Multimedia information recommendation method and device, equipment and storage medium

Abstract

The invention discloses a multimedia information recommendation method and device, equipment and a storage medium, and the method and device can be used for scenes of cloud technology, artificial intelligence, smart traffic, auxiliary driving and the like. Iteration processing is performed on an estimation model comprising a target estimation network and a reference estimation network which are applied to different bidding mechanisms, and each round of iteration: for each piece of sample bidding information, on the basis of a bidding mechanism associated with a corresponding sample bidding result, a matched estimation network is adopted to estimate the bidding result of each piece of sample bidding information; obtaining an estimated probability set of successful recommendation of the corresponding sample multimedia information based on each estimated bid; if the pre-estimated network is a reference pre-estimated network, performing parameter adjustment on the reference pre-estimated network based on the pre-estimated probability set; and if the pre-estimation network is a target pre-estimation network, performing parameter adjustment on the pre-estimation model on a mapping probability set of a pre-estimation probability set based on each specified reference pre-estimation probability set so as to improve the information recommendation precision.

Description

Technical Field

[0001] This application relates to the field of Internet technologies, and in particular, to a method, apparatus, device, and storage medium for recommending multimedia information. Background Art

[0002] With the development of Internet technologies, bidders for Internet multimedia information can obtain display opportunities for multimedia information through bidding to achieve real-time recommendation of multimedia information; for example, advertisers rely on Real-Time Bidding (RTB) technology to recommend advertisements in real time.

[0003] Under related technologies, for a bidder, usually according to a preset probability distribution function, the historical bids and historical bidding situations (such as: bidding success, bidding failure) of multiple historical multimedia information are respectively fitted to obtain a historical probability distribution fitted by the corresponding bidding success of each historical bid, and then based on this historical probability distribution, the target bid corresponding to the maximum historical probability is adopted to achieve the recommendation of the target multimedia information to be recommended.

[0004] However, in practical applications, when the distribution of the preset probability distribution function does not match the distribution of the true probability distribution of the historical bids and their historical bidding situations of multiple historical multimedia information, it is difficult to obtain a reliable historical probability distribution based on the unreasonable preset probability distribution function, and further the accuracy of the target bid obtained based on this historical probability distribution is also difficult to guarantee, resulting in the inability to achieve accurate recommendation of the target multimedia information based on the inaccurate target bid. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for recommending multimedia information, which is used to optimize the training process of the prediction model corresponding to the multimedia information to improve the recommendation accuracy of the multimedia information.

[0006] In a first aspect, this application provides a method for recommending multimedia information, including:

[0007] Obtain multiple pieces of sample bidding information and corresponding sample bidding results; each piece of sample bidding information includes at least: sample multimedia information and object information of a sample object;

[0008] Based on the multiple pieces of sample bidding information, perform at least one round of iterative processing on the prediction model to be trained; wherein, the prediction model includes: a target prediction network applied to different bidding mechanisms and at least one reference prediction network, and each round of iteration includes:

[0009] For each piece of sample bidding information, based on the bidding mechanism associated with the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid price;

[0010] If the prediction network is the reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities;

[0011] If the prediction network is the target prediction network, the parameters of the prediction model are adjusted based on the mapped probability sets of the set of prediction probabilities with respect to the specified reference prediction probability sets respectively.

[0012] In a second aspect, the present application provides a multimedia information recommendation device, including:

[0013] An acquisition module, configured to acquire multiple pieces of sample bidding information and corresponding sample bidding results; each piece of sample bidding information at least includes: sample multimedia information and object information of the sample object;

[0014] A processing module, configured to perform at least one round of iterative processing on a prediction model to be trained based on the multiple pieces of sample bidding information; wherein, the prediction model includes: a target prediction network and at least one reference prediction network applied to different bidding mechanisms, and each round of iteration includes: for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid price; if the prediction network is the reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities; if the prediction network is the target prediction network, the parameters of the prediction model are adjusted based on the mapped probability sets of the set of prediction probabilities with respect to the specified reference prediction probability sets respectively.

[0015] In a third aspect, the present application provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the method for recommending any multimedia information in the first aspect above is implemented.

[0016] In a fourth aspect, the present application provides a computer storage medium, and computer program instructions are stored in the computer-readable storage medium, and when the computer program instructions are executed by the processor, the method for recommending any multimedia information in the first aspect above is implemented.

[0017] In a fifth aspect, a computer program product provided by an embodiment of the present application includes computer program instructions, and when the computer program instructions are executed by a processor, the method for recommending any multimedia information in the first aspect above is implemented.

[0018] The beneficial effects of this application are as follows:

[0019] In the embodiments of this application, for bidders, multiple sample bidding information and corresponding sample bidding results are obtained, and based on this multiple sample bidding information, at least one round of iterative processing is performed on the prediction model to be trained. Each round of iteration includes: for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid; wherein, each piece of sample bidding information includes: sample multimedia information and object information of the sample object, and the prediction model includes: a target prediction network and at least one reference prediction network using different bidding mechanisms. In this way, a prediction network is preset for each bidding mechanism, and based on the matching of the bidding mechanisms, the association relationship between the sample bidding information and the prediction network is constructed. Furthermore, in the iterative training process of the prediction model, the prediction model learns the association relationship and association rules between the input data for the multiple pieces of sample bidding information and their corresponding sample bidding results. Correspondingly, based on the learned association information, the prediction model respectively outputs a set of prediction probabilities for successfully bidding on each piece of sample bidding information based on each predicted bid; in this way, the trained prediction model can adaptively output a relatively accurate set of prediction probabilities, that is, this set of prediction probabilities can more comprehensively, flexibly, and accurately represent the successful bidding probability of the corresponding sample bidding information for different predicted bids. Furthermore, based on the high-precision set of prediction probabilities, a relatively accurate target bid can be obtained, thereby improving the recommendation accuracy for the target multimedia information to be recommended.

[0020] In addition, to further improve the prediction accuracy of the target prediction network in the prediction model, for each round of iteration, after obtaining the set of prediction probabilities corresponding to a piece of sample bidding information, the following operations will also be performed:

[0021] In one case, if the prediction network is a reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities. In other words, when the bidding mechanism corresponding to the sample bidding information matches the bidding mechanism applied by the reference prediction network, the parameters of the corresponding reference prediction network are adjusted based on the corresponding set of prediction probabilities, so that the corresponding reference prediction network can learn more effective information under the corresponding bidding mechanism to improve the prediction accuracy for the corresponding bidding mechanism.

[0022] In another case, if the prediction network is the target prediction network, the parameters of the prediction model are adjusted based on the mapping probability sets of the prediction probability set with respect to the specified reference prediction probability sets respectively. In other words, when the bidding mechanism corresponding to the sample bidding information matches the bidding mechanism applied by the reference prediction network, the specified reference prediction probability sets will also be obtained (for example, the reference prediction probability sets obtained by any reference prediction network other than the target prediction network after processing the corresponding sample bidding information). Understandably, between each obtained reference prediction probability set and the prediction probability set of the target prediction network, different bidding mechanisms are respectively corresponding. Thus, based on the mapping probability sets of the prediction probability set with respect to the specified reference prediction probability sets respectively, the mechanism migration from the bidding mechanisms respectively corresponding to the specified reference prediction probability sets to the bidding mechanism corresponding to this prediction probability set is realized. By combining multiple bidding mechanisms, the parameters of the prediction model are adjusted to improve the prediction accuracy of the target prediction network and the reference prediction network in the prediction model, realize the optimization of the training process of the prediction model, and thus improve the recommendation accuracy for multimedia information based on the prediction model.

[0023] Other features and advantages of the present application will be described in the following specification. And, partly, they will become obvious from the specification, or be understood by implementing the present application. The objectives and other advantages of the present application can be achieved and obtained through the structures specifically pointed out in the written specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0025] Figure 1 It is a schematic diagram of an optional application scenario in an embodiment of the present application;

[0026] Figure 2 It is a schematic flowchart of a method for recommending multimedia information provided by an embodiment of the present application;

[0027] Figures 3A - 3B It is a schematic diagram of a bidding mechanism in an embodiment of the present application;

[0028] Figures 4A - 4B It is a schematic structural diagram of a possible prediction network in an embodiment of the present application;

[0029] Figure 5 It is a schematic structural diagram of a possible feature embedding module in an embodiment of the present application;

[0030] Figure 6 It is a schematic diagram of the mapping process of a possible reference probability set in an embodiment of the present application;

[0031] Figure 7 This is a schematic diagram of the possible process of obtaining the resource overlap interval in the embodiments of the present application;

[0032] Figure 8 This is a schematic diagram of the possible process of obtaining the target bid price in the embodiments of the present application;

[0033] Figure 9 This is an example diagram of the possible process of training the prediction model in the embodiments of the present application;

[0034] Figure 10 This is a schematic structural diagram of the multimedia information recommendation device provided in the embodiments of the present application;

[0035] Figure 11 This is a schematic structural diagram of a computer device provided in the embodiments of the present application. Detailed implementation manners

[0036] In order to make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0037] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be fully or partially implemented by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of the overall module or unit that includes the function of the module or unit.

[0038] In the embodiments of the present application, the processing of the collection, storage, use, processing, transmission, provision, and disclosure of the user's personal information complies with the provisions of relevant laws and regulations and does not violate public order and good customs.

[0039] To facilitate the understanding of the technical solutions provided in the embodiments of the present application, some key terms used in the embodiments of the present application will be explained first.

[0040] The embodiments of the present application relate to artificial intelligence technologies, mainly including computer vision technology, speech technology, natural language processing technology, machine learning technology, and autonomous driving technology in artificial intelligence technologies.

[0041] Artificial Intelligence (AI): It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, enabling the machines to have the functions of perception, reasoning and decision-making.

[0042] Artificial intelligence technology is an interdisciplinary subject, covering 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, pre-trained model technology, operation / interaction systems, mechatronics, etc. Among them, the pre-trained model, also known as the large model or the foundation model, can be widely applied to downstream tasks in various directions of artificial intelligence after fine-tuning. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0043] Computer Vision Technology (CV) Computer vision is a science that studies how to enable machines to "see". Further, it refers to using cameras and computers to replace human eyes for tasks such as target recognition and measurement in machine vision, and further performing graphic processing to make the images processed by the computer more suitable for human eyes to observe or be transmitted to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies, and attempts to build artificial intelligence systems that can obtain information from images or multi-dimensional data. The large model technology has brought important changes to the development of computer vision technology. Pre-trained models in the visual field such as swin-transformer, ViT, V-MOE, and MAE can be quickly and widely applied to downstream specific tasks after fine-tuning. Computer vision technology usually includes technologies such as image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, etc., and also includes common biometric recognition technologies such as face recognition and fingerprint recognition.

