Content recommendation method and related device

Abstract

The invention discloses a content recommendation method and a related device. The content recommendation method and the related device can be applied to various scenes such as cloud technology, artificial intelligence and vehicle-mounted scenes. And performing recommendation prediction on the obtained multiple candidate contents through the target recommendation model to obtain a recommendation prediction value of each candidate content. Respectively performing interest exploration on the plurality of candidate contents through an interest exploration model to obtain a first prediction probability value of each candidate content; and fusing the first prediction probability value of each candidate content with the recommended prediction value to obtain a fused prediction value. And determining the candidate content of which the fusion predicted value meets a preset recommendation condition in the plurality of candidate contents as a recommendation result. According to the content recommendation method, the fusion prediction value obtained by fusing the first prediction probability values can accurately reflect the probability that the candidate content is the new interest content of the interaction object, and the method provided by the invention can improve the accuracy of interest exploration, so that the possibility of success of interest exploration is improved; and the improvement range of the interest exploration effect is expanded.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular to a content recommendation method and related devices. Background Art

[0002] In recent years, with the explosive development of mobile Internet and big data technologies, recommendation systems have been widely used in an increasing number of Internet products. Recommendation systems recommend content of interest to users, for example, news recommendations (for news content), movie and book recommendations (for movies and books), video recommendations (for videos), and product recommendations (for products).

[0003] After a recommendation system has been running for a while, the content recommended to the interacting party often falls into an information cocoon, meaning that the content recommended to the interacting party is limited to a few points of interest. To this end, interest exploration is necessary. Related technologies mainly use the similarities between content categories to conduct interest exploration. For example, in the video recommendation scenario, if the interacting party's main interest is game videos, other content with high similarity to games, such as high-quality animation and 2D videos, can be used as candidate content for interest exploration. After careful sorting and scoring, the high-scoring videos are selected and recommended to the interacting party.

[0004] Exploring interests based on similarities between content categories is based on the characteristics of certain interacting objects: if they are interested in a certain type of content, they are also interested in content similar to that type, and thus, the current interacting object is inferred to have the same characteristics. However, inferring the current interacting object's interests based on other interacting objects is difficult to reflect their true interests. This method is relatively crude and, in practice, can only be forced into the final recommendation results to test interest. This has a high probability of exploration failure, and therefore has limited improvement in interest exploration effectiveness. Summary of the Invention

[0005] In order to solve the above technical problems, the present application provides a content recommendation method and related devices, which can increase the possibility of successful interest exploration and expand the improvement range of interest exploration effects.

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

[0007] In one aspect, an embodiment of the present application provides a content recommendation method, the method comprising:

[0008] Obtain multiple candidate contents;

[0009] Performing recommendation prediction on the plurality of candidate contents by using a target recommendation model to obtain a recommendation prediction value for each of the plurality of candidate contents;

[0010] Performing interest exploration on the plurality of candidate contents respectively through an interest exploration model, obtaining a first predicted probability value for each candidate content, the first predicted probability value being used to reflect the probability that the corresponding candidate content is new interest content of the interactive object, the new interest content being content whose satisfaction reaches a satisfaction threshold and to which the classification belongs does not belong to the historical interest classification set of the interactive object, the interest exploration model being trained based on a plurality of sample contents of the interactive object and sample labels corresponding to the plurality of sample contents, the sample labels being used to indicate the interaction between the interactive object and the corresponding sample content;

[0011] For each candidate content, fusing the first prediction probability value and the recommendation prediction value to obtain a fused prediction value, where the fused prediction value is positively correlated with the first prediction probability value;

[0012] The candidate content whose fusion prediction value meets the preset recommendation condition among the multiple candidate contents is determined as the recommendation result.

[0013] On the one hand, an embodiment of the present application provides a content recommendation device, which includes an acquisition unit, a recommendation prediction unit, an interest exploration unit, a fusion unit, and a determination unit:

[0014] The acquisition unit is configured to acquire a plurality of candidate contents;

[0015] The recommendation prediction unit is configured to perform recommendation prediction on the plurality of candidate contents using a target recommendation model to obtain a recommendation prediction value for each of the plurality of candidate contents;

[0016] The interest exploration unit is configured to perform interest exploration on each of the plurality of candidate contents through an interest exploration model to obtain a first predicted probability value for each of the candidate contents, wherein the first predicted probability value is used to reflect the probability that the corresponding candidate content is a new interest content of the interactive object, wherein the new interest content is content whose satisfaction reaches a satisfaction threshold and whose classification does not belong to the historical interest classification set of the interactive object, and the interest exploration model is trained based on a plurality of sample contents of the interactive object and sample labels corresponding to the plurality of sample contents, respectively, and wherein the sample labels are used to indicate the interaction between the interactive object and the corresponding sample content;

[0017] The fusion unit is configured to fuse the first prediction probability value and the recommendation prediction value for each candidate content to obtain a fused prediction value, where the fused prediction value is positively correlated with the first prediction probability value;

[0018] The determining unit is configured to determine, among the plurality of candidate contents, a candidate content whose fusion prediction value satisfies a preset recommendation condition as a recommendation result.

[0019] In one aspect, an embodiment of the present application provides a computer device, comprising a processor and a memory:

[0020] The memory is used to store a computer program and transmit the computer program to the processor;

[0021] The processor is configured to execute the method described in any one of the preceding aspects according to instructions in the computer program.

[0022] In one aspect, an embodiment of the present application provides a computer-readable storage medium for storing a computer program. When the computer program is executed by a processor, the processor executes the method described in any one of the aforementioned aspects.

[0023] In one aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any of the aforementioned aspects when executed by a processor.

[0024] As can be seen from the above technical solution, when performing content recommendation, the present application can add an interest exploration model to the existing target recommendation model to assist in content recommendation. Specifically, multiple candidate content can be obtained, and recommendation predictions can be performed on the multiple candidate content using the target recommendation model to obtain a recommendation prediction value for each of the multiple candidate content. The interest exploration model can perform interest exploration on each of the multiple candidate content, obtaining a first prediction probability value for each candidate content. The interest exploration model is trained based on multiple sample content of an interactive object and the sample labels corresponding to the multiple sample content. The sample labels are used to indicate the interaction between the interactive object and the corresponding sample content. Therefore, the interest exploration model accurately learns the interests of the interactive object. The first prediction probability value obtained based on the interest exploration model can accurately reflect the probability that the corresponding candidate content is a new interest content for the interactive object. The higher the first prediction probability value of the candidate content, the more likely the candidate content is a new interest of the interactive object. Therefore, for each candidate content, the first prediction probability value and the recommendation prediction value are fused to obtain a fused prediction value. The fused prediction value is positively correlated with the first prediction probability value, thereby adjusting the recommendation prediction value to obtain a fused prediction value, thereby increasing the fused prediction value of the new interest content of the interactive object. In this way, when candidate content whose fused prediction value meets the preset recommendation conditions among multiple candidate content is determined as a recommendation result, candidate content belonging to new interest content can be ranked at the top, helping to explore new interests. Because the first predicted probability value can accurately reflect the probability that the candidate content is the new interest content of the interactive object, the fused prediction value adjusted based on the first predicted probability value can also accurately reflect the possibility that the candidate content is the new interest content of the interactive object. In this way, the first predicted probability value can be used to rank the new interest content of the interactive object at the top, improving the accuracy of interest exploration, thereby increasing the possibility of successful interest exploration and expanding the improvement of interest exploration effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technical members in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0026] Figure 1 An application scenario architecture diagram of a content recommendation method provided in an embodiment of the present application;

[0027] Figure 2 A flowchart of a content recommendation method provided in an embodiment of the present application;

[0028] Figure 3A schematic diagram of content recommendation provided in an embodiment of the present application;

[0029] Figure 4 Another content recommendation diagram provided in an embodiment of the present application;

[0030] Figure 5 A flowchart of a method for training an interest exploration model provided in an embodiment of the present application;

[0031] Figure 6 A data graph of physical duration and average stay duration provided in an embodiment of the present application;

[0032] Figure 7 A schematic diagram of a satisfaction curve provided in an embodiment of the present application;

[0033] Figure 8 A data comparison chart of the average number of exposure categories per person for content recommendations for interest exploration and content recommendations for non-interest exploration provided in an embodiment of the present application;

[0034] Figure 9 A data comparison chart showing the average number of clicks per category for content recommendations for interest exploration and content recommendations for non-interest exploration, provided in an embodiment of the present application;

[0035] Figure 10 A data comparison chart of click-through rates for content recommendations for interest exploration and content recommendations for non-interest exploration provided in an embodiment of the present application;

[0036] Figure 11 A structural diagram of a content recommendation device provided in an embodiment of the present application;

[0037] Figure 12 A structural diagram of a terminal provided in an embodiment of the present application;

[0038] Figure 13 A structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The embodiments of the present application are described below with reference to the accompanying drawings.

[0040] Recommendation systems can recommend content of interest to interacting parties. However, after a period of operation, the content recommended to the same interacting party tends to become trapped in an information cocoon, with the content recommended by the recommendation system being limited to specific points of interest.

