Content recommendation method and related device

Abstract

The embodiment of the invention discloses a content recommendation method and a related device, and the method comprises the steps: analyzing the corresponding similarity between a to-be-analyzed content and a plurality of contents according to the modal information corresponding to the plurality of contents with behavior information and the to-be-analyzed content in combination with the two dimensions of the modal information and the behavior information, and carrying out the analysis of the similarity between the to-be-analyzed content and the plurality of contents. Therefore, based on the similarity, the behavior information corresponding to the plurality of contents can be accurately and reasonably referenced, and the behavior information corresponding to the to-be-analyzed contents is formed. The relevance between the behavior information and the behavior information corresponding to the multiple pieces of content meets the similarity of the content to be analyzed and the multiple pieces of content in the two information dimensions of behavior information and modal information; therefore, the behavior information corresponding to the to-be-analyzed content can accurately represent the selection behavior for selecting the to-be-analyzed content in the content recommendation process, then the content recommendation model can effectively recommend the to-be-analyzed content based on the behavior information, and the content recommendation efficiency and universality are improved.

Description

Technical Field

[0001] This application relates to the technical field of data analysis, and in particular, to a content recommendation method and related devices. Background Art

[0002] Recommendation algorithms are widely used in various application scenarios. For example, when shopping online, recommendation algorithms can analyze users' historical purchase behaviors to recommend products with a higher purchase intention to users.

[0003] In the related art, during the training process of the recommendation algorithm, each recommendable content generates corresponding behavior information, which is used to represent the behavior of the user selecting the corresponding content. After determining the historical behavior information based on the user's historical behaviors through the algorithm, the historical behavior information can be matched with the behavior information corresponding to each recommendable content, and the recommendable content corresponding to the behavior information with a higher matching degree is determined as the actually recommended content.

[0004] However, since the behavior information in the related art corresponds one-to-one with the recommendable content, when new recommendable content appears, it is necessary to retrain the recommendation algorithm to obtain the behavior information corresponding to the new recommendable content, so as to be able to recommend the new recommendable content. Therefore, the recommendation algorithm in the related art can only recommend the recommendable content involved in the training process, and when applied to a new content domain, it cannot accurately recommend the new content in the new content domain, has poor versatility, and requires a large amount of training resources. Summary of the Invention

[0005] To solve the above technical problems, this application provides a content recommendation algorithm, which can analyze the similarities between the content to be analyzed and multiple contents in terms of behavior information and modal information according to the modal information respectively corresponding to the multiple contents with behavior information, so as to accurately construct the behavior information of the content to be analyzed based on the behavior information corresponding to the multiple contents, enable the content recommendation model to accurately analyze the probability of the content to be analyzed being selected based on the behavior information corresponding to the content to be analyzed, and then can recommend the content to be analyzed.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] In a first aspect, the embodiments of this application disclose a content recommendation method, and the method includes:

[0008] Obtain a set of recommendable contents, where multiple contents included in the set of recommendable contents respectively have corresponding behavior information, the behavior information is used to represent the selection behavior of selecting the corresponding content, and the content recommendation model is used to determine the content with the highest probability of being selected from the set of recommendable contents based on the behavior information;

[0009] Through a similarity analysis model, according to the modal information corresponding to the content to be analyzed and the multiple pieces of content respectively, determine the similarity between the content to be analyzed and the multiple pieces of content. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any one of the multiple pieces of content, and the modal information is used to characterize the corresponding content;

[0010] According to the similarity between the content to be analyzed and the multiple pieces of content, and the behavior information corresponding to the multiple pieces of content respectively, generate the behavior information corresponding to the content to be analyzed, and add the content to be analyzed to the set of recommendable content. The similarity between the content to be analyzed and the target content is used to characterize the reference strength of the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

[0011] In a second aspect, an embodiment of the present application discloses a content recommendation device, which includes a first acquisition unit, a first determination unit, and a generation unit:

[0012] The first acquisition unit is configured to acquire a set of recommendable content. Each of the multiple pieces of content included in the set of recommendable content has corresponding behavior information. The behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model is used to determine the content with the highest selection probability from the set of recommendable content based on the behavior information;

[0013] The first determination unit is configured to, through a similarity analysis model, according to the modal information corresponding to the content to be analyzed and the multiple pieces of content respectively, determine the similarity between the content to be analyzed and the multiple pieces of content. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any one of the multiple pieces of content, and the modal information is used to characterize the corresponding content;

[0014] The generation unit is configured to, according to the similarity between the content to be analyzed and the multiple pieces of content, and the behavior information corresponding to the multiple pieces of content respectively, generate the behavior information corresponding to the content to be analyzed, and add the content to be analyzed to the set of recommendable content. The similarity between the content to be analyzed and the target content is used to characterize the reference strength of the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

[0015] In a possible implementation manner, the device further includes a first training unit:

[0016] The first training unit is configured to train the similarity analysis model according to the similarity between the behavior information respectively corresponding to the multiple contents and the modality information respectively corresponding to the multiple contents, so that the first similarity determined by the similarity analysis model according to the modality information respectively corresponding to the multiple contents is greater than the second similarity. The first similarity is the similarity between the target content and the first content set, and the second similarity is the similarity between the target content and the second content set. The similarity between the target content and the first content set in the dimension of behavior information is greater than the similarity between the target content and the second content set in the dimension of behavior information. The first content set and the second content set are content sets divided from the multiple contents.

[0017] In a possible implementation manner, the first training unit is specifically configured to:

[0018] Take the multiple contents as the target content respectively, determine multiple first contents in the multiple contents that have the highest similarity with the target content in the dimension of behavior information to form the first content set, and determine multiple second contents in the multiple contents to form the second content set. The multiple second contents do not include the first content;

[0019] Through the initial similarity analysis model, determine the first similarity and the second similarity according to the modality information respectively corresponding to the multiple contents;

[0020] Construct a loss function corresponding to the initial similarity analysis model according to the first similarity and the second similarity. The loss function is positively correlated with the first similarity and negatively correlated with the second similarity;

[0021] Adjust the model parameters of the initial similarity analysis model according to the loss function so that the loss function is greater than a loss threshold, and obtain the similarity analysis model. The loss threshold satisfies that the first similarity is greater than the second similarity.

[0022] In a possible implementation manner, the content recommendation model is trained through the following method:

[0023] Obtain historical behavior records and the initial behavior information respectively corresponding to the multiple contents. The historical behavior records are used to record multiple selection behaviors performed by a first target object in a historical period. The multiple selection behaviors are used to select multiple selected contents. The historical selection behavior records have corresponding sample contents. The sample content is the content selected by the first target object in the next selection behavior after completing the multiple selection behaviors. The multiple contents include the multiple selected contents and the sample contents;

[0024] Through an initial content recommendation model, historical behavior information is determined according to the initial behavior information corresponding to the multiple selected contents, and the historical behavior information is used to characterize the multiple selection behaviors;

[0025] Adjust the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to the multiple contents, so that the matching degree between the historical behavior information determined by the initial content recommendation model and the initial behavior information corresponding to the sample content is maximized, and the content recommendation model and the behavior information corresponding to the multiple contents are obtained.

[0026] In a possible implementation manner, the historical behavior record is used to record multiple selection behaviors performed by the first target object in a historical period and the execution order of the multiple selection behaviors. Determining the historical behavior information according to the initial behavior information corresponding to the multiple selected contents includes:

[0027] According to the execution order corresponding to the multiple selection behaviors, order information corresponding to the multiple selection behaviors is determined, and the order information is used to characterize the execution order of the corresponding selection behavior;

[0028] According to the initial behavior information of the selected content corresponding to the target selection behavior and the order information corresponding to the target selection behavior, behavior representation information corresponding to the target selection behavior is generated. The behavior representation information corresponding to the target selection behavior is used to characterize the target selection behavior executed in the target execution order. The target selection behavior is any one of the multiple selection behaviors, and the target execution order is the execution order of the target selection behavior in the multiple selection behaviors;

[0029] According to the behavior representation information corresponding to the multiple selection behaviors, the historical behavior information is determined;

[0030] The adjusting the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to the multiple contents includes:

[0031] Adjust the model parameters corresponding to the initial content recommendation model, the initial behavior information corresponding to the multiple contents, and the execution order information corresponding to the multiple execution orders. The multiple execution orders are the execution orders corresponding to the multiple selection behaviors.

[0032] In a possible implementation manner, the initial content recommendation model includes N layers of information processing layers, and the N layers of information processing layers are used to perform N rounds of information processing. Determining the historical behavior information according to the behavior representation information corresponding to the multiple selection behaviors includes:

[0033] In the i-th round of information processing, obtain the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing, and the sub-behavior information respectively corresponding to the multiple selection behaviors in the first round of information processing is the behavior characterization information;

[0034] According to the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing, determine the similarities respectively corresponding to the target selection behavior and multiple reference behaviors in the dimension of sub-behavior information, where the multiple reference behaviors are the first M executed selection behaviors among the multiple selection behaviors, and the target selection behavior is the M-th executed selection behavior;

[0035] According to the similarities respectively corresponding to the target selection behavior and the multiple reference behaviors in the dimension of sub-behavior information, and the sub-behavior information respectively corresponding to the multiple reference behaviors in the i-th round of information processing, determine the target sub-behavior information corresponding to the target selection behavior in the (i + 1)-th round of information processing. The similarity between the target selection behavior and the target reference behavior in the dimension of sub-behavior information is used to represent the reference strength of the sub-behavior information corresponding to the target reference behavior in the i-th round of information processing when determining the target sub-behavior information, and the target reference behavior is any one of the multiple reference behaviors;

[0036] Determine the sub-behavior information corresponding to the last selection behavior output in the N-th round of information processing as the historical behavior information, where the last selection behavior is the latest executed selection behavior among the multiple historical behaviors.

[0037] In a possible implementation manner, the device further includes a second acquisition unit, a second determination unit, and a third determination unit:

[0038] The second acquisition unit is configured to acquire a historical behavior record to be analyzed, where the historical behavior record to be analyzed is used to record multiple historical selection behaviors executed by the object to be recommended in a historical period, and the multiple historical selection behaviors are used to select multiple historical contents, and the multiple historical contents are included in the content set to be recommended;

[0039] The second determination unit is configured to, through the content recommendation model, determine target historical behavior information according to the behavior information respectively corresponding to the multiple historical contents, where the target historical behavior information is used to represent the multiple historical selection behaviors;

[0040] The third determination unit is configured to, based on the highest matching degree between the target historical behavior information and the behavior information corresponding to the content to be analyzed in the content set to be recommended, determine the content to be analyzed as the content with the highest probability of being selected by the object to be recommended.

