Media content recommendation method, apparatus and device, and computer storage medium
By training the content recommendation model and adjusting the repeated recommendation weight, the problem of repeated exposure of media content in the recommendation system is solved, improving recommendation accuracy and user experience.
Patent Information
- Application Number
- CN202310517396.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-09
- Publication Date
- 2024-11-12
AI Technical Summary
There is a problem of repeated exposure of media content in existing recommendation systems, resulting in low recommendation accuracy and lack of effective solutions.
By training the content recommendation model, the interaction operation probability is estimated based on the characteristics of the target object and the characteristics of the media content, and the model loss value is adjusted through the repeated recommendation weight to reduce the probability deviation of the media content interaction of repeated exposure, thereby improving recommendation accuracy.
It effectively reduces the repeated exposure of media content, improves the accuracy and user experience of media content recommendations, and optimizes the learning effect of content recommendation models.
Smart Images

Figure CN118939861A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of media content recommendation, and provides a media content recommendation method, device, equipment and computer storage medium. Background Art
[0002] With the development of Internet technology, a variety of content services can be provided through the Internet. Correspondingly, service providers can recommend content that may be of interest through recommendation systems, such as recommending products, games, and coupons in the form of multimedia content. Therefore, how to make accurate recommendations is a problem that recommendation systems pay more attention to.
[0003] When the recommendation system receives a content recommendation request triggered by a target object, it selects the most suitable media content from the media content library and returns it to the target object for display exposure at the corresponding display position.
[0004] At present, there is repeated exposure in the media content that is finally displayed and exposed. Specifically, the same or similar media content is exposed to the same target object multiple times within a certain period of time. Excessive repeated exposure will not only reduce the novelty of the displayed media content, but also cause content reading troubles to the target object, and waste the display opportunities of other media content.
[0005] However, there is no good solution for the situation of repeated exposure, so the recommendation accuracy of current recommendation systems is still low. Summary of the invention
[0006] The embodiments of the present application provide a media content recommendation method, apparatus, device and computer storage medium for improving the repeated exposure of media content and improving the accuracy of media content recommendation.
[0007] On the one hand, a media content recommendation method is provided, the method comprising:
[0008] In response to a content recommendation request sent by a target object, obtaining a plurality of candidate media contents to be recommended;
[0009] Obtaining, by means of a trained content recommendation model, respectively estimated probabilities of the target object performing interactive operations on the multiple candidate media contents based on the object features of the target object and the content features corresponding to each of the multiple candidate media contents;
[0010] Determine target media content from the plurality of candidate media content based on the obtained plurality of probability estimates, and recommend the target media content to a target object;
[0011] Among them, the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample. Each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object in the corresponding training sample. Each interaction probability deviation represents: the deviation between the estimated probability of the sample object performing an interactive operation on the sample media content in the corresponding training sample and the true value of the probability.
[0012] In one aspect, a media content recommendation device is provided, the device comprising:
[0013] A content acquisition unit, configured to obtain a plurality of candidate media contents to be recommended in response to a content recommendation request sent by a target object;
[0014] A probability estimation unit, configured to obtain, through a trained content recommendation model, probability estimation values of the target object performing an interactive operation on the multiple candidate media contents based on the object features of the target object and the content features corresponding to the multiple candidate media contents;
[0015] A recommendation unit, configured to determine target media content from the plurality of candidate media content based on the obtained plurality of probability estimates, and recommend the target media content to a target object;
[0016] Among them, the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample. Each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object in the corresponding training sample. Each interaction probability deviation represents: the deviation between the estimated probability of the sample object performing an interactive operation on the sample media content in the corresponding training sample and the true value of the probability.
[0017] On the one hand, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the above methods when executing the computer program.
[0018] In one aspect, a computer storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.
[0019] In one aspect, a computer program product is provided, the computer program product comprising a computer program, the computer program being stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device performs the steps of any of the above methods.
[0020] In an embodiment of the present application, when training a content recommendation model, the repeated recommendation weight of each training sample is obtained according to the repetition degree of the sample media content in each training sample being recommended to the sample object, and the model loss value of the content recommendation model is determined based on the repeated recommendation weight and interaction probability deviation corresponding to each training sample, that is, for each training sample, when calculating the sample loss value brought by the sample to the entire model, it will be based on the respective repeated recommendation weight, and the higher the repetition degree, the greater the repeated recommendation weight, which is equivalent to increasing the sample loss value corresponding to the training sample with a higher degree of repetition, and increasing the proportion of such training samples in the overall training samples, so that the training process has a stronger learning degree for such training samples, and the content recommendation model finally obtained can reduce the interaction probability deviation for media content with a higher degree of repetition, and improve the probability estimate of such media content, so as to achieve the purpose of improving the repeated exposure of media content and improving the accuracy of media content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0022] Figure 1a and Figure 1b The parameter changes as the number of repeated exposures increases according to the embodiment of the present application;
[0023] Figure 2a and Figure 2b A schematic diagram of an application scenario provided for an embodiment of the present application;
[0024] Figure 3 A schematic diagram of the training process of the content recommendation model provided in the embodiment of the present application;
[0025] Figure 4 A schematic diagram of a process for determining a repeated recommendation weight provided in an embodiment of the present application;
[0026] Figure 5a to Figure 5d A diagram showing examples of repeated recommendation weights provided in an embodiment of the present application;
[0027] Figure 6 A schematic diagram of another process for determining the weight of repeated recommendations provided in an embodiment of the present application;
[0028] Figure 7 A flowchart of a media content recommendation method provided in an embodiment of the present application;
[0029] Figure 8 A schematic diagram of a structure of a content recommendation model provided in an embodiment of the present application;
[0030] Fig. 9 A schematic diagram of the structure of a media content recommendation device provided in an embodiment of the present application;
[0031] Fig.10 A schematic diagram of the structure of a computer device provided in an embodiment of the present application;
[0032] Fig.11 A schematic diagram of the composition structure of another computer device to which an embodiment of the present application is applied. DETAILED DESCRIPTION
[0033] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below in conjunction with the drawings in the embodiment of the present application. Obviously, the described embodiment is only a part of the embodiment of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other arbitrarily. In addition, although the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in an order different from that here.
[0034] It is understandable that in the following specific implementations of the present application, data related to the target object, such as interactive behavior data, is involved. When the various embodiments of the present application are applied to specific products or technologies, relevant licenses or consents need to be obtained, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions. For example, when it is necessary to obtain relevant data, relevant volunteers can be recruited and relevant agreements for volunteer authorization data can be signed, and then the data of these volunteers can be used for implementation; or, by implementing within the scope of the authorized organization, the following implementation methods are implemented by using the data of internal members of the organization to make relevant recommendations to internal members; or, the relevant data used in the specific implementation are all simulated data, such as simulated data generated in a virtual scene.
[0035] To facilitate 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 here:
[0036] Media content: also known as multimedia information, refers to information that users can interact with, such as advertisements, web pages, etc. Taking advertisements as an example, advertisements can be displayed through web pages or advertisement display positions in the application interface. Users can click on the advertisement to enter the corresponding page and perform corresponding operations, such as downloading the application or placing an order.
[0037] Each media content is composed of one or more content materials, and different media content can be generated using the same content material in different content forms. In addition, media content can be associated with a recommendation, and different media content can be associated with the same recommendation, for example, advertisement A and advertisement B are both used to promote the same product.
[0038] Interactive operation: refers to certain operations performed by users on media content. For example, when the media content is an advertisement, the interactive operation can be users clicking on the advertisement, users downloading the application in the advertisement, users purchasing the goods in the advertisement, or users activating the products in the advertisement.
[0039] Probability estimate: represents the probability of a user interacting with the media content after the media content is recommended to the user as predicted by the content recommendation model. Taking advertising as an example, the probability estimate can also be called the predicted click-through rate (pCTR). The problem it solves is to predict the probability of a user clicking on a specific ad in a specific ad position under specific circumstances, thereby assisting in the decision-making of whether to recommend the ad to the user and the specific ad to recommend to the user.
[0040] Interaction probability deviation: refers to the deviation between the actual value of the probability that the user will interact with the media content after the media content is recommended to the user, and the estimated probability value predicted by the content recommendation model. The interaction probability deviation can characterize the accuracy of the content recommendation model. The smaller the interaction probability deviation, the smaller the difference between the estimated probability value predicted by the content recommendation model and the actual probability value, and thus the higher the prediction accuracy of the content recommendation model.
[0041] The embodiments of the present application involve machine learning (ML) technology, which is mainly designed based on machine learning in artificial intelligence.
[0042] Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.
[0043] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0044] Machine learning is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.
[0045] Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies. Artificial neural networks (ANN) abstract the human brain neural network from the perspective of information processing, establish a simple model, and form different networks according to different connection methods. A neural network is a computational model composed of a large number of nodes (or neurons) connected to each other. Each node represents a specific output function, called an activation function. The connection between each two nodes represents a weighted value for the signal passing through the connection, called a weight, which is equivalent to the memory of the artificial neural network. The output of the network varies depending on the network connection method, weight value and activation function. The network itself is usually an approximation of a certain algorithm or function in nature, or it may be an expression of a logical strategy.
