Media information recommendation method and device, equipment, storage medium and program product

Abstract

The invention provides a media information recommendation method and device, equipment, a storage medium and a program product. The method comprises the steps of obtaining an information feature set and a behavior feature set; performing importance evaluation on each media information feature in the information feature set to obtain information importance corresponding to each media information feature, and performing feature transformation on each media information feature in the information feature set based on the information importance to obtain a target information feature set; performing importance degree evaluation on each historical behavior feature in the behavior feature set to obtain a behavior importance degree corresponding to each historical behavior feature, and performing feature transformation on each historical behavior feature in the behavior feature set based on the behavior importance degree to obtain a target behavior feature set; and performing media information recommendation on the to-be-recommended object in combination with the target behavior feature set and the target information feature set. According to the invention, the recommendation accuracy of the media information can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method, device, equipment, storage medium and program product for recommending media information. Background Art

[0002] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. Artificial intelligence technology is an interdisciplinary subject that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies.

[0003] In the related art, the features of the object to be recommended are usually constructed based on the basic static information of the object to be recommended, such as the age and gender of the object to be recommended, and media information is recommended based on the constructed features of the object to be recommended. However, for objects to be recommended with the same basic static information, there may be different preferences for media information, which leads to low recommendation accuracy. Summary of the invention

[0004] The embodiments of the present application provide a method, device, electronic device, computer-readable storage medium, and computer program product for recommending media information, which can effectively improve the accuracy of recommending media information.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] The present application provides a method for recommending media information, including:

[0007] Acquire an information feature set including a plurality of media information features and a behavior feature set including a plurality of historical behavior features;

[0008] The multiple media information features are features of media information interacted by the to-be-recommended object at multiple different historical moments, and the multiple historical behavior features are features corresponding to the interactive behaviors of the to-be-recommended object at each of the historical moments;

[0009] Performing importance evaluation on each of the media information features in the information feature set to obtain information importance corresponding to each of the media information features, and performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set;

[0010] Performing importance evaluation on each of the historical behavior features in the behavior feature set to obtain the behavior importance corresponding to each of the historical behavior features, and performing feature transformation on each of the historical behavior features in the behavior feature set based on the behavior importance to obtain a target behavior feature set;

[0011] The target behavior feature set and the target information feature set are combined to recommend media information to the object to be recommended.

[0012] The present application provides a media information recommendation device, including:

[0013] An acquisition module, used to acquire an information feature set including a plurality of media information features and a behavior feature set including a plurality of historical behavior features; wherein the plurality of media information features are features of media information interacted with by the object to be recommended at a plurality of different historical moments, and the plurality of historical behavior features are features corresponding to the interactive behaviors of the object to be recommended at each of the historical moments;

[0014] a media evaluation module, configured to evaluate the importance of each of the media information features in the information feature set, obtain the information importance corresponding to each of the media information features, and perform feature transformation on each of the media information features in the information feature set based on the information importance, to obtain a target information feature set;

[0015] A behavior evaluation module is used to evaluate the importance of each of the historical behavior features in the behavior feature set to obtain the behavior importance corresponding to each of the historical behavior features, and based on the behavior importance, perform feature transformation on each of the historical behavior features in the behavior feature set to obtain a target behavior feature set;

[0016] The recommendation module is used to recommend media information to the object to be recommended by combining the target behavior feature set and the target information feature set.

[0017] In the above scheme, the media information feature includes multiple sub-information features, each of which includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature; the above-mentioned media evaluation module is also used to perform the following processing for each media information feature: for each sub-information feature in the media information feature, the values ​​of each feature element in the sub-information feature are weighted averaged to obtain a reference value of the sub-information feature; combined with the reference values ​​of each sub-information feature in the media information feature, the target importance of each sub-information feature in the media information feature is determined.

[0018] In the above scheme, the above media evaluation module is also used to construct a first reference feature of the media information feature based on the reference values ​​of each of the sub-information features in the media information feature; wherein the feature elements in the first reference feature correspond one-to-one to the sub-information features in the media information feature, and the values ​​of the feature elements in the first reference feature are equal to the reference values ​​of the corresponding sub-information features in the media information feature; the first reference feature is dimensionally reduced to obtain a reduced-dimensional feature, and the reduced-dimensional feature is stretched to obtain a target feature; wherein the target feature has the same dimension as the first reference feature, and the feature elements in the target feature correspond one-to-one to the feature elements in the first reference feature; and the values ​​of each feature element in the target feature are respectively determined as the target importance of the corresponding sub-information feature.

[0019] In the above scheme, the above media evaluation module is also used to obtain compression features, and the compression features are used to compress the feature dimension of the first reference feature; the feature product of the compression features and the first reference feature is determined as a candidate dimensionality reduction feature; and the candidate dimensionality reduction feature is nonlinearly transformed to obtain the dimensionality reduction feature.

[0020] In the above scheme, the above media evaluation module is also used to obtain a stretching feature, and the stretching feature is used to stretch the feature dimension of the dimensionality reduction feature; the product of the dimensionality reduction feature and the stretching feature is determined as a candidate target feature; the values ​​of each feature element in the candidate target feature are normalized respectively to obtain the normalized value corresponding to each feature element; the value of each feature element in the candidate target feature is replaced by the corresponding normalized value to obtain the target feature.

[0021] In the above scheme, the media information feature includes multiple sub-information features, each of which includes multiple feature elements, and the information importance includes the target importance of each of the sub-information features in the corresponding media information feature; the above media evaluation module is also used to multiply each of the sub-information features in the media information feature by the corresponding target importance for each of the media information features in the information feature set to obtain updated sub-information features corresponding to each of the sub-information features; replace each of the sub-information features of each of the media information features in the information feature set with the corresponding updated sub-information features to obtain a candidate information feature set; and perform linear transformation on each of the updated sub-information features in the candidate information feature set to obtain the target information feature set.

[0022] In the above scheme, the media information feature includes multiple sub-information features, each of which includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature; the above media evaluation module is also used to multiply each sub-information feature in the media information feature by the corresponding target importance for each media information feature in the information feature set, so as to obtain updated sub-information features corresponding to each sub-information feature; and replace each sub-information feature of each media information feature in the information feature set with the corresponding updated sub-information feature, so as to obtain the target information feature set.

[0023] In the above scheme, the above media evaluation module is also used to obtain multiple importance prediction networks, and different importance prediction networks correspond to different network parameters; for each importance prediction network, the importance prediction network is called to perform importance evaluation on each media information feature in the information feature set to obtain the information importance corresponding to each media information feature; the above media evaluation module is also used to obtain multiple feature transformation networks, and the feature transformation networks correspond to the importance prediction networks one by one; for each feature transformation network, the feature transformation network is called to perform feature transformation on each media information feature in the information feature set based on the corresponding information importance to obtain a candidate information feature set corresponding to the feature transformation network; and feature fusion is performed on each candidate information feature set to obtain the target information feature set.

[0024] In the above scheme, the above media evaluation module is also used to obtain an importance prediction network, and call the importance prediction network to perform importance evaluation on each of the media information features in the information feature set to obtain the information importance corresponding to each of the media information features; the above media evaluation module is also used to obtain a feature transformation network, and call the feature transformation network to perform feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set corresponding to the feature transformation network.

[0025] In the above scheme, the above historical behavior characteristics include multiple sub-behavior characteristics, each of which includes multiple characteristic elements, and the behavior importance includes the target importance of each sub-behavior characteristic in the corresponding historical behavior characteristics; the above behavior evaluation module is also used to perform the following processing for each historical behavior characteristic: for each sub-behavior characteristic in the historical behavior characteristics, weighted average the values ​​of each characteristic element in the sub-behavior characteristic to obtain a reference value of the sub-behavior characteristic; and determine the target importance of each sub-behavior characteristic in the historical behavior characteristics in combination with the reference values ​​of each sub-behavior characteristic in the historical behavior characteristics.

[0026] In the above scheme, the above-mentioned behavior evaluation module is also used to construct a second reference feature of the historical behavior feature based on the reference values ​​of each sub-behavior feature in the historical behavior feature; wherein the feature elements in the second reference feature correspond one-to-one to the sub-behavior features in the historical behavior feature, and the values ​​of the feature elements in the second reference feature are equal to the reference values ​​of the corresponding sub-behavior features in the historical behavior feature; the second reference feature is dimensionally reduced to obtain the reduced-dimensional feature of the second reference feature, and the reduced-dimensional feature of the second reference feature is stretched to obtain the target feature of the second reference feature; wherein the target feature of the second reference feature has the same dimension as that of the second reference feature, and the feature elements in the target feature of the second reference feature correspond one-to-one to the feature elements in the second reference feature; the values ​​of each feature element in the second reference feature are respectively determined as the target importance of the corresponding sub-behavior feature.

[0027] In the above scheme, the historical behavior feature includes multiple sub-behavior features, each of which includes multiple feature elements, and the behavior importance includes the target importance of each sub-behavior feature in the corresponding historical behavior feature; the above-mentioned behavior evaluation module is also used to multiply each sub-behavior feature in the historical behavior feature by the corresponding target importance for each historical behavior feature in the behavior feature set to obtain an updated sub-behavior feature corresponding to each sub-behavior feature; replace each sub-behavior feature of each historical behavior feature in the behavior feature set with the corresponding updated sub-behavior feature to obtain a candidate behavior feature set; and perform linear transformation on each updated sub-behavior feature in the candidate behavior feature set to obtain the target behavior feature set.

[0028] In the above scheme, the target behavior feature set includes multiple target sub-behavior features, the target information feature set includes multiple target sub-information features, and the above recommendation module is also used to perform feature fusion on the multiple target sub-behavior features and the multiple target sub-information features to obtain target fusion features; call the recommendation network to predict the media information to be recommended of the object to be recommended based on the target fusion features to obtain the media information to be recommended of the object to be recommended; and send the media information to be recommended to the object to be recommended.

[0029] An embodiment of the present application provides an electronic device, including:

[0030] A memory for storing computer executable instructions or computer programs;

[0031] The processor is used to implement the media information recommendation method provided in the embodiment of the present application when executing the computer executable instructions or computer programs stored in the memory.

[0032] An embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for causing a processor to execute instructions to implement a method for recommending media information provided in an embodiment of the present application.

