Content recommendation method and apparatus, electronic device, storage medium, and program product

Abstract

The application discloses a content recommendation method and device, an electronic device, a storage medium and a program product; the application determines a plurality of to-be-recommended contents based on a content recommendation request of a to-be-recommended object, extracts interaction feature information of the to-be-recommended contents in a current recommendation scene, the current recommendation scene is a scene of content recommendation to the to-be-recommended object; obtains historical interaction feature information of the to-be-recommended contents in a historical recommendation scene; based on the current interaction feature information and the historical interaction feature information, predicts interaction probability information of the to-be-recommended object producing an interaction behavior on the to-be-recommended content; determines target recommendation content from the plurality of to-be-recommended contents according to the interaction probability information; and recommends the target recommendation content to the to-be-recommended object. In the application, the target recommendation content that may trigger the to-be-recommended object to produce an interaction behavior can be screened out. Therefore, the application can improve the accuracy of content recommendation.

Description

Technical Field

[0001] This application relates to the field of computers, and specifically to a content recommendation method, apparatus, electronic device, storage medium, and program product. Background Technology

[0002] With the rapid development of internet technology, content promotion via the internet has become one of the main channels for information dissemination. This promotion method can proactively push information to users, allowing them to access relevant content without actively searching.

[0003] However, due to the vast size and diverse needs of the internet user base, the current indiscriminate content promotion strategy inevitably reaches a large number of non-target audiences during information dissemination. While this broad coverage helps expand the reach of information, it also leads to a decrease in the precision of the promotion. This decrease in precision not only wastes promotional resources but may also affect user experience, thereby reducing the overall promotional effectiveness. Summary of the Invention

[0004] This application provides a content recommendation method, apparatus, electronic device, storage medium, and program product, which can filter out target recommendation content that may trigger interactive behavior in the target user and accurately recommend the target content to the target user. This can improve the accuracy of content recommendation, save promotional resources, enhance user experience, and thus further improve the overall promotional effect.

[0005] This application provides a content recommendation method, including:

[0006] Based on the content recommendation request of the object to be recommended, multiple pieces of content to be recommended are identified, and the interaction feature information of the content to be recommended in the current recommendation scenario is extracted. The current recommendation scenario is a scenario of recommending content to the object to be recommended.

[0007] Obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios;

[0008] Based on current and historical interaction features, predict the probability of the recommended object interacting with the recommended content.

[0009] Based on interaction probability information, the target recommended content is determined from multiple content to be recommended;

[0010] Recommend target content to the target audience.

[0011] This application also provides a content recommendation device, including:

[0012] The information processing unit is used to determine multiple recommended contents based on the content recommendation request of the recommended object, and extract the interaction feature information of the recommended contents in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the recommended object;

[0013] The information acquisition unit is used to acquire historical interaction feature information of the content to be recommended in historical recommendation scenarios;

[0014] The interaction prediction unit is used to predict the probability of the recommended object interacting with the recommended content based on the current interaction feature information and historical interaction feature information.

[0015] The recommendation determination unit is used to determine the target recommended content from multiple content to be recommended based on interaction probability information;

[0016] The content recommendation unit is used to recommend target content to the target audience.

[0017] In some embodiments, the information processing unit includes an information acquisition subunit, an attribute feature extraction subunit, an environmental feature extraction subunit, a content feature extraction subunit, and an information determination subunit:

[0018] The information acquisition subunit is used to acquire the object attribute information of the object to be recommended and the interaction context information of the object to be recommended in the current recommendation scenario.

[0019] The attribute feature extraction subunit is used to perform feature extraction processing on object attribute information to obtain the object attribute features of the object to be recommended.

[0020] The environmental feature extraction subunit is used to extract features from the interaction context information to obtain the environmental features of the object to be recommended.

[0021] The content feature extraction subunit is used to extract features from the content to be recommended, thereby obtaining content feature information.

[0022] The information determination subunit is used to determine the interaction feature information of the content to be recommended in the current recommendation scenario based on object attribute features, environmental features, and content feature information.

[0023] In some embodiments, the content feature extraction subunit includes a feature extraction module and a feature enhancement module:

[0024] The feature extraction module is used to extract features from the content to be recommended, and obtain initial content feature information;

[0025] The feature enhancement module is used to perform feature enhancement processing on the initial content feature information to obtain the content feature information.

[0026] In some embodiments, the information acquisition unit is configured to:

[0027] Determine the historical feature storage area corresponding to the content to be recommended. The historical feature storage area stores multiple interaction feature information of the content to be recommended in historical recommendation scenarios.

[0028] Retrieve multiple interaction feature information from historical recommendation scenarios from the historical feature storage area;

[0029] By fusing multiple interaction feature information from historical recommendation scenarios, we can obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0030] In some embodiments, after extracting the interaction feature information of the content to be recommended in the current recommendation scenario, the device further includes an information storage unit for:

[0031] The interaction feature information of the current recommendation scenario is stored in the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area and obtain the updated historical feature storage area. The updated historical feature storage area is used for the content recommendation scenario of the content to be recommended in the future time period.

[0032] In some embodiments, the storage structure of the historical feature storage area corresponding to the content to be recommended is a queue;

[0033] Information storage unit, used for:

[0034] Determine the current remaining storage space of the historical feature storage area corresponding to the content to be recommended;

[0035] Based on the remaining storage space and the target storage space required for the interaction feature information in the current recommendation scenario, the interaction feature information in the historical recommendation scenario in the historical feature storage area is deleted in a first-in-first-out order to obtain the processed historical feature storage area.

[0036] The interaction feature information of the current recommendation scenario is stored in the processed historical feature storage area.

[0037] In some embodiments, the interactive prediction unit is configured to:

[0038] The target interaction feature information is obtained by fusing the current interaction feature information and the historical interaction feature information.

[0039] Based on the target interaction feature information, the interaction probability information of the object to be recommended and the content to be recommended is predicted in multiple interaction type operation tasks.

[0040] In some embodiments, the information processing unit is configured to:

[0041] By using a content processing model, interactive feature information of the content to be recommended in the current recommendation scenario is extracted;

[0042] Before extracting the interaction feature information of the content to be recommended in the current recommendation scenario through the content processing model, the device also includes a content processing model training unit, used for:

[0043] Obtain the first training data, which includes the interaction label information between the sample recommendation object and the sample content, the object attribute information of the sample recommendation object, the interaction context information, and the content information of the sample content;

[0044] Through the content processing model, feature extraction is performed on the object attribute information, interaction context information, and content information of the sample recommendation object, respectively, to obtain sample object attribute features, sample environment features, and sample content feature information.

[0045] By using a content processing model, feature interaction is performed on sample object attribute features, sample environment features, and sample content features to obtain sample interaction feature information.

[0046] Based on the sample interaction feature information, predict the sample interaction probability information of the sample recommendation object generating interactive behavior on the sample content;

[0047] Based on the sample interaction probability information and interaction tag information, the model parameters of the content processing model are adjusted to obtain the trained content processing model.

[0048] In some embodiments, the content processing model training unit is used to perform feature extraction processing on the content information of the sample content, including:

[0049] By using a content processing model, feature extraction is performed on the content information of the sample content to obtain initial sample feature information;

[0050] The content information of the sample is processed to obtain semantic residual features;

[0051] The semantic residual features and the initial sample feature information are fused to obtain the sample content feature information.

[0052] In some embodiments, the feature extraction module is used for:

[0053] By using a content feature extraction model, features are extracted from the content to be recommended to obtain initial content feature information;

[0054] Before obtaining initial content feature information by extracting features from the content to be recommended using a content feature extraction model, the device also includes a content feature extraction model training unit, used for:

[0055] Obtain the second training data, which includes multiple sample contents, as well as positive and negative sample contents.

[0056] Using a content feature extraction model, feature extraction is performed on the positive and negative sample content of the sample content and the sample recommended content, respectively, to obtain the semantic features of the sample content, the semantic features of the positive sample content, and the semantic features of the negative sample content.

[0057] Based on the semantic features of the sample content and the semantic features of the positive sample content, the positive similarity between the sample content and the positive sample content is determined, and based on the semantic features of the positive sample content and the semantic features of the negative sample content, the negative similarity between the sample content and the negative sample content is determined.

[0058] Based on the positive and negative similarities of multiple sample contents, the model parameters of the content feature extraction model are adjusted to obtain the trained content feature extraction model.

[0059] In some embodiments, the apparatus further includes a model evaluation unit for:

[0060] Obtain the performance evaluation dataset, which includes multiple evaluation items;

[0061] The trained content feature extraction model is used to extract features from the evaluation content to obtain its semantic features.

[0062] Based on the semantic features of the evaluation content and the semantic features of the reference evaluation content, similar evaluation content that is similar to the evaluation content is determined from the reference evaluation content. The reference evaluation content refers to other evaluation content outside the evaluation content.

[0063] Based on the content information of the evaluation content and the content information of similar evaluation content, the similarity between the evaluation content and the similar evaluation content is evaluated to obtain information on the degree of similarity between the similar evaluation content and the evaluation content.

