Multimedia information recommendation method, device, program product, equipment and medium

Abstract

The present invention provides a multimedia information recommendation method, apparatus, program product, device and medium, the method comprising: obtaining first object feedback data for target multimedia information, and predicting a first score corresponding to each first sub-object feedback data based on the first object feedback data; determining a fusion weight parameter corresponding to each first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended; and fusing the first score corresponding to each first sub-object feedback data based on the fusion weight parameter corresponding to each first sub-object feedback data to obtain a second score of the target multimedia information, thereby using the second score of the target multimedia information to recommend multimedia information to the object to be recommended, thereby achieving more accurate recommendation, and at the same time, the hardware overhead required by the multi-objective weight search network is smaller.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multimedia information recommendation method, device, program product, equipment and medium. Background Art

[0002] The neural network model based on machine learning can be used to recommend various videos. Artificial Intelligence (AI) is a comprehensive technology in computer science. By studying the design principles and implementation methods of various intelligent machines, the machines are given the functions of perception, reasoning and decision-making. Artificial Intelligence technology is a comprehensive discipline that covers a wide range of fields, such as natural language processing technology and machine learning / deep learning. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0003] In traditional technology, accurate video recommendation often requires collecting a large amount of user operation records. However, it is often difficult to collect enough labeled samples for new users for traditional machine learning to extract recommended information from the data, which can easily lead to model overfitting and introduce new noise, affecting the data processing effect of the model. At the same time, the hardware overhead required to achieve accurate video recommendation is relatively large. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a multimedia information recommendation method, apparatus, program product, device and medium, which can determine the fusion weight parameters corresponding to each first sub-object feedback data through a multi-objective weight search network, and fuse the first scores corresponding to each first sub-object feedback data to obtain the second score of the target multimedia information, and use the second score to recommend multimedia information to the object to be recommended.

[0005] An embodiment of the present invention provides a multimedia information recommendation method, the method comprising:

[0006] Acquire first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types;

[0007] Based on the first object feedback data, predict a first score corresponding to each first sub-object feedback data;

[0008] Based on the first object feature data of the object to be recommended, a fusion weight parameter corresponding to each of the first sub-object feedback data is determined through a multi-objective weight search network; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information;

[0009] Based on the fusion weight parameter corresponding to each of the first sub-object feedback data, the first score corresponding to each of the first sub-object feedback data is fused to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

[0010] The embodiment of the present invention further provides a multimedia information recommendation device, the device comprising:

[0011] An information transmission module, configured to obtain first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types;

[0012] An information processing module, configured to predict, based on the first object feedback data, a first score corresponding to each of the first sub-object feedback data;

[0013] The information processing module is used to determine the fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information;

[0014] The information processing module is used to fuse the first scores corresponding to each of the first sub-object feedback data based on the fusion weight parameters corresponding to each of the first sub-object feedback data to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

[0015] In the above scheme,

[0016] The information processing module is used to obtain a training sample set, wherein the training sample set includes the real object feedback data of the sample object for the sample multimedia information;

[0017] The information processing module is used to use the training sample set to update the network parameters of the initial multi-objective weighted search network through an evolutionary strategy algorithm, and obtain the multi-objective weighted search network through the updated network parameters.

[0018] In the above scheme,

[0019] The information processing module is used to calculate the network parameter update direction of the initial multi-objective weight search network according to the noise standard deviation parameter of the initial multi-objective weight search network;

[0020] The information processing module is used to configure a first random noise parameter for each sample object in the training sample set according to the noise standard deviation parameter;

[0021] The information processing module is used to determine the fitness score corresponding to each of the sample objects according to the first random noise parameter, the network parameter of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object;

[0022] The information processing module is used to share the real object feedback data of the sample object among all sample objects in the training sample set;

[0023] The information processing module is used to configure a second random noise parameter for each sample object in the training sample set;

[0024] The information processing module is used to determine the first gradient change parameters of different rounds of the initial multi-objective weight search network according to the second random noise parameter, the network parameters of the initial multi-objective weight search network and the noise standard deviation parameter;

[0025] The information processing module is used to update the network parameters of the initial multi-objective weight search network according to the network parameter update direction of the initial multi-objective weight search network and using the first gradient change parameters of different rounds of the initial multi-objective weight search network.

[0026] In the above scheme,

[0027] The information processing module is used to use the training sample set to update the network parameters of the initial multi-objective weighted search network through a meta-learning framework based on an evolutionary strategy algorithm, and obtain the multi-objective weighted search network based on the updated network parameters.

[0028] In the above scheme,

[0029] The information processing module is used to extract a first training sample set of the target task number from the training sample set; wherein each of the first training sample sets corresponds to a sample object;

[0030] Based on at least one set of task numbers, configuring a support set for the initial multi-objective weighted search network;

[0031] Calculating a network parameter update direction of the initial multi-objective weight search network according to a noise standard deviation parameter of the initial multi-objective weight search network;

[0032] According to the noise standard deviation parameter, configuring a third random noise parameter for each sample object in the support set;

[0033] According to the third random noise parameter, the network parameter of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, determine the fitness score corresponding to each of the sample objects in the support set, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object;

[0034] Sharing the real object feedback data of the sample object among all sample objects in the support set;

[0035] configuring a fourth random noise parameter for each sample object in the support set;

[0036] Determining second gradient change parameters of different rounds of the initial multi-objective weight search network according to the fourth random noise parameter, the network parameter of the initial multi-objective weight search network, and the noise standard deviation parameter;

[0037] According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the second gradient change parameters of different rounds of the initial multi-objective weight search network to obtain the first network parameters of the initial multi-objective weight search network.

[0038] In the above scheme,

[0039] The information processing module is used to configure a query set for the initial multi-objective weighted search network based on at least one task number set;

[0040] Calculating a network parameter update direction of the initial multi-objective weight search network according to a noise standard deviation parameter of the initial multi-objective weight search network;

[0041] According to the noise standard deviation parameter, configuring a fifth random noise parameter for each sample object in the query set;

[0042] According to the fifth random noise parameter, the network parameter of the initial multi-objective weighted search network and the noise standard deviation parameter, a fitness score corresponding to each of the sample objects in the query set is determined by a fitness function corresponding to the initial multi-objective weighted search network, and the fitness score corresponding to the sample object is used as the real object feedback data of the sample object;

[0043] sharing the real object feedback data of the sample object among all sample objects in the query set, and configuring a sixth random noise parameter for each sample object in the query set;

[0044] Determining third gradient change parameters of different rounds of the initial multi-objective weight search network according to the sixth random noise parameter, the network parameters of the initial multi-objective weight search network, and the noise standard deviation parameter;

[0045] According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the third gradient change parameters of different rounds of the initial multi-objective weight search network and the first network parameters of the initial multi-objective weight search network to obtain the second network parameters of the initial multi-objective weight search network.

[0046] The embodiment of the present invention further provides a computer program product, including a computer program or an instruction, wherein when the computer program or the instruction is executed by a processor, the aforementioned multimedia information recommendation method is implemented.

[0047] An embodiment of the present invention further provides an electronic device, the electronic device comprising:

[0048] A memory for storing executable instructions;

[0049] The processor is used to implement the preceding data processing method when running the executable instructions stored in the memory.

[0050] An embodiment of the present invention further provides a computer-readable storage medium storing executable instructions, wherein the executable instructions, when executed by a processor, implement a preceding multimedia information recommendation method or a preceding data processing method.

[0051] The embodiments of the present invention have the following beneficial effects:

[0052] The embodiment of the present invention obtains first object feedback data for target multimedia information, and predicts a first score corresponding to each first sub-object feedback data based on the first object feedback data; determines a fusion weight parameter corresponding to each first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended; and fuses the first score corresponding to each first sub-object feedback data based on the fusion weight parameter corresponding to each first sub-object feedback data to obtain a second score of the target multimedia information. Thus, it is possible to fuse the first score corresponding to each first sub-object feedback data to obtain the second score of the target multimedia information, and use the second score to accurately recommend multimedia information to the object to be recommended, while achieving the effect of stronger generalization and data processing capabilities of the multi-objective weight search network and lower hardware overhead. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a scenario schematic diagram of a multimedia information recommendation method provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of the structure of a multimedia information recommendation device provided by an embodiment of the present invention;

[0055] Figure 3 An optional flowchart of a multimedia information recommendation method provided in an embodiment of the present invention;

[0056] Figure 4 An optional flowchart of a multimedia information recommendation method provided in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of a multi-objective weighted search network structure in an embodiment of the present invention;

[0058] Figure 6 A schematic diagram of an optional processing process of the multimedia information recommendation method in an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of the effect of the evolution strategy algorithm in an embodiment of the present invention;

[0060] Figure 8 A schematic diagram of an optional processing process of the multimedia information recommendation method in an embodiment of the present invention;

[0061] Fig. 9A A schematic diagram of the effect of the multimedia information recommendation method according to an embodiment of the present invention;

[0062] Fig. 9B Schematic diagram of the effect of the multimedia information recommendation method in an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0064] 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.