[0033] The embodiment of the present application provides a computer program product, which includes a computer program or a computer executable instruction, and the computer program or the computer executable instruction is stored in a computer-readable storage medium. The processor of the electronic device reads the computer executable instruction from the computer-readable storage medium, and the processor executes the computer executable instruction, so that the electronic device executes the media information recommendation method described in the embodiment of the present application.

[0034] The embodiments of the present application have the following beneficial effects:

[0035] In the process of recommending media information to the recommended object, the interactive behavior of the recommended object at each historical moment and the interacted media information are comprehensively considered, so as to comprehensively characterize the dynamic characteristics of the recommended object, and by determining the behavioral importance and information importance of the interactive behavior of the recommended object at each historical moment and the interacted media information, the interactive behavior of the recommended object at each historical moment and the interacted media information are further transformed according to the behavioral importance and the information importance to obtain a target behavior feature set and a target information feature set, so that the target behavior feature set and the target information feature set determined based on the behavioral importance and the information importance can accurately reflect the preference of the recommended object for different media information at different historical moments, and media information recommendation is performed on the recommended object through the highly credible target behavior feature set and the target information feature set, thereby effectively improving the accuracy of media information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the architecture of the media information recommendation system provided in the embodiment of the present application;

[0037] Figure 2 is a schematic diagram of the structure of an electronic device for media information recommendation provided in an embodiment of the present application;

[0038] Figure 3 to Figure 4 It is a flowchart of a method for recommending media information provided in an embodiment of the present application;

[0039] Figures 5 and 6 It is a schematic diagram of the principle of the importance prediction network provided in the embodiment of the present application;

[0040] Figure 7 is a schematic diagram of the structure of the importance prediction network provided in the embodiment of the present application;

[0041] Figures 8 to 12 It is a flowchart of a method for recommending media information provided in an embodiment of the present application;

[0042] Fig.13 It is a schematic diagram of an interface of a method for recommending media information provided in an embodiment of the present application;

[0043] Fig.14 It is a schematic diagram of the principle of the method for recommending media information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0045] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0046] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0048] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0049] 1) Machine Learning (ML): It is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance.

[0050] 2) Artificial Intelligence (AI): It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. Artificial intelligence technology is an interdisciplinary subject that covers a wide range of fields, including both hardware-level technology and software-level technology. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies.

[0051] 3) Convolutional Neural Networks (CNN): It is a type of feed forward neural network (FNN) that includes convolution calculations and has a deep structure. It is one of the representative algorithms of deep learning. Convolutional neural networks have representation learning capabilities and can perform shift-invariant classification on input images according to their hierarchical structure.

[0052] 4) In response to: used to indicate the conditions or states on which the executed operations depend. When the dependent conditions or states are met, one or more operations executed may be in real time or have a set delay. Unless otherwise specified, there is no restriction on the order in which the multiple operations executed are executed.

[0053] 5) Convolutional layer: Each convolutional layer in a convolutional neural network consists of several convolutional units, and the parameters of each convolutional unit are optimized through the back propagation algorithm. The purpose of the convolution operation is to extract different features of the input. The first convolutional layer may only extract some low-level features such as edges, lines, and corners. More layers of the network can iteratively extract more complex features from low-level features.

[0054] 6) Pooling layer: After the convolution layer performs feature extraction, the output feature map is passed to the pooling layer for feature selection and information filtering. The pooling layer contains a pre-set pooling function, which replaces the result of a single point in the feature map with the feature map statistics of its adjacent area. The pooling layer selects the pooling area in the same way as the convolution kernel scans the feature map, which is controlled by the pooling size, step size, and padding.

[0055] 7) Fully-Connected Layer: The fully connected layer in a convolutional neural network is equivalent to the hidden layer in a traditional feedforward neural network. The fully connected layer is located at the end of the hidden layer of the convolutional neural network and only transmits signals to other fully connected layers. The feature map loses its spatial topological structure in the fully connected layer, is expanded into a vector and passes through the activation function.

[0056] During the implementation of the embodiments of the present application, the applicant discovered that the related technology has the following problems:

[0057] With the continuous development of media information platforms such as online shopping platforms, recommendation systems have become an irreplaceable and important part of e-commerce. Recommendation systems can learn the preference information hidden in users' historical behaviors, thereby making further predictions on media information behaviors, helping users select satisfactory media information, and promoting the revenue of e-commerce platforms.

[0058] In the related art, the features of the object to be recommended are usually constructed based on the basic static information of the object to be recommended, such as the age and gender of the object to be recommended, and media information is recommended based on the constructed features of the object to be recommended. However, for objects to be recommended with the same basic static information, there may be different preferences for media information, which leads to low recommendation accuracy.

[0059] The embodiments of the present application provide a method, device, electronic device, computer-readable storage medium, and computer program product for recommending media information, which can effectively improve the accuracy of media information recommendation. The following describes an exemplary application of the media information recommendation system provided by the embodiments of the present application.

[0060] See also Figure 1 , Figure 11 is a schematic diagram of the architecture of a media information recommendation system 100 provided in an embodiment of the present application. A terminal (terminal 400 is shown as an example) is connected to a server 200 via a network 300. The network 300 may be a wide area network or a local area network, or a combination of the two.

[0061] The terminal 400 is used for the user to use the client 410 to display media information on a graphical interface 410-1 (graphic interface 410-1 is shown as an example). The terminal 400 and the server 200 are connected to each other via a wired or wireless network.

[0062] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The terminal 400 may be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart TV, a smart watch, a car terminal, etc., but is not limited thereto. The electronic device provided in the embodiment of the present application may be implemented as a terminal or as a server. The terminal and the server may be directly or indirectly connected via wired or wireless communication, which is not limited in the embodiment of the present application.

[0063] In some embodiments, the server 200 obtains an information feature set including multiple media information features and a behavior feature set including multiple historical behavior features, determines the information importance and the behavior importance, determines a target information feature set based on the information importance, determines a target behavior feature set based on the behavior importance, and sends the target behavior feature set and the target information feature set to the terminal 400. The terminal 400 combines the target behavior feature set and the target information feature set to make media information recommendations for the recommended object.

[0064] In other embodiments, the terminal 400 obtains an information feature set including multiple media information features and a behavior feature set including multiple historical behavior features, determines the information importance and the behavior importance, determines a target information feature set based on the information importance, determines a target behavior feature set based on the behavior importance, and recommends media information to the recommended object by combining the target behavior feature set and the target information feature set.

[0065] In other embodiments, the embodiments of the present application can be implemented with the aid of cloud technology. Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or a local area network to achieve data computing, storage, processing, and sharing.

[0066] Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, and application technology based on the cloud computing business model. It can form a resource pool that can be used on demand and is flexible and convenient. Cloud computing technology will become an important support. The background services of the technical network system require a large amount of computing and storage resources.

[0067] See also Figure 2 , Figure 2 is a schematic diagram of the structure of an electronic device 500 for media information recommendation provided in an embodiment of the present application, wherein: Figure 2 The electronic device 500 shown may be Figure 1 The server 200 or the terminal 400 in Figure 2 The electronic device 500 shown includes: at least one processor 410, a memory 450, and at least one network interface 420. The various components in the electronic device 500 are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to the data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 440 is not described in detail. Figure 2 Various buses are labeled as bus system 440 .

[0068] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0069] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0070] The memory 450 includes a volatile memory or a non-volatile memory, and may also include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0071] In some embodiments, memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplarily described below.

[0072] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0073] The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include: Bluetooth, Wireless Fidelity (WiFi), and Universal Serial Bus (USB).

[0074] In some embodiments, the media information recommendation device provided in the embodiments of the present application can be implemented in software. Figure 2 The media information recommendation device 455 stored in the memory 450 is shown, which can be software in the form of a program and a plug-in, etc., and includes the following software modules: an acquisition module 4551, a media evaluation module 4552, a behavior evaluation module 4553, and a recommendation module 4554. These modules are logical, and therefore can be arbitrarily combined or further split according to the functions implemented. The functions of each module will be described below.

[0075] In other embodiments, the media information recommendation device provided in the embodiments of the present application can be implemented in hardware. As an example, the media information recommendation device provided in the embodiments of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the media information recommendation method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor can adopt one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), or other electronic components.

[0076] In some embodiments, the terminal or server can implement the media information recommendation method provided in the embodiments of the present application by running a computer program or a computer executable instruction. For example, the computer program can be a native program (e.g., a dedicated recommendation program) or a software module in the operating system, for example, a recommendation module that can be embedded in any program (such as an instant messaging client, an album program, an electronic map client, a navigation client); for example, it can be a native application (APP, Application), that is, a program that needs to be installed in the operating system to run. In short, the above-mentioned computer program can be any form of application, module or plug-in.

[0077] The method for recommending media information provided by the embodiment of the present application will be described in conjunction with the exemplary application and implementation of the server or terminal provided by the embodiment of the present application.

[0078] See also Figure 3 , Figure 3 is a flowchart of a method for recommending media information provided by an embodiment of the present application, which will be combined with Figure 3 Steps 101 to 106 are shown for illustration. The media information recommendation method provided in the embodiment of the present application can be implemented by a server or a terminal alone, or by a server and a terminal in collaboration. The following description will be made by taking the server alone as an example.

[0079] In step 101, an information feature set including a plurality of media information features and a behavior feature set including a plurality of historical behavior features are obtained.

[0080] In some embodiments, the multiple media information features are features of media information interacted by the object to be recommended at multiple different historical moments, and the multiple historical behavior features are features corresponding to the interactive behaviors of the object to be recommended at each historical moment.

[0081] In some embodiments, the above media information includes recommendation information of various media types such as video information, audio information, and text information.

[0082] In some embodiments, the object to be recommended may be a client of a user to whom media information needs to be recommended. The user to be recommended may be a user among sample users or a new user. Any user to whom an item needs to be recommended may be considered a user to be recommended.

[0083] In some embodiments, the above step 101 can be implemented in the following ways: obtaining the media information interacted by the object to be recommended at multiple different historical moments, and performing feature extraction on the media information interacted at multiple different historical moments to obtain an information feature set including multiple media information features; obtaining the interactive behavior of the object to be recommended at each historical moment, and performing feature extraction on the interactive behavior of the object to be recommended at each historical moment to obtain a behavior feature set including multiple historical behavior features.

