining the optimized model can include:

[0136] Step S401: Obtain preset outer layer parameters corresponding to the neural network model, where the preset outer layer parameters include at least one of the following: hyperparameters and initialization parameters.

[0137] For the model parameters of the neural network model, they can be implemented as the model internal parameters of the neural network model (which can be the model weights of the neural network model). When optimizing the neural network model, not only the model internal parameters of the neural network model can be optimized, but also the model external parameters of the neural network model can be optimized. Among them, the model external parameters can refer to the model parameters that need to be manually configured. For example, the model external parameters can include at least one of the following: hyperparameters, initialization parameters, and the initialization parameters can be the initial unit matrix, initial unit vector, etc. of the neural network model.

[0138] Specifically, in order to ensure the optimization quality and effect of the neural network model, preset outer layer parameters corresponding to the neural network model can be obtained. In some instances, the preset outer layer parameters can be obtained through human-computer interaction operations, or the preset outer layer parameters can be stored in a preset area. At this time, the preset outer layer parameters can be obtained by accessing the preset area, which effectively ensures the accuracy and reliability of obtaining the preset outer layer parameters.

[0139] Step S402: Optimize the preset outer layer parameters to obtain optimized outer layer parameters corresponding to the preset outer layer parameters.

[0140] After obtaining the preset outer layer parameters, the preset outer layer parameters can be optimized, so that the optimized outer layer parameters corresponding to the preset outer layer parameters can be obtained. In some instances, the optimized outer layer parameters can be obtained by optimizing the preset outer layer parameters through a preset optimization algorithm or machine learning model. Alternatively, the optimized outer layer parameters can be obtained by optimizing the preset outer layer parameters through an objective function corresponding to the click-through rate prediction request. At this time, optimizing the preset outer layer parameters to obtain the optimized outer layer parameters corresponding to the preset outer layer parameters may include: determining the objective function corresponding to the click-through rate prediction request based on the optimized model parameters, the objective function being a binary classification function related to the preset outer layer parameters and used to identify whether to perform a click operation; optimizing the preset outer layer parameters based on the objective function to determine the optimized outer layer parameters corresponding to the preset outer layer parameters.

[0141] In order to accurately preset the outer layer parameters for optimization, after obtaining the optimized model parameters, the objective function corresponding to the click rate prediction request can be determined based on the optimized model parameters. The objective function can be a binary classification function related to the preset outer layer parameters and used to identify whether to perform a click operation. In some instances, the objective function can be stored in a preset area. After obtaining the optimized model parameters, the preset area can be accessed based on the optimized model parameters, and the objective function corresponding to the click rate prediction request can be obtained. After obtaining the target function Afterwards, based on the objective function The initialization outer layer parameters are optimized so as to determine the optimized outer layer parameters corresponding to the initialization outer layer parameters.

[0142] Specifically, this embodiment does not limit the specific implementation method of optimizing the preset outer layer parameters based on the objective function. In some instances, since the value of the objective function is related to the value of the preset outer layer parameters, the value of the objective function may change with the change of the preset outer layer parameters. Therefore, the optimized outer layer parameters corresponding to the preset outer layer parameters can be determined based on the function value of the objective function. At this time, optimizing the preset outer layer parameters based on the objective function, determining the optimized outer layer parameters corresponding to the preset outer layer parameters may include: obtaining the function change trend of the objective function; estimating the parameter change trend of the preset outer layer parameters based on the function change trend, and then determining the optimized outer layer parameters corresponding to the initialized outer layer parameters based on the function change trend and the parameter change trend, thereby effectively realizing the stable optimization operation of the initialized outer layer parameters.

[0143] In other instances, the optimized outer layer parameters can also be determined by initializing the gradient of the outer layer parameters. In this case, the preset outer layer parameters are optimized based on the objective function, and determining the optimized outer layer parameters corresponding to the preset outer layer parameters may include: determining the gradient of the preset outer layer parameters based on the objective function; determining the outer layer iteration step for Riemann optimization of the objective function; iteratively updating the preset outer layer parameters based on the gradient of the preset outer layer parameters and the outer layer iteration step and using the Riemann gradient descent method to obtain the minimum preset outer layer parameters; and determining the minimum preset outer layer parameters as the optimized outer layer parameters.

[0144] Among them, the objective function is a function related to the preset outer layer parameter θ After obtaining the target function Afterwards, we can use the objective function To determine the gradient of the preset outer layer parameters In some instances, in order to simplify the gradient of the preset outer parameters, the matrix representation in the gradient can be rewritten into a preset sequence. At this time, based on the objective function, determining the gradient of the preset outer parameters may include: based on the objective function, determining the original gradient of the preset outer parameters, the original gradient including a matrix representation whose time complexity is greater than a preset threshold; rewriting the matrix representation in the original gradient into a preset sequence to obtain the gradient of the initialization parameter, wherein the error between the gradient of the initialization parameter determined based on the preset sequence and the gradient of the initialization parameter determined based on the matrix representation is less than or equal to the preset threshold.

[0145] Specifically, for the gradient of the preset outer layer parameters in, The operation can be easily calculated using automatic differentiation techniques, since Then you can get Then, the chain rule can be used to convert the above Expanded to The following relationship can be obtained through the above expanded formula:

[0146]

[0147] By readjusting the above formula, we can get the following formula:

[0148]

[0149] in, So that we can obtain This allows the gradient of the preset outer layer parameters to be implicitly calculated as

[0150] For the gradient of the above preset outer layer parameters, (IJ) -1 The calculation of is very difficult. Specifically, when the number of parameters of the neural network model is N, the matrix J can be a matrix of N×N, then the matrix is ​​difficult to store, making the inverse matrix (IJ) -1 The time complexity is O(N 3 ).