[0064] Based on the similarity information corresponding to similar assessment content, the target assessment content is determined from the similar assessment content;

[0065] Based on the number of similar evaluation contents corresponding to multiple evaluation contents and the number of target evaluation contents corresponding to multiple evaluation contents, the evaluation result information of the trained content feature extraction model is determined, so as to adjust the trained content feature extraction model according to the evaluation result information.

[0066] This application also provides an electronic device, including a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to execute the steps in any of the content recommendation methods provided in this application.

[0067] This application also provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to perform the steps in any of the content recommendation methods provided in this application.

[0068] This application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in any of the content recommendation methods provided in this application.

[0069] This application embodiment can determine multiple pieces of content to be recommended based on the content recommendation request of the object to be recommended, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the object to be recommended; obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios; predict the interaction probability information of the object to be recommended generating interaction behavior on the content to be recommended based on the current interaction feature information and the historical interaction feature information; determine the target recommended content from the multiple pieces of content to be recommended based on the interaction probability information; and recommend the target recommended content to the object to be recommended.

[0070] In this application, multiple pieces of content to be recommended can be identified based on the content recommendation request of the target user. Then, the interaction feature information of these pieces of content in the current recommendation scenario is extracted, while historical interaction feature information of each piece of content in historical scenarios is also obtained. Combining the current and historical interaction feature information allows for a better characterization of the content to be recommended, thereby more accurately predicting the probability of the target user interacting with each piece of content. This helps to filter out target content that is likely to induce interaction with the target user, and then accurately recommends this target content to the target user. This improves the accuracy of content recommendation, saves promotional resources, enhances user experience, and ultimately improves the overall promotional effect. Attached Figure Description

[0071] 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.

[0072] Figure 1a This is a schematic diagram of a scenario illustrating the content recommendation method provided in an embodiment of this application;

[0073] Figure 1b This is a flowchart illustrating the content recommendation method provided in an embodiment of this application;

[0074] Figure 2aThis is a flowchart illustrating the application of the content recommendation method provided in this application embodiment in an advertising recommendation scenario;

[0075] Figure 2b This is a schematic diagram of a scenario for the training content processing model provided in an embodiment of this application;

[0076] Figure 2c This is a schematic diagram of a scenario for the training content feature extraction model provided in an embodiment of this application;

[0077] Figure 2d This is a schematic diagram illustrating the application of the content recommendation method provided in this application embodiment in a personalized recommendation scenario;

[0078] Figure 3 This is a schematic diagram of the structure of the content recommendation device provided in the embodiments of this application;

[0079] Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0080] 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.

[0081] This application provides a content recommendation method, apparatus, electronic device, storage medium, and program product.

[0082] Specifically, the content recommendation device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0083] In some embodiments, the content recommendation device can also be integrated into multiple electronic devices. For example, the content recommendation device can be integrated into multiple servers, and the content recommendation method of this application can be implemented by multiple servers.

[0084] In some embodiments, the terminal can also be used as a server to implement some or all of the functions of a server.

[0085] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0086] It is understood that in the specific implementation of this application, data related to the objects to be recommended are involved. When the 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.

[0087] For example, refer to Figure 1a This electronic device can determine multiple pieces of content to be recommended based on the content recommendation request of the object to be recommended, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, which is a scenario of recommending content to the object to be recommended; obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios; predict the interaction probability information of the object to be recommended on the content to be recommended based on the current interaction feature information and the historical interaction feature information; determine the target recommended content from the multiple pieces of content to be recommended based on the interaction probability information; and recommend the target recommended content to the object to be recommended.

[0088] This application can identify multiple pieces of content to be recommended based on the content recommendation request of the target user. Next, it extracts the interaction feature information of these pieces of content in the current recommendation scenario, and simultaneously obtains the historical interaction feature information of each piece of content in historical scenarios. Combining the current and historical interaction feature information allows for a better characterization of the content to be recommended, thereby more accurately predicting the probability of the target user interacting with each piece of content. This helps to filter out target content that is likely to induce interaction with the target user, and then accurately recommends this target content to the target user. This improves the accuracy of content recommendation, saves promotional resources, enhances user experience, and ultimately improves the overall promotional effect.

[0089] The following sections provide detailed descriptions of each example. It should be noted that the sequence numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0090] In this embodiment, a content recommendation method is provided, such as... Figure 1b As shown, the specific process of this content recommendation method can be summarized as follows:

[0091] 101. Based on the content recommendation request of the object to be recommended, determine multiple pieces of content to be recommended, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the object to be recommended.

[0092] The "object to be recommended" refers to the entity that is waiting to receive recommended content.

[0093] For example, the target audience for recommendations could be users browsing web pages, watching videos, reading e-articles, or users using applications, engaging in social media interactions, or searching for products on e-commerce platforms, and so on. A content recommendation request is a request to send recommended content to the target audience.

[0094] For example, content recommendation requests can be generated while the person to be recommended is browsing web pages, watching videos, or reading e-articles, or they can be triggered when the person to be recommended is using an application, engaging in social media interactions, or conducting product searches.

[0095] Content to be recommended refers to content that needs to be pushed out.

[0096] For example, the types of content to be recommended can be advertisements, articles, videos, music, products, social media posts, trending topics, promotional activities, etc.

[0097] The current recommendation scenario is one where content is recommended to the target audience.

[0098] For example, the current recommendation scenario can refer to the process of pushing content to the object to be recommended, or it can refer to the specific scenario in which the object to be recommended is waiting to receive the recommended content. This specific scenario can include not only the basic characteristics and historical behavior data of the object to be recommended, but also environmental factors such as the time and location of the recommendation, the type of device receiving the recommended content and its network information, etc.

[0099] Interactive feature information is the information obtained after the content to be recommended interacts with the object to be recommended in the current recommendation scenario.

[0100] In some embodiments, in order to quickly screen out content matching the object to be recommended, thereby reducing the computational burden in the subsequent content selection process, multiple pieces of content to be recommended are determined based on the content recommendation request of the object to be recommended, including:

[0101] Based on the content recommendation request of the object to be recommended, obtain multiple pieces of content that match the object to be recommended.

[0102] In some embodiments, since the performance of the content to be recommended may vary in different recommendation scenarios, in order to achieve accurate recommendation of the content to be recommended, the scenario of recommending the content to be recommended to the target audience is considered, so as to better select the content to be recommended that matches the current recommendation scenario, improve the relevance and effectiveness of the recommendation, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, including:

[0103] Obtain the object attribute information of the object to be recommended and the interaction context information of the object to be recommended in the current recommendation scenario;

[0104] The object attribute information is processed by feature extraction to obtain the object attribute features of the object to be recommended;

[0105] Feature extraction processing is performed on the interaction context information to obtain the environmental features of the object to be recommended;

[0106] Feature extraction is performed on the recommended content to obtain content feature information;

[0107] Based on object attribute features, environmental features, and content features, determine the interaction features of the content to be recommended in the current recommendation scenario.

[0108] Among them, object attribute information is used to describe the relevant data of the object to be recommended. This relevant data may include the basic information of the object to be recommended, historical behavior data, etc.

[0109] Interactive contextual information refers to the contextual information of how the recommended content is received by the target device in the current recommendation scenario. For example, interactive contextual information may include the time and location of the recommendation, as well as environmental factors such as the type of device the target device uses to receive the recommended content and its network information.

[0110] Object attribute features refer to the data obtained after feature extraction of the object attribute information of the object to be recommended. They can effectively describe the object attribute information of the object to be recommended.

[0111] Environmental features refer to the data obtained after feature extraction of the interaction context information of the object to be recommended. They can effectively describe the interaction context information of the object to be recommended in the current recommendation scenario.

[0112] Content feature information refers to the data obtained after feature extraction of the content to be recommended. It can reflect the interaction preferences of the content to be recommended in historical recommendation scenarios and is helpful to provide valuable reference for the current recommendation scenario.

[0113] In some embodiments, considering the potential changes in the content to be recommended, in order to ensure that the current content to be recommended can identify potential interaction preferences after feature extraction, thereby achieving more accurate recommendations, feature extraction is performed on the content to be recommended to obtain content feature information, including:

[0114] Feature extraction is performed on the content to be recommended to obtain initial content feature information;

[0115] The initial content feature information is subjected to feature enhancement processing to obtain the content feature information.

[0116] Among them, the initial content feature information is the semantic information of the content to be recommended, which can reflect the characteristics of the content to be recommended in terms of theme, content, meaning, relevance, etc.

[0117] Content feature information is the information obtained by enhancing the semantic features of the content to be recommended, so that it can reflect the interaction preferences of the content to be recommended in historical recommendation scenarios.

[0118] In some embodiments, to enable the content processing model to learn the potential interaction preferences of the content to be recommended, thereby ensuring that the current content to be recommended can accurately identify these preferences after feature extraction, the interaction feature information of the content to be recommended in the current recommendation scenario is extracted, including:

[0119] By using a content processing model, interactive feature information of the content to be recommended in the current recommendation scenario is extracted;

[0120] Before extracting the interaction feature information of the content to be recommended in the current recommendation scenario through the content processing model, it also includes:

[0121] Obtain the first training data, which includes the interaction label information between the sample recommendation object and the sample content, the object attribute information of the sample recommendation object, the interaction context information, and the content information of the sample content;

[0122] Through the content processing model, feature extraction is performed on the object attribute information, interaction context information, and content information of the sample recommendation object, respectively, to obtain sample object attribute features, sample environment features, and sample content feature information.