[0041] In a possible implementation, the device further includes a recommendation unit:

[0042] The recommendation unit is configured to recommend the content to be analyzed to the object to be recommended based on determining that the content to be analyzed is the content with the highest selection probability for the object to be recommended.

[0043] In a possible implementation, the device further includes an update unit:

[0044] The update unit is configured to update the model parameters corresponding to the content recommendation model and the behavior information corresponding to the content in the set of content that can be recommended, so that the matching degree between the historical behavior information determined by the updated content recommendation model based on the target historical behavior record and the behavior information corresponding to the updated content to be analyzed is the highest, and the sample content corresponding to the target historical behavior record is the content to be analyzed.

[0045] In a possible implementation, the content to be analyzed is any one of a plurality of cross-domain contents, the plurality of cross-domain contents correspond to a first content field, the plurality of contents correspond to a second content field, and the device further includes a third acquisition unit, a fourth determination unit, and an adjustment unit:

[0046] The third acquisition unit is configured to acquire cross-domain historical behavior records, where the cross-domain historical behavior records are used to record a plurality of cross-domain selection behaviors performed by a second target object in a historical period, the plurality of cross-domain selection behaviors are used to select the content in the plurality of cross-domain contents, and the sample content corresponding to the cross-domain historical behavior records is the target cross-domain content in the plurality of cross-domain contents;

[0047] The fourth determination unit is configured to determine cross-domain historical behavior information through the initial content recommendation model according to the behavior information corresponding to the content selected by the plurality of cross-domain selection behaviors, where the cross-domain historical behavior information is used to characterize the plurality of cross-domain selection behaviors;

[0048] The adjustment unit is configured to adjust the model parameters corresponding to the initial content recommendation model and the behavior information corresponding to the plurality of cross-domain contents respectively, so that the matching degree between the cross-domain historical behavior information determined by the initial content recommendation model and the behavior information corresponding to the target cross-domain content is the largest, to obtain a cross-domain content recommendation model and the updated behavior information corresponding to the plurality of cross-domain contents respectively, and the cross-domain content recommendation model is configured to determine the cross-domain content with the highest selection probability from the plurality of cross-domain contents according to the updated behavior information corresponding to the plurality of cross-domain contents respectively.

[0049] In a third aspect, an embodiment of the present application discloses a computer device, and the computer device includes a processor and a memory:

[0050] The memory is used to store a computer program and transmit the computer program to the processor;

[0051] The processor is used to execute the content recommendation method described in any one of the first aspect according to the instructions in the computer program;

[0052] In a fourth aspect, an embodiment of the present application discloses a computer-readable storage medium, which is used to store a computer program, and the computer program is used to execute the content recommendation method described in any one of the first aspect;

[0053] In a fifth aspect, an embodiment of the present application discloses a computer program product including a computer program. When it runs on a computer device, it causes the computer device to execute the content recommendation method described in any one of the first aspect.

[0054] As can be seen from the above technical solutions, when generating the behavior information corresponding to the new content in this application, it is not necessary to train the content recommendation model. Instead, it is possible to analyze the similarity between the content of the generated behavior information and the content of the behavior information to be generated in terms of both modality information and behavior information based on the modality information, and generate the new behavior information by referring to the generated behavior information based on this similarity. In this application, the content recommendation model can predict the content with the highest selection probability based on the behavior information corresponding to multiple contents in the content set to be recommended, so as to perform content recommendation. Through pre-training, a similarity analysis model can be obtained. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any content among the multiple contents. That is, on the one hand, this similarity is determined based on the modality information, so it can represent a certain similarity in the modality information dimension; on the other hand, this similarity analysis model can make the determined similarity represent the similarity in behavior information. Therefore, the similarity determined based on this similarity analysis model can represent the similarity in both the modality information and behavior information dimensions. Furthermore, based on this similarity analysis model, the similarity between the content to be analyzed and multiple contents can be determined according to the modality information of the content to be analyzed and the multiple contents, and the behavior information corresponding to the content to be analyzed can be generated by referring to the behavior information corresponding to the multiple contents based on this similarity. Thus, the behavior information corresponding to the content to be analyzed can accurately and reasonably represent the selection behavior of selecting the content to be analyzed. At the same time, since the behavior information corresponding to multiple contents is suitable for the content recommendation model to perform content recommendation, the behavior information corresponding to the content to be analyzed determined in this way can enable the content recommendation model to accurately analyze whether the content to be analyzed is the content with the highest selection probability, so as to reasonably recommend the content to be analyzed. Since the above process only requires simple similarity analysis based on the modality information used to represent the content itself and is easy to obtain, and there is no need to generate the behavior information corresponding to the new content through model training, it is possible to quickly and efficiently implement content recommendation for the new content while ensuring the accuracy of content recommendation, reducing the training cost and consumption required for content recommendation of the new content. Brief Description of the Drawings

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0056] Figure 1Schematic diagram of a content recommendation method in an actual application scenario provided by an embodiment of the present application;

[0057] Figure 2 Flowchart of a content recommendation method provided by an embodiment of the present application;

[0058] Figure 3 Schematic diagram of a content recommendation method provided by an embodiment of the present application;

[0059] Figure 4 Flowchart of a content recommendation method in an actual application scenario provided by an embodiment of the present application;

[0060] Figure 5 Schematic diagram of a content recommendation method in an actual application scenario provided by an embodiment of the present application;

[0061] Figure 6 Block diagram of a content recommendation device provided by an embodiment of the present application;

[0062] Figure 7 Structural diagram of a terminal provided by an embodiment of the present application;

[0063] Figure 8 Structural diagram of a server provided by an embodiment of the present application. Detailed implementation manners

[0064] The embodiments of the present application will be described below with reference to the accompanying drawings.

[0065] Automated and intelligent content recommendation technologies have been widely applied in multiple fields. For example, in the online shopping field, the content can be the products in the shopping platform. By analyzing the products purchased by users in the historical period, the products with a relatively high probability of being selected by users next time can be predicted for recommendation; in the short video field, the content can be short videos. By analyzing the short videos watched by users in the historical period, the short videos that users are more interested in can be predicted for recommendation.

[0066] In the related art, content recommendation is usually achieved through content recommendation algorithms. During the algorithm training process, a content recommendation model for content recommendation and behavior information corresponding to multiple contents will be generated. The behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model can analyze the matching degree between the user's historical selection behavior and the selection behavior for each content through this behavior information, so as to determine the content with the highest selection probability for recommendation.

[0067] However, when new content appears, since this content has not participated in the training process, there is no accurate behavior information corresponding to this content. In the related art, it is necessary to retrain the content recommendation model based on the new content, resulting in a large amount of resources being consumed for model training when making content recommendations for new content. Furthermore, when the content to be recommended changes, it is difficult for the related art to efficiently provide an accurate content recommendation function, and the generality of content recommendation is poor, making it difficult to achieve content recommendation across content domains.

[0068] To solve the above technical problems, the present application provides a content recommendation method. According to the modality information respectively corresponding to multiple contents with behavior information and the content to be analyzed, by combining the two dimensions of modality information and behavior information, the similarity respectively corresponding between the content to be analyzed and the multiple contents can be analyzed. Thus, based on this similarity, the behavior information respectively corresponding to the multiple contents can be relatively accurately and reasonably referred to, constituting the behavior information corresponding to the content to be analyzed, such that the relevance between the behavior information and the behavior information respectively corresponding to the multiple contents satisfies the similarity between the content to be analyzed and the multiple contents in the two information dimensions of behavior information and modality information. As a result, the behavior information corresponding to the content to be analyzed can accurately represent the selection behavior of selecting the content to be analyzed during the content recommendation process, and furthermore, the content recommendation model can effectively recommend the content to be analyzed based on this behavior information. It can be seen that the process of determining behavior information in the present application does not require re-training the model, and the accurate constitution of the behavior information corresponding to the new content can be completed through a relatively efficient similarity analysis method, improving the content recommendation efficiency and generality.

[0069] It can be understood that this method can be applied to a computer device, which is a computer device capable of performing content recommendation, such as a terminal device or a server. This method can be independently executed by the terminal device or the server, or can also be applied to a network scenario where the terminal device and the server communicate, and is executed in cooperation with the terminal device and the server. Among them, the terminal device can be a device such as a mobile phone, a tablet computer, a notebook computer, or a desktop computer. The terminal device can also include various virtual reality devices, such as augmented reality (AR) devices, such as AR glasses, AR screens, etc., and can also include virtual reality (VR) devices, such as head-mounted VR glasses, etc. The server can be understood as an application server or a Web server. In actual deployment, this server can be an independent server, a cluster server, or a cloud server, etc.

[0070] This application also relates to artificial intelligence technology. Artificial Intelligence (AI) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, enabling the machines to have the functions of perception, reasoning, and decision-making.

[0071] Artificial intelligence technology is an interdisciplinary subject with a wide range of fields involved, including both hardware-level and software-level technologies. Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and intelligent transportation.

[0072] This application mainly relates to machine learning technology among them. Machine Learning (ML) is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning.

[0073] This application can efficiently complete the effective training of content recommendation models and similarity analysis models by using machine learning technology.

[0074] To facilitate the understanding of the technical solution provided by this application, next, a content recommendation method provided by this application will be introduced in combination with an actual application scenario.

[0075] See Figure 1 , Figure 1A schematic diagram of a content recommendation method in an actual application scenario provided by this application. In this actual application scenario, the computer device can be a server 101 with content recommendation capabilities. Content 1, Content 2, and Content 3 are contents that participated in the model training process and thus have corresponding behavior information. The content to be analyzed is a new content that does not yet have behavior information. Each content has corresponding modality information, which is used to characterize the content itself. For example, it can be the image information, text information, etc. corresponding to the content.

[0076] The server 101 can determine the corresponding similarities between the content to be analyzed and Content 1, Content 2, and Content 3 respectively through a similarity analysis model, as Figure 1 shown. This similarity analysis model can satisfy that the similarity determined based on the modality information corresponding to any two contents is positively correlated with the similarity between these two contents in the dimension of behavior information. That is, if the similarity between two contents in the dimension of behavior information is higher, then the similarity determined by the similarity analysis model based on the modality information of these two contents is higher. For example, the similarity determined based on the modality information of Content A and Content B is proportional to the similarity B between Content A and Content B in the dimension of behavior information.