[0041] The recommendation system's limited recommendations for interactive users to specific interests can cause several problems. First, interactive users will only receive content related to specific interests, resulting in a narrow range of content. After a period of time, these users may experience interest fatigue, which can lead to a high probability of user withdrawal from the recommendation system, leading to retention issues. This retention issue can refer to the increased probability of user churn caused by interest fatigue. Second, after a recommendation system has been running for a period of time, interactive users will only receive content related to specific interests, reducing the types of content available (such as advertisements and videos), and thus reducing their commercial value. Third, the narrow range of content recommended by the recommendation system for interactive users means that the recommendation system collects less information about the content of interest to the interactive user. The less information about the interactive user's interests, the more the recommendation system's recommendations will be limited to specific interests, which can lead to a vicious cycle in the recommendation system's ecosystem. To address this issue, interest exploration is necessary. Related technologies primarily utilize similarities between content categories for interest exploration. Interest exploration refers to the process of discovering content that may be of interest to the interactive user through exploration.

[0042] Related technologies use similarities between content categories to explore interests. This is primarily achieved by analyzing the interaction data of other interacting objects. Specifically, if the interaction data of other interacting objects reveals that they are already interested in a certain type of content, then they are likely interested in similar content. Therefore, it is inferred that the current interacting object also has similar characteristics. Therefore, when recommending content to the current interacting object, the similarities between content categories are used to explore the current interacting object's interests and achieve content recommendations for the current interacting object.

[0043] However, the characteristics of other interacting subjects don't fully represent the characteristics of the current interacting subject. Therefore, the interests of the interacting subject, derived from interest exploration based on similarity between content categories, are unlikely to accurately reflect the interacting subject's true interests. In actual interest exploration, interest exploration is often limited to forcing content similar to the interacting subject's existing interests into the final recommendation results. This leads to a high probability of exploration failure, thus limiting the effectiveness of interest exploration. This failure forces the interacting subject to continue selecting content that already interests them, making it difficult to address the problem of the recommendation system falling into an information cocoon. For example, the interacting subject's primary interest is sports videos, such as football and basketball games. Fitness videos and game commentary videos are highly similar to sports videos, so the system attempts to recommend videos of these highly similar categories to the interacting subject. However, the interacting subject only enjoys watching sports videos, such as football and basketball games, and does not enjoy fitness. The interacting subject prefers traveling over fitness.

[0044] To address the aforementioned technical issues and enhance the effectiveness of interest exploration, embodiments of the present application provide a content recommendation method that incorporates interest exploration. Specifically, during content recommendation for an interactive object, interest exploration is used to obtain new content of interest to the interactive object, thereby prioritizing the new content in the recommended results. During interest exploration, a first predicted probability value is obtained using an interest exploration model. The interest exploration model is trained based on multiple sample contents of the interactive object and sample labels corresponding to the multiple sample contents. The sample labels indicate the interaction between the interactive object and the corresponding sample contents. Therefore, the interest exploration model accurately learns the interests of the interactive object. The first predicted probability value obtained based on the interest exploration model accurately represents the probability that the candidate content is the new content of interest to the interactive object. The higher the first predicted probability value, the more likely the corresponding candidate content is the new content of interest to the interactive object. Therefore, the recommendation prediction value is adjusted based on the first predicted probability value to obtain a fused prediction value, thereby increasing the fused prediction value of the new content of interest to the interactive object. Thus, when recommending content, the new content of interest to the interactive object is more likely to be recommended to the interactive object. This reduces the likelihood of interest exploration failure and enhances the effectiveness of interest exploration.

[0045] It should be noted that the content recommendation method provided in the embodiment of the present application can be applied to various content recommendation scenarios. The content recommendation scenarios can be, for example, video recommendation, picture and text recommendation, music recommendation, material recommendation, web page recommendation and other scenarios, and the embodiment of the present application does not limit this.

[0046] The content recommendation method provided in the embodiments of the present application can be executed by a computer device, which can be, for example, a server or a terminal. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminals include but are not limited to smartphones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc.

[0047] like Figure 1 As shown, Figure 1 An application scenario architecture diagram of a content recommendation method is shown, which may include a server 100 and a terminal 200. The server 100 can provide content recommendation services for the terminal 200, wherein the server 200 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.

[0048] Terminal 200 may install an application with a content recommendation function, which can recommend content through the recommendation system. In the embodiment of the present application, the application can be various types of applications, such as video applications, music applications, news applications, etc. Video applications may include short video applications and long video applications, and music applications may include music players, radio applications, etc.

[0049] After the interactive object opens an application with a content recommendation function on the terminal 200, the application with a content recommendation function can recommend content to the interactive object through the content recommendation function. The server 100 can provide content recommendation services, so it triggers the terminal 200 to send a content recommendation request to the server 100. The server 100 executes the content recommendation method based on the content recommendation request, obtains the recommendation result, and returns the recommendation result to the terminal 200, so that the recommendation result is displayed to the interactive object on the terminal 200.

[0050] The terminal 100 includes, but is not limited to, a smartphone, a computer, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, etc. The server 100 and the terminal 200 may be connected directly or indirectly via wired or wireless communication, which is not limited in this application. For example, the server 100 and the terminal 200 may be connected via a network, which may be a wired or wireless network.

[0051] The specific process of the server 100 executing the content recommendation method provided in the embodiment of the present application may be: the server 100 may first obtain multiple candidate contents, and the candidate contents may refer to the contents waiting to be selected. In the embodiment of the present application, it is necessary to select the contents recommended to the interactive object from the candidate contents. In the embodiment of the present application, the process of content recommendation includes processes such as recall, rough sorting and fine sorting. Recall corresponds to a large number of contents. Rough sorting will screen the recalled contents and roughly sort the screened contents. Fine sorting will further screen the results of the rough sorting and give the final sorting displayed on the terminal 200. The method provided in the embodiment of the present application can be applied to the process of fine sorting. In this case, the candidate content may be the content selected by the rough sorting.

[0052] It should be noted that in different application scenarios, the types of candidate content may be different. For example, in the scenario of video recommendation, the candidate content may be video; in the scenario of music recommendation, the candidate content may be audio.

[0053] The server 100 inputs multiple candidate contents into the target recommendation model, performs recommendation prediction based on the target recommendation model, and obtains recommendation prediction values ​​corresponding to the multiple candidate contents. Each candidate content input into the target recommendation model has a corresponding recommendation prediction value.

[0054] Candidate content may be new content of interest to the interactive subject, or it may be content of historical interest to the interactive subject. Historical content of interest may refer to content whose corresponding classification belongs to a set of historical interest classifications. A set of historical interest classifications may be a set of categories that the interactive subject has already determined to be of interest to. Therefore, historical content of interest may also refer to content within a category that the interactive subject has already determined to be of interest to. In one possible implementation, by analyzing the interactive subject's historical interaction data, if it is determined that the interactive subject is interested in film and television content, and thus film and television is added to the set of historical interest classifications, then related film and television content, such as film review videos and news about new film releases, may be considered as content of historical interest to the interactive subject. New content of interest may refer to content whose corresponding classification does not belong to the set of historical interest classifications but that the interactive subject may be interested in. For example, although popular science content does not belong to the set of historical interest classifications, analysis indicates that the interactive subject may be interested in popular science content. Therefore, popular science content, such as popular science videos and articles, may be considered as new content of interest to the interactive subject. Therefore, the server 100 may perform interest exploration on multiple candidate content to identify new content of interest.

[0055] Specifically, the server 100 may input multiple candidate contents into the interest exploration model, conduct interest exploration based on the interest exploration model, and obtain first predicted probability values ​​corresponding to each of the multiple candidate contents. The interest exploration model is trained based on multiple sample contents of an interactive object and sample labels corresponding to each of the multiple sample contents. The sample labels are used to indicate the interaction between the interactive object and the corresponding sample content. Therefore, the interest exploration model more accurately learns the points of interest of the interactive object. The first predicted probability value obtained based on the interest exploration model is used to accurately reflect the probability that the corresponding candidate content is a new content of interest for the interactive object. The higher the first predicted probability value, the more likely the corresponding candidate content is a new content of interest for the interactive object.

[0056] Therefore, server 100 can fuse the recommendation prediction values ​​and the first prediction probability values ​​corresponding to the multiple candidate contents to obtain fused prediction values ​​corresponding to each of the multiple candidate contents. The fused prediction value is positively correlated with the first prediction probability value. Adjusting the recommendation prediction value based on the first prediction probability value can improve the fused prediction value of the newly interesting content. The candidate content whose fused prediction value meets the preset recommendation criteria is then determined as the recommendation result.

[0057] Next, the server 100 may send the recommendation result to the terminal 200 , and the terminal 200 may recommend content to the interaction object based on the recommendation result sent by the server 100 .

[0058] When terminal 200 recommends content based on the recommendation results obtained above, since the fused prediction value is obtained using the recommendation prediction value and the first prediction probability value, and the first prediction probability value reflects the probability that the candidate content is new content of interest to the interactive subject, the higher the first prediction probability value of the candidate content, the higher the fused prediction value corresponding to the candidate content, and the more likely the candidate content will be displayed at a front position on terminal 200. Placing new content of interest to the interactive subject at the front can increase the likelihood of successful interest exploration and improve the effectiveness of interest exploration.

[0059] exist Figure 1 In the example of content recommendation scenario, video recommendation is used. The recommended results displayed include videos of football games, videos of basketball games, videos of travel guides, and videos of football game reviews, etc. Among them, videos of travel guides are new interest content obtained through interest exploration, so they are ranked higher in the recommendation results.