[0123] By using a content processing model, feature interaction is performed on sample object attribute features, sample environment features, and sample content features to obtain sample interaction feature information.

[0124] Based on the sample interaction feature information, predict the sample interaction probability information of the sample recommendation object generating interactive behavior on the sample content;

[0125] Based on the sample interaction probability information and interaction tag information, the model parameters of the content processing model are adjusted to obtain the trained content processing model.

[0126] Content processing models are models capable of extracting interactive feature information of the content to be recommended in the current recommendation scenario. For example, content processing models can be deep learning models, hybrid models, embedding-based models, feature engineering-based models, collaborative filtering models, etc.

[0127] The first training data is used to train the content processing model so that the model can extract features that reflect the interaction preferences of the content to be recommended in its historical recommendation scenarios, and make accurate predictions based on these features, which is beneficial to provide higher accuracy for subsequent recommendations.

[0128] Sample recommendation objects refer to objects that have been recommended in historical recommendation scenarios. The interaction records between these objects and their corresponding sample content can be used to train the content processing model. By analyzing these interaction records between sample recommendation objects and sample content, the content processing model can learn how to better understand and predict users' preferences for different content. Sample content refers to the content recommended to sample recommendation objects in historical recommendation scenarios. For example, the types of sample content can be advertisements, articles, videos, music, products, social media posts, trending topics, promotional activities, etc.

[0129] Interaction tag information refers to the tag information corresponding to the interactive behavior of the sample recommendation object on the sample content. For example, the interaction tag information can be 1 or 0, where 1 indicates that the sample recommendation object has generated an interactive behavior on the sample content, and 0 indicates that the sample recommendation object has not generated an interactive behavior on the sample content.

[0130] Interactive behavior refers to the interactive operations performed by the sample recommendation object on the sample recommendation object. For example, interactive behavior can be clicking, browsing, viewing time, saving, liking, sharing, purchasing, etc.

[0131] The object attribute information of the sample recommendation object is the relevant data used to describe the sample recommendation object. This relevant data may include the basic information of the sample recommendation object, historical behavior data, etc.

[0132] The contextual information of a sample recommendation object refers to the contextual information in which the sample recommendation object receives the sample content. For example, contextual information may include the time and location of the recommendation, as well as environmental factors such as the type of device the sample recommendation object uses to receive the sample content and its network information.

[0133] The content information of the sample content is used to describe the semantic information of the sample content, which can reflect the characteristics of the sample content in terms of theme, content, meaning, relevance, etc.

[0134] Sample object attribute features refer to the data obtained after feature extraction of the object attribute information of the sample recommended object, which can effectively describe the object attribute information of the sample recommended object.

[0135] Sample environment features refer to the data obtained after feature extraction of the interaction context information of the sample recommendation object. It can effectively describe the interaction request information of the sample recommendation object when receiving sample content.

[0136] Sample content feature information refers to the data obtained after feature extraction of sample content. It can reflect the interaction preferences of sample recommended content in historical recommendation scenarios learned by the content processing model in the previous training process.

[0137] Sample interaction feature information is feature data obtained by integrating sample object attribute features, sample environment features, and sample content features. This sample interaction feature information references the interaction preferences of sample recommended content in historical recommendation scenarios learned by the content processing model during the previous training process, in order to predict the interaction behavior of sample recommended objects on sample content.

[0138] Sample interaction probability information is the probability that a sample recommendation object will generate interactive behavior on the sample content, which is predicted by the content processing model based on the sample interaction feature information.

[0139] The model parameters of the content processing model are used to adjust the performance of the content processing model in the feature extraction process.

[0140] The trained content processing model is obtained by calculating the loss generated from sample interaction probability information and interaction label information, and adjusting the model parameters based on this loss. Through this training, the content processing model can identify and learn the interaction preferences between sample recommendation objects and sample content.

[0141] In some embodiments, in order to enable the content processing model to retain the original content information of the sample content during training, and thus fully consider the impact of the content to be recommended on the recommendation result during recommendation, feature extraction processing is performed on the content information of the sample content to obtain sample content feature information, including:

[0142] By using a content processing model, feature enhancement processing is performed on the content information of the sample content to obtain initial sample feature information;

[0143] The content information of the sample is processed to obtain semantic residual features;

[0144] The semantic residual features and the initial sample feature information are fused to obtain the sample content feature information.

[0145] The initial sample feature information is obtained by performing feature enhancement processing on the content information of the sample content through the content processing model. This initial sample feature information can reflect the interaction preferences of the sample recommended content in historical recommendation scenarios during the previous learning process of the content processing model.

[0146] Semantic residual features preserve semantic information in the sample content that was not fully captured. By processing the sample content through residual connections, semantic details that were missed in the content information can be captured. In this way, semantic residual features supplement and enhance these details, thereby improving the content processing model's ability to deeply understand and predict content information.

[0147] In some embodiments, in order to accurately capture the semantic features of the content to be recommended during the feature extraction process, feature extraction is performed on the content to be recommended to obtain initial content feature information, including:

[0148] By using a content feature extraction model, features are extracted from the content to be recommended to obtain initial content feature information;

[0149] Before obtaining initial content feature information by extracting features from the content to be recommended using a content feature extraction model, the process also includes:

[0150] Obtain the second training data, which includes multiple sample contents, as well as positive and negative sample contents.

[0151] Using a content feature extraction model, feature extraction is performed on the sample content, the positive sample content, and the negative sample content, respectively, to obtain the semantic features of the sample content, the semantic features of the positive sample content, and the semantic features of the negative sample content.

[0152] Based on the semantic features of the sample content and the semantic features of the positive sample content, the positive similarity between the sample content and the positive sample content is determined, and based on the semantic features of the positive sample content and the semantic features of the negative sample content, the negative similarity between the sample content and the negative sample content is determined.

[0153] Based on the positive and negative similarities of multiple sample contents, the model parameters of the content feature extraction model are adjusted to obtain the trained content feature extraction model.

[0154] Among them, the content feature extraction model is a model that can extract information about the content to be recommended at the semantic level. It can reflect the characteristics of the content to be recommended, such as its theme, content, meaning, and relevance.

[0155] The second training data is used to train the content feature extraction model, enabling the model to accurately extract content information from the semantic level.

[0156] Sample content refers to the samples used to train the content feature extraction model, ensuring that the model can accurately extract information at the semantic level. For example, sample content can be of various types, such as advertisements, articles, videos, music, products, social media posts, trending topics, and promotional activities.

[0157] Positive sample content refers to content that is similar to sample content.

[0158] Negative sample content refers to content that is dissimilar to sample content.

[0159] The semantic features of sample content can characterize the information of sample content at the semantic level.

[0160] The semantic features of positive sample content can characterize the information of positive sample content at the semantic level.

[0161] The semantic features of negative sample content can characterize the information of negative sample content at the semantic level.

[0162] Positive similarity refers to the degree of semantic similarity between sample content and positive sample content.

[0163] Negative similarity refers to the degree of semantic similarity between sample content and negative sample content.

[0164] The trained content feature extraction model is obtained by evaluating the positive and negative similarity of multiple sample contents, adjusting the model parameters, and optimizing its ability to extract features from sample contents. This results in a trained model that can effectively distinguish between positive and negative sample contents and accurately extract semantic features.

[0165] In some embodiments, to enable the trained content feature extraction model to accurately extract the semantic features of the content to be recommended, the method further includes:

[0166] Obtain the performance evaluation dataset, which includes multiple evaluation items;

[0167] The trained content feature extraction model is used to extract features from the evaluation content to obtain its semantic features.

[0168] Based on the semantic features of the evaluation content and the semantic features of the reference evaluation content, similar evaluation content that is similar to the evaluation content is determined from the reference evaluation content. The reference evaluation content refers to other evaluation content outside the evaluation content.

[0169] Based on the content information of the evaluation content and the content information of similar evaluation content, the similarity between the evaluation content and the similar evaluation content is evaluated to obtain information on the degree of similarity between the similar evaluation content and the evaluation content.

[0170] Based on the similarity information corresponding to similar assessment content, the target assessment content is determined from the similar assessment content;

[0171] Based on the number of similar evaluation contents corresponding to multiple evaluation contents and the number of target evaluation contents corresponding to multiple evaluation contents, the evaluation result information of the trained content feature extraction model is determined, so as to adjust the trained content feature extraction model according to the evaluation result information.

[0172] The performance evaluation dataset is a dataset consisting of a set of evaluation content used to evaluate the performance of the trained content feature extraction model.

[0173] The evaluation content is used to test the ability of the trained content feature extraction model to extract information at the semantic level. Since the trained content feature extraction model aims to extract semantic information, the evaluation content can be of the same category as the content to be recommended, or it can be of a different category. This diverse evaluation method helps to comprehensively examine the performance of the trained content feature extraction model in different contexts.

[0174] The initial content feature information of the evaluation content is the semantic information of the evaluation content extracted by the trained content feature extraction model.