[0077] Therefore, the similarity determined through this similarity analysis model can characterize the similarity in behavior information between the content to be analyzed and multiple contents based on these two dimensions of behavior information and modality information. Thus, the behavior information of multiple contents can be referred to based on this similarity to determine the behavior information corresponding to the content to be analyzed. For example, the server 101 can use Similarity 1, Similarity 2, and Similarity 3 as weights to fuse Behavior Information 1, Behavior Information 2, and Behavior Information 3 to obtain the behavior information corresponding to the content to be analyzed, so that the parts in each generated behavior information that are more closely related to the content to be analyzed can be fused into the behavior information corresponding to the content to be analyzed, thereby obtaining more accurate behavior information. Since Behavior Information 1, Behavior Information 2, and Behavior Information 3 are all behavior information applicable to the content recommendation model, the behavior information determined in this way also satisfies the recommendation logic of the content recommendation model. Thus, the content recommendation model can effectively recommend the content to be analyzed based on the behavior information corresponding to the content to be analyzed.

[0078] It can be seen from this that when facing new content, this application does not need to retrain the model. Instead, it can perform similarity analysis based on easily obtainable modality information, and determine the behavior information corresponding to the new content based on the analyzed similarity and the already generated behavior information. Thus, on the premise of ensuring the accuracy of content recommendation, the content recommendation model can be more efficiently applied to the content recommendation of new content, improving the versatility of the content recommendation model.

[0079] Next, in combination with the accompanying drawings, the content recommendation method provided by this application will be introduced in detail.

[0080] See Figure 2 , Figure 2 which is a flowchart of a content recommendation method provided by an embodiment of this application. In this embodiment, the computer device can be any of the above computer devices with content recommendation functions. The method includes:

[0081] S201: Obtain a set of recommendable content.

[0082] The multiple contents included in the set of recommendable content respectively have corresponding behavior information. The behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model is used to determine the content with the highest selection probability from the set of recommendable content based on the behavior information. That is, the multiple contents are the contents that the content recommendation model can already accurately recommend. The behavior information corresponding to the multiple contents respectively is the behavior information that can accurately characterize the selection behavior, which can be generated during the model training process or generated through the technical solution of this application. The content recommendation model can analyze, based on the behavior information, which selection behavior the object has the highest probability of performing, so as to determine the content corresponding to this selection behavior as the content with the highest selection probability, and then recommend this content to the object to achieve content recommendation.

[0083] In this application, the content recommendation model can include multiple models. The content can be content in multiple fields, such as commodity content, short video content, etc., or can be content in multiple forms, such as video content, image content, text content, etc. The object can be any object that can select content, such as a user who shops or browses short videos. The selection behavior for the content can also include multiple types. The selection behaviors corresponding to the content in different fields can be different. For example, it can be the browsing behavior for short videos, the purchase behavior for commodities, etc., which is not limited here.

[0084] S202: Through the similarity analysis model, determine the similarity between the content to be analyzed and the multiple contents according to the modal information corresponding to the content to be analyzed and the multiple contents respectively.

[0085] It can be understood that each content usually has corresponding modal information. The modal information is used to characterize the corresponding content, so that the details of the content can be understood through the modal information. For example, the modal information can be a commodity picture, a commodity name, a short video image, etc. When the modal information of two contents is relatively similar, to a certain extent, it can indicate that the two contents themselves are relatively similar. However, when recommending similar contents only based on the similarity of the modal information, the probability of inaccurate content recommendation is relatively high.

[0086] For example, in the digital field, the modal information of "apple" usually refers to the content of mobile phones. In the food field, the modal information of "apple" usually refers to the content of fruits. Therefore, when there is behavioral information corresponding to the content of mobile phones in the digital field, if only the similarity in the dimension of modal information is used to determine the behavioral information corresponding to the new content in the food field, it will lead to relatively similar behavioral information for the two completely different selection behaviors of "selecting fruits" and "selecting mobile phones", resulting in the content recommendation model being unable to accurately recommend these two contents based on the behavioral information.

[0087] The present application can pre-train a similarity analysis model through the behavioral information and modal information respectively corresponding to multiple contents. When determining the similarity based on the modal information, the similarity determined by this similarity analysis model is positively correlated with the similarity between the behavioral information corresponding to different contents, that is, this similarity analysis model has the ability to analyze the similarity in the dimension of behavioral information through modal information. At the same time, since this similarity is determined based on the modal information, this similarity can reflect the similarity in the modal information to a certain extent. Therefore, the similarity determined by the similarity analysis model can represent the similarity of multiple contents in the two dimensions of behavioral information and modal information. Therefore, even if the modal information corresponding to different contents is relatively similar, the present application can accurately analyze the differences between different contents in the dimension of behavioral information.

[0088] For example, in the embodiment of the present application, the similarity between the content to be analyzed and multiple contents can be determined through the similarity analysis model, where the similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavioral information, and the target content can be any content among the multiple contents.

[0089] S203: Generate the behavioral information corresponding to the content to be analyzed according to the similarity between the content to be analyzed and multiple contents, and the behavioral information respectively corresponding to the multiple contents, and add the content to be analyzed to the set of recommendable contents.

[0090] From the above analysis, it can be seen that the similarity determined by this application can represent the comprehensive similarity between the content to be analyzed and multiple contents in terms of both behavioral information and modal information. Thus, the computer device can reasonably refer to the generated behavioral information based on this similarity, and accurately extract from it the information part that is relatively closely related to the content to be analyzed, thereby constituting the behavioral information corresponding to the content to be analyzed, enabling this behavioral information to more accurately represent the selection behavior of choosing the content to be analyzed. The similarity between the content to be analyzed and the target content is used to represent the degree of reference to the behavioral information corresponding to the target content when generating the behavioral information corresponding to the content to be analyzed. That is, the higher the similarity, the higher the correlation between the behavioral information corresponding to the content to be analyzed and the behavioral information of the target content, and the more information parts are extracted from the behavioral information corresponding to the target content.

[0091] For example, the method for determining the behavioral information of the content to be analyzed can be shown by the following formula:

[0092]

[0093] Wherein, is the modal information corresponding to the content to be analyzed, is the modal information corresponding to the first content among multiple contents, |V S | is the total number of multiple contents, is the similarity determined by the similarity analysis model based on the modal information corresponding to the content to be analyzed and the modal information corresponding to the first content. This formula means to select the m contents with the highest similarity from multiple contents for reference to determine the behavioral information corresponding to the content to be analyzed, and the reference method is shown by the following formula: Wherein,

[0094]

[0095]

[0096] Wherein, score i,j refers to the reference weight of the j-th content among multiple contents when constructing the behavioral information of the content to be analyzed, which is used to control the degree of reference to the j-th content. It can be seen from the formula that the reference weight is positively correlated with the similarity, is the finally determined behavioral information of the content to be analyzed, is the behavioral information corresponding to the k-th content among multiple contents.

[0097] In addition, since the behavior information corresponding to multiple contents is already applicable to content recommendation by means of a content recommendation model, and this behavior information is the behavior information that the content recommendation model can accurately analyze, the behavior information corresponding to the content to be analyzed constructed based on this behavior information also has the information characteristics that can enable the content recommendation model to accurately analyze.

[0098] In summary, the behavior information corresponding to the content to be analyzed determined by this application can enable the content recommendation model to accurately recommend the content to be analyzed. Furthermore, the computer device can add the content to be analyzed to the set of recommendable contents, so that when there is a content recommendation requirement subsequently, the content recommendation model can complete the recommendation of the content to be analyzed.

[0099] It can be seen from the above technical solution that when generating the behavior information corresponding to the new content in this application, there is no need to train the content recommendation model. Instead, it is possible to analyze the similarity between the content of the generated behavior information and the content of the behavior information to be generated in terms of both the modality information and the behavior information based on the modality information, and generate the new behavior information by referring to the generated behavior information based on this similarity, so that the behavior information corresponding to the content to be analyzed can more accurately and reasonably represent the selection behavior of selecting the content to be analyzed. At the same time, since the behavior information corresponding to multiple contents is applicable to content recommendation by the content recommendation model, furthermore, the behavior information corresponding to the content to be analyzed determined by this method can enable the content recommendation model to accurately analyze whether the content to be analyzed is the content with the highest selection probability, so as to reasonably recommend the content to be analyzed. Since the above process only needs to perform a simple similarity analysis based on the modality information used to represent the content itself and is easy to obtain, and there is no need to generate the behavior information corresponding to the new content by means of model training, it is possible to quickly and efficiently realize the content recommendation for the new content while ensuring the accuracy of the content recommendation, and reduce the training cost and consumption required for content recommendation for the new content.

[0100] As mentioned above, the similarity analysis model can be pre-trained by the computer device or can be obtained by training when the computer device analyzes the behavior information corresponding to the content to be analyzed. Next, the training process of the similarity analysis model will be introduced in detail.

[0101] In a possible implementation manner, before determining the similarity between the content to be analyzed and multiple contents according to the modality information corresponding to the content to be analyzed and the multiple contents respectively, the computer device can train the similarity analysis model in the following manner:

[0102] A computer device can train a similarity analysis model based on the similarity between the behavior information corresponding to multiple contents and the modality information corresponding to multiple contents, so that the first similarity determined by the similarity analysis model according to the modality information corresponding to multiple contents is greater than the second similarity. Among them, the first similarity is the similarity between the target content and the first content set, the second similarity is the similarity between the target content and the second content set, the similarity between the target content and the first content set in the dimension of behavior information is greater than the similarity between the target content and the second content set in the dimension of behavior information, the target content is any content among multiple contents, and the first content set and the second content set are content sets divided from multiple contents. That is, the computer device can use the similarity between behavior information as the basis for measuring whether the output of the similarity analysis model is accurate, so as to adjust the model parameters corresponding to the similarity analysis model and make the similarity analysis model achieve the above similarity analysis effect.

[0103] The similarity analysis model can be any model capable of performing similarity analysis based on modality information. For example, it can be a sentence vector contrast learning model (Simple Contrastive Learning of Sentence Embeddings, abbreviated as SIMCSE) used in the field of natural language processing, etc., which is not limited here.

[0104] In actual application of the similarity analysis model, the computer device can directly call the pre-trained similarity analysis model, or can retrain the similarity analysis model based on the updated content in the current recommended content set, so that the similarity analysis model matches the continuously updated content characteristics to perform more accurate analysis of similarity, thereby further improving the generation accuracy of behavior information.