[0044] The key technologies of speech technology include automatic speech recognition (ASR), text-to-speech (TTS), and voiceprint recognition technology. Enabling computers to listen, see, speak, and sense is the future development direction of human-computer interaction, and among them, speech has become one of the most promising human-computer interaction methods. Large model technology has brought about a revolution in the development of speech technology. Pre-trained models such as WavLM and UniSpeech that follow the Transformer architecture have strong generalization and versatility and can excellently complete speech processing tasks in various directions.

[0045] Natural language processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can enable effective communication between humans and computers using natural language. Natural language processing involves natural language, that is, the language people use in daily life, and is closely related to linguistics research; at the same time, it involves computer science and mathematics; furthermore, natural language processing, as an important technology for model training in the field of artificial intelligence, pre-trained models have evolved from large language models (LLMs) in the NLP field. After fine-tuning, large language models can be widely applied to downstream tasks. Natural language processing technology usually includes technologies such as text processing, semantic understanding, machine translation, robot question answering, and knowledge graphs.

[0046] Machine learning: It is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning. Pre-trained models are the latest development results of deep learning, integrating the above technologies.

[0047] Autonomous driving technology refers to the vehicle's ability to drive itself without driver operation. It usually includes technologies such as high-precision maps, environmental perception, computer vision, behavior decision-making, path planning, and motion control. Autonomous driving includes multiple development paths such as single-vehicle intelligence, vehicle-road collaboration, and networked cloud control. Autonomous driving technology has broad application prospects. Currently, in addition to the fields of logistics, public transportation, taxis, and intelligent transportation, it will be further developed in the future.

[0048] With the research and progress of artificial intelligence technology, artificial intelligence technology has been studied and applied in multiple fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, driverless, autonomous driving, drones, digital twins, virtual humans, robots, artificial intelligence generated content (AIGC), conversational interaction, intelligent healthcare, intelligent customer service, vehicle networking, autonomous driving, intelligent transportation, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0049] In the embodiments of the present application, artificial intelligence technology is applied to the field of intelligent recommendation, specifically for extracting features from bidding information or bidding requests, performing feature analysis on the extracted features, performing bid price estimation on the estimated probability set of bidding information or bidding requests, etc., so as to train an estimation model to achieve the recommendation of multimedia information. For example, advertisements are recommended through the target bidding mechanism corresponding to the target estimation network to improve the recommendation accuracy.

[0050] To facilitate the understanding of the technical solutions provided in the embodiments of the present application, some key terms used in the embodiments of the present application are explained below.

[0051] Multimedia information: It is a human-computer interactive information communication and dissemination medium that combines two or more media. Media includes text, pictures, sounds, films, etc. In the embodiments of the present application, multimedia information can be content such as advertisements, articles, news, videos, music, etc. The present application mainly takes advertisements as an example for illustration.

[0052] Display opportunity: The value brought by winning the display opportunity of multimedia information and being exposed to the multimedia platform. The evaluation of the display opportunity is related to click-through rate, conversion rate, etc.

[0053] Bidding mechanism: A mechanism in which multiple bidders for multimedia information (such as advertisers) bid for a display opportunity of multimedia information. The highest bidder will win the display opportunity, and all bids for the display opportunity will be sorted from high to low, and the transaction will be made according to the bid at the specified position. Under different bidding mechanisms, the highest bidder wins the display opportunity, and the difference lies in the transaction price for the display opportunity. For example, the bidding mechanism can include: first price auction, second price auction, K-price auction, etc.

[0054] First Price Auction (FPA): A bidding mechanism in which multiple bidders for multimedia information bid for a display opportunity of multimedia information. The highest bidder wins the display opportunity and the transaction is made according to the highest bid, that is, the transaction price is the highest bid.

[0055] Second Price Auction (SPA): A competition mechanism in which multiple bidders for multimedia information bid to compete for a display opportunity for multimedia information. The highest bidder wins the display opportunity and closes the deal at the second-highest bid, that is, the transaction price is the second-highest bid.

[0056] K-Price Auction (KPA): A competition mechanism in which multiple bidders for multimedia information bid to compete for a display opportunity for multimedia information. The highest bidder wins the display opportunity and closes the deal at the K-th (K>2) highest bid, that is, the transaction price is the K-th highest bid.

[0057] It should be noted that in the embodiments of the present application, when the bidding mechanism applied by the target prediction network is the first-price auction, the bidding mechanism applied by the reference prediction network can be the second-price auction or the K-price auction; or, when the bidding mechanism applied by the target prediction network is the second-price auction, the bidding mechanism applied by the reference prediction network can be the first-price auction or the K-price auction, and so on; specifically, it can be set according to actual application requirements, and the embodiments of the present application do not make specific limitations in this regard. For the convenience of those skilled in the art to better understand the inventive concept of the embodiments of the present application, the following mainly elaborates on a target prediction network applied to the first-price auction and a reference prediction network applied to the second-price auction.

[0058] Revenue: The benefit that a bidder for multimedia information (such as an advertiser) can obtain after winning the display opportunity for multimedia information, which can be measured by the difference between the value of the display opportunity and the transaction price.

[0059] Winning price: Also known as the market price, which refers to the highest external bid. When the bid of a bidder for multimedia information (such as an advertiser) is greater than the winner, the bidder can obtain the display opportunity for multimedia information.

[0060] Bid shading: In the first-price auction scenario, while ensuring that the bid is higher than the winning price (i.e., the market price), the bid is reduced as much as possible to obtain more additional revenue.

[0061] Prediction probability: The probability of winning a display opportunity when bidding for a display opportunity.

[0062] Expected revenue: The mathematical expectation of the revenue generated by bidding for a display opportunity once, and its value is equal to the difference between the prediction probability corresponding to the bid, the value of the display opportunity, and the corresponding bid.

[0063] Click-Through-Rate (CTR): It can be used to measure the display effect of multimedia information and can also be called the click-through rate. Taking multimedia information as an advertisement as an example, the click-through rate can be the actual number of clicks on the advertisement divided by the display volume of the advertisement.

[0064] Conversion Rate (CVR): An indicator for measuring the display effect of multimedia information. Taking the multimedia information as an advertisement for example, the conversion rate can be the conversion rate from an object clicking on the advertisement to becoming a valid activation or registration or even a paying object.

[0065] The following briefly introduces the design concept of the embodiments of the present application.

[0066] In the field of advertising bidding, advertisers rely on real-time bidding technology to recommend advertisements. As a conventional advertising bidding mechanism, that is, the first-price auction, when a designated advertiser wants to push an advertisement to a corresponding advertising space based on a designated bid price to recommend it to an object interacting with the advertising space, the designated bid price given by it needs to be greater than the bid prices given by other advertisers to obtain the display opportunity of the advertising space and achieve real-time recommendation of the advertisement.

[0067] In the related method, the designated advertiser fits the historical bid prices of historical advertisements and the corresponding historical bidding situations (such as: bidding success, bidding failure) based on a preset probability distribution function to obtain a historical probability distribution representing the historical bidding success probability based on each historical bid price. Then, based on this historical probability distribution, the target bid price corresponding to the maximum historical probability is adopted to achieve real-time recommendation of the advertisement.

[0068] However, in the above related method, the advertiser needs to preset a probability distribution function to assume the distribution situation of the historical probability distribution. When the distribution situation of the preset probability distribution function does not match the real distribution situation, it is difficult to obtain a reliable historical probability distribution based on the unreasonable preset probability distribution function, resulting in difficulty in achieving real-time recommendation of the advertisement based on the inaccurate historical probability distribution. Secondly, as time goes by, the real distribution situation corresponding to the designated advertiser will also change, and there is usually a deviation between the real distribution situation corresponding to this change and the distribution situation of the preset probability distribution function, that is, this method also has the problems of weak generalization ability and inability to handle dynamic complex scenarios. In addition, the advertising revenue of the advertiser will also become lower as the estimated bid price increases. In an extreme case, when the estimated bid price is infinitely high, the advertiser can stably obtain the display opportunity of the advertisement, but the corresponding advertising revenue is also infinitely low. In this way, the meaning of advertising recommendation is lost. In other words, the estimated bid price obtained by abandoning the advertising revenue is also inaccurate.

[0069] In view of this, the embodiments of the present application provide a method for recommending multimedia information based on artificial intelligence. A prediction network is preset for each bidding mechanism. Based on the matching of the bidding mechanisms, for the target bidding mechanism, the information of other bidding mechanisms is flexibly applied to optimize the training process of the prediction model corresponding to the multimedia information, so as to improve the recommendation accuracy of the multimedia information.

[0070] Specifically, the embodiment of the present application provides a method for constructing a prediction model based on deep learning. Based on multiple pieces of sample bidding information obtained, at least one round of iterative processing is performed on the prediction model to be trained. The prediction model includes a target prediction network and at least one reference prediction network applied to different bidding mechanisms. Among them, in each round of iteration, for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully bidding the sample bidding information based on each predicted bid. Each sample bidding result corresponds to a bidding mechanism. In this way, combining the idea of deep learning, a prediction model including prediction networks applied to different bidding mechanisms is constructed, and based on the bidding mechanisms corresponding to each sample bidding information, the prediction model can at least establish an association between different sample bidding information and the prediction network based on the matching prediction network, so that the prediction model can output a set of prediction probabilities for the corresponding bidding information. Compared with the related methods mentioned above, there is no need to assume a preset probability distribution function, but the prediction network inside the prediction model is used to learn information, effectively improving the generalization ability. In addition, for the prediction model with multiple prediction networks applied to different bidding mechanisms, and based on the bidding mechanism corresponding to the sample bidding information, the corresponding prediction network is selected for probability prediction, which can adapt to different bidding mechanisms (such as: first-price auction, second-price auction, K-price auction), and improve the generalization ability under different bidding mechanisms.