[0084] As an example, the information feature set includes multiple media information features, and the multiple media information features are features of the media information interacted by the object to be recommended at multiple different historical moments. For example, the information feature set includes features of the media information interacted by the object to be recommended at historical moment A (for example, video A1 and audio A2 interacted by the object to be recommended at historical moment A), and features of the media information interacted by the object to be recommended at historical moment B (for example, video B1 and audio B2 interacted by the object to be recommended at historical moment B).

[0085] As an example, the behavior feature set includes multiple historical behavior features, and the multiple historical behavior features are features corresponding to the interactive behaviors of the object to be recommended at each historical moment. For example, the historical behavior features include features corresponding to the interactive behavior of the object to be recommended at historical moment A (the object to be recommended switches the interactive media information at historical moment A), and features corresponding to the interactive behavior of the object to be recommended at historical moment B (the object to be recommended lowers the volume at historical moment B).

[0086] In step 102, importance evaluation is performed on each media information feature in the information feature set to obtain the information importance corresponding to each media information feature.

[0087] In some embodiments, the media information feature includes a plurality of sub-information features, each sub-information feature includes a plurality of feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature.

[0088] In some embodiments, the information importance corresponding to the media information feature is used to indicate the relative importance of the media information feature, and the information importance of the media information feature is associated with the weight of the media information feature in the information feature set.

[0089] As an example, the expression of the media information feature may be:

[0090]

[0091] Among them, M is used to indicate the characteristics of media information, M1, M2, ..., M n Used to indicate multiple sub-information features included in the media information feature, M 11 、M 12 、M 13 Multiple feature elements used to indicate sub-information feature M1, M 21 、M 22 、M 23 Multiple feature elements used to indicate sub-information feature M2, M n1 、M n2 、M n3 Used to indicate sub-information feature M n Multiple characteristic elements.

[0092] Continuing from the above example, the information importance of media information feature M includes sub-information features M1, M2, ..., M n importance of the target.

[0093] In some embodiments, see Figure 4 , Figure 4 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3 The step 102 shown can be performed separately for each media information feature. Figure 4 Steps 1021 to 1022 are implemented as shown.

[0094] In step 1021, for each sub-information feature in the media information feature, the value of each feature element in the sub-information feature is weighted averaged to obtain a reference value of the sub-information feature.

[0095] Continuing from the above example, for the sub-information feature M1 in the media information feature, the feature element M 11 、M 12 、M 13 The values ​​of are weighted averaged to obtain the reference value of the sub-information feature M1.

[0096] Continuing with the above example, the expression of the reference value of the sub-information feature M1 can be:

[0097]

[0098] Among them, M1′ is used to indicate the reference value of the sub-information feature M1, and M1′1 is used to indicate the feature element M 11 The value of M1′2 is used to indicate the characteristic element M 12 The value of M1′3 is used to indicate the characteristic element M 13 The value of .

[0099] Continuing with the above example, for the sub-information feature M2 in the media information feature, the feature element M 21 、M 22 、M 23 The values ​​of are weighted averaged to obtain the reference value of the sub-information feature M2.

[0100] Continuing with the above example, the expression of the reference value of the sub-information feature M2 can be:

[0101]

[0102] Among them, M2′ is used to indicate the reference value of the sub-information feature M2, and M2′1 is used to indicate the feature element M 21 The value of M2′2 is used to indicate the characteristic element M 22 The value of M2′3 is used to indicate the characteristic element M 23 The value of .

[0103] In step 1022, the target importance of each sub-information feature in the media information feature is determined in combination with the reference value of each sub-information feature in the media information feature.

[0104] In some embodiments, the above step 1022 may be implemented in the following manner: the reference value of each sub-information feature in the media information feature is determined as the target importance of the corresponding sub-information feature.

[0105] In other embodiments, the above-mentioned step 1022 can be implemented in the following manner: constructing a first reference feature of the media information feature based on the reference values ​​of each sub-information feature in the media information feature; performing dimensionality reduction on the first reference feature to obtain a reduced dimensionality feature, and performing stretching processing on the reduced dimensionality feature to obtain a target feature; and determining the value of each feature element in the target feature as the target importance of the corresponding sub-information feature.

[0106] In some embodiments, the feature elements in the first reference feature correspond one-to-one to the sub-information features in the media information feature, and the value of the feature element in the first reference feature is equal to the reference value of the corresponding sub-information feature in the media information feature.

[0107] As an example, the expression of the first reference feature may be:

[0108] C={M1′,M2′…M′ n} (4)

[0109] Wherein, C is used to indicate the first reference feature, M2′ is used to indicate the reference value of the sub-information feature M2, M1′ is used to indicate the reference value of the sub-information feature M1, and M′ n Used to indicate sub-information feature M n reference value.

[0110] Continuing with the above example, the feature element M1′ in the first reference feature corresponds one-to-one to the sub-information feature M1 in the media information feature, and the value of the feature element M1′ in the first reference feature is equal to the reference value M1′ of the corresponding sub-information feature in the media information feature.

[0111] In some embodiments, the target feature has the same dimension as the first reference feature, and feature elements in the target feature correspond one-to-one to feature elements in the first reference feature.

[0112] In some embodiments, the above-mentioned dimensionality reduction of the first reference feature to obtain the reduced dimensionality feature can be achieved in the following way: obtaining a compressed feature; multiplying the feature product of the compressed feature and the first reference feature to determine the candidate reduced dimensionality feature; performing a nonlinear transformation on the candidate reduced dimensionality feature to obtain the reduced dimensionality feature.

[0113] In some embodiments, the compression feature is used to compress the feature dimension of the first reference feature.

[0114] As an example, the expression of the dimension reduction feature can be:

[0115] R sq =f sq (Z) = a1(W sq ·Z T ) (5)

[0116] Among them, R sq Used to indicate the dimension reduction feature, W sq Used to indicate compression characteristics, Z T Used to indicate the first reference feature, W sq ·Z T It is used to indicate the candidate dimension reduction feature, a1 is used to indicate the nonlinear transformation activation function, f sq Used to indicate the compression function, R sq The dimension is K / r; the compressed feature W sq The dimension is

[0117] In some embodiments, the above-mentioned stretching processing of the reduced-dimensionality feature to obtain the target feature can be achieved in the following manner: obtaining the stretched feature; determining the product of the reduced-dimensionality feature and the stretched feature as the candidate target feature; normalizing the value of each feature element in the candidate target feature to obtain the normalized value corresponding to each feature element; replacing the value of each feature element in the candidate target feature with the corresponding normalized value to obtain the target feature.

[0118] In some embodiments, the stretching feature is used to stretch the feature dimension of the dimensionality reduction feature.

[0119] As an example, the above target feature can be expressed as:

[0120] R ex =f ex (R sq )=a2(W ex ·R sq ) (6)

[0121] Among them, R ex Used to indicate target features, R sq Used to indicate the dimension reduction feature, W ex is used to indicate the stretching feature, a2 is used to indicate the normalization function, W ex ·R sq Used to indicate candidate target features, R ex Used to scale K features, target feature R ex The dimension is a2 is the sigmoid function

[0122] In some embodiments, the above step 102 can be implemented in the following manner: obtain multiple importance prediction networks, and different importance prediction networks correspond to different network parameters; for each importance prediction network, call the importance prediction network, evaluate the importance of each media information feature in the information feature set, and obtain the information importance corresponding to each media information feature.

[0123] As an example, see Figure 5 , Figure 5It is a principle schematic diagram of the importance prediction network provided in the embodiment of the present application, obtaining importance prediction network 1 and importance prediction network 2, and the network parameters corresponding to importance prediction network 1 and importance prediction network 2 are different; for importance prediction network 1, importance prediction network 1 is called, and each media information feature (S1, S2, S3) in the information feature set is evaluated for importance, and the information importance corresponding to each media information feature (S1, S2, S3) is obtained. For importance prediction network 2, importance prediction network 2 is called, and each media information feature (S1, S2, S3) in the information feature set is evaluated for importance, and the information importance corresponding to each media information feature (S1, S2, S3) is obtained.

[0124] In other embodiments, the above step 102 can be implemented as follows: obtain an importance prediction network, call the importance prediction network, evaluate the importance of each media information feature in the information feature set, and obtain the information importance corresponding to each media information feature.

[0125] As an example, see Figure 6 , Figure 6 It is a principle diagram of the importance prediction network provided in an embodiment of the present application. Importance prediction network 1 is obtained, importance prediction network 1 is called, importance evaluation is performed on each media information feature (S1, S2, S3) in the information feature set, and the information importance corresponding to each media information feature (S1, S2, S3) is obtained.

[0126] In some embodiments, see Figure 7 , Figure 7 It is a structural diagram of the importance prediction network provided in an embodiment of the present application. The above-mentioned importance prediction network may include a pooling layer, a compression layer and a stretching layer. The above-mentioned calling of the importance prediction network to evaluate the importance of each media information feature in the information feature set to obtain the information importance corresponding to each media information feature can be achieved in the following manner: calling the pooling layer, for each sub-information feature in the media information feature, performing weighted averaging on the value of each feature element in the sub-information feature to obtain a reference value of the sub-information feature, and constructing a first reference feature of the media information feature based on the reference value of each sub-information feature in the media information feature; calling the compression layer to reduce the dimension of the first reference feature to obtain a reduced dimension feature; calling the stretching layer to stretch the reduced dimension feature to obtain a target feature; and determining the value of each feature element in the target feature as the target importance of the corresponding sub-information feature.

[0127] As an example, see Figure 7, calling the pooling layer 11, for each sub-information feature in the media information feature, performing weighted averaging on the values ​​of each feature element in the sub-information feature to obtain a reference value of the sub-information feature, and constructing a first reference feature of the media information feature based on the reference value of each sub-information feature in the media information feature; calling the compression layer 12, performing dimensionality reduction on the first reference feature to obtain a reduced dimensionality feature; calling the stretching layer 13, performing stretching on the reduced dimensionality feature to obtain a target feature; and determining the value of each feature element in the target feature as the target importance of the corresponding sub-information feature.

[0128] In this way, by determining the target importance of each sub-information feature in the media information feature, the reference value of the sub-information features at each historical moment in the media information feature can be accurately measured. By making the target importance of high efficiency, media information recommendations are made to the recommended objects, thereby effectively improving the accuracy of media information recommendations.

[0129] In step 103, based on the information importance, feature transformation is performed on each media information feature in the information feature set to obtain a target information feature set.