[0105] Specifically, when training a similarity analysis model based on the similarity between the behavior information corresponding to multiple contents and the modality information corresponding to multiple contents, the computer device can train the model in the following way:

[0106] First, the computer device can use multiple contents as target contents respectively, determine multiple first contents with the highest similarity to the target contents in the dimension of behavior information from the multiple contents to form a first content set, and determine multiple second contents from the multiple contents to form a second content set. The multiple second contents do not include the first contents, that is, the second contents are not the multiple contents with the highest similarity to the target contents in the dimension of behavior information, and can be regarded as contents with relatively low similarity to the target contents in the dimension of behavior information. Among them, the number of the first contents can be adjusted based on the training effect during the training process of the similarity model, which is not limited here. Thus, the similarity between the target content and the first content set in the dimension of behavior information is higher than the similarity between the target content and the second content set in the dimension of behavior information.

[0107] Then, the computer device can use the initial similarity analysis model to determine a first similarity and a second similarity according to the modality information corresponding to the multiple contents respectively. The computer device can construct a loss function corresponding to the initial similarity analysis model based on the first similarity and the second similarity, making the loss function positively correlated with the first similarity and negatively correlated with the second similarity. Thus, the greater the model loss characterized by the loss function, the higher the positive correlation between the similarity determined by the initial similarity analysis model based on the modality information and the similarity in the dimension of behavior information, that is, the stronger the ability of the initial similarity analysis model to analyze the similarity in the dimension of behavior information based on the modality information.

[0108] Therefore, the computer device can adjust the model parameters of the initial similarity analysis model according to the loss function, making the loss function greater than the loss threshold. In this process, the initial similarity analysis model can learn how to accurately analyze the behavior information similarity based on the modality information, and then obtain a similarity analysis model. The loss threshold satisfies that the first similarity is greater than the second similarity.

[0109] Among them, the determination method of the loss function can be as follows:

[0110] The computer device can regard the target content as v i , and through the similarity in behavior information, k contents most similar to the target content can be obtained to form a first content set Identify the first first content corresponding to the target content v i . Then, regard as a positive pair, and regard as a negative pair. For the target content v i , the loss function can be expressed as:

[0111]

[0112] Among them, is the first similarity degree, is the second similarity degree, N is the number of the actually extracted second contents, t i is the content v i corresponding modal information. Thus, it can be seen that this loss function is positively correlated with the first similarity degree and negatively correlated with the second similarity degree.

[0113] The above content details the training process and model principle of the similarity analysis model. Next, the recommendation method and model training method for content recommendation by the content recommendation model will be introduced in detail.

[0114] First, the training process of the content recommendation model will be introduced in detail. In a possible implementation manner, the content recommendation model is trained through the following method:

[0115] The computer device can first obtain the historical behavior records and the initial behavior information corresponding to multiple contents. Among them, the historical behavior records are used to record multiple selection behaviors performed by the first target object in the historical period. The multiple selection behaviors are used to select multiple selected contents. The first target object can be any of the above objects capable of performing selection behaviors. Through this historical behavior record, the selection behavior characteristics of the first target object in the historical period can be characterized.

[0116] The historical selection behavior record has corresponding sample content. The sample content is the content selected by the first target object for the next selection behavior after performing multiple selection behaviors. For example, if the first target object purchases apples and bananas and then purchases oranges, the two behaviors of purchasing apples and purchasing bananas can be used as multiple selection behaviors performed in the historical period, and oranges are the sample content corresponding to the historical behavior record. In other words, the sample content is the content that the first target object actually most wants to select after performing multiple selection behaviors in the historical period, that is, the content that the first target object is most interested in. The multiple contents include multiple selected contents and sample content. That is, in the training process, the computer device can generate the behavior information corresponding to the selected contents and sample content, so that these contents can be recommended by the content recommendation model.

[0117] The computer device can determine the historical behavior information through the initial content recommendation model according to the initial behavior information corresponding to multiple selected contents. The historical behavior information is used to characterize multiple selection behaviors. That is, first, the initial content recommendation model can analyze the behavior characteristics represented by multiple selection behaviors in the historical period to generate the historical behavior information.

[0118] When determining the content with the highest selection probability, the initial content recommendation model can determine the matching degree between the historical behavior information and the behavior information corresponding to each content. This matching degree is used to represent the probability of performing the selection behavior corresponding to each content after multiple selection behaviors. The higher the matching degree, the greater the probability of selecting this content. Therefore, the greater the probability of this content being selected. As mentioned above, the sample content is the content with the highest selection probability of the first target object in fact. Therefore, in order to improve the accuracy of content recommendation, the computer device can adjust the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to each of the multiple contents, so that the matching degree between the historical behavior information determined by the initial content recommendation model and the initial behavior information corresponding to the sample content is maximized. Thus, during the adjustment process, the initial content recommendation model can learn how to accurately analyze the probability of content being selected, and then obtain the content recommendation model and the behavior information corresponding to each of the multiple contents.

[0119] Among them, the model parameters corresponding to the initial content recommendation model are used to generate historical behavior information, and the initial behavior information corresponding to each of the multiple contents is the basis for generating historical behavior information and performing behavior information matching. Therefore, during the adjustment process, the model parameters and these behavior information can be adjusted so that the content recommendation model can fully learn how to accurately analyze the probability of content being selected.

[0120] It can be understood that in addition to the content selected by the object in the historical period, the selection order of the object's selected content also plays a key role in determining the next content that the object is interested in. For example, assume that in the historical period, the object first selects the content in content field 1 and then selects the content in content field 2. Then, generally, the content field to which the object's next selected content belongs is closer to content field 2; conversely, if the object first selects the content in content field 2 and then selects the content in content field 1, then the object's next selected content is likely to be close to content field 1.

[0121] Based on this, in a possible implementation manner, in order to further improve the accuracy of content recommendation by the content recommendation model, the computer device can enable the content recommendation model to refer to both the content selected in the historical period and the selection order of the selected content for content recommendation. In this implementation manner, the historical behavior record is used to record the multiple selection behaviors performed by the first target object in the historical period and the execution order of the multiple selection behaviors.

[0122] When determining historical behavior information based on the initial behavior information corresponding to multiple selected contents respectively, in addition to including the mapping between the content and the behavior information, there is also a mapping relationship between the execution order of the selection behaviors and the order information. The computer device can determine the order information corresponding to multiple selection behaviors according to the execution order corresponding to multiple selection behaviors respectively. The order information is used to characterize the execution order of the corresponding selection behavior, and the order information corresponding to different execution orders is different. Thus, by adding the order information, the execution order characteristics of multiple selection behaviors can be characterized.

[0123] Taking the target selection behavior among multiple selection behaviors as an example, according to the initial behavior information of the selected content corresponding to the target selection behavior and the order information corresponding to the target selection behavior, the behavior representation information corresponding to the target selection behavior can be generated. The behavior representation information corresponding to the target selection behavior is used to characterize the target selection behavior executed in the target execution order, such as the behavior of "selecting content A as the sixth selection behavior among ten selection behaviors". The target selection behavior can be any one of multiple selection behaviors, and the target execution order is the execution order of the target selection behavior among multiple selection behaviors.

[0124] The initial content recommendation model can determine historical behavior information according to the behavior representation information corresponding to multiple selection behaviors respectively, so as to analyze the content selection probability from two dimensions: the content selected in the historical period and the order of selecting the content, achieving the effect of enriching the model reference dimension and improving the model recommendation accuracy.

[0125] At the same time, as can be seen from the above content, in this implementation, there is an additional adjustable dimension of "order information", that is, when adjusting the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to multiple contents respectively, the computer device can adjust the model parameters corresponding to the initial content recommendation model, the initial behavior information corresponding to multiple contents respectively, and the order information corresponding to multiple execution orders respectively. The multiple execution orders are the execution orders corresponding to multiple selection behaviors respectively. Since the adjustable dimension for model training increases, to a certain extent, the granularity of model training is more refined. Thus, based on the behavior information and order information, the content recommendation model obtained through training can perform more accurate content recommendation.

[0126] The model structure of this content recommendation model can also include various types. Next, a model architecture for determining historical behavior information in a content recommendation model will be introduced in detail. In this implementation, the initial content recommendation model includes N layers of information processing layers, and the N layers of information processing layers are used to perform N rounds of information processing. The information processing layer can include various types, such as a self-attention mechanism layer, etc. When determining historical behavior information according to the behavior representation information corresponding to multiple selection behaviors respectively, the following information processing can be performed through the initial content recommendation model. Taking the i-th round of information processing as an example.

[0127] In the i-th round of information processing, sub-behavior information respectively corresponding to multiple selection behaviors in the i-th round of information processing can be obtained, where the sub-behavior information respectively corresponding to the multiple selection behaviors in the first round of information processing is behavior characterization information, that is, the behavior characterization information determined by the multiple selection behaviors through the above method is the input of the first round of information processing.

[0128] According to the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing, the similarities respectively corresponding to the target selection behavior and multiple reference behaviors in the dimension of sub-behavior information can be determined. The multiple reference behaviors are the first M executed selection behaviors among the multiple selection behaviors, and the target selection behavior is the M-th executed selection behavior, that is, the reference behaviors are from the first selection behavior to the target selection behavior. Then, according to the similarities respectively corresponding to the target selection behavior and the multiple reference behaviors in the dimension of sub-behavior information, and the sub-behavior information respectively corresponding to the multiple reference behaviors in the i-th round of information processing, the target sub-behavior information corresponding to the target selection behavior in the (i + 1)-th round of information processing can be determined. The similarity between the target selection behavior and the target reference behavior in the dimension of sub-behavior information is used to represent the reference strength of the sub-behavior information corresponding to the target reference behavior in the i-th round of information processing when determining the target sub-behavior information, and the target reference behavior is any one of the multiple reference behaviors. The purpose of this information processing method is to fuse the behavior characteristics of the target selection behavior and the behavior characteristics of the selection behaviors executed before the target selection behavior, so that the sub-behavior information corresponding to the output target selection behavior can represent the behavior characteristics of this series of selection behaviors from the first selection behavior to the target selection behavior in the execution order of the multiple selection behaviors.

