Artificial intelligence-based media information recommendation method, device and readable storage medium

Abstract

The embodiment of the application discloses a media information recommendation method and device based on artificial intelligence and a readable storage medium. The trained retention prediction model can be used to predict the retention indication value of each consumed media based on the first object feature of the to-be-recommended object and the first media feature of each consumed media of the to-be-recommended object. Each consumed media is sorted according to the retention indication value of each consumed media to obtain the recommendation order of each consumed media. Based on the recommendation order, the media information corresponding to each consumed media is recommended to the to-be-recommended object. The embodiment of the application can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and auxiliary driving. The scheme can effectively find out the real retention reasons with uncertainty and sparsity, so that the media information that really affects the retention of the consumer user can be preferentially recommended to the to-be-recommended object, thereby improving the recommendation accuracy of the media information and the overall retention rate of the media platform.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a media information recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence and machine learning technologies, these technologies are being used in more and more fields. Among them, media information recommendation (such as WeChat official account recommendation, video account recommendation, WeChat official account article recommendation, video recommendation, etc.) is an important application of artificial intelligence.

[0003] To achieve more accurate media information (such as articles and videos) recommendations so that media information consumers can obtain higher quality content, existing technologies use global features such as click-through rate and total click duration to model and predict retention rates, thereby identifying the media information that affects consumer retention.

[0004] However, the inventors of this application discovered during the actual research and development process that: due to the uncertainty and sparsity of retention reasons, and the excessive focus on global features in the prior art, it is impossible to effectively identify the real factors affecting consumer user retention, thus leading to the problem of inaccurate media information recommendations. Summary of the Invention

[0005] This application provides a media information recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product based on artificial intelligence. It can effectively identify the real reasons for retention that have uncertainty and sparsity, so that media information that truly affects the retention of consumers can be prioritized for recommendation to the target audience, thereby improving the accuracy of media information recommendation.

[0006] In a first aspect, embodiments of this application provide a media information recommendation method based on artificial intelligence, the method comprising:

[0007] Obtain the first object feature of the object to be recommended and the first media feature of each consumed media of the object to be recommended;

[0008] By using a trained retention prediction model, a retention indication value for each consumed media is obtained based on the first object feature and the first media feature. The trained retention prediction model is learned based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object.

[0009] The consumed media are sorted according to their retention index values ​​to obtain a recommended ranking of the consumed media.

[0010] Based on the recommendation ranking, media information corresponding to each consumed media is recommended to the object to be recommended.

[0011] Secondly, embodiments of this application provide a media information recommendation device, the media information recommendation device comprising:

[0012] The acquisition unit is used to acquire the first object feature of the object to be recommended and the first media feature of each consumed media of the object to be recommended.

[0013] The prediction unit is used to predict, based on the first object features and the first media features, the retention indication value of each consumed media by using a trained retention prediction model, wherein the trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object.

[0014] The sorting unit is used to sort the consumed media according to the retention indication value of each consumed media to obtain the recommended sort of the consumed media.

[0015] The recommendation unit is used to recommend media information corresponding to each consumed media to the object to be recommended based on the recommendation ranking.

[0016] Thirdly, embodiments of this application also provide a computer device, the computer device including a processor and a memory, the memory storing a computer program, and the processor executing any of the artificial intelligence-based media information recommendation methods provided in embodiments of this application when calling the computer program in the memory.

[0017] Fourthly, embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to execute the aforementioned artificial intelligence-based media information recommendation method.

[0018] Fifthly, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement any of the artificial intelligence-based media information recommendation methods provided in the embodiments of this invention.

[0019] From the above, it can be concluded that the embodiments of this application have the following beneficial effects:

[0020] This application embodiment uses a trained retention prediction model to predict the retention indicator value of each consumed media using the first object feature and the first media feature. It can aggregate the retention probability of the pair of objects to be recommended and consumed media (i.e., it can aggregate the user and media information pair), thereby predicting the media information that truly affects user retention (i.e., predicting the real reason for user retention). This avoids the problem of not being able to effectively find the real retention reasons with uncertainty and sparsity caused by directly using global feature modeling. Therefore, after sorting each consumed media using the retention indicator value of each consumed media, the media information corresponding to each consumed media is recommended to the object to be recommended. This can make the media information that truly affects the retention of consuming users preferentially recommended to the object to be recommended, thereby improving the accuracy of media information recommendation and improving the overall retention rate of the media platform. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram illustrating an application scenario of the media information recommendation method based on artificial intelligence provided in this application embodiment;

[0023] Figure 2 This is a schematic flowchart of an embodiment of the media information recommendation method based on artificial intelligence provided in this application.

[0024] Figure 3 This is a schematic diagram of a process for training the retention prediction model provided in the embodiments of this application;

[0025] Figure 4 This is a schematic diagram of the offline and online prediction application architecture of the retention prediction model based on the MIL framework in the embodiments of this application;

[0026] Figure 5 This is a schematic diagram of the training framework of the retention prediction model provided in the embodiments of this application;

[0027] Figure 6 This is an illustrative diagram illustrating the media information recommendation process provided in an embodiment of this application;

[0028] Figure 7 A schematic diagram of an embodiment of the media information recommendation device in this application;

[0029] Figure 8 A schematic diagram of the structure of the computer device involved in the embodiments of this application. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features, rather than being used to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0032] It is understood that in the specific implementation of this application, user information such as user account, location, consumed media, and consumption behavior (such as click behavior, reading behavior, and viewing behavior) related to the media information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0033] This application provides a media information recommendation method, apparatus, computer device, and computer-readable storage medium based on artificial intelligence. The media information recommendation apparatus can be integrated into a computer device, which can be a server or a user terminal, etc. The user terminal includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, and aircraft.

[0034] The AI-based media information recommendation method of this embodiment can be implemented by a server or jointly by a terminal and a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and AI platforms, but is not limited to these. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0035] The following example illustrates this AI-based media information recommendation method, implemented jointly by a terminal and a server.

[0036] refer to Figure 1 The media information recommendation system based on artificial intelligence provided in this embodiment of the invention includes a terminal 101 and a server 102, etc.; the terminal 101 and the server 102 are connected through a network, such as through a wired or wireless network.

[0037] The terminal 101 is used to acquire and send to the server 102 the first object features of the object to be recommended and the first media features of each consumed media of the object to be recommended. The server 102 can receive the first object features and first media features sent by the terminal 101; use a trained retention prediction model to predict the retention indication value of each consumed media based on the first object features and first media features, wherein the trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object; sort the consumed media according to the retention indication value of each consumed media to obtain a recommendation ranking; and recommend the media information corresponding to each consumed media to the object to be recommended based on the recommendation ranking.

[0038] The AI-based media information recommendation method provided in this embodiment may specifically involve AI cloud services. The following description takes a computer device as the execution subject of the AI-based media information recommendation method. For the sake of simplicity, the execution subject will be omitted in the following text.

[0039] Artificial intelligence cloud services are generally also known as AIaaS (AI as a Service). This is currently a mainstream service model for artificial intelligence platforms. Specifically, AIaaS platforms break down several common AI services and provide them as independent or packaged services in the cloud. This service model is similar to opening an AI-themed marketplace: all developers can access and use one or more AI services provided by the platform through API interfaces. Some experienced developers can also use the AI ​​framework and AI infrastructure provided by the platform to deploy and maintain their own dedicated cloud AI services.

[0040] The following is a detailed description in conjunction with the accompanying drawings. It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments. Although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be performed in a different order than that shown in the drawings.

[0041] like Figure 2As shown, the specific process of this AI-based media information recommendation method can be summarized in steps 201 to 204, where:

[0042] 201. Obtain the first object characteristics of the object to be recommended and the first media characteristics of each consumed media of the object to be recommended;

[0043] Media information includes, but is not limited to, various types of information such as public accounts, video accounts, articles, news information, and videos.

[0044] The "target audience" refers to the audience to whom media information is to be recommended. For example, it could be a consumer who wants to be recommended a public account, a consumer who wants to be recommended a video account, a consumer who wants to be recommended a public account article, or a consumer who wants to be recommended a video.