[0130] In some embodiments, the above-mentioned media information feature includes multiple sub-information features, the sub-information feature includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature.

[0131] In some embodiments, the information importance corresponding to the media information feature is used to indicate the relative importance of the media information feature, and the information importance of the media information feature is associated with the weight of the media information feature in the information feature set.

[0132] In some embodiments, see Figure 8 , Figure 8 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3 The illustrated step 103 may be performed by Figure 8 Steps 1031A to 1033A are implemented as shown.

[0133] In step 1031A, for each media information feature in the information feature set, each sub-information feature in the media information feature is multiplied by the corresponding target importance to obtain an updated sub-information feature corresponding to each sub-information feature.

[0134] As an example, the expression for updating the sub-information feature can be:

[0135] M Gi =Q i M i (7)

[0136] Among them, MGi Used to indicate the update sub-information characteristics, Q i It is used to indicate the target importance corresponding to the sub-information feature, M i Used to indicate sub-information characteristics.

[0137] In step 1032A, each sub-information feature of each media information feature in the information feature set is replaced with a corresponding updated sub-information feature to obtain a candidate information feature set.

[0138] As an example, the expression of the candidate information feature set can be:

[0139] M G ={M G1 , M G2 …M Gi} (8)

[0140] Among them, M G Used to indicate the candidate information feature set, M G1 To M Gi Used to indicate the characteristics of each updated sub-information.

[0141] In step 1033A, each updated sub-information feature in the candidate information feature set is linearly transformed to obtain a target information feature set.

[0142] In some embodiments, the above step 1033A can be implemented in the following manner: obtain a linear transformation network, call the linear transformation network, perform linear transformation on each updated sub-information feature in the candidate information feature set, and obtain a target information feature set.

[0143] In some embodiments, the above-mentioned linear transformation network can be various types of linear transformation networks such as a multi-head self-attention network and a self-attention network. The specific form of the linear transformation network does not constitute a limitation on the embodiments of the present application.

[0144] In this way, by performing feature transformation on each media information feature in the information feature set based on the information importance, a target information feature set is obtained, so that the target information feature set can accurately reflect the information importance between different sub-information features, and the weights between the sub-information features are reasonably allocated according to the information importance. By using a high-efficiency target information feature set, media information is recommended for the recommended object, thereby effectively improving the recommendation accuracy of the media information.

[0145] In some embodiments, see Fig. 9 , Fig. 9 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3 The illustrated step 103 may be performed by Fig. 9Steps 1031B to 1032B are implemented as shown.

[0146] In step 1031B, for each media information feature in the information feature set, each sub-information feature in the media information feature is multiplied by the corresponding target importance to obtain an updated sub-information feature corresponding to each sub-information feature.

[0147] As an example, the expression for updating the sub-information feature can be:

[0148] M Gi =Q i M i (9)

[0149] Among them, M Gi Used to indicate the update sub-information characteristics, Q i It is used to indicate the target importance corresponding to the sub-information feature, M i Used to indicate sub-information characteristics.

[0150] In step 1032B, each sub-information feature of each media information feature in the information feature set is replaced with a corresponding updated sub-information feature to obtain a target information feature set.

[0151] As an example, the expression of the target information feature set can be:

[0152] M G ={M G1 , M G2 …M Gi} (10)

[0153] Among them, M G Used to indicate the target information feature set, M G1 To M Gi Used to indicate the characteristics of each updated sub-information.

[0154] In this way, by performing feature transformation on each media information feature in the information feature set based on the information importance, a target information feature set is obtained, so that the target information feature set can accurately reflect the information importance between different sub-information features, and the weights between the sub-information features are reasonably allocated according to the information importance. By using a high-efficiency target information feature set, media information is recommended for the recommended object, thereby effectively improving the recommendation accuracy of the media information.

[0155] In some embodiments, the above step 103 can be implemented in the following manner: obtain multiple feature transformation networks, and the feature transformation networks correspond one-to-one to the importance prediction networks; for each feature transformation network, call the feature transformation network, and based on the corresponding information importance, perform feature transformation on each media information feature in the information feature set to obtain a candidate information feature set corresponding to the feature transformation network; perform feature fusion on each candidate information feature set to obtain a target information feature set.

[0156] In some embodiments, the feature fusion mentioned above may refer to a process of splicing at least two different features to obtain a fused feature.

[0157] As an example, see Figure 5 , obtain feature transformation network 3 and feature transformation network 4, feature transformation network 3 corresponds to importance prediction network 1, feature transformation network 4 corresponds to importance prediction network 2, for feature transformation network 3, call feature transformation network 3, based on the information importance corresponding to importance prediction network 1, perform feature transformation on each media information feature in the information feature set, and obtain the candidate information feature set corresponding to feature transformation network 3; for feature transformation network 4, call feature transformation network 4, based on the information importance corresponding to importance prediction network 2, perform feature transformation on each media information feature in the information feature set, and obtain the candidate information feature set corresponding to feature transformation network 4. Perform feature fusion on the candidate information feature set corresponding to feature transformation network 3 and the candidate information feature set corresponding to feature transformation network 4 to obtain the target information feature set.

[0158] In this way, feature transformation is performed separately through multiple feature transformation networks corresponding one by one to the importance prediction network, and the transformed candidate information feature sets are fused to obtain the target information feature set, so that the weight of the target information feature set can be determined in combination with the prediction accuracy of each importance prediction network, so that the target information feature set finally determined is more accurate.

[0159] In other embodiments, the above step 103 can be implemented in the following manner: obtain a feature transformation network, call the feature transformation network, perform feature transformation on each media information feature in the information feature set based on the information importance, and obtain a target information feature set corresponding to the feature transformation network.

[0160] As an example, see Figure 6 , obtain a feature transformation network 3, and call the feature transformation network 3, based on the information importance corresponding to the importance prediction network 1, perform feature transformation on each media information feature in the information feature set, and obtain a target information feature set corresponding to the feature transformation network.

[0161] In step 104, the importance of each historical behavior feature in the behavior feature set is evaluated to obtain the behavior importance corresponding to each historical behavior feature.

[0162] In some embodiments, the historical behavior feature includes a plurality of sub-behavior features, each of which includes a plurality of feature elements, and the behavior importance includes the target importance of each sub-behavior feature in the corresponding historical behavior feature.

[0163] In some embodiments, the behavior importance corresponding to the historical behavior feature is used to indicate the relative importance of the historical behavior feature, and the behavior importance of the historical behavior feature is associated with the weight of the historical behavior feature in the behavior feature set.

[0164] As an example, the expression of historical behavior characteristics can be:

[0165]

[0166] Among them, N is used to indicate historical behavior characteristics, N1, N2...N n Used to indicate multiple sub-behavior features included in the historical behavior feature, N 11 、N 12 、N 13 Multiple feature elements used to indicate sub-behavior feature N1, N 21 、N 22 、N 23 Multiple feature elements used to indicate sub-behavior feature N2, N n1 、N n2 、N n3 Used to indicate sub-behavior characteristics N n Multiple characteristic elements.

[0167] Continuing from the above example, the behavior importance of historical behavior feature N includes sub-behavior features N1, N2, ... N n importance of the target.

[0168] In some embodiments, see Fig.10 , Fig.10 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3 The step 104 shown can be performed for each historical behavior feature separately. Fig.10 Steps 1041 to 1042 are shown to be implemented.

[0169] In step 1041, for each sub-behavior feature in the historical behavior feature, the value of each characteristic element in the sub-behavior feature is weighted averaged to obtain a reference value of the sub-behavior feature.

[0170] Continuing with the above example, for the sub-behavior feature N1 in the historical behavior feature, the feature element N 11 、N 12 、N 13 The values ​​of are weighted averaged to obtain the reference value of sub-behavior feature N1.

[0171] Continuing with the above example, the expression for the reference value of sub-behavior feature N1 can be:

[0172]

[0173] Among them, N′1 is used to indicate the reference value of sub-behavior feature N1, N′ 11 Used to indicate the characteristic element N 11 The value of N′ 12 Used to indicate the characteristic element N 12 The value of N′ 13 Used to indicate the characteristic element N 13 The value of .

[0174] Continuing with the above example, for the sub-behavior feature N2 in the historical behavior feature, the feature element N 21 、N 22 、N 23 The values ​​of are weighted averaged to obtain the reference value of sub-behavior feature N2.

[0175] Continuing with the above example, the expression for the reference value of sub-behavior feature N2 can be:

[0176]

[0177] Among them, N′2 is used to indicate the reference value of the sub-behavior feature N2, N′ 21 Used to indicate the characteristic element N 21 The value of N′ 22 Used to indicate the characteristic element N 22 The value of N′ 23 Used to indicate the characteristic element N 23 The value of .

[0178] In step 1042, the target importance of each sub-behavior feature in the historical behavior feature is determined in combination with the reference value of each sub-behavior feature in the historical behavior feature.

[0179] In some embodiments, the above step 1022 may be implemented in the following manner: the reference value of each sub-behavior feature in the historical behavior feature is determined as the target importance of each sub-behavior feature in the historical behavior feature.

[0180] In other embodiments, the above-mentioned step 1042 can be implemented in the following manner: constructing a second reference feature of the historical behavior feature based on the reference value of each sub-behavior feature in the historical behavior feature; reducing the dimension of the second reference feature to obtain a reduced dimension feature of the second reference feature, and stretching the reduced dimension feature of the second reference feature to obtain a target feature of the second reference feature; and determining the value of each feature element in the second reference feature as the target importance of the corresponding sub-behavior feature.

[0181] In some embodiments, the target feature of the second reference feature has the same dimension as the second reference feature, and feature elements in the target feature of the second reference feature correspond one-to-one to feature elements in the second reference feature.

[0182] In some embodiments, the feature elements in the second reference feature correspond one-to-one to the sub-behavior features in the historical behavior feature, and the value of the feature element in the second reference feature is equal to the reference value of the corresponding sub-behavior feature in the historical behavior feature.

[0183] As an example, the expression of the second reference feature may be:

[0184] E={N′1,N′2 . . . N′ n} (14)

[0185] Wherein, E is used to indicate the second reference feature, N′2 is used to indicate the reference value of the sub-information feature N2, N′1 is used to indicate the reference value of the sub-information feature N1, and N′ n Used to indicate sub-information feature N n reference value.