[0046] The embodiments of the present application involve the prediction of the probability estimate of interactive operations on media content in the form of advertisements, so as to recommend media content to the target object based on the probability estimate of each media content. When predicting the probability estimate of media content, it is necessary to adopt a content recommendation model based on deep learning, that is, the embodiment of the present application adopts a machine learning method to obtain a content recommendation model for predicting the probability estimate of interactive operations on media content. The content recommendation model is based on machine learning to achieve the ability to process and understand the media content currently suitable for recommendation. Taking advertisements as an example, the probability estimate of each candidate advertisement currently being recommended to the target object can be comprehensively evaluated based on the object characteristics of the target object, the advertisement characteristics of each candidate advertisement, the advertisement position characteristics, and the traffic characteristics, etc., expressing the possibility of the target object performing interactive operations on each current candidate advertisement, thereby measuring the suitability of each candidate advertisement currently being recommended to the target object based on this, and finally recommending media content accordingly.
[0047] Specifically, the prediction of the probability estimate value in the embodiment of the present application can be divided into two parts, including a training part and an application part. Among them, the training part involves the technical field of machine learning. In the training part, the artificial neural network model (i.e., the content recommendation model mentioned later) is trained through the technology of machine learning, so that the artificial neural network model is trained based on the training sample data such as the object characteristics, content characteristics, content display position characteristics, and traffic characteristics given in the embodiment of the present application, and the model parameters are continuously adjusted through the optimization algorithm until the model converges; the application part is used to use the artificial neural network model trained in the training part to predict the probability estimate value of each media content when the content recommendation is triggered during actual use, so as to rank the recommendation degree of each individual content. In addition, it should be noted that the artificial neural network model in the embodiment of the present application can be either online training or offline training, which is not specifically limited here. In this article, offline training is used as an example for illustration.
[0048] In the related art, the media content that is finally displayed and exposed is repeatedly exposed, which is specifically manifested as multiple exposures of the same or similar media content to the same target object within a certain time range. Excessive repeated exposure will not only reduce the novelty of the displayed media content, but also cause content reading troubles to the target object, and at the same time waste the display opportunities of other media content.
[0049] Although there is currently a method of sending the number of repeated exposures as a freshness feature into the model for training in the hope of reducing the estimation bias to alleviate the phenomenon of repeated exposure, see Figure 1a and Figure 1b As shown, Figure 1a The change in user share as the number of repeated exposures increases. Figure 1bThe figure shows the change of CTR with the increase of repeated exposure times. It can be seen that with the increase of repeated exposure times, the click-through rate gradually decreases, and the user share decreases by an order of magnitude. In addition, the proportion of repeated exposure data in the overall data is low, so a small amount of data samples has little impact on model training. And with low click-through rate, the training process is very easy to be biased by other data samples with high proportion and high click-through rate. In addition, since the distinction between repeated exposure samples and other data samples is not high, it will also cause difficulties for the content recommendation model to distinguish repeated exposure samples during training. In summary, the current method of introducing freshness features is not effective in improving repeated exposure.
[0050] Therefore, this application takes into account that the proportion of repeated exposure samples is small, and therefore wants to increase the impact of such small samples on the training process of the content recommendation model. It is necessary to balance the sample size under different exposure times to avoid the first exposure samples with a large proportion bringing high prediction results of the content recommendation model.
[0051] Based on this, an embodiment of the present application provides a media content recommendation method, in which, when training a content recommendation model, the repeated recommendation weight of each training sample is obtained according to the repetition degree of the sample media content in each training sample being recommended to the sample object, and the model loss value of the content recommendation model is determined based on the repeated recommendation weight and interaction probability deviation corresponding to each training sample, that is, for each training sample, when calculating the sample loss value brought by the sample to the entire model, it will be based on the respective repeated recommendation weight, and the higher the repetition degree, the greater the repeated recommendation weight, which is equivalent to increasing the sample loss value corresponding to the training sample with a higher degree of repetition, increasing the proportion of such training samples in the overall training samples, balancing the sample size at different degrees of repetition, so that the training process has a stronger learning degree for such training samples, and the content recommendation model finally obtained can reduce the interaction probability deviation for media content with a higher degree of repetition, and improve the probability estimate of such media content, so as to achieve the purpose of improving the repeated exposure of media content and improving the accuracy of media content recommendation.
[0052] In addition, considering that the learning weights are increased for media content with a high degree of re-exposure during the training process, when the obtained content recommendation model recommends media content online, it will also cause deviations in the prediction of the overall media content by the content recommendation model. Therefore, in an embodiment of the present application, when making online recommendations, upsampling will be performed on the deviations caused by the training process to compensate for the deviations, thereby improving the accuracy of the final probability estimate.
[0053] The following briefly introduces the application scenarios to which the technical solutions of the embodiments of the present application can be applied. It should be noted that the application scenarios introduced below are only used to illustrate the embodiments of the present application and are not limited. In the specific implementation process, the technical solutions provided by the embodiments of the present application can be flexibly applied according to actual needs.
[0054] The solution provided in the embodiment of the present application can be applied to most media content recommendation scenarios, such as advertising recommendation scenarios. Figure 2a As shown, it is a schematic diagram of an application scenario provided by an embodiment of the present application. In this scenario, a terminal device 101 and a server 102 may be included.
[0055] The terminal device 101 can be, for example, a mobile phone, a tablet computer (PAD), a laptop computer, a desktop computer, a smart TV, an intelligent vehicle-mounted device, and an intelligent wearable device, and any device capable of presenting media content. The terminal device 101 can be installed with a target application, and the target should have the function of presenting media content, such as an instant messaging application, a music application, a video application, a short video application, a news application, or a shopping application. The application involved in the embodiment of the present application can be a software client, or a client such as a web page or a small program, and the server 102 is a server corresponding to the software or a web page, a small program, etc., without limiting the specific type of the client. The server 102 can be, for example, an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, i.e., content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms, but is not limited thereto.
[0056] It should be noted that the media content recommendation method in the embodiment of the present application can be executed by the terminal device 101 or the server 102 alone, or can be executed by the server 102 and the terminal device 101 together. For example, when there is an exposure opportunity for media content by the terminal device 101, a content recommendation request is triggered. Then, based on the content recommendation request, the terminal device 101 itself selects the target media content that is most suitable for the current exposure opportunity from the candidate media content through the content recommendation model, and displays it through the exposure opportunity. Alternatively, when there is an exposure opportunity for media content by the terminal device 101, a content recommendation request is triggered to be sent to the server 102. Then, based on the content recommendation request, the server 102 selects the target media content that is most suitable for the current exposure opportunity from the candidate media content through the content recommendation model, and returns the target media content to the terminal device 101, and then displays it through the exposure opportunity. Alternatively, when the terminal device 101 has an exposure opportunity for media content, it triggers a content recommendation request to be sent to the server 102. The server 102 predicts the probability estimate corresponding to the candidate media content based on the content recommendation request through the content recommendation model, and then returns it to the terminal device 101. The terminal device 101 can then select the target media content that is most suitable for the current exposure opportunity based on the probability estimate, and display it through the exposure opportunity. This application is not specifically limited here, and the following mainly uses the execution of the server 102 as an example for illustration.
[0057] Both the server 102 and the terminal device 101 may include one or more processors, memories, and interactive I / O interfaces, etc. In addition, the server 102 may also be configured with a database, which may be used to store historical recommendation sequences of target objects, interactive data, and trained model parameters, etc. Among them, the memories of the server 102 and the terminal device 101 may also store program instructions required to be executed in the media content recommendation method provided in the embodiment of the present application, and these program instructions, when executed by the processor, can be used to implement the media content recommendation process provided in the embodiment of the present application, for example, the server 102 may store program instructions for the prediction of probability estimates and the selection of media content related processes, so that the processor of the server 102 can execute the related processes.
[0058] In a possible implementation, the method of the embodiment of the present application can be applied to an advertising recommendation scenario. Since the prediction process of the probability estimate involves the use of the attribute information of the target object and the interactive data, in order to achieve more accurate advertising recommendations to the target object, the target object can be requested to authorize the use of the relevant data. If the target object allows, see Figure 2bAs shown in the figure, when there is an opportunity for advertising exposure, such as when the target opens a web page containing an advertising slot, when there is an advertising slot when sliding in the circle of friends, or when the public account page contains an advertising slot, an advertising exposure opportunity is triggered. Correspondingly, the advertising platform will receive an advertising request from the content server (the backend service corresponding to the front end), and then select the most suitable advertisement from the advertising library to return. All advertising exposure, click and conversion data generated in this process will be stored, and input features and category labels will be generated for model training through extraction, conversion and loading. The trained models will be loaded into the server of the advertising platform for online real-time advertising click-through rate prediction. For each advertisement displayed, the model estimates the probability of clicking the advertisement. The positive sample of the click will be marked with label 1, and the sample label of the unclicked sample will be marked with label 0. Among them, by loading the trained content recommendation model, the object characteristics of the target object can be obtained based on the relevant data of the target object, and the probability estimate of each candidate advertisement being clicked by the target object can be comprehensively predicted by combining the advertising characteristics, advertising position characteristics and traffic characteristics of each candidate advertisement, so as to select the advertisement recommended to the target object in the current advertising exposure opportunity and present it in the corresponding advertising position. Since the content recommendation model improves the learning weight of repeated exposure advertisements during training, it can effectively improve the accuracy of the probability estimate and improve the accuracy of advertising recommendations.
[0059] In the embodiment of the present application, the terminal device 101 and the server 102 can be directly or indirectly connected to each other through one or more networks 103. The network 103 can be a wired network or a wireless network, for example, the wireless network can be a mobile cellular network, or a wireless fidelity (Wireless-Fidelity, WIFI) network, and of course, it can also be other possible networks, which are not limited in the embodiment of the present application.
[0060] It should be noted that Figure 2a The figure is only an example. In fact, the number of terminal devices and servers is not limited and is not specifically limited in the embodiments of the present application.