[0151] For the objective function, the original gradient of the preset outer layer parameters can be determined based on the objective function. The original gradient can be implemented as The matrix representation of the time complexity included in the original gradient obtained is greater than the preset threshold, and the matrix representation is (IJ) -1In order to reduce the complexity of the gradient of the preset outer layer parameter, the matrix representation in the original gradient is rewritten as a preset sequence. The preset sequence can be implemented as a Neumann sequence, so that the gradient of the preset outer layer parameter can be obtained, wherein the error between the gradient of the preset outer layer parameter determined based on the preset sequence and the gradient of the preset outer layer parameter determined based on the matrix representation is less than or equal to the preset threshold, that is, the gradient of the preset outer layer parameter determined by the preset sequence is The gradient of the preset outer layer parameters determined based on the matrix representation can be After the rewriting operation of the preset sequence, the gradient of the preset outer layer parameter obtained satisfies the following relationship in, Used to identify the error between the gradient of the preset outer layer parameters determined based on the preset sequence and the gradient of the preset outer layer parameters determined based on the matrix representation, wherein c is a preset constant and δ is the acceptable error basic information.

[0152] In order to accurately determine the optimized outer layer parameters, the outer layer iteration step length K for performing the Riemann optimization operation on the objective function can be determined. The outer layer iteration step length K can be a parameter stored in a preset area. The outer layer iteration step length for performing the Riemann optimization operation on the objective function can be determined by accessing the preset area. After obtaining the gradient and the outer layer iteration step length of the initial outer layer function, the preset outer layer parameters can be iteratively updated based on the gradient and the outer layer iteration step length of the preset outer layer parameters and using the Riemann gradient descent method, so as to obtain the minimum preset outer layer parameters.

[0153] Specifically, the specific implementation process and implementation effect of iteratively updating the preset outer layer parameters are similar to the specific implementation process and implementation effect of iteratively processing the loss function in the above embodiment. For details, please refer to the above statement content, which will not be repeated here. After obtaining the minimum preset outer layer parameter, the minimum preset outer layer parameter can be directly determined as the optimized outer layer parameter, which effectively ensures the accuracy and reliability of determining the optimized outer layer parameter.

[0154] Step S403: updating the neural network model based on the optimized model parameters and the optimized outer layer parameters to determine the optimized model.

[0155] After obtaining the optimized model parameters and optimized outer layer parameters, the neural network model can be updated based on the optimized model parameters and optimized outer layer parameters. In some instances, the optimized model parameters can be directly used to replace the model parameters in the neural network model, and the optimized outer layer parameters can be used to replace the initial model external parameters in the neural network model, so that the optimized model can be stably obtained. Compared with the neural network model corresponding to the model parameters, the reasoning accuracy of the optimized model is higher than the reasoning accuracy of the neural network model; or, the reasoning efficiency of the optimized model is higher than the reasoning efficiency of the neural network model, so that in the process of click-through rate prediction based on the optimized model, the accuracy and reliability of click-through rate prediction can be effectively improved.

[0156] In this embodiment, by obtaining preset outer layer parameters corresponding to the neural network model, and then optimizing the preset outer layer parameters, the optimized outer layer parameters corresponding to the preset outer layer parameters are obtained, and the neural network model is updated based on the optimized model parameters and the optimized outer layer parameters to determine the optimized model, thereby effectively ensuring the flexibility and reliability of the optimization operation of the neural network model and further improving the reasoning accuracy of the neural network model.

[0157] In specific applications, in the application scenario of click-through rate prediction, the prediction operation of click-through rate can be implemented through a pre-trained click-through rate prediction model. Considering that the model optimization task in the click-through rate prediction model can be modeled as a Riemann two-layer optimization problem, for example: the hyperparameter optimization task of the click-through rate prediction model, etc. However, the current Riemann two-layer optimization method has high time and space complexity, that is, directly using automatic differentiation technology to solve the Riemann two-layer optimization problem will introduce computational pressure and space storage pressure.

[0158] In order to solve the above technical problems, this application embodiment provides a click-through rate prediction method based on Riemann implicit differentiation technology, which can solve the Riemann two-level optimization problem corresponding to the above click-through rate prediction model through Riemann implicit technology, which can greatly reduce the time and space complexity. Specifically, in order to clearly understand the prediction principle of click-through rate, the modeling process of the Riemann two-level optimization problem is first explained below:

[0159] For the click-through rate prediction model, the optimization tasks of the click-through rate prediction model may include: an inner optimization task for training the model parameters of the neural network (for example: model weights), and an outer optimization task for training hyperparameters (for example: regularization coefficients, sample weights, etc.). The optimization tasks of the click-through rate prediction model can be modeled as a Riemann two-layer optimization problem, such as: Riemann meta-learning tasks, Riemann hyperparameter optimization tasks, and Riemann meta-optimization tasks; among them, the Riemann meta-learning task can effectively achieve rapid adaptation in small sample recognition tasks; the Riemann hyperparameter optimization task can be used to solve recognition problems with long-tail distributions to alleviate the problem of low recognition accuracy of tail samples.