[0060] It should be noted that in the specific implementation of this application, user information and other related data may be involved in the entire process. When the above embodiments of this application are applied to specific products or technologies, the user's separate consent or separate permission is required, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0061] It should be noted that the method provided in the embodiments of the present application may involve artificial intelligence technology, and the embodiments of the present application can automatically implement content recommendation based on 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 similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0062] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, pre-trained models, operating / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be fine-tuned and widely applied to downstream tasks across various AI disciplines. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0063] In the embodiment of the present application, the interest exploration model is the key to achieving content recommendation. When training the interest exploration model, machine learning may also be involved. Machine Learning (ML) is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formula-based learning. In the embodiment of the present application, machine learning can be used to train the interest exploration model.

[0064] Next, the content recommendation method provided by the embodiment of the present application will be introduced with reference to the accompanying drawings, taking the computer device being a server as an example. Figure 2 , Figure 2 A flow chart of a content recommendation method is shown, the method comprising the following steps:

[0065] S201: Acquire multiple candidate contents.

[0066] When the interactive object uses an application with a content recommendation function, the interactive object may trigger a content recommendation request through the terminal. The server can make content recommendations for the interactive object in response to the content recommendation request sent by the terminal. When the server makes content recommendations, it can first obtain multiple candidate content based on the content recommendation request. The candidate content may refer to content waiting to be selected. If the content recommendation method provided in the embodiment of the present application is applied to the process of fine sorting, the candidate content may be the content selected in the rough sorting. Among them, the interactive object may be an object that can interact with the candidate content, and the interactive object may be a user, for example.

[0067] It is understandable that the types of candidate content may be different in different application scenarios. For example, in the scenario of video recommendation, the candidate content may be video; in the scenario of music recommendation, the candidate content may be audio; in the scenario of picture recommendation, the candidate content may be picture.

[0068] The candidate content can be generated manually or by a machine, and the embodiments of the present application do not limit the method of generating the candidate content. In one possible implementation, in the scenario of video recommendation, the candidate content can be generated by the video creator based on video editing software. Video editing refers to the process of remixing the added pictures, background music, special effects, scenes and other materials with the video, cutting and merging the video source, and generating videos with different expressiveness through secondary encoding. In another possible implementation, in the scenario of picture recommendation, the candidate content can be drawn by a machine using artificial intelligence drawing software.

[0069] S202: Perform recommendation prediction on multiple candidate contents using a target recommendation model to obtain a recommendation prediction value for each candidate content in the multiple candidate contents.

[0070] After obtaining multiple candidate content, the server can determine the interactive subject's level of interest in each candidate content. In an embodiment of the present application, the server can use the target recommendation model to perform recommendation predictions on the multiple candidate content, obtaining a recommendation prediction value for each of the multiple candidate content. The recommendation prediction value can reflect the interactive subject's level of interest in the candidate content, which can be used to subsequently determine the recommendation result.

[0071] In the embodiment of the present application, the target refers to an indicator for recommendation prediction. The target can be an estimated click-through rate, a playback duration, a completion rate, etc., and the embodiment of the present application does not limit this. The target recommendation model can be a single-target recommendation model or a multi-target recommendation model. Depending on the actual usage scenario, the target selected by the target recommendation model will also be different. For example, in the scenario of recommending short videos, the target can be the playback duration of the video; in the scenario of music recommendation, the target can be the user's completion rate of the music. The user's completion rate of the music can refer to the ratio of the playback duration to the physical duration. For example, if the user plays the music for 3 seconds and the physical duration of the music is 30 seconds in total, then the user's completion rate of the music is 3 / 30=0.1.

[0072] The single-target recommendation model is a target recommendation model that makes recommendation predictions for a single target. For example, the single target may be an estimated click-through rate, that is, the single-target recommendation model can make recommendation predictions based on the estimated click-through rates of multiple candidate contents, and obtain a recommendation prediction value related to the estimated click-through rate for each candidate content in the multiple candidate contents. The estimated click-through rate may refer to the estimated probability that the interactive object will click on this content. It is an estimated probability value, and the estimated click-through rate reflects the degree of attention paid by the interactive object to a certain content. For example, the server obtains three candidate contents, namely candidate content A, candidate content B, and candidate content C. The estimated click-through rate of candidate content A is 60%, the estimated click-through rate of candidate content B is 56%, and the estimated click-through rate of candidate content C is 41%. The estimated click-through rate of candidate content A is higher than that of candidate content B, and the estimated click-through rate of candidate content B is higher than that of candidate content C. This shows that the interacting object is most interested in candidate content A, followed by candidate content B, and finally candidate content C. Therefore, the recommendation prediction value of candidate content A obtained by using the single-target recommendation model based on estimated click-through rate for recommendation prediction may be higher than the recommendation prediction value of candidate content B, and the recommendation prediction value of candidate content B may be higher than the recommendation prediction value of candidate content C.

[0073] For another example, a single target can be the playback duration, that is, the single-target recommendation model can make recommendation predictions based on the playback duration of multiple candidate contents, and obtain the recommendation prediction value related to the playback duration of each candidate content in the multiple candidate contents. For example, the server obtains three candidate contents, namely candidate content A, candidate content B, and candidate content C. The playback duration of candidate content A is 20 minutes, the playback duration of candidate content B is 5 minutes, and the playback duration of candidate content C is 5 seconds. The playback duration of candidate content A is longer than that of candidate content B, and the playback duration of candidate content B is longer than that of candidate content C. This shows that the interactive object is most interested in candidate content A, followed by candidate content B, and finally candidate content C. Therefore, the recommendation prediction value of candidate content A obtained by using the single-target recommendation model based on playback duration may be higher than the recommendation prediction value of candidate content B, and the recommendation prediction value of candidate content B may be higher than the recommendation prediction value of candidate content C.

[0074] The multi-objective recommendation model is a target recommendation model that performs recommendation predictions for multiple targets. For example, the multiple targets may be estimated click-through rates and playback durations, that is, the multi-objective recommendation model can simultaneously perform recommendation predictions based on the estimated click-through rates and playback durations of multiple candidate contents, and obtain a recommendation prediction value for each candidate content in the multiple candidate contents. For example, the server obtains three candidate contents, namely candidate content A, candidate content B, and candidate content C. The estimated click-through rate of candidate content A is 80%, and the playback duration is 5 seconds; the estimated click-through rate of candidate content B is 40%, and the playback duration is 3 minutes; the estimated click-through rate of candidate content C is 10%, and the playback duration is 20 minutes. The multi-objective recommendation model can use the estimated click-through rate and playback duration to obtain the recommendation prediction value of the three candidate contents that combines the estimated click-through rate and playback duration. The recommendation prediction value of candidate content A may be higher than that of candidate content B and candidate content C.

[0075] It should be noted that the network structure of the target recommendation model in the embodiment of the present application can have multiple options, for example, it can be a deep learning model of various network structures. In one possible implementation, taking the target recommendation model as a multi-target recommendation model as an example, the network structure of the multi-target recommendation model can be a multi-gate mixture of experts (MMoE) network. The multi-gate mixture of experts network is a commonly used model for multi-target learning and can be composed of an embedding layer, an expert network, a gated network, and a task tower network.

[0076] S203: Perform interest exploration on multiple candidate contents respectively through the interest exploration model to obtain a first predicted probability value for each candidate content.

[0077] The multiple candidate contents may include new content of interest to the interactive object, and may also include historical content of interest to the interactive object. Historical content of interest may refer to content whose corresponding classification belongs to the historical interest classification set. The historical interest classification set may be a set consisting of classifications that the interactive object has determined to be interested in. Therefore, historical content of interest may also be content under the classification that the interactive object has determined to be interested in. New content of interest may refer to content whose corresponding classification does not belong to the historical interest classification set and that the interactive object may be interested in. In order to avoid limiting the content recommended to the interactive object to only a few points of interest and to recommend more new content of interest to the interactive object, the embodiment of the present application may also perform interest exploration on multiple candidate contents separately through the interest exploration model to obtain a first predicted probability value for each candidate content. Among them, the interest exploration model is trained based on multiple sample contents of the interactive object and sample labels corresponding to the multiple sample contents, and the sample labels are used to indicate the interaction between the interactive object and the corresponding sample content.

[0078] Interest exploration can be a process of obtaining content that may be of interest to the interactive object through exploration. The interest exploration model is a model used for interest exploration. The interest exploration model can be used to conduct interest exploration on multiple candidate contents respectively to obtain a first predicted probability value for each candidate content. The first predicted probability value is used to reflect the probability that the corresponding candidate content is a new interest content for the interactive object. The new interest content is content whose satisfaction reaches a satisfaction threshold and whose category does not belong to the historical interest classification set of the interactive object. The satisfaction threshold can be a preset value for determining whether the interactive object is satisfied with the new interest content. It can be set according to actual needs. The embodiment of the present application does not limit the specific value of the satisfaction threshold. If the interactive object's satisfaction with the new interest content reaches the satisfaction threshold, it can indicate that the interactive object is interested in the new interest content. Otherwise, it indicates that the interactive object is not interested in the new interest content.

[0079] It should be noted that there are multiple options for the network structure of the interest exploration model. In one possible implementation, the network structure of the interest exploration model can be a combination of a gated network and a task tower network, or more specifically, a combination of a gated network and a task tower network.