[0129] Thus, in the process of layer-by-layer information processing, the sub-behavior information corresponding to each selection behavior can continuously extract and fuse the behavior characteristics of the reference behaviors, so that the correlation between the multiple selection behaviors can be continuously extracted based on the similarity in the dimension of sub-behavior information and continuously transmitted to the sub-behavior information corresponding to the selection behavior with a later execution order. Furthermore, after N rounds of information processing, the sub-behavior information corresponding to the last executed selection behavior can accurately represent the behavior characteristics of this series of selection behaviors from the first selection behavior to the last selection behavior in the execution order of the multiple selection behaviors. It can be seen that this sub-behavior information can play the role of representing historical behavior information.

[0130] Based on this, the model can determine the sub-behavior information corresponding to the end selection behavior in the output of the Nth round of information processing as historical behavior information, where the end selection behavior is the latest executed selection behavior among multiple historical behaviors. In this way, the determined historical behavior information can accurately represent multiple selection behaviors from two dimensions: the selection content and the content selection order of multiple selection behaviors, so that the content recommendation model can perform reasonable behavior information matching to accurately analyze the content with the highest selection probability.

[0131] As Figure 3 shown, Figure 3 it shows the method of determining historical behavior information through N self-attention mechanism layers. The content recommendation model can be a deep learning model for sequential recommendation (Self-Attentive Sequential Recommendation, abbreviated as SASRec). First, historical behavior records can be obtained as the content selected for the sth u selection, where s Figure 3 is 10 in u . The input behavior matrix of the lth self-attention layer can be expressed as where the behavior matrix refers to a matrix composed of sub-behavior information corresponding to multiple selection behaviors, and is the sub-behavior information corresponding to the ith selection behavior. Then, we obtain the behavior matrix of the (l + 1)th layer o and after passing through N self-attention mechanism layers, historical behavior information u o can be obtained (the sub-behavior information of the last selection behavior in the Nth layer is used as historical behavior information), which is expressed by the formula as follows:

[0132]

[0133]

[0134] where refers to selecting the sub-behavior information corresponding to the end selection behavior from the sub-behavior information output by the Nth self-attention mechanism layer.

[0135] The above content mainly introduces the training methods and operation principles of multiple models. Next, how to perform content recommendation based on these models in practical applications will be introduced.

[0136] In a possible implementation, when making content recommendations for a target object to be recommended, a computer device may first obtain a historical behavior record to be analyzed, which is used to record multiple historical selection behaviors performed by the target object to be recommended during a historical period. The multiple historical selection behaviors are used to select multiple historical contents, and the set of recommendable contents includes the multiple historical contents.

[0137] The computer device can use a content recommendation model to determine target historical behavior information based on the behavior information corresponding to the multiple historical contents respectively. The target historical behavior information is used to represent the multiple historical selection behaviors. Then, the matching degree between the target historical behavior information and the behavior information corresponding to the multiple contents included in the set of recommendable contents can be analyzed to analyze the probability of each of the multiple contents being selected. The greater the matching degree, the higher the probability of being selected.

[0138] Therefore, based on the fact that in the set of recommendable contents, the matching degree between the target historical behavior information and the behavior information corresponding to the content to be analyzed is the highest, the computer device can determine the content to be analyzed as the content with the highest probability of being selected by the target object to be recommended. Thus, in subsequent processes, the computer device can recommend the content to be analyzed to the target object to be recommended to best meet the content needs of the target object to be recommended. For example, the historical behavior record to be analyzed can be the record of the target object to be recommended watching short videos. Through this record, the short video with the highest probability of being watched by the target object to be recommended next can be analyzed, and then this short video can be recommended to the target object to be recommended to bring a better video watching experience to the target object to be recommended.

[0139] Among them, after determining the content with the highest probability of being selected, it can be recommended to the target object to be recommended by other operating parties, or the computer device can also perform the recommendation itself to further improve the efficiency and convenience of content recommendation and reduce the dependence on manual operations. For example, in a possible implementation, based on determining that the content to be analyzed is the content with the highest probability of being selected by the target object to be recommended, the computer device can directly recommend the content to be analyzed to the target object to be recommended without the intervention of other operating parties. For example, in the field of short videos, after determining the short video with the highest probability of being selected by the target object to be recommended, the computer device can directly push this short video to the target object to be recommended; in the field of online shopping, after determining the product with the highest probability of being selected by the target object to be recommended, this product can be directly displayed on the home page of the online shopping platform corresponding to the target object to be recommended for the target object to be recommended to purchase.

[0140] As can be seen from the above, the present application can quickly, efficiently, and accurately process the behavior information of new content. The processing device can directly use the content recommendation model that has been trained to accurately recommend new content based on the behavior information corresponding to the new content. In addition, in other implementation manners, in order to further improve the recommendation accuracy and rationality of new content recommendations, the computer device can also train the content recommendation model based on the behavior information corresponding to the new content, so that the content recommendation model can further learn the recommendation method of the new content.

[0141] Among them, the training methods based on the behavior of new content can include the following two:

[0142] Method 1:

[0143] In a possible implementation manner, the computer device can adjust based on the existing content recommendation model. The computer device can update the model parameters corresponding to the content recommendation model and the behavior information corresponding to the content in the recommended content set, so that the matching degree between the historical behavior information determined based on the target historical behavior record and the behavior information corresponding to the updated content to be analyzed is the highest. The sample content corresponding to the target historical behavior record is the content to be analyzed.

[0144] That is, in this implementation manner, the computer device does not need to retrain the content recommendation model, but can obtain the corresponding target historical behavior record for the content to be analyzed, and adjust the content recommendation model and behavior information obtained through the above training process based on the target historical behavior record, so that the content recommendation model can further learn how to accurately recommend the content to be analyzed and make up for the model knowledge that was not learned during the initial training process. Since this adjustment method is based on the content recommendation model that has been trained a large number of times, the training efficiency is relatively high compared to retraining. Therefore, the content recommendation model can be iterated quickly and efficiently during the content update process, so that the content recommendation model can maintain an accurate content recommendation effect even when the content is continuously updated.

[0145] Method 2:

[0146] As mentioned above, the content may include content in multiple content domains, and there may be significant differences between the content in different content domains. For example, there are significant differences between the content in the fruit domain and the content in the digital domain. Although the behavior information determined in the above manner can more accurately represent the selection behavior of selecting the content to be analyzed, it is still generated based on the behavior information of multiple generated contents and will be affected by the content domains to which the multiple contents belong. When there are significant differences between the content to be analyzed and the content domains to which the multiple contents belong, the content domains to which the multiple contents belong may affect the accuracy of the recommendation for the content to be analyzed.

[0147] Based on this, in a possible implementation manner, in order to further improve the recommendation accuracy for new content, the computer device can re-train the model for the content domain to which the new content belongs to obtain a content recommendation model for this content domain, and implement accurate content recommendation for the new content domain through this content recommendation model.

[0148] In this implementation manner, the content to be analyzed can be any one of multiple cross-domain contents. The multiple cross-domain contents correspond to the first content domain, and the multiple contents correspond to the second content domain, that is, the cross-domain contents and the multiple contents involved in the model training process are contents in different content domains.

[0149] After the computer device determines the behavior information corresponding to multiple new cross-domain contents based on the behavior information of the existing multiple contents in the above manner, it can obtain a cross-domain historical behavior record. The cross-domain historical behavior record is used to record multiple cross-domain selection behaviors performed by the second target object in the historical period. The multiple cross-domain selection behaviors are used to select the content in the multiple cross-domain contents. The sample content corresponding to the cross-domain historical behavior record is the target cross-domain content in the multiple cross-domain contents. That is, through the cross-domain historical behavior record, the content selection characteristics of the second target object in this first content domain can be characterized. Since the above content recommendation model is trained based on the content in multiple second content domains, this content recommendation model has strong pertinence for the second content domain. Although through the above manner, compared with the determination method in the related art that only determines behavior information based on modal information, it can also have a relatively high accuracy when recommending content for the content in the first content domain, but relatively speaking, the accuracy may be insufficient.

[0150] In this implementation manner, the computer device can determine cross-domain historical behavior information through an initial content recommendation model (that is, an initial model that has not undergone model training. At this time, the initial content recommendation model does not have pertinence in terms of content domain), according to the behavior information corresponding to the content selected by the multiple cross-domain selection behaviors. The cross-domain historical behavior information is used to characterize the multiple cross-domain selection behaviors.

[0151] Then, the computer device can adjust the model parameters corresponding to the initial content recommendation model and the behavior information corresponding to multiple cross-domain contents, so that the matching degree between the cross-domain historical behavior information determined by the initial content recommendation model and the behavior information corresponding to the target cross-domain content is maximized, obtaining the cross-domain content recommendation model and the updated behavior information corresponding to multiple cross-domain contents respectively. Thus, it can be seen from the above model training principle that the cross-domain content recommendation model can more specifically learn the content selection method of the object in the first content field, reducing the influence of the behavior information of multiple contents in the second content field on the content recommendation in the first content field. Therefore, the cross-domain content recommendation model can be used to determine the cross-domain content with the highest selection probability from multiple cross-domain contents according to the updated behavior information corresponding to multiple cross-domain contents respectively, that is, it can be used for more accurate content recommendation for the first content field.

[0152] In addition, since through the content recommendation method of the present application, the part with a relatively high correlation with the behavior information of cross-domain content can be extracted from the behavior information of multiple contents to generate the behavior information of cross-domain content, when training the cross-domain content recommendation model, the initial behavior information corresponding to multiple cross-domain contents already has a certain degree of accuracy. Compared with the randomly generated initial behavior information, more efficient model training can be carried out, and the efficiency of model fitting is higher, thus saving the resource consumption required for training the content recommendation model in the new field.

[0153] Combining the above various model training methods can meet different content recommendation requirements. For example, it can meet the efficient content recommendation for multiple content fields, or it can meet the high-precision content recommendation for a certain specific content field.

[0154] To facilitate the understanding of the technical solution provided by the present application, next, a content recommendation method provided by an embodiment of the present application will be introduced as a whole in combination with an actual application scenario.

[0155] See Figure 4 , Figure 4 is a flowchart of a content recommendation method in an actual application scenario provided by an embodiment of the present application. In this actual application scenario, the computer device can be a server for managing video push on a video platform, the content is a recommendable video, and the object is a user of the video platform. The method includes:

[0156] S401: Train a content recommendation model and a similarity analysis model through the historical behavior records in the first content field.