[0160] Specifically, the Riemann two-level optimization problem can be modeled as:

[0161]

[0162] The first formula in the above formula (1) can be expressed as To represent In the application scenario of click-through rate prediction, θ is a preset model external parameter, which can be implemented as a preset hyperparameter and / or initialization parameter of the click-through rate prediction model, that is, θ is a parameter set manually to be optimized. The binary classification objective function for whether a click operation is performed is optimized for the outer layer of the model. It can be expressed as a binary objective function of whether the click operation is performed when minimizing the model external parameter θ; X is the model internal parameter located on the Riemann manifold, is a Riemann manifold, It is expressed as the objective function for optimizing the model internal parameters. In the CTR prediction scenario, the above objective function can be implemented as the loss function of the CTR prediction model, X * (θ) is the minimum model internal parameter corresponding to the minimum value of the objective function.

[0163] Then, the memory optimization process of the click rate prediction model can be modeled in a unified way to obtain the following formula:

[0164]

[0165] Among them, X (t+1) It refers to the model internal parameters obtained after t+1 iterations of update operations. X can be realized as a vector or matrix. is a recycling operation in the Riemann manifold, used to convert the parameter f θ (X (t) ) is mapped to the Riemann manifold. Specifically, Can be implemented as P is the input parameter of the recycling operation, which can be realized as the parameter of the cutting space. θ (X (t) ) is the model internal parameter X obtained for the t-th iteration operation (t) Implementation function for update operation.

[0166] And, the above formula (2) can be rewritten as:

[0167]

[0168] Substituting the above formula (3) into formula (1), the Riemann two-level optimization problem can be remodeled as:

[0169]

[0170] Based on the above formula (4), the click rate prediction model can be optimized. Specifically, the inner optimization problem can be solved first to obtain the optimized model internal parameter X * , and then the above optimized model internal parameter X * Substitute it into the outer layer objective function to solve the outer layer objective function and obtain the optimized model external parameters, that is, the optimization operation of the click rate prediction model may include: solving the inner layer optimization problem and solving the outer layer optimization problem, wherein the solution process of the inner layer optimization problem may include the following steps:

[0171] Step 11: Optimize the model internal parameters to obtain the optimized model internal parameters.

[0172] When the model internal parameter is X, the optimized model internal parameter X is obtained through the above formula (3): * , specifically, the optimized model internal parameter X * It can be determined by the loss function established as follows:

[0173]

[0174] Based on the above loss function, the gradient of the model internal parameter X is calculated, and the gradient of the model internal parameter X is obtained as:

[0175]

[0176] In the above formula (6), I is the unit matrix, T is the preset inner layer iteration step size, is the Heather matrix. Specifically, use the loss function The reasons for calculating the inner layer parameter gradient are: 1) The loss function The gradient of the model internal parameter X contains a quadratic term It can improve the convergence performance of the model internal parameter X; 2) The loss function It is equivalent to solving the equality constraints; solving the objective function is equivalent to infinite iterations of the model internal parameter X, which can alleviate the bias problem in the two-layer optimization.

[0177] In determining the gradient of model parameters After that, we can use the Riemann gradient descent method and the gradient of the model parameters And the inner iteration step length T has an effect on the loss function Perform iterative processing to stably obtain the minimum value of the loss function Then the minimum value can be The corresponding model internal parameter value is determined as the optimized model internal parameter X * , which effectively ensures the optimization of the internal parameters X of the model * The accuracy and reliability of the determination.

[0178] Step 12: Update the model internal parameters based on the optimized model internal parameters.

[0179] After obtaining the optimized model internal parameter X * When , the optimized model internal parameters can be realized as Then, the Riemann gradient descent algorithm can be used to update the model internal parameter X. The update operation can be implemented based on the following update function:

[0180]

[0181] Among them, Γ X () is the recycling operation function on the Riemann manifold, which is used to map the iterative model parameters to the Riemann manifold, so that the iterative model parameters on the Riemann manifold can be obtained, ξ is the preset optimization step size, π X () is an orthogonal projection operation on the Riemann manifold, which takes the gradient in the Euclidean space as input and outputs the tangent space The update vector for the Riemann parameters on .

[0182] On the other hand, the solution process of the outer optimization problem can include the following steps:

[0183] Step 21: Derivation and calculation of model extrinsic parameter gradients.

[0184] Among them, the outer layer parameter θ can be implemented as a hyperparameter and / or initialization parameter of the click rate prediction model. Through the chain rule, the outer layer parameter gradient can be calculated as follows:

[0185]

[0186] in, The operation can be easily calculated using automatic differentiation techniques, since Then you can get Then, by using the chain rule, the above Expanded to The following relationship can be obtained through the above expanded formula:

[0187]

[0188] By readjusting the above formula (10), we can get the following formula:

[0189]

[0190] in, So that we can obtain This allows the gradient of the model's external parameters to be implicitly calculated as

[0191] Step 22: Optimize the model external parameters to obtain optimized model external parameters.

[0192] After obtaining the gradient of the model extrinsic parameter, the model extrinsic parameter can be optimized based on the gradient of the model extrinsic parameter to obtain the optimized model extrinsic parameter. For the gradient of the model extrinsic parameter, (IJ) -1 The calculation of is very difficult. Specifically, when the number of parameters of the neural network model is N and the matrix J is an N×N matrix, then the matrix is ​​difficult to store, making the inverse matrix (IJ) -1 The time complexity is O(N 3 ).