[0065] Before further describing the embodiments of the present invention in detail, the nouns and terms involved in the embodiments of the present invention are described. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations.

[0066] 1) Neural Network (NN): Artificial Neural Network (ANN), referred to as neural network or quasi-neural network, is a mathematical model or computational model that imitates the structure and function of biological neural networks (the central nervous system of animals, especially the brain) in the field of machine learning and cognitive science, and is used to estimate or approximate functions.

[0067] 2) Model parameters: It is a quantity that uses universal variables to establish the relationship between functions and variables. In artificial neural networks, model parameters are usually real number matrices.

[0068] 3) Model training, multi-classification learning of image datasets. The model can be constructed using various types of deep learning frameworks, and a multi-classification model is formed by combining multiple layers of neural network layers such as CNN. The input of the model is a three-channel or original channel matrix formed by reading the image through tools such as openCV. The model output is a multi-classification probability, and the web page category is finally output through algorithms such as softmax. During training, the model approaches the correct trend through objective functions such as cross entropy.

[0069] 4) Multi-task Learning: In the field of machine learning, by jointly learning and evolving multiple related tasks at the same time, better model accuracy can be achieved than that of a single task. Multiple tasks help each other by sharing the representation layer. This training method is called multi-task learning, also called joint learning.

[0070] 5) Meta-learning, also known as learning to learn, refers to the process of learning how to learn. Traditional machine learning problems are to learn a mathematical model for prediction from scratch, which is far from the process of human learning and accumulating historical experience (also known as meta-knowledge) to guide new learning tasks. Meta-learning is the learning and training process of learning different machine learning tasks, and learning how to train a model faster and better.

[0071] 6) N-way K-shot, a common training setting for small sample learning in the classification field. In the training phase, N categories are extracted from the training set, with K samples in each category. A total of N*K samples constitute a meta-task as the support set of the model. Then, a batch of samples are extracted from the remaining data of the N categories as the query set of the model. Such a task is called the N-way K-shot problem.

[0072] 7) Task, the model training and testing unit of meta-learning. It consists of a support set and a query set. For example, according to the 5-way 5-shot experimental setting, 5 categories are randomly selected from the data set, and 5 samples are randomly selected from each category to form a support set. A certain number of samples (for example, 15 samples) from the same category are selected to form a query set, and finally a task is formed.

[0073] 8) In response, it is 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 of execution of the multiple operations executed.

[0074] 9) Terminals, including but not limited to: ordinary terminals and dedicated terminals, wherein the ordinary terminals maintain a long connection and / or a short connection with the sending channel, and the dedicated terminals maintain a long connection with the sending channel.

[0075] 10) Client: a carrier that implements specific functions in a terminal. For example, a mobile client (APP) is a carrier of specific functions in a mobile terminal, such as executing payment and consumption functions or video playback functions.

[0076] 11) Multi-objective fusion sorting: In the recommendation system, the user's estimated values ​​of various indicators of candidate items (such as estimated click-through rate, comment rate, etc.) are comprehensively utilized and fused into a score that can measure the user's satisfaction with the recommendation through some method (such as weighted summation, etc.). The candidate items are then sorted according to their scores and finally recommended to the corresponding users.

[0077] Among them, the embodiments of the present invention can be implemented in combination with cloud technology. Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software and network in a wide area network or a local area network to realize data calculation, storage, processing and sharing. It can also be understood as a general term for network technology, information technology, integration technology, management platform technology and application technology based on cloud computing business model applications. The background services of the technical network system require a large amount of computing and storage resources, such as video websites, picture websites and more portal websites, so cloud technology needs to be supported by cloud computing.

[0078] It should be noted that cloud computing is a computing model that distributes computing tasks on a resource pool composed of a large number of computers, so that various application systems can obtain computing power, storage space and information services as needed. The network that provides resources is called the "cloud". From the user's point of view, the resources in the "cloud" can be infinitely expanded, and can be obtained at any time, used on demand, expanded at any time, and paid for by use. As a basic capability provider of cloud computing, a cloud computing resource pool platform will be established, referred to as a cloud platform, generally referred to as Infrastructure as a Service (IaaS), and various types of virtual resources will be deployed in the resource pool for external customers to choose to use. The cloud computing resource pool mainly includes: computing devices (which can be virtualized machines, including operating systems), storage devices and network devices.

[0079] Figure 1 A schematic diagram of a multimedia information processing method provided by an embodiment of the present invention is shown in FIG. Figure 1, the terminal (including the terminal 10-1 and the terminal 10-2) is provided with a corresponding client capable of playing the embedded multimedia information, and the terminal is connected to the server 200 through the network 300, and the network 300 can be a wide area network or a local area network, or a combination of the two, and a wireless link is used to realize data transmission, wherein the multimedia information includes but is not limited to videos, pictures, GIF animations and advertising information. Among them, the types of multimedia information obtained by the terminal (including the terminal 10-1 and the terminal 10-2) from the corresponding server 200 through the network 300 can be the same or different, for example: the terminal (including the terminal 10-1 and the terminal 10-2) can obtain various types of multimedia advertisements placed by advertisers from the corresponding server 200 through the network 300, and can also obtain short videos or long videos provided by video operators from the corresponding server 200 through the network 300 for viewing, and the specific types of multimedia information are not limited by the present invention. Different multimedia information can be stored in the server 200, wherein the multimedia information used as an advertisement can be content in different dynamic formats, such as gif, mp4, mov, etc.

[0080] In the process of the terminal (terminal 10-1 and / or terminal 10-2) acquiring and displaying the corresponding service with embedded multimedia information from the server 200 through the network 300, the user can perform different operations on the multimedia information presented in the multimedia information playback window through the terminal (terminal 10-1 and / or terminal 10-2), resulting in different user behaviors. For example, when the multimedia information is a video advertisement, the user can share and / or like the exposed short video while watching the information, or collect the product provided in the video advertisement by clicking the hyperlink in the short video. When the multimedia information is a dynamic GIF advertisement, during the exposure of the advertisement through the terminal (terminal 10-1 and / or terminal 10-2), the user can forward and / or comment on the advertisement, or jump to the corresponding product purchase link page through the GIF advertisement.

[0081] As an example, when the server 200 determines which multimedia information to recommend to the user's terminal 10-1 or 10-2 for playback, it is necessary to timely adjust the multimedia information to be played, such as replacing any multimedia information in the multimedia information set to be played, so as to adapt to the viewing needs of different target users. Taking short video information as an example, the multi-objective weighted search network provided by the present invention can be applied to short video playback. In short video playback, different short video information from different data sources are usually processed, and finally the corresponding different information information and the corresponding short video recommendation process corresponding to the recommended video are presented on the user interface UI (User Interface). The accuracy and timeliness of the characteristics of different information information directly affect the user experience. The background database of video playback receives a large amount of video data from different sources every day, and the different information information obtained for recommending information to target users can also be called by other applications (for example, the recommendation results of the short video recommendation process are migrated to the long video recommendation process or the news recommendation process). Of course, the multi-objective weighted search network matching the corresponding target user can also be migrated to different video recommendation processes (for example, web video recommendation process, mini program video recommendation process or long video client video recommendation process).