[0186] Continuing with the above example, the feature element N′1 in the second reference feature corresponds one-to-one to the sub-information feature N1 in the historical behavior feature, and the value of the feature element N′1 in the second reference feature is equal to the reference value N′1 of the corresponding sub-information feature in the historical behavior feature.

[0187] In some embodiments, the above-mentioned dimensionality reduction of the second reference feature to obtain the reduced dimensionality feature of the second reference feature can be achieved in the following way: obtaining a compressed feature; multiplying the compressed feature and the feature product of the second reference feature to determine a candidate reduced dimensionality feature; performing a nonlinear transformation on the candidate reduced dimensionality feature to obtain the reduced dimensionality feature of the second reference feature.

[0188] In some embodiments, the compression feature is used to compress the feature dimension of the first reference feature.

[0189] In some embodiments, the above-mentioned stretching processing of the reduced-dimensionality feature of the second reference feature to obtain the target feature of the second reference feature can be achieved in the following manner: obtaining the stretched feature; determining the product of the reduced-dimensionality feature and the stretched feature of the second reference feature as the candidate target feature; normalizing the values ​​of each feature element in the candidate target feature to obtain the normalized value corresponding to each feature element; replacing the value of each feature element in the candidate target feature with the corresponding normalized value to obtain the target feature of the second reference feature.

[0190] In some embodiments, the stretching feature is used to stretch the feature dimension of the dimensionality reduction feature.

[0191] In some embodiments, the above step 104 can be implemented in the following manner: obtain multiple importance prediction networks, different importance prediction networks correspond to different network parameters; for each importance prediction network, call the importance prediction network, perform importance evaluation on each historical behavior feature in the behavior feature set, and obtain the behavior importance corresponding to each historical behavior feature.

[0192] In other embodiments, the above step 104 can also be implemented in the following way: obtain an importance prediction network, call the importance prediction network, evaluate the importance of each historical behavior feature in the behavior feature set, and obtain the behavior importance corresponding to each historical behavior feature.

[0193] In this way, by determining the target importance of each sub-behavior feature in the historical behavior feature, the reference value of the sub-behavior features at each historical moment in the historical behavior feature can be accurately measured. By making the target importance of high efficiency, media information is recommended for the recommended object, which effectively improves the accuracy of media information recommendation.

[0194] In step 105, based on the behavior importance, feature transformation is performed on each historical behavior feature in the behavior feature set to obtain a target behavior feature set.

[0195] In some embodiments, the above historical behavior feature includes a plurality of sub-behavior features, each sub-behavior feature includes a plurality of feature elements, and the behavior importance includes the target importance of each sub-behavior feature in the corresponding historical behavior feature.

[0196] In some embodiments, the behavior importance corresponding to the historical behavior feature is used to indicate the relative importance of the historical behavior feature, and the behavior importance of the historical behavior feature is associated with the weight of the historical behavior feature in the behavior feature set.

[0197] In some embodiments, see Fig.11 , Fig.11 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3The illustrated step 105 may be performed by Fig.11 Steps 1051 to 1053 are implemented as shown.

[0198] In step 1051, for each historical behavior feature in the behavior feature set, each sub-behavior feature in the historical behavior feature is multiplied by the corresponding target importance to obtain an updated sub-behavior feature corresponding to each sub-behavior feature.

[0199] As an example, the expression for updating the sub-behavior feature can be:

[0200] N Gi =Q i N i (15)

[0201] Among them, N Gi Used to indicate the update sub-information characteristics, Q i It is used to indicate the target importance corresponding to the sub-behavior feature, N i Used to indicate sub-behavior characteristics.

[0202] In step 1052, each sub-behavior feature of each historical behavior feature in the behavior feature set is replaced with a corresponding updated sub-behavior feature to obtain a candidate behavior feature set.

[0203] As an example, the expression of the candidate behavior feature set may be:

[0204] N G = {N G1 , N G2 …N Gi} (16)

[0205] Among them, N G It is used to indicate the candidate behavior feature set, N G1 To N Gi Used to indicate the behavior characteristics of each update sub-item.

[0206] In step 1053, each updated sub-behavior feature in the candidate behavior feature set is linearly transformed to obtain a target behavior feature set.

[0207] In some embodiments, the above step 1053 can be implemented in the following manner: obtain a linear transformation network, call the linear transformation network, perform linear transformation on each updated sub-behavior feature in the candidate behavior feature set, and obtain a target behavior feature set.

[0208] In some embodiments, the above-mentioned linear transformation network can be various types of linear transformation networks such as a multi-head self-attention network and a self-attention network. The specific form of the linear transformation network does not constitute a limitation on the embodiments of the present application.

[0209] In this way, by performing feature transformation on each sub-behavior feature in the behavior feature set based on the behavior importance, a target behavior feature set is obtained, so that the target behavior feature set can accurately reflect the behavior importance between different sub-behavior features, and the weights between the sub-behavior features are reasonably allocated according to the behavior importance. By using a high-efficiency target behavior feature set, media information is recommended for the recommended object, thereby effectively improving the accuracy of media information recommendation.

[0210] In step 106, media information is recommended to the recommended object in combination with the target behavior feature set and the target information feature set.

[0211] In some embodiments, the target behavior feature set includes multiple target sub-behavior features, and the target information feature set includes multiple target sub-information features.

[0212] In some embodiments, see Fig.12 , Fig.12 It is a flowchart of a method for recommending media information provided in an embodiment of the present application. Figure 3 Step 106 shown may be performed by Fig.12 Steps 1061 to 1063 are implemented as shown.

[0213] In step 1061, feature fusion is performed on multiple target sub-behavior features and multiple target sub-information features to obtain target fusion features.

[0214] In some embodiments, the feature fusion refers to a process of combining two different features to obtain a fused feature.

[0215] As an example, the expression of the above target fusion feature can be:

[0216] D={D1,D2} (17)

[0217] Among them, D is used to indicate the target fusion feature, D1 is used to indicate multiple target sub-behavior features, and D2 is used to indicate multiple target sub-information features.

[0218] In step 1062, the recommendation network is called to predict the media information to be recommended of the object to be recommended based on the target fusion feature, so as to obtain the media information to be recommended of the object to be recommended.

[0219] In some embodiments, the recommendation network is used to recommend media information to a recommendation object, and the recommendation performance of the recommendation network is positively correlated with the accuracy of the input of the recommendation network.

[0220] In some embodiments, the above-mentioned recommendation network can be a traditional recommendation model or a recommendation system based on a deep learning model. The specific form of the above-mentioned recommendation network does not constitute a limitation on the embodiments of the present application.

[0221] In step 1063, the media information to be recommended is sent to the object to be recommended.

[0222] In some embodiments, the above step 1063 may be implemented in the following manner: sending the media information to be recommended to the terminal where the object to be recommended is located.

[0223] As an example, when the terminal where the object to be recommended is located is a terminal running an instant messaging APP, the media information to be recommended may be a video to be recommended, and the video to be recommended may be sent to the terminal running the instant messaging APP.

[0224] In this way, by evaluating the importance of the features of the media information interacted by the recommended object at multiple different historical moments and the features corresponding to the interactive behaviors of the recommended object at each historical moment, we can obtain the information importance corresponding to each media information feature and the behavior importance corresponding to each historical behavior feature. Based on the information importance and the behavior importance, we can transform the media information features and the historical behavior features, respectively, to obtain a target information feature set and a target behavior feature set. Combined with the target information feature set and the target behavior feature set, media information recommendation is performed for the recommended object. In this way, in the process of recommending media information to the recommended object, the interactive behavior of the recommended object at each historical moment and the interacted media information are comprehensively considered, so as to comprehensively characterize the dynamic characteristics of the recommended object, and by determining the behavioral importance and information importance of the interactive behavior of the recommended object at each historical moment and the interacted media information, the interactive behavior of the recommended object at each historical moment and the interacted media information are further transformed according to the behavioral importance and the information importance to obtain the target behavior feature set and the target information feature set, so that the target behavior feature set and the target information feature set determined based on the behavioral importance and the information importance can accurately reflect the preference of the recommended object for different media information at different historical moments, and media information recommendation is performed on the recommended object through the highly credible target behavior feature set and the target information feature set, thereby effectively improving the accuracy of media information recommendation.

[0225] Next, an exemplary application of the embodiment of the present application in an actual application scenario of media information recommendation will be described.

[0226] Sequential Recommendation (SR) is a method that models the evolution of user interests based on their historical behaviors on items and recommends the next item to the user.

[0227] At present, the modeling of sequence recommendation mainly adopts the self-attention mechanism (Self-Attention). On this basis, further considering some side information (Side Information) in addition to the Item ID attribute during sequence modeling can significantly improve the effect. These side information usually cover a variety of Item attribute features: category, author, exposure, interaction, etc., and often also include behavioral information such as event type and occurrence time of the user object's behavior. Item ID and SideInfo are collectively referred to as attributes of behavior below, among which the features belonging to the Item side are collectively referred to as item attributes. The media information recommendation method provided in the embodiment of the present application can evaluate the importance of different attributes according to the needs of different modeling stages, thereby improving the overall recommendation effect.

[0228] In some embodiments, see Fig.13 , Fig.13 It is a schematic diagram of the interface of the media information recommendation method provided by the embodiment of the present application. The embodiment of the present application is suitable for recommendation business scenarios, and can be applied to the recall / rough sorting / fine sorting stages to build sequence recommendations based on user temporal behaviors. The embodiment of the present application is based on the importance prediction network (SENet) and the feature transformation network (SA). In the multi-head attention processing process of the self-attention mechanism of sequence modeling, independent SENets are introduced to perform independent feature importance evaluation on multiple attribute feature representations, and then in the process of extracting sequence correlations in each attention head, all attribute features can be considered globally, and attribute features can be screened in a targeted manner. By introducing the importance evaluation of behavioral attribute features, the limitations brought about by the indiscriminate unified fusion and separate processing of all attribute features are avoided. Different items and different attention heads can focus on different attribute features, thereby improving the effect of sequence modeling.

[0229] In the SA multi-head attention processing process, multiple linear transformations (Head) are often performed based on a unified behavior representation sequence. Attention calculation is performed on the Query, Key, and Value obtained from each linear transformation, and then the multi-head aggregation results are spliced ​​to obtain the updated representation of the current position and input into the next layer.