[0193] For the objective function, the original gradient of the model external parameter can be determined based on the objective function. The original gradient can be implemented as Among them, the matrix representation of the time complexity included in the original gradient obtained is greater than the preset threshold, and the matrix representation is (IJ) -1 In order to reduce the complexity of the gradient of the model external parameters, the matrix representation (IJ) in the original gradient is -1 Rewrite to preset sequence The preset sequence can be implemented as a Neumann sequence, that is, the Neumann sequence is used to calculate the inverse of the matrix, so that the implicit gradient of the model external parameter can be obtained as:

[0194]

[0195] According to the above formula (11), the first K terms can be used to estimate the formula. The above K value can be between 2 and 5, that is, K is 2, 3, 4, 5, etc., and then the above formula (11) can be rewritten as the following formula:

[0196]

[0197] The error between the gradient of the model extrinsic parameter determined based on the preset sequence and the gradient of the model extrinsic parameter determined based on the matrix representation is less than or equal to the preset threshold, that is, the gradient of the model extrinsic parameter determined by the preset sequence is The gradient of the model extrinsic parameters determined based on the matrix representation can be After the preset sequence of rewriting operations, the gradient of the model external parameters obtained satisfies the following relationship in, It is used to identify the error between the gradient of the model extrinsic parameter determined based on the preset sequence and the gradient of the model extrinsic parameter determined based on the matrix representation. c is a preset constant. Specifically, c can be calculated by the following formula: The above-mentioned S, β1, β2, γ1, λ1, α1, α2, ρ2, and Ι can be obtained through calculation operations, and δ is the acceptable error information. This shows that the error of the above-mentioned click-through rate prediction operation is bounded and can be accepted and tolerated by users.

[0198] In order to accurately determine the optimized outer layer parameters, the outer layer iteration step length K for performing the Riemann optimization operation on the objective function can be determined. The outer layer iteration step length K can be a parameter stored in a preset area. The outer layer iteration step length for performing the Riemann optimization operation on the objective function can be determined by accessing the preset area. After obtaining the gradient of the model extrinsic parameters and the outer layer iteration step length, the model extrinsic parameters can be iteratively updated based on the gradient of the model extrinsic parameters and the outer layer iteration step length, and using the Riemann gradient descent method, so as to obtain the minimum model extrinsic parameters.

[0199] Specifically, in order to calculate the optimized model external parameters by the gradient of the model external parameters, intermediate variables v and p can be introduced, and the above intermediate variables v and intermediate variables p can be initialized as Then the intermediate variables are updated by iterating K times:

[0200] v i+1 =v i J,p i+1 =p i +v i+1 (13)

[0201] After calculation using the intermediate variables v and p, after K iterations, the implicit gradient can be estimated as:

[0202]

[0203] Among them, the p value can be obtained by calculation, and the above-mentioned K value is the iteration step size, which can be a preset value. After obtaining the above-mentioned implicit gradient, the implicit gradient can be analyzed and processed by the gradient descent algorithm to obtain the minimum model extrinsic parameter θ. After obtaining the minimum model extrinsic parameter, the minimum model extrinsic parameter can be directly determined as the optimized external layer parameter, which effectively ensures the accuracy and reliability of determining the optimized external layer parameter.

[0204] After obtaining the optimized outer layer parameters and the optimized model internal parameters, the optimization operation of the click rate prediction model can be implemented based on the above optimized outer layer parameters and the optimized model internal parameters, which can specifically include the following steps:

[0205] Step 31: Optimize the click rate prediction model based on the optimized outer layer parameters and the optimized model internal parameters to obtain the optimized prediction model.

[0206] Step 32: In response to a click rate prediction request for the object to be predicted, a click rate prediction operation may be performed on the object to be predicted using the optimized prediction model, thereby obtaining a click rate prediction result for the object to be predicted.

[0207] Tests have shown that the click-through rate prediction operation implemented by the optimized prediction model can not only ensure the accuracy of the click-through rate prediction results, but also reduce the error rate of the click-through rate prediction on the data set, thereby ensuring the practicality of the method.

[0208] The technical solution provided by the present application embodiment optimizes the click-through rate prediction model through the Riemann implicit method, decoupling the calculation of the gradient of the outer parameter and the optimization process of the gradient of the inner parameter, which can greatly reduce the time complexity and space complexity. In addition, in the process of optimizing the click-through rate prediction model, it is not necessary to save the calculation graph involved in the gradient optimization process of the inner parameter, which greatly reduces the demand for computing storage space. In addition, in the process of determining the optimized outer parameters and the optimized model internal parameters based on implicit gradients, it is not necessary to expand and differentiate the optimization operation of the inner parameter, thereby greatly reducing the time complexity of the outer gradient, thus achieving efficient optimization of the click-through rate prediction model, and in the process of performing click-through rate prediction operations using the optimized click-through rate prediction model, the accuracy and reliability of the click-through rate prediction operation can be improved, further improving the practicality of the method.

[0209] Figure 5A flow chart of a sample identification method provided by an embodiment of the present invention; see attached Figure 5 As shown, this embodiment provides a sample identification method, the execution subject of the method is a sample identification device. Specifically, the sample identification device can be implemented as software, or a combination of software and hardware. When the sample identification device is implemented as hardware, it can be various electronic devices that can implement sample identification operations, including but not limited to personal computers, servers, databases, etc. When the sample identification device is implemented as software, it can be installed in the electronic devices listed above. Based on the above sample identification device, the sample identification operation can be implemented. Specifically, the sample identification method may include:

[0210] Step S501: Obtain samples to be identified.

[0211] Step S502: Determine a neural network model for implementing a sample recognition operation, where the neural network model has corresponding model parameters.

[0212] Step S503: Map the model parameters to the topological space, and perform Riemann optimization on the model parameters to obtain optimized model parameters.