[0071] In addition, in the embodiment of the present application, a method for training a prediction model based on mechanism transfer is also provided. For each iteration process, after obtaining a set of prediction probabilities corresponding to a piece of sample bidding information, in one case, if the corresponding prediction network is a reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities to specifically improve the prediction accuracy under the bidding mechanism corresponding to the reference prediction network. In another case, if the corresponding prediction network is a target prediction network, the parameters of the prediction model are adjusted based on the mapping probability sets of the set of prediction probabilities for each specified reference prediction probability set, that is, combining the bidding mechanisms of the reference prediction networks corresponding to each reference prediction probability set, and performing mechanism transfer for the bidding mechanism corresponding to the target prediction network to adjust the parameters of the prediction model, so as to adaptively improve the prediction accuracy under multiple bidding mechanisms corresponding to the prediction model, realize the optimization of the training process of the prediction model, and thus improve the recommendation accuracy for multimedia information based on the prediction model.

[0072] Next, a simple introduction is made to the application scenarios applicable to the technical solutions of the embodiments of the present application. It should be noted that the application scenarios described below are only used to illustrate the embodiments of the present application rather than to limit them. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.

[0073] Such as Figure 1As shown, it is a schematic diagram of the application scenario of the embodiment of the present application. The application scenario diagram includes a terminal device 110 and a server 120.

[0074] In the embodiment of the present application, the terminal device 110 includes, but is not limited to, devices such as mobile phones, tablet computers, laptop computers, desktop computers, e-book readers, intelligent voice interaction devices, intelligent household appliances, vehicle-mounted terminals, and intelligent terminals; a bidding client for recommending multimedia information can be installed on the terminal device, and the client can be software (such as browsers, chat software, short video software, news information software, etc.), or a web page, a small program, etc. The server 120 is a background server corresponding to the software, web page, small program, etc. client, or a server dedicated to multimedia information recommendation, and the present application does not make specific limitations. The server 120 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network, CDN), and big data and artificial intelligence platforms.

[0075] It should be noted that the method for recommending multimedia information in each embodiment of the present application can be executed by an electronic device, and the electronic device can be the terminal device 110 or the server 120, that is, the method can be executed alone by the terminal device 110 or the server 120, or can be jointly executed by the terminal device 110 and the server 120. For example, when jointly executed by the terminal device 110 and the server 120, first, the server 120 performs iterative processing on the pre-estimation model to be trained based on multiple sample bidding information, obtains the trained pre-estimation model and deploys it. Furthermore, the terminal device 110 sends a bidding request containing target bidding information to the server 120, the server 120 receives the target bidding information, uses the trained pre-estimation model, obtains the target bidding result of the target bidding information, that is, based on each pre-estimated bid, the set of pre-estimated probabilities of successfully recommending the target multimedia information in the target bidding information to the target object in the target bidding information, and selects the pre-estimated bid corresponding to the pre-estimated probability that meets the preset recommendation condition as the target bid, and uses the target bid to recommend the target multimedia information.

[0076] Taking multimedia information as an example of an advertisement, for instance, a designated advertiser hopes to push an advertisement for purchasing books to the advertisement slot on the home page of a chat software. The purpose of recommending this advertisement is to promote the books and increase the purchase volume of the books. It is easy to understand that in order to achieve the corresponding purpose, the designated advertiser needs to adopt a corresponding bid to compete for the display opportunity of this advertisement slot; and there will be many advertisers interested in this advertisement slot, and the actual bidding mechanisms selected for this advertisement slot (such as: first-price auction, second-price auction, K-price auction) are also diverse. Then the designated advertiser needs to estimate a relatively accurate bid in order to increase the probability of obtaining the display opportunity of this advertisement slot, so as to achieve precise recommendation of the advertisement.

[0077] Therefore, in the embodiments of the present application, by constructing a target estimation network and at least one reference estimation network applied to different bidding mechanisms, and based on multiple sample bidding information, at least one round of iterative processing is performed on the estimation model to be trained to solve the above problems.

[0078] Specifically, each round of iteration is as follows: For each sample bidding information, the server 120, based on the bidding mechanism (such as: first-price auction, second-price auction, K-price auction) associated with the corresponding sample bidding result (such as: winning the bid based on the sample bid, losing the bid based on the sample bid), uses a matching estimation network to obtain a set of estimated probabilities of winning the bid for the sample bidding information based on each estimated bid: If the estimation network is a reference estimation network, the parameters of the reference estimation network are adjusted based on the set of estimated probabilities; if the estimation network is a target estimation network, the parameters of the estimation model are adjusted based on the set of mapping probabilities of the set of estimated probabilities for each designated set of reference estimated probabilities.

[0079] In an alternative embodiment, the terminal device 110 and the server 120 can communicate through a communication network.

[0080] In an alternative embodiment, the communication network is a wired network or a wireless network.

[0081] It should be noted that Figure 1 The above is only an example. In fact, the number of terminal devices and servers is not limited and is not specifically defined in the embodiments of the present application.

[0082] In the embodiments of the present application, when the number of servers is multiple, the multiple servers can form a blockchain, and the server is a node on the blockchain; such as the recommendation method of multimedia information disclosed in the embodiments of the present application, including the sample bidding information, sample bidding result, estimation model, bidding mechanism, set of estimated probabilities corresponding to each sample bidding information, etc.

[0083] In addition, the usage scenarios to which the embodiments of the present application can be applied include, but are not limited to, scenarios such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving.

[0084] Next, in combination with the above-described application scenarios, the method for recommending multimedia information provided by the exemplary embodiments of the present application will be described with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown for the convenience of understanding the spirit and principle of the present application, and the embodiments of the present application are not limited in this regard.

[0085] Refer to Figure 2 As shown, it is a flowchart of the implementation of a method for recommending multimedia information provided by an embodiment of the present application. Taking the server as the execution entity as an example, the specific implementation process of this method is as follows: steps 201 to 202:

[0086] Step 201: Obtain multiple pieces of sample bidding information and corresponding sample bidding results; each piece of sample bidding information includes: sample multimedia information and object information of the sample object.

[0087] Among them, the sample multimedia information refers to any type of information stream, such as: advertisements, articles, news, videos, music, etc.

[0088] Taking advertisements as an example below, the same principle applies to other multimedia information, and the text will not be repeated.

[0089] In the embodiments of the present application, the object information of the sample object in the sample bidding information is: the object information of the object interacting with the sample multimedia information. In addition, the sample bidding information may also include the object information of other objects, such as: the bidder of the corresponding sample bidding information, the release object of the corresponding sample bidding information, the production object of the corresponding sample bidding information. Among them, the bidder may be the object for which the advertisement is to be recommended (such as an advertiser, a bidder, etc.), the release object may be the object that releases the advertisement (such as an advertisement publisher, etc.), and the production object may be the object that produces the advertisement (such as an advertisement producer, etc.); in addition, the foregoing objects may be any one and combination of users, individuals, and organizations, or any corresponding devices (such as: terminal devices, servers). Here, the inventor has found through practice that generally, the sample bidding information includes the object information and multimedia information of the sample object and the release object, which can achieve better results.

[0090] Optionally, the sample bidding information further includes display information that can be used to display the sample bidding information, such as advertisement position information (such as: the position on the web page for displaying advertisements, etc.).

[0091] Each sample bidding information corresponds to a sample bidding result. A sample bidding result includes: for a corresponding piece of sample bidding information, based on the sample bid, the sample probability of successfully recommending the sample multimedia information therein to the corresponding sample object (i.e., the sample probability of winning the bid for this piece of sample bidding information); in other words, there is a corresponding relationship between the sample bidding result and the sample bidding information. The sample bidding result at least includes: based on a bidding mechanism, the sample bid given for the corresponding sample bidding information, and the sample probability of winning the bid based on this sample bid. Among them, the sample bidding result can be an advertisement bidding log (such as the sample probability of winning the bid corresponding to the sample bid of a historical advertisement, the sample probability of losing the bid corresponding to the sample bid of a historical advertisement), and the sample probability can be set to 1 (winning the bid) or 0 (losing the bid) according to actual application requirements.

[0092] In one implementation, multiple pieces of sample bidding information and their corresponding sample bidding results can be obtained through multiple historical direct or indirect interactions between the bidder and the terminal device. For example, the bidder has historically entered information bidding instructions carrying bidding requirements into the terminal device multiple times, and then the terminal device responds to each information bidding instruction to obtain the corresponding sample bidding information and the corresponding sample bidding result; for another example, the bidder has historically sent information bidding instructions carrying bidding requirements to the terminal device multiple times through other communication devices connected to the terminal device, and then the terminal device responds to each information bidding instruction to obtain the corresponding sample bidding information and the corresponding sample bidding result. In addition, the terminal device can also directly obtain multiple pieces of sample bidding information and their corresponding sample bidding results from the database.

[0093] It should be noted that the multiple pieces of sample bidding information and their corresponding sample bidding results obtained can correspond to the same bidder or different bidders. The embodiments of the present application do not make specific limitations on this. The following mainly takes the case of corresponding to the same bidder as an example.

[0094] In the embodiments of the present application, each sample bidding result is obtained by applying a bidding mechanism to bid on the corresponding sample bidding information; the bidding mechanism can include: a target bidding mechanism and at least one reference bidding mechanism. For example, taking advertisement bidding as an example, a target bidding mechanism can be a first-price auction, that is, the bid of the bidder is higher than the bids of other bidders, and this bid is the transaction price corresponding to winning the bid; a target bidding mechanism can be a second-price auction, that is, the bid of the bidder is higher than the bids of other bidders, and the second-highest bid is the transaction price corresponding to winning the bid; a reference bidding mechanism can be a K (K>2)-price auction, that is, the bid of the bidder is higher than the bids of other bidders, and the Kth bid ranked from high to low among all the bids is used as the transaction price corresponding to winning the bid.

[0095] SeeFigure 3A As shown in the figure, it is a schematic diagram of an auction mechanism. Taking the first-price auction as an example, n bidders respectively adopt different bids to recommend corresponding multimedia information; the bids corresponding to the n bidders are sorted in descending order, and the sorting result is: Bid 1 > Bid 2 >... > Bid n. Then, the bidder corresponding to the highest (largest) Bid 1 is the winning bidder, and the highest (largest) Bid 1 is used as the transaction price (i.e., the transaction price) for Bidder 1 to successfully bid for Multimedia Information 1.

[0096] See Figure 3B As shown in the figure, it is a schematic diagram of an auction mechanism. Taking the K-price auction mechanism as an example, n bidders respectively adopt different bids to recommend corresponding multimedia information; the bids corresponding to the n bidders are sorted in descending order, and the sorting result is: Bid 1 > Bid 2 >... > Bid k >... > Bid n. Then, the bidder corresponding to the highest (largest) Bid 1 is the winning bidder, and the Kth (the Kth largest) Bid K is used as the transaction price for Bidder 1 to successfully bid for Multimedia Information 1.