[0157] Steps S401 - S404 are the training and adjustment processes of the content recommendation model, as Figure 5As shown, first, the server can perform pre-training through the historical behavior records of the first content domain to obtain a content recommendation model and a similarity analysis model. The similarity analysis model can be trained through the historical behavior records and modal information of multiple content domains, enabling it to accurately output the similarity between contents based on modal information, and thus output the similarity between contents in terms of both modal information and behavior information. Through model training, the behavior information of the content corresponding to the first content domain can be obtained. The content recommendation model can be an ID-based Sequential Model, where each content corresponds to a unique ID, and the similarity analysis model can be a Cross-domain ID Matcher (CDIM).

[0158] S402: Determine the similarity through the similarity analysis model according to the modal information corresponding to the videos in the second content domain and the videos in the first content domain respectively.

[0159] This similarity can represent the similarity between the contents of the two domains in terms of both modal information and behavior information, so that the behavior information in the first content domain can be accurately referenced based on this similarity.

[0160] S403: Determine the behavior information corresponding to the videos in the second content domain according to the similarity and the behavior information corresponding to the videos in the first content domain respectively.

[0161] The server can select the m contents corresponding to the first content domain with the highest similarity according to the similarity, and input these contents into the ID Embedding Generator. The ID Embedding Generator is used to generate the behavior information of the new content based on the similarity as a reference weight and referring to the already generated behavior information, so that the behavior information corresponding to the content in the second content domain can be obtained.

[0162] S404: Update the content recommendation model according to the historical behavior records and behavior information of the second content domain.

[0163] To enable the content recommendation model to more accurately recommend content in the second content field, the server can adjust the content recommendation model based on the behavior information corresponding to the second content field obtained through the above methods. For example, it can adjust the model parameters of the encoder part in the sequence model, the behavior information corresponding to each piece of content, and the sequence information corresponding to each execution order. The encoder part is used to generate historical behavior information based on the behavior information and the sequence information. In addition to this adjustment method, the server can also retrain the content recommendation model based on the behavior information of the second content field for targeted recommendation of the content in the second content field.

[0164] S405: Obtain the historical behavior record to be analyzed corresponding to the object to be recommended.

[0165] The historical behavior record to be analyzed can be used to record the videos viewed by the object to be recommended in the historical period and the order of viewing the videos.

[0166] S406: Determine the video to be recommended with the highest selection probability according to the historical behavior record to be analyzed.

[0167] Through the content recommendation model, the historical behavior information corresponding to the historical behavior record to be analyzed can be determined, and the video to be recommended with the highest selection probability can be determined by the matching degree between the historical behavior information and the behavior information corresponding to multiple videos respectively.

[0168] S407: Recommend the video to be recommended to the object to be recommended.

[0169] As can be seen from the above content, the following are the improvement points of this application in actual applications:

[0170] (1) The behavior information of new content can be quickly generated using the generated behavior information, and this process does not require retraining the content recommendation model, improving the generation efficiency of the behavior information.

[0171] (2) The similarities in the two dimensions of modality information and behavior information analyzed by the similarity analysis model can accurately reference the generated behavior information, ensuring the accuracy of the behavior information of new content, and thus ensuring the recommendation accuracy for new content.

[0172] (3) The content recommendation model can be adjusted based on the behavior information of new content, ensuring that the content recommendation model can effectively recommend content in each content field.

[0173] Based on the content recommendation method provided in the above embodiments, this application also provides a content recommendation device. See Figure 6 , Figure 6The block diagram of a content recommendation device provided by this application. The content recommendation device 600 includes a first acquisition unit 601, a first determination unit 602, and a generation unit 603:

[0174] The first acquisition unit 601 is configured to acquire a set of recommendable content. Each of the multiple pieces of content included in the set of recommendable content has corresponding behavior information, and the behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model is used to determine the content with the highest selection probability from the set of recommendable content based on the behavior information;

[0175] The first determination unit 602 is configured to determine the similarity between the content to be analyzed and the multiple pieces of content according to the modality information corresponding to the content to be analyzed and the multiple pieces of content through a similarity analysis model. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any one of the multiple pieces of content, and the modality information is used to characterize the corresponding content;

[0176] The generation unit 603 is configured to generate the behavior information corresponding to the content to be analyzed according to the similarity between the content to be analyzed and the multiple pieces of content, and the behavior information corresponding to the multiple pieces of content respectively, and add the content to be analyzed to the set of recommendable content. The similarity between the content to be analyzed and the target content is used to characterize the reference strength of the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

[0177] In a possible implementation manner, the device further includes a first training unit:

[0178] The first training unit is configured to train the similarity analysis model according to the similarity between the behavior information corresponding to the multiple pieces of content respectively and the modality information corresponding to the multiple pieces of content, so that the first similarity determined by the similarity analysis model according to the modality information corresponding to the multiple pieces of content is greater than the second similarity. The first similarity is the similarity between the target content and the first content set, and the second similarity is the similarity between the target content and the second content set. The similarity between the target content and the first content set in the dimension of behavior information is greater than the similarity between the target content and the second content set in the dimension of behavior information. The first content set and the second content set are content sets divided from the multiple pieces of content.

[0179] In a possible implementation manner, the first training unit is specifically configured to:

[0180] Using the multiple contents as the target content respectively, determining multiple first contents with the highest similarity to the target content among the multiple contents in the dimension of behavior information to form the first content set, and determining multiple second contents among the multiple contents to form the second content set, where the multiple second contents do not include the first content;

[0181] Through the initial similarity analysis model, determine the first similarity and the second similarity according to the modality information respectively corresponding to the multiple contents;

[0182] Construct a loss function corresponding to the initial similarity analysis model according to the first similarity and the second similarity, where the loss function is positively correlated with the first similarity and negatively correlated with the second similarity;

[0183] Adjust the model parameters of the initial similarity analysis model according to the loss function to make the loss function greater than the loss threshold, and obtain the similarity analysis model, where the loss threshold satisfies that the first similarity is greater than the second similarity.

[0184] In a possible implementation, the content recommendation model is trained in the following way:

[0185] Obtain the historical behavior record and the initial behavior information respectively corresponding to the multiple contents. The historical behavior record is used to record multiple selection behaviors performed by a first target object in a historical period. The multiple selection behaviors are used to select multiple selected contents. The historical selection behavior record has corresponding sample contents. The sample content is the content selected by the first target object for the next selection behavior after completing the multiple selection behaviors. The multiple contents include the multiple selected contents and the sample content;

[0186] Through the initial content recommendation model, determine the historical behavior information according to the initial behavior information respectively corresponding to the multiple selected contents, where the historical behavior information is used to characterize the multiple selection behaviors;

[0187] Adjust the model parameters corresponding to the initial content recommendation model and the initial behavior information respectively corresponding to the multiple contents to maximize the matching degree between the historical behavior information determined by the initial content recommendation model and the initial behavior information corresponding to the sample content, and obtain the content recommendation model and the behavior information respectively corresponding to the multiple contents.

[0188] In a possible implementation, the historical behavior record is used to record multiple selection behaviors performed by the first target object in a historical period, and the execution order of the multiple selection behaviors. Determining historical behavior information according to the initial behavior information respectively corresponding to the multiple selected contents includes:

[0189] Determine the sequence information respectively corresponding to the multiple selection behaviors according to the execution order respectively corresponding to the multiple selection behaviors. The sequence information is used to represent the execution order of the corresponding selection behavior;

[0190] Generate the behavior representation information corresponding to the target selection behavior according to the initial behavior information of the selected content corresponding to the target selection behavior and the sequence information corresponding to the target selection behavior. The behavior representation information corresponding to the target selection behavior is used to represent the target selection behavior executed in the target execution order. The target selection behavior is any one of the multiple selection behaviors, and the target execution order is the execution order of the target selection behavior among the multiple selection behaviors;

[0191] Determine the historical behavior information according to the behavior representation information respectively corresponding to the multiple selection behaviors;

[0192] Adjusting the model parameters corresponding to the initial content recommendation model and the initial behavior information respectively corresponding to the multiple contents includes:

[0193] Adjust the model parameters corresponding to the initial content recommendation model, the initial behavior information respectively corresponding to the multiple contents, and the execution sequence information respectively corresponding to the multiple execution sequences. The multiple execution sequences are the execution sequences respectively corresponding to the multiple selection behaviors.

[0194] In a possible implementation, the initial content recommendation model includes N information processing layers. The N information processing layers are used to perform N rounds of information processing. Determining the historical behavior information according to the behavior representation information respectively corresponding to the multiple selection behaviors includes:

[0195] In the i-th round of information processing, obtain the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing. The sub-behavior information respectively corresponding to the multiple selection behaviors in the first round of information processing is the behavior representation information;

[0196] Determine the similarities respectively corresponding to the target selection behavior and multiple reference behaviors in the dimension of sub-behavior information according to the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing. The multiple reference behaviors are the first M selection behaviors executed among the multiple selection behaviors, and the target selection behavior is the M-th executed selection behavior;

[0197] Determine the target sub - behavior information corresponding to the target selection behavior in the (i + 1)-th round of information processing based on the similarities respectively corresponding to the target selection behavior and the multiple reference behaviors to be referred to in the dimension of sub - behavior information, and the sub - behavior information respectively corresponding to the multiple reference behaviors in the i - th round of information processing. The similarity between the target selection behavior and the target reference behavior in the dimension of sub - behavior information is used to characterize the reference strength of the sub - behavior information corresponding to the target reference behavior in the i - th round of information processing when determining the target sub - behavior information. The target reference behavior is any one of the multiple reference behaviors;

[0198] Determine the sub - behavior information corresponding to the end selection behavior output in the N - th round of information processing as the historical behavior information. The end selection behavior is the selection behavior executed latest among the multiple historical behaviors.

[0199] In a possible implementation manner, the device further includes a second acquisition unit, a second determination unit, and a third determination unit:

[0200] The second acquisition unit is configured to acquire a historical behavior record to be analyzed, where the historical behavior record to be analyzed is used to record multiple historical selection behaviors executed by the object to be recommended in a historical period. The multiple historical selection behaviors are used to select multiple historical contents, and the multiple historical contents are included in the set of recommendable contents;

[0201] The second determination unit is configured to determine, through the content recommendation model, target historical behavior information based on the behavior information respectively corresponding to the multiple historical contents. The target historical behavior information is used to characterize the multiple historical selection behaviors;

[0202] The third determination unit is configured to determine the content to be analyzed as the content with the highest probability of being selected by the object to be recommended based on the fact that the matching degree between the target historical behavior information and the behavior information corresponding to the content to be analyzed is the highest in the set of recommendable contents.