[0045] Here, the first object feature is the object feature of the object to be recommended, specifically it can be a feature vector used to represent the object to be recommended. For example, in step 201, information such as the gender and location of the object to be recommended can be represented as a vector to obtain the first object feature.

[0046] In this context, "consumed media" refers to media information that the target audience has consumed (e.g., clicked, read, or watched) within a past timeframe (e.g., the past 3 days). The criteria for determining whether a consumption has occurred can be determined based on the actual business scenario; this embodiment does not impose specific restrictions on the criteria. For example, if the media information is an article from a WeChat Official Account of a mini-program (or application), consumed media could specifically be articles clicked by the target audience within that mini-program within the past 3 days. Similarly, if the media information is an article from a WeChat Official Account of a mini-program (or application), consumed media could specifically be articles read for more than 1 minute within that mini-program within the past 3 days. Furthermore, if the media information is a video account from a short video application, consumed media could specifically be videos watched for more than 5 seconds within that short video application within the past 3 days.

[0047] Among them, the first media feature is the feature vector of the consumed media of the object to be recommended, which is used to characterize the features of the consumed media.

[0048] For example, in step 201, the first media feature can be obtained as follows: obtain the media identifier and media type of each consumed media; and perform vector representation based on the media identifier and media type of each consumed media to obtain the first media feature of each consumed media. For example, if the media information is an article and the consumed media information is an article from a public account that the recommended user clicked on in a certain mini-program in the past 3 days (e.g., article a), then the article ID and article category of article a can be vectorized to obtain the media feature of article a, which can be used as the first media feature of the consumed media (article a). As another example, if the media information is a video account and the consumed media information is a video account that the recommended user watched for more than 5 seconds in a certain short video application in the past 3 days (e.g., video account b), then the video account ID and video account category of video account b can be vectorized to obtain the media feature of video account b, which can be used as the first media feature of the consumed media (video account b).

[0049] 202. Using the trained retention prediction model, predictions are made based on the first object features and the first media features to obtain the retention indicator value for each consumed media.

[0050] The trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object. The detailed learning process will be described in detail later and will not be repeated here.

[0051] Among them, retention indicators are indicators used to reflect whether users are retained. For example, retention indicators can be retention scores, retention rates, etc.

[0052] Please refer to Figure 4 For example, the first object feature and the first media feature can be input into a trained retention prediction model. The prediction layer in the trained retention prediction model then predicts the retention indicator value for each consumed media based on the first object feature and the first media feature of each consumed media. For example, as... Figure 4 As shown, for the first object feature (object feature m) of user m input into the trained retention prediction model, the first media feature of user m's consumed media 1 (media feature 1), the first media feature of consumed media 2 (media feature 2), the first media feature of consumed media 3 (media feature 3), ..., the first media feature of consumed media n (media feature n) will be jointly predicted with media feature 1 and object feature m to obtain the retention score of user m's consumed media 1 (e.g., ...). Figure 4 The "Retention Score 1" shown in the figure) is obtained by jointly predicting the retention score of the consumed media 2 of user m by media feature 2 and object feature m (e.g., Figure 4The "Retention Score 2" shown in the figure) is obtained by jointly predicting the retention score of the consumed media 3 for user m with media feature 3 and object feature m (e.g., Figure 4 The "Retention Score 3" shown in the figure, ..., is the retention score of the consumed media n of user m, which is obtained by jointly predicting media feature n and object feature m (e.g., Figure 4 The "retention score n" shown in the figure.

[0053] 203. Sort each consumed media according to its retention index value to obtain a recommended ranking of each consumed media.

[0054] In step 203, you can sort all consumed media or only sort the consumed media with higher retention index values; there is no specific restriction here.

[0055] Depending on the information used for ranking, there are multiple ways to determine the recommended ranking of each consumed media in step 203. Taking the ranking of only consumed media with higher retention index values ​​as an example, the methods include:

[0056] (1) Sort the consumed media from highest to lowest according to their retention index value. At this time, step 203 may specifically include the following steps 2031A to 2032A:

[0057] 2031A. Obtain target consumed media whose retention indicator values ​​meet predetermined conditions from each consumed media.

[0058] The pre-defined conditions can be set according to the actual business scenario requirements, and there are no specific restrictions here. For example, it can be that the retention indicator value ranks in the top N among all the consumed media of the object to be recommended.

[0059] 2032A. Determine the recommended ranking of target consumed media based on the retention indicator value of the target consumed media.

[0060] For example, the top N consumed media with the highest retention index can be selected from all consumed media and used as target consumed media.

[0061] (2) Based on the retention index of consumed media and the comprehensive score of click-through rate of each consumed media, sort them from high to low. At this time, step 203 may specifically include the following steps 2031B to 2032B:

[0062] 2031B: Obtain the click-through rate of each consumed media.

[0063] The click-through rate of consumed media can be determined based on the exposure and clicks of consumed media.

[0064] 2032B. Based on the retention indicator and click-through rate of each consumed media, determine the recommended ranking of each consumed media.

[0065] For example, the top N consumed media outlets by retention indicator (CI) can be selected from all consumed media. Then, for each of the top N consumed media outlets by CI, their click-through rate (CTR) and retention indicator (CI) are weighted to obtain a comprehensive weighted score. Finally, the top N consumed media outlets by CI are ranked from highest to lowest according to their comprehensive weighted scores, resulting in a recommended ranking for each of the top N consumed media outlets by CI.

[0066] 204. Based on the recommendation ranking, recommend media information corresponding to each consumed media to the target audience.

[0067] Corresponding to the sorting method in step 203, there are also multiple methods for recommending media information in step 204, including, for example:

[0068] (1) Sort the consumed media according to their retention index values ​​from highest to lowest to obtain the recommended ranking of the target consumed media. Then, in step 204, media information corresponding to the target consumed media can be recommended to the target audience according to the recommended ranking of the target consumed media. Further, to achieve more accurate media information (such as articles and videos) recommendations so that media information consumers can obtain higher-quality content, media information of the same type as the target consumed media can also be recommended to users. In this case, step 204 can specifically include: obtaining the media type of the target consumed media; obtaining media information of the media type as the media information corresponding to the target consumed media; and recommending media information of the media type to the target audience according to the recommended ranking of the target consumed media.

[0069] (2) Based on the retention indicator value of the consumed media and the comprehensive score of the click-through rate of each consumed media, sort them from high to low to obtain the recommendation ranking of each consumed media with the top N retention indicators. At this time, in step 204, media information corresponding to each consumed media with the top N retention indicators can be recommended to the target audience according to the recommendation ranking of each consumed media with the top N retention indicators. Further, in order to achieve more accurate media information (such as articles and videos) recommendations so that media information consuming users can obtain higher quality content, media information of the same type as the consumed media with the top N retention indicators can also be recommended to users. At this time, step 204 can specifically include: obtaining media information of the same type as each consumed media with the top N retention indicators as the media information corresponding to each consumed media with the top N retention indicators; recommending media information of the same type as each consumed media with the top N retention indicators to the target audience according to the recommendation ranking of each consumed media with the top N retention indicators.

[0070] The training process of the retention prediction model is described below. For example, the trained retention prediction model can be obtained by training the retention prediction model to be trained through the following steps 301 to 305, wherein:

[0071] 301. Obtain the annotation retention indicator value of the first sample object.

[0072] The first sample object is the object used to train the retention prediction model.

[0073] The labeled retention indicator (CPI) reflects the actual online retention of the first sample object (e.g., the user's actual retention score in subsequent time periods, such as the next 3 days). The CPI is introduced as a supervisory signal to guide the training of the retention prediction model, thereby improving the prediction accuracy of the CPI of the trained model and increasing the confidence of the CPI (e.g., the retention indicator of consumed media, or the final retention indicator of aggregated users).

[0074] like Figure 4 As shown, for example, the labeled retention indicator value of the first sample object can be determined based on its subsequent media consumption and subsequent media exposure. For instance, the retention probability (or retention score, etc.) of the first sample object in a subsequent time period (such as within 3 days after the consumption of previously consumed media) can be calculated and used as the labeled retention indicator value for the first sample object. Here, subsequent media consumption refers to the number of times the first sample object consumes previously consumed media again in a subsequent time period, such as subsequent media clicks, subsequent media reads, and subsequent media views.