[0097] It should be noted that the same principle applies to other auction mechanisms, and the text will not repeat it here.

[0098] Step 202: Perform at least one round of iterative processing on the prediction model to be trained based on multiple sample bidding information; among them, the prediction model includes: a target prediction network applied to different auction mechanisms and at least one reference prediction network, and each round of iteration includes: Steps 2021 to 2023.

[0099] It should be noted that the following takes the target prediction network applying the target auction mechanism of the first-price auction and at least one reference prediction network applying the reference auction mechanisms of the second-price auction or the K (K > 2)-price auction as an example to summarize the design idea of performing at least one round of iterative processing on the prediction model to be trained provided by the present application as follows.

[0100] Specifically, the auction mechanism can be set according to different actual recommendation scenarios. Therefore, for the same bidder, in different set scenarios, multimedia information is usually recommended based on different auction mechanisms. That is, there is a certain similarity among the bidders corresponding to different auction mechanisms. Correspondingly, the same publishing object, the same production object (or the same auction platform) can also recommend multimedia information based on different auction mechanisms. It is easy to understand that there are commonalities between the target auction mechanism and the reference auction mechanisms.

[0101] For example, in an advertising bidding scenario, first-price auctions and second-price auctions share a similar group of users (i.e., the objects interacting with the advertisements), advertisers (i.e., bidders), and bidding environments, forming the prerequisite for the transfer effectiveness from second-price auctions to first-price auctions. From the perspective of the user group, based on the inventor's creative labor in analyzing actual data, the overlap rate of objects under different bidding mechanisms is approximately 20% to 30%, which provides a good foundation for the transfer. From the perspective of advertisers, based on the inventor's creative labor in analyzing actual data, the overlap rate of these advertisers exceeds 80%, which provides a relatively good foundation for the transfer. Further, the overlap of advertisers also results in some similar advertisements between the environments corresponding to first-price auctions and second-price auctions, thus providing a good foundation for the migration. From the perspective of the bidding environment, for a specific advertising alliance or Internet advertising DSP (Demand-Side Platform), the bidding environments, including the bidding opponents and bidding preferences, such as those in first-price auctions and second-price auctions, are similar, which also provides a good foundation for the migration. Thus, on the one hand, under a first-price auction, for example, the bidders encountered by the designated bidders in the embodiments of the present application are almost the same as those encountered under a second-price auction. On the other hand, the fluctuations in the bidding environment, such as the preferences of other bidders except the designated bidders, that is, the opportunity to obtain an advertisement display. Therefore, through the inventor's creative labor, it is found that in the embodiments of the present application, there is a commonality between the target bidding mechanism and the reference bidding mechanism, and this commonality provides the basic conditions for the mechanism migration between different bidding mechanisms.

[0102] Further, based on the iterative idea outlined above, the following provides a detailed description of one round of the iterative process in combination with steps 2021 to 2023 below.

[0103] Step 2021: For each sample bidding information, based on the bidding mechanism corresponding to the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid.

[0104] Among them, the bidding mechanism includes a target bidding mechanism and at least one reference bidding mechanism. The prediction network can be a target prediction network or a reference prediction network. The target prediction network corresponds to the target bidding mechanism, and the reference prediction network corresponds to the reference bidding mechanism. The target bidding mechanism is the bidding mechanism applicable to the prediction model in the application stage, and the reference bidding mechanism is the bidding mechanism used for auxiliary training of the prediction model in the training stage.

[0105] In one embodiment, a matching prediction network is adopted, and for each sample bidding information, the following operations are respectively performed: Based on the sample bidding result corresponding to a piece of sample bidding information, a prediction network matching the bidding mechanism of the sample bidding result is obtained, and then the feature extraction sub-network in the prediction network is used to obtain the sample bidding features of the sample bidding information, and based on the prediction parameters in the prediction network, feature mapping processing is performed on the sample bidding features to obtain a set of prediction probabilities of successfully bidding for the sample bidding information based on each predicted bid price.

[0106] See Figure 4A As shown, it is a schematic structural diagram of a possible prediction network in an embodiment of the present application. Taking the prediction network as the target prediction network or a reference prediction network as an example, in the prediction network, first, the feature extraction sub-network is used to obtain the sample bidding features g(r i ) of the sample bidding information, and then based on the prediction parameters in the prediction network, feature mapping processing is performed on the sample bidding information to obtain a set of prediction probabilities of successfully bidding for the sample bidding information based on each predicted bid price.

[0107] Among them, as Figure 4A shown, the prediction sub-network is the network layer in the above prediction network, which is used to perform feature analysis processing on the sample bidding features to obtain corresponding feature processing results, and the Softmax layer is used to perform normalization processing on the feature processing results to obtain corresponding normalization processing results, that is, the set of prediction probabilities. In the embodiment of the present application, the set of prediction probabilities can be characterized as a discrete probability distribution, and of course, it can also be characterized as a continuous probability distribution. Here, the discrete probability distribution is taken as an example for the following description.

[0108] More specifically, in the feature extraction sub-network in the above prediction network, a feature embedding module can be used to extract the deep space features of the sample bidding information to obtain corresponding sample initial features. Based on this, the feature extraction sub-network in the prediction network is further used to perform weight mapping processing on the sample initial features based on the weight mapping parameters in the feature extraction sub-network to obtain corresponding feature mapping weights, and based on the feature extraction parameters in the feature extraction sub-network, perform feature interaction processing on the sample initial features to obtain corresponding sample interaction features. Then, based on the feature mapping weights and the sample interaction features, the sample bidding features of the sample bidding information are obtained.

[0109] Among them, the above weight mapping parameters represent: in the historical round of iteration, the positive influence of each feature mapping weight on the corresponding sample bidding result; the above feature extraction parameters represent: in the historical round of iteration, the positive influence of each sample interaction feature on the corresponding sample bidding result.

[0110] See Figure 4BAs shown in the figure, it is a schematic structural diagram of a possible prediction network in an embodiment of the present application. Taking the prediction network as the target prediction network or a reference prediction network as an example, for the feature extraction sub-network in the prediction network, the sample bidding information is input into the feature embedding module to extract deep spatial features, and the sample initial feature r output by the feature embedding module is obtained. i , r i is respectively input into the expert cluster and the gating unit G. The expert cluster corresponds to the corresponding bidding mechanism and contains n sub-expert networks (such as: E f_1 , …, E f_n ). Each sub-expert network respectively focuses on different feature dimensions in the sample initial feature, so that on the corresponding feature dimensions, based on the corresponding feature extraction parameters, some effective information in the sample initial feature is strengthened. The output results of the n sub-expert networks are selectively spliced to obtain the sample interaction feature E(r i ), and then E(r i ) is input into the gating unit G. The gating unit G can be a single-layer feedforward network with a Softmax activation function. Then, through the single-layer feedforward network with the Softmax activation function (combined with the linear layer), using the weight mapping parameter, the weight mapping process is performed on the sample initial feature r i (and the effective expert weights of the expert cluster are extracted), and the corresponding feature mapping weight w(r i ) is obtained. Then, the weighted sum of the feature mapping weight w(r i ) and the sample interaction feature E(r i ) is calculated to obtain the sample bidding feature g(r i ) of the sample bidding information.

[0111] Exemplarily, for the sample initial feature r i : Taking the target prediction network as an example, the feature extraction idea of the expert cluster in the target prediction network can be characterized as: Correspondingly, taking the reference prediction network as an example, the feature extraction idea of the expert cluster in the reference prediction network can be characterized as:

[0112] Among them, n s and n f are respectively the numbers of sub-expert networks in the expert clusters of the target prediction network and the reference prediction network, d is the dimension of the input representation, and Es and Ef are matrices composed of all selected vectors of the expert clusters in the target prediction network and the reference prediction network respectively.

[0113] Exemplarily, for the sample initial feature r i : Taking the target prediction network as an example, the feature mapping extraction idea in the target prediction network can be characterized as: w f (ri ) = Softmax(W f r i ); Correspondingly, taking the reference prediction network as an example, the feature extraction idea of the expert cluster in the reference prediction network can be characterized as: w s (r i ) = Softmax(W s r i ).

[0114] Among them, and are the weight mapping parameters (i.e., parameter matrices) in the target prediction network and the reference prediction network respectively.

[0115] Exemplarily, for the sample initial feature r i : Taking the target prediction network as an example, the feature selection idea of the gating unit in the target prediction network can be characterized as: g f (r i ) = w f (r i )E f (r i ); Correspondingly, taking the reference prediction network as an example, the feature selection idea of the gating unit in the reference prediction network can be characterized as: g s (r i ) = w s (r i )Es(r i ).

[0116] Among them, g s (r i ) and g f (r i ) are the sample bidding features output by the gating unit G in the target prediction network and the reference prediction network respectively.

[0117] It can be seen that in the embodiments of the present application, by applying the decentralized mixture of experts (DMoE) structure, different prediction networks contain different expert clusters. In other words, different bidding tasks (such as bidding tasks corresponding to different bidding mechanisms) correspond to different expert clusters in the prediction model. Since the sample bidding information may come from the bidding environments of different bidding mechanisms (such as first-price auction, second-price auction, K-price auction), each expert cluster can be used to reduce the negative transfer of the sample bidding information from different bidding environments.

[0118] Furthermore, in the above feature embedding module, the corresponding sample initial features are obtained by extracting the deep spatial features of the feature embedding module.

[0119] See Figure 5As shown in the figure, it is a schematic structural diagram of a possible feature embedding module in an embodiment of the present application. Taking the target prediction network or a reference prediction network as an example of the prediction network, for the feature extraction sub-network in the prediction network, sample bidding information is input into the feature embedding module. The sample bidding information may include: sample media features corresponding to sample multimedia information (such as: advertisement features), sample bidding features corresponding to the object information of the sample object (such as: advertisement position features, object features of advertisement position-related objects, etc.), and sample publisher features corresponding to the sample publisher object (such as: publisher features). Specifically, using the shared embedding layer as shown in Figure 5 For the i-th sample bidding information x among multiple sample bidding information, as shown in i , the sample bidding information x i (x i ∈x) can be embedded into a low-dimensional dense vector representation as: h ij = embed(x ij ) ∈ R d , where d is the dimension of the embedding vector.