[0080] It is understandable that the interest exploration model can be a network structure independent of the target recommendation model, or it can form a complete network structure together with the target recommendation model, that is, the interest exploration model can also share part of the network structure with the target recommendation model. In one possible implementation, when the target recommendation model is a multi-target recommendation model, and the multi-target recommendation model is MMoE, the interest exploration model shares the embedding layer and expert network with the multi-target recommendation model, and the gating network and task tower network of the interest exploration model are independent of the gating network and task tower network of the multi-target recommendation model.

[0081] Related technologies mainly use the similarity between the categories to which the content belongs to perform interest exploration. In addition to using the similarity between the categories to which the content belongs to perform interest exploration, related technologies can also perform interest exploration based on the Multi-armed Bandit (MAB) algorithm. The Multi-armed Bandit algorithm is an exploration algorithm that continuously explores according to a certain strategy, collects feedback in real time, and continuously updates the algorithm exploration strategy based on the feedback, thereby maximizing benefits in the process of continuous exploration. The Multi-armed Bandit algorithm can recommend content that is as consistent as possible with the interests of the interacting object based on the historical interests of the interacting object. On the other hand, it can also recommend content that is not historically interesting but that the interacting object is also interested in. However, the interest exploration method based on MAB cannot fully utilize the various high-dimensional sparse features of the interacting object and the candidate content, and does not have the ability to explicitly learn interest exploration, resulting in poor interest exploration results through the Multi-armed Bandit algorithm.

[0082] The interest exploration model provided by the embodiment of the present application can make full use of various high-dimensional sparse features of interactive objects and candidate contents (such as features related to interactive objects and features related to candidate contents) through the embedding layer. Compared with the interest exploration based on the multi-armed bandit algorithm, the interest exploration model provided by the present application improves the effect of interest exploration by making full use of features. In addition, the interest exploration model is obtained by training based on multiple sample contents of interactive objects and sample labels corresponding to multiple sample contents. The sample labels serve as supervision information when training the interest exploration model. Since the sample labels are used for supervision during the training process of the interest exploration model, the trained interest exploration model has stronger interest exploration capabilities. When training the interest exploration model, the gating network and task tower network of the interest exploration model have strong fitting capabilities. At the same time, by introducing the interest exploration target, the effect of interest exploration is improved through strong fitting capabilities and the introduction of the interest exploration target. The interest exploration target refers to the target used to predict whether the candidate content is the new interest content of the interactive object.

[0083] It should be understood that the higher the first predicted probability value corresponding to the candidate content, the more likely the candidate content is a new interest content that the interactive object is interested in. For example, the server obtains three candidate contents, namely candidate content A, candidate content B, and candidate content C. Interest exploration is performed through the interest exploration model to obtain the first predicted probability value A corresponding to candidate content A, the first predicted probability value B corresponding to candidate content B, and the first predicted probability value C corresponding to candidate content C. The historical interest classification set of the interactive object includes the classification of candidate content A, and candidate content A is the historical interest content of the interactive object. The historical interest classification set of the interactive object does not include the classification of candidate content B and the classification of candidate content C, then the first predicted probability value A will be relatively lower than the first predicted probability value B and the first predicted probability value C. If the interactive object's satisfaction with candidate content B does not reach the satisfaction threshold, and the interactive object's satisfaction with candidate content C reaches the satisfaction threshold, then the first predicted probability value C will be relatively higher than the first predicted probability value B.

[0084] S204: For each candidate content, fuse the first prediction probability value and the recommended prediction value to obtain a fused prediction value.

[0085] For each candidate content, the server may fuse the first predicted probability value and the recommended predicted value to obtain a fused predicted value. The fused predicted value is positively correlated with the first predicted probability value. The higher the first predicted probability value of the candidate content, the higher the fused predicted value of the candidate content. This fused predicted value is then adjusted to the recommended predicted value, thereby increasing the fused predicted value of the new content of interest to the interactive subject.

[0086] It should be noted that there are multiple ways to fuse the first predicted probability value and the recommended predicted value. In one possible implementation, the first predicted probability value and the recommended predicted value can be fused via a result fusion layer. The result fusion layer can be a neural network layer for fusing the first predicted probability value and the recommended predicted value. When fusing the first predicted probability value and the recommended predicted value via the result fusion layer, the first predicted probability value and the recommended predicted value for each candidate content need only be input into the result fusion layer. The result fusion layer can then fuse the first predicted probability value and the recommended predicted value to obtain a fused predicted value. The fused predicted value can be obtained quickly and conveniently via the result fusion layer.

[0087] In another possible implementation, the first predicted probability value and the recommended predicted value can be fused using a fusion calculation formula to obtain a fused predicted value. When the first predicted probability value and the recommended predicted value are fused using the fusion calculation formula, the first predicted probability value and the recommended predicted value of each candidate content can be substituted into the fusion calculation formula to calculate the corresponding fused predicted value. The fusion calculation formula provides a formalized framework for the fusion of the first predicted probability value and the recommended predicted value. The fusion calculation formula provides guidance for the fusion of the first predicted probability value and the recommended predicted value, thereby making it possible to fuse the first predicted probability value and the recommended predicted value using mathematical methods.

[0088] In the embodiment of the present application, different fusion calculation formulas can be selected. For example, the fusion calculation formula can be a power multiplication calculation formula. When the fusion calculation formula is a power multiplication calculation formula, the first predicted probability value can be multiplied by the recommended predicted value in a power multiplication manner to obtain a fusion prediction value. In the power multiplication calculation formula, the exponent of the first predicted probability value is the weight of the first predicted probability value. For ease of understanding, the embodiment of the present application provides a power multiplication calculation formula, and the expression of the power multiplication calculation formula is:

[0089] s1=s0·p α

[0090] Where s1 is the fused prediction value calculated using the power multiplication formula, s0 is the recommended prediction value, p is the first prediction probability value, and α is the weight of the first prediction probability value. A larger value of α results in a greater influence of the first prediction probability value on the fused prediction value. The optimal value of α can be determined through online parameter adjustment experiments.

[0091] In the power multiplication formula of the present embodiment, the exponent of the first predicted probability value is the weight of the first predicted probability value. When fusing the first predicted probability value and the recommendation prediction value using the power multiplication formula, the higher the weight of the first predicted probability value, the greater its influence on the fused prediction value. The resulting fused prediction value can accurately assess the potential interaction between the candidate content and the interactive object during content recommendation.

[0092] S205: Determine, among the multiple candidate contents, the candidate content whose fusion prediction value meets the preset recommendation condition as the recommendation result.

[0093] The server determines as a recommendation result the candidate content whose fused prediction value satisfies a preset recommendation condition among the multiple candidate content. The preset recommendation condition may be a pre-set condition for selecting candidate content to recommend to the interactive object. This application does not limit the preset recommendation condition. For example, the preset recommendation condition may be that the fused prediction value is ranked in the top K, or that the fused prediction value reaches a prediction value threshold, etc.

[0094] It should be noted that there may be multiple implementations of S205, and the implementations of S205 may be different depending on the preset recommendation conditions. In one possible implementation, the preset recommendation condition may be that the fusion prediction value is ranked in the top K. At this time, the method of determining the candidate content whose fusion prediction value meets the preset recommendation condition among multiple candidate contents as the recommendation result may be that the server sorts the multiple candidate contents in descending order according to the fusion prediction value. The larger the fusion prediction value, the higher the order. The server may determine the candidate content ranked in the top K as the recommendation result. K may be set according to actual needs. For example, K may be 20. The embodiment of the present application does not limit the value of K.

[0095] In another possible implementation, the preset recommendation condition may be that the fusion prediction value reaches a prediction value threshold. At this time, the method of determining the candidate content whose fusion prediction value meets the preset recommendation condition among multiple candidate contents as the recommendation result may be that the server determines the candidate content whose fusion prediction value reaches the prediction value threshold among multiple candidate contents as the recommendation result.

[0096] In some cases, the user may have pre-set categories for content they are not interested in. If candidate content falls into these categories, we should avoid recommending this candidate content to them. For example, if the user is trying to lose weight, they may have pre-set a preference for not browsing food-related content. In this case, the server will exclude food-related candidate content when determining recommendation results and recommend only those candidate content that does not fall into the food category and has a fusion prediction value greater than the prediction value threshold.

[0097] After the server determines the recommendation result, it can send the recommendation result to the terminal corresponding to the content recommendation request. After the terminal receives the recommendation result, it can display the recommendation result on the terminal. When displaying the recommendation result, the recommendation result can be displayed in the form of a card (for example, a double-column stream card). When the content recommendation scenario is a double-column stream video recommendation under video recommendation, the recommendation result can be displayed in the form of a double-column stream card.