[0075] By utilizing the subsequent media consumption and exposure of the first sample object, the labeled retention indicator value of the first sample object can be determined. This allows us to use the media consumption (e.g., the number of times users click on consumed media again, read consumed media again, or watch consumed media again) and media exposure of the first sample object in subsequent time periods (e.g., within 3 days after the consumed media behavior) as a supervisory signal to guide the training of the retention prediction model. This enables more accurate quantification of retention, increases the confidence of the retention indicator value, and reduces the randomness and noise of the user retention indicator value.

[0076] For different forms of consumption behavior, there are multiple ways to obtain the labeled retention indication value in step 301, including, for example:

[0077] (1) Consumption behavior is click behavior. Based on the number of media clicks and media exposures of the first sample object in a subsequent time period (e.g., within 3 days after the click behavior of the consumed media), the retention probability of the first sample object in the subsequent time period is calculated as the labeled retention indicator value of the first sample object. At this time, step 301 may specifically include the following steps 3011A to 3013A, wherein:

[0078] 3011A. Obtain the subsequent media clicks of the first sample object.

[0079] The subsequent time period is the preset duration after the first sample object has consumed the consumed media, for example, within 3 days after the consumed media has been consumed.

[0080] Subsequent media clicks refer to the number of times the first sample subject clicks on consumed media again in a subsequent time period, such as the number of times user u clicks on consumed media in the next three days.

[0081] 3012A. Obtain the subsequent media exposure of the first sample object.

[0082] The subsequent media exposure count refers to the number of consumed media items shown to the first sample object in the subsequent time period. For example, showing user u the number of consumed media items to be exposed in the next three days.

[0083] 3013A. Based on subsequent media clicks and subsequent media exposures, determine the label retention indicator value for the first sample object.

[0084] In some embodiments, the ratio between subsequent media clicks and subsequent media exposures can be used as the retention probability of the first sample object in a subsequent time period, and thus serve as the labeled retention indicator value for the first sample object.

[0085] In some embodiments, as shown in Formula 1 below, the subsequent media clicks and subsequent media exposures can be substituted into the retention indicator calculation formula shown in Formula 1 to obtain the calculation result as the labeled retention indicator for the first sample object. For example, the number of consumed media clicks by user u in the next three days and the number of consumed media exposures given to user u in the next three days can be substituted into Formula 1 to calculate the retention score of user u in the next three days, which is used as the labeled retention indicator for user u, that is, the actual retention score of user u in the next three days.

[0086] W u =alog(1+cn) u )+blog(1+in u ) Formula 1

[0087] In Formula 1, cn u This refers to subsequent media clicks (e.g., the number of media clicks a user will consume in the next three days). u It represents subsequent media exposure (e.g., the number of consumed media items shown to user u in the next three days), a and b are hyperparameters, and Wu is the retention indicator value.

[0088] (2) Consumption behavior is reading behavior. Based on the number of media reads and media exposures of the first sample object in a subsequent time period (e.g., within 3 days after the consumed media reading behavior), the retention probability of the first sample object in the subsequent time period is calculated as the labeled retention indicator value of the first sample object. At this time, step 301 may specifically include the following steps 3011B to 3013B, wherein:

[0089] 3011B. Obtain the subsequent media readership of the first sample object.

[0090] The subsequent time period is the preset duration after the first sample object has consumed the consumed media, for example, within 3 days after the consumed media has been consumed.

[0091] Subsequent media reading volume refers to the number of times the first sample subject reads the consumed media again in a subsequent time period, such as the number of times user u reads the consumed media in the next three days.

[0092] 3012B, Obtain the subsequent media exposure of the first sample object.

[0093] The subsequent media exposure count refers to the number of consumed media items shown to the first sample object in the subsequent time period. For example, showing user u the number of consumed media items to be exposed in the next three days.

[0094] 3013B. Based on subsequent media readership and subsequent media exposure, determine the annotation retention indicator value for the first sample object.

[0095] In some embodiments, the ratio between subsequent media readership and subsequent media exposure can be used as the retention probability of the first sample object in a subsequent time period, and thus serve as the labeled retention indicator value for the first sample object.

[0096] In some embodiments, referring to Formula 1, subsequent media readership (in in Formula 1 at this time) can be used to calculate the media readership. u Specifically, the number of subsequent media reads and subsequent media exposures are substituted into the retention indicator calculation formula shown in Formula 1, and the calculation result is used as the labeled retention indicator value of the first sample object.

[0097] 302. Using the retention prediction model to be trained, predictions are made based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object to obtain the retention indicator value of each consumed media sample.

[0098] The second object feature is the object feature of the first sample object, which can specifically be a feature vector used to characterize the first sample object.

[0099] Among them, the consumed media sample is the media information that the first sample object consumed (such as clicks, reading, or viewing) within a past time period (e.g., within the past 3 days).

[0100] The second media feature is used to characterize the consumed media sample. The method for obtaining the second media feature can be found in the explanation of the first media feature, and will not be repeated here.

[0101] To facilitate understanding, the retention prediction model used in this implementation will be introduced below, such as... Figure 4 As shown, the retention prediction model includes a prediction layer and an aggregation layer. For example, the retention prediction model can specifically be a classifier using a multi-instance learning (MIL) mechanism. Figure 4 Taking the retention prediction model (using a Multi-instance Learning classifier as an example) as an example, the working principle of the retention prediction model is as follows:

[0102] ①Prediction layer (reference) Figure 4The "UI retention score" section is used to predict and generate retention scores for each media based on object characteristics and media characteristics of each media, serving as retention indicators for each media. For example, based on the second object characteristics and the second media characteristics of each consumed media sample of the first sample object, a retention indicator for each consumed media sample is obtained. Similarly, based on the first object characteristics and the first media characteristics of each consumed media of the object to be recommended, a retention indicator for each consumed media is obtained.

[0103] ② Polymer layer (reference) Figure 4 The "instance-level attention" section is used to aggregate the retention indicator values ​​of all consumed media for a user (e.g., the first sample object) to obtain the user's final retention score, which serves as the user's retention indicator value. For example, aggregating the retention indicator values ​​of all consumed media samples for the first sample object yields the first sample object's third retention indicator value. Exemplarily, the aggregation layer can aggregate the retention indicator values ​​of all consumed media for the user (i.e., the first sample object) according to the following Formula 2 to obtain the user's final retention score, which serves as the user's retention indicator value.

[0104]

[0105] In Formula 2, u is a given user, I us Let Q be the set of media consumed by user u, Q be the object characteristics of user u, and K be the set of media consumed by user u. ui Let be the feature vector of the consumed media i for user u, where d is the dimension of the feature, and S is the feature vector. ui It is the retention score of user u's consumed media i (predicted by the prediction layer of the retention prediction model), y MILu It is the final retention score of user u (obtained by aggregation from the aggregation layer of the retention prediction model).

[0106] For example, the second object features and the second media features of each consumed media sample of the first sample object can be input into the retention prediction model to be trained (e.g., Figure 4 In the Multi-instance Learning classifier shown, the prediction layer of the retention prediction model to be trained (such as...) Figure 4 The “UI retention score” shown is predicted based on the second object features and the second media features of each consumed media sample of the first sample object, and the retention score of each consumed media sample of the first sample object is generated as the retention indicator value of each consumed media sample.

[0107] 303. By aggregating the retention indicator values ​​of each consumed media sample through the retention prediction model to be trained, the third retention indicator value of the first sample object is obtained.

[0108] The third retention indicator value is the retention indicator value obtained by aggregating the retention indicator values ​​of each consumed media sample of the first sample object by the retention prediction model. It is used to reflect the user retention situation predicted by the retention prediction model offline (such as the user retention score predicted by the retention prediction model in the subsequent time period, such as the next 3 days).

[0109] For example, the aggregation layer of the retention prediction model to be trained (such as...) Figure 4 The "instance-level attention" shown aggregates the retention indicator values ​​of each consumed media sample to obtain the third retention indicator value of the first sample object.

[0110] 304. Based on the labeled retention indicator value and the third retention indicator value, determine the retention loss function of the retention prediction model to be trained.