[0061] The following describes the media content recommendation method provided by the exemplary embodiment of the present application in combination with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the implementation of the present application is not limited in this respect.
[0062] Since the media content recommendation process of the embodiment of the present application needs to be based on a content recommendation model, and the content recommendation model needs to be trained before being put into use, before introducing the content recommendation model of the embodiment of the present application, the training process of the content recommendation model used in the embodiment of the present application is first introduced here.
[0063] See Figure 3 FIG. 1 is a schematic diagram of a training process of a content recommendation model provided in an embodiment of the present application, and the process includes the following steps:
[0064] Step 301: Obtain multiple training samples.
[0065] In an embodiment of the present application, training samples can be collected from a historical recommendation process, that is, each training sample can be a media content historically recommended to an object. These objects are referred to as sample objects in the present application embodiment, and these media contents are referred to as sample media contents. The label of each training sample can be the true value of the probability that the sample object will perform an interactive operation on the sample media content after the sample media content is recommended to the sample object. The true value of the probability represents the actual probability that the sample object will perform an interactive operation on the sample media content.
[0066] In a possible implementation, the interactive operation may refer to a specific operation, such as clicking, liking, or commenting on the sample media content, or purchasing, activating, or downloading the recommended items of the sample media content. In actual scenarios, in this case, there are two situations in which the sample object performs interactive operations on the sample media content, namely, performing interactive operations and not performing interactive operations. Therefore, the true value of the probability usually has two situations, namely, 0 and 100%, or is represented by 0 and 1, and can be set according to the actual situation of the sample object performing interactive operations on the sample media content.
[0067] In a possible implementation, the interactive operation may also refer to a combination of multiple operations, such as a series of predetermined actions such as clicking and activating. Then the value of the probability true value may be determined based on the operation conditions of each operation, such as the mean or weighted value of the probability true values of each operation. Taking the mean as an example, if the interactive operation is designated as the click and purchase operation after the sample media content is recommended to the sample object, and the sample object clicks and purchases the sample media content, the probability of each of the two operations is 100%, and the mean of the two, that is, the probability true value of the sample, is 100%. If the sample object clicks on the sample media content but does not purchase the product associated with the sample media content, and the probability of the click operation is 100% and the probability of the purchase operation is 0, then the mean of the two, that is, the probability true value of the sample, is 50%.
[0068] In the embodiment of the present application, the operation data of each object on the media content in the actual scene is continuously performed, and usable training samples will be continuously generated. In order to improve the accuracy of the online recommendation of the content recommendation model, the content recommendation model can be continuously trained and updated.
[0069] In a possible implementation, the training of the content recommendation model may be triggered based on certain conditions. When the training conditions are met, the acquisition of training samples is triggered to train the content recommendation model. The training conditions may be set according to the needs of the actual scenario. For example, the accuracy of the content recommendation model used online may be tested. When the accuracy is lower than a certain accuracy threshold, the training conditions are considered to be met and the content recommendation model needs to be trained. Alternatively, the number of training samples may be tested. When the number of collected training samples reaches a certain number threshold, the training conditions are considered to be met and the content recommendation model needs to be trained.
[0070] In a possible implementation, the training of the content recommendation model may also be performed periodically, that is, the training cycle may be pre-set, and training samples may be collected during a training cycle, and when a training cycle ends, the training samples are used to train the content recommendation model. The training cycle may be set based on actual needs or system performance, for example, to 2 hours, 1 day, or one month, etc., and the present application embodiment does not limit this.
[0071] In the embodiment of the present application, a training sample may include the following information:
[0072] (1) Feature information of the sample object, such as the sample object’s gender, age, historical interaction data, and interests and hobbies.
[0073] (2) Feature information of the sample media content. For example, when the sample media content is an advertisement, it may include information such as the advertiser identity (ID), advertising industry category, advertisement type, and optimization target.
[0074] (3) Contextual features of this recommendation, including but not limited to information such as the time and number of times the media content is pulled.
[0075] (4) Traffic-side features of this recommendation, including but not limited to information such as the content display position and display scene of this recommendation. The display scene can be used to distinguish the source of the content display position, such as the public account scene, the circle of friends scene, or the music scene.
[0076] (5) Freshness feature, which is used to characterize the degree of repeated exposure of the current sample media content. For example, the number of repeated recommendations of the media content can be used as the freshness feature. The number of repeated recommendations refers to the number of times a media content is recommended to an object within a certain period of time.
[0077] Of course, any other possible information may also be included, and the embodiments of the present application do not limit this.
[0078] In the embodiment of the present application, the obtained N training samples can be expressed as follows:
[0079] {x1, x2, …, x i , …, x N}
[0080] The corresponding sample labels can be expressed as:
[0081] {y1, y2, …, y i , …, y N}
[0082] Among them, x i represents the i-th training sample, y i Represents the true value of the probability of the i-th training sample.
[0083] Step 302: Based on the historical recommendation sequences of the sample objects included in each training sample, respectively determine the repeated recommendation weights corresponding to each training sample.
[0084] In an embodiment of the present application, in order to improve the situation of repeated exposure, a new repetitive exposure weight (REW) algorithm is provided. By increasing the weight of the small sample part, the sample size under different exposure times is balanced through sample weighting to avoid the first exposure advertising sample with a relatively large proportion carrying a high probability estimate when it is exposed again. The main factor for distinguishing small samples is whether it is a repeated exposure sample, which can be characterized according to factors such as the number of repeated exposures and the length of the interval between the most recent repeated exposures.
[0085] For each training sample, which is a recommendation of sample media content for a sample object, the repeated exposure weight corresponding to the current training sample can be obtained by combining the historical recommendation sequence of the sample object to determine whether the current training sample is a repeated exposure sample, and combining the repeated degree of exposure of the sample media content in the current training sample in the historical recommendation sequence. In the repeated exposure scenario, the higher the repeated exposure degree, the more the learning degree of such samples should be increased, so the repeated recommendation weight should be higher, that is, the repeated recommendation weight can be considered as a penalty coefficient for repeated exposure samples.
[0086] In a possible implementation, the repeated recommendation weight may be expressed as:
[0087]
[0088] Among them, REW(x i ,y i) represents the repeated recommendation weight of the i-th training sample.
[0089] It can be seen that for samples that are repeatedly exposed and have not been interactively operated, the above repeated recommendation weight is greater than 1, which is equivalent to increasing the proportion weight of such samples. In other words, if the current training sample is a repeated exposure sample and y i =0, then weighted learning should be performed on the negative sample to achieve the purpose of suppressing repeated exposure.
[0090] It should be noted that the calculation process of the repeated recommendation weight will be specifically introduced in the subsequent embodiments, so it will not be described in detail here.
[0091] In the embodiment of the present application, the content recommendation model can be trained multiple times by obtaining various training samples until a certain convergence condition is finally reached. Since each training process is similar in this process, the following mainly illustrates one training process.
[0092] Step 303: Obtain probability estimates corresponding to each of the multiple training samples through the content recommendation model used this time, based on content features of the sample media content included in each of the multiple training samples and object features of the sample objects.
[0093] Specifically, the entire training process can be performed iteratively in batches. Before training, the configuration information related to each batch can be pre-configured, such as the number of training times, the number of batches for each training, the number of samples extracted for each batch, and the learning rate.
[0094] During each training (here refers to the smallest unit of training), a specified number of training samples can be extracted, and through the current content recommendation model (the initial model for the first training, and the model after the last parameter adjustment for subsequent training), the probability of interactive operations is predicted based on the content characteristics of the sample media content included in each of these training samples, as well as the object characteristics of the sample objects, to obtain the probability estimate corresponding to each training sample.
[0095] Step 304: Based on the obtained probability estimates and corresponding probability true values, the interaction probability deviations corresponding to the plurality of training samples are determined respectively.
[0096] Specifically, the interaction probability deviation is the deviation value between the estimated value and the actual value, which can characterize the estimation accuracy of the current content recommendation model.
[0097] Step 305: Obtain a model loss value based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample.
[0098] Specifically, based on the obtained interaction probability deviation, the model loss value of the content recommendation model relative to the estimated accuracy can be obtained. The purpose of the embodiment of the present application is to increase the model loss weight of the negative samples with repeated exposure to achieve the purpose of increasing the sample proportion.
[0099] In a possible implementation, the model loss value L can be calculated using the following calculation method:
[0100]
[0101] BCE(y i , f θ (x i ))=-y i logδ(f θ (x i ))-(1-y i )log(1-δ(f θ (x i )))
[0102] Among them, f θ (x i ) represents the training sample x i The probability estimate obtained, δ() represents the sigmoid function, and it should be noted that the loss function is specifically a binary cross entropy loss (BCE) function as an example. The physical meaning of BCE is to make the estimated f θ (x i ) is close to sample y i In actual scenarios, in addition to this loss function, any other possible loss function can also be used for calculation, and the embodiments of the present application do not limit this.
[0103] Step 306: Determine whether the iteration termination condition is met.
[0104] Specifically, the iteration termination condition may include at least one of the following conditions:
[0105] (1) When the model loss value L is not greater than the preset loss value threshold, that is, when the model accuracy is sufficient, the iterative training can be stopped.
[0106] (2) The number of training sessions reaches the upper limit of the number of training sessions.
[0107] If the result of step 306 is yes, that is, the characterization model training is completed, the training process ends, and the content recommendation model can be put into the subsequent probability estimation process.