[0203] In a possible implementation manner, the device further includes a recommendation unit:

[0204] The recommendation unit is configured to recommend the content to be analyzed to the object to be recommended based on determining that the content to be analyzed is the content with the highest probability of being selected by the object to be recommended.

[0205] In a possible implementation manner, the device further includes an update unit:

[0206] The update unit is configured to update the model parameters corresponding to the content recommendation model and the behavior information corresponding to the content in the set of recommendable content, so that the updated content recommendation model has the highest degree of matching between the historical behavior information determined based on the target historical behavior record and the behavior information corresponding to the content to be analyzed after update, and the sample content corresponding to the target historical behavior record is the content to be analyzed.

[0207] In a possible implementation manner, the content to be analyzed is any one of multiple cross-domain contents, the multiple cross-domain contents correspond to a first content field, the multiple contents correspond to a second content field, and the apparatus further includes a third acquisition unit, a fourth determination unit, and an adjustment unit:

[0208] The third acquisition unit is configured to acquire a cross-domain historical behavior record, where the cross-domain historical behavior record is used to record multiple cross-domain selection behaviors performed by a second target object in a historical period, the multiple cross-domain selection behaviors are used to select content from the multiple cross-domain contents, and the sample content corresponding to the cross-domain historical behavior record is the target cross-domain content in the multiple cross-domain contents;

[0209] The fourth determination unit is configured to determine cross-domain historical behavior information through the initial content recommendation model according to the behavior information corresponding to the content selected by the multiple cross-domain selection behaviors, where the cross-domain historical behavior information is used to characterize the multiple cross-domain selection behaviors;

[0210] The adjustment unit is configured to adjust the model parameters corresponding to the initial content recommendation model and the behavior information corresponding to the multiple cross-domain contents respectively, so that the degree of matching between the cross-domain historical behavior information determined by the initial content recommendation model and the behavior information corresponding to the target cross-domain content is the largest, to obtain a cross-domain content recommendation model and the updated behavior information corresponding to the multiple cross-domain contents respectively, and the cross-domain content recommendation model is used to determine the cross-domain content with the highest selection probability from the multiple cross-domain contents according to the updated behavior information corresponding to the multiple cross-domain contents respectively.

[0211] An embodiment of this application further provides a computer device. Please refer to Figure 7 As shown, this computer device may be a terminal device. Taking the terminal device as a mobile phone as an example:

[0212] Figure 7 Shown is a block diagram of a part of the structure of a mobile phone related to the terminal device provided in the embodiment of this application. Refer to Figure 7, The mobile phone includes components such as a Radio Frequency (RF) circuit 710, a memory 720, an input unit 730, a display unit 740, sensors 750, an audio circuit 760, a Wireless Fidelity (WiFi) module 770, a processor 780, and a power supply 790. Those skilled in the art can understand that Figure 7 the mobile phone structure shown in

[0213] does not limit the mobile phone and may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements. Figure 7 The following specifically introduces each component of the mobile phone:

[0214] The RF circuit 710 can be used for receiving and transmitting signals during information reception or call processes. Specifically, after receiving the downlink information from the base station, it is given to the processor 780 for processing; in addition, it sends the designed uplink data to the base station. Generally, the RF circuit 710 includes but is not limited to antennas, at least one amplifier, a transceiver, a coupler, a Low Noise Amplifier (LNA), a duplexer, etc. In addition, the RF circuit 710 can also communicate with the network and other devices through wireless communication. The above wireless communication can use any communication standard or protocol, including but not limited to the Global System of Mobile communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, Short Messaging Service (SMS), etc.

[0215] The memory 720 can be used to store software programs and modules. The processor 780 executes various functional applications and data processing of the mobile phone by running the software programs and modules stored in the memory 720. The memory 720 may mainly include a program storage area and a data storage area. Among them, the program storage area can store the operating system, application programs required for at least one function (such as the sound playback function, the image playback function, etc.); the data storage area can store the data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory 720 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0216] The input unit 730 can be used to receive input digital or character information, and generate key signal inputs related to the user settings and function controls of the mobile phone. Specifically, the input unit 730 may include a touch panel 731 and other input devices 732. The touch panel 731, also known as a touch screen, can collect the touch operations of the user on or near it (such as the operations of the user using a finger, a stylus, or any suitable object or accessory on or near the touch panel 731), and drive the corresponding connection device according to a pre-set program. Optionally, the touch panel 731 may include two parts: a touch detection device and a touch controller. Among them, the touch detection device detects the touch position of the user, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into contact coordinates, and then sends it to the processor 780, and can receive and execute the commands sent by the processor 780. In addition, various types such as resistive, capacitive, infrared, and surface acoustic wave can be used to implement the touch panel 731. In addition to the touch panel 731, the input unit 730 may also include other input devices 732. Specifically, the other input devices 732 may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control keys, power on / off keys, etc.), a trackball, a mouse, a joystick, etc.

[0217] The display unit 740 can be used to display information input by the user or information provided to the user, as well as various menus of the mobile phone. The display unit 740 may include a display panel 741. Optionally, the display panel 741 can be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc. Further, the touch panel 731 can cover the display panel 741. When the touch panel 731 detects a touch operation on or near it, it is transmitted to the processor 780 to determine the type of touch event. Subsequently, the processor 780 provides a corresponding visual output on the display panel 741 according to the type of touch event. Although in Figure 7 the touch panel 731 and the display panel 741 are implemented as two independent components to realize the input and input functions of the mobile phone, in some embodiments, the touch panel 731 and the display panel 741 can be integrated to realize the input and output functions of the mobile phone.

[0218] The mobile phone may further include at least one sensor 750, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor. Among them, the ambient light sensor can adjust the brightness of the display panel 741 according to the brightness of the ambient light, and the proximity sensor can turn off the display panel 741 and / or the backlight when the mobile phone is moved to the ear. As a kind of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in all directions (generally three axes). When stationary, it can detect the magnitude and direction of gravity, and can be used for applications that identify the posture of the mobile phone (such as horizontal and vertical screen switching, related games, magnetometer attitude calibration), vibration recognition related functions (such as pedometer, tapping), etc.; as for other sensors that the mobile phone can also be configured with, such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors, etc., they will not be elaborated here.

[0219] The audio circuit 760, the speaker 761, and the microphone 762 can provide an audio interface between the user and the mobile phone. The audio circuit 760 can transmit the electrical signal converted from the received audio data to the speaker 761, and the speaker 761 converts it into a sound signal for output; on the other hand, the microphone 762 converts the collected sound signal into an electrical signal, which is received by the audio circuit 760 and then converted into audio data. After the audio data is output to the processor 780 for processing, it is sent to another mobile phone through the RF circuit 710, for example, or the audio data is output to the memory 720 for further processing.

[0220] WiFi belongs to short-range wireless transmission technology. The mobile phone can help users send and receive emails, browse the web, and access streaming media through the WiFi module 770. It provides users with wireless broadband Internet access. Although Figure 7The WiFi module 770 is shown, but it can be understood that it does not belong to the essential components of the mobile phone and can be completely omitted within the scope of not changing the essence of the invention as needed.

[0221] The processor 780 is the control center of the mobile phone, connecting various parts of the entire mobile phone through various interfaces and circuits. By running or executing software programs and / or modules stored in the memory 720, and by calling data stored in the memory 720, it executes various functions of the mobile phone and processes data, thereby performing an overall detection of the mobile phone. Optionally, the processor 780 may include one or more processing units; preferably, the processor 780 may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modem processor mainly processes wireless communications. It can be understood that the above-mentioned modem processor may not be integrated into the processor 780 either.

[0222] The mobile phone further includes a power supply 790 (such as a battery) for supplying power to each component. Preferably, the power supply can be logically connected to the processor 780 through a power management system, so as to implement functions such as management of charging, discharging, and power consumption management through the power management system.

[0223] Although not shown, the mobile phone may further include a camera, a Bluetooth module, etc., which will not be elaborated here.

[0224] In this embodiment, the processor 780 included in the terminal device further has the following functions:

[0225] Obtain a set of recommended content. Each of the multiple pieces of content included in the set of recommended content has corresponding behavior information, and the behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model is used to determine the content with the highest selection probability from the set of recommended content based on the behavior information;

[0226] Through a similarity analysis model, according to the modality information corresponding to the content to be analyzed and the multiple pieces of content respectively, determine the similarity between the content to be analyzed and the multiple pieces of content. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any one of the multiple pieces of content, and the modality information is used to characterize the corresponding content;

[0227] Generate the behavior information corresponding to the content to be analyzed according to the similarity between the content to be analyzed and the multiple pieces of content, and the behavior information corresponding to the multiple pieces of content respectively, and add the content to be analyzed to the set of content that can be recommended. The similarity between the content to be analyzed and the target content is used to characterize the degree of reference to the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

[0228] The embodiment of the present application further provides a server. Please refer to Figure 8 as shown in Figure 8 is a structural diagram of the server 800 provided by the embodiment of the present application. The server 800 may vary greatly due to configuration or performance, and may include one or more central processing units (CPUs) 822 (for example, one or more processors) and a memory 832, and one or more storage media 830 for storing application programs 842 or data 844 (for example, one or more mass storage devices). Among them, the memory 832 and the storage media 830 may be transient storage or persistent storage. The program stored in the storage media 830 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Further, the central processing unit 822 may be configured to communicate with the storage media 830 and execute a series of instruction operations in the storage media 830 on the server 800.

[0229] The server 800 may further include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input / output interfaces 858, and / or one or more operating systems 841, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM and so on.

[0230] The steps performed by the server in the above embodiments may be based on Figure 8 the server structure shown in

[0231] The embodiment of the present application further provides a computer-readable storage medium for storing a computer program, and the computer program is used to execute any one of the content recommendation methods described in the foregoing embodiments.

[0232] The embodiments of the present application also provide a computer program product including a computer program. When it runs on a computer device, it causes the computer device to execute the content recommendation method described in any one of the above embodiments.

[0233] It can be understood that in the specific implementation of the present application, it involves relevant data such as user information (such as historical behavior information, purchase information, video viewing information). When the above embodiments of the present application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0234] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium can be at least one of the following media: read-only memory (abbreviation: ROM), RAM, magnetic disk, or optical disc, etc., which can store program codes.