[0082] As an example, the server 200 is used to deploy a corresponding multi-objective weighted search network to implement the multimedia information recommendation method provided by the present invention, or to deploy a multimedia information recommendation device to implement the multimedia information recommendation method. Specifically, the multimedia information recommendation can be implemented by the following method: obtaining first object feedback data for the target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types; based on the first object feedback data, predicting a first score corresponding to each of the first sub-object feedback data; based on the first object feature data of the object to be recommended, determining a fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weighted search network; In the method, the multi-objective weighted search network is obtained by updating the network parameters of the initial multi-objective weighted search network by evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information; based on the fusion weight parameter corresponding to each of the first sub-object feedback data, the first score corresponding to each of the first sub-object feedback data is fused to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

[0083] Furthermore, the terminal (terminal 10-1 and / or terminal 10-2) can also display and output the multimedia information to be recommended that matches the target user. Taking short video playback as an example, the multi-objective weighted search network provided by the present invention can be applied to the short video recommendation environment to process the corresponding object feature data, determine the fusion weight parameter corresponding to each sub-object feedback data, and usually process different short multimedia information from different data sources in short video playback, and finally present the corresponding different multimedia information on the user interface UI (User Interface). The accuracy and timeliness of multimedia information recommendation directly affect the user experience. The background database of video playback receives a large amount of multimedia information data from different sources every day, and the different multimedia information obtained for multimedia information recommendation to users can also be called by other applications (for example, the recommendation results of the short video recommendation process are migrated to the recommendation process or news recommendation process in the instant messaging client). Of course, the multi-objective weighted search network that matches the corresponding target user can also be migrated to different video recommendation processes (for example, web video recommendation process, mini-program video recommendation process, or video recommendation process of the client in the instant messaging client).

[0084] Among them, the multimedia information recommendation method provided in the embodiment of the present invention is implemented based on artificial intelligence. 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. In other words, artificial intelligence is a comprehensive technology in computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.

[0085] Artificial intelligence technology is a comprehensive discipline 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. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0086] In the embodiments of the present invention, the artificial intelligence software technologies mainly involved include the above-mentioned speech processing technology and machine learning. For example, it may involve speech recognition technology (Automatic Speech Recognition, ASR) in speech technology (Speech Technology), including speech signal preprocessing (Speech signal preprocessing), speech signal frequency domain analysis (Speech signal frequency analyzing), speech signal feature extraction (Speech signal feature extraction), speech signal feature matching / recognition (Speech signal feature matching / recognition), speech training (Speech training), etc.

[0087] For example, it may involve machine learning (ML), which is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. 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. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning usually includes technologies such as deep learning. Deep learning includes artificial neural networks, such as convolutional neural networks (CNN), recurrent neural networks (RNN), deep neural networks (DNN), etc.

[0088] It can be understood that the multimedia information recommendation method and voice processing provided by the present invention can be applied to intelligent devices (Intelligent device), and the intelligent device can be any device with information display function, for example, it can be a smart terminal, smart home device (such as smart speakers, smart washing machines, etc.), smart wearable devices (such as smart watches), in-vehicle intelligent central control system (displaying multimedia information to users through small programs that perform different tasks) or AI smart medical equipment (displaying treatment cases by displaying multimedia information), etc.

[0089] The structure of the multimedia information recommendation device of the embodiment of the present invention is described in detail below. The multimedia information recommendation device can be implemented in various forms, such as a dedicated terminal with a multimedia information recommendation processing function, or a server provided with a multimedia information recommendation device processing function, such as the preceding Figure 1 Server 200 in. Figure 2 The schematic diagram of the structure of the multimedia information recommendation device provided by the embodiment of the present invention can be understood as follows: Figure 2 Only an exemplary structure of the multimedia information recommendation device is shown, not all structures, and can be implemented as needed. Figure 2 Partial or complete structure shown.

[0090] The multimedia information recommendation device provided in the embodiment of the present invention includes: at least one processor 201, a memory 202, a user interface 203 and at least one network interface 204. The various components in the multimedia information recommendation device are coupled together through a bus system 205. It can be understood that the bus system 205 is used to realize the connection and communication between these components. In addition to the data bus, the bus system 205 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 205 is not used in the example of FIG. Figure 2 Various buses are labeled as bus system 205 .

[0091] The user interface 203 may include a display, a keyboard, a mouse, a trackball, a click wheel, keys, buttons, a touch pad or a touch screen.

[0092] It is understood that the memory 202 can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. The memory 202 in the embodiment of the present invention can store data to support the operation of the terminal (such as 10-1). Examples of these data include: any computer program for operating on the terminal (such as 10-1), such as an operating system and an application. Among them, the operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, etc., which are used to implement various basic services and process hardware-based tasks. The application can include various applications.

[0093] In some embodiments, the multimedia information recommendation device provided in the embodiment of the present invention can be implemented in a combination of software and hardware. As an example, the multimedia information recommendation device provided in the embodiment of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the multimedia information recommendation method provided in the embodiment of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPL Ds), field programmable gate arrays (FPGAs), or other electronic components.

[0094] As an example of a multimedia information recommendation device provided in an embodiment of the present invention being implemented by a combination of software and hardware, the multimedia information recommendation device provided in an embodiment of the present invention can be directly embodied as a combination of software modules executed by a processor 201. The software module can be located in a storage medium, and the storage medium is located in a memory 202. The processor 201 reads the executable instructions included in the software module in the memory 202, and combines with necessary hardware (for example, including the processor 201 and other components connected to the bus 205) to complete the multimedia information recommendation method provided in an embodiment of the present invention.

[0095] As an example, the processor 201 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.

[0096] As an example of a multimedia information recommendation device provided in an embodiment of the present invention being implemented by hardware, the device provided in an embodiment of the present invention can be directly executed by a processor 201 in the form of a hardware decoding processor, for example, 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 to implement the multimedia information recommendation method provided in an embodiment of the present invention.

[0097] The memory 202 in the embodiment of the present invention is used to store various types of data to support the operation of the multimedia information recommendation device. Examples of such data include: any executable instructions for operating on the multimedia information recommendation device, such as executable instructions, and the program for implementing the multimedia information recommendation method of the embodiment of the present invention can be included in the executable instructions.

[0098] In other embodiments, the multimedia information recommendation device provided by the embodiment of the present invention can be implemented in software. Figure 2 The multimedia information recommendation device stored in the memory 202 is shown, which may be software in the form of a program and a plug-in, and includes a series of modules. As an example of a program stored in the memory 202, a multimedia information recommendation device may be included, and the multimedia information recommendation device includes the following software modules:

[0099] Information transmission module 2081 and information processing module 2082. When the software modules in the multimedia information recommendation device are read into the RAM by the processor 201 and executed, the multimedia information recommendation method provided by the embodiment of the present invention will be implemented, wherein the functions of each software module in the multimedia information recommendation device include: an information transmission module, which is used to obtain the behavior parameter information of the target user in response to the multimedia information recommendation request;

[0100] The information transmission module 2081 is used to obtain first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types;

[0101] An information processing module 2082 is configured to predict, based on the first object feedback data, a first score corresponding to each of the first sub-object feedback data;

[0102] The information processing module 2082 is used to determine the fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information;

[0103] The information processing module 2082 is used to fuse the first scores corresponding to each of the first sub-object feedback data based on the fusion weight parameters corresponding to each of the first sub-object feedback data to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

[0104] according to Figure 2 The electronic device shown, in one aspect of the present invention, the present invention further provides a computer program product or a computer program, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. The processor of the electronic device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the electronic device executes the different embodiments and combinations of embodiments provided in various optional implementations of the multimedia information recommendation method described above.

[0105] Combination Figure 2 The multimedia information recommendation device shown in the figure illustrates the multimedia information recommendation method provided by the embodiment of the present invention. Figure 3 , Figure 3 This is an optional flow chart of the multimedia information recommendation method provided by the embodiment of the present invention. It can be understood that: Figure 3 The steps shown can be executed by various electronic devices running the multimedia information recommendation device, for example, a terminal, a server or a server cluster with a multimedia information recommendation device, wherein the terminal with the multimedia information recommendation device can be the preceding sequence Figure 2 The electronic device with multimedia information recommendation device in the embodiment shown in the figure is as follows. Figure 3 The steps shown are explained.

[0106] Step 301: The multimedia information recommendation apparatus obtains first object feedback data for target multimedia information.

[0107] In some embodiments, before obtaining the first object feedback data for the target multimedia information, the user can first trigger the client that presents the multimedia information. When the client is triggered and starts to recommend different multimedia information to the user, the server receives the multimedia information recommendation request sent by the terminal. The server responds to the multimedia information recommendation request and obtains the multimedia information to be recommended in the multimedia information data source. Using the multimedia information recommendation method provided by the present invention, these multimedia information to be recommended can be sorted according to the second score corresponding to each multimedia information, thereby recommending the most suitable multimedia information to the user.