[0108] Step 307: If the result of step 306 is no, the content recommendation model used this time is adjusted based on the model loss value, and the next training process is entered, that is, jumping to step 303 to continue execution.
[0109] Specifically, the model parameters may be adjusted by, for example, batch stochastic gradient descent. Of course, other possible model optimization methods may also be used, such as gradient descent or Newton's method.
[0110] Next, the calculation process of the repeated recommendation weights of the training samples is introduced.
[0111] Specifically, each element in the historical recommendation sequence of each sample object is a media content and its corresponding first recommendation time. For each training sample, the repetition degree of the training sample can be determined by the historical recommendation sequence of the sample object in the training sample.
[0112] In a possible implementation, the consideration of the degree of repetition can be characterized from the following dimensions:
[0113] (1) The number of repeated exposures, i.e., taking the recommendation time of the current training sample as the benchmark, the number of historical recommended contents that are repeated with the sample media content of the training sample is determined from the historical recommended contents before the recommendation time in the historical recommendation sequence of the sample object.
[0114] (2) The time interval between the repeated media content that meets the preset conditions and the recommendation time of the current training sample, such as the time interval between the historical recommended content that was the most recent repeated recommendation of the current training sample and the sample media content of the training sample, or the time interval between the historical recommended content that has been interactively operated and the sample media content of the training sample, where the historical recommended content that has been interactively operated can refer to the same media content as the sample media content of the current training sample, or it can be different media content.
[0115] Of course, in addition to the two dimensions mentioned above, other possible dimensions may also be included, and the embodiments of the present application do not limit this.
[0116] In the embodiment of the present application, since the process of calculating the repeated recommendation weight is similar for each training sample, a training sample is taken as an example to introduce the process of calculating the repeated recommendation weight. Figure 4 As shown, it is a schematic diagram of the process of determining the repeated recommendation weight provided in an embodiment of the present application.
[0117] Step 3021a: Obtain duplicate media content that is duplicated with sample media content in the training sample from the historical recommendation sequence of the sample object in the current training sample.
[0118] It should be noted that, for a sample object, the historical recommendation sequence is the media content of the sample object being recommended during a historical period, and for a training sample, its corresponding recommendation moment may be any moment within the historical period. Therefore, when selecting the repeated media content of a training sample, the recommendation moment of the training sample can be used as the reference moment, and whether there is repeated media content can be determined from the historical recommendation sequence before the reference moment.
[0119] See also Figure 5a As shown in the figure, it is a schematic diagram of the historical recommendation sequence. If the recommendation time of the current training sample is t cur , there are still other recommended contents after this moment, but when determining the repeated media content of the current training sample, t cur The dividing line is determined in the historical recommendation sequence before this moment, and the same is true for other training samples.
[0120] In a possible implementation, repeated media content may be selected based on content ID, that is, when the historical recommended content in the historical recommendation sequence has the same content ID as the sample media content in the current training sample, the historical recommended content is considered to be repeated media content of the sample media content.
[0121] In one possible implementation, taking into account that in actual scenarios, when the content publisher is delivering media content, due to considerations of delivery cost, communication effect or other aspects, a variety of materials or content forms may be used to promote the same recommendation. The media contents that are finally delivered have the same or similar visual and textual content at the material level, but belong to different media content and have different content IDs and material IDs. This will bring great challenges to subsequent model training and is also one of the reasons for repeated exposure. Therefore, in order to improve the problem of repeated exposure, it is also necessary to model these material information with similar images and texts, explore their essential semantic features, and identify the same or similar media content from different levels through the same fingerprint ID.
[0122] Specifically, the same or similar media content can be characterized from the following dimensions:
[0123] (1) Material fingerprint: From the perspective of the final media content, the same media content may have various types of changes such as size, form or delivery account, but the material fingerprint can cover the differences at this level, thus being used to more accurately measure the duplication index of media content. Therefore, when the materials or creative ideas of two media contents have visual similarities, including similar images, videos and texts, they are assigned the same material fingerprint ID.
[0124] See also Figure 5b The following is an example of an ad with the same material fingerprint ID but different material IDs. Figure 5b The two advertisements in the video both recommend the same bed sheet, but the patterns are different. However, it is not difficult to see from the pictures that these are different styles of the same bed sheet taken by the same merchant, and they should be given the same material fingerprint ID.
[0125] (2) Recommended item fingerprint, or product fingerprint. When the recommended items associated with two media contents are the same, they are assigned the same recommended item fingerprint ID. Taking advertising as an example, when the advertising objects (i.e., the advertised products) corresponding to the two advertisements are the same, they are assigned the same product fingerprint ID. Recommended item fingerprints can be obtained through clustering based on the original information and external material of the recommended items, and are used to better characterize the same product. From the perspective of user experience, it is currently hoped that the same promoted product will be exposed repeatedly in a short period of time.
[0126] See also Figure 5c As shown, this is an example of an advertisement with the same product fingerprint but different material ID and advertising ID. It can be seen that both Ads 1 and 2 are promotional advertisements for the XX game, but Ad 1 uses different materials from Ad 2, which ultimately generates two different advertisements, resulting in different advertising IDs. However, the products associated with the two advertisements are the same.
[0127] Of course, there may be other possible fingerprint dimensions, which are not limited in the embodiments of the present application.
[0128] Therefore, when selecting duplicate media content, the selection can be made based on the above fingerprint dimension.
[0129] In a possible implementation manner, when selecting the repeated media content, at least one of the following methods may be used:
[0130] (1) Selection based on material fingerprint identification, that is, repeated media content that repeats the material fingerprint identification of the sample media content can be obtained from the historical recommendation sequence of the sample object in the training sample.
[0131] For example, in the historical recommendation sequence, media content A and media content B are included, and their content IDs are different, but their material fingerprint identifiers are both ID1. If the material fingerprint identifier of the sample media content in the current training sample is also ID1, then media content A and media content B are both duplicate media content of the sample media content.
[0132] (2) Selection based on the recommendation fingerprint, that is, repeated media content with the recommendation fingerprint of the sample media content can be obtained from the historical recommendation sequence of the sample object in the training sample.
[0133] For example, in the historical recommendation sequence, media content C and media content D are included, and their content IDs are different, but their recommendation fingerprint identifiers are both ID2. If the recommendation fingerprint identifier of the sample media content in the current training sample is also ID2, then media content C and media content D are both duplicate media content of the sample media content.
[0134] The above two selection methods can also be used in combination, that is, the media content is considered to be duplicate only when the material fingerprint identifier and the recommendation fingerprint identifier are the same.
[0135] Exemplarily, in the historical recommendation sequence, media content E, media content F and media content G are included, and the content IDs of the three are different. The material fingerprint identifiers of media content E and media content F are both ID10, the recommended object fingerprint identifier of media content E is ID20, the recommended object fingerprint identifiers of media content F and media content G are ID21, and the material fingerprint identifier of media content G is ID11. The material fingerprint identifier of the sample media content in the current training sample is also ID10, and the recommended object fingerprint identifier is ID21, then media content F is a repeated media content of the sample media content.
[0136] Of course, in actual scenarios, when other fingerprint dimensions exist, selection can also be made from other fingerprint dimensions, and the embodiments of the present application do not limit this.
[0137] Step 3022a: Based on the obtained number of repeated media contents and the time interval between the first recommended time of each repeated media content and the second recommended time of the sample media content, obtain a first sub-weight in the dimension of the number of repetitions.
[0138] In the embodiment of the present application, the first sub-weight is a repetition degree index measured from the dimension of the number of repetitions, and its calculation can be performed in combination with the number of repeated media contents and the actual situation during each repetition, and here the repetition degree is measured by the time interval between each repetition. Among them, the first sub-weight is positively correlated with the number and negatively correlated with the time interval.
[0139] Specifically, the recommendation moment of each historical media content in the historical recommendation sequence is collectively referred to as the first recommendation moment, and the recommendation moment of the sample media content in the current training sample is referred to as the second recommendation moment. For each repeated media content, the initial sub-weight corresponding to each repeated media content can be determined based on the time interval between the first recommendation moment and the second recommendation moment of each repeated media content and the timing attenuation amplitude coefficient, wherein the timing attenuation amplitude coefficient is mainly used to control the amplitude of the timing attenuation, and the larger the timing attenuation amplitude coefficient, the faster the timing attenuation.
[0140] Finally, based on the obtained initial sub-weights corresponding to the respective repeated media contents and the quantity of the repeated media contents, weighted processing is performed to obtain a target sub-weight.
[0141] The target sub-weight can be the final first sub-weight. Alternatively, in order to avoid the first sub-weight value being too large or too small to cause a large estimation deviation to the model, a threshold can be cut off for the first sub-weight, that is, a weight value range [c min , c max ], and then after obtaining the target sub-weight, the final first sub-weight can be determined based on the size relationship between the target sub-weight and the preset weight value range.
[0142] Specifically, if the target sub-weight is within the weight value range, the target sub-weight is used as the first sub-weight; if the target sub-weight is greater than the upper limit of the weight value range, the upper limit is used as the first sub-weight; if the target sub-weight is greater than the lower limit of the weight value range, the lower limit is used as the first sub-weight. In this way, the value of the first sub-weight can be limited to the weight value range, avoiding the first sub-weight value being too large or too small, which will bring a large estimation deviation to the model and prevent the generation of abnormal values.