[0120] Then, using the shared base layer as shown in Figure 5 , the output features of the shared embedding layer are concatenated for the embedding vectors, which is expressed as: Here, the concatenated embedding vectors will pass through the shared bottom layer and be output as: r i = f shared_bottom (v i ), where r i is used to represent the network output of v i passing through the shared bottom layer, and is used as the sample initial feature r i initially extracted by the shared module for multiple sample bidding information. Of course, in the embodiment of the present application, mainly the sample initial feature corresponding to a single sample bidding information is taken as an example for detailed description, and the case of multiple splicings can be obtained in the same way and will not be elaborated here.

[0121] In summary, for a single prediction network, it can be used to obtain a set of prediction probabilities corresponding to sample bidding information that matches the bidding mechanism.

[0122] Step 2022: If the prediction network is a reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities.

[0123] Among them, the reference prediction network is an auxiliary prediction network, and the sample bidding result includes: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid.

[0124] In the embodiment of the present application, if the prediction network is a reference prediction network, the parameters of the reference prediction network are adjusted based on the degree of approximation of the set of prediction probabilities to the sample bidding result.

[0125] In a possible implementation, if the set of estimated probabilities is characterized as a discrete probability distribution, a probability density function (PDF) can be used to convert it into a PDF form, such as a continuous probability distribution. Specifically, the process of the above parameter adjustment can include the following two cases.

[0126] Case 1: When the sample bidding result is a successful bid, based on the positive approach degree of the estimated probability corresponding to the sample bid price in the set of estimated probabilities relative to the sample probability, a reference loss value is calculated, and the reference prediction network is adjusted with this reference loss value.

[0127] Specifically, for the sample bidding information that adopts the reference bidding mechanism and has a successful bid, it can be considered that the sample bid price in the corresponding sample bidding result is relatively accurate. Then, in the obtained sets of estimated probabilities, the estimated probability consistent with the sample bid price is selected, and the corresponding sample probability is maximized to obtain the corresponding reference loss value.

[0128] Optionally, the above reference loss value can be obtained by using the cross-entropy method, and the sample bidding information with a successful bid can be considered that its corresponding sample bid price is uncensored.

[0129] Exemplarily, in the scenario of adopting the reference bidding mechanism and having a successful bid, the positive approach degree of the above estimated probability relative to the sample probability can be characterized as:

[0130]

[0131] Where is the reference loss value corresponding to the scenario of adopting the reference bidding mechanism and having a successful bid, (x k , z k ) represents the kth sample bidding information x k and its corresponding sample bid price z k , and p z (z k | x k ) represents the conditional probability that for the given kth sample bidding information x k , its estimated bid price is the sample bid price z k .

[0132] Case 2: When the sample bidding result is a failed bid, based on the negative approach degree of the estimated probabilities corresponding to at least one estimated bid price greater than the sample bid price in the set of estimated probabilities relative to the sample probability, a reference loss value is calculated, and the reference prediction network is adjusted with this loss value.

[0133] Specifically, for the sample bidding information that adopts the reference bidding mechanism and fails in the bidding, it can be considered that the sample bid in the corresponding sample bidding result is relatively small. Then, among the obtained estimated probability sets, select the estimated probabilities greater than the sample bid, and maximize their respective corresponding sample probabilities to obtain the corresponding reference loss value, that is: in the PDF distribution corresponding to the estimated probability set, maximize the cumulative distribution function (CDF) of the right half of the sample bid.

[0134] Optionally, the above reference loss value can be obtained by using the cross-entropy method, and for the sample bidding information that succeeds in the bidding, it can be considered that the corresponding sample bid has a right censoring.

[0135] Exemplarily, in the scenario of adopting the reference bidding mechanism and failing in the bidding, the degree of reverse approach of the above estimated probability relative to the sample probability can be characterized as:

[0136]

[0137] Among them, is the reference loss value corresponding to the scenario of adopting the reference bidding mechanism and failing in the bidding. (x k , b k ) represents the k-th sample bidding information x k and its corresponding sample bid b k , p z represents the probability density function of the sample transaction price (such as: transaction price, market price, etc.), W(b k |x k ) represents the estimated probability of successful bidding recommended for the k-th sample bidding information x k with the sample bid b k , s k represents the sample bidding result of losing the bid corresponding to the k-th sample bidding information, which can be specifically implemented by an event indicator (if the bidding fails, s k is 1, if the bidding succeeds, s k is 0).

[0138] Combining the above two described situations, the parameter adjustment of the reference prediction network based on the estimated probability set can be characterized as:

[0139] Among them, is the reference loss value corresponding to the reference bidding mechanism, is the reference loss value corresponding to the scenario of adopting the reference bidding mechanism and succeeding in the bidding, is the reference loss value corresponding to the scenario of adopting the reference bidding mechanism and failing in the bidding.

[0140] It should be noted that in the embodiments of the present application, when performing step 2022, the parameter adjustment of the reference prediction network can be performed once based on multiple sample bidding information, or the parameter adjustment of the reference prediction network can be performed once based on each sample bidding information, which can be set according to the actual situation here.

[0141] Step 2023: If the prediction network is the target prediction network, then based on each specified reference prediction probability set for the mapping probability set of the prediction probability set, the parameter adjustment of the prediction model is performed.

[0142] Among them, the target prediction network is the prediction network that obtains the target bid during the application process, and the sample bidding result includes: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid.

[0143] In the embodiments of the present application, if the prediction network is the target prediction network, then at least one specified reference prediction probability set is obtained, and then based on the difference between each obtained reference prediction probability set's respective mapping probability set for the prediction probability set and the prediction probability set, the parameter adjustment of the prediction model is performed. Among them, each reference prediction probability set is: obtained by using a specified reference prediction network and successfully bidding on the sample bidding information based on each prediction bid.

[0144] It should be noted that the above-mentioned one reference prediction network can be specified according to the actual application requirements; for example, in the advertising bidding mechanism, if the target bidding mechanism adopted by the target prediction network is the first-price auction, then a reference prediction network adopting the second-price auction as the reference bidding mechanism can be specified.

[0145] In a specific implementation manner, based on the prediction probability set, the mapping process is respectively performed on each obtained reference prediction probability set to obtain the mapping probability set of each reference prediction probability, and according to the set difference between each mapping probability set and the prediction probability set, the resource coincidence interval of each set difference is obtained at each prediction bid, and then, under the constraint of the resource coincidence interval, based on the degree of approximation of the prediction probability set to the sample bidding result, the parameter adjustment of the prediction model is performed.

[0146] See Figure 6 As shown, it is a schematic diagram of the mapping process of a possible reference probability set in the embodiments of the present application. Taking one reference prediction probability set and one prediction probability set as an example, it can be seen that the mapping process is performed on one reference prediction probability set to the prediction probability set to obtain the mapping probability set of one reference prediction probability set and the set difference between the mapping probability set and the prediction probability set.

[0147] Optionally, to further improve the accuracy of obtaining the mapping probability set, for each obtained reference estimated probability, the following operations are respectively performed: obtain the estimated continuous distribution corresponding to the estimated probability set, and the reference continuous distribution corresponding to a reference probability set, and then perform a distribution translation process on the reference continuous distribution based on the distribution distance between the reference continuous distribution and the estimated continuous distribution to obtain the corresponding mapping probability set.

[0148] The above process realizes the distribution mapping from the reference estimated probability to the estimated probability. The mapping idea comes from the fact that different estimation networks adopt different bidding mechanisms, and there are some commonalities in different bidding mechanisms. To improve the recommendation accuracy of the target estimation network, therefore, through the mapping method, the bidding mechanisms of other reference estimation networks are combined. For example, under the advertising bidding mechanism, there are differences between the first-price auction and the second-price auction. If we want to ensure the estimation accuracy of the target estimation network corresponding to the first-price auction, we learn the effective information in the second-price auction. Here, through the mapping method, the deviation between the second-price auction and the first-price auction is minimized as much as possible to obtain more effective information.

[0149] Specifically, taking a reference estimated probability set and an estimated probability set as an example, the estimated continuous distribution corresponding to the estimated probability set is denoted as p f , and the reference continuous distribution corresponding to the reference probability set is denoted as p s . Here, each estimated resource range in p s is moved to the vicinity of each estimated resource range in p f to obtain the corresponding mapping probability set p′f.

[0150] Exemplarily, the mapping probability set p′f can be obtained by calculating the Wasserstein (Earth Mover's Distance) distance between the two distributions to minimize the distance between the two distributions. Here, the Wasserstein distance between the above two distributions p f and p s can be characterized as

[0151]

[0152] where WD(p f , p s ) is used to represent the distribution distance between p f and p s , and Γ(p f , p s ) represents the joint distribution between p f and p s .

[0153] Further, after performing a distribution translation process on the reference continuous distribution based on the distribution distance between the reference continuous distribution and the estimated continuous distribution to obtain the corresponding mapping probability set, in order to obtain a more accurate resource overlap interval, an embodiment of the present application proposes a heuristic method for obtaining the resource overlap interval.

[0154] See Figure 7 As shown, it is a schematic diagram of a possible process for obtaining a resource overlap interval in an embodiment of the present application. Taking a reference estimation probability set and an estimation probability set as an example, the estimated continuous distribution corresponding to the estimation probability set is denoted as p f , and the mapped continuous distribution of the mapped probability set corresponding to the reference probability set is denoted as p'f. Perform an intersection operation on the estimated continuous distribution p f and the mapped continuous distribution p'f. The result of the intersection operation can be referred to Figure 7 the lower half part connected by the dashed line in, and obtain the part of the intersection result above the threshold ω. Based on the left endpoint value δ lc and the right endpoint value δ rc in this part, obtain the resource overlap interval formed by the left endpoint and the right endpoint.

[0155] In a possible implementation manner, based on the resource overlap interval, parameter adjustment of the estimation model may include the following two cases.

[0156] Case 1, when the sample bidding result is a successful bid, parameter adjustment of the estimation model is performed based on the positive approaching degree of at least one estimated probability in the estimated probability set that satisfies the first preset condition with respect to the sample probability. Among them, the first screening condition includes: less than the sample bid price and not less than the left endpoint of the resource overlap interval.

[0157] Specifically, for the sample bidding information that adopts the target bidding mechanism and has a successful bid, it can be considered that the sample bid price in the corresponding sample bidding result is relatively large. Then, under the constraint of the left endpoint of the resource overlap interval, select the estimated probabilities less than the sample bid price from the obtained estimated probability sets, and maximize their corresponding sample probabilities to obtain the corresponding target loss value.

[0158] Optionally, the above target loss value can be obtained by using the cross-entropy method, and the sample bidding information with a successful bid can be considered to have a left censoring of its corresponding sample bid price.