[0108] Among them, the first object feedback data includes at least two first sub-object feedback data of different feedback types. Taking the multimedia information as a short video as an example, the user can share and / or like the exposed short video while watching the information, or collect the product provided in the video advertisement by clicking the hyperlink in the short video. When the multimedia information is a dynamic GIF advertisement, during the exposure of the advertisement through the terminal (terminal 10-1 and / or terminal 10-2), the user can forward and / or comment on the advertisement, or jump to the corresponding product purchase link page through the GIF advertisement. Each operation performed on the short video information can be regarded as the first sub-object feedback data of a feedback type.

[0109] Step 302: The multimedia information recommendation apparatus predicts a first score corresponding to each of the first sub-object feedback data based on the first object feedback data.

[0110] Step 303: The multimedia information recommendation device determines the fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended.

[0111] In step 303, the first object feature data may include but is not limited to: age features, gender features, consumption level features and user context features, wherein for any user who is the object to be recommended, the context of the user may include but is not limited to the time, place, mood, etc. when the user visits the recommendation system, which is very important for improving the accuracy of the recommendation. Among them, the time context features of the user's visit to the recommendation system can be used to reflect the changes in interests reflected in the user's recent behavior through the user's recent and long-term behaviors, thereby ensuring the continuity of the user's interest prediction during the recommendation process.

[0112] The multi-objective weighted search network used in the recommendation process is obtained by updating the network parameters of the initial multi-objective weighted search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information. Among them, the sample objects may include: one or more of low-consumption users, new users and high-consumption users. Specifically, new users are users who register and use the client for the first time; low-consumption users are users who have registered and used the short video client, but generate less user behavior information; high-consumption users are users who have registered and used the short video client and generate more user behavior information during the use of the client.

[0113] For example, in the recommendation environment of a short video client, the threshold of the number of user behavior information can be set to 10. When the number of user behavior information of a certain user is less than or equal to 10, the user can be regarded as a low-consumption user; when the number of user behavior information of a certain user is greater than 10, the user is regarded as a high-consumption user. It should be noted that for different multimedia information recommendation environments, the threshold of user behavior information can be dynamically adjusted, and the embodiments of the present invention do not impose specific limitations.

[0114] Before using the multi-objective weighted search network to implement multimedia information recommendation, the multi-objective weighted search network needs to be trained to determine the network parameters of the multi-objective weighted search network.

[0115] Among them, taking multimedia information as short videos as an example, before using the multi-objective weighted search network, it is necessary to train the multi-objective weighted search network to determine the network parameters of the multi-objective weighted search network. The determined network parameters of the multi-objective weighted search network can adapt to the current multimedia information recommendation environment, see Figure 4 , Figure 4 This is an optional flow chart of the multimedia information recommendation method provided by the embodiment of the present invention. It can be understood that: Figure 4 The steps shown may be executed by various servers running the multimedia information recommendation device, such as a dedicated terminal, server or server cluster with a video processing function.

[0116] Before introducing the multi-objective weighted search network training process provided by the present invention, the defects in the relevant information recommendation method technology are first explained. When implementing the recommendation of different types of multimedia information, the relevant technology can have two methods: single-objective optimization and multi-objective optimization. Take the single-objective optimization of the basic feature being the user's click-through rate as an example. After seeing a video or article, will the user click to watch or read it. However, in actual situations, users have a variety of behaviors on a product, such as: likes, forwarding, collections, comments, etc. The problem of optimizing the user's CTR for a single goal is that content recommendation to users based solely on the user's click behavior may not achieve the ideal recommendation effect. Users may click on popular content when they see it, which will result in a higher CTR, but the recommended content based on CTR cannot meet the actual needs of users.

[0117] There are four ways to perform multi-objective optimization. Taking the estimation of the probability of users reading, liking, forwarding, commenting, etc. as an example, it can include:

[0118] 1) Change sample weights, 2) Multi-model score fusion, 3) Ranking learning (Learning To Rank, LTR), 4) Multi-task learning (Multi-Task Learning, MTL). In multi-objective optimization, multi-objective fusion is an important step in generating ranking results. The multiple feedback behavior indicators predicted by the MTL model are fused into a score that reflects the user's satisfaction with the short video (i.e., the second score of the target multimedia information in the present invention). The recommended short videos are sorted by the score, thereby achieving the effect of maximizing the satisfaction of the recommendation system. There are two main methods for multi-objective fusion in the related art: formula fusion and ranking learning. Among them, formula fusion has the advantages of simple model and low cost compared to ranking learning and is widely used. Formula fusion is mainly divided into two parts: agent function design and fusion weight search. Agent functions mostly use linear weighted fusion or weighted multiplication fusion. The search for fusion weights generally includes manual parameter setting, grid search, and Bayesian optimization methods. Manual parameter setting is to select several groups of solutions as fusion candidate vectors based on subjective experience, and select an optimal solution through A / B testing. Grid search mainly searches for every possible combination of hyperparameter sets through exhaustive iteration, and uses cross-validation to select the optimal hyperparameter combination as the fusion weight. Bayesian optimization mainly builds a probability model based on the past evaluation results of the objective function, and infers the hyperparameter values ​​of extreme points as fusion weights by updating the model.

[0119] However, no matter which target optimization method is used, the training process will suffer from low parameter search efficiency and high experimental resource consumption. Grid search is an exhaustive search method for specifying parameter values, and the search efficiency is low. Although the Bayesian optimization method uses the information of the previous search point to guide the next parameter collection process, which improves the efficiency of the search, the entire process still requires multiple sets of parameters to be online to collect samples, which consumes a lot of resources.

[0120] At the same time, it is impossible to perform personalized modeling for special users. Since the results of grid search and Bayesian optimization search are still fixed parameters, personalized modeling cannot be performed, and adaptive adjustments cannot be made based on changes in the distribution of ranking scores, making the recommendation strategy for multimedia information too single, affecting the user experience.

[0121] More importantly, accurate video recommendations often require the collection of a large number of user operation records. Recommendation systems in real-world scenarios are often accompanied by the problem of few-sample learning, such as the recommendation of low-spending users (users with less sample information can become low-spending users) and new users. When the number of training samples is small, the model cannot fit the real behavior of these users. Therefore, it is often difficult to collect enough labeled samples for new users for traditional machine learning to extract information that can be recommended from the data, which can easily lead to model overfitting and introduce new noise, affecting the data processing effect of the model.

[0122] To address these deficiencies, refer to Figure 4 The multimedia information recommendation method provided in the embodiment of the present invention can be used for non-real-time multi-objective weighted search network training, and can adjust the recommendation order of any multimedia information (including various video types such as TV series, movies, short videos, etc.). Figure 4 The steps shown are explained.

[0123] Step 401: The multimedia information recommendation apparatus obtains a training sample set, where the training sample set includes the real object feedback data of the sample object for sample multimedia information.

[0124] The training samples include information on the operation behaviors of different types of users on multimedia information. It should be noted that different types of users are often generated due to different viewing habits of users. For users who often watch short videos, the viewing records can record different operations on short videos. The satisfaction evaluation of short video recommendation is based on relevant indicators of clicks, browsing depth (stay time), add-to-cart, favorites, purchase of goods in short videos, repeated purchase of goods in short videos, and praise of goods in short videos. Different recommendation systems, different periods, and different product forms often have different indicators; new users and users with fewer usage records often have less information on the operation behaviors of multimedia information, so the training samples formed are also less. During the training process, data in the training sample set can be extracted. Through different training processes, the trained multi-objective weight search network can process the features of different types of users to meet different recommendation environments for multimedia information recommendation.

[0125] Taking short videos as multimedia information, in order to achieve more personalized recommendations, when recommending target multimedia information, the multi-objective ranking model can be used to predict various indicators of the target multimedia information, where different indicators can correspond to different feedback types, specifically including: completion rate, consumption time, like rate, comment rate, purchase rate and favorable rate, etc. One or more. Thus, the first sub-object feedback data corresponding to different feedback types of the target multimedia information can be obtained, and based on the first sub-object feedback data, the corresponding first score can be predicted.