[0143] In a possible implementation, it is assumed that the currently exposed media cur content is represented as (A cur , t cur ), the historical exposure sequence is represented as {(A1, t1), (A2, t2), …, (A N , t N )}, where (A i , t i ) represents the i-th recommended media content and its recommendation time in history. The first sub-weight based on the repetition number dimension can be calculated by the following formula:
[0144]
[0145] Among them, w countRepresents the first sub-weight based on the repetition number dimension, that is, the penalty coefficient for the training sample in the repeated exposure number dimension, c max Characterizes the upper limit of the weight range, c min It represents the lower limit of the weight value range, M is the number of repeated media content, and γ is the temporal attenuation amplitude coefficient.
[0146] Step 3023a: Based on the time interval between the first recommended time and the second recommended time of the repeated media content that meets the preset condition in the obtained repeated media content, obtain a second sub-weight in the time interval dimension.
[0147] In the embodiment of the present application, the second sub-weight is a repetition index measured from the dimension of time interval, and its calculation can be performed in combination with the time interval between a certain repeated media content and the current sample media content. The second sub-weight is negatively correlated with the time interval.
[0148] Specifically, the preset condition can be set according to the needs of the actual scene, for example, it can be set to the repeated media content that was recommended repeatedly the most recently, or the repeated media content that has been interactively operated by the sample object.
[0149] Taking the repeated media content that meets the preset conditions as the most recently recommended repeated media content as an example, the time interval is the time interval between it and the current sample media content. The second sub-weight can also be determined based on the time interval and the temporal attenuation amplitude coefficient, and the weight value range of the second sub-weight [r min , r max ] is compared, and when it is within the weight value range, the value determined based on the time interval and the timing attenuation amplitude coefficient is determined as the second sub-weight; or, when the determined value is greater than the upper limit value of the weight value range, the upper limit value is determined as the second sub-weight; or, when the determined value is less than the lower limit value of the weight value range, the lower limit value is determined as the second sub-weight.
[0150] In a possible implementation, the second sub-weight based on the time interval dimension can be calculated by the following formula:
[0151]
[0152] Among them, w recency Represents the second sub-weight based on the time interval dimension, that is, the penalty coefficient for the training sample in the time interval dimension, t a Indicates the recommended time of the repeated media content that meets the conditions, r max Characterizes the upper limit of the weight range, r min The lower limit of the weight value range.
[0153] Step 3024a: Perform weighted processing based on the first sub-weight and the second sub-weight to obtain a repeated recommendation weight.
[0154] In the embodiment of the present application, different dimension weight values can be assigned to the repetition number dimension and the time interval dimension respectively, so that the weighted sum can be performed based on the above-obtained sub-weights and the dimension weight values corresponding to each dimension to obtain the repeated recommendation weight. Among them, the dimension weight value can be set according to the importance of different dimensions to the repeated recommendation weight, can also be obtained based on experimental measurements, or can also be obtained based on parameter adjustment during the training process.
[0155] In a possible implementation, the repeated recommendation weight may be:
[0156] w=α×w count +(1-α)×w recency
[0157] Among them, α is w count The corresponding dimension weight value, (1-α) is w recency The corresponding dimension weight value.
[0158] The following is a specific example to illustrate the calculation of repeated recommendation weights. Figure 5d , which is an example diagram of calculating the repeated recommendation weight using the REW algorithm.
[0159] The sample media content of the current training set sample is represented as a circle. Different shapes represent different media content. It can be seen that in the historical recommendation sequence, the sample media content was repeatedly recommended 3 times, namely, the 1st, 3rd and 5th recommendations. Here, γ is 0.5 as an example.
[0160] Then, according to the above-mentioned calculation method of the first sub-weight based on the dimension of the number of repetitions, it is specifically as follows:
[0161]
[0162] It can be seen that the more historical exposure times there are and the closer the time interval to the current sample media content is, the greater the final first sub-weight is.
[0163] Similarly, according to the above-mentioned calculation method of the second sub-weight based on the time interval dimension, it is specifically:
[0164]
[0165] It can be seen that the smaller the time interval corresponding to the most recent recommendation, the larger the second sub-weight is. Finally, based on the obtained w count and w recencyThe weights are added to obtain the repeated recommendation weights.
[0166] As shown above, in the embodiment of the present application, the definition of repeated recommendation can be the same from different fingerprint dimensions. Therefore, when calculating the repeated recommendation weight of a training sample, it can also be performed separately according to the degree of repetition of different fingerprint dimensions, and the final repeated recommendation weight can be obtained by combining the degree of repetition of different fingerprint dimensions.
[0167] Similarly, since the process of calculating the repeated recommendation weight is similar for each training sample, a training sample is used as an example to introduce the process of calculating the repeated recommendation weight. Figure 6 As shown, it is another schematic diagram of the determination process of repeated recommendation weight provided by the embodiment of the present application. Here, the above-mentioned examples of material fingerprint identification and recommendation object fingerprint identification are used to calculate the repeated recommendation weight from these two fingerprint dimensions. Correspondingly, the historical recommendation sequence can be specifically divided into a first historical recommendation subsequence composed of material fingerprint identification associated with media content, and a second historical recommendation subsequence composed of recommendation object fingerprint identification associated with media content.
[0168] Step 3021b: Obtain repeated media content with the same material fingerprint identifier as the corresponding sample media content from the first historical recommendation subsequence of the corresponding sample object, and obtain a third sub-weight based on the material repetition dimension based on the second recommendation time of the sample media content and the first recommendation time of each repeated media content.
[0169] Step 3022b: Obtain repeated media content with the same recommendation fingerprint as the sample media content from the second historical recommendation subsequence of the sample object, and obtain a fourth sub-weight based on the recommendation repetition dimension based on the second recommendation time and the first recommendation time of each repeated media content.
[0170] Specifically, the third sub-weight and the fourth sub-weight can be obtained by Figure 4 The calculation method shown is used to calculate, that is, Figure 4 The historical recommendation sequence in the illustrated embodiment can be replaced by the first historical recommendation subsequence and the second historical recommendation subsequence here, and the sub-weights of the material fingerprint identification and the recommendation fingerprint identification on their respective dimensions can be obtained. Therefore, the calculation process will not be described in detail here.
[0171] Step 3023b: Perform weighted processing based on the third sub-weight and the fourth sub-weight to obtain a repeated recommendation weight.
[0172] Specifically, the material fingerprint identifier and the recommended object fingerprint identifier may be assigned respective dimensional weight values, and then a weighted sum may be performed based on the third sub-weight and the fourth sub-weight and the respective dimensional weight values to obtain the repeated recommendation weight.
[0173] In a possible implementation, the dimensional weight values of the material fingerprint identifier and the recommendation fingerprint identifier may be the same, that is, the material fingerprint identifier and the recommendation fingerprint identifier are deemed to have equal importance, and the repeated recommendation weight may be the sum of the third sub-weight and the fourth sub-weight.
[0174] In the embodiment of the present application, after completing Figure 3 After the training process shown, the content recommendation model can be used in the actual online recommendation process, because the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample, and as shown above, each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object. The higher the degree of repetition, the greater the repeated recommendation weight, so that the degree of learning for the training sample is stronger during training, so that the final content recommendation model can reduce the interaction probability deviation for media content with a higher degree of repetition, and improve the probability estimate of such media content, so as to achieve the purpose of improving the repeated exposure of media content and improving the accuracy of media content recommendation.
[0175] See also Figure 7 FIG. 1 is a flow chart of a media content recommendation method provided in an embodiment of the present application. The specific implementation process of the method is as follows:
[0176] Step 701: In response to a content recommendation request sent by a target object, a plurality of candidate media contents to be recommended are obtained.
[0177] In an embodiment of the present application, a content recommendation request may be triggered by an exposure opportunity of media content contained in a page browsed by the target object. For example, when the target object opens a new page, and the page contains a display position for media content, a content recommendation request will be triggered; or, when the target object slides on a page, and there is a display position for media content in the page after sliding, a content recommendation request will be triggered; or, when other operations are performed so that there is an exposure opportunity, a content recommendation request will also be triggered. The embodiment of the present application does not limit the specific triggering page for the content recommendation request. Taking the media content as an advertisement as an example, when there is an advertisement on the page, a content recommendation request will not be triggered, and the specific page where the advertisement position is located is not limited. It can be a public account page, a mini program page, or a circle of friends page, etc.
[0178] Specifically, when a content recommendation request is received, all media content in the media content library can be used as multiple candidate media content to be recommended; or, considering that the model calculation process consumes performance, a large number of candidate media content may increase the recommendation delay, so a rough selection can be performed on the media content library based on certain selection conditions to select a small number of candidate media content and input them into the subsequent content recommendation model for probability estimation. For example, a rough selection can be performed based on the characteristics of the current advertising position or the characteristics of the target object. Of course, other selection criteria can also be used, and the embodiments of the present application do not limit this.
[0179] Step 702: Obtaining estimated probabilities of the target object performing interactive operations on the multiple candidate media contents respectively based on the object features of the target object and the content features corresponding to the multiple candidate media contents through the trained content recommendation model.
[0180] In the embodiment of the present application, for each candidate media content obtained, it is necessary to combine object characteristics and content characteristics, etc. to predict the probability estimate of the target object's interactive operation on multiple candidate media contents after recommending each candidate media content to the target object. The target object usually interacts with the content that he or she is interested in. Therefore, to some extent, the probability estimate also reflects the possibility of the target object's interest in multiple candidate media contents.
[0181] Specifically, the content recommendation model can adopt the deep neural network (DNN) commonly used in machine learning algorithms. DNN is a neural network with multiple hidden layers.
[0182] In one possible implementation, see Figure 8 , which is a schematic diagram of a structure of a content recommendation model provided by an embodiment of the present application. The bottom layer of the model is the input layer, and the data input by the input layer may include various prediction-related features, which are similar to the features used in the above-mentioned training samples, so they will not be described in detail here.