[0175] The reference evaluation content refers to all evaluation content in the performance evaluation dataset other than the evaluation content itself.

[0176] The initial content feature information of the reference evaluation content is extracted at the semantic level through a trained content feature extraction model.

[0177] Similar evaluation content refers to reference evaluation content that is similar to the evaluation content in terms of semantic features.

[0178] The content information of the assessment content is the information that describes the assessment content.

[0179] The content information of similar assessment content is information that describes similar assessment content.

[0180] Similarity information is used to indicate the degree of similarity between similar assessment content and assessment content.

[0181] In some embodiments, in order to evaluate the ability of the trained content feature extraction model to extract semantic information, the similarity between the evaluation content and similar evaluation content is evaluated based on the content information of the evaluation content and the content information of similar evaluation content, to obtain similarity information between the similar evaluation content and the evaluation content, including:

[0182] Determine the content similarity between the content information of the assessment content and the content information of similar assessment content;

[0183] Based on content similarity, the similarity between the evaluation content and similar evaluation content is evaluated to obtain information on the degree of similarity between the similar evaluation content and the evaluation content.

[0184] Content similarity reflects the degree of similarity between the evaluated content and similar evaluated content. For example, content similarity can be calculated using methods such as cosine similarity, semantic similarity, Euclidean distance, and Jaccard similarity.

[0185] The target evaluation content is the evaluation content that is similar to the evaluation content, selected from the similar evaluation content based on the similarity information corresponding to the similar evaluation content.

[0186] The evaluation results are used to indicate the performance of the trained content feature extraction model in extracting information at the semantic level.

[0187] 102. Obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0188] Among them, historical recommendation scenarios refer to scenarios in which content to be recommended was recommended before the current recommendation scenario.

[0189] For example, when the content to be recommended is new, since it has not been recommended or interacted with in the past, its interaction feature information in the historical recommendation scenario is 0. When the content to be recommended is content that has already been recommended, at least one historical interaction feature information of the content in the historical recommendation scenario can be obtained.

[0190] In some embodiments, the historical recommendation scenario may include multiple recommendation periods, and the content to be recommended corresponds to at least one historical interaction feature information in each recommendation period.

[0191] The recommendation period refers to the time frame within which content is recommended. For example, holidays could be used as a recommendation period, or content could be recommended every 7 days, and so on.

[0192] In some embodiments, in order to determine at least one historical interaction feature information corresponding to the content to be recommended during the recommendation period, the method further includes:

[0193] Obtain the content processing model corresponding to the recommendation cycle;

[0194] Feature extraction is performed on the content to be recommended to obtain initial content feature information;

[0195] By using the content processing model corresponding to the recommendation cycle, the initial content feature information is enhanced to obtain the historical interaction feature information corresponding to the recommendation cycle.

[0196] Historical interaction feature information can reflect the interaction preferences between the content to be recommended and historical recommendation objects in historical recommendation scenarios. Historical recommendation objects refer to the objects that recommended the content to be recommended in historical recommendation scenarios.

[0197] For example, historical interaction feature information can be the interaction feature information of the historical recommendation scenario closest to the current recommendation scenario, or it can be any one of multiple interaction feature information of historical recommendation scenarios within a preset time range from the current recommendation scenario, or it can be the historical interaction feature information obtained by averaging multiple interaction feature information of historical recommendation scenarios.

[0198] In some embodiments, in order to comprehensively consider various interaction feature information of the content to be recommended in historical recommendation scenarios when recommending content, thereby achieving more accurate recommendations, the historical interaction feature information of the content to be recommended in historical recommendation scenarios is obtained, including:

[0199] Determine the historical feature storage area corresponding to the content to be recommended. The historical feature storage area stores multiple interaction feature information of the content to be recommended in historical recommendation scenarios.

[0200] Retrieve multiple interaction feature information from historical recommendation scenarios from the historical feature storage area;

[0201] By fusing multiple interaction feature information from historical recommendation scenarios, we can obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0202] The historical feature storage area is used to manage and store multiple interaction feature information of the content to be recommended in historical recommendation scenarios. For example, the historical feature storage area can be a database, table, message queue, etc.

[0203] Interaction feature information can reflect the interaction preferences between the content to be recommended and the historical recommendation objects in the corresponding historical recommendation scenarios. Different historical recommendation scenarios correspond to different interaction feature information.

[0204] In some embodiments, in order to obtain comprehensive historical interaction feature information, multiple interaction feature information from historical recommendation scenarios are fused to obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios, including:

[0205] The historical interaction feature information of the content to be recommended is obtained by averaging multiple interaction feature information in historical recommendation scenarios.

[0206] In some embodiments, in order to obtain the interaction feature information of the content to be recommended in the latest historical recommendation scenario when recommending content to be recommended in a future time period, which is beneficial for more accurate recommendation of the content to be recommended, after extracting the interaction feature information of the content to be recommended in the current recommendation scenario, the method further includes:

[0207] The interaction feature information of the current recommendation scenario is stored in the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area and obtain the updated historical feature storage area. The updated historical feature storage area is used for the content recommendation scenario of the content to be recommended in the future time period.

[0208] The updated historical feature storage area refers to the area obtained after updating the interactive feature information in the historical feature storage area.

[0209] The future time period refers to the time period following the current recommended scenario.

[0210] In some embodiments, in order to dynamically update the interaction feature information in the historical feature storage area so as to reflect the latest user preferences and behaviors in a timely manner, thereby improving the accuracy and real-time performance of recommendations, the storage structure of the historical feature storage area corresponding to the content to be recommended is a queue.

[0211] The interaction feature information of the current recommendation scenario is stored in the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area, including:

[0212] Determine the current remaining storage space of the historical feature storage area corresponding to the content to be recommended;

[0213] Based on the remaining storage space and the target storage space required for the interaction feature information in the current recommendation scenario, the interaction feature information in the historical recommendation scenario in the historical feature storage area is deleted in a first-in-first-out order to obtain the processed historical feature storage area.

[0214] The interaction feature information of the current recommendation scenario is stored in the processed historical feature storage area.

[0215] In this context, "queue" refers to the management of interactive feature information in the historical feature storage area according to a first-in, first-out (FIFO) order.

[0216] First-In, First-Out (FIFO) order means that the first piece of data to enter the queue is processed or removed first. Under this order, data processing follows the "first-come, first-served" principle. That is, the first piece of data added to the queue is the first to be retrieved, and subsequent data will be processed only after the preceding data has been processed.

[0217] The current remaining storage space refers to the unused storage capacity within the historical feature storage area corresponding to the content to be recommended. It reflects the amount of interactive feature information that can still be accommodated in this historical feature storage area. The size of the current remaining storage space typically depends on the difference between the total capacity of the historical feature storage area and the capacity of the already stored interactive feature information.

[0218] The target storage space refers to the space occupied by the interaction feature information in the current recommendation scenario in the historical feature storage area.

[0219] The processed historical feature storage area refers to the area obtained after storing the interaction feature information of the current recommendation scenario into the historical feature storage area.

[0220] 103. Based on current interaction feature information and historical interaction feature information, predict the interaction probability information of the object to be recommended on the content to be recommended.

[0221] Among them, the interaction probability information is the probability that the object to be recommended will interact with the content to be recommended in the current recommendation scenario by combining the current interaction feature information and the historical interaction feature information.

[0222] In some embodiments, in order to comprehensively consider the interaction probability information of the object to be recommended interacting with the content to be recommended on multiple interaction type operation tasks when recommending content, the interaction probability information of the object to be recommended interacting with the content to be recommended is predicted based on current interaction feature information and historical interaction feature information, including:

[0223] The target interaction feature information is obtained by fusing the current interaction feature information and the historical interaction feature information.

[0224] Based on the target interaction feature information, the interaction probability information of the object to be recommended and the content to be recommended is predicted in multiple interaction type operation tasks.

[0225] Among them, target interaction feature information refers to the feature information obtained by fusing current interaction feature information and historical interaction feature information.

[0226] Multiple interaction type operation tasks are tasks established for different types of interaction operations, used to predict the probability that the object to be recommended will perform various interaction operations on the content to be recommended.

[0227] Interactive actions can specifically include clicking, browsing, viewing time, saving, liking, sharing, and purchasing.

[0228] In some embodiments, in order to predict the click probability and conversion probability of the object to be recommended on the content to be recommended, thereby achieving more accurate content recommendation, interaction probability information of the object to be recommended generating interactive behavior on the content to be recommended is predicted based on feature interaction information, including:

[0229] Based on feature interaction information, predict the click probability of the object to be recommended clicking on the content to be recommended;

[0230] Based on feature interaction information, predict the conversion probability of the recommended object for the recommended content;

[0231] Based on interaction probability information, target recommended content is determined from multiple candidate recommended content, including:

[0232] Based on click probability and conversion probability, target recommended content is determined from multiple potential recommended content.

[0233] Click probability refers to the predicted probability that the target audience will click on the recommended content, which is used to measure the target audience's interest in the recommended content.

[0234] Conversion probability refers to the predicted probability that the recommended object will convert in response to the recommended content. It is used to measure the degree of influence of the recommended content on the recommended object.