[0230] In some embodiments, see Fig.14 , Fig.14It is a schematic diagram of the principle of the media information recommendation method provided by the embodiment of the present application. Call the importance prediction network, and perform importance prediction based on a behavior feature set (S1, S2, S3) including multiple historical behavior features and an information feature set including multiple media information features to obtain the behavior importance corresponding to the historical behavior features and the information importance corresponding to the media information features; based on the behavior importance corresponding to the historical behavior features, call the score prediction layer and the score evaluation layer with different parameters respectively, perform feature transformation, obtain feature 1 and feature 2 respectively, and perform feature splicing on feature 1 and feature 2 to obtain spliced ​​features; based on the information importance corresponding to the media information features, perform feature transformation on the information feature set to obtain a transformed feature set; combine the transformed feature set and the spliced ​​features to recommend media information.

[0231] First, let's introduce the calculation steps of the importance prediction network structure to evaluate the importance of input features (such as Fig.14 Importance prediction network part), mainly divided into four steps: Squeeze: Assume that the embodiment of the present application has K attribute features, for a certain attribute feature vector Fk, its dimension is dk, the embodiment of the present application first needs to obtain a statistical summary value zk of this feature in this step. In fact, the embodiment of the present application obtains this value through some pooling methods, such as directly calculating the mean of the feature vector Fk, that is:

[0232]

[0233] Where fik represents the i-th element of the feature vector Fk. Thus, for K features, the embodiment of the present application obtains a K-dimensional statistical row vector Z, where zk represents the k-th element. Assuming the compression ratio is r, after obtaining the statistical vector, the embodiment of the present application performs the following compression and dimensionality reduction operations:

[0234] R sq =f sq (Z) = a1(W sq ·Z T ) (19)

[0235] Among them, R sq is the compressed intermediate result, with a dimension of K / r; the parameter matrix Wsq has a dimension of a1 is a nonlinear activation function.

[0236] In some embodiments, Excitation: Excitation is a dimensionality increase operation performed on the intermediate result of compression and dimensionality reduction, that is:

[0237] R ex =f ex (R sq )=a2(W ex·R sq ) (20)

[0238] Among them, R ex is the stretched K-dimensional weight vector, which is used to scale K features. The dimension of the parameter matrix Wex is a2 is the sigmoid function Directly obtain the feature importance weight between 0 and 1.

[0239] In some embodiments, reweighting: K features are multiplied by a weight vector R ex The corresponding weights for the kth feature are:

[0240] SE(F k )=[R ex ] k ·F k (twenty one)

[0241] Among them, [R ex ] k Represents the kth element of the weight vector.

[0242] In some embodiments, see Fig.14 , then describe from bottom to top Fig.14 The model operation in , for a user object's behavior sequence, is arranged from the farthest to the closest to the current time, and we get For the behavior of a user object with time step t>1, the embodiment of the present application can predict the user's behavior at time step t based on the behavior history sequence of the first t-1 time steps in the sequence. Assuming that there are K attribute features for each behavior, for the behavior at time step t, each attribute is respectively obtained through the Embedding layer to obtain the corresponding vector representation. For the kth attribute feature, its vector representation is The dimension is d k , then the vector of the behavior at time step t can be represented by the concatenation of the vectors of all attributes, that is:

[0243]

[0244] Among them, [...] represents vector concatenation, and the dimension of the input vector at time step t is In the Multi-Head Attention module, assuming there are H heads in total, for one of the heads h, its input first uses a dedicated SENet to assign a feature importance weight between 0 and 1 to each attribute feature, that is:

[0245]

[0246] Then perform linear transformation to obtain Query, Key and Value, namely:

[0247]

[0248] in, and The dimension is d X ×d QK The parameter matrix of The dimension is d X ×d V The parameter matrix of V =d X / H,d QK Generally, it can be taken as V Similarly, the calculation of its self-attention score (i.e., the relevance score of the current time step t to all T behavior sequences) can be expressed as:

[0249]

[0250] Among them, the attention score result of the behavior after time step t is directly set to negative infinity, so that its weight is zero when taking softmax, and the attention weight of T dimension can be expressed as:

[0251]

[0252] Then for time step t, the output from the perspective of attention head h can be expressed as

[0253]

[0254] Then all attention heads are concatenated and processed by FFN (Feed-Forward Network), namely:

[0255]

[0256] in Represents the concatenated vector output by H heads, W1 is of dimension d X ×d FFN The parameter matrix of b1 is of dimension d FFN The parameter vector of W2 is of dimension d. FFN ×d X The parameter matrix of b2 is of dimension d X The parameter vector of .

[0257] Usually, in order to improve training stability, residual links are added to SA units with Layer>1:

[0258] X′ t =X′ t+Dropout(LN([X′ t ] l-1 )) (29)

[0259] Among them, [X′ t ] l-1 Represents the output of the previous SA layer. Dropout means randomly setting the parameters to zero according to the set ratio. In addition to adding residual links, LayerNorm is often used to keep the distribution of the input of the current SA layer (the output of the previous layer) constant, that is:

[0260]

[0261] Among them, α l-1 and β l-1 are two trainable parameters, ε is a small smoothing term, and is the output vector [X′ t ] l-1 The mean and variance calculated from all elements in , that is:

[0262]

[0263] In the actual training process, SA may be stacked with multiple layers. In the above description, SENet only processes the bottommost original sequence input layer (Layer=1), and residual links, Dropout, and LN are only effective for SA with Layer>1.

[0264] Finally, the final loss function is calculated. After all item IDs and other attribute features on the item side are converted into vectors, the set of all item attributes including the item ID is recorded as All items are processed by an independent SENet to obtain the representation matrix of all items. For item j, its representation vector can be recorded as:

[0265]

[0266] Its dimension is For time step t, the final output vector of SA is [X′ t ] L , where L represents the last layer of the stacked SA, then the estimated probability distribution of all items at the current time step can be calculated by Softmax as follows:

[0267]

[0268] Among them, W Table The dimension d X ×dI The parameter matrix of The dimension is N×d I The item representation matrix, where N represents the number of all possible items, and the next item in the sequence is The row number in is denoted by n t+1 , then the loss function of the current time step is:

[0269]

[0270] For the sequence The loss of the sequence can be obtained by summing the losses of the previous T-1 time steps, that is,

[0271]

[0272] In this way, the embodiment of the present application introduces a mechanism for evaluating the importance of various attribute features in the representation of different Items. In the multi-head self-attention mechanism, through independent attribute importance evaluation, the self-attention head can perform targeted and differentiated screening of the attributes of the behavior when evaluating the correlation between sequences from different aspects. It can focus only on a part of the relevant attribute dimensions, dynamically decouple the ItemID representation and the attribute feature representation during Query extraction, and at the same time take into account the global information of the behavior attributes, avoiding the limitation of the attention mechanism Query and Key only using specific attributes.

[0273] In this way, by evaluating the importance of the features of the media information interacted by the recommended object at multiple different historical moments and the features corresponding to the interactive behaviors of the recommended object at each historical moment, we can obtain the information importance corresponding to each media information feature and the behavior importance corresponding to each historical behavior feature. Based on the information importance and the behavior importance, we can transform the media information features and the historical behavior features, respectively, to obtain a target information feature set and a target behavior feature set. Combined with the target information feature set and the target behavior feature set, media information recommendation is performed for the recommended object. In this way, in the process of recommending media information to the recommended object, the interactive behavior of the recommended object at each historical moment and the interacted media information are comprehensively considered, so as to comprehensively characterize the dynamic characteristics of the recommended object, and by determining the behavioral importance and information importance of the interactive behavior of the recommended object at each historical moment and the interacted media information, the interactive behavior of the recommended object at each historical moment and the interacted media information are further transformed according to the behavioral importance and the information importance to obtain the target behavior feature set and the target information feature set, so that the target behavior feature set and the target information feature set determined based on the behavioral importance and the information importance can accurately reflect the preference of the recommended object for different media information at different historical moments, and media information recommendation is performed on the recommended object through the highly credible target behavior feature set and the target information feature set, thereby effectively improving the accuracy of media information recommendation.

[0274] It is understandable that in the embodiments of the present application, when data related to media information is involved, when the embodiments of the present application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0275] The following is a description of an exemplary structure of the media information recommendation device 455 provided in the embodiment of the present application implemented as a software module. In some embodiments, for example Figure 3As shown, the software modules in the media information recommendation device 455 stored in the memory 450 may include: an acquisition module, used to acquire an information feature set including multiple media information features, and a behavior feature set including multiple historical behavior features; wherein the multiple media information features are features of the media information interacted by the object to be recommended at multiple different historical moments, and the multiple historical behavior features are features corresponding to the interactive behaviors of the object to be recommended at each of the historical moments; a media evaluation module, used to evaluate the importance of each of the media information features in the information feature set, obtain the information importance corresponding to each of the media information features, and based on the information importance, perform feature transformation on each of the media information features in the information feature set to obtain a target information feature set; a behavior evaluation module, used to evaluate the importance of each of the historical behavior features in the behavior feature set, obtain the behavior importance corresponding to each of the historical behavior features, and based on the behavior importance, perform feature transformation on each of the historical behavior features in the behavior feature set to obtain a target behavior feature set; and a recommendation module, used to recommend media information to the object to be recommended in combination with the target behavior feature set and the target information feature set.

[0276] In some embodiments, the media information feature includes multiple sub-information features, each sub-information feature includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature; the above-mentioned media evaluation module is also used to perform the following processing for each media information feature: for each sub-information feature in the media information feature, the values ​​of each feature element in the sub-information feature are weighted averaged to obtain a reference value of the sub-information feature; combined with the reference values ​​of each sub-information feature in the media information feature, the target importance of each sub-information feature in the media information feature is determined.

[0277] In some embodiments, the above-mentioned media evaluation module is also used to construct a first reference feature of the media information feature based on the reference values ​​of each sub-information feature in the media information feature; wherein the feature elements in the first reference feature correspond one-to-one to the sub-information features in the media information feature, and the values ​​of the feature elements in the first reference feature are equal to the reference values ​​of the corresponding sub-information features in the media information feature; the first reference feature is dimensionally reduced to obtain a reduced-dimensional feature, and the reduced-dimensional feature is stretched to obtain a target feature; wherein the target feature has the same dimension as the first reference feature, and the feature elements in the target feature correspond one-to-one to the feature elements in the first reference feature; the values ​​of each feature element in the target feature are respectively determined as the target importance of the corresponding sub-information feature.