[0213] Among them, in the application scenario of sample recognition, the model parameters can be implemented as weight information of the neural network model.

[0214] Step S504: Update the neural network model based on the optimized model parameters to obtain the optimized model.

[0215] Step S505: Process the sample to be identified based on the optimized model to obtain the identification result.

[0216] The specific implementation methods and implementation effects of the above steps in this embodiment are the same as those in the above Figure 2 The specific implementation method and implementation effect of steps S201 to S205 in the illustrated embodiment are similar, and the details can be referred to the above statements, which will not be repeated here.

[0217] In some other instances, with respect to the above-mentioned step S503, in order to ensure the optimization quality and effect of the neural network model, when performing Riemann optimization operation on the model parameters, not only can the model parameters be mapped to the topological space, but the sample features of the samples to be identified can also be obtained, and then the sample features and model parameters can be mapped to the topological space, and then the model parameters can be Riemann optimized based on the sample features mapped on the Riemann manifold, so as to obtain the optimized model parameters, which effectively ensures the quality and effect of the Riemann optimization operation on the model parameters in the application scenario of sample identification.

[0218] In addition, the sample identification method in this embodiment may also include the above Figure 2-Figure 4 For the method of the embodiment shown in the figure, the part not described in detail in this embodiment can be referred to Figure 2-Figure 4 The implementation process and technical effects of this technical solution refer to Figure 2-Figure 4 The description in the illustrated embodiment will not be repeated here.

[0219] Figure 6 A flowchart of an information recommendation method provided by an embodiment of the present invention; see attached Figure 6 As shown, this embodiment provides an information recommendation method, the execution subject of the method is an information recommendation device. Specifically, the information recommendation device can be implemented as software, or a combination of software and hardware. When the information recommendation device is implemented as hardware, it can be various electronic devices that can implement information recommendation operations, including but not limited to personal computers, servers, databases, etc. When the information recommendation device is implemented as software, it can be installed in the electronic devices listed above. Based on the above information recommendation device, the information recommendation operation can be implemented. Specifically, the information recommendation method may include:

[0220] Step S601: Obtain the object to be recommended, and the object to be recommended has corresponding portrait information.

[0221] Step S602: Determine a neural network model for implementing the advertisement recommendation operation, where the neural network model has corresponding model parameters.

[0222] Step S603: Map the model parameters and the portrait information to the topological space, and perform Riemann optimization on the model parameters to obtain the optimized model parameters.

[0223] Step S604: Update the neural network model based on the optimized model parameters to obtain the optimized model.

[0224] Step S605: Based on the optimized model and portrait information, determine the recommendation information that matches the object to be recommended.

[0225] The specific implementation methods and implementation effects of the above steps in this embodiment are the same as those in the above Figure 2 The specific implementation method and implementation effect of steps S201 to S205 in the illustrated embodiment are similar, and the details can be referred to the above statements, which will not be repeated here.

[0226] In addition, the sample identification method in this embodiment may also include the above Figure 2-Figure 4 For the method of the embodiment shown in the figure, the part not described in detail in this embodiment can be referred to Figure 2-Figure 4 The implementation process and technical effects of this technical solution refer to Figure 2-Figure 4 The description in the illustrated embodiment will not be repeated here.

[0227] Figure 7 A schematic diagram of a click rate prediction device provided by an embodiment of the present invention; Figure 7 As shown, this embodiment provides a click rate prediction device, which is used to perform Figure 2 The click rate prediction method shown in the figure, the click rate prediction device may include:

[0228] A first acquisition module 11 is used to acquire a click rate prediction request for a preset object;

[0229] A first determination module 12 is used to determine a neural network model for implementing a click rate prediction operation based on a click rate prediction request, wherein the neural network model has model parameters corresponding thereto;

[0230] The first processing module 13 is used to map the model parameters to the topological space and perform Riemann optimization on the model parameters to obtain the optimized model parameters;

[0231] A first processing module 13 is used to update the neural network model based on the optimized model parameters to obtain the optimized model;

[0232] The first processing module 13 is further used to perform a click rate prediction operation based on the optimized model to obtain a click rate prediction result corresponding to a preset object.

[0233] In some instances, when the first processing module 13 performs Riemann optimization on the model parameters to obtain the optimized model parameters, the first processing module 13 is used to execute: constructing a loss function corresponding to the neural network model; performing Riemann optimization on the loss function to obtain the minimum value of the loss function; and determining the target parameter corresponding to the minimum value as the optimized model parameter.

[0234] In some instances, when the first processing module 13 constructs a loss function corresponding to the model parameters, the first processing module 13 is used to execute: mapping the model parameters to the Riemann space to obtain mapping parameters corresponding to the model parameters; and constructing a loss function corresponding to the neural network model based on the model parameters and the mapping parameters.

[0235] In some instances, when the first processing module 13 performs Riemann optimization on the loss function to obtain the minimum value of the loss function, the first processing module 13 is used to execute: based on the loss function, determine the gradient of the model parameters; determine the inner iteration step for Riemann optimization of the loss function; based on the gradient of the model parameters and the inner iteration step, and using the Riemann gradient descent method, iteratively process the loss function to obtain the minimum value of the loss function.

[0236] In some instances, in the process of iteratively processing the loss function based on the gradient of the model parameters and the inner iteration step and using the gradient descent method, the first processing module 13 in this embodiment is also used to perform the following steps: mapping the gradient of the model parameters to the tangent space to obtain the mapping gradient; determining the iterative model parameters corresponding to the model parameters based on the mapping gradient; and updating the model parameters to the iterative model parameters.