[0111] For example, the retention loss function can be shown in Formula 3 below. The labeled retention indicator value and the third retention indicator value can be substituted into Formula 3 below to determine the retention loss function of the retention prediction model to be trained.

[0112]

[0113] In formula 3, y u It refers to the retention indicator value (such as the actual retention score of user u over the next three days), y MILu It is the third retention indicator (such as the retention score of user u predicted by the retention prediction model for the next three days), MSE() is the mean squared error loss function, L MIL It is the retention loss value of the retention prediction model to be trained.

[0114] 305. Based on the retention loss function, train the retention prediction model to be trained to obtain the trained retention prediction model.

[0115] Please refer to Figure 5 The retention prediction model can be referenced. Figure 5 In the "CMIL Predictor" section, there are several ways to train the model in step 305, including, for example:

[0116] (1) Training is performed using the retention loss function. In this case, step 305 may specifically include the following step 3051A:

[0117] 3051A. Using the retention loss function, the model parameters of the retention prediction model to be trained are adjusted to obtain the trained retention prediction model.

[0118] For example, with the goal of minimizing the retention loss function, the model parameters of the retention prediction model to be trained (such as the model parameters of the prediction layer and the model parameters of the aggregation layer) are adjusted until the training stops, and the trained retention prediction model is obtained.

[0119] Since the trained retention prediction model is trained on the retention prediction model to be trained based on the loss between the third retention indicator and the labeled retention indicator of the first sample object, the third retention indicator is obtained by aggregating the retention indicator of each consumed media sample of the first sample object, and the retention indicator of each consumed media sample is predicted based on the second object feature of the first sample object and the second media feature of each consumed media sample; it can be seen that the retention prediction model aggregates the retention score / retention probability of the recommended object and the consumed media pair (i.e., user-media information item pair) through the attention mechanism, thereby predicting the user's true retention. Based on this retention prediction model, the retention score / retention probability of the user-media information item pair can be learned, avoiding the problem of not being able to effectively find the real retention reasons with uncertainty and sparsity due to directly using global feature modeling, and improving the credibility of the aha item that truly affects user retention.

[0120] (2) Training is performed using the retention loss function and the retention bias function. In this case, step 305 may specifically include the following step 3051B. The training process of the retention prediction model also includes the following steps A1 to A3:

[0121] A1. Based on the weight parameters of the retention prediction model to be trained, the consumed media samples are divided to obtain the first consumed media and the second consumed media.

[0122] The weight parameter indicates the weight value of each consumed media sample, with the weight value of the first consumed media being greater than that of the second consumed media.

[0123] Among them, the first consumer media is the consumer media sample with a relatively high weight value among all consumer media samples.

[0124] Among them, the second consumer media is the consumer media sample with a relatively low weight value among all consumer media samples.

[0125] The retention prediction model to be trained can learn the weight of each consumed medium (i.e., user-user consumed medium pair), meaning the weight parameters of the retention prediction model to be trained include the weight of each consumed medium (i.e., user-user consumed medium pair). For example, firstly, the weight values ​​of each consumed medium sample are determined using the weight parameters of the retention prediction model to be trained. Then, according to a predetermined grouping ratio (e.g., a group with relatively higher weight values ​​accounting for 10% and another group with relatively lower weight values ​​accounting for 90%; or the ratios are 20% and 80%, or 30% and 70%, etc.), all consumed media samples of the first sample object are divided into two groups. At this point, the group of consumed media samples with relatively higher weight values ​​is designated as the first consumed medium (referred to as the positive item in this paper), and the other group of consumed media samples with relatively lower weight values ​​is designated as the second consumed medium (referred to as the negative item in this paper).

[0126] A2. Based on the retention indication values ​​of the first consumer media and each consumed media sample, obtain the first retention indication value of the first sample object; and based on the retention indication values ​​of the second consumer media and each consumed media sample, obtain the second retention indication value of the first sample object.

[0127] The first retention indicator value is the retention indicator value of the first sample object obtained by aggregating and predicting the retention indicator values ​​of the first consumption medium (i.e., positive item).

[0128] The second retention indicator value is the retention indicator value of the first sample object obtained by aggregating and predicting the retention indicator values ​​of the second consumption medium (i.e., negative items).

[0129] For example, positive and negative items can be hidden separately, and then aggregated and predicted through the aggregation layer of the retention prediction model to be trained to obtain a first retention indicator value and a second retention indicator value. For instance, after the prediction layer outputs the retention indicator value for each consumed media sample, the retention indicator value for the first consumed media is selected from these values ​​and input into the aggregation layer; through the aggregation layer, the retention indicator value of the first consumed media (i.e., the positive item) is used for aggregated prediction to obtain the first retention indicator value. Similarly, after the prediction layer outputs the retention indicator value for each consumed media sample, the retention indicator value for the second consumed media is selected from these values ​​and input into the aggregation layer; through the aggregation layer, the retention indicator value of the second consumed media (i.e., the negative item) is used for aggregated prediction to obtain the second retention indicator value.

[0130] A3. Based on the first retention indicator value and the second retention indicator value, construct the retention bias function of the retention prediction model to be trained.

[0131] The retention bias function is used to increase the deviation between the first retention indicator value and the second retention indicator value.

[0132] For example, the retention bias function can be shown in Formula 4 below. The third retention indicator value, the first retention indicator value, and the second retention indicator value can be substituted into Formula 4 below to determine the retention bias function of the retention prediction model to be trained.

[0133]

[0134] In Formula 4, U is the user set, and y MIL It is the third retention indicator (such as the retention score of user u predicted by the retention prediction model for the next three days, for reference). Figure 5 in "y MIL ”), y CMILpos y CMILneg These are the first retention indicator values ​​(i.e., the retention indicator values ​​of the first sample object obtained by performing hidden prediction on negative items, see reference). Figure 5 in "y CMIL The second retention indicator (i.e., the retention indicator value of the first sample object obtained by performing hidden prediction on the positive item, see reference) Figure 5 in "y CMIL ”), d(y MIL ,y CMILpos ) is y MIL With y MILpos The absolute deviation between them, d(y) MIL ,y CMILneg ) is y MIL With y CMILneg The absolute deviation between them, L CL This is the retention bias value of the retention prediction model to be trained (reference). Figure 5 "L" CL "); m is a hyperparameter that is adjusted within [0,3] to widen the distribution span between positive and negative samples.

[0135] In the course of actual research and development, the inventors of this application concluded that media information that influences user retention is often a sparse portion of a large amount of information. Furthermore, when dividing the consumed media samples, the proportion of the first consumed media can be made greater than the proportion of the second consumed media, in order to extract the truly influential, sparse media information from a large amount of media information.

[0136] Thus, for reference Figure 5 When training the retention prediction model, the retention bias function is maximized (see reference). Figure 5 "L" JSThis allows for the comparison learning method to increase the absolute deviation d(y) between the first retention indicator and the second retention indicator. MILpos ,y MILneg This allows for the exploration of scarce media information that influences user retention, thereby improving the accuracy of retention indicators for each consumed media item in the recommended content. This enables the prioritization of media information that influences user retention, improving user retention rates and providing users with higher-quality content. Therefore, a retention prediction model based on multi-instance learning, incorporating a retention bias function, can effectively use "multi-instance learning based on contrastive learning" to find more accurate and reliable media information that influences user retention (referred to as AHA items in this paper).

[0137] Step 305 may specifically include the following step 3051B:

[0138] 3051B. Based on the retention bias function and retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0139] For example, with the goal of minimizing the retention loss function and maximizing the retention bias function, the model parameters of the retention prediction model to be trained (such as the model parameters of the prediction layer and the model parameters of the aggregation layer) are adjusted until the training stop condition is met, and the trained retention prediction model is obtained.

[0140] As can be seen, given the uncertainty and sparsity of retention reasons, contrastive learning is introduced through a retention bias function. This allows the attention mechanism of the retention prediction model to be guided by contrastive learning to uncover the real reasons that affect user retention, thereby improving the accuracy of retention indicators for each consumed media of the target audience.