[0159] Exemplarily, in the scenario of adopting the target bidding mechanism and having a successful bid, the positive approaching degree of the above estimated probability with respect to the sample probability can be characterized as:

[0160]

[0161] where, (x k ,bk ) represents the k-th sample bidding information x k and its corresponding sample bid b k , p z represents the probability density function of the sample transaction price (i.e., market price, etc.), p i represents the probability density function that the estimated bid corresponding to index i is the sample transaction price, b represents any estimated bid in the interval (δ lc , b k ), i represents the specific bin index of different estimated bids b, and the estimated bid δ represented by the left endpoint value of the resource overlap interval lc corresponding index is sample bid b k corresponding index is

[0162] Case 2, when the sample bidding result is a bidding failure, based on the concentration of the estimated probabilities, for at least one estimated probability that satisfies the second preset condition, the degree of reverse approach to the sample probability is used to adjust the parameters of the estimation model; among them, the second screening condition includes: greater than the sample bid and not greater than the right endpoint of the resource overlap interval.

[0163] Specifically, for the sample bidding information that adopts the target bidding mechanism and has a bidding failure, it can be considered that the sample bid in the corresponding sample bidding result is relatively small. Then, under the constraint of the left endpoint of the resource overlap interval, from the obtained estimated probability sets, select the estimated probabilities greater than the sample bid, and maximize their corresponding sample probabilities to obtain the corresponding target loss value.

[0164] Optionally, the above target loss value can be obtained by using the cross-entropy method, and for the sample bidding information with a successful bid, it can be considered that the corresponding sample bid has a right censoring.

[0165] Exemplarily, in the scenario of adopting the target bidding mechanism and having a bidding failure, the degree of reverse approach of the above estimated probability to the sample probability can be characterized as:

[0166]

[0167] Among them, (x k , b k ) represents the k-th sample bidding information x k and its corresponding sample bid b k , p z represents the probability density function of the sample transaction price (such as: transaction price, market price, etc.), p i represents the probability density function that the estimated bid corresponding to index i is the market price, b represents the interval (b k , δ rc)Any estimated bid on, where i represents the specific binning index for different estimated bids b, and the estimated bid δ represented by the right endpoint value of the resource overlap interval rc The corresponding index is Sample bid b k The corresponding index is

[0168] Combining the above two situations, the target loss value corresponding to the target bidding mechanism can be represented as:

[0169] Among them, is the target loss value corresponding to adopting the target bidding mechanism, is the target loss value corresponding to adopting the target bidding mechanism and in the case of successful bidding, is the target loss value corresponding to adopting the target bidding mechanism and in the case of failed bidding.

[0170] Therefore, based on step 2023 and combined with step 2022, the model loss value used to adjust the parameters of the prediction model can be obtained, which can be represented as:

[0171] Among them, is the model loss value used to adjust the parameters of the prediction model, is the target loss value corresponding to adopting the target bidding mechanism, is the reference loss value corresponding to adopting the target bidding mechanism, α f and α s are the weights set for the target bidding mechanism and the reference bidding mechanism respectively.

[0172] It should be noted that in the embodiments of the present application, when performing step 2023, the parameters of the prediction model can be adjusted once based on multiple sample bidding information, or the parameters of the prediction model can be adjusted once based on each sample bidding information, which can be set according to the actual situation here.

[0173] Optionally, after performing iterative processing on a specified threshold for the prediction model, or when the model loss value meets a preset condition, a trained prediction model is obtained. This prediction model can be used to obtain a relatively accurate set of prediction probabilities for the target bidding information to be recommended, that is, to obtain the prediction probabilities of the target bidding information for successful bidding at different estimated bids. Based on this, a relatively accurate interaction rate (such as: yield rate) can be obtained. Here, a bidding masking algorithm can be used to obtain the target bid with the maximum interaction rate, and then the target bid is used to achieve precise recommendation for the target multimedia information in the target bidding information.

[0174] See Figure 8As shown in the figure, it is a schematic diagram of a possible process for obtaining a target bid price. Using a trained prediction model, based on the target prediction network therein, a target bid price result of target bid price information is obtained. Among them, the target bid price information includes: target multimedia information and a target object, and the target bid price result includes: a set of predicted probabilities of successfully bidding for the target bid price information based on each predicted bid price. Then, based on the target bid price result, the predicted bid price corresponding to the predicted probability that meets the preset recommendation condition is selected as the target bid price, and then the target multimedia information is recommended to the target object using the target bid price.

[0175] Specifically, in the process of obtaining the target bid price result of the target bid price information, the target prediction network in the trained prediction model is used. Based on the feature extraction sub-network in the target prediction network, the target bid price feature of the target bid price information is obtained. Then, the prediction sub-network and the Softmax layer in the target prediction network are used to perform feature mapping processing on the target bid price feature based on the prediction parameters in the target prediction network, and a set of predicted probabilities of successfully bidding for the target bid price information based on each predicted bid price is obtained as the corresponding target bid price result. Then, the target bid price for pushing the target multimedia information is selected from it to improve the recommendation accuracy of the target multimedia information.

[0176] Based on the above embodiments, an example in the field of advertising bidding is used below to illustrate the training process of the prediction model in the embodiments of the present application. Refer to Figure 9 As shown in the figure, it is an example diagram of the process of training the prediction model in the embodiments of the present application, specifically including:

[0177] It should be noted that here, taking the combination of second-price auctions to improve the accuracy of the set of predicted probabilities in first-price auctions as an example for detailed description; and the method of the embodiments of the present application can also be to combine first-price auctions to improve the accuracy of the set of predicted probabilities in second-price auctions, or to combine multiple bidding mechanisms to improve the accuracy of the set of predicted probabilities in first-price auctions or any bidding mechanism. The embodiments of the present application do not make specific limitations on this.

[0178] Specifically, first, multiple pieces of sample bid price information and corresponding sample bid price results are obtained. Each piece of sample bid price information includes: feature information such as sample advertisements, sample ad spaces, sample users, sample publishers, etc., and the sample results include: sample bid prices and probabilities of successful bidding. And the multiple pieces of sample bid price information are used as the model input of the prediction model to be trained.

[0179] Then, in each training process of the prediction model, a (first-price & second-price) shared embedding layer is used. For the i-th sample bid price information x among multiple pieces of sample bid price information i , the sample bid price information x i (x i ∈x) can be embedded into a low-dimensional dense vector representation as: hij = embed(x ij ) ∈ R d , where d is the dimension of the embedding vector. Then, a shared base layer is used to concatenate the embedding vectors of the output features of the (monovalent & divalent) shared embedding layer, denoted as: Here, the concatenated embedding vectors will pass through the shared base layer and be output as: r i = f shared_bottom (v i ), where r i is used to represent the network output of v i passing through the shared base layer, as the initial sample feature r i initially extracted by the shared module from multiple sample bidding messages.

[0180] Secondly, the initial sample feature r i is respectively input into the corresponding expert clusters (such as: monovalent expert cluster, divalent expert cluster), and of course, it can also be synchronously input into the monovalent expert cluster and the divalent expert cluster. For example, using the monovalent expert cluster can obtain Then, the corresponding gating unit G is used to obtain the sample bidding feature g f (r i ) = w f (r i ) E f (r i ), where w f (r i ) is the weight mapping parameter, w f (r i ) = Softmax(W f r i ). Again, for example, using the divalent expert cluster can obtain Then, the corresponding gating unit G is used to obtain the sample bidding feature g s (r i ) = w s (r i ) E s (r i ), where w s (r i ) is the weight mapping parameter, w s (r i ) = Softmax(W s r i ). The embodiments of this application do not specifically limit the number of selected expert clusters.

[0181] Further, after obtaining the sample bidding features, the first-price network and the softmax layer in the prediction sub-network are used to perform feature mapping on the sample bidding features, and a set of predicted probabilities of successful bidding for the sample bidding information based on each predicted bid is obtained. This set of predicted probabilities is characterized as a discrete distribution, ensuring that the predicted bids cover a limited precision, and further improving the prediction accuracy and training efficiency.

[0182] Again, for the reference prediction network (second-price auction) corresponding to the auxiliary task, a corresponding set of reference predicted probabilities is obtained; in the scenario of successful bidding (no censoring in the second-price auction), the cross-entropy method is used to obtain the corresponding reference loss value based on the set of reference predicted probabilities: Among them, (x k , z k ) represents the k-th sample bidding information x k and its corresponding sample bid z k , and p z (z k |x k ) represents the conditional probability that for the given k-th sample bidding information x k , its predicted bid is the sample bid z k ; in the scenario of failed bidding (right censoring in the second-price auction), the corresponding reference loss value is obtained based on the set of reference predicted probabilities: Among them, (x k , b k ) represents the k-th sample bidding information x k and its corresponding sample bid b k , and p z represents the probability density function of the sample transaction price (such as: transaction price, market price, etc.), and W(b k |x k ) represents the predicted probability of successful bidding for the k-th sample bidding information x k recommended at the sample bid b k , and s k represents the sample bidding result of losing the bid corresponding to the k-th sample bidding information, which can be specifically implemented using an event indicator (if the bidding fails, then s k is 1, if the bidding is successful, then s k is 0). Combining the above two situations, the parameter adjustment of the reference prediction network based on the set of reference predicted probabilities can be characterized as:

[0183] Finally, for the target prediction network corresponding to the main task (first-price auction), a corresponding set of target prediction probabilities is obtained. Additionally, a set of reference prediction probabilities output by the reference prediction network corresponding to the auxiliary task (second-price auction) is obtained. The predicted continuous distribution corresponding to the prediction probability set is denoted as p f , and the reference continuous distribution corresponding to the reference probability set is denoted as p s . Here, by calculating the Wasserstein distance between the two distributions p f and p s , it can be characterized in the metric space (X, d) as where WD(p f , p s ) is used to represent the distribution distance between p f and p s , Γ(p f , p s ) represents the joint distribution between p f and p s . Move each predicted resource range in p s to the vicinity of each predicted resource range in p f to obtain the corresponding mapped probability set p′f. Then, perform an intersection operation on the predicted continuous distribution p f and the mapped continuous distribution p′f, and obtain the part of the intersection result above the threshold ω. Based on the left endpoint value δ lc and the right endpoint value δ rc in this part, obtain the resource overlap interval formed by the left and right endpoints.