[0098] As can be seen from the above technical solution, when performing content recommendation, the present application can add an interest exploration model to the existing target recommendation model to assist in content recommendation. Specifically, multiple candidate content can be obtained, and recommendation predictions can be performed on the multiple candidate content using the target recommendation model to obtain a recommendation prediction value for each of the multiple candidate content. The interest exploration model can perform interest exploration on each of the multiple candidate content, obtaining a first prediction probability value for each candidate content. The interest exploration model is trained based on multiple sample content of an interactive object and the sample labels corresponding to the multiple sample content. The sample labels are used to indicate the interaction between the interactive object and the corresponding sample content. Therefore, the interest exploration model accurately learns the interests of the interactive object. The first prediction probability value obtained based on the interest exploration model can accurately reflect the probability that the corresponding candidate content is a new interest content for the interactive object. The higher the first prediction probability value of the candidate content, the more likely the candidate content is a new interest of the interactive object. Therefore, for each candidate content, the first prediction probability value and the recommendation prediction value are fused to obtain a fused prediction value. The fused prediction value is positively correlated with the first prediction probability value, thereby adjusting the recommendation prediction value to obtain a fused prediction value, thereby increasing the fused prediction value of the new interest content of the interactive object. In this way, when candidate content whose fused prediction value meets the preset recommendation conditions among multiple candidate content is determined as a recommendation result, candidate content belonging to new interest content can be ranked at the top, helping to explore new interests. Because the first predicted probability value can accurately reflect the probability that the candidate content is the new interest content of the interactive object, the fused prediction value adjusted based on the first predicted probability value can also accurately reflect the possibility that the candidate content is the new interest content of the interactive object. In this way, the first predicted probability value can be used to rank the new interest content of the interactive object at the top, improving the accuracy of interest exploration, thereby increasing the possibility of successful interest exploration and expanding the improvement of interest exploration effect.

[0099] Next, an example is given of the content recommendation process of an embodiment of the present application in a video recommendation scenario. The terminal can be a smartphone, and the smartphone of the interactive object has a video application installed. When the interactive object opens the video application, the video application can send a content recommendation request to the server. Based on the content recommendation request, the server obtains multiple videos (candidate content) from the storage database corresponding to the video application. The server predicts the recommendation of multiple videos using the target recommendation model and obtains a recommendation prediction value for each video.

[0100] In related technologies, the recommendation prediction value can be directly used to recommend content to the interactive object. Figure 3 A content recommendation diagram is shown in Figure 3In the recommended results, the multiple videos are presented to the interacting user as a two-column stream of cards. Clicking on a video takes them to the immersive page. The top four videos with the highest predicted recommendation values ​​are a football match video, a basketball match video, a football game commentary video, and a basketball game commentary video. Therefore, these four videos are recommended to the interacting user. When displaying the recommended results, the four videos are arranged according to their corresponding predicted recommendation values, with videos with higher predicted recommendation values ​​appearing higher in the ranking.

[0101] When recommending content to an interactive object, the server can use an interest exploration model to explore the interests of multiple videos separately to broaden the range of recommended videos and avoid being limited to a few points of interest. The server then obtains a first predicted probability value for each video, which reflects the probability that the corresponding video represents new content of interest to the interactive object. For each video, the server then fuses the first predicted probability value with the recommendation prediction value to obtain a fused prediction value. The server then determines as a recommendation result the video whose fused prediction value satisfies the preset recommendation criteria among the multiple videos.

[0102] The embodiment of the present application can use the fusion prediction value to recommend content to the interactive object. Figure 4 This is a diagram of another type of content recommendation. In the recommendation results, the top four videos with the highest fusion prediction values ​​are a football match video, a basketball game video, a travel guide video, and a football game review video. Therefore, these four videos can be recommended to the interacting user. When displaying the recommendation results, these four videos are arranged according to their corresponding fusion prediction values, with videos with higher fusion prediction values ​​being ranked higher.

[0103] By comparison Figure 3 and Figure 4 As can be seen, after exploring the interests of the interactive subject, travel guides are ranked higher in content recommendations. The fused prediction value is obtained by adjusting the first prediction probability value. The first prediction probability value reflects the likelihood that the candidate content is a new interest for the interactive subject. The fused prediction value also reflects the likelihood that the candidate content is a new interest for the interactive subject. Among multiple videos, the travel guide has a relatively high first prediction probability value, indicating a greater likelihood of being a new interest. Therefore, it is ranked relatively high in content recommendations for the interactive subject. Figure 4 The corresponding content recommendation diagram enables the new interest content that the interacting object is interested in to be placed in the front, thereby improving the accuracy of interest exploration and further increasing the possibility of successful interest exploration.

[0104] By combining interest exploration in the content recommendation process, the recommendation system can recommend videos with a wider range of interests and in line with the interests of the interacting objects, increase the distribution of videos of potential interest to users, and improve the richness of recommended content as a whole, thereby improving the user experience.

[0105] The interest exploration model can accurately explore the interests of the interactive object. The interest exploration model can be trained by a computer device, which can be a server or a terminal. Next, the training method of the interest exploration model provided by the embodiment of the present application is introduced by taking the server training interest exploration model as an example and combining the accompanying drawings. Among them, the server for training the interest exploration model can be the same as or different from the server for content recommendation. Figure 5 , Figure 5 A flowchart of a method for training an interest exploration model is shown, the method comprising the following steps:

[0106] S501: Acquire multiple sample contents and sample labels corresponding to the multiple sample contents.

[0107] The server can obtain multiple sample contents and sample labels corresponding to the multiple sample contents in the historical data. The sample content can be the content used when training the interest exploration model, and the type of the sample content is related to the application scenario of the interest exploration model. For example, when the interest exploration model is used to explore the interest of videos, the type of the sample content is video; when the interest exploration model is used to explore the interest of pictures, the type of the sample content is picture. The sample content has a corresponding sample label, and the sample label is used to indicate the interaction between the interactive object and the corresponding sample content, thereby reflecting whether the sample content is new interest content. In the scenario where the interest exploration model is used to explore the interest of videos, the sample content can be a video.

[0108] S502: Perform interest exploration on multiple sample contents respectively through the initial network model to obtain a second predicted probability value for each sample content.

[0109] Multiple sample contents are input into the initial network model. The initial network model then performs interest exploration on each of the sample contents to obtain a second predicted probability value corresponding to each sample content. The second predicted probability value, obtained by the initial network model, represents the probability that the sample content is a new content of interest to the interacting subject.

[0110] S503: For each sample content, determine the degree of difference between the second predicted probability value of the sample content and the sample label of the sample content.

[0111] The sample label is used to indicate the interaction between the interactive object and the corresponding sample content, and thus reflects whether the sample content is new interest sample content. The second predicted probability value is used to indicate the probability that the sample content is new interest sample content. The server can determine the degree of difference between the second predicted probability value of each sample content and the sample label of the sample content.

[0112] S504: Determine the sample training weight of each sample content.

[0113] When training the interest exploration model, the training weights corresponding to different sample contents may be different. In order to train an interest exploration model that can explore new interest content, the sample training weight of the sample content should be determined based on the following principle, that is, among multiple sample contents, the sample training weight of the new interest sample content is greater than the sample training weight of the historical interest sample content. The new interest sample content may refer to sample content whose corresponding classification does not belong to the historical interest classification set and the interactive object is interested in (that is, the satisfaction reaches the satisfaction threshold), that is, the interactive object's satisfaction with the new interest sample content reaches the satisfaction threshold, and the historical interest classification set does not include the classification to which the new interest sample content belongs. The historical interest sample content may refer to sample content whose corresponding classification belongs to the historical interest classification set, that is, the historical interest classification set includes the classification to which the historical interest sample content belongs. The embodiment of the present application does not limit the value of the sample training weight of the new interest sample content and the value of the sample training weight of the historical interest sample content. It can be a fixed value or a dynamic value, as long as it is ensured that the sample training weight of the new interest sample content is greater than the sample training weight of the historical interest sample content.

[0114] For example, in the video recommendation scenario, video A belongs to the new interest sample content, and video B belongs to the historical interest sample content. The sample training weight of video A is greater than the sample training weight of video B. In this case, the sample training weight of video A can be 3, and the sample training weight of video B can be 1.

[0115] It should be noted that the server can determine the sample training weight of each sample content through a variety of methods, and this application does not limit the method for determining the sample training weight of each sample content. In one possible implementation method, the method for determining the sample training weight of each sample content can be that the server obtains the initial weight of each sample content, and the initial weight is the initial weight corresponding to the sample content when training the interest exploration model. The initial weight of the sample content can be set manually, and the initial weight of each sample content can be the same. The embodiment of this application does not limit the value of the initial weight, for example, it can be 1.

[0116] Next, the server may determine whether the sample content belongs to new interest sample content or historical interest sample content, and then achieve a sample training weight of the new interest sample content greater than the sample training weight of the historical interest sample content by at least one of increasing the initial weight of the new interest sample content and decreasing the initial weight of the historical interest sample content. Specifically, if the server determines that the sample content belongs to new interest sample content, the server may increase the initial weight of the sample content to obtain a sample training weight for the sample content; if the server determines that the sample content belongs to historical interest sample content, the server may decrease the initial weight of the sample content or keep the initial weight unchanged to obtain a sample training weight for the sample content, thereby making the sample training weight of the new interest sample content greater than the sample training weight of the historical interest sample content.