[0235] It should be noted that the embodiments in this specification are all described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0236] As described above, it is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A content recommendation method, characterized in that, The method includes: Obtaining a set of recommendable content, where multiple pieces of content included in the set of recommendable content respectively have corresponding behavior information, and the behavior information is used to characterize the selection behavior of selecting the corresponding content. A content recommendation model is used to determine, based on the behavior information, the content with the highest selection probability from the set of recommendable content; Through a similarity analysis model, according to the modality information respectively corresponding to the content to be analyzed and the multiple pieces of content, determining the similarity between the content to be analyzed and the multiple pieces of content. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any piece of content among the multiple pieces of content, and the modality information is used to characterize the corresponding content; According to the similarity between the content to be analyzed and the multiple pieces of content, and the behavior information respectively corresponding to the multiple pieces of content, generating the behavior information corresponding to the content to be analyzed, and adding the content to be analyzed to the set of recommendable content. The similarity between the content to be analyzed and the target content is used to characterize the reference strength of the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

2. The method according to claim 1, characterized in that, Before the step of determining, through the similarity analysis model, the similarity between the content to be analyzed and the multiple pieces of content according to the modality information respectively corresponding to the content to be analyzed and the multiple pieces of content, the method further includes: Training the similarity analysis model according to the similarity between the behavior information respectively corresponding to the multiple pieces of content and the modality information respectively corresponding to the multiple pieces of content, so that the first similarity determined by the similarity analysis model according to the modality information respectively corresponding to the multiple pieces of content is greater than the second similarity. The first similarity is the similarity between the target content and the first content set, and the second similarity is the similarity between the target content and the second content set. The similarity between the target content and the first content set in the dimension of behavior information is greater than the similarity between the target content and the second content set in the dimension of behavior information. The first content set and the second content set are content sets divided from the multiple pieces of content.

3. The method according to claim 2, characterized in that The training the similarity analysis model according to the similarity between the behavior information respectively corresponding to the multiple pieces of content and the modality information respectively corresponding to the multiple pieces of content includes: Taking the multiple pieces of content respectively as the target content, determining multiple first pieces of content with the highest similarity to the target content in the dimension of behavior information among the multiple pieces of content to form the first content set, and determining multiple second pieces of content among the multiple pieces of content to form the second content set. The multiple second pieces of content do not include the first pieces of content; Determining the first similarity and the second similarity through an initial similarity analysis model according to the modality information respectively corresponding to the multiple pieces of content; Constructing a loss function corresponding to the initial similarity analysis model according to the first similarity and the second similarity. The loss function is positively correlated with the first similarity and negatively correlated with the second similarity; Adjust the model parameters of the initial similarity analysis model according to the loss function, so that the loss function is greater than the loss threshold, and obtain the similarity analysis model, where the loss threshold satisfies that the first similarity is greater than the second similarity.

4. The method according to claim 1, wherein The content recommendation model is trained in the following way: Obtain historical behavior records and initial behavior information corresponding to the multiple contents respectively. The historical behavior records are used to record multiple selection behaviors performed by a first target object in a historical period. The multiple selection behaviors are used to select multiple selected contents. The historical selection behavior records have corresponding sample contents. The sample content is the content selected by the first target object in the next selection behavior after completing the multiple selection behaviors. The multiple contents include the multiple selected contents and the sample content; Through the initial content recommendation model, determine historical behavior information according to the initial behavior information corresponding to the multiple selected contents respectively. The historical behavior information is used to characterize the multiple selection behaviors; Adjust the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to the multiple contents respectively, so that the matching degree between the historical behavior information determined by the initial content recommendation model and the initial behavior information corresponding to the sample content is maximized, and obtain the content recommendation model and the behavior information corresponding to the multiple contents respectively.

5. The method according to claim 4, characterized in that The historical behavior records are used to record multiple selection behaviors performed by the first target object in a historical period and the execution order of the multiple selection behaviors. Determining historical behavior information according to the initial behavior information corresponding to the multiple selected contents respectively includes: Determine the order information corresponding to the multiple selection behaviors respectively according to the execution order corresponding to the multiple selection behaviors. The order information is used to characterize the execution order of the corresponding selection behavior; Generate behavior characterization information corresponding to a target selection behavior according to the initial behavior information of the selected content corresponding to the target selection behavior and the order information corresponding to the target selection behavior. The behavior characterization information corresponding to the target selection behavior is used to characterize the target selection behavior executed in the target execution order. The target selection behavior is any one of the multiple selection behaviors, and the target execution order is the execution order of the target selection behavior in the multiple selection behaviors; Determine the historical behavior information according to the behavior characterization information corresponding to the multiple selection behaviors respectively; Adjusting the model parameters corresponding to the initial content recommendation model and the initial behavior information corresponding to the multiple contents respectively includes: Adjust the model parameters corresponding to the initial content recommendation model, the initial behavior information corresponding to the multiple contents respectively, and the execution order information corresponding to the multiple execution orders. The multiple execution orders are the execution orders corresponding to the multiple selection behaviors respectively.

6. The method according to claim 5, wherein The initial content recommendation model includes N information processing layers. The N information processing layers are used to perform N rounds of information processing. Determining the historical behavior information according to the behavior characterization information corresponding to the multiple selection behaviors respectively includes: In the i-th round of information processing, obtain the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing. The sub-behavior information respectively corresponding to the multiple selection behaviors in the first round of information processing is the behavior characterization information; According to the sub-behavior information respectively corresponding to the multiple selection behaviors in the i-th round of information processing, determine the similarities respectively corresponding to the target selection behavior and multiple reference behaviors in the dimension of sub-behavior information. The multiple reference behaviors are the first M executed selection behaviors among the multiple selection behaviors, and the target selection behavior is the M-th executed selection behavior; According to the similarities respectively corresponding to the target selection behavior and the multiple reference behaviors in the dimension of sub-behavior information, and the sub-behavior information respectively corresponding to the multiple reference behaviors in the i-th round of information processing, determine the target sub-behavior information corresponding to the target selection behavior in the (i + 1)-th round of information processing. The similarity between the target selection behavior and the target reference behavior in the dimension of sub-behavior information is used to represent the reference strength of the sub-behavior information corresponding to the target reference behavior in the i-th round of information processing when determining the target sub-behavior information. The target reference behavior is any one of the multiple reference behaviors; Determine the sub-behavior information corresponding to the last selection behavior output in the N-th round of information processing as the historical behavior information. The last selection behavior is the latest executed selection behavior among the multiple historical behaviors.

7. The method according to claim 4, characterized in that The method further includes: Obtain a historical behavior record to be analyzed, where the historical behavior record to be analyzed is used to record multiple historical selection behaviors executed by the object to be recommended in a historical period. The multiple historical selection behaviors are used to select multiple historical contents, and the multiple historical contents are included in the set of recommendable contents; Through the content recommendation model, determine the target historical behavior information according to the behavior information respectively corresponding to the multiple historical contents. The target historical behavior information is used to represent the multiple historical selection behaviors; Based on the highest matching degree between the target historical behavior information and the behavior information corresponding to the content to be analyzed in the set of recommendable contents, determine the content to be analyzed as the content with the highest probability of being selected by the object to be recommended.

8. The method according to claim 7, wherein The method further includes: Based on determining that the content to be analyzed is the content with the highest probability of being selected by the object to be recommended, recommend the content to be analyzed to the object to be recommended.

9. The method according to claim 4, characterized in that, The method further includes: Update the model parameters corresponding to the content recommendation model and the behavior information corresponding to the contents in the set of recommendable contents, so that the historical behavior information determined by the updated content recommendation model based on the target historical behavior record has the highest matching degree with the behavior information corresponding to the updated content to be analyzed. The sample content corresponding to the target historical behavior record is the content to be analyzed.

10. The method according to claim 4, characterized in that, The content to be analyzed is any one of multiple cross-domain contents. The multiple cross-domain contents correspond to a first content field, and the multiple contents correspond to a second content field. The method further includes: Obtain cross - domain historical behavior records, where the cross - domain historical behavior records are used to record multiple cross - domain selection behaviors performed by a second target object during a historical period. The multiple cross - domain selection behaviors are used to select content from the multiple cross - domain contents, and the sample content corresponding to the cross - domain historical behavior records is the target cross - domain content among the multiple cross - domain contents; Through the initial content recommendation model, determine cross - domain historical behavior information according to the behavior information corresponding to the content selected by the multiple cross - domain selection behaviors. The cross - domain historical behavior information is used to characterize the multiple cross - domain selection behaviors; Adjust the model parameters corresponding to the initial content recommendation model and the behavior information corresponding to the multiple cross - domain contents respectively, so that the matching degree between the cross - domain historical behavior information determined by the initial content recommendation model and the behavior information corresponding to the target cross - domain content is maximized, to obtain a cross - domain content recommendation model and the updated behavior information corresponding to the multiple cross - domain contents respectively. The cross - domain content recommendation model is used to determine the cross - domain content with the highest selection probability from the multiple cross - domain contents according to the updated behavior information corresponding to the multiple cross - domain contents respectively.

11. A content recommendation device, characterized in that, The device includes a first acquisition unit, a first determination unit, and a generation unit: The first acquisition unit is used to acquire a set of recommendable content. Each of the multiple contents included in the set of recommendable content has corresponding behavior information. The behavior information is used to characterize the selection behavior of selecting the corresponding content. The content recommendation model is used to determine the content with the highest selection probability from the set of recommendable content based on the behavior information; The first determination unit is used to determine the similarity between the content to be analyzed and the multiple contents through a similarity analysis model according to the modality information corresponding to the content to be analyzed and the multiple contents respectively. The similarity between the content to be analyzed and the target content is positively correlated with the similarity between the content to be analyzed and the target content in the dimension of behavior information. The target content is any one of the multiple contents, and the modality information is used to characterize the corresponding content; The generation unit is used to generate the behavior information corresponding to the content to be analyzed according to the similarity between the content to be analyzed and the multiple contents, and the behavior information corresponding to the multiple contents respectively, and add the content to be analyzed to the set of recommendable content. The similarity between the content to be analyzed and the target content is used to characterize the reference strength of the behavior information corresponding to the target content when generating the behavior information corresponding to the content to be analyzed.

12. A computer device, characterized in that, The computer device includes a processor and a memory: The memory is used to store a computer program and transmit the computer program to the processor; The processor is used to execute the content recommendation method according to any one of claims 1 - 10 based on the instructions in the computer program.

13. A computer-readable storage medium, characterized in that, The computer - readable storage medium is used to store a computer program, and the computer program is used to execute the content recommendation method according to any one of claims 1 - 10.

14. A computer program product including a computer program, which, when running on a computer device, causes the computer device to execute the content recommendation method according to any one of claims 1-10.

Images