[0126] Furthermore, in order to fuse the prediction results of multiple indicators (which may correspond to the first score corresponding to each first sub-object feedback data) into a score that can measure the user's satisfaction with the target multimedia information (which may correspond to the second score of the target multimedia information) for recommending the target multimedia information, for example, the above operation may be performed for different target multimedia information to obtain the second score of each target multimedia information, and the target multimedia information may be sorted accordingly, and then based on the sorting results of each target multimedia information, the head (for example, the first one in the sorting result or the first two in the sorting result) target multimedia information may be preferentially recommended to the user terminal. In an exemplary embodiment, the first score corresponding to each first sub-object feedback data may be fused by a target fusion function to obtain the second score of the target multimedia information, and the expression of the target fusion function may be as shown in Formula 1.

[0127] Fus score =i∑(α i *log(pred scorei +β i )+γ i ) Formula 1

[0128] Among them, pred scorei Represents the first score corresponding to the first sub-object feedback data, Fus score The second score characterizing the target multimedia information, α, β and γ are hyperparameters. α, as the fusion weight parameter of the fusion target, measures the user's emphasis on each target, which will directly affect the accuracy of the fusion function. To this end, it is necessary to dynamically provide personalized fusion weight parameters for each user through the multi-target weight search network provided by the present invention, so that the recommended multimedia information will be more accurate and the user will have a better viewing experience.

[0129] In some embodiments of the present invention, the model structure of the multi-objective sorting model of the present invention can be replaced by: Long Short-Term Memory Network (LSTM), Gated Recurrent Unit Network Model (GRU) Model, Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Deep Neural Network (DNN) and Three-factor Model (Fama-French 3-factor model). The specific model structure can be flexibly selected according to different types of multimedia information. For example, when the multimedia information is a short video, the multi-objective sorting model can be constructed based on the long short-term memory network. When the multimedia information is a video advertisement, the multi-objective sorting model can be constructed based on the recurrent neural network. There is no specific limitation on this in the implementation of the present invention. It should be noted that no matter which model structure the multi-objective sorting model uses, the multi-objective weight search network provided by the present invention can be used to dynamically provide personalized fusion weight parameters for each user.

[0130] See also Figure 5 , Figure 5The figure is a schematic diagram of the structure of the multi-objective weight search network in the embodiment of the present invention, and the multi-objective weight search network (MOWSN) used in the present invention. Among them, MOWSN is a perceptron model. Specifically, the perceptron is a linear classification model for two-category classification. The input is the feature vector of the instance, and the output is the category of the instance, which takes the binary values ​​of +1 and -1. The perceptron corresponds to the separating hyperplane that divides the instance into positive and negative categories in the input space (feature space). It is a discriminant model that aims to find the separating hyperplane that linearly divides the training data. To this end, a loss function based on misclassification is introduced, and the gradient descent method is used to minimize the loss function to obtain the perceptron model, and the perceptron model is used to classify the new input. Given an input X=(X1.X2…X n ), where n represents the dimension of the input data, and xi represents a feature of the user, including the user's age, gender, consumption level, user context features, etc. The output of the model is the weight α of each target k (k∈{1,2,3...m}), where m represents the number of objectives in the multi-objective ranking model. k , can be calculated by formula 2 as follows:

[0131] α k =W T X+b Formula 2

[0132] where W∈{w1,w2,w i ....w n} are the parameters of the perceptron model, and b represents the bias of the perceptron model.

[0133] Step 402: The multimedia information recommendation device uses the training sample set to update the network parameters of the initial multi-objective weighted search network through an evolutionary strategy algorithm, and obtains the multi-objective weighted search network through the updated network parameters.

[0134] Furthermore, accurate video recommendation often requires the collection of a large number of user operation records. Recommendation systems are often accompanied by a few-sample learning problem in real-world scenarios, such as the recommendation of low-spending users (users with less sample information can become low-spending users) and new users. When the number of training samples is small, the model cannot fit the real behavior of these users. Therefore, for this situation, step 403 can also be performed.

[0135] Step 403: The multimedia information recommendation device uses the training sample set to update the network parameters of the initial multi-objective weighted search network through a meta-learning framework based on an evolutionary strategy algorithm, and obtains the multi-objective weighted search network based on the updated network parameters.

[0136] The following describes the training methods of the two multi-objective weighted search networks involved in step 402 and step 403, wherein, in the technical solution of the present invention, the evolutionary learning algorithm can use the evolution strategy (ES) algorithm. Other algorithms can also be used for implementation, and the embodiment of the present invention does not make any limitation to this.

[0137] When executing step 402, the network parameters of the initial multi-objective weighted search network are updated by the evolutionary strategy algorithm. In the process of obtaining the multi-objective weighted search network by the updated network parameters, the evolutionary strategy algorithm used is referenced to Table 1. Figure 6 The different processing steps in the paper illustrate the process of updating the network parameters of the initial multi-objective weight search network through the evolution strategy algorithm.

[0138] Table 1

[0139]

[0140]

[0141] refer to Figure 6 , Figure 6 The following is a schematic diagram of an optional processing process of the multimedia information recommendation method in an embodiment of the present invention, which specifically includes the following steps:

[0142] Step 601: Calculate the network parameter update direction of the initial multi-objective weight search network according to the noise standard deviation parameter of the initial multi-objective weight search network.

[0143] Among them, since the noise ε follows a Gaussian distribution, the update direction of the network parameter W can be obtained by

[0144] Formula 3 represents:

[0145]

[0146] Step 602: Obtain a noise standard deviation parameter and a learning rate parameter of the multi-objective weight search network.

[0147] As shown in the first row of Table 1, when the evolution strategy algorithm is used, the noise standard deviation parameter obtained is σ, the learning rate parameter of the multi-objective weight search network is θ, and the network parameter W0 of the initial multi-objective weight search network can also be obtained.

[0148] In some embodiments of the present invention, when obtaining the noise standard deviation parameter and the learning rate parameter of the multi-objective weight search network, in order to ensure that the parameters of the multi-objective weight search network finally determined are more reasonable, different learning rate parameters can be set according to the different types of recommendation environments of the multi-objective weight search network. For example, the learning rate parameter in the short video recommendation environment including entertainment information can be smaller than the learning rate parameter of the short video recommendation of shopping promotion, so as to achieve accurate sorting of the short video recommendations of shopping promotion and guide users to generate more purchasing behaviors.

[0149] Step 603: According to the noise standard deviation parameter, configure a first random noise parameter for each sample object in the training sample set.

[0150] Among them, the target users involved in Table 1 are Figure 6 The sample objects in the processing steps shown will not be repeated later. As shown in rows 4-5 of Table 1, for each target user i=1, ...n, it is necessary to configure random noise parameters, such as randomly generating ε1...εn~N(0,I). Before configuring the first random noise parameter, as shown in row 2 of Table 1, it is necessary to initialize n target users using a known random seed and network parameter W0. Since the random numbers generated during the training process are all pseudo-random numbers, they are a series of values ​​generated by the algorithm. Therefore, it is necessary to give the function a random value as the initial value, and continuously iterate based on this benchmark to obtain a series of random numbers. This initial value is called a random seed, for example, the known random seed in Schedule 1 at the start of the server training of the multi-target weighted search network can be used.

[0151] Step 604: According to the first random noise parameter, the network parameter of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, determine the fitness score corresponding to each of the sample objects, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object.

[0152] Among them, as shown in the sixth row of Table 1, by calculating the fitness function F i =F(W t +σε i ), for each target user i=1, ...n, the corresponding fitness score can be calculated. Through the processing of step 604, the scalar return value Fi of each sample object can be used as the real object feedback data of the sample object.

[0153] Step 605: Share the real object feedback data of the sample object among all sample objects in the training sample set.

[0154] The present invention needs to explain that the evolutionary strategy algorithm is to perform global optimization by simulating evolution in nature, and follows Darwin's evolutionary theory of natural selection and survival of the fittest. Therefore, when the model training based on the evolutionary strategy algorithm is realized, it only depends on the user feedback data. Through step 605, the real object feedback data of the sample object is shared among all sample objects, so that it can be independent of the back propagation of the error, and is therefore more suitable for the optimization process of the neural network model including multi-stage and multi-strategy. At the same time, the calculation using the evolutionary strategy algorithm is simple and efficient, can be highly parallelized, and has low resource loss, so it can also effectively reduce the hardware overhead of model training, and is also convenient for deploying the multi-objective weight search network in the mobile terminal, so as to realize the large-scale application of the multi-objective weight search network.