[0237] In some instances, when the first processing module 13 determines the optimized model based on the optimized model parameters, the first processing module 13 is used to execute: obtaining preset outer layer parameters corresponding to the neural network model, the preset outer layer parameters including at least one of the following: hyperparameters, initialization parameters; optimizing the preset outer layer parameters to obtain optimized outer layer parameters corresponding to the preset outer layer parameters; updating the neural network model based on the optimized model parameters and the optimized outer layer parameters to determine the optimized model.

[0238] In some instances, when the first processing module 13 optimizes the preset outer layer parameters to obtain optimized outer layer parameters corresponding to the preset outer layer parameters, the first processing module 13 is used to execute: based on the optimized model parameters, determining the objective function corresponding to the click-through rate prediction request, the objective function being a binary classification function related to the preset outer layer parameters and used to identify whether to perform a click operation; optimizing the preset outer layer parameters based on the objective function, and determining the optimized outer layer parameters corresponding to the preset outer layer parameters.

[0239] In some instances, when the first processing module 13 optimizes the preset outer layer parameters based on the objective function and determines the optimized outer layer parameters corresponding to the preset outer layer parameters, the first processing module 13 is used to execute: based on the objective function, determine the gradient of the preset outer layer parameters; determine the outer layer iteration step for Riemann optimization of the objective function; based on the gradient of the preset outer layer parameters and the outer layer iteration step, and using the Riemann gradient descent method, iteratively update the preset outer layer parameters to obtain the minimum preset outer layer parameters; determine the minimum preset outer layer parameters as the optimized outer layer parameters.

[0240] In some instances, when the first processing module 13 determines the gradient of a preset outer parameter based on an objective function, the first processing module 13 is used to perform: determining the original gradient of the preset outer parameter based on the objective function, the original gradient including a matrix representation having a time complexity greater than a preset threshold; rewriting the matrix representation in the original gradient into a preset sequence to obtain the gradient of the initialization parameter, wherein the error between the gradient of the initialization parameter determined based on the preset sequence and the gradient of the initialization parameter determined based on the matrix representation is less than or equal to the preset threshold.

[0241] Figure 7 The device shown can perform Figure 1-Figure 4For the method of the embodiment shown in the figure, the part not described in detail in this embodiment can be referred to Figure 1-Figure 4 The implementation process and technical effects of this technical solution refer to Figure 1-Figure 4 The description in the illustrated embodiment will not be repeated here.

[0242] In one possible design, Figure 7 The structure of the click rate prediction device shown in the figure can be implemented as an electronic device, which can be a controller, a personal computer, a server, or other devices. Figure 8 As shown, the electronic device may include: a first processor 21 and a first memory 22. The first memory 22 is used to store the corresponding electronic device to execute the above Figure 1-Figure 4 The program of the click rate prediction method provided in the illustrated embodiment, the first processor 21 is configured to execute the program stored in the first memory 22 .

[0243] The program includes one or more computer instructions, wherein one or more computer instructions can implement the following steps when executed by the first processor 21: obtaining a click rate prediction request for a preset object; based on the click rate prediction request, determining a neural network model for implementing a click rate prediction operation, wherein the neural network model corresponds to model parameters; mapping the model parameters to a topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; updating the neural network model based on the optimized model parameters to obtain an optimized model; performing a click rate prediction operation based on the optimized model to obtain a click rate prediction result corresponding to the preset object.

[0244] Furthermore, the first processor 21 is also used to execute the aforementioned Figure 1-Figure 4 All or part of the steps in the illustrated embodiments.

[0245] The structure of the electronic device may further include a first communication interface 23 for the electronic device to communicate with other devices or a communication network.

[0246] In addition, an embodiment of the present invention provides a computer storage medium for storing computer software instructions used by electronic devices, which includes instructions for executing the above Figure 1-Figure 4 The procedures involved in the method for predicting click rate in the illustrated embodiment.

[0247] In addition, an embodiment of the present invention provides a computer program product, including: a computer-readable storage medium storing computer instructions, when the computer instructions are executed by one or more processors, causing the one or more processors to execute the above Figure 1-Figure 4 The steps in the method for predicting click rate in the method embodiment are shown.

[0248] Fig. 9A schematic diagram of a sample identification device provided by an embodiment of the present invention; Fig. 9 As shown, this embodiment provides a sample identification device, which is used to perform Figure 5 The sample identification method shown, the sample identification device may include:

[0249] The second acquisition module 31 is used to acquire a sample to be identified;

[0250] A second determination module 32 is used to determine a neural network model for implementing a sample recognition operation, and the neural network model corresponds to model parameters;

[0251] The second processing module 33 is used to map the model parameters to the topological space and perform Riemann optimization on the model parameters to obtain the optimized model parameters;

[0252] The second processing module 33 is used to update the neural network model based on the optimized model parameters to obtain the optimized model;

[0253] The second processing module 33 is further used to process the sample to be identified based on the optimized model to obtain the identification result.

[0254] Fig. 9 The device shown can perform Figure 5 For the method of the embodiment shown in the figure, the part not described in detail in this embodiment can be referred to Figure 5 The implementation process and technical effects of this technical solution refer to Figure 5 The description in the illustrated embodiment will not be repeated here.

[0255] In one possible design, Fig. 9 The structure of the sample identification device shown can be implemented as an electronic device, which can be a controller, a personal computer, a server, or other devices. Fig.10 As shown, the electronic device may include: a second processor 41 and a second memory 42. The second memory 42 is used to store the corresponding electronic device executing the above Figure 5 The program of the sample identification method provided in the illustrated embodiment, the second processor 41 is configured to execute the program stored in the second memory 42 .