[0117] For example, the server obtains the initial weights of three videos, namely the initial weight of video A, the initial weight of video B, and the initial weight of video C. The initial weight of video A, the initial weight of video B, and the initial weight of video C are all 1. After obtaining the initial weights, the server can determine whether the three videos belong to new interest sample content or historical interest sample content. If the server determines that video A belongs to new interest sample content, video B and video C belong to historical interest sample content. In one possible implementation method, the server can increase the initial weight of video A to 3 (that is, the sample training weight of video A is 3), and maintain the initial weight of video B and the initial weight of video C unchanged (that is, the sample training weights of video B and video C are 1), and obtain the sample training weights corresponding to video A, video B, and video C respectively. In another possible implementation, the server can increase the initial weight of video A to 3 (i.e., the sample training weight of video A is 3), and reduce the initial weight of video B and the initial weight of video C to 0.5 (i.e., the sample training weights of video B and video C are 0.5), and obtain the sample training weights corresponding to video A, video B, and video C respectively.

[0118] In addition, the server can maintain the initial weight of video A at 1 (that is, the sample training weight of video A is 1), and reduce the initial weights of video B and video C to 0.5 (that is, the sample training weights of video B and video C are 0.5), and obtain the sample training weights corresponding to video A, video B and video C respectively.

[0119] The above method obtains an initial weight for each sample content. After determining whether the sample content belongs to new interest sample content or historical interest sample content, the initial weight corresponding to the sample content is adjusted. This adjustment makes the sample training weight of the new interest sample content greater than the sample training weight of the historical interest sample content. Adjusting the sample training weight based on whether the sample content belongs to new interest sample content or historical interest sample content can make new interest sample content that meets the needs of the interactive object play a more important role in model training, allowing the ultimately trained interest exploration model to more accurately predict the new interest content of the interactive object.

[0120] It should be noted that the server can enhance the initial weight of the sample content through a variety of methods, and the embodiment of the present application does not limit the method for enhancing the initial weight of the sample content. In one possible implementation, in order to reflect the importance of different sample content in the training process, the initial weights of different sample content can be enhanced to different degrees. The satisfaction of the interactive object with different sample content may be different. The higher the satisfaction of the sample content, the more it can help the interest exploration model learn how to explore the new interest content of the interactive object. Therefore, the server can enhance the initial weight of the sample content based on the satisfaction of the interactive object with the sample content to obtain the sample training weight of the sample content. The higher the satisfaction of the interactive object with the sample content, the higher the sample training weight obtained after enhancement.

[0121] For example, in a video recommendation scenario, the initial weights of Video A and Video B are both 1. The server determines that both Video A and Video B are new interest sample content. The interacting audience is more satisfied with Video A than with Video B. The server can increase the initial weight of Video A to 3 and the initial weight of Video B to 2 to obtain the sample training weights for Video A and Video B.

[0122] The above method can enhance the initial weight of sample content based on the interaction object's satisfaction with the sample content. The higher the interaction object's satisfaction with the sample content, the higher the sample training weight of the sample content, indicating that the sample content is more important in model training. Using the interaction object's satisfaction with the sample content as a condition for setting the sample training weight, the sample training weight of the sample content is dynamically set. The interest exploration model trained with the dynamically set sample training weight can more accurately explore the interests of the interaction object.

[0123] The goal of the interest exploration model is to discover new interests for the interacting party. Accordingly, during the training process, the interest exploration model can learn to explore new interests based on new interest sample content. Determining whether a sample is new requires first determining whether the corresponding category of the sample belongs to the set of historical interest categories and secondly determining whether the interacting party is interested in the sample. This determination is relatively difficult.

[0124] In an embodiment of the present application, the interaction between an interactive object and sample content can be used to determine whether the interactive object is interested in the sample content. For example, when an interactive object clicks on the sample content, it can indicate, to a certain extent, that the interactive object is interested in the sample content. The degree of the interactive object's interest in the sample content can be determined by the number of clicks on the sample content, the number of favorites, likes, shares, and followers generated by the interactive object. In one possible implementation, whether the interactive object is interested in the sample content can be determined by the number of clicks on the sample content or the number of favorites, likes, shares, and followers generated by the interactive object. For another example, the degree of interest in the sample content can be reflected by the length of time the interactive object stays on the sample content. If the interactive object stays on the sample content for a short time, it indicates that the interactive object is less interested in the sample content; if the interactive object stays on the sample content for a long time, or even plays it several times, it indicates that the interactive object is more interested in the sample content.

[0125] The length of time that an interactive object stays on the sample content will be affected by the physical length of the sample content. The longer the physical length of the sample content, the longer the interactive object's stay will tend to be. For example, the length of time that an interactive object stays on a video with a physical length of 10 seconds is 9 seconds; the length of time that an interactive object stays on a video with a physical length of 60 seconds is 10 seconds. Since the latter has a longer physical length, the length of time that the interactive object stays on the former is also longer than the former, but from the perspective of the degree of interest of the interactive object, the degree of interest of the interactive object in the former should be greater than that in the latter. In one possible implementation method, the physical length of the sample content and the length of time that the interactive object stays on the sample content can be considered simultaneously to determine whether the interactive object is interested in the sample content. Based on this, the server can determine whether the sample content belongs to new interest sample content by the following method:

[0126] The server can obtain the physical duration of the sample content and the duration that the interactive object stays on the sample content. The physical duration can be the time required for the interactive object to fully browse the sample content. The physical duration of the sample content is different in different scenarios. In the scenario of video recommendation, the sample content can be a video, and the physical duration of the sample content is the length of the video. In the scenario of audio recommendation, the sample content can be audio, and the physical duration of the sample content is the length of the audio. In the scenario of news recommendation, the sample content can be news, and the physical duration of the sample content is the time required for the interactive object to fully browse the news. The duration of stay can be the length of time the interactive object stays on the sample content. The duration of stay of the interactive object on the sample content can reflect the interactive object's satisfaction with the sample content. The longer the interactive object stays on the sample content, the higher the interactive object's satisfaction with the sample content and the greater the interactive object's interest in the sample content.

[0127] The server predicts the average duration of the sample content based on its physical duration, obtaining the average dwell time. This average dwell time is used to reflect the satisfaction threshold. For sample content of a given physical duration, if the dwell time of an interacting subject exceeds the satisfaction threshold, it can be considered that the interacting subject has a high level of interest in the sample content. If the dwell time of an interacting subject on the sample content reaches the average dwell time, and the category to which the sample content belongs is not in the historical interest category set, the sample content is determined to be newly interesting sample content.

[0128] The above method predicts the average duration through the physical duration of the sample content to obtain the average stay duration of the sample content, and the average stay duration is used to reflect the satisfaction threshold. When the interactive object's stay duration on the sample content is greater than the average stay duration, it proves that the interactive object has a high degree of interest in the sample content. If the category to which the sample content belongs does not exist in the historical interest classification set, it is determined that the sample content belongs to new interest sample content. By combining the physical duration and the stay duration to determine whether the sample content belongs to the new interest sample content, the determination bias caused by the physical duration can be avoided, so that it can be more accurately determined whether the sample content belongs to the new interest sample content of the interactive object.

[0129] In one possible implementation, the server can use the fitted satisfaction curve to predict the average duration of the sample content's physical duration, thereby obtaining the average dwell time of the sample content. The fitted satisfaction curve can more accurately and quickly predict the average dwell time corresponding to sample content of any physical duration.

[0130] It should be noted that the server can obtain the satisfaction curve by fitting through a variety of methods, and this application does not limit the method of fitting the satisfaction curve. In one possible implementation, the server can perform duration statistics on multiple sample contents, and obtain multiple data points through duration statistics. Each of the multiple data points is used to reflect the average stay time under the corresponding physical duration. Then, according to the satisfaction function, a curve is fitted to the multiple data points to obtain the satisfaction curve. The satisfaction function is a function that calculates the average stay time based on the physical duration, and the satisfaction function meets at least the following conditions: the values ​​of the physical duration and the average stay time in the satisfaction function are respectively greater than zero; the average stay time in the satisfaction function increases monotonically relative to the physical duration; when the physical duration in the satisfaction function approaches infinity, the average duration calculated based on the satisfaction function and the physical duration approaches a constant greater than zero. Among them, the satisfaction curve can be expressed as T=f(t), where t represents the physical duration, T represents the average stay time, and f() represents the satisfaction function.

[0131] When performing duration statistics on multiple sample contents, all sample contents with the same physical duration can be obtained, and the average duration can be obtained based on the duration corresponding to these sample contents, thereby obtaining a data point consisting of the physical duration and the average duration. Any data point can be expressed as (t i ,T i ), i=1,2,…,M,t i Indicates physical duration, T i Represents the physical duration t i The corresponding average residence time, M is the number of sampled data points. For all sampled data points, Figure 6 A data graph of physical duration and average stay duration provided in the embodiment of the present application. Figure 6 In the example, the horizontal axis is the physical duration of the sample content in seconds, and the vertical axis is the average dwell time in seconds. Starting from a physical duration of 5 seconds, a data point is obtained every 1 second until the physical duration reaches 600 seconds. A total of 596 data points are obtained, that is, =596. Each data point has a corresponding average dwell time. Through 596 data points, we can get Figure 6 The data graph shown. Figure 6 It can be seen that Figure 6 The data points in the data graph shown have the following pattern: when the physical duration is less than 300 seconds, the average stay time of the interactive object increases with the increase of the physical duration. When the physical duration exceeds 300 seconds, the average stay time remains basically unchanged.