[0155] Step 606: configure a second random noise parameter for each sample object in the training sample set.

[0156] Since the real object feedback data of the sample object is shared among all sample objects in the training sample set through step 605, the real object feedback data may have an impact on the behavior of other sample objects, so the noise parameters need to be readjusted. As shown in the 10th row of Table 1, the second random noise parameters ε1…εn~N(0,I) are continuously reconfigured using the known random seed, and the training process is implemented using the second random noise parameters.

[0157] Step 607: Determine the first gradient change parameters of different rounds of the initial multi-objective weight search network according to the second random noise parameter, the network parameters of the initial multi-objective weight search network and the noise standard deviation parameter.

[0158] Step 608: According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the first gradient change parameters of different rounds of the initial multi-objective weight search network.

[0159] In some embodiments, the recommendation effect can be measured by one or more of the number and duration of videos actually watched by the target user, that is, the real object feedback data may include one or more of the number and duration of videos actually watched. Multiple target users are highly parallel at the same time, and different target users communicate only through a scalar Fi. Through the evolutionary strategy algorithm shown in Table 1, the network parameter W can be updated according to the user's real-time feedback behavior, so that personalized and highly accurate target fusion parameters can be obtained. In the updating process, as shown in row 11 of Table 1, the network parameters of the initial multi-objective weighted search network can be updated as follows: Among them, the learning rate θ can be used to control the round of the first gradient change parameter and adjust the accuracy of the multi-objective weight search network.

[0160] Among them, reference Figure 7 , Figure 7 FIG. 4 is a schematic diagram of the effect of the evolutionary strategy algorithm in the embodiment of the present invention. The following takes the round number as 4 to illustrate the processing effect of the evolutionary strategy algorithm in the embodiment of the present invention. Figure 7 As shown in the figure, after 4 iterations (learning rate θ is 4), reward increases from -0.13 to 0.4. In order to better perform the update of network parameters through computer programs, refer to Table 1. When optimizing the update results of the first network parameters, the evolutionary strategy algorithm that can be used is shown in Table 1, and θ = 0.01, σ = 0.5 are set. The fitness function F(*) measures the user's real consumption behavior score when the network parameter W is given. This score is the user's real-time feedback.

[0161] Next, the process of updating the network parameters of the initial multi-objective weight search network based on the meta-learning framework of the evolution strategy algorithm provided in step 403 is described. Figure 6In the processing process shown in FIG. 1 , the trained multi-objective weight search network can already process the feature vectors of target objects with a large sample capacity (greater than the sample capacity threshold), but for a small sample object with a small sample capacity (less than or equal to the sample capacity threshold), it is still unable to accurately process the feature vectors. Therefore, more targeted training can be performed by extracting samples of small sample users in the training sample set. This training process uses fine-tuning and a meta-learning framework to process the multi-objective weight search network. The principle of the fine-tuning method is to modify the output layer using known network structures and known network parameters, and use samples of small sample users (samples corresponding to users with sample capacity less than or equal to the sample capacity threshold) as training samples to fine-tune the parameters of all layers before the last layer in the multi-objective weight search network, and increase the learning rate of the last layer of the network to achieve the purpose of reducing training costs. The meta-learning framework is used in the training process of this fine-tuning method. The meta-learning framework is briefly described below. The learning method based on meta-learning consists of a meta-learner and a base-learner. The meta-learner mainly learns the commonalities between different tasks through a small number of samples. The base learner mainly generates a neural network that is more suitable for the characteristics of the task after performing at least one gradient iteration using a small number of samples of a task based on the knowledge learned by the meta-learner. For example, the meta-learner in the Model-Agnostic Meta-Learning (MAML) algorithm is responsible for seeking the initialization parameters of the base learner, and its training process is to minimize the meta-loss of the meta-learner on a large number of target tasks (i.e., base learners). The base learner is the prediction model used by the target task, which is given initialization parameters by the meta-learner and then trained through a small number of gradient iterations; for example, in the parameter generation algorithm based on meta-learning, the meta-learner learns to generate the common network parameters of the classification layer of the base learner, and the new task uses a small number of sample gradient iterations to generate the base learner of the task.

[0162] pass Figure 6 The training process shown in the figure enables the multi-objective weighted search network to recommend videos to users with abundant samples. However, for users with few samples, the multi-objective weighted search network can be trained based on the meta-learning framework of the evolutionary strategy algorithm. Please refer to Table 2.

[0163] Table 2

[0164]

[0165]

[0166] Among them, p(Τ) is the set of tasks. Tasks are divided into users, and each task represents all training samples of the user. Set θ a =0.02,θ b =0.01,σ=0.6.

[0167] Combined with the processing process shown in Table 2, refer to Figure 8 , Figure 8 The following is a schematic diagram of an optional processing process of the multimedia information recommendation method in an embodiment of the present invention, which specifically includes the following steps:

[0168] Step 801: extracting a first training sample set of a target number of tasks from the training sample set; wherein each of the first training sample sets corresponds to a sample object.

[0169] Step 802: Based on at least one task number set, configure a support set for the initial multi-objective weighted search network.

[0170] Specifically, a batch of tasks can be randomly selected, and the support set support-set can be used to train the base learner in each task to achieve the effect of updating the network parameters of the initial multi-objective weight search network. After the base learner training of this batch of tasks is completed, the training samples in the query set query-set are used to update the meta learner using the sum of the test results. This process is called slow weight update.

[0171] As shown in Table 2, a set of tasks T is randomly selected from p(T) t , T t Each task in is denoted as task i For each task i , extract ns (preferably set to 5) samples to generate the support-set of the task, recorded as support-task i , the rest are query-sets, denoted as query-task i .

[0172] Since the meta-learning framework shown in the figure still uses the evolutionary strategy algorithm to update the parameters of the multi-objective weight search network, the method shown in Table 1 is used to update the parameters of the multi-objective weight search network through the base learner and the meta-learner, as shown in Table 2. In this embodiment, the base learner executes the loop process of 7-9 in Table 2 to update the network parameters of the initial multi-objective weight search network, and finally determines the network parameters of the multi-objective weight search network. The meta-learner executes the loop process of 10-11 in Table 2, initializes the network parameters of the base learner, and then trains through a small number of gradient iterations.

[0173] Step 803: Calculate the network parameter update direction of the initial multi-objective weight search network according to the noise standard deviation parameter of the initial multi-objective weight search network.

[0174] Step 804: According to the noise standard deviation parameter, configure a third random noise parameter for each sample object in the support set.

[0175] Among them, the target users involved in Table 1 are Figure 8 The sample objects in the processing steps shown will not be repeated later. As shown in rows 4-5 in Table 1, for each target user i=1,...n, it is necessary to configure the third machine noise parameters, such as randomly generating ε1...εn~N(0,I). This process can start the server to train the multi-target weight search network with the known random seed in schedule 1.

[0176] Step 805: According to the third random noise parameter, the network parameters of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, determine the fitness score corresponding to each of the sample objects in the support set, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object.

[0177] Step 806: Share the real object feedback data of the sample object among all sample objects in the support set.

[0178] Among them, as shown in the sixth row of Table 1, by calculating the fitness function F i =F(W t +σε i ), for each target user i=1, ...n, the corresponding fitness score can be calculated. Through the processing of steps 805 and 806, the scalar return value Fi of each sample object can be used as the real object feedback data of the sample object. This process can also be independent of the back propagation of errors. Therefore, the meta-learning framework based on the evolution strategy algorithm is also applicable to the optimization process of neural network models containing multiple stages and multiple strategies. At the same time, the evolution strategy algorithm is simple and efficient in calculation, highly parallelizable, and has low resource consumption, so it can also effectively reduce the hardware overhead of model training.

[0179] Step 807: configuring a fourth random noise parameter for each sample object in the support set.

[0180] Since the real object feedback data of the sample object is shared among all sample objects in the training sample set through step 807, the real object feedback data may have an impact on the behavior of other sample objects, so the noise parameters need to be readjusted. As shown in the 10th row of Table 1, the fourth random noise parameter ε1…εn~N(0,I) is continuously reconfigured using the known random seed, and the model training is continued using the fourth random noise parameter.