[0141] (3) Training is performed using the retention loss function and the divergence function. In this case, step 305 may specifically include the following step 3051C. The training process of the retention prediction model also includes the following steps B1 to B2:

[0142] B1. Using a trained causal reasoning model, predictions are made based on each consumed media sample to obtain the fourth retention indicator value for the first sample object.

[0143] The fourth retention indicator value is a retention indicator value predicted by the trained causal inference model based on the second object features of the first sample object and the second media features of each consumed media sample.

[0144] For ease of understanding, the causal reasoning model in this implementation will be introduced below. This causal reasoning model includes a selection layer and a reasoning layer. For example, the causal reasoning model can specifically be rationale-MIL (…). Figure 5 Taking the rationale-MIL causal inference model as an example, the working principle of the causal inference model is as follows:

[0145] ① Select layer (reference) Figure 5 The "Rationale Selection" section is used to determine media information that has a high impact on users based on the object characteristics of users (such as the first sample object) and the media characteristics of subsequently consumed media. For example, the retained media of the first sample object is determined based on the second object characteristics and the second media characteristics of each consumed media sample.

[0146] ②Inference layer (reference) Figure 5 The "Rationale-MIL Predictor" section is used to predict user retention indicators based on the object characteristics of users (such as the first sample object) and the media characteristics of media information that has a high impact on users. For example, based on the second object characteristics and the media characteristics of the retention media of the first sample object, a fourth retention indicator value for the first sample object is obtained.

[0147] For example, step B1 may specifically include: determining the retained media of the first sample object based on each consumed media sample using a trained causal inference model; and predicting the fourth retention indicator value of the first sample object based on the features of the second object and the media features of the retained media of the first sample object using the trained causal inference model.

[0148] For example, such as Figure 5 As shown, the second object features of the first sample object and the second media features of each consumed media sample (refer to...) can be combined. Figure 5 The "consumed media samples" are input into the trained causal inference model. The selection layer of the trained causal inference model first identifies media information with a high impact on user retention (such as...). Figure 5 The "Retention Media" shown in the image is used as the retention media of the first sample object. Then, through the inference layer of the trained causal inference model, a prediction is made based on the features of the second object and the media features of the retention media of the first sample object to obtain the fourth retention indicator value of the first sample object (see reference). Figure 5 in "y RMIL ”).

[0149] For example, a causal reasoning model can be trained through the following steps C1 to C4:

[0150] C1. Obtain the true retention indicator value of the second sample object, the third object characteristics of the second sample object, the third media characteristics of each historical consumption media of the second sample object, and the subsequent consumption media of the second sample object.

[0151] The second sample object is the object used to train the causal inference model.

[0152] The True Retention Indicator (TRI) is a retention indicator that reflects the actual online retention of the second sample object. The method for obtaining the TRI can be found in the previous explanation regarding the annotation of retention indicators, and will not be repeated here.

[0153] The third object feature is the object feature of the second sample object, which can be a feature vector used to characterize the second sample object.

[0154] Among them, historical consumption media refers to media information in which the second sample subject engaged in consumption behavior (such as clicking, reading, or watching) within a past time period (e.g., within the past 3 days).

[0155] The third media feature is used to characterize the features of historical consumption media. The method for obtaining the third media feature can be found in the relevant explanation of the first media feature, and will not be repeated here.

[0156] Among them, subsequent consumption media refers to media information that the second sample subject consumes again in a subsequent time period, such as media information that the user clicks again in the next few days.

[0157] C2. Using the causal reasoning model to be trained, determine the retention media of the second sample object based on the characteristics of the third object and the characteristics of the third media.

[0158] C3. Using the causal reasoning model to be trained, predictions are made based on the media characteristics of the third object and the retention media of the second sample object to obtain the fifth retention indicator value of the second sample object.

[0159] The fifth retention indicator value is the retention indicator value predicted by the causal inference model to be trained based on the third object features of the second sample object and the third media features of each historical consumption media.

[0160] C4. The causal inference model to be trained is trained based on the loss between the true retention indicator and the fifth retention indicator to obtain the trained causal inference model.

[0161] For example, such as Figure 5 As shown, firstly, the true retention indicator value can be calculated (refer to...). Figure 5 "Subsequent consumption media" and the fifth retention indicator (reference) Figure 5 in "y RMILLosses between (reference) Figure 5 "L" RMIL Then, with the goal of minimizing the loss between the true retention indicator and the fifth retention indicator, the causal inference model to be trained is trained to obtain the trained causal inference model.

[0162] As can be seen, high-potential external information is introduced by using the loss between the true retention indicator and the fifth retention indicator to find media information with a high impact on user retention. Specifically: since the causal inference model minimizes the loss between the true retention indicator and the fifth retention indicator, and the true retention indicator is determined based on the subsequent media consumption and subsequent media exposure of the second sample object (i.e., the true retention indicator reflects the actual online retention of the second sample object), the causal inference model can use the media information of retained users in the next few days as high-potential external information to learn and find media information with a high impact on user retention. This media information with a high impact on user retention constitutes powerful causal media information for retention. Thus, by establishing a divergence function based on the fourth retention indicator and the labeled retention indicator, more accurate and reliable ahaitems can be found through contrast-based multi-instance learning guided by the JS distribution.

[0163] B2. Determine the divergence function of the retention prediction model to be trained using the fourth retention indicator value and the third retention indicator value.

[0164] For example, the divergence function can be as shown in Formula 5. The fourth retention indicator value and the labeled retention indicator value can be substituted into Formula 5 to determine the divergence bias function of the retention prediction model to be trained.

[0165]

[0166] In Formula 5, U is the user set, and y MIL It is the predicted output (i.e., the third retention indicator) based on a contrastive multi-instance learning framework, y RMIL It is the predicted output of the causal inference model (i.e., the fourth retention indicator value), and JS is the traditional JS divergence formula.

[0167] Step 305 may specifically include the following step 3051C:

[0168] 3051C. Based on the retention loss function and divergence function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0169] For example, with the goal of minimizing the retention loss function and minimizing the divergence function, the model parameters of the retention prediction model to be trained (such as the model parameters of the prediction layer and the model parameters of the aggregation layer) are adjusted until the training stops, and the trained retention prediction model is obtained.

[0170] It is evident that, given the significant differences between offline training and online services, introducing high-potential external information through the divergence function enhances the effective supervision and guidance of the learning process. This improves the credibility of the identified aha items that truly influence user retention. Consequently, the attention mechanism of the retention prediction model can be guided by contrastive learning to uncover the real reasons affecting user retention, thereby improving the accuracy of retention indicators for each consumed media of the recommended target.

[0171] (4) Use the retention loss function, retention bias function and divergence function for training. At this time, step 305 may specifically include the following step 3051D. The training process of the retention prediction model also includes steps A1 to A3 and steps B1 to B2.

[0172] 3051D, based on the divergence function, retention bias function and retention loss function, trains the retention prediction model to be trained, and obtains the trained retention prediction model.

[0173] For example, with the goal of minimizing the retention loss function, maximizing the retention bias function, and minimizing the divergence function, the model parameters of the retention prediction model to be trained (such as the model parameters of the prediction layer and the model parameters of the aggregation layer) are adjusted until the training stop condition is met, and the trained retention prediction model is obtained.

[0174] As can be seen from the above, by using the trained retention prediction model and the first object feature and the first media feature to predict the retention indicator value of each consumed media, the retention probability of the pair of the object to be recommended and the consumed media (i.e., the user and media information pair) can be aggregated, thereby predicting the media information that truly affects user retention (i.e., predicting the real reason for user retention). This avoids the problem of not being able to effectively find the real reasons for retention due to uncertainty and sparsity caused by directly using global feature modeling. Therefore, after sorting the consumed media by the retention indicator value of each consumed media, recommending the corresponding media information of each consumed media to the object to be recommended can prioritize the media information that truly affects the retention of consuming users, improve the accuracy of media information recommendation, and improve the overall retention rate of the media platform.