[0132] In a possible implementation, the server can Figure 6 The curve fitting of the data points in Figure 7The satisfaction curve shown in Figure 7 A schematic diagram of a satisfaction curve provided in an embodiment of the present application is shown. Figure 7 The horizontal axis is the physical duration in seconds; the vertical axis is the average stay time in seconds. Figure 7 The thick curve in is the satisfaction curve. For any physical duration, Figure 7 The satisfaction curve shown estimates the average stay time corresponding to the physical duration, and the average stay time can be used as a satisfaction threshold.

[0133] Based on the rules of the above data points, in order to fit a satisfaction curve that conforms to the above rules, the satisfaction function must meet at least the following conditions: first, the values ​​of the physical duration and the average stay duration in the satisfaction function are greater than zero respectively; second, the average stay duration in the satisfaction function increases monotonically relative to the physical duration; third, when the physical duration in the satisfaction function approaches infinity, the average duration calculated based on the satisfaction function and the physical duration approaches a constant greater than zero.

[0134] It should be noted that any function that meets the above three conditions can be used as a satisfaction function. When the above three conditions are met, different functional forms of the satisfaction function can be designed. In addition to the above three conditions, the satisfaction function must be able to fit the actual data well, that is, the fitting error must be as small as possible. After continuous verification, in one possible implementation method, the functional form of the satisfaction function designed in the embodiment of the present application is as follows:

[0135]

[0136] Among them, t is the physical duration, f(t) is the average stay time calculated when the physical duration is t, a, b, c and d are the parameters to be learned in the satisfaction function, and all the data points (t i ,T i ) is fitted.

[0137] When learning the parameters in the satisfaction function, in order to better fit the actual situation, the learning goal is to make the fitting error as small as possible. To this end, the embodiment of the present application provides the following fitting goals:

[0138]

[0139] The values ​​of c and d, t i is the physical duration of the i-th data point obtained by sampling, T i is the average residence time of the i-th data point obtained by sampling, is the average length of stay calculated based on the satisfaction function and physical duration.

[0140] The algorithms used for fitting may include multiple ones. In one possible implementation, the server may use all the acquired data points (t i ,T i ), and the satisfaction curve is obtained by least square fitting.

[0141] Through the satisfaction function that meets the conditions, curve fitting is performed on multiple sampled data points. The satisfaction function can accurately reflect the data patterns of multiple data points, and then a more accurate satisfaction curve can be obtained by fitting, so as to improve the prediction accuracy of the average stay time.

[0142] S505 : Based on the sample training weight of each sample content, perform weighted summation on the difference degrees corresponding to the multiple sample contents to obtain a target loss function.

[0143] The sample training weights for new interest sample content are different from those for historical interest sample content. The sample training weights for new interest sample content are larger than those for historical interest sample content. Based on the sample training weights for each sample content, the weighted sum of the corresponding differences across multiple sample contents is used to generate the target loss function.

[0144] The embodiment of the present application does not limit the specific form of the target loss function. In one possible implementation, the target loss function can be weighted binary cross entropy, mean square error, etc. Taking the weighted binary cross entropy as an example, the target loss function can be calculated using the following formula:

[0145]

[0146] Among them, L(θ) is the target loss function, θ is the parameter to be learned by the model, N is the total number of sample contents, i is a positive integer greater than 0 and less than or equal to N, and w i Represents the sample training weight of the i-th sample content. i Indicates the sample label of the i-th sample content, indicating the click of the interactive object on the sample content. If the i-th sample content is clicked by the interactive object, then y i =1, if the i-th sample content is not clicked by the interactive object, then y i =0. Represents the second predicted probability value corresponding to the i-th sample content.

[0147] In the above objective loss function, w i It can be determined by S504. In one possible implementation, w i It can be expressed by the following formula:

[0148]

[0149] Among them, w i Represents the sample training weight of the i-th sample content. i Indicates the sample label of the i-th sample content, indicating the click of the interactive object on the sample content. If the i-th sample content is clicked by the interactive object, then y i =1, otherwise y i =0. x i represents the duration of the interaction object's stay on the i-th sample content, f(t i ) represents the average length of stay of the i-th sample content obtained through the satisfaction curve. i Indicates the category to which the content of the i-th sample belongs, A i Represents the historical interest classification set of the interaction object. i =1 and x i >f(t i ) indicates that after the interactive object clicks on the i-th sample content, the interactive object's stay time on the i-th sample content is greater than the average stay time. Indicates that the category to which the i-th sample content belongs does not belong to the historical interest category set, that is, the i-th sample content belongs to the new interest sample content. At this time, the sample training weight of the i-th sample content is 3.0; at other times, the i-th sample content may belong to the historical interest sample content, or the i-th sample content may not be clicked by the interactive object or the satisfaction threshold is not reached after being clicked by the interactive object. At this time, the sample training weight of the i-th sample content is 1.0.

[0150] S506: Optimize the model parameters of the initial network model using the target loss function to obtain an interest exploration model.

[0151] The server optimizes the model parameters of the initial network model based on the target loss function to obtain an interest exploration model. In one possible implementation, the server can stop optimizing the model parameters of the initial network model when a training cutoff condition is reached. The training cutoff condition can be a pre-set number of training rounds, or when the model obtained after optimizing the model parameters of the initial network model reaches a preset accuracy. When the training cutoff condition is met, training is stopped, and the interest exploration model is obtained. The trained interest exploration model is used in online services to conduct interest exploration in the content recommendation process.

[0152] The training method of the above-mentioned interest exploration model obtains sample content and corresponding sample labels. The sample content can include new interest sample content and historical interest sample content, and the sample labels are used to indicate the interaction between the interactive object and the corresponding sample content. When training the initial network model, the initial network model is used to explore the interest of the sample content to obtain a second predicted probability value for each sample content. The second predicted probability value is the probability that the sample content obtained by the initial network model is the new interest sample content of the interactive object. The degree of difference between the second predicted probability value and the sample label of the sample content can reflect the prediction accuracy of the initial network model for the sample content. Since the interest exploration model is to explore the interest of the interactive object, the sample training weight of the new interest sample content of the interactive object in the target loss function is greater than the sample training weight of the historical interest sample content in the loss function. Using the sample training weight of each sample content, the degree of difference corresponding to multiple sample contents is weighted and summed to obtain the target loss function, and the target loss function is used to optimize the model parameters of the initial network model to obtain the interest exploration model. Since the sample training weight of new interest sample content is larger in the target loss function, the model adjusted by the target loss function has a stronger learning ability for new interest sample content than for historical interest sample content. The trained interest exploration model can more accurately explore the interests of interactive objects.

[0153] Applying the interest exploration model obtained through the above training to content recommendation can improve the recommendation system's ability to explore new interests of users. In the content recommendation method provided in the embodiment of the present application, content recommendations for interest exploration have certain improvements in the click-through rate of recommendation results, the average number of clicked categories per person, and the average number of exposed categories per person, compared to content recommendations without interest exploration. The average number of clicked categories per person can be the number of categories to which the recommended results clicked on average by each interactive object belong, and the average number of exposed categories per person can be the average number of categories to which candidate content belongs in the recommended results of each interactive object.

[0154] Next, in the context of video recommendations, we will compare the data between content recommendations for interest exploration and content recommendations for non-interest exploration. Figure 8 A data comparison chart of the average number of exposure categories per person for content recommendation for interest exploration and content recommendation for non-interest exploration provided in an embodiment of the present application. Figure 8 The horizontal axis is the date, and the vertical axis is the percentage increase in the number of categories exposed per capita in Group B (experimental group) compared to Group A (control group). Figure 8In the figure, the line corresponding to Group A represents the data corresponding to content recommendations for those with no interest in exploration, and the line corresponding to Group B represents the data corresponding to content recommendations for those with interest in exploration. By comparing the lines of Groups A and B, it is clear that the data on the number of categories of exposure per person for content recommendations for those with interest in exploration (Group B) has significantly increased compared to the data on content recommendations for those with no interest in exploration (Group A). ​​The average increase in the number of categories of exposure per person for content recommendations for those with interest in exploration is approximately 0.55%.

[0155] In the embodiment of the present application, the average number of exposure categories per person for content recommendations for interest exploration increases more significantly, indicating that more categories of recommendation results are recommended to the interactive objects, and the interactive objects have more categories to choose from.

[0156] Figure 9 A data comparison chart of the average number of click categories per person for content recommendation for interest exploration and content recommendation for non-interest exploration provided in an embodiment of the present application. Figure 9 The horizontal axis is the date, and the vertical axis is the percentage increase in the number of categories clicked per person in Group B (experimental group) compared to Group A (control group). Figure 9 In the figure, the line corresponding to Group A represents the data corresponding to content recommendations for those with no interest in exploration, and the line corresponding to Group B represents the data corresponding to content recommendations for those with interest in exploration. By comparing the lines of Groups A and B, it is clear that the data on the number of categories clicked per person for content recommendations for those with interest in exploration (Group B) has significantly increased compared to the data on content recommendations for those with no interest in exploration (Group A). ​​The average increase in the number of categories clicked per person for content recommendations for those with interest in exploration is approximately 1.27%.

[0157] In the embodiment of the present application, the average number of clicks on categories for content recommendations for interest exploration increases more significantly, indicating that the interactive objects clicked on the newly explored categories, that is, they were interested in the newly explored recommendation results.