[0278] In some embodiments, the above-mentioned media evaluation module is also used to obtain a compression feature, and the compression feature is used to compress the feature dimension of the first reference feature; the feature product of the compression feature and the first reference feature is determined as a candidate dimensionality reduction feature; and the candidate dimensionality reduction feature is nonlinearly transformed to obtain a dimensionality reduction feature.

[0279] In some embodiments, the above-mentioned media evaluation module is also used to obtain stretching features, and the stretching features are used to stretch the feature dimensions of the reduced dimension features; the product of the reduced dimension features and the stretching features is determined as the candidate target feature; the values ​​of each feature element in the candidate target feature are normalized respectively to obtain the normalized values ​​corresponding to each feature element; the values ​​of each feature element in the candidate target feature are replaced with the corresponding normalized values ​​to obtain the target feature.

[0280] In some embodiments, the media information feature includes multiple sub-information features, the sub-information feature includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature; the above-mentioned media evaluation module is also used to multiply each sub-information feature in the media information feature by the corresponding target importance for each media information feature in the information feature set, so as to obtain the updated sub-information features corresponding to each sub-information feature; replace each sub-information feature of each media information feature in the information feature set with the corresponding updated sub-information feature, so as to obtain a candidate information feature set; and perform linear transformation on each updated sub-information feature in the candidate information feature set to obtain a target information feature set.

[0281] In some embodiments, the media information feature includes multiple sub-information features, the sub-information feature includes multiple feature elements, and the information importance includes the target importance of each sub-information feature in the corresponding media information feature; the above-mentioned media evaluation module is also used to multiply each sub-information feature in the media information feature by the corresponding target importance for each media information feature in the information feature set, so as to obtain the updated sub-information features corresponding to each sub-information feature; and replace each sub-information feature of each media information feature in the information feature set with the corresponding updated sub-information feature to obtain the target information feature set.

[0282] In some embodiments, the above-mentioned media evaluation module is also used to obtain multiple importance prediction networks, and different importance prediction networks correspond to different network parameters; for each importance prediction network, the importance prediction network is called to perform importance evaluation on each media information feature in the information feature set to obtain the information importance corresponding to each media information feature; the above-mentioned media evaluation module is also used to obtain multiple feature transformation networks, and the feature transformation network corresponds to the importance prediction network one by one; for each feature transformation network, the feature transformation network is called to perform feature transformation on each media information feature in the information feature set based on the corresponding information importance to obtain a candidate information feature set corresponding to the feature transformation network; and feature fusion is performed on each candidate information feature set to obtain a target information feature set.

[0283] In some embodiments, the above-mentioned media evaluation module is also used to obtain an importance prediction network, and call the importance prediction network to perform importance evaluation on each media information feature in the information feature set to obtain the information importance corresponding to each media information feature; the above-mentioned media evaluation module is also used to obtain a feature transformation network, and call the feature transformation network to perform feature transformation on each media information feature in the information feature set based on the information importance to obtain a target information feature set corresponding to the feature transformation network.

[0284] In some embodiments, the above-mentioned historical behavior characteristics include multiple sub-behavior characteristics, the sub-behavior characteristics include multiple characteristic elements, and the behavior importance includes the target importance of each sub-behavior characteristic in the corresponding historical behavior characteristics; the above-mentioned behavior evaluation module is also used to perform the following processing for each historical behavior characteristic: for each sub-behavior characteristic in the historical behavior characteristic, the values ​​of each characteristic element in the sub-behavior characteristic are weighted averaged to obtain a reference value of the sub-behavior characteristic; combined with the reference values ​​of each sub-behavior characteristic in the historical behavior characteristic, the target importance of each sub-behavior characteristic in the historical behavior characteristic is determined.

[0285] In some embodiments, the above-mentioned behavior evaluation module is also used to construct a second reference feature of the historical behavior feature based on the reference values ​​of each sub-behavior feature in the historical behavior feature; wherein the feature elements in the second reference feature correspond one-to-one to the sub-behavior features in the historical behavior feature, and the values ​​of the feature elements in the second reference feature are equal to the reference values ​​of the corresponding sub-behavior features in the historical behavior feature; the second reference feature is dimensionally reduced to obtain the reduced-dimensional feature of the second reference feature, and the reduced-dimensional feature of the second reference feature is stretched to obtain the target feature of the second reference feature; wherein the target feature of the second reference feature has the same dimension as the second reference feature, and the feature elements in the target feature of the second reference feature correspond one-to-one to the feature elements in the second reference feature; the values ​​of each feature element in the second reference feature are respectively determined as the target importance of the corresponding sub-behavior feature.

[0286] In some embodiments, the historical behavior feature includes multiple sub-behavior features, the sub-behavior feature includes multiple feature elements, and the behavior importance includes the target importance of each sub-behavior feature in the corresponding historical behavior feature; the above-mentioned behavior evaluation module is also used to multiply each sub-behavior feature in the historical behavior feature by the corresponding target importance for each historical behavior feature in the behavior feature set, to obtain the updated sub-behavior feature corresponding to each sub-behavior feature; replace each sub-behavior feature of each historical behavior feature in the behavior feature set with the corresponding updated sub-behavior feature, to obtain a candidate behavior feature set; perform linear transformation on each updated sub-behavior feature in the candidate behavior feature set, to obtain a target behavior feature set.

[0287] In some embodiments, the target behavior feature set includes multiple target sub-behavior features, the target information feature set includes multiple target sub-information features, and the above-mentioned recommendation module is also used to perform feature fusion on multiple target sub-behavior features and multiple target sub-information features to obtain target fusion features; call the recommendation network to predict the media information to be recommended for the recommended object based on the target fusion features to obtain the media information to be recommended for the recommended object; and send the media information to be recommended to the recommended object.

[0288] The embodiment of the present application provides a computer program product, which includes a computer program or a computer executable instruction, and the computer program or the computer executable instruction is stored in a computer-readable storage medium. The processor of the electronic device reads the computer executable instruction from the computer-readable storage medium, and the processor executes the computer executable instruction, so that the electronic device executes the media information recommendation method described in the embodiment of the present application.

[0289] The present application embodiment provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, the processor will be caused to execute the media information recommendation method provided by the present application embodiment, for example, Figure 3 A recommended method of displaying media information.

[0290] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEP ROM, flash memory, magnetic surface storage, optical disk, or CD-ROM; or it may be various electronic devices including one or any combination of the above memories.

[0291] In some embodiments, computer executable instructions may be in the form of a program, software, software module, script or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine or other unit suitable for use in a computing environment.

[0292] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file storing other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0293] As an example, computer executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed at multiple sites and interconnected by a communication network.

[0294] In summary, the embodiments of the present application have the following beneficial effects:

[0295] (1) By evaluating the importance of the features of the media information interacted by the recommended object at multiple different historical moments and the features corresponding to the interactive behaviors of the recommended object at each historical moment, the information importance corresponding to each media information feature and the behavior importance corresponding to each historical behavior feature are obtained. Based on the information importance and the behavior importance, the media information features and the historical behavior features are transformed to obtain a target information feature set and a target behavior feature set. The target information feature set and the target behavior feature set are combined to recommend media information to the recommended object. In this way, in the process of recommending media information to the recommended object, the interactive behavior of the recommended object at each historical moment and the interacted media information are comprehensively considered, so as to comprehensively characterize the dynamic characteristics of the recommended object, and by determining the behavioral importance and information importance of the interactive behavior of the recommended object at each historical moment and the interacted media information, the interactive behavior of the recommended object at each historical moment and the interacted media information are further transformed according to the behavioral importance and the information importance to obtain the target behavior feature set and the target information feature set, so that the target behavior feature set and the target information feature set determined based on the behavioral importance and the information importance can accurately reflect the preference of the recommended object for different media information at different historical moments, and media information recommendation is performed on the recommended object through the highly credible target behavior feature set and the target information feature set, thereby effectively improving the accuracy of media information recommendation.

[0296] (2) By determining the target importance of each sub-information feature in the media information feature, the reference value of the sub-information features at each historical moment in the media information feature can be accurately measured. By making the target importance of high efficiency the media information recommendation for the recommended object, the recommendation accuracy of the media information is effectively improved.

[0297] (3) By transforming the features of each media information feature in the information feature set based on the information importance, a target information feature set is obtained, so that the target information feature set can accurately reflect the information importance between different sub-information features, and the weights between the sub-information features are reasonably allocated according to the information importance. By making the target information feature set with high efficiency, media information is recommended to the recommended object, thereby effectively improving the recommendation accuracy of media information.

[0298] (4) Feature transformation is performed respectively through multiple feature transformation networks corresponding to the importance prediction network, and the transformed candidate information feature sets are fused to obtain the target information feature set, so that the weight of the target information feature set can be determined in combination with the prediction accuracy of each importance prediction network, so that the target information feature set finally determined is more accurate.

[0299] (5) By determining the target importance of each sub-behavior feature in the historical behavior feature, the reference value of the sub-behavior feature at each historical moment in the historical behavior feature can be accurately measured. By making the target importance of high efficiency the media information recommendation for the recommended object, the recommendation accuracy of the media information is effectively improved.

[0300] (6) By performing feature transformation on each sub-behavior feature in the behavior feature set based on the behavior importance, a target behavior feature set is obtained, so that the target behavior feature set can accurately reflect the behavior importance between different sub-behavior features, and the weights between the sub-behavior features are reasonably allocated according to the behavior importance. By making the target behavior feature set with high efficiency, media information is recommended to the recommended object, thereby effectively improving the accuracy of media information recommendation.

[0301] (7) The embodiment of the present application introduces a mechanism for evaluating the importance of various attribute features in the representation of different items. In the multi-head self-attention mechanism, through independent attribute importance evaluation, the self-attention head can perform targeted and differentiated screening of the attributes of the behavior when evaluating the correlation between sequences from different aspects. It can focus on only a part of the relevant attribute dimensions and dynamically decouple the ItemID representation and attribute feature representation during query extraction. At the same time, it can take into account the global information of the behavior attributes, avoiding the limitation that the attention mechanism Query and Key only use specific attributes.