[0235] For example, conversion probability specifically refers to the conversion rate of the target audience after they come into contact with the recommended content and perform interactive operations such as browsing, watching, collecting, liking, and sharing. It can also refer to the conversion rate of the target audience purchasing related items through the recommended content.

[0236] 104. Based on the interaction probability information, determine the target recommended content from multiple content to be recommended.

[0237] Among them, the target recommendation content is the recommendation content that needs to be recommended to the target audience.

[0238] For example, the target recommended content can be the content to be recommended corresponding to the highest interaction probability information, or it can be any one of the top n recommended content in terms of interaction probability information, or it can be the recommended content filtered by weighted interaction probability information.

[0239] 105. Recommend target content to the target audience.

[0240] As can be seen from the above, the embodiments of this application can determine multiple pieces of content to be recommended based on the content recommendation request of the object to be recommended, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the object to be recommended; obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios; predict the interaction probability information of the object to be recommended generating interaction behavior on the content to be recommended based on the current interaction feature information and the historical interaction feature information; determine the target recommended content from the multiple pieces of content to be recommended based on the interaction probability information; and recommend the target recommended content to the object to be recommended.

[0241] Therefore, this solution can identify multiple pieces of content to be recommended based on the content recommendation request of the target audience. Next, it extracts the interaction features of these pieces of content in the current recommendation scenario, and simultaneously obtains their historical interaction features in historical scenarios. Combining the current and historical interaction features allows for a better characterization of the content to be recommended, and thus more accurately predicts the probability of the target audience interacting with each piece of content. This helps to filter out target content that is likely to induce interaction and accurately recommend it to the target audience. This improves the accuracy of content recommendation, saves promotional resources, enhances user experience, and ultimately improves the overall promotional effect.

[0242] The method described in the above embodiments will be further described in detail below.

[0243] In this embodiment, the method of this application embodiment will be described in detail using the recommendation of advertisements as an example.

[0244] like Figure 2a As shown, the specific process of a content recommendation method is as follows:

[0245] 201. Based on the content recommendation request of the object to be recommended, determine multiple pieces of content to be recommended, and extract the interaction feature information of the content to be recommended in the current recommendation scenario, which is a scenario of recommending content to the object to be recommended.

[0246] In some embodiments, the interaction feature information of the content to be recommended in the current recommendation scenario is extracted, including:

[0247] Obtain the object attribute information of the object to be recommended and the interaction context information of the object to be recommended in the current recommendation scenario;

[0248] The object attribute information is processed by feature extraction to obtain the object attribute features of the object to be recommended;

[0249] Feature extraction processing is performed on the interaction context information to obtain the environmental features of the object to be recommended;

[0250] Feature extraction is performed on the recommended content to obtain content feature information;

[0251] Based on object attribute features, environmental features, and content features, determine the interaction features of the content to be recommended in the current recommendation scenario.

[0252] In some embodiments, feature extraction is performed on the content to be recommended to obtain content feature information, including:

[0253] Feature extraction is performed on the content to be recommended to obtain initial content feature information;

[0254] The initial content feature information is subjected to feature enhancement processing to obtain the content feature information.

[0255] In some embodiments, the interaction feature information of the content to be recommended in the current recommendation scenario is extracted, including:

[0256] By using a content processing model, interactive feature information of the content to be recommended in the current recommendation scenario is extracted;

[0257] Before extracting the interaction feature information of the content to be recommended in the current recommendation scenario through the content processing model, it also includes:

[0258] Obtain the first training data, which includes the interaction label information between the sample recommendation object and the sample content, the object attribute information of the sample recommendation object, the interaction context information, and the content information of the sample content;

[0259] Using a content processing model, feature extraction is performed on the object attribute information and interaction context information of the sample recommendation object to obtain sample object attribute features and sample environment features.

[0260] The sample content information is processed to extract features to obtain initial sample feature information; the sample content information is also processed to obtain semantic residual features; the semantic residual features and the initial sample feature information are then fused to obtain sample content feature information.

[0261] By using a content processing model, feature interaction is performed on sample object attribute features, sample environment features, and sample content features to obtain sample interaction feature information.

[0262] Based on the sample interaction feature information, predict the probability information of sample recommendation objects interacting with samples that generate interactive behavior in multiple interaction type operation tasks.

[0263] Based on the sample interaction probability information and interaction tag information, the model parameters of the content processing model are adjusted to obtain the trained content processing model.

[0264] For example, taking multiple interaction type operation tasks, including Click-Through Rate (CTR) and Conversion Rate (CVR), with the sample content information being the template ad content embedding and the initial sample feature information being the ad scenario embedding, we can model the ad content embedding based on the Click-Through Rate (CTR) or Conversion Rate (CVR) task, and calculate the initial sample feature information, i.e., the ad scenario embedding, to capture business information in the ad scenario.

[0265] like Figure 2b As shown, the input to the content processing model includes sample ad content embedding, object attribute information of sample recommended objects, and interaction context information. The object attribute information includes basic information about the sample recommended objects, such as interests and ad click sequences; the interaction context information includes location and device type. For the object attribute information and interaction context information, an embedding layer is constructed based on the deep neural network-based content processing model to map the feature input to a low-dimensional dense embedding space. Similarly, an embedding layer is constructed based on the deep neural network-based content processing model to map the ad content embedding. Through the supervision signal of the CTR / CVR task, the semantic space of the ad embedding learns the personalized recommendation signal that the content processing model focuses on, while the original semantic space information of the ad content embedding is preserved through residual connections.

[0266] Then, an interactive module fuses sample content features, sample object attribute features, and sample environment features to perform feature cross-learning, thereby achieving CTR / CVR prediction. The content processing model is trained using the Cross Entropy Loss function, the mathematical formula of which is as follows:

[0267] Loss=-∑(y×logp+(1-y)×log(1-p));

[0268] Where y represents the interaction tag information of the sample content in CTR / CVR, i.e., whether it was clicked or converted; p represents the sample interaction probability information of the content processing model. The smaller the value of the cross-entropy loss function, the closer the distribution of the sample interaction probability information predicted by the content processing model is to the distribution of the interaction tag information. During training, the gradient optimization algorithm (Adam) is used to minimize the cross-entropy loss function, enabling the content processing model to learn personalized recommendation information. Finally, the area under the curve (AUC) metric is used to evaluate the performance of the content processing model. In actual business, the AUC metric can reach 0.92. The interaction feature information in historical recommendation scenarios is produced by the Embedding Layer of advertising features. In some embodiments, feature extraction is performed on the content to be recommended to obtain initial content feature information, including:

[0269] By using a content feature extraction model, features are extracted from the content to be recommended to obtain initial content feature information;

[0270] Before obtaining initial content feature information by extracting features from the content to be recommended using a content feature extraction model, the process also includes:

[0271] Obtain the second training data, which includes multiple sample contents, as well as positive and negative sample contents.

[0272] Using a content feature extraction model, feature extraction is performed on the sample content, the positive sample content, and the negative sample content, respectively, to obtain the semantic features of the sample content, the semantic features of the positive sample content, and the semantic features of the negative sample content.

[0273] Based on the semantic features of the sample content and the semantic features of the positive sample content, the positive similarity between the sample content and the positive sample content is determined, and based on the semantic features of the positive sample content and the semantic features of the negative sample content, the negative similarity between the sample content and the negative sample content is determined.

[0274] Based on the positive and negative similarities of multiple sample contents, the model parameters of the content feature extraction model are adjusted to obtain the trained content feature extraction model.

[0275] Among them, the content feature extraction model is based on metric learning to model the similarity of advertising content and obtain initial content feature information (ad content embedding). The embedding space of this initial content feature information reflects the similarity of advertising content.

[0276] For example, such as Figure 2cAs shown, an encoding model for the advertising content embedding, i.e., a content feature extraction model, is designed. A computer vision network (Vision Transformer, ViT) based on a self-attention network (Transformer architecture) is used to extract features from the image and video content of the sample advertisement. The Transformer network is used to extract features from the text content of the sample advertisement. Then, the multimodal features of the advertisement are fused through an interaction module. The interaction module is generally implemented based on a cross-attention network. Finally, the sample advertisement content embedding is output.

[0277] During the training of the content feature extraction model, positive and negative sample ads are identified based on the sample ads in the second training data.

[0278] Based on the embedding of sample advertisement content, and using cosine similarity as the metric function, the content feature extraction model is trained using triple loss. Following the principles of metric learning, the higher the cosine similarity of the embeddings of two similar and positive sample advertisements after processing by the content feature extraction model, the lower the cosine similarity of the embeddings of two dissimilar and negative sample advertisements. The formula for the triple loss function used to train the content feature extraction model is as follows:

[0279] Loss = max(d(a , p)-d(a , n)+margin,0)

[0280] Where a, p, and n represent sample ads, positive sample ads, and negative sample ads, respectively. Among the previously labeled ad pairs, sample ad pairs marked as similar are positive sample ads to each other, and sample ad pairs marked as dissimilar are negative sample ads to each other.