[0184] Furthermore, in the scenario of successful bidding (left censoring in first-price auction), using the cross-entropy method, based on the target prediction probability set, obtain the corresponding target loss value:

[0185] where (x k , b k ) represents the k-th sample bidding information x k and its corresponding sample bid b k , p z represents the probability density function of the sample transaction price (such as: transaction price, market price, etc.), p i represents the probability density function that the predicted bid corresponding to index i is the sample transaction price, b represents any predicted bid in the interval (δ lc , b k ), i represents the specific bin index of different predicted bids b, and the predicted bid δ lc characterized by the left endpoint value of the resource overlap interval corresponds to the index sample bid bk The corresponding index is In the scenario of auction failure (right censored under the first-price auction), using the cross-entropy method, based on the target estimated probability set, obtain the corresponding target loss value:

[0186] where (x k , b k ) represents the k-th sample bidding information x k and its corresponding sample bid b k , p z represents the probability density function of the sample transaction price (such as: transaction price, market price, etc.), p i represents the probability density function that the estimated bid corresponding to index i is the market price, b represents any estimated bid in the interval (b k , δ rc ), i represents the specific bin index of different estimated bids b, and the estimated bid δ characterized by the right endpoint value of the resource overlap interval rc The corresponding index is The index corresponding to the sample bid b k is Combining the above two situations, the target loss value corresponding to the first-price bidding mechanism can be characterized as:

[0187] Furthermore, under the first-price auction, combining the target loss value and the reference loss value, obtain the model loss value used to adjust the parameters of the prediction model, which can be characterized as:

[0188] In summary, the embodiments of the present application provide a method for constructing a prediction model based on deep learning. Based on multiple acquired sample bidding information, at least one round of iterative processing is performed on the prediction model to be trained. The prediction model includes a target prediction network and at least one reference prediction network applied to different bidding mechanisms. Among them, in each round of iteration, for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, a matching prediction network is used to obtain a set of prediction probabilities for successfully bidding on the sample bidding information based on each predicted bid. In this way, combining the idea of deep learning, a prediction model including prediction networks applied to different bidding mechanisms is constructed, and based on the bidding mechanisms corresponding to each sample bidding information, the prediction model can at least output a corresponding set of prediction probabilities based on the matching prediction network. Compared with the related methods mentioned above, there is no need to assume a preset probability distribution function, but the prediction network inside the prediction model is used to learn information, effectively improving the generalization ability. In addition, for the prediction model, multiple prediction networks applied to different bidding mechanisms are set, and based on the bidding mechanism corresponding to the sample bidding information, the corresponding prediction network is selected for probability prediction, which can adapt to different bidding mechanisms (such as different advertising bidding mechanisms) and improve the generalization ability under different bidding mechanisms.

[0189] In addition, in the embodiments of the present application, a method for training a prediction model based on mechanism transfer is also provided. For each iteration process, after obtaining a set of prediction probabilities corresponding to a piece of sample bidding information, in one case, if the corresponding prediction network is a reference prediction network, the parameters of the reference prediction network are adjusted based on the set of prediction probabilities to specifically improve the prediction accuracy under the bidding mechanism corresponding to the reference prediction network. In another case, if the corresponding prediction network is a target prediction network, the parameters of the prediction model are adjusted based on the mapping probability sets of the set of prediction probabilities for each specified reference probability set, that is, combining the bidding mechanisms of the reference prediction networks corresponding to each reference probability set, the mechanism transfer of the bidding mechanism corresponding to the target prediction network is performed to adjust the parameters of the prediction model, so as to adaptively improve the prediction accuracy of the prediction model under multiple bidding mechanisms, realize the optimization of the training process of the prediction model, and thus improve the recommendation accuracy of multimedia information based on the prediction model.

[0190] See Figure 10 As described above, based on the same inventive concept, the embodiments of the present application also provide a multimedia information recommendation device, which includes:

[0191] An acquisition module 1001, configured to acquire multiple pieces of sample bidding information and corresponding sample bidding results; each piece of sample bidding information includes: sample multimedia information and object information of the sample object;

[0192] A processing module 1002 is configured to perform at least one round of iterative processing on a to-be-trained prediction model based on the multiple sample bidding information; wherein, the prediction model includes: a target prediction network applied to different bidding mechanisms and at least one reference prediction network. Each round of iteration includes: for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, using a matching prediction network to obtain a set of prediction probabilities for successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid; if the prediction network is the reference prediction network, parameter adjustment is performed on the reference prediction network based on the set of prediction probabilities; if the prediction network is the target prediction network, parameter adjustment is performed on the prediction model based on the set of mapping probabilities of the set of prediction probabilities with respect to each specified set of reference prediction probabilities.

[0193] Optionally, the processing module 1002 is specifically configured to:

[0194] If the prediction network is a reference prediction network, parameter adjustment is performed on the reference prediction network based on the degree of approximation of the set of prediction probabilities to the sample bidding result.

[0195] Optionally, the sample bidding result includes: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid.

[0196] Then, the processing module 1002 is configured to perform parameter adjustment on the reference prediction network based on the degree of approximation of the set of prediction probabilities to the sample bidding result, and is specifically configured to:

[0197] When the sample bidding result is a successful bid, parameter adjustment is performed on the reference prediction network based on the positive degree of approximation of the prediction probability corresponding to the sample bid in the set of prediction probabilities to the sample probability.

[0198] When the sample bidding result is a failed bid, parameter adjustment is performed on the reference prediction network based on the negative degree of approximation of the prediction probabilities corresponding to at least one predicted bid greater than the sample bid in the set of prediction probabilities to the sample probability.

[0199] Optionally, the processing module 1002 is specifically configured to:

[0200] If the prediction network is the target prediction network, at least one specified set of reference prediction probabilities is obtained; wherein, each set of reference prediction probabilities is obtained by a reference prediction network using a specified bidding mechanism and successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid.

[0201] Based on the differences between the mapping probability sets of each obtained reference prediction probability set with respect to the prediction probability set and the prediction probability set, the parameters of the prediction model are adjusted.

[0202] Optionally, the processing module 1002 is configured to adjust the parameters of the prediction model based on the differences between the mapping probability sets of each obtained reference prediction probability set with respect to the prediction probability set and the prediction probability set respectively, and specifically:

[0203] Based on the prediction probability set, perform mapping processing on each obtained reference prediction probability set respectively to obtain the mapping probability set of each reference prediction probability;

[0204] According to the set differences between each mapping probability set and the prediction probability set respectively, obtain the resource coincidence intervals of each set difference at each predicted bid price;

[0205] Under the constraint of the resource coincidence interval, adjust the parameters of the prediction model based on the degree of approximation of the prediction probability set to the sample bidding result.

[0206] Optionally, the processing module 1002 is configured to perform mapping processing on each obtained reference prediction probability set respectively based on the prediction probability set to obtain the mapping probability set of each reference prediction probability, and specifically:

[0207] For each obtained reference prediction probability, perform the following operations respectively:

[0208] Obtain the predicted continuous distribution corresponding to the prediction probability set and the reference continuous distribution corresponding to the reference probability set;

[0209] Based on the distribution distance of the reference continuous distribution with respect to the predicted continuous distribution, perform distribution translation processing on the reference continuous distribution to obtain the corresponding mapping probability set.

[0210] Optionally, the sample bidding result includes: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid price;

[0211] Then the processing module 1002 is configured to adjust the parameters of the prediction model based on the degree of approximation of the prediction probability set to the sample bidding result under the constraint of the resource coincidence interval, and specifically:

[0212] When the sample bidding result is a successful bid, based on at least one estimated probability in the estimated probability set that satisfies the first preset condition, with respect to the positive approach degree of the sample probability, the parameters of the estimation model are adjusted; wherein, the first screening condition includes: being less than the sample bid price and not less than the left endpoint of the resource overlap interval.

[0213] When the sample bidding result is a failed bid, based on at least one estimated probability in the estimated probability set that satisfies the second preset condition, with respect to the reverse approach degree of the sample probability, the parameters of the estimation model are adjusted; wherein, the second screening condition includes: being greater than the sample bid price and not greater than the right endpoint of the resource overlap interval.

[0214] Optionally, the processing module 1002 is specifically configured to:

[0215] For each sample bidding information, the following operations are respectively performed:

[0216] Based on the sample bidding result corresponding to the sample bidding information, an estimation network matching the bidding mechanism of the sample bidding result is obtained;

[0217] Using the feature extraction sub-network in the estimation network, the sample bidding feature of the sample bidding information is obtained;

[0218] Based on the estimation parameters in the estimation network, feature mapping processing is performed on the sample bidding feature, and an estimated probability set for successfully bidding the sample bidding information based on each estimated bid price is obtained.

[0219] Optionally, the processing module 1002 is used to obtain the sample bidding feature of the sample bidding information by using the feature extraction sub-network in the estimation network, and is specifically configured to:

[0220] Extract the deep space feature of the sample bidding information to obtain the corresponding sample initial feature;

[0221] Using the feature extraction sub-network in the estimation network, based on the weight mapping parameters in the feature extraction sub-network, weight mapping processing is performed on the sample initial feature to obtain the corresponding feature mapping weight; wherein, the weight mapping parameters represent: in the historical round of iteration, the positive influence of each feature mapping weight on the corresponding sample bidding result; and,

[0222] Based on the feature extraction parameters in the feature extraction sub-network, feature interaction processing is performed on the sample initial feature to obtain the corresponding sample interaction feature; wherein, the feature extraction parameters represent: in the historical round of iteration, the positive influence of each sample interaction feature on the corresponding sample bidding result;

[0223] Based on the feature mapping weights and the sample interaction features, obtain the sample bidding features of the sample bidding information.

[0224] Optionally, after performing at least one round of iterative processing on the to-be-trained prediction model based on the multiple sample bidding information, the processing module 1002 further includes:

[0225] In response to receiving a bidding request carrying target bidding information, use the target prediction network in the trained prediction model to obtain a set of predicted probabilities of successfully recommending the target multimedia information in the target bidding information to the target object in the target bidding information based on each predicted bid;

[0226] Select, from the set of predicted probabilities, the predicted bid corresponding to the predicted probability that meets the preset recommendation condition as the target bid;

[0227] Use the target bid to recommend the target multimedia information to the target object.

[0228] Optionally, the processing module 1002 is configured to use the target prediction network in the trained prediction model to obtain a set of predicted probabilities of successfully recommending the target multimedia information in the target bidding information to the target object in the target bidding information based on each predicted bid, and specifically is configured to:

[0229] Use the target prediction network in the trained prediction model to extract corresponding target bidding features for the target bidding information based on the feature extraction sub-network in the target prediction network; wherein, the target bidding information at least includes: target multimedia information and object information of the target object;

[0230] Based on the prediction parameters in the target prediction network, perform feature mapping processing on the target bidding features to obtain a set of predicted probabilities of successfully recommending the target multimedia information to the target object based on each predicted bid.