[0183] See also Figure 8As shown in the figure, the traffic side features are sent to a factorization machine (FM) for processing, and then the results are input to the fully connected layer (FC layer). The output result of the fully connected layer is superimposed with the initial field weight of the field-weighted fifield factorization machine (FwFFM) to obtain a new field weight (New fieldweight), and the input freshness features, object features, context features, and content features are processed accordingly based on the New field weight. The output result of FwFFM is vectorized by the embedding layer, thereby converting it from a high-dimensional feature to a low-dimensional universal feature vector. The universal feature vector is input into different scenario parameter layers, and is assigned scene features in different scenarios, thereby obtaining scene feature vectors in different scenarios, such as Figure 8 The shared Tower shown in the figure is a general scenario parameter layer, and the feeds Tower, Details Tower, etc. are parameter layers for specific recommendation scenarios. The scenario feature vectors output by each scenario parameter layer are concat processed through the connection operation layer (ConcatLayer).
[0184] In addition, see Figure 8 Another data processing branch in the data processing branch inputs the basic attributes of the object (such as age, gender, etc.), content features and context features into a FM for processing, and inputs the result into the fully connected layer. And, in another data processing branch, the display position features of the content display position of the current requested media content are input into the linear regression (LR) layer for linear regression processing. Among them, the LR processing method can be, for example, sum, average or concatenate.
[0185] Finally, the general scene feature vector, the feature vector output by the Concat Layer, and the output results of the above two branches are superimposed to obtain the final fused feature vector, and the fused feature vector is mapped to the range of (0, 1) by the sigmoid function to obtain the probability estimate. The final probability estimate result can be marked as f θ, where θ is the hyperparameter of the model. It is worth mentioning that during the training process, the probability estimate obtained here combined with the label of the training sample can be used to calculate the current model loss value.
[0186] It should be noted that in actual scenarios, in addition to Figure 8 In addition to the model structure shown, any other possible model structure may also be used, and the embodiments of the present application are not limited to this.
[0187] Step 703: Based on the obtained multiple probability estimates, determine target media content from multiple candidate media content, and recommend the target media content to the target object.
[0188] In an embodiment of the present application, after obtaining the probability estimate value, the target media content recommended to the target object can be selected based on the probability estimate value, for example, the candidate media content with the highest probability estimate value can be used as the target media content, or the target media content can be comprehensively decided based on other selection factors, and finally the most suitable target media content can be recommended to the target object, and displayed and exposed on the content display position of the current page browsed by the target object.
[0189] In actual scenarios, content publishers publish media content through content delivery platforms and pay the media content platform based on the final delivery effect of the media content. Therefore, for the recommendation system, its purpose is to recommend media content to users, and then achieve the dissemination effect of media content through interactive operations such as user clicks on media content, thereby obtaining profits from the content publisher. In other words, the click-through rate (CTR) of users on media content is a major factor that determines the final profit of the content delivery platform. Taking the media content of a specific page as an example, when a user opens the page, if the media content recommendation is triggered, the recommendation system of the media content platform will select appropriate media content from the content library for display. When the media content is displayed to the user or the user clicks on the media content or even activates the application (application, APP) recommended by the media content or places an order, the content delivery platform will automatically deduct fees from the content publisher to collect the corresponding fees. It can be seen that accurate media content recommendations can help content publishers make recommendations to potential users more quickly, thereby improving the efficiency of media content delivery, maximizing resource utilization, and improving user experience, achieving a win-win situation for content publishers, content delivery platforms, and users.
[0190] The recommendation of media content not only requires the sorting of candidate media content in the candidate queue, but also requires accurate probability estimates of interactive operations such as click-through rate and conversion rate to ensure the stable revenue of the content delivery platform. In the embodiment of the present application, due to the unilateral suppression of samples with repeated exposure during the training phase, the repeated recommendation of media content is reduced, but at the same time, it will also cause the content recommendation model to estimate the overall data biased, and ultimately the overall revenue will be partially reduced.
[0191] Therefore, in order to ensure the estimated accuracy of the overall data, when it is actually used online, the overall data will be calibrated by upsampling the overall positive samples.
[0192] Specifically, for the above training process, the repeated recommendation weight of each training sample is:
[0193]
[0194] That is, when the rew algorithm is used for calculation, the weight of the current training sample is w count +w recency , and in other cases except this, the weight of the current training sample is 1, for example, when there is no repeated exposure, the weight of the current training sample is 1.
[0195] Then, the estimated bias brought by the above weights to the overall data set is:
[0196]
[0197] Therefore, in order to modify the bias, in the embodiment of the present application, the correction purpose is achieved by correcting the positive samples, that is, probabilistic upsampling is performed on all positive samples.
[0198] Specifically, after a content recommendation model is trained, the probability deviation compensation rate of the content recommendation model is obtained based on the repeated recommendation weights corresponding to each training sample and the interaction probability deviations obtained during the last training of the content recommendation model. In the process of online recommendation, after the probability estimate value of each candidate media content is obtained, the multiple probability estimate values are corrected based on the probability deviation compensation rate to obtain multiple probability correction values accordingly. Finally, based on the multiple probability correction values, the target media content is determined from the multiple candidate media content, thereby achieving the purpose of correcting the probability estimate value.
[0199] Among them, the probability deviation compensation rate can be obtained based on the deviation of all positive samples in the training process. Then, from each training sample, a positive sample whose probability true value is greater than a preset probability threshold can be obtained, such as a training sample with a label of 1. Then, based on the repeated recommendation weights corresponding to each positive sample, the interaction probability deviations corresponding to each positive sample can be weighted and summed to obtain the first factor of the probability deviation compensation rate, and the interaction probability deviations corresponding to each training sample can be summed to obtain the second factor of the probability deviation compensation rate. Then, the ratio of the first factor to the second factor is determined as the probability deviation compensation rate. Of course, this is only one implementation method of the probability deviation compensation rate. In actual scenarios, other possible calculation methods can also be used to obtain the probability deviation compensation rate.
[0200] In a possible implementation, the sampling rate, that is, the probability deviation compensation rate μ, may be:
[0201]
[0202] Then for each candidate media content, the probability correction value after correction can be expressed as:
[0203]
[0204] Among them, pctr is the probability estimate, and pctr′ is the probability correction value.
[0205] In summary, in order to reduce the repeated display of the same media content to the user side, that is, to solve the problem of estimation bias of the recommendation system for repeatedly exposed media content, the embodiment of the present application designs a model optimization solution based on sample enhancement. On the one hand, the weight of the small sample part is increased, and the sample size under different exposure times is balanced through sample weighting to avoid the first exposure media content sample with a relatively large proportion carrying a high estimation value when it is exposed again; on the other hand, we introduce sample resampling as another supplement to the sample imbalance problem. Based on the combination of sample weighting and sample resampling, the stability of the overall market is guaranteed to the maximum extent, while improving the indicators related to repeated exposure, avoiding the fluctuation of system returns.
[0206] By experimenting with this solution, it can be clearly felt that the repeated exposure phenomenon is significantly improved, as shown in the following Table 1.
[0207]
[0208] Table 1
[0209] It can be seen that from the material fingerprint dimension, the proportion of repeated exposures dropped from 29.1% to 24.84%, a significant decrease of 14.7%, and the model-estimated pCTR bias dropped from 6.88% to 2%, a decrease of 71%. From the product fingerprint dimension, the proportion of repeated exposures dropped from 40.5% to 38.27%, a significant decrease of 5.5%. The model-estimated pCTR bias dropped from 6.17% to 3.07%, a decrease of 50%. It can be seen that the optimization effect before and after the improvement is very significant.
[0210] See also Fig. 9 Based on the same inventive concept, the embodiment of the present application further provides a media content recommendation device 90, which includes:
[0211] The content acquisition unit 901 is used to obtain a plurality of candidate media contents to be recommended in response to a content recommendation request sent by a target object;
[0212] The probability estimation unit 902 is used to obtain the probability estimation values of the target object performing interactive operations on the multiple candidate media contents respectively based on the object characteristics of the target object and the content characteristics corresponding to the multiple candidate media contents through the trained content recommendation model;
[0213] A recommendation unit 903, configured to determine target media content from multiple candidate media content based on the obtained multiple probability estimates, and recommend the target media content to a target object;
[0214] Among them, the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample. Each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object in the corresponding training sample. Each interaction probability deviation represents: the deviation between the estimated probability of the sample object performing an interactive operation on the sample media content in the corresponding training sample and the true value of the probability.
[0215] In a possible implementation, the device further includes a model training unit 904, which is used to:
[0216] Based on the historical recommendation sequences of the sample objects included in each training sample, respectively determine the repeated recommendation weights corresponding to each training sample;
[0217] The content recommendation model is trained multiple times through various training samples. Each training process includes the following steps:
[0218] Through the content recommendation model used this time, based on the content features of the sample media content included in each of the multiple training samples and the object features of the sample object, the probability estimation values corresponding to each of the multiple training samples are obtained respectively, and based on the obtained probability estimation values and the corresponding true probability values, the interaction probability deviations corresponding to each of the multiple training samples are determined respectively;
[0219] Based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample, the model loss value is obtained, and the parameters of the content recommendation model used this time are adjusted based on the model loss value.