[0256] The program includes one or more computer instructions, wherein one or more computer instructions can implement the following steps when executed by the second processor 41: obtaining a sample to be identified; determining a neural network model for implementing a sample identification operation, wherein the neural network model has corresponding model parameters; mapping the model parameters to a topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; updating the neural network model based on the optimized model parameters to obtain an optimized model; processing the sample to be identified based on the optimized model to obtain an identification result.

[0257] Furthermore, the second processor 41 is also used to execute the aforementioned Figure 5 All or part of the steps in the illustrated embodiments.

[0258] The structure of the electronic device may further include a second communication interface 43 for the electronic device to communicate with other devices or a communication network.

[0259] In addition, an embodiment of the present invention provides a computer storage medium for storing computer software instructions used by electronic devices, which includes instructions for executing the above Figure 5 The procedures involved in the sample identification method in the illustrated embodiment.

[0260] In addition, an embodiment of the present invention provides a computer program product, including: a computer-readable storage medium storing computer instructions, when the computer instructions are executed by one or more processors, causing the one or more processors to execute the above Figure 5 The steps in the sample identification method in the method embodiment are shown.

[0261] Fig.11 A schematic diagram of the structure of an information recommendation device provided by an embodiment of the present invention; Fig.11 As shown, this embodiment provides an information recommendation device, the sample identification device is used to perform Figure 6 The information recommendation method shown, the information recommendation device may include:

[0262] The third acquisition module 51 is used to acquire the object to be recommended, and the object to be recommended has corresponding portrait information;

[0263] A third determination module 52 is used to determine a neural network model for implementing an advertisement recommendation operation, and the neural network model corresponds to model parameters;

[0264] The third processing module 53 is used to map the model parameters and the image information to the topological space, and perform Riemann optimization on the model parameters to obtain the optimized model parameters;

[0265] A third processing module 53 is used to update the neural network model based on the optimized model parameters to obtain an optimized model;

[0266] The third processing module 53 is further used to determine recommendation information matching the object to be recommended based on the optimized model and portrait information.

[0267] Fig.11 The device shown can perform Figure 6 For the method of the embodiment shown in the figure, the part not described in detail in this embodiment can be referred to Figure 6 The implementation process and technical effects of this technical solution refer to Figure 6 The description in the exemplary embodiments is omitted here.

[0268] In one possible design, Fig.11 The structure of the information recommendation device shown can be implemented as an electronic device, which can be a controller, a personal computer, a server, or other devices. Fig.12 As shown, the electronic device may include: a third processor 61 and a third memory 62. The third memory 62 is used to store the corresponding electronic device executing the above Figure 6 The program of the information recommendation method provided in the illustrated embodiment, the third processor 61 is configured to execute the program stored in the third memory 62 .

[0269] The program includes one or more computer instructions, wherein one or more computer instructions can implement the following steps when executed by the third processor 61: obtaining an object to be recommended, corresponding to portrait information of the object to be recommended; determining a neural network model for implementing the advertising recommendation operation, corresponding to the neural network model is model parameters; mapping the model parameters and portrait information to the topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; updating the neural network model based on the optimized model parameters to obtain the optimized model; and determining recommendation information matching the object to be recommended based on the optimized model and portrait information.

[0270] Furthermore, the third processor 61 is also used to execute the aforementioned Figure 6 All or part of the steps in the illustrated embodiments.

[0271] The structure of the electronic device may further include a third communication interface 63 for the electronic device to communicate with other devices or a communication network.

[0272] In addition, an embodiment of the present invention provides a computer storage medium for storing computer software instructions used by electronic devices, which includes instructions for executing the above Figure 6 The procedures involved in the information recommendation method in the illustrated embodiment.

[0273] In addition, an embodiment of the present invention provides a computer program product, including: a computer-readable storage medium storing computer instructions, when the computer instructions are executed by one or more processors, causing the one or more processors to execute the above Figure 6 The steps in the information recommendation method in the method embodiment are shown.

[0274] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0275] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0276] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0277] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable device to produce a machine, so that the instructions executed by the processor of the computer or other programmable device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0278] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable device so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide the functions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps of the functions specified in the blocks or multiple blocks. In a typical configuration, the computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory. The memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0279] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0280] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting click rate, characterized in that: include: Get click rate prediction request for preset objects; Based on the click-through rate prediction request, determining a neural network model for implementing a click-through rate prediction operation, wherein the neural network model corresponds to model parameters; Mapping the model parameters to the topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; Update the neural network model based on the optimized model parameters to obtain an optimized model; A click rate prediction operation is performed based on the optimized model to obtain a click rate prediction result corresponding to the preset object.

2. The method according to claim 1, characterized in that Performing Riemann optimization on the model parameters to obtain optimized model parameters includes: Constructing a loss function corresponding to the neural network model; Performing Riemann optimization on the loss function to obtain a minimum value of the loss function; The target parameter corresponding to the minimum value is determined as the optimized model parameter.

3. The method according to claim 2, characterized in that Constructing a loss function corresponding to the model parameters, including: Mapping the model parameters into the Riemann space to obtain mapping parameters corresponding to the model parameters; Based on the model parameters and the mapping parameters, a loss function corresponding to the neural network model is constructed.