[0175] For ease of understanding, combined with Figure 4 , Figure 5 and Figure 6Taking the user account as the object to be recommended (in this case, the first sample object is the first sample account), the media information as a public account (i.e., consumed media is consumed public accounts, and consumed media samples are consumed public account samples), and recommending public accounts (such as public accounts of the same type that the user is interested in) to the user as an example, the media information recommendation process in this application embodiment is explained as follows: Figure 6 As shown, the media information recommendation process is as follows:

[0176] 601. Based on the second object features of the first sample account and the second media features of each consumed public account sample of the first sample account, train the retention prediction model to be trained to obtain the trained retention prediction model.

[0177] For example, taking training using the retention loss function, retention bias function, and divergence function as an example, the training process can specifically include:

[0178] (1.1) Obtain the annotation retention indication value of the first sample account.

[0179] (1.2) The retention prediction model to be trained is used to predict the retention index value of each consumed public account sample based on the second object feature of the first sample account and the second media feature of each consumed public account sample of the first sample account.

[0180] (1.3) By using the retention prediction model to be trained, the retention indicator values ​​of each consumed public account sample are aggregated to obtain the third retention indicator value of the first sample account.

[0181] (1.4) Based on the labeled retention indicator value and the third retention indicator value, determine the retention loss function of the retention prediction model to be trained.

[0182] (1.5) Based on the retention indication values ​​of the first consumer media and each consumed public account sample, obtain the first retention indication value of the first sample account; and based on the retention indication values ​​of the second consumer media and each consumed public account sample, obtain the second retention indication value of the first sample account.

[0183] (1.6) Based on the first retention indicator value and the second retention indicator value, construct the retention bias function of the retention prediction model to be trained.

[0184] (1.7) Using the trained causal reasoning model, the fourth retention indicator value of the first sample account is determined based on the subsequent consumption media of the first sample account and the samples of each consumed public account.

[0185] (1.8) Determine the divergence function of the retention prediction model to be trained by using the fourth retention indicator value and the labeled retention indicator value.

[0186] (1.9) Based on the divergence function, retention bias function and retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0187] For a detailed explanation of the training process in step 601, please refer to the relevant instructions in steps 301 to 305, which will not be repeated here.

[0188] 602. Obtain the first object characteristics of the user account and the first media characteristics of each public account consumed by the user account.

[0189] 603. Using the trained retention prediction model, predictions are made based on the first object feature and the first media feature to obtain the retention indicator value of each consumed public account.

[0190] 604. Sort the consumed public accounts according to their retention index values ​​to obtain a recommended ranking of the consumed public accounts.

[0191] 605. Based on recommendation ranking, recommend similar public accounts to user accounts that have already been consumed.

[0192] Furthermore, to verify the credibility of the media information (aha items) affecting user retention identified by the trained retention prediction model in this embodiment, and its impact on user retention, this embodiment conducted offline tests on two real datasets: Zhihurec and Lookalike. Zhihurec is an open-source question-answering dataset containing 7963 users and 81214 answer posts; the Lookalike offline dataset contains 2.5 million users and 1.5 million recommended articles. Related experimental tests were conducted, including:

[0193] 1. Verify the prediction accuracy of the models, as shown in Table 1. Table 1 shows the prediction accuracy of several models, including MLP, URO(AVG), URO(MIL), URO(MIL+MSS), URO(CMIL), and URO(RCMIL), on the zhihurec and Lookalike datasets. For each model in Table 1, the model is first used to predict user retention scores to determine whether the user will remain active in the next three days. Then, based on the actual online retention data for the next three days, the model's offline prediction accuracy for the next three days is calculated. The prediction accuracy of the models is calculated based on the AUC (Area Under Curve) metric.

[0194] Table 1

[0195]

[0196] Among them, the MLP model consists of only a few simple fully connected network layers; URO(MIL) contains only a multi-instance learning module and is used as the base model; URO(CMIL) is a multi-instance learning framework after introducing contrastive learning; URO(MIL+MSS) is URO(CMIL) with the addition of supervision signals on the basis of URO(CMIL); and URO(RCMIL) represents the URO(CMIL) model guided by the retained causal inference model.

[0197] The experimental results in Table 1 show that although the MLP model has a high offline prediction rate for three-day retention, it cannot be used to guide online systems to improve retention because the MLP model cannot generate an independent retention score for each item (i.e., media information). URO (AVG) assumes that each item has an equal impact on whether a user ultimately stays, while URO (MIL) assumes that the impact of each item on retention is not equal and aggregates the retention scores of each item through instance-level attention. Experimental results demonstrate that the impact of items on user retention is not equal, and the multi-instance learning framework that incorporates the number of clicks and impressions in the next three days as guidance performs better. Offline implementation results show that multi-instance learning with a contrastive learning mechanism has better retention prediction capabilities, and the contrastive multi-instance learning framework guided by the retention causal inference model has better robustness and confidence.

[0198] 2. Online user retention analysis was conducted on Lookalike for URO (RCMIL) deployment (see reference). Figure 4 The experimental results are shown in Table 2. Specifically, URO (RCMIL) was used on the Lookalike platform to predict the UI score (i.e., retention indicator value) of the user's consumed media. Figure 4 The "Retention Score" section reorders the user's top-k exposure list based on the UI score, ensuring that users with higher UI scores receive higher rankings / ranks in the user's top-k exposure list (see reference). Figure 4 (See the "Online Prediction" section in Table 2); The percentages in Table 2 represent the percentage increase in day-one retention and three-day retention compared to recommending media information to users before adjusting the sorting of the user's topk exposure list, using UI scores after adjusting the sorting of the user's topk exposure list.

[0199] Table 2

[0200]

[0201] As shown in Table 2, the experimental results demonstrate that introducing an attention mechanism based on contrastive learning and a causal reasoning model to mine retention-related items has a positive effect on online retention rate. It can effectively identify the real retention reasons with uncertainty and sparsity and effectively improve the retention rate.

[0202] To better implement the AI-based media information recommendation method in this application embodiment, this application embodiment also provides a media information recommendation device, which can be integrated into computer equipment, such as servers or terminals.

[0203] For example, such as Figure 7 As shown, Figure 7 This is a schematic diagram of an embodiment of the media information recommendation device in this application. The media information recommendation device may include an acquisition unit 701, a prediction unit 702, a sorting unit 703, a recommendation unit 704, etc., as follows:

[0204] The acquisition unit 701 is used to acquire the first object feature of the object to be recommended and the first media feature of each consumed media of the object to be recommended.

[0205] The prediction unit 702 is used to make predictions based on the first object features and the first media features through a trained retention prediction model to obtain the retention indicator value of each consumed media. The trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object.

[0206] The sorting unit 703 is used to sort each consumed media according to the retention indicator value of each consumed media to obtain the recommended sort of each consumed media.

[0207] Recommendation unit 704 is used to recommend media information corresponding to each consumed media to the object to be recommended based on the recommendation ranking.

[0208] In some embodiments, the media information recommendation device further includes a training unit (not shown in the figures), which is specifically used for:

[0209] Obtain the annotation retention indicator value of the first sample object;

[0210] The retention prediction model to be trained predicts the retention index value of each consumed media sample based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object.

[0211] By aggregating the retention indicator values ​​of each consumed media sample using the retention prediction model to be trained, the third retention indicator value of the first sample object is obtained.

[0212] Based on the labeled retention indicator and the third retention indicator, the retention loss function of the retention prediction model to be trained is determined;

[0213] Based on the retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0214] In some embodiments, the training unit is specifically used for:

[0215] Based on the weight parameters of the retention prediction model to be trained, each consumed media sample is divided into a first consumed media and a second consumed media. The weight parameters are used to indicate the weight value of each consumed media sample, and the weight value of the first consumed media is greater than the weight value of the second consumed media.

[0216] Based on the retention indication values ​​of the first consumer media and each consumed media sample, obtain the first retention indication value of the first sample object; and based on the retention indication values ​​of the second consumer media and each consumed media sample, obtain the second retention indication value of the first sample object.

[0217] Based on the first retention indicator value and the second retention indicator value, a retention bias function is constructed for the retention prediction model to be trained, wherein the retention bias function is used to increase the deviation between the first retention indicator value and the second retention indicator value;

[0218] Based on the retention bias function and the retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0219] In some embodiments, the training unit is specifically used for:

[0220] Based on the trained causal reasoning model, the fourth retention indicator value of the first sample object is determined according to the subsequent consumed media and each consumed media sample of the first sample object.