[0220] In a possible implementation, each element in the historical recommendation sequence is a media content and its corresponding first recommendation time; the model training unit 904 is specifically used to:
[0221] For each training sample, perform the following operations:
[0222] For a training sample, obtaining repeated media content that is repeated with sample media content in the training sample from a historical recommendation sequence of a sample object in the training sample;
[0223] Based on the obtained number of repeated media contents and the time interval between the first recommended time of each repeated media content and the second recommended time of the sample media content, obtain a first sub-weight in the dimension of the number of repetitions;
[0224] Based on the time interval between the first recommended time and the second recommended time of the repeated media content that meets the preset condition in the obtained repeated media content, obtaining a second sub-weight in the time interval dimension;
[0225] A weighted process is performed based on the first sub-weight and the second sub-weight to obtain a repeated recommendation weight.
[0226] In a possible implementation manner, each media content is associated with a material fingerprint of the content material used by the media content, and a recommendation fingerprint of the recommendation corresponding to the media content; the model training unit 904 is specifically configured to:
[0227] Obtaining repeated media content that is repeated with the material fingerprint identification of the sample media content from the historical recommendation sequence of the sample object in the training sample; and / or,
[0228] From the historical recommendation sequence of the sample object in the training sample, duplicate media content that is duplicated with the recommendation object fingerprint of the sample media content is obtained.
[0229] In a possible implementation, the model training unit 904 is specifically configured to:
[0230] Determine the initial sub-weights corresponding to the respective repeated media contents based on the time interval between the first recommended time and the second recommended time of the respective repeated media contents and the temporal attenuation amplitude coefficient;
[0231] Perform weighted processing based on the initial sub-weights and quantities corresponding to each of the repeated media contents to obtain a target sub-weight;
[0232] Based on the size relationship between the target sub-weight and a preset weight value range, a first sub-weight is determined.
[0233] In a possible implementation, the model training unit 904 is specifically configured to:
[0234] If the target sub-weight is within the weight value range, the target sub-weight is used as the first sub-weight;
[0235] If the target sub-weight is greater than the upper limit of the weight range, the upper limit is used as the first sub-weight;
[0236] If the target sub-weight is greater than the lower limit of the weight value range, the lower limit is used as the first sub-weight.
[0237] In a possible implementation, the first sub-weight is positively correlated with the quantity and negatively correlated with the time interval, and the second sub-weight is negatively correlated with the time interval.
[0238] In a possible implementation, the historical recommendation sequence includes a first historical recommendation subsequence consisting of material fingerprint identifiers associated with the media content and a second historical recommendation subsequence consisting of recommendation object fingerprint identifiers associated with the media content; then the model training unit 904 is specifically configured to:
[0239] For each training sample, perform the following operations:
[0240] For a training sample, obtaining repeated media content with the same material fingerprint identifier as the corresponding sample media content from the first historical recommendation subsequence of the corresponding sample object, and obtaining a third sub-weight based on the material repetition dimension based on the second recommendation time of the sample media content and the first recommendation time of each repeated media content;
[0241] Obtaining repeated media content having the same recommendation fingerprint as the sample media content from the second historical recommendation subsequence of the sample object, and obtaining a fourth sub-weight based on the recommendation repetition dimension based on the second recommendation time and the first recommendation time of each repeated media content;
[0242] A weighted processing is performed based on the third sub-weight and the fourth sub-weight to obtain a repeated recommendation weight.
[0243] In a possible implementation manner, the device further includes an estimation compensation unit 905, configured to:
[0244] Based on the repeated recommendation weights corresponding to the respective training samples and the respective interaction probability deviations obtained by the content recommendation model during the last training, the probability deviation compensation rate of the content recommendation model is obtained;
[0245] The recommendation unit is specifically configured to respectively correct the multiple probability estimation values based on the probability deviation compensation rate to obtain multiple probability correction values accordingly; and determine the target media content from the multiple candidate media contents based on the multiple probability correction values.
[0246] In a possible implementation manner, the estimation compensation unit 905 is specifically configured to:
[0247] From each training sample, obtain a positive sample whose true probability value is greater than a preset probability threshold;
[0248] Based on the repeated recommendation weights corresponding to each positive sample, the interaction probability deviations corresponding to each positive sample are weighted and summed to obtain the first factor of the probability deviation compensation rate;
[0249] Based on the summation of the interaction probability deviations corresponding to each training sample, a second factor of the probability deviation compensation rate is obtained;
[0250] The ratio of the first factor to the second factor is determined as the probability deviation compensation rate.
[0251] Through the above-mentioned device, the proportion of such training samples in the overall training samples is increased, so that the training process has a stronger learning degree for such training samples. The final content recommendation model can reduce the interaction probability deviation for media content with a higher degree of repetition, and improve the probability estimation value of such media content, so as to achieve the purpose of improving the repeated exposure of media content and improving the accuracy of media content recommendation.
[0252] The device can be used to execute the methods shown in the various embodiments of the present application. Therefore, for the functions that can be implemented by the various functional modules of the device, reference can be made to the description of the aforementioned embodiments and no further details will be given.
[0253] See also Fig.10 Based on the same technical concept, the present application also provides a computer device. In one embodiment, the computer device can be Figure 2a The server shown in FIG. Fig.10 As shown, it includes a memory 1001 , a communication module 1003 and one or more processors 1002 .
[0254] The memory 1001 is used to store computer programs executed by the processor 1002. The memory 1001 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and programs required to run the functions in the embodiments of the present application, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.
[0255] The memory 1001 may be a volatile memory, such as a random-access memory (RAM); the memory 1001 may 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 1001 may be any other medium that can be used to carry or store the 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 1001 may be a combination of the above memories.
[0256] The processor 1002 may include one or more central processing units (CPU) or a digital processing unit, etc. The processor 1002 is configured to implement the above media content recommendation method when calling the computer program stored in the memory 1001 .
[0257] The communication module 1003 is used to communicate with terminal devices and other servers.
[0258] The specific connection medium between the memory 1001, the communication module 1003 and the processor 1002 is not limited in the embodiment of the present application. Fig.10 In the embodiment, the memory 1001 and the processor 1002 are connected via a bus 1004. The bus 1004 is connected to the processor 1002 via a bus 1004. Fig.10 The connections between the other components are only for illustration and are not intended to be limiting. The bus 1004 can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Fig.10 The diagram shows that only one thick line is used, but this does not mean that there is only one bus or one type of bus.
[0259] The memory 1001 stores a computer storage medium, and the computer storage medium stores computer executable instructions. The computer executable instructions are used to implement the media content recommendation method of the embodiment of the present application. The processor 1002 is used to execute the media content recommendation method of each of the above embodiments.
[0260] In another embodiment, the computer device may also be a terminal device, such as Figure 2a In this embodiment, the structure of the computer device can be as follows: Fig.11 As shown, it includes: a communication component 1110, a memory 1120, a display unit 1130, a camera 1140, a sensor 1150, an audio circuit 1160, a Bluetooth module 1170, a processor 1180 and other components.
[0261] The communication component 1110 is used to communicate with the server. In some embodiments, a circuit wireless fidelity (Wireless Fidelity, WiFi) module may be included. The WiFi module belongs to a short-range wireless transmission technology. The computer device can help the user to send and receive information through the WiFi module.
[0262] The memory 1120 can be used to store software programs and data. The processor 1180 executes various functions and data processing of the terminal device by running the software programs or data stored in the memory 1120. The memory 1120 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. The memory 1120 stores an operating system that enables the terminal device to run. In the present application, the memory 1120 can store an operating system and various application programs, and can also store code for executing the media content recommendation method of the embodiment of the present application.
[0263] The display unit 1130 can also be used to display information input by the user or information provided to the user and a graphical user interface (GUI) of various menus of the terminal device. Specifically, the display unit 1130 may include a display screen 1132 disposed on the front of the terminal device. The display screen 1132 may be configured in the form of a liquid crystal display, a light emitting diode, etc. The display unit 1130 may be used to display various media content display pages in the embodiments of the present application.
[0264] The display unit 1130 can also be used to receive input digital or character information and generate signal input related to user settings and function control of the terminal device. Specifically, the display unit 1130 may include a touch screen 1131 arranged on the front of the terminal device, which can collect user touch operations on or near it, such as clicking a button, dragging a scroll box, etc.
[0265] The touch screen 1131 can be covered on the display screen 1132, or the touch screen 1131 and the display screen 1132 can be integrated to realize the input and output functions of the terminal device, and the integrated touch screen can be referred to as a touch display screen. In this application, the display unit 1130 can display the application and the corresponding operation steps.
[0266] The camera 1140 can be used to capture still images, and users can post comments on the images taken by the camera 1140 through the application. The camera 1140 can be one or more. The object generates an optical image through the lens and projects it onto the photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the light signal into an electrical signal, and then passes the electrical signal to the processor 1180 to convert it into a digital image signal.
[0267] The terminal device may further include at least one sensor 1150, such as an acceleration sensor 1151, a distance sensor 1152, a fingerprint sensor 1153, and a temperature sensor 1154. The terminal device may also be configured with other sensors such as a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, and a motion sensor.
[0268] The audio circuit 1160, the speaker 1161, and the microphone 1162 can provide an audio interface between the user and the terminal device. The audio circuit 1160 can transmit the electrical signal converted from the received audio data to the speaker 1161, which is converted into a sound signal for output. The terminal device can also be equipped with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 1162 converts the collected sound signal into an electrical signal, which is received by the audio circuit 1160 and converted into audio data, and then the audio data is output to the communication component 1110 to be sent to, for example, another terminal device, or the audio data is output to the memory 1120 for further processing.