[0281] d(·) is the metric function, namely the cosine phase distance, which is calculated based on the sample ad content embedding. Taking d(a,p) as an example, a and p are the sample ad content embeddings obtained by encoding the sample ad and positive sample ad through the content feature extraction model, respectively.

[0282] d(a,p) = 1 - cosine(a,p), where margin is a constant greater than 0 used to control the minimum interval between positive and negative samples, usually set to 0.5.

[0283] Based on the above loss function, the gradient optimization algorithm (Adam) is used to train the content feature extraction model. The content feature extraction model learns to map the ad content to a low-dimensional embedding semantic space, so that the embedding vectors of ads with similar content are closer together, and the embedding vectors of ads with dissimilar content are farther apart.

[0284] This application embodiment uses a content feature extraction model and a content processing model to obtain the interaction feature information of advertisements in historical recommendation scenarios. The online recommendation system uses this interaction feature information for personalized recommendations. For the content feature extraction model, after achieving satisfactory results, it is frozen and its model parameters are no longer updated. For the content feature extraction model, periodic updates are performed by collecting click-conversion logs of the content to be recommended; the period window can be set to 7 days.

[0285] In some embodiments, the method further includes:

[0286] Obtain the performance evaluation dataset, which includes multiple evaluation items;

[0287] The trained content feature extraction model is used to extract features from the evaluation content to obtain its semantic features.

[0288] Based on the semantic features of the evaluation content and the semantic features of the reference evaluation content, similar evaluation content that is similar to the evaluation content is determined from the reference evaluation content. The reference evaluation content refers to other evaluation content outside the evaluation content.

[0289] Based on the content information of the evaluation content and the content information of similar evaluation content, the similarity between the evaluation content and the similar evaluation content is evaluated to obtain information on the degree of similarity between the similar evaluation content and the evaluation content.

[0290] Based on the similarity information corresponding to similar assessment content, the target assessment content is determined from the similar assessment content;

[0291] Based on the number of similar evaluation contents corresponding to multiple evaluation contents and the number of target evaluation contents corresponding to multiple evaluation contents, the evaluation result information of the trained content feature extraction model is determined, so as to adjust the trained content feature extraction model according to the evaluation result information.

[0292] For example, the P@10 metric is used to evaluate the performance of the trained content feature extraction model. On the performance evaluation dataset, for each evaluation ad, the nearest neighbor retrieval is performed based on the ad content embedding to recall 10 similar evaluation ads with the closest embedding distance. Then, the similarity is manually labeled to obtain the similarity information between the similar evaluation ads and the evaluation ad. Here, n is the number of target evaluation ads labeled as similar among the 10 recalled evaluation ads, which is used to evaluate the ability of the ad content embedding to represent content similarity.

[0293] In actual business operations, this metric was evaluated as 0.95, demonstrating excellent performance. Finally, the model parameters of the content feature extraction model were frozen and fixed, and the ad content embedding results were used as input to the content processing model.

[0294] Where n is the number of target evaluation ads.

[0295] 202. Obtain multiple interaction feature information from the historical feature storage area corresponding to the content to be recommended in the historical recommendation scenario. The historical feature storage area stores multiple interaction feature information of the content to be recommended in the historical recommendation scenario.

[0296] For example, taking an advertisement as the content to be recommended, the historical feature storage area stores multiple interaction feature information (ad embeddings) of the same advertisement in historical recommendation scenarios. The data dimension of the ad embedding is d, and the storage space of the historical feature storage area is k, indicating that the historical feature storage area can store k ad embeddings of the same advertisement in historical recommendation scenarios. The ad embeddings in the historical feature storage area are written in k update processes, following a first-in-first-out queue mechanism. In practical scenarios, k can be set to 2 and d can be set to 64; the range of values ​​for k and d is not restricted here.

[0297] In some embodiments, multiple interaction feature information from historical recommendation scenarios is obtained from the historical feature storage area corresponding to the content to be recommended, including:

[0298] In the current recommendation scenario, i interaction feature information from the historical feature storage area corresponding to the content to be recommended can be obtained from the historical recommendation scenario. The i interaction feature information is less than the k interaction feature information.

[0299] Furthermore, the interaction feature information (ad scene embedding) of the ad to be recommended in the current recommendation scenario is written into the historical feature storage area, and the ad scene embedding is updated. Multiple ad scene embeddings of the same ad follow the first-in-first-out queue mechanism, and no more than k are stored.

[0300] 203. Merge multiple interaction feature information from historical recommendation scenarios to obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0301] In some embodiments, the fusion of multiple interaction feature information from historical recommendation scenarios includes:

[0302] The average value of the i interaction features in the historical recommendation scenario is calculated to obtain the historical interaction features of the content to be recommended in the historical recommendation scenario.

[0303] In this embodiment, the interaction feature acquisition operation in historical recommendation scenarios does not require embedding feature calculation; querying the historical feature storage area suffices, thus meeting the high real-time requirements of the recommendation system. By averaging the embeddings of the most recent i updated ads obtained in each query, the real-time changes in user interest in ads within the recommendation system can be effectively captured.

[0304] 204. The current interaction feature information and the historical interaction feature information are fused together to obtain the target interaction feature information; and based on the target interaction feature information, the interaction probability information of the object to be recommended interacting with the content to be recommended in multiple interaction type operation tasks is predicted.

[0305] 205. Based on the interaction probability information corresponding to multiple interaction type operation tasks, determine the target recommended content from multiple content to be recommended.

[0306] 206. Recommend target content to the target audience.

[0307] like Figure 2d As shown, this application, after obtaining a content recommendation request from a target object, first identifies multiple target content items. For each target content item, it retrieves the interaction feature information of that item in historical recommendation scenarios from the historical feature storage area corresponding to that item, and averages these interaction feature information to obtain the historical interaction feature information of the target content in the historical recommendation scenarios. Simultaneously, it calculates the interaction feature information of the target content in the current recommendation scenario and writes it into the historical feature storage area corresponding to that item. Combining the historical interaction feature information and the current interaction feature information, a content processing model predicts the interaction probability information of the target object interacting with the target content in multiple interaction type operation tasks. Based on the interaction probability information, the target recommended content is determined from the multiple target content items and promoted to the target object.

[0308] Thus, taking recommended advertising as an example, this application proposes an asynchronous online service framework for advertising scenario embedding. This framework asynchronously separates the acquisition and updating of advertising scenario embedding, satisfying the real-time requirements of recommendation systems while allowing for more complex modeling and appropriate updates to the advertising scenario embedding to capture the latest personalized recommendation signals. Simultaneously, a two-stage complex embedding modeling method is designed. In the first stage, metric learning is used to model the similarity of advertising content, resulting in the first-stage advertising content embedding, which reflects the similarity of advertising content. In the second stage, the advertising content embedding obtained in the first stage is modeled based on CTR or CVR tasks, resulting in the second-stage advertising scenario embedding. This second-stage embedding retains, to a certain extent, the similarity relationships of advertising content while capturing the business signals of the advertising recommendation scenario, enabling it to play a better role in personalized recommendation systems.

[0309] The asynchronous online service framework of this application only needs to query the historical feature storage area to obtain the interaction feature information (ad embedding) in the historical recommendation scenario for each request, and the time consumption can be completed in milliseconds. Compared with the previous real-time calculation service time consumption performance and stability improvement, the framework is very significant. The complex content processing model designed in this application has stronger representation ability and better information capture for recommendation business, which significantly improves the personalized recommendation effect in actual recommendation business.

[0310] As can be seen from the above, this application can improve the accuracy of content recommendations, save promotional resources, enhance user experience, and thus further improve the overall promotional effect.

[0311] To better implement the above methods, this application also provides a content recommendation device, which can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, or personal computer; the server can be a single server or a server cluster composed of multiple servers.

[0312] For example, in this embodiment, the method of this application embodiment will be described in detail by taking the content recommendation device specifically integrated into an electronic device as an example.

[0313] For example, such as Figure 3 As shown, the content recommendation device may include an information processing unit 301, an information acquisition unit 302, an interaction prediction unit 303, a recommendation determination unit 304, and a content recommendation unit 305, as follows:

[0314] (I) Information processing unit 301.

[0315] The information processing unit 301 is used to determine multiple recommended contents based on the content recommendation request of the object to be recommended, and extract the interaction feature information of the recommended contents in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the object to be recommended.

[0316] In some embodiments, the information processing unit includes an information acquisition subunit, an attribute feature extraction subunit, an environmental feature extraction subunit, a content feature extraction subunit, and an information determination subunit.

[0317] The information acquisition subunit is used to acquire the object attribute information of the object to be recommended and the interaction context information of the object to be recommended in the current recommendation scenario.

[0318] The attribute feature extraction subunit is used to perform feature extraction processing on object attribute information to obtain the object attribute features of the object to be recommended.

[0319] The environmental feature extraction subunit is used to extract features from the interaction context information to obtain the environmental features of the object to be recommended.

[0320] The content feature extraction subunit is used to extract features from the content to be recommended, thereby obtaining content feature information.

[0321] The information determination subunit is used to determine the interaction feature information of the content to be recommended in the current recommendation scenario based on object attribute features, environmental features, and content feature information.