[0181] Step 808: Determine the second gradient change parameters of different rounds of the initial multi-objective weight search network according to the fourth random noise parameter, the network parameters of the initial multi-objective weight search network and the noise standard deviation parameter.

[0182] Through the process shown in Table 2, use support-task i , use the evolutionary strategy algorithm to update the network parameters and obtain the network parameters W' for this task i For T t For each task in i Construct a sample set and use the evolution strategy algorithm to update the network parameter w. Thus, the trained multi-objective weight search network solves the problem of poor fitting ability caused by the lack of labeled samples.

[0183] Step 809: According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the second gradient change parameters of different rounds of the initial multi-objective weight search network to obtain the first network parameters of the initial multi-objective weight search network.

[0184] The recommendation effect can be measured by the number and duration of videos actually watched by the user. Multiple target users are highly concurrent, and different target users communicate only through a scalar Fi. Through the evolutionary strategy algorithm shown in Table 1, the network parameter W can be updated according to the user's real-time feedback behavior, so that personalized and highly accurate target fusion parameters can be obtained. In the updating process, as shown in the 8th row of Table 2, the network parameters of the initial multi-target weighted search network can be updated as follows: Among them, the learning rate θ can be used to control the rounds of the second gradient change parameters and adjust the accuracy of the multi-objective weight search network.

[0185] because Figure 8The processing steps shown can be used to process the training samples of users with few samples. When using the model-agnostic meta-learning algorithm, the meta-learner is responsible for finding the initialization parameters of the base learner. The training process is to minimize the meta-loss of the meta-learner on a large number of target tasks (i.e., base learners). The base learner is the prediction model used by the target task. The meta-learner gives it initialization parameters and then trains it through a small number of gradient iterations. Therefore, the meta-learner also needs to be trained.

[0186] As shown in Table 2, in this embodiment, the base learner executes the loop process of 7-9 in Table 2 to update the network parameters of the initial multi-objective weight search network and finally determine the network parameters of the multi-objective weight search network. The meta learner executes the loop process of 10-11 in Table 2 to initialize the network parameters of the base learner and then train them through a small number of gradient iterations. In this process, the parameter update still uses the evolution strategy algorithm. Specifically, referring to Table 2, a query set query set can be configured for the initial multi-objective weight search network based on at least one task number set; the network parameter update direction of the initial multi-objective weight search network is calculated according to the noise standard deviation parameter of the initial multi-objective weight search network; according to the noise standard deviation parameter, a fifth random noise parameter is configured for each sample object in the query set, wherein the configuration method of the fifth random noise parameter is shown in rows 4-5 of Table 1. For each target user i=1,...n, the fifth random noise parameter needs to be configured, for example, ε1...εn~N(0,I) is randomly generated. This process can start the server to train the multi-objective weight search network with the known random seed in schedule 1.

[0187] Afterwards, according to the fifth random noise parameter, the network parameters of the initial multi-objective weighted search network and the noise standard deviation parameter, the fitness score corresponding to each of the sample objects in the query set is determined through the fitness function corresponding to the initial multi-objective weighted search network, and the fitness score corresponding to the sample object is used as the real object feedback data of the sample object; the real object feedback data of the sample object is shared among all the sample objects in the query set. Specifically, as shown in the sixth row of Table 1, by calculating the fitness function F i =F(W t +σε i ), for each target user i=1,…n, the corresponding fitness score can be calculated, and the scalar return value Fi of each sample object can be used as the real object feedback data of the sample object.

[0188] Finally, a sixth random noise parameter is configured for each sample object in the query set; the third gradient change parameter of different rounds of the initial multi-objective weight search network is determined according to the sixth random noise parameter, the network parameter of the initial multi-objective weight search network, and the noise standard deviation parameter; according to the network parameter update direction of the initial multi-objective weight search network, the network parameter of the initial multi-objective weight search network is updated using the third gradient change parameter of different rounds of the initial multi-objective weight search network and the first network parameter of the initial multi-objective weight search network to obtain the second network parameter of the initial multi-objective weight search network. As shown in rows 10-11 of Table 2, the meta-learner uses T t All query-tasks i , using the third gradient change parameters of different rounds of the initial multi-objective weight search network and the first network parameters of the initial multi-objective weight search network, by Continue to update W t+1 , the updated W t+1 can continue to be used by the base learner, as shown in row 8 of Table 2 The network parameters of the multi-objective weight search network are continuously updated in this manner until the network parameters of the target weight search network are determined, thereby ending the training of the target weight search network.

[0189] Therefore, for users with less sample information and new users, Figure 8 The method shown can solve the defect of poor fitting ability of multi-objective weighted search network and realize accurate multimedia information recommendation for users with less sample information and new users.

[0190] Step 305: The multimedia information recommendation device fuses the first scores corresponding to each of the first sub-object feedback data based on the fusion weight parameters corresponding to each of the first sub-object feedback data to obtain the second score of the target multimedia information. The second score of the target multimedia information is used to recommend the multimedia information to the object to be recommended.

[0191] Among them, the second score of the target multimedia information is used to characterize the order of recommending multimedia information to the user. For example, the two multimedia information are video advertisements A and B. When the second score of the target multimedia information of A is 1 and the second score of the target multimedia information of B is 2, it means that the current user may be more interested in video advertisement B. Therefore, according to the second score of the target multimedia information, advertisement A is replaced by advertisement B to configure more playback traffic for advertisement B, increase the exposure rate of advertisement B, and enable users to have a better viewing experience.

[0192] It should be noted that, since the multi-objective weighted search network can be set as a multi-objective weighted search network of different dimensions (for example, a multi-objective weighted search network of the user dimension and a multi-objective weighted search network of the video dimension to be recommended can be set), at least two multi-objective weighted search networks can be used in a superimposed manner to meet the needs of different scenarios. Specifically, the number of multi-objective weighted search networks can be determined according to the type of multimedia information recommendation environment; when the number of the multi-objective weighted search networks is two, the fusion weight parameters corresponding to the two multi-objective weighted search networks are determined, wherein the first fusion weight parameter corresponds to the behavioral information dimension of the first object, and the second fusion weight parameter corresponds to the multimedia information dimension; based on the first fusion weight parameter and the second fusion weight parameter, the corresponding second score is determined. For the multimedia information dimension, the multimedia information to be recommended is obtained from the multimedia information data source; the different types of sub-information included in the multimedia information to be recommended are processed through a multi-objective weight search network to determine the identification embedding features, text embedding features, and statistical features that match the multimedia information to be recommended; the attribute types of the multimedia information to be recommended are converted to determine the multimedia information attribute features corresponding to the multimedia information to be recommended, and the user attribute features corresponding to the multimedia information to be recommended are determined; feature fusion processing is performed on the identification embedding features, text embedding features, statistical features, multimedia information attribute features, and user attribute features; based on the result of the feature fusion processing, the fused weight information is determined through a multi-objective weight search network, so that the multimedia information to be recommended to the user can be more accurately recommended.

[0193] The following takes the video news information recommendation environment in the short video playback interface as an example to illustrate the multimedia information recommendation method provided by the embodiment of the present invention, wherein: Fig. 9A FIG. 1 is a schematic diagram showing the effect of the multimedia information recommendation method according to an embodiment of the present invention, wherein: Fig. 9AAs shown, the video news information playback interface can be displayed in the corresponding APP, or it can be triggered by the instant messaging client applet (the multi-objective weighted search network can be encapsulated in the corresponding APP after training or saved in the instant messaging client applet in the form of a plug-in, and the recommendation environment is the recommendation of news information). With the continuous development and increase of short video application products, the carrying capacity of video news video information is far greater than that of text information. Video news information can be continuously recommended to users through the corresponding application. For example, the "Take a look" entrance included in the discovery page of the instant messaging client application, or the audio recommendation entrance of the audio application, or the video recommendation entrance of the video application, or the live broadcast recommendation entrance of the live broadcast application, etc. When the target terminal runs the target application according to the user operation and controls the target application to display the application page including the trigger entrance for triggering the opening of the recommended content display page, the trigger operation of the trigger entrance can be detected. When a trigger operation corresponding to the trigger entry is generated, a recommendation request is sent to the server, and after receiving the recommended content fed back by the server in response to the recommendation request, the multimedia information recommendation method provided by the present invention uses a multi-objective weight search network to determine the corresponding fusion weight parameter, calculates the second score of different target multimedia information, and displays the recommended content according to the second score of the target multimedia information on the recommended content display page, such as Fig. 9A As shown, for the recommended object (that is, the short video viewing user) 1, the second score of short video A is 1, and the second score of short video B is 2. Therefore, the short video B is first recommended to user 1. Fig. 9B FIG. 1 is a schematic diagram showing the effect of the multimedia information recommendation method according to an embodiment of the present invention. Fig. 9B As shown, for the recommended object 2, the second score of short video A is 2, and the second score of short video B is 1. Therefore, the short video A is first recommended to the user. In this way, short videos can be accurately recommended to different recommended objects. At the same time, when the recommended object 2 is a new user of the short video client (a user who registers and uses the short video client for the first time), the probability of the new user browsing the recommended short video increases by 8%, and the duration increases by 1.2%. Therefore, the multimedia information recommendation method of the present invention can also achieve effective recommendations for new users.