[0269] The Bluetooth module 1170 is used to exchange information with other Bluetooth devices having Bluetooth modules through the Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable computer device (such as a smart watch) that also has a Bluetooth module through the Bluetooth module 1170 to exchange data.
[0270] The processor 1180 is the control center of the terminal device. It uses various interfaces and lines to connect various parts of the entire terminal. It executes various functions of the terminal device and processes data by running or executing software programs stored in the memory 1120 and calling data stored in the memory 1120. In some embodiments, the processor 1180 may include one or more processing units; the processor 1180 may also integrate an application processor and a baseband processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the baseband processor mainly processes wireless communications. It is understandable that the above-mentioned baseband processor may not be integrated into the processor 1180. In the present application, the processor 1180 can run the operating system, application programs, user interface display and touch response, as well as the media content recommendation method of the embodiment of the present application. In addition, the processor 1180 is coupled to the display unit 1130.
[0271] Based on the same inventive concept, an embodiment of the present application further provides a storage medium storing a computer program. When the computer program is executed on a computer, the computer executes the steps of the media content recommendation method according to various exemplary embodiments of the present application described above in this specification.
[0272] In some possible implementations, various aspects of the media content recommendation method provided by the present application may also be implemented in the form of a computer program product, which includes a computer program. When the program product is run on a computer device, the computer program is used to enable the computer device to execute the steps of the media content recommendation method according to various exemplary implementations of the present application described above in this specification. For example, the computer device may execute the steps of each embodiment.
[0273] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with 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 disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.
[0274] The program product of the embodiment of the present application may adopt a portable compact disk read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto. In the present application, the readable storage medium may be any tangible medium containing or storing a program, and the computer program included therein may be used by or in combination with a command execution system, apparatus or device.
[0275] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein a readable computer program is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A readable signal medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with a command execution system, apparatus, or device.
[0276] The computer program embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
[0277] Computer programs for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and conventional procedural programming languages such as "C" language or similar programming languages.
[0278] It should be noted that, although several units or subunits of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided into multiple units to be embodied.
[0279] In addition, although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.
[0280] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0281] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.
[0282] 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 equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A media content recommendation method, characterized in that: The method comprises: In response to a content recommendation request sent by a target object, obtaining a plurality of candidate media contents to be recommended; Obtaining, by means of a trained content recommendation model, respectively estimated probabilities of the target object performing interactive operations on the multiple candidate media contents based on the object features of the target object and the content features corresponding to each of the multiple candidate media contents; Determine target media content from the plurality of candidate media content based on the obtained plurality of probability estimates, and recommend the target media content to a target object; Among them, the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample. Each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object in the corresponding training sample. Each interaction probability deviation represents: the deviation between the estimated probability of the sample object performing an interactive operation on the sample media content in the corresponding training sample and the true value of the probability.
2. The method according to claim 1, characterized in that The content recommendation model is trained through the following training process: Determining respectively the repeated recommendation weights corresponding to the respective training samples based on the historical recommendation sequences of the sample objects respectively included in the respective training samples; The content recommendation model is trained multiple times through the training samples, and each training process includes the following steps: By using the content recommendation model used this time, based on the content features of the sample media content included in each of the multiple training samples and the object features of the sample object, the probability estimation values corresponding to each of the multiple training samples are obtained respectively, and based on the obtained probability estimation values and the corresponding true probability values, the interaction probability deviations corresponding to each of the multiple training samples are determined respectively; Based on the repeated recommendation weights and interaction probability deviations corresponding to each of the training samples, a model loss value is obtained, and based on the model loss value, the parameters of the content recommendation model used this time are adjusted.
3. The method according to claim 2, characterized in that Each element in the historical recommendation sequence is a media content and its corresponding first recommendation time; Then, based on the historical recommendation sequences of the sample objects included in each training sample, respectively determining the repeated recommendation weights corresponding to each training sample includes: For each training sample, perform the following operations respectively: For a training sample, obtaining repeated media content that is repeated with sample media content in the training sample from a historical recommendation sequence of a sample object in the training sample; Based on the number of obtained repeated media contents and the time interval between the first recommended time of each of the repeated media contents and the second recommended time of the sample media content, obtain a first sub-weight in the dimension of the number of repetitions; Based on the time interval between the first recommended time and the second recommended time of the repeated media content that meets the preset condition in the obtained repeated media content, obtaining a second sub-weight in the time interval dimension; The repeated recommendation weight is obtained by performing weighted processing based on the first sub-weight and the second sub-weight.
4. The method according to claim 3, characterized in that Each media content is associated with a material fingerprint of the content material used by itself and a recommendation fingerprint of its corresponding recommendation; Then obtaining repeated media content that is repeated with the sample media content in the training sample from the historical recommendation sequence of the sample object in the training sample includes: Obtaining repeated media content that is repeated with the material fingerprint identifier of the sample media content from the historical recommendation sequence of the sample object in the training sample; and / or, From the historical recommendation sequence of the sample object in the training sample, duplicate media content that is duplicated with the recommendation object fingerprint of the sample media content is obtained.
5. The method according to claim 3, characterized in that Based on the number of obtained repeated media contents and the time interval between the first recommended time of each of the repeated media contents and the second recommended time of the sample media content, a first sub-weight in the dimension of the number of repetitions is obtained, including: Determining initial sub-weights corresponding to the respective repeated media contents based on a time interval between a first recommended time and the second recommended time of the respective repeated media contents and a temporal attenuation amplitude coefficient; Performing weighted processing based on the initial sub-weights corresponding to the respective repeated media contents and the quantity to obtain a target sub-weight; The first sub-weight is determined based on a size relationship between the target sub-weight and a preset weight value range.
6. The method according to claim 5, characterized in that Determining the first sub-weight based on a magnitude relationship between the target sub-weight and a preset weight value range includes: If the target sub-weight is within the weight value range, the target sub-weight is used as the first sub-weight; If the target sub-weight is greater than the upper limit of the weight value range, the upper limit is used as the first sub-weight; If the target sub-weight is greater than the lower limit of the weight value range, the lower limit is used as the first sub-weight.
7. The method according to claim 3, characterized in that The first sub-weight is positively correlated with the quantity and negatively correlated with the time interval, and the second sub-weight is negatively correlated with the time interval.
8. The method according to claim 2, characterized in that The historical recommendation sequence includes a first historical recommendation subsequence consisting of material fingerprint identifiers associated with the media content and a second historical recommendation subsequence consisting of recommendation object fingerprint identifiers associated with the media content; Then, based on the historical recommendation sequences of the sample objects included in each training sample, respectively determining the repeated recommendation weights corresponding to each training sample includes: For each training sample, perform the following operations respectively: For a training sample, obtaining repeated media content with the same material fingerprint identifier as the corresponding sample media content from the first historical recommendation subsequence of the corresponding sample object, and obtaining a third sub-weight based on the material repetition dimension based on the second recommendation time of the sample media content and the first recommendation time of each repeated media content; Obtaining repeated media content having the same recommendation fingerprint as the sample media content from the second historical recommendation subsequence of the sample object, and obtaining a fourth sub-weight based on a recommendation repetition dimension based on the second recommendation time and the first recommendation time of each repeated media content; The repeated recommendation weight is obtained by performing weighted processing based on the third sub-weight and the fourth sub-weight.
9. The method according to any one of claims 1 to 8, characterized in that: Before determining the target media content from the plurality of candidate media content based on the obtained plurality of probability estimates, the method further includes: Obtaining a probability deviation compensation rate of the content recommendation model based on the repeated recommendation weights corresponding to the respective training samples and the respective interaction probability deviations obtained during the last training of the content recommendation model; Then, based on the obtained multiple probability estimates, determining the target media content from the multiple candidate media content includes: Based on the probability deviation compensation rate, the plurality of probability estimation values are respectively corrected to obtain a plurality of probability correction values accordingly; Based on the plurality of probability revision values, target media content is determined from the plurality of candidate media content.
10. The method according to claim 9, characterized in that Based on the repeated recommendation weights corresponding to the training samples and the interaction probability deviations obtained by the content recommendation model in the last training, the probability deviation compensation rate of the content recommendation model is obtained, including: Obtaining, from each of the training samples, a positive sample whose true probability value is greater than a preset probability threshold; Based on the repeated recommendation weights corresponding to the positive samples, weighted summing processing is performed on the interaction probability deviations corresponding to the positive samples to obtain the first factor of the probability deviation compensation rate; A second factor of the probability deviation compensation rate is obtained by summing up the interaction probability deviations corresponding to the training samples; The ratio of the first factor to the second factor is determined as the probability deviation compensation rate.
11. A media content recommendation device, characterized in that: The device comprises: A content acquisition unit, configured to obtain a plurality of candidate media contents to be recommended in response to a content recommendation request sent by a target object; A probability estimation unit, configured to obtain, through a trained content recommendation model, probability estimation values of the target object performing an interactive operation on the multiple candidate media contents based on the object features of the target object and the content features corresponding to the multiple candidate media contents; A recommendation unit, configured to determine target media content from the plurality of candidate media content based on the obtained plurality of probability estimates, and recommend the target media content to a target object; Among them, the model loss value of the content recommendation model is determined based on the repeated recommendation weights and interaction probability deviations corresponding to each training sample. Each repeated recommendation weight is positively correlated with the degree of repetition of the sample media content being recommended to the sample object in the corresponding training sample. Each interaction probability deviation represents: the deviation between the estimated probability of the sample object performing an interactive operation on the sample media content in the corresponding training sample and the true value of the probability.
12. A computer device comprising a memory, a processor and a computer program stored in 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 10 are implemented.
13. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.
14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.