[0221] The divergence function of the retention prediction model to be trained is determined by using the fourth retention indicator value and the labeled retention indicator value.

[0222] Based on the divergence function, retention bias function, and retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0223] In some embodiments, the training unit is specifically used for:

[0224] By using a trained causal reasoning model, a fourth retention indicator value for the first sample object is obtained based on predictions made from each consumed media sample.

[0225] The divergence function of the retention prediction model to be trained is determined by using the fourth retention indicator value and the third retention indicator value.

[0226] Based on the retention loss function and divergence function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

[0227] In some embodiments, the training unit is specifically used for:

[0228] Based on the trained causal reasoning model and each consumed media sample, the retained media of the first sample object is determined.

[0229] By using a trained causal reasoning model, a fourth retention indicator value for the first sample object is obtained by predicting based on the characteristics of the second object and the media characteristics of the retained media of the first sample object.

[0230] In some embodiments, the training unit is specifically used for:

[0231] Obtain the true retention indicator value of the second sample object, the third object characteristics of the second sample object, the third media characteristics of each historical consumption media of the second sample object, and the subsequent consumption media of the second sample object;

[0232] Using the causal reasoning model to be trained, the retention media of the second sample object is determined based on the characteristics of the third object and the characteristics of the third media.

[0233] By using the causal reasoning model to be trained, predictions are made based on the media characteristics of the retained media of the third object and the second sample object, and the fifth retention indicator value of the second sample object is obtained.

[0234] The causal inference model to be trained is trained based on the loss between the true retention indicator and the fifth retention indicator to obtain the trained causal inference model.

[0235] In some embodiments, the training unit is specifically used for:

[0236] Obtain the subsequent media clicks of the first sample object;

[0237] Obtain the subsequent media exposure of the first sample object;

[0238] Based on subsequent media clicks and subsequent media exposures, the label retention indicator value for the first sample object is determined.

[0239] In some embodiments, each consumed media includes target consumed media whose retention indicator value meets predetermined conditions, and the sorting unit 703 is specifically used for:

[0240] From each consumed media, identify target consumed media whose retention indicator values ​​meet predetermined conditions;

[0241] Based on the retention indicator value of the target consumed media, determine the recommendation ranking of the target consumed media;

[0242] In some embodiments, the recommendation unit 704 is specifically used for:

[0243] Based on the recommended ranking of the target media already consumed, recommend media information corresponding to the target media already consumed to the target audience.

[0244] In some embodiments, the recommendation unit 704 is specifically used for:

[0245] Obtain the media types of the target media that have been consumed;

[0246] Obtain media information for different media types to serve as the media information corresponding to the target consumed media;

[0247] Based on the recommended ranking of media consumed by the target audience, recommend media information of media types to the target audience.

[0248] In some embodiments, the sorting unit 703 is specifically used for:

[0249] Obtain the click-through rate of each consumed media outlet;

[0250] The recommended ranking of each consumed media is determined based on the retention indicator and click-through rate of each consumed media.

[0251] Therefore, the media information recommendation device provided in this application embodiment can bring the following technical effects: by using a trained retention prediction model, the retention indicator value of each consumed media is predicted using the first object feature and the first media feature; the retention probability of the pair of the object to be recommended and the consumed media (i.e., the user and media information pair) can be aggregated, thereby predicting the media information that truly affects user retention (i.e., predicting the real reason for user retention), avoiding the problem of not being able to effectively find the real reason for retention with uncertainty and sparsity due to directly using global feature modeling; therefore, after sorting each consumed media using the retention indicator value of each consumed media, the media information corresponding to each consumed media is recommended to the object to be recommended, so that the media information that truly affects the retention of consuming users can be recommended to the object to be recommended first, thereby improving the accuracy of media information recommendation and improving the overall retention rate of the media platform.

[0252] Furthermore, to better implement the AI-based media information recommendation method in the embodiments of this application, this application also provides a computer device, such as... Figure 8 As shown, it illustrates a structural schematic diagram of the computer device involved in the embodiments of this application, specifically:

[0253] The computer device may include components such as a processor 801 with one or more processing cores, a memory 802 with one or more computer-readable storage media, a power supply 803, and an input unit 804. Those skilled in the art will understand that... Figure 8 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0254] The processor 801 is the control center of the computer device. It connects various parts of the computer device via various interfaces and lines. By running or executing software programs and / or modules stored in the memory 802, and by calling data stored in the memory 802, it performs various functions of the computer device and processes data, thereby performing overall detection of the computer device. Optionally, the processor 801 may include one or more processing cores; preferably, the processor 801 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into the processor 801.

[0255] The memory 802 can be used to store software programs and modules. The processor 801 executes various functional applications and data processing by running the software programs and modules stored in the memory 802. The memory 802 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, video playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 802 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 802 may also include a memory controller to provide the processor 801 with access to the memory 802.

[0256] The computer device also includes a power supply 803 that supplies power to the various components. Preferably, the power supply 803 can be logically connected to the processor 801 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 803 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0257] The computer device may also include an input unit 804, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0258] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 801 in the computer device loads the executable files corresponding to the processes of one or more application programs into the memory 802 according to the following instructions, and the processor 801 runs the application programs stored in the memory 802 to realize various functions, as follows:

[0259] Obtain the first object characteristics of the object to be recommended and the first media characteristics of each consumed media of the object to be recommended;

[0260] By using a trained retention prediction model, a retention indicator value for each consumed media is obtained based on the first object feature and the first media feature. The trained retention prediction model is learned based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object.

[0261] The consumed media are sorted according to their retention index values ​​to obtain a recommended ranking of the consumed media.

[0262] Based on the recommendation ranking, media information corresponding to each consumed media is recommended to the target audience.

[0263] For details of each of the above operations, please refer to the previous embodiments, which will not be repeated here.

[0264] Therefore, the computer device in this embodiment can bring the following technical effects: by using a trained retention prediction model, the retention indicator value of each consumed media is predicted using the first object feature and the first media feature; the retention probability of the pair of recommended objects and consumed media (i.e., the user and media information pair) can be aggregated, thereby predicting the media information that truly affects user retention (i.e., predicting the real reason for user retention), avoiding the problem of not being able to effectively find the real reason for retention due to uncertainty and sparsity caused by directly using global feature modeling; therefore, after sorting each consumed media using the retention indicator value of each consumed media, the media information corresponding to each consumed media is recommended to the recommended object, so that the media information that truly affects the retention of consuming users can be recommended to the recommended object first, improving the accuracy of media information recommendation and improving the overall retention rate of the media platform.

[0265] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0266] Therefore, embodiments of this application provide a computer-readable storage medium storing a computer program that can be loaded by a processor to execute the steps of any of the object detection methods provided in the embodiments of this application. For example, the computer program can execute the following steps:

[0267] Obtain the first object characteristics of the object to be recommended and the first media characteristics of each consumed media of the object to be recommended;

[0268] By using a trained retention prediction model, a retention indicator value for each consumed media is obtained based on the first object feature and the first media feature. The trained retention prediction model is learned based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object.

[0269] The consumed media are sorted according to their retention index values ​​to obtain a recommended ranking of the consumed media.

[0270] Based on the recommendation ranking, media information corresponding to each consumed media is recommended to the target audience.

[0271] As can be seen, the computer program can be loaded by the processor to execute the steps in any of the artificial intelligence-based media information recommendation methods provided in the embodiments of this application. Therefore, the computer-readable storage medium of the embodiments of this application can bring the following technical effects: By using a trained retention prediction model, the retention indicator value of each consumed media is predicted using the first object feature and the first media feature; the retention probability of the pair of objects to be recommended and consumed media (i.e., the user and media information pair) can be aggregated, thereby predicting the media information that truly affects user retention (i.e., predicting the true reason for user retention), avoiding the problem of failing to effectively find the true retention reason with uncertainty and sparsity due to directly using global feature modeling; therefore, after sorting each consumed media using the retention indicator value of each consumed media, recommending the media information corresponding to each consumed media to the object to be recommended can prioritize recommending the media information that truly affects the retention of consuming users to the object to be recommended, improving the accuracy of media information recommendation and increasing the overall retention rate of the media platform.