[0302] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A method for recommending media information, characterized in that: The method comprises: Acquire an information feature set including a plurality of media information features and a behavior feature set including a plurality of historical behavior features; The multiple media information features are features of media information interacted by the to-be-recommended object at multiple different historical moments, and the multiple historical behavior features are features corresponding to the interactive behaviors of the to-be-recommended object at each of the historical moments; Performing importance evaluation on each of the media information features in the information feature set to obtain information importance corresponding to each of the media information features, and performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set; Performing importance evaluation on each of the historical behavior features in the behavior feature set to obtain the behavior importance corresponding to each of the historical behavior features, and performing feature transformation on each of the historical behavior features in the behavior feature set based on the behavior importance to obtain a target behavior feature set; The target behavior feature set and the target information feature set are combined to recommend media information to the object to be recommended.

2. The method according to claim 1, characterized in that The media information feature includes a plurality of sub-information features, each of the sub-information features includes a plurality of feature elements, and the information importance includes a target importance of each of the sub-information features in the corresponding media information feature; The step of evaluating the importance of each of the media information features in the information feature set to obtain the information importance corresponding to each of the media information features includes: The following processing is performed for each of the media information features: For each of the sub-information features in the media information feature, weighted average the values ​​of each of the feature elements in the sub-information feature to obtain a reference value of the sub-information feature; The target importance of each of the sub-information features in the media information feature is determined in combination with the reference value of each of the sub-information features in the media information feature.

3. The method according to claim 2, characterized in that The determining the target importance of each of the sub-information features in the media information feature by combining the reference value of each of the sub-information features in the media information feature includes: Constructing a first reference feature of the media information feature based on the reference value of each of the sub-information features in the media information feature; The feature elements in the first reference feature correspond to the sub-information features in the media information feature one by one, and the value of the feature element in the first reference feature is equal to the reference value of the corresponding sub-information feature in the media information feature; Performing dimensionality reduction on the first reference feature to obtain a reduced dimensionality feature, and performing stretching processing on the reduced dimensionality feature to obtain a target feature; The target feature has the same dimension as the first reference feature, and feature elements in the target feature correspond one-to-one to feature elements in the first reference feature; The value of each characteristic element in the target feature is respectively determined as the target importance of the corresponding sub-information feature.

4. The method according to claim 3, characterized in that The step of reducing the dimension of the first reference feature to obtain a reduced-dimensional feature includes: Acquire a compression feature, where the compression feature is used to compress the feature dimension of the first reference feature; Multiplying the feature product of the compression feature and the first reference feature to determine a candidate dimensionality reduction feature; Perform nonlinear transformation on the candidate dimensionality reduction feature to obtain the dimensionality reduction feature.

5. The method according to claim 3, characterized in that: The stretching process of the dimension reduction feature to obtain the target feature includes: Acquire a stretching feature, wherein the stretching feature is used to stretch the feature dimension of the dimensionality reduction feature; Determine the product of the dimension reduction feature and the stretching feature as a candidate target feature; Normalizing the values ​​of each feature element in the candidate target feature to obtain a normalized value corresponding to each feature element; The value of each feature element in the candidate target feature is replaced by the corresponding normalized value to obtain the target feature.

6. The method according to claim 1, characterized in that The media information feature includes a plurality of sub-information features, each of which includes a plurality of feature elements, and the information importance includes a target importance of each of the sub-information features in the corresponding media information feature; The step of performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set includes: For each of the media information features in the information feature set, multiply each of the sub-information features in the media information feature by the corresponding target importance, to obtain updated sub-information features corresponding to each of the sub-information features; Replacing each of the sub-information features of each of the media information features in the information feature set with the corresponding updated sub-information features to obtain a candidate information feature set; Performing a linear transformation on each of the updated sub-information features in the candidate information feature set to obtain the target information feature set.

7. The method according to claim 1, characterized in that The media information feature includes a plurality of sub-information features, each of which includes a plurality of feature elements, and the information importance includes a target importance of each of the sub-information features in the corresponding media information feature; The step of performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set includes: For each of the media information features in the information feature set, multiply each of the sub-information features in the media information feature by the corresponding target importance, to obtain updated sub-information features corresponding to each of the sub-information features; Each of the sub-information features of each of the media information features in the information feature set is replaced with the corresponding updated sub-information feature to obtain the target information feature set.

8. The method according to claim 1, characterized in that: The step of evaluating the importance of each of the media information features in the information feature set to obtain the information importance corresponding to each of the media information features includes: Acquire multiple importance prediction networks, where different importance prediction networks correspond to different network parameters; For each of the importance prediction networks, calling the importance prediction network, performing importance evaluation on each of the media information features in the information feature set, and obtaining the information importance corresponding to each of the media information features; The step of performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set includes: Acquire multiple feature transformation networks, wherein the feature transformation networks correspond one to one to the importance prediction networks; For each of the feature transformation networks, calling the feature transformation network, and performing feature transformation on each of the media information features in the information feature set based on the corresponding information importance, to obtain a candidate information feature set corresponding to the feature transformation network; Feature fusion is performed on each of the candidate information feature sets to obtain the target information feature set.

9. The method according to claim 1, characterized in that: The step of evaluating the importance of each of the media information features in the information feature set to obtain the information importance corresponding to each of the media information features includes: Obtaining an importance prediction network, and calling the importance prediction network to evaluate the importance of each of the media information features in the information feature set, and obtaining the information importance corresponding to each of the media information features; The step of performing feature transformation on each of the media information features in the information feature set based on the information importance to obtain a target information feature set includes: A feature transformation network is obtained, and the feature transformation network is called to perform feature transformation on each of the media information features in the information feature set based on the information importance, so as to obtain a target information feature set corresponding to the feature transformation network.

10. The method according to claim 1, characterized in that The historical behavior feature includes a plurality of sub-behavior features, each of which includes a plurality of feature elements, and the behavior importance includes the target importance of each of the sub-behavior features in the corresponding historical behavior feature; The evaluating the importance of each of the historical behavior features in the behavior feature set to obtain the behavior importance corresponding to each of the historical behavior features includes: The following processing is performed for each of the historical behavior characteristics: For each of the sub-behavior features in the historical behavior feature, weighted average the values ​​of each of the characteristic elements in the sub-behavior feature to obtain a reference value of the sub-behavior feature; The target importance of each of the sub-behavior features in the historical behavior feature is determined in combination with the reference value of each of the sub-behavior features in the historical behavior feature.

11. The method according to claim 10, characterized in that The determining the target importance of each sub-behavior feature in the historical behavior feature by combining the reference value of each sub-behavior feature in the historical behavior feature includes: Constructing a second reference feature of the historical behavior feature based on the reference value of each of the sub-behavior features in the historical behavior feature; The characteristic elements in the second reference characteristic correspond to the sub-behavior characteristics in the historical behavior characteristic one by one, and the value of the characteristic element in the second reference characteristic is equal to the reference value of the corresponding sub-behavior characteristic in the historical behavior characteristic; Performing dimensionality reduction on the second reference feature to obtain a reduced dimensionality feature of the second reference feature, and performing stretching processing on the reduced dimensionality feature of the second reference feature to obtain a target feature of the second reference feature; The target feature of the second reference feature has the same dimension as the second reference feature, and the feature elements in the target feature of the second reference feature correspond one-to-one to the feature elements in the second reference feature; The value of each feature element in the second reference feature is determined as the target importance of the corresponding sub-behavior feature.

12. The method according to claim 1, characterized in that The historical behavior feature includes a plurality of sub-behavior features, each of which includes a plurality of feature elements, and the behavior importance includes the target importance of each of the sub-behavior features in the corresponding historical behavior feature; The step of performing feature transformation on each of the historical behavior features in the behavior feature set based on the behavior importance to obtain a target behavior feature set includes: For each of the historical behavior features in the behavior feature set, multiply each of the sub-behavior features in the historical behavior features by the corresponding target importance, to obtain updated sub-behavior features corresponding to each of the sub-behavior features; Replacing each of the sub-behavior features of each of the historical behavior features in the behavior feature set with the corresponding updated sub-behavior features to obtain a candidate behavior feature set; Performing linear transformation on each of the updated sub-behavior features in the candidate behavior feature set to obtain the target behavior feature set.

13. The method according to claim 1, characterized in that The target behavior feature set includes a plurality of target sub-behavior features, the target information feature set includes a plurality of target sub-information features, and combining the target behavior feature set and the target information feature set to recommend media information to the to-be-recommended object includes: Performing feature fusion on the multiple target sub-behavior features and the multiple target sub-information features to obtain a target fusion feature; Calling a recommendation network to predict the media information to be recommended of the object to be recommended based on the target fusion feature, and obtaining the media information to be recommended of the object to be recommended; The media information to be recommended is sent to the object to be recommended.

14. A media information recommendation device, characterized in that: The device comprises: An acquisition module, used to acquire an information feature set including a plurality of media information features and a behavior feature set including a plurality of historical behavior features; wherein the plurality of media information features are features of media information interacted with by the object to be recommended at a plurality of different historical moments, and the plurality of historical behavior features are features corresponding to the interactive behaviors of the object to be recommended at each of the historical moments; a media evaluation module, configured to evaluate the importance of each of the media information features in the information feature set, obtain the information importance corresponding to each of the media information features, and perform feature transformation on each of the media information features in the information feature set based on the information importance, to obtain a target information feature set; A behavior evaluation module is used to evaluate the importance of each of the historical behavior features in the behavior feature set to obtain the behavior importance corresponding to each of the historical behavior features, and based on the behavior importance, perform feature transformation on each of the historical behavior features in the behavior feature set to obtain a target behavior feature set; The recommendation module is used to recommend media information to the object to be recommended by combining the target behavior feature set and the target information feature set.

15. An electronic device, characterized in that: The electronic device comprises: A memory for storing computer executable instructions or computer programs; A processor is used to implement the media information recommendation method according to any one of claims 1 to 13 when executing the computer executable instructions or computer programs stored in the memory.

16. A computer-readable storage medium storing computer-executable instructions, characterized in that: When the computer executable instructions are executed by a processor, the media information recommendation method according to any one of claims 1 to 13 is implemented.

17. A computer program product comprising a computer program or computer executable instructions, characterized in that When the computer program or computer executable instructions are executed by a processor, the method for recommending media information according to any one of claims 1 to 13 is implemented.

Images