[0322] In some embodiments, the content feature extraction subunit includes a feature extraction module and a feature enhancement module:

[0323] The feature extraction module is used to extract features from the content to be recommended, and obtain initial content feature information;

[0324] The feature enhancement module is used to enhance the initial content feature information to obtain the content feature information.

[0325] In some embodiments, the information processing unit is configured to:

[0326] By using a content processing model, interactive feature information of the content to be recommended in the current recommendation scenario is extracted;

[0327] Before extracting the interaction feature information of the content to be recommended in the current recommendation scenario through the content processing model, the device also includes a content processing model training unit, used for:

[0328] Obtain the first training data, which includes the interaction label information between the sample recommendation object and the sample content, the object attribute information of the sample recommendation object, the interaction context information, and the content information of the sample content;

[0329] Through the content processing model, feature extraction is performed on the object attribute information, interaction context information, and content information of the sample recommendation object, respectively, to obtain sample object attribute features, sample environment features, and sample content feature information.

[0330] By using a content processing model, feature interaction is performed on sample object attribute features, sample environment features, and sample content features to obtain sample interaction feature information.

[0331] Based on the sample interaction feature information, predict the sample interaction probability information of the sample recommendation object generating interactive behavior on the sample content;

[0332] Based on the sample interaction probability information and interaction tag information, the model parameters of the content processing model are adjusted to obtain the trained content processing model.

[0333] In some embodiments, the content processing model training unit is used to perform feature extraction processing on the content information of the sample content, including:

[0334] By using a content processing model, feature extraction is performed on the content information of the sample content to obtain initial sample feature information;

[0335] The content information of the sample is processed to obtain semantic residual features;

[0336] The semantic residual features and the initial sample feature information are fused to obtain the sample content feature information.

[0337] In some embodiments, the feature extraction module is used for:

[0338] By using a content feature extraction model, features are extracted from the content to be recommended to obtain initial content feature information;

[0339] Before obtaining initial content feature information by extracting features from the content to be recommended using a content feature extraction model, the device also includes a content feature extraction model training unit, used for:

[0340] Obtain the second training data, which includes multiple sample contents, as well as positive and negative sample contents.

[0341] Using a content feature extraction model, feature extraction is performed on the sample content, the positive sample content, and the negative sample content, respectively, to obtain the semantic features of the sample content, the semantic features of the positive sample content, and the semantic features of the negative sample content.

[0342] Based on the semantic features of the sample content and the semantic features of the positive sample content, the positive similarity between the sample content and the positive sample content is determined, and based on the semantic features of the positive sample content and the semantic features of the negative sample content, the negative similarity between the sample content and the negative sample content is determined.

[0343] Based on the positive and negative similarities of multiple sample contents, the model parameters of the content feature extraction model are adjusted to obtain the trained content feature extraction model.

[0344] In some embodiments, the apparatus further includes a model evaluation unit for:

[0345] Obtain the performance evaluation dataset, which includes multiple evaluation items;

[0346] The trained content feature extraction model is used to extract features from the evaluation content to obtain its semantic features.

[0347] Based on the semantic features of the evaluation content and the semantic features of the reference evaluation content, similar evaluation content that is similar to the evaluation content is determined from the reference evaluation content. The reference evaluation content refers to other evaluation content outside the evaluation content.

[0348] Based on the content information of the evaluation content and the content information of similar evaluation content, the similarity between the evaluation content and the similar evaluation content is evaluated to obtain information on the degree of similarity between the similar evaluation content and the evaluation content.

[0349] Based on the similarity information corresponding to similar assessment content, the target assessment content is determined from the similar assessment content;

[0350] Based on the number of similar evaluation contents corresponding to multiple evaluation contents and the number of target evaluation contents corresponding to multiple evaluation contents, the evaluation result information of the trained content feature extraction model is determined, so as to adjust the trained content feature extraction model according to the evaluation result information.

[0351] (II) Information Acquisition Unit 302.

[0352] The information acquisition unit 302 is used to acquire historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0353] In some embodiments, the information acquisition unit is configured to:

[0354] Determine the historical feature storage area corresponding to the content to be recommended. The historical feature storage area stores multiple interaction feature information of the content to be recommended in historical recommendation scenarios.

[0355] Retrieve multiple interaction feature information from historical recommendation scenarios from the historical feature storage area;

[0356] By fusing multiple interaction feature information from historical recommendation scenarios, we can obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios.

[0357] In some embodiments, after extracting the interaction feature information of the content to be recommended in the current recommendation scenario, the device further includes an information storage unit for:

[0358] The interaction feature information of the current recommendation scenario is stored in the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area and obtain the updated historical feature storage area. The updated historical feature storage area is used for the content recommendation scenario of the content to be recommended in the future time period.

[0359] In some embodiments, the storage structure of the historical feature storage area corresponding to the content to be recommended is a queue;

[0360] Information storage unit, used for:

[0361] Determine the current remaining storage space of the historical feature storage area corresponding to the content to be recommended;

[0362] Based on the remaining storage space and the target storage space required for the interaction feature information in the current recommendation scenario, the interaction feature information in the historical recommendation scenario in the historical feature storage area is deleted in a first-in-first-out order to obtain the processed historical feature storage area.

[0363] The interaction feature information of the current recommendation scenario is stored in the processed historical feature storage area.

[0364] (III) Interactive prediction unit 303.

[0365] The interaction prediction unit 303 is used to predict the interaction probability information of the object to be recommended on the content to be recommended based on the current interaction feature information and the historical interaction feature information.

[0366] In some embodiments, the interactive prediction unit is configured to:

[0367] The target interaction feature information is obtained by fusing the current interaction feature information and the historical interaction feature information.

[0368] Based on the target interaction feature information, the interaction probability information of the object to be recommended and the content to be recommended is predicted in multiple interaction type operation tasks.

[0369] (iv) Recommendation for determining Unit 304.

[0370] The recommendation determination unit 304 is used to determine the target recommended content from multiple content to be recommended based on interaction probability information.

[0371] (V) Content Recommendation Unit 305.

[0372] Content recommendation unit 305 is used to recommend target content to the object to be recommended.

[0373] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0374] As can be seen from the above, the content recommendation device in this embodiment consists of an information processing unit that determines multiple content items to be recommended based on the content recommendation request of the object to be recommended, and extracts the interaction feature information of the content items to be recommended in the current recommendation scenario, where the current recommendation scenario is a scenario of recommending content to the object to be recommended; an information acquisition unit that acquires the historical interaction feature information of the content items to be recommended in historical recommendation scenarios; an interaction prediction unit that predicts the interaction probability information of the object to be recommended on the content items to be recommended based on the current interaction feature information and the historical interaction feature information; a recommendation determination unit that determines the target recommended content from the multiple content items to be recommended based on the interaction probability information; and a content recommendation unit that recommends the target recommended content to the object to be recommended.

[0375] Therefore, this application embodiment can filter out target recommendation content that may prompt the recommended object to interact, and accurately recommend the target recommendation content to the recommended object. This can improve the accuracy of content recommendation, save promotional resources, enhance user experience, and thus further improve the overall promotion effect.

[0376] This application also provides an electronic device, which can be a terminal, a server, or other similar device. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0377] In some embodiments, the content recommendation device can also be integrated into multiple electronic devices. For example, the content recommendation device can be integrated into multiple servers, and the content recommendation method of this application can be implemented by multiple servers.

[0378] In this embodiment, a server will be used as an example for detailed description. For example, ... Figure 4 As shown, it illustrates a schematic diagram of the server structure involved in an embodiment of this application. Specifically:

[0379] The server may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, an input module 404, and a communication module 405. Those skilled in the art will understand that... Figure 4 The server architecture shown does not constitute a limitation on the server and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Wherein:

[0380] The processor 401 is the control center of the server, connecting various parts of the server via various interfaces and lines. It performs various server functions and processes data by running or executing software programs and / or modules stored in the memory 402, and by calling data stored in the memory 402. In some embodiments, the processor 401 may include one or more processing cores; in some embodiments, the processor 401 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 401.

[0381] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the server, etc. In addition, the memory 402 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 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0382] The server also includes a power supply 403 that supplies power to the various components. In some embodiments, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 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.

[0383] The server may also include an input module 404, which can be used to receive input numeric or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0384] The server may also include a communication module 405. In some embodiments, the communication module 405 may include a wireless module, through which the server can perform short-range wireless transmission, thereby providing users with wireless broadband internet access. For example, the communication module 405 can be used to help users send and receive emails, browse web pages, and access streaming media.

[0385] Although not shown, the server may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the server loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows:

[0386] Based on the content recommendation request of the object to be recommended, multiple pieces of content to be recommended are identified, and the interaction feature information of the content to be recommended in the current recommendation scenario is extracted. The current recommendation scenario is a scenario of recommending content to the object to be recommended.

[0387] Obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios;

[0388] Based on current and historical interaction features, predict the probability of the recommended object interacting with the recommended content.

[0389] Based on interaction probability information, the target recommended content is determined from multiple content to be recommended;

[0390] Recommend target content to the target audience.