4. The method according to claim 2, characterized in that: Performing Riemann optimization on the loss function to obtain the minimum value of the loss function includes: Based on the loss function, determining the gradient of the model parameters; Determining an inner iteration step size for performing Riemann optimization on the loss function; Based on the gradient of the model parameters and the inner layer iteration step, the loss function is iteratively processed using the Riemann gradient descent method to obtain the minimum value of the loss function.

5. The method according to claim 4, characterized in that In the process of iteratively processing the loss function based on the gradient of the model parameters and the inner layer iteration step size and using the gradient descent method, the method further includes: Mapping the gradient of the model parameters to the tangent space to obtain a mapping gradient; Based on the mapping gradient, determining iterative model parameters corresponding to the model parameters; The model parameters are updated to the iterative model parameters.

6. The method according to any one of claims 1 to 5, characterized in that: Determining the optimized model based on the optimized model parameters includes: Acquire preset outer layer parameters corresponding to the neural network model, wherein the preset outer layer parameters include at least one of the following: hyperparameters and initialization parameters; Optimizing the preset outer layer parameters to obtain optimized outer layer parameters corresponding to the preset outer layer parameters; The neural network model is updated based on the optimized model parameters and the optimized outer layer parameters to determine the optimized model.

7. The method according to claim 6, characterized in that Optimizing the preset outer layer parameters to obtain optimized outer layer parameters corresponding to the preset outer layer parameters includes: Based on the optimized model parameters, determining an objective function corresponding to the click rate prediction request, wherein the objective function is a binary classification function related to the preset outer layer parameters and used to identify whether a click operation is performed; The preset outer layer parameters are optimized based on the objective function to determine optimized outer layer parameters corresponding to the preset outer layer parameters.

8. The method according to claim 7, characterized in that Optimizing the preset outer layer parameters based on the objective function and determining optimized outer layer parameters corresponding to the preset outer layer parameters includes: Based on the objective function, determining the gradient of the preset outer layer parameter; Determining an outer iteration step size for performing Riemann optimization on the objective function; Based on the gradient of the preset outer layer parameter and the outer layer iteration step, the preset outer layer parameter is iteratively updated by using the Riemann gradient descent method to obtain the minimum preset outer layer parameter; The minimum preset outer layer parameter is determined as the optimized outer layer parameter.

9. The method according to claim 8, characterized in that Based on the objective function, determining the gradient of the preset outer layer parameter includes: Based on the objective function, determining the original gradient of the preset outer layer parameter, wherein the original gradient includes a matrix representation whose time complexity is greater than a preset threshold; The matrix representation in the original gradient is rewritten into a preset sequence to obtain a gradient of the initialization parameter, wherein an error between a gradient of the initialization parameter determined based on the preset sequence and a gradient of the initialization parameter determined based on the matrix representation is less than or equal to a preset threshold.

10. A sample identification method, characterized in that: include: Obtain samples to be identified; Determining a neural network model for implementing a sample recognition operation, wherein the neural network model corresponds to model parameters; Mapping the model parameters to the topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; Update the neural network model based on the optimized model parameters to obtain an optimized model; The sample to be identified is processed based on the optimized model to obtain an identification result.

11. An information recommendation method, characterized in that: include: Obtaining an object to be recommended, wherein the object to be recommended has corresponding portrait information; Determining a neural network model for implementing an advertisement recommendation operation, wherein the neural network model corresponds to model parameters; Mapping the model parameters and the portrait information to a topological space, and performing Riemann optimization on the model parameters to obtain optimized model parameters; Update the neural network model based on the optimized model parameters to obtain an optimized model; Based on the optimized model and the portrait information, recommendation information matching the object to be recommended is determined.

12. A click rate prediction device, characterized in that: include: A first acquisition module, used to acquire a click rate prediction request for a preset object; A first determination module is used to determine a neural network model for implementing a click-through rate prediction operation based on the click-through rate prediction request, wherein the neural network model corresponds to a model parameter; A first processing module is used to map the model parameters to a topological space and perform Riemann optimization on the model parameters to obtain optimized model parameters; The first processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model; The first processing module is further used to perform a click-through rate prediction operation based on the optimized model to obtain a click-through rate prediction result corresponding to the preset object.

13. A sample identification device, characterized in that: include: The second acquisition module is used to acquire the sample to be identified; A second determination module is used to determine a neural network model for implementing a sample recognition operation, wherein the neural network model corresponds to model parameters; A second processing module is used to map the model parameters to the topological space and perform Riemann optimization on the model parameters to obtain optimized model parameters; The second processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model; The second processing module is further used to process the sample to be identified based on the optimized model to obtain an identification result.

14. An information recommendation device, characterized in that: include: The third acquisition module is used to acquire the object to be recommended, and the object to be recommended has corresponding portrait information; A third determination module is used to determine a neural network model for implementing an advertisement recommendation operation, wherein the neural network model corresponds to model parameters; A third processing module is used to map the model parameters and the portrait information to a topological space, and perform Riemann optimization on the model parameters to obtain optimized model parameters; The third processing module is used to update the neural network model based on the optimized model parameters to obtain an optimized model; The third processing module is further used to determine recommendation information matching the object to be recommended based on the optimized model and the portrait information.

15. An electronic device, characterized in that: include: A memory, a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement the method as described in any one of claims 1-11.

16. A computer storage medium, characterized in that: Used to store a computer program, which enables a computer to implement the method of any one of claims 1 to 11 when executed by the computer.

17. A computer program product, characterized in that include: A computer program, when executed by a processor of an electronic device, causes the processor to execute the steps of any one of the methods of claims 1-11.

Images