[0194] The present invention has the following beneficial technical effects:

[0195] The embodiment of the present invention obtains first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types; based on the first object feedback data, predicts a first score corresponding to each first sub-object feedback data; based on the first object feature data of the object to be recommended, determines the fusion weight parameter corresponding to each first sub-object feedback data through a multi-objective weight search network; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information; based on the fusion weight parameter corresponding to each first sub-object feedback data, the first score corresponding to each first sub-object feedback data is fused to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended. In this way, it is possible to perform fusion processing on the first score corresponding to the feedback data of each first sub-object to obtain the second score of the target multimedia information, and use the second score to accurately recommend multimedia information to the object to be recommended. At the same time, the generalization ability and data processing capability of the multi-target weighted search network are stronger, and the hardware overhead is smaller. It is also convenient to deploy the multi-target weighted search network in a mobile terminal to realize the large-scale application of the multi-target weighted search network.

[0196] The above descriptions are merely embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multimedia information recommendation method, characterized in that: The method comprises: Acquire first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types; Based on the first object feedback data, predict a first score corresponding to each first sub-object feedback data; Based on the first object feature data of the object to be recommended, a fusion weight parameter corresponding to each of the first sub-object feedback data is determined through a multi-objective weight search network; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information; Based on the fusion weight parameter corresponding to each of the first sub-object feedback data, the first score corresponding to each of the first sub-object feedback data is fused to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

2. The method according to claim 1, characterized in that Before determining the fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended, the method further includes: Acquire a training sample set, wherein the training sample set includes the real object feedback data of the sample object for the sample multimedia information; The training sample set is used to update the network parameters of the initial multi-objective weighted search network through an evolutionary strategy algorithm, and the multi-objective weighted search network is obtained through the updated network parameters.

3. The method according to claim 2, characterized in that The method of using the training sample set to update the network parameters of the initial multi-objective weighted search network through an evolutionary strategy algorithm includes: Calculating a network parameter update direction of the initial multi-objective weight search network according to a noise standard deviation parameter of the initial multi-objective weight search network; According to the noise standard deviation parameter, configuring a first random noise parameter for each sample object in the training sample set; According to the first random noise parameter, the network parameter of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, determine the fitness score corresponding to each of the sample objects, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object; Sharing the real object feedback data of the sample object among all sample objects in the training sample set; configuring a second random noise parameter for each sample object in the training sample set; Determining first gradient change parameters of different rounds of the initial multi-objective weight search network according to the second random noise parameter, the network parameter of the initial multi-objective weight search network, and the noise standard deviation parameter; According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the first gradient change parameters of different rounds of the initial multi-objective weight search network.

4. The method according to claim 2, characterized in that: The method of using the training sample set to update the network parameters of the initial multi-objective weighted search network through an evolutionary strategy algorithm, and obtaining the multi-objective weighted search network through the updated network parameters, comprises: The training sample set is used to update the network parameters of the initial multi-objective weighted search network through a meta-learning framework based on an evolutionary strategy algorithm, and the multi-objective weighted search network is obtained based on the updated network parameters.

5. The method according to claim 4, characterized in that The method of using the training sample set to update the network parameters of the initial multi-objective weighted search network by using an evolutionary strategy algorithm in a meta-learning manner includes: Extracting a first training sample set of the target number of tasks from the training sample set; wherein each of the first training sample sets corresponds to a sample object; Based on at least one set of task numbers, configuring a support set for the initial multi-objective weighted search network; Calculating a network parameter update direction of the initial multi-objective weight search network according to a noise standard deviation parameter of the initial multi-objective weight search network; According to the noise standard deviation parameter, configuring a third random noise parameter for each sample object in the support set; According to the third random noise parameter, the network parameter of the initial multi-objective weight search network and the noise standard deviation parameter, and through the fitness function corresponding to the initial multi-objective weight search network, determine the fitness score corresponding to each of the sample objects in the support set, and use the fitness score corresponding to the sample object as the real object feedback data of the sample object; Sharing the real object feedback data of the sample object among all sample objects in the support set; configuring a fourth random noise parameter for each sample object in the support set; Determining second gradient change parameters of different rounds of the initial multi-objective weight search network according to the fourth random noise parameter, the network parameter of the initial multi-objective weight search network, and the noise standard deviation parameter; According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the second gradient change parameters of different rounds of the initial multi-objective weight search network to obtain the first network parameters of the initial multi-objective weight search network.

6. The method according to claim 5, characterized in that The method further comprises: Based on at least one set of task numbers, configuring a query set for the initial multi-objective weighted search network; Calculating a network parameter update direction of the initial multi-objective weight search network according to a noise standard deviation parameter of the initial multi-objective weight search network; According to the noise standard deviation parameter, configuring a fifth random noise parameter for each sample object in the query set; According to the fifth random noise parameter, the network parameter of the initial multi-objective weighted search network and the noise standard deviation parameter, a fitness score corresponding to each of the sample objects in the query set is determined by a fitness function corresponding to the initial multi-objective weighted search network, and the fitness score corresponding to the sample object is used as the real object feedback data of the sample object; sharing the real object feedback data of the sample object among all sample objects in the query set, and configuring a sixth random noise parameter for each sample object in the query set; Determining third gradient change parameters of different rounds of the initial multi-objective weight search network according to the sixth random noise parameter, the network parameters of the initial multi-objective weight search network, and the noise standard deviation parameter; According to the network parameter update direction of the initial multi-objective weight search network, the network parameters of the initial multi-objective weight search network are updated using the third gradient change parameters of different rounds of the initial multi-objective weight search network and the first network parameters of the initial multi-objective weight search network to obtain the second network parameters of the initial multi-objective weight search network.

7. A multimedia information recommendation device, characterized in that: The device comprises: An information transmission module, configured to obtain first object feedback data for target multimedia information, wherein the first object feedback data includes at least two first sub-object feedback data of different feedback types; An information processing module, configured to predict, based on the first object feedback data, a first score corresponding to each of the first sub-object feedback data; The information processing module is used to determine the fusion weight parameter corresponding to each of the first sub-object feedback data through a multi-objective weight search network based on the first object feature data of the object to be recommended; wherein the multi-objective weight search network is obtained by updating the network parameters of the initial multi-objective weight search network through evolutionary learning based on the second object feature data of the object to be recommended and the real object feedback data of the sample object for the sample multimedia information; each real object feedback data for the sample multimedia information is the object feedback data of the sample object for the recommended sample multimedia information; The information processing module is used to fuse the first scores corresponding to each of the first sub-object feedback data based on the fusion weight parameters corresponding to each of the first sub-object feedback data to obtain the second score of the target multimedia information, and the second score of the target multimedia information is used to recommend multimedia information to the object to be recommended.

8. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, the multimedia information recommendation method according to any one of claims 1 to 6 is implemented.

9. An electronic device, characterized in that: The electronic device comprises: A memory for storing executable instructions; The processor is used to implement the multimedia information recommendation method according to any one of claims 1 to 6 when running the executable instructions stored in the memory.

10. A computer-readable storage medium storing executable instructions, characterized in that: When the executable instructions are executed by the processor, the multimedia information recommendation method according to any one of claims 1 to 6 is implemented.

Images