[0391] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0392] As shown above, combining current and historical interaction features allows for a better characterization of the content to be recommended. This enables more accurate prediction of the probability of the target user interacting with each piece of content, thus filtering out target content that is likely to induce interaction and accurately recommending it to the target user. This improves the accuracy of content recommendation, saves promotional resources, enhances user experience, and ultimately improves the overall promotional effect.

[0393] 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.

[0394] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the content recommendation methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0395] Based on the content recommendation request of the object to be recommended, multiple pieces of content to be recommended are identified, and the interaction feature information of the content to be recommended in the current recommendation scenario is extracted. The current recommendation scenario is a scenario of recommending content to the object to be recommended.

[0396] Obtain historical interaction feature information of the content to be recommended in historical recommendation scenarios;

[0397] Based on current and historical interaction features, predict the probability of the recommended object interacting with the recommended content.

[0398] Based on interaction probability information, the target recommended content is determined from multiple content to be recommended;

[0399] Recommend target content to the target audience.

[0400] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0401] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the methods provided in the various optional implementations of the content recommendation aspect provided in the above embodiments. The computer program / instructions are stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the methods provided in the various optional implementations of the content recommendation aspect provided in the above embodiments.

[0402] Since the instructions stored in the storage medium can execute the steps of any of the content recommendation methods provided in the embodiments of this application, the beneficial effects that any of the content recommendation methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0403] The foregoing has provided a detailed description of a content recommendation method, apparatus, electronic device, storage medium, and program product 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 methods 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 content recommendation method, characterized in that, include: Based on the content recommendation request of the object to be recommended, multiple pieces of content to be recommended are determined, and the interaction feature information of the content to be recommended in the current recommendation scenario is extracted, wherein the current recommendation scenario is a scenario in which content is recommended to the object to be recommended; Obtain the historical interaction feature information of the content to be recommended in historical recommendation scenarios; Based on the current interaction feature information and the historical interaction feature information, predict the interaction probability information of the object to be recommended on the content to be recommended; Based on the interaction probability information, the target recommended content is determined from the plurality of content to be recommended; The target recommended content is recommended to the object to be recommended.

2. The content recommendation method as described in claim 1, characterized in that, The step of extracting the interaction feature information of the content to be recommended in the current recommendation scenario includes: Obtain the object attribute information of the object to be recommended and the interaction context information of the object to be recommended in the current recommendation scenario; The object attribute information is subjected to feature extraction processing to obtain the object attribute features of the object to be recommended; The interaction context information is processed by feature extraction to obtain the environmental features of the object to be recommended; Feature extraction is performed on the content to be recommended to obtain content feature information; Based on the object attribute features, the environmental features, and the content feature information, the interaction feature information of the content to be recommended in the current recommendation scenario is determined.

3. The content recommendation method as described in claim 2, characterized in that, The step of extracting features from the content to be recommended to obtain content feature information includes: Feature extraction is performed on the content to be recommended to obtain initial content feature information; The initial content feature information is subjected to feature enhancement processing to obtain content feature information.

4. The content recommendation method as described in claim 1, characterized in that, The step of obtaining the historical interaction feature information of the content to be recommended in historical recommendation scenarios includes: Determine the historical feature storage area corresponding to the content to be recommended, wherein the historical feature storage area stores multiple interaction feature information of the content to be recommended in historical recommendation scenarios; Obtain multiple interaction feature information from the historical recommendation scenarios from the historical feature storage area; The interaction feature information of the content to be recommended is obtained by fusing multiple interaction feature information in the historical recommendation scenario.

5. The content recommendation method as described in claim 4, characterized in that, After extracting the interaction feature information of the content to be recommended in the current recommendation scenario, the method further includes: The interaction feature information of the current recommendation scenario is stored in the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area and obtain the updated historical feature storage area. The updated historical feature storage area is used for the content recommendation scenario of the content to be recommended in a future time period.

6. The content recommendation method as described in claim 5, characterized in that, The storage structure of the historical feature storage area corresponding to the content to be recommended is a queue; The step of storing the interaction feature information of the current recommendation scenario into the historical feature storage area corresponding to the content to be recommended, so as to update the historical feature storage area, includes: Determine the current remaining storage space of the historical feature storage area corresponding to the content to be recommended; Based on the current remaining storage space and the target storage space required for the interaction feature information under the current recommendation scenario, the interaction feature information under the historical recommendation scenario in the historical feature storage area is deleted in a first-in-first-out order to obtain the processed historical feature storage area. The interaction feature information of the current recommendation scenario is stored in the processed historical feature storage area.

7. The content recommendation method as described in claim 1, characterized in that, The step of predicting the probability information of the target object interacting with the content to be recommended based on the current interaction feature information and the historical interaction feature information includes: The current interaction feature information and the historical interaction feature information are fused together to obtain the target interaction feature information; Based on the target interaction feature information, the interaction probability information of the object to be recommended interacting with the content to be recommended in multiple interaction type operation tasks is predicted respectively.

8. The content recommendation method as described in claim 1, characterized in that, The step of extracting the interaction feature information of the content to be recommended in the current recommendation scenario includes: The interactive feature information of the content to be recommended in the current recommendation scenario is extracted using a content processing model; Before extracting the interaction feature information of the content to be recommended in the current recommendation scenario through the content processing model, the method further includes: Obtain first training data, which includes interaction tag information between sample recommendation objects and sample content, object attribute information of the sample recommendation objects, interaction context information, and content information of the sample content; Using a content processing model, feature extraction is performed on the object attribute information, interaction context information, and content information of the sample recommendation object, respectively, to obtain sample object attribute features, sample environment features, and sample content feature information. Through a content processing model, feature interaction is performed on the sample object attribute features, the sample environment features, and the sample content feature information to obtain sample interaction feature information; Based on the sample interaction feature information, predict the sample interaction probability information of the sample recommendation object generating interactive behavior on the sample content; Based on the sample interaction probability information and the interaction tag information, the model parameters of the content processing model are adjusted to obtain the trained content processing model.

9. The content recommendation method as described in claim 8, characterized in that, Feature extraction processing is performed on the content information of the sample content, including: The content processing model is used to extract features from the content information of the sample content to obtain initial sample feature information. The content information of the sample is subjected to residual processing to obtain semantic residual features; The semantic residual features and the initial sample feature information are fused to obtain the sample content feature information.

10. The content recommendation method as described in claim 3, characterized in that, The step of extracting features from the content to be recommended to obtain initial content feature information includes: The content feature extraction model is used to extract features from the content to be recommended to obtain initial content feature information; Before extracting features from the content to be recommended using a content feature extraction model to obtain initial content feature information, the method further includes: Obtain second training data, which includes multiple sample contents, as well as positive and negative sample contents of the sample contents; Using the content feature extraction model, feature extraction processing is performed on the sample content, the positive sample content, and the negative sample content of the sample content to obtain the semantic features of the sample content, the semantic features of the positive sample content, and the semantic features of the negative sample content. Based on the semantic features of the sample content and the semantic features of the positive sample content, the positive similarity between the sample content and the positive sample content is determined, and based on the semantic features of the positive sample content and the semantic features of the negative sample content, the negative similarity between the sample content and the negative sample content is determined. Based on the positive and negative similarities of the multiple sample contents, the model parameters of the content feature extraction model are adjusted to obtain the trained content feature extraction model.

11. The content recommendation method as described in claim 10, characterized in that, The method further includes: Obtain a performance evaluation dataset, which includes multiple evaluation items; The trained content feature extraction model is used to extract features from the evaluation content to obtain the semantic features of the evaluation content. Based on the semantic features of the evaluation content and the semantic features of the reference evaluation content, similar evaluation content that is similar to the evaluation content is determined from the reference evaluation content, wherein the reference evaluation content is other evaluation content besides the evaluation content. Based on the content information of the evaluation content and the content information of the similar evaluation content, the similarity between the evaluation content and the similar evaluation content is evaluated to obtain the similarity degree information between the similar evaluation content and the evaluation content. Based on the similarity information corresponding to the similar evaluation content, the target evaluation content is determined from the similar evaluation content; Based on the number of similar evaluation contents corresponding to the multiple evaluation contents and the number of target evaluation contents corresponding to the multiple evaluation contents, the evaluation result information of the trained content feature extraction model is determined, so as to adjust the trained content feature extraction model according to the evaluation result information.

12. A content recommendation device, characterized in that, include: An information processing unit is used to determine multiple pieces of content to be recommended based on the content recommendation request of the object to be recommended, and to extract the interaction feature information of the content to be recommended in the current recommendation scenario, wherein the current recommendation scenario is a scenario in which content is recommended to the object to be recommended; The information acquisition unit is used to acquire historical interaction feature information of the content to be recommended in historical recommendation scenarios; An interaction prediction unit is used to predict the probability information of the object to be recommended generating interactive behavior on the content to be recommended, based on the current interaction feature information and the historical interaction feature information. The recommendation determination unit is used to determine the target recommended content from the plurality of content to be recommended based on the interaction probability information; The content recommendation unit is used to recommend the target content to the object to be recommended.

13. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to perform the steps in the content recommendation method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the content recommendation method according to any one of claims 1 to 11.

15. 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 steps of the content recommendation method according to any one of claims 1 to 11.

Images