[0272] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0273] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0274] According to the AI-based media information recommendation method of this application, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above embodiments.

[0275] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process and beneficial effects of the media information recommendation device, computer-readable storage medium, computer equipment and its corresponding units described above can be referred to the description of the artificial intelligence-based media information recommendation method in the above embodiments, and will not be repeated here.

[0276] The foregoing has provided a detailed description of an artificial intelligence-based media information recommendation method, apparatus, computer device, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A media information recommendation method based on artificial intelligence, characterized in that, The method includes: Obtain the first object feature of the object to be recommended and the first media feature of each consumed media of the object to be recommended; By using a trained retention prediction model, based on the first object features and the first media features, a retention indicator value for each consumed media is obtained. The trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object. The trained retention prediction model is trained through the following steps: Obtain the labeled retention indicator value of the first sample object; using the retention prediction model to be trained, predict the retention indicator value of each consumed media sample based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object; aggregate the retention indicator values ​​of each consumed media sample using the retention prediction model to be trained to obtain the third retention indicator value of the first sample object; determine the retention loss function of the retention prediction model to be trained based on the labeled retention indicator value and the third retention indicator value; divide the consumed media samples based on the weight parameters of the retention prediction model to be trained to obtain the first consumed media and the second consumed media, wherein the weight parameters are used to indicate the retention loss function of each consumed media sample. The weight values ​​of the samples are determined, with the weight value of the first consumed media being greater than that of the second consumed media; based on the retention indicator values ​​of the first consumed media and each consumed media sample, a first retention indicator value for the first sample object is obtained; and based on the retention indicator values ​​of the second consumed media and each consumed media sample, a second retention indicator value for the first sample object is obtained; based on the first retention indicator value and the second retention indicator value, a retention bias function for the retention prediction model to be trained is constructed, wherein the retention bias function is used to increase the deviation between the first retention indicator value and the second retention indicator value; based on the retention bias function and the retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model; The consumed media are sorted according to their retention index values ​​to obtain a recommended ranking of the consumed media. Based on the recommendation ranking, media information corresponding to each consumed media is recommended to the object to be recommended.

2. The media information recommendation method based on artificial intelligence according to claim 1, characterized in that, The method further includes: Based on the subsequent consumption media of the first sample object and each consumed media sample, a fourth retention indicator value of the first sample object is determined using a trained causal reasoning model. The divergence function of the retention prediction model to be trained is determined by the fourth retention indicator value and the labeled retention indicator value. The step of training the retention prediction model to be trained based on the retention bias function and the retention loss function to obtain the trained retention prediction model includes: The retention prediction model to be trained is trained based on the divergence function, the retention bias function, and the retention loss function to obtain the trained retention prediction model.

3. The media information recommendation method based on artificial intelligence according to claim 1, characterized in that, The method further includes: By using a trained causal reasoning model, a fourth retention indicator value for the first sample object is obtained based on each consumed media sample; The divergence function of the retention prediction model to be trained is determined by the fourth retention indicator value and the third retention indicator value. The step of training the retention prediction model to be trained based on the retention loss function to obtain the trained retention prediction model includes: Based on the retention loss function and the divergence function, the retention prediction model to be trained is trained to obtain the trained retention prediction model.

4. The media information recommendation method based on artificial intelligence according to claim 3, characterized in that, The step of predicting the fourth retention indicator value of the first sample object based on each consumed media sample using a trained causal inference model includes: Based on the trained causal reasoning model and the consumed media samples, the retained media of the first sample object is determined. By using a trained causal reasoning model, a fourth retention indicator value for the first sample object is obtained by predicting based on the characteristics of the second object and the media characteristics of the retention media of the first sample object.

5. The media information recommendation method based on artificial intelligence according to claim 4, characterized in that, The method further includes: Obtain the true retention indicator value of the second sample object, the third object characteristics of the second sample object, the third media characteristics of each historical consumption media of the second sample object, and the subsequent consumption media of the second sample object; Based on the characteristics of the third object and the characteristics of the third media, the retention media of the second sample object is determined using the causal reasoning model to be trained. By using the causal reasoning model to be trained, based on the media characteristics of the retained media of the third object and the second sample object, a prediction is made to obtain the fifth retention indicator value of the second sample object. The causal inference model to be trained is trained based on the loss between the true retention indicator value and the fifth retention indicator value to obtain the trained causal inference model.

6. The media information recommendation method based on artificial intelligence according to claim 1, characterized in that, The step of obtaining the labeled retention indication value of the first sample object includes: Obtain the subsequent media clicks of the first sample object; Obtain the subsequent media exposure of the first sample object; Based on the subsequent media clicks and subsequent media exposures, the annotation retention indicator value of the first sample object is determined.

7. The media information recommendation method based on artificial intelligence according to claim 1, characterized in that, Each consumed media includes target consumed media whose retention indicator values ​​meet predetermined conditions. The step of sorting the consumed media according to their retention indicator values ​​to obtain a recommended ranking of the consumed media includes: From each of the consumed media, obtain the target consumed media whose retention indicator value meets the predetermined conditions; Based on the retention indicator value of the target consumed media, determine the recommended ranking of the target consumed media; The step of recommending media information corresponding to each consumed media to the object to be recommended based on the recommendation ranking includes: Based on the recommended ranking of the target consumed media, media information corresponding to the target consumed media is recommended to the object to be recommended.

8. The media information recommendation method based on artificial intelligence according to claim 7, characterized in that, The step of recommending media information corresponding to the target consumed media to the target object according to the recommended ranking of the target consumed media includes: Obtain the media type of the target consumed media; Obtain media information of the media type as the media information corresponding to the target consumed media; Based on the recommended ranking of the target consumed media, media information of the media type is recommended to the target to be recommended.

9. The media information recommendation method based on artificial intelligence according to claim 1, characterized in that, The step of sorting the consumed media according to their retention indicators to obtain a recommended ranking of the consumed media includes: Obtain the click-through rate of each of the consumed media; The recommended ranking of each consumed media is determined based on the retention index and the click-through rate of each consumed media.

10. A media information recommendation device, characterized in that, The media information recommendation device includes: The acquisition unit is used to acquire the first object feature of the object to be recommended and the first media feature of each consumed media of the object to be recommended. The prediction unit is configured to predict, based on the first object features and the first media features, the retention indication value of each consumed media using a trained retention prediction model. The trained retention prediction model is learned based on the second object features of the first sample object and the second media features of each consumed media sample of the first sample object. The trained retention prediction model is trained through the following steps: Obtain the labeled retention indicator value of the first sample object; using the retention prediction model to be trained, predict the retention indicator value of each consumed media sample based on the second object feature of the first sample object and the second media feature of each consumed media sample of the first sample object; aggregate the retention indicator values ​​of each consumed media sample using the retention prediction model to be trained to obtain the third retention indicator value of the first sample object; determine the retention loss function of the retention prediction model to be trained based on the labeled retention indicator value and the third retention indicator value; divide the consumed media samples based on the weight parameters of the retention prediction model to be trained to obtain the first consumed media and the second consumed media, wherein the weight parameters are used to indicate the retention loss function of each consumed media sample. The weight values ​​of the samples are determined, with the weight value of the first consumed media being greater than that of the second consumed media; based on the retention indicator values ​​of the first consumed media and each consumed media sample, a first retention indicator value for the first sample object is obtained; and based on the retention indicator values ​​of the second consumed media and each consumed media sample, a second retention indicator value for the first sample object is obtained; based on the first retention indicator value and the second retention indicator value, a retention bias function for the retention prediction model to be trained is constructed, wherein the retention bias function is used to increase the deviation between the first retention indicator value and the second retention indicator value; based on the retention bias function and the retention loss function, the retention prediction model to be trained is trained to obtain the trained retention prediction model; The sorting unit is used to sort the consumed media according to the retention indication value of each consumed media to obtain the recommended sort of the consumed media. The recommendation unit is used to recommend media information corresponding to each consumed media to the object to be recommended based on the recommendation ranking.

11. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the artificial intelligence-based media information recommendation method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1-9.

13. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the artificial intelligence-based media information recommendation method as described in any one of claims 1 to 9.

Images