[0231] This device can be used to execute the methods shown in the embodiments of the present application. Therefore, for the functions that can be realized by each functional module of this device, reference can be made to the description of the foregoing embodiments, and details are not repeated here.

[0232] Please refer to Figure 11 As shown, based on the same technical concept, an embodiment of the present application further provides a computer device 1100, and this computer device 1100 can be Figure 1 the terminal device or server shown, and this computer device 1100 may include a memory 1101 and a processor 1102.

[0233] The memory 1101 is used to store computer programs executed by the processor 1102. The memory 1101 mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function, etc.; the data storage area can store data created according to the use of the computer device, etc. The processor 1102 can be a central processing unit (CPU), or a digital processing unit, etc. In the embodiments of the present application, the specific connection medium between the memory 1101 and the processor 1102 is not limited. In the embodiments of the present application Figure 11 it is connected between the memory 1101 and the processor 1102 through a bus 1103. The bus 1103 is represented by a thick line in Figure 11 For the connection manners between other components, only schematic illustrations are provided and are not to be taken as limitations. The bus 1103 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 11 in it only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0234] The memory 1101 can be a volatile memory, such as a random-access memory (RAM); the memory 1101 can also be a non-volatile memory, such as a read-only memory, a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD), or the memory 1101 is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 1101 can be a combination of the above memories.

[0235] The processor 1102 is used to execute the methods performed by the devices in the embodiments of the present application when calling the computer programs stored in the memory 1101.

[0236] In some possible implementation manners, various aspects of the method provided in the present application can also be implemented in the form of a program product, which includes program code. When the program product runs on a computer device, the program code is used to cause the computer device to execute the steps in the methods according to various exemplary implementation manners of the present application described above in this specification. For example, the computer device can execute the methods performed by the devices in the embodiments of the present application.

[0237] The program product may employ any combination of one or more readable media. The readable media may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0238] Although the preferred embodiments of the present application have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present application.

[0239] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for recommending multimedia information, characterized in that, Including: Obtaining multiple pieces of sample bidding information and corresponding sample bidding results; each piece of sample bidding information at least includes: sample multimedia information and object information of a sample object; Based on the multiple pieces of sample bidding information, performing at least one round of iterative processing on a to-be-trained prediction model; wherein, the prediction model includes: a target prediction network applied to different bidding mechanisms and at least one reference prediction network, and each round of iteration includes: For each piece of sample bidding information, based on the bidding mechanism associated with the corresponding sample bidding result, using a matching prediction network to obtain a set of prediction probabilities of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid; If the prediction network is the reference prediction network, adjusting the parameters of the reference prediction network based on the set of prediction probabilities; If the prediction network is the target prediction network, adjusting the parameters of the prediction model based on the set of mapped probabilities of the set of prediction probabilities with respect to each specified set of reference prediction probabilities.

2. The method according to claim 1, characterized in that, The "if the prediction network is the reference prediction network, adjusting the parameters of the reference prediction network based on the set of prediction probabilities" includes: If the prediction network is the reference prediction network, adjusting the parameters of the reference prediction network based on the degree of approximation of the set of prediction probabilities to the sample bidding result.

3. The method according to claim 2, characterized in that, The sample bidding result includes: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid; Then the "adjusting the parameters of the reference prediction network based on the degree of approximation of the set of prediction probabilities to the sample bidding result" includes: When the sample bidding result is a successful bid, adjusting the parameters of the reference prediction network based on the positive degree of approximation of the prediction probability corresponding to the sample bid in the set of prediction probabilities to the sample probability; When the sample bidding result is a failed bid, adjusting the parameters of the reference prediction network based on the reverse degree of approximation of the prediction probabilities corresponding to at least one predicted bid greater than the sample bid in the set of prediction probabilities to the sample probability.

4. The method according to claim 1, characterized in that, The "if the prediction network is the target prediction network, adjusting the parameters of the prediction model based on the set of mapped probabilities of the set of prediction probabilities with respect to each specified set of reference prediction probabilities" includes: If the prediction network is the target prediction network, obtaining at least one specified set of reference prediction probabilities; wherein, each set of reference prediction probabilities is obtained by a reference prediction network adopting a specified bidding mechanism and successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid; Adjusting the parameters of the prediction model based on the difference between the set of mapped probabilities of each obtained set of reference prediction probabilities with respect to the set of prediction probabilities and the set of prediction probabilities.

5. The method according to claim 4, wherein The "adjusting the parameters of the prediction model based on the differences between the set of mapped probabilities of each obtained set of reference prediction probabilities with respect to the set of prediction probabilities and the set of prediction probabilities respectively" includes: Based on the set of estimated probabilities, perform mapping processing on each obtained reference set of estimated probabilities respectively to obtain the set of mapped probabilities for each respective reference estimated probability; According to the set differences between the sets of mapped probabilities and the set of estimated probabilities respectively, obtain the resource overlap intervals of the set differences at each estimated bid price; Under the constraint of the resource overlap interval, based on the degree of approximation of the set of estimated probabilities to the sample bidding results, adjust the parameters of the estimation model.

6. The method according to claim 5, wherein The performing mapping processing on each obtained reference set of estimated probabilities respectively based on the set of estimated probabilities to obtain the set of mapped probabilities for each respective reference estimated probability includes: For each obtained reference estimated probability, perform the following operations respectively: Obtain the estimated continuous distribution corresponding to the set of estimated probabilities and the reference continuous distribution corresponding to the reference set of probabilities; Based on the distribution distance of the reference continuous distribution relative to the estimated continuous distribution, perform distribution translation processing on the reference continuous distribution to obtain the corresponding set of mapped probabilities.

7. The method according to claim 5, characterized in that, The sample bidding results include: the sample probability of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on the sample bid price; Then the adjusting the parameters of the estimation model based on the degree of approximation of the set of estimated probabilities to the sample bidding results under the constraint of the resource overlap interval includes: When the sample bidding result is a successful bid, adjust the parameters of the estimation model based on the positive degree of approximation of at least one estimated probability in the set of estimated probabilities that satisfies the first preset condition to the sample probability; wherein, the first screening condition includes: being less than the sample bid price and not less than the left endpoint of the resource overlap interval; When the sample bidding result is a failed bid, adjust the parameters of the estimation model based on the negative degree of approximation of at least one estimated probability in the set of estimated probabilities that satisfies the second preset condition to the sample probability; wherein, the second screening condition includes: being greater than the sample bid price and not greater than the right endpoint of the resource overlap interval.

8. The method according to any one of claims 1 to 7, characterized in that, The obtaining the set of estimated probabilities of successfully recommending the corresponding sample multimedia information based on each estimated bid price by using a matching estimation network based on the bidding mechanism associated with the corresponding sample bidding result for each sample bidding information includes: For each sample bidding information, perform the following operations respectively: Based on the sample bidding result corresponding to the sample bidding information, obtain an estimation network that matches the bidding mechanism of the sample bidding result; Use the feature extraction sub-network in the estimation network to obtain the sample bidding feature of the sample bidding information; Based on the estimation parameters in the estimation network, perform feature mapping processing on the sample bidding feature to obtain the set of estimated probabilities of successfully bidding on the sample bidding information based on each estimated bid price.

9. The method according to claim 8, wherein The using the feature extraction sub-network in the estimation network to obtain the sample bidding feature of the sample bidding information includes: Extract the deep spatial feature of the sample bidding information to obtain the corresponding sample initial feature; Using the feature extraction sub-network in the prediction network, based on the weight mapping parameters in the feature extraction sub-network, perform weight mapping processing on the initial sample features to obtain corresponding feature mapping weights; wherein, the weight mapping parameters represent: in the historical round of iteration, the positive influence of each feature mapping weight on the corresponding sample bidding result; and, Based on the feature extraction parameters in the feature extraction sub-network, perform feature interaction processing on the initial sample features to obtain corresponding sample interaction features; wherein, the feature extraction parameters represent: in the historical round of iteration, the positive influence of each sample interaction feature on the corresponding sample bidding result; Based on the feature mapping weights and the sample interaction features, obtain the sample bidding features of the sample bidding information.

10. The method according to any one of claims 1 to 7, characterized in that, After performing at least one round of iteration processing on the prediction model to be trained based on the multiple sample bidding information, it further includes: In response to receiving a bidding request carrying target bidding information, use the target prediction network in the trained prediction model to obtain a set of prediction probabilities for successfully recommending the target multimedia information in the target bidding information to the target object in the target bidding information based on each predicted bid; Select, from the set of prediction probabilities, the predicted bid corresponding to the prediction probability that meets the preset recommendation condition as the target bid; Use the target bid to recommend the target multimedia information to the target object.

11. The method according to claim 10, wherein The step of using the target prediction network in the trained prediction model to obtain a set of prediction probabilities for successfully recommending the target multimedia information in the target bidding information to the target object in the target bidding information based on each predicted bid includes: Use the target prediction network in the trained prediction model, based on the feature extraction sub-network in the target prediction network, extract corresponding target bidding features for the target bidding information; wherein, the target bidding information at least includes: target multimedia information and object information of the target object; Based on the prediction parameters in the target prediction network, perform feature mapping processing on the target bidding features to obtain a set of prediction probabilities for successfully recommending the target multimedia information to the target object based on each predicted bid.

12. A recommendation device for multimedia information, characterized in that, It includes: An acquisition module, configured to acquire multiple sample bidding information and corresponding sample bidding results; each sample bidding information at least includes: sample multimedia information and object information of the sample object; A processing module, configured to perform at least one round of iterative processing on a to-be-trained prediction model based on the multiple pieces of sample bidding information; wherein, the prediction model includes: a target prediction network applied to different bidding mechanisms and at least one reference prediction network, and each round of iteration includes: for each piece of sample bidding information, based on the bidding mechanism of the corresponding sample bidding result, using a matching prediction network to obtain a set of prediction probabilities of successfully recommending the corresponding sample multimedia information to the corresponding sample object based on each predicted bid; if the prediction network is the reference prediction network, parameter adjustment is performed on the reference prediction network based on the set of prediction probabilities; if the prediction network is the target prediction network, parameter adjustment is performed on the prediction model based on the mapping probability sets of the set of prediction probabilities with respect to each specified reference prediction probability set.

13. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.

14. A computer storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. A computer program product, comprising computer program instructions, characterized in that, When the computer program instructions are executed by the processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images