[0158] Figure 10 A data comparison chart of click-through rates of content recommendations for interest exploration and content recommendations without interest exploration provided in an embodiment of the present application. Figure 10 The horizontal axis is the date, and the vertical axis is the percentage of click-through rate increase of group B (experimental group) compared with group A (control group). Figure 10 In the figure, the line corresponding to Group A represents the data corresponding to content recommendations for those with no interest in exploration, and the line corresponding to Group B represents the data corresponding to content recommendations for those with interest in exploration. By comparing the lines of Groups A and B, it is clear that the click-through rate of content recommendations for those with interest in exploration (Group B) has been significantly improved compared to the data for content recommendations for those with no interest in exploration (Group A). ​​The click-through rate of content recommendations for those with interest in exploration has increased by an average of approximately 1.06%.

[0159] In the embodiment of the present application, the click-through rate of content recommendations for interest exploration is increased more significantly, indicating that the interactive object has clicked on the newly explored recommendation results. The click behavior can to a certain extent indicate that the interactive object is interested in the clicked recommendation result.

[0160] It should be noted that, based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods.

[0161] based on Figure 2 Corresponding to the content recommendation method provided in the embodiment, the embodiment of the present application also provides a content recommendation device 1100. Figure 11 As shown, the content recommendation device 1100 includes an acquisition unit 1101, a recommendation prediction unit 1102, an interest exploration unit 1103, a fusion unit 1104 and a determination unit 1105:

[0162] The acquisition unit 1101 is configured to acquire multiple candidate contents;

[0163] The recommendation prediction unit 1102 is configured to perform recommendation prediction on the plurality of candidate contents using a target recommendation model to obtain a recommendation prediction value for each of the plurality of candidate contents;

[0164] The interest exploration unit 1103 is configured to perform interest exploration on each of the plurality of candidate contents through an interest exploration model to obtain a first predicted probability value for each of the candidate contents, wherein the first predicted probability value is used to reflect the probability that the corresponding candidate content is a new interest content of the interactive object, wherein the new interest content is content whose satisfaction reaches a satisfaction threshold and whose classification does not belong to the historical interest classification set of the interactive object, and the interest exploration model is trained based on a plurality of sample contents of the interactive object and sample labels corresponding to the plurality of sample contents, respectively, and wherein the sample labels are used to indicate the interaction between the interactive object and the corresponding sample content;

[0165] The fusion unit 1104 is configured to fuse the first prediction probability value and the recommendation prediction value for each candidate content to obtain a fused prediction value, where the fused prediction value is positively correlated with the first prediction probability value;

[0166] The determining unit 1105 is configured to determine, among the plurality of candidate contents, a candidate content whose fusion prediction value satisfies a preset recommendation condition as a recommendation result.

[0167] In a possible implementation, the fusion unit is specifically configured to:

[0168] fusing the first predicted probability value and the recommended predicted value using a fusion calculation formula to obtain a fused predicted value;

[0169] Alternatively, the first prediction probability value and the recommended prediction value are fused using a result fusion layer to obtain a fused prediction value.

[0170] In a possible implementation, the fusion unit is specifically configured to:

[0171] Using the power multiplication calculation formula, the first predicted probability value is multiplied by the recommended prediction value in a power multiplication manner to obtain the fused prediction value. In the power multiplication calculation formula, the exponent of the first predicted probability value is the weight of the first predicted probability value.

[0172] In a possible implementation, the interest exploration model is obtained by training through a model training unit, and the model training unit is specifically used to:

[0173] Obtaining a plurality of sample contents and sample labels corresponding to the plurality of sample contents;

[0174] Performing interest exploration on the plurality of sample contents respectively through the initial network model to obtain a second predicted probability value for each of the sample contents;

[0175] For each of the sample contents, determining a degree of difference between the second predicted probability value of the sample content and the sample label of the sample content;

[0176] Determining a sample training weight for each of the sample contents, wherein the sample training weight of the new interest sample content among the plurality of sample contents is greater than the sample training weight of the historical interest sample content, the new interest sample content being sample content whose satisfaction reaches the satisfaction threshold and whose category does not belong to the historical interest category set, the historical interest category set including the category to which the historical interest sample content belongs;

[0177] Based on the sample training weight of each of the sample contents, weighted summation is performed on the difference degrees corresponding to the plurality of sample contents to obtain a target loss function;

[0178] The target loss function is used to optimize the model parameters of the initial network model to obtain the interest exploration model.

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

[0180] Obtaining an initial weight for each of the sample contents;

[0181] For each of the sample contents, if it is determined that the sample content belongs to the new interest sample content, the initial weight of the sample content is enhanced to obtain a sample training weight of the sample content;

[0182] If it is determined that the sample content belongs to the historical interest sample content, the initial weight of the sample content is weakened or kept unchanged to obtain a sample training weight of the sample content.

[0183] In a possible implementation, the model training unit is specifically configured to:

[0184] Based on the interactive object's satisfaction with the sample content, the initial weight of the sample content is enhanced to obtain a sample training weight of the sample content. The enhanced sample training weight is positively correlated with the interactive object's satisfaction with the sample content.

[0185] In a possible implementation, the model training unit is specifically configured to:

[0186] Obtaining the physical duration of the sample content and the duration of the interactive object's stay on the sample content, and obtaining the historical interest classification set, wherein the duration of the interactive object's stay on the sample content is used to reflect the interactive object's satisfaction with the sample content;

[0187] Performing an average duration prediction based on the physical duration of the sample content to obtain an average dwelling duration of the sample content, wherein the average dwelling duration of the sample content is used to reflect the satisfaction threshold;

[0188] If the duration of the interaction object's stay on the sample content reaches the average duration, and the category to which the sample content belongs does not exist in the historical interest category set, it is determined that the sample content belongs to new interest sample content.

[0189] In a possible implementation, the model training unit is specifically configured to:

[0190] The average duration of the physical duration of the sample content is predicted by fitting the obtained satisfaction curve to obtain the average stay duration of the sample content.

[0191] In a possible implementation, the model training unit further includes a fitting subunit, and the fitting subunit is specifically configured to:

[0192] Performing duration statistics on the plurality of sample contents to obtain a plurality of data points, wherein each of the plurality of data points is used to reflect an average duration of stay under a corresponding physical duration;

[0193] According to the satisfaction function, curve fitting is performed on the multiple data points to obtain the satisfaction curve, the satisfaction function is a function that calculates the average stay time based on the physical time, and the satisfaction function satisfies at least the following conditions: the values ​​of the physical time and the average stay time in the satisfaction function are respectively greater than zero; the average stay time in the satisfaction function increases monotonically relative to the physical time; when the physical time in the satisfaction function approaches infinity, the average time calculated based on the satisfaction function and the physical time approaches a constant greater than zero.

[0194] In a possible implementation manner, the determining unit is specifically configured to:

[0195] Sorting the plurality of candidate contents in descending order of the fusion prediction value; determining the candidate contents ranked in the top K as the recommendation results, wherein the preset recommendation condition is that the candidate contents ranked in the top K of the fusion prediction value;

[0196] Alternatively, a candidate content whose fused prediction value reaches a prediction value threshold among the plurality of candidate contents is determined as the recommendation result, and the preset recommendation condition is that the fused prediction value reaches the prediction value threshold.

[0197] As can be seen from the above technical solution, when performing content recommendation, the present application can add an interest exploration model to the existing target recommendation model to assist in content recommendation. Specifically, multiple candidate content can be obtained, and recommendation predictions can be performed on the multiple candidate content using the target recommendation model to obtain a recommendation prediction value for each of the multiple candidate content. The interest exploration model can perform interest exploration on each of the multiple candidate content, obtaining a first prediction probability value for each candidate content. The interest exploration model is trained based on multiple sample content of an interactive object and the sample labels corresponding to the multiple sample content. The sample labels are used to indicate the interaction between the interactive object and the corresponding sample content. Therefore, the interest exploration model accurately learns the interests of the interactive object. The first prediction probability value obtained based on the interest exploration model can accurately reflect the probability that the corresponding candidate content is a new interest content for the interactive object. The higher the first prediction probability value of the candidate content, the more likely the candidate content is a new interest of the interactive object. Therefore, for each candidate content, the first prediction probability value and the recommendation prediction value are fused to obtain a fused prediction value. The fused prediction value is positively correlated with the first prediction probability value, thereby adjusting the recommendation prediction value to obtain a fused prediction value, thereby increasing the fused prediction value of the new interest content of the interactive object. In this way, when candidate content whose fused prediction value meets the preset recommendation conditions among multiple candidate content is determined as a recommendation result, candidate content belonging to new interest content can be ranked at the top, helping to explore new interests. Because the first predicted probability value can accurately reflect the probability that the candidate content is the new interest content of the interactive object, the fused prediction value adjusted based on the first predicted probability value can also accurately reflect the possibility that the candidate content is the new interest content of the interactive object. In this way, the first predicted probability value can be used to rank the new interest content of the interactive object at the top, improving the accuracy of interest exploration, thereby increasing the possibility of successful interest exploration and expanding the improvement of interest exploration effect.

[0198] The embodiment further provides a computer device that can execute the content recommendation method. The computer device can be a terminal. Figure 12 The structure diagram of a terminal provided by an embodiment of the present application is shown. Figure 12 In this example, a smartphone is used as the terminal:

[0199] refer to Figure 12The smartphone includes components such as a radio frequency (RF) circuit 1210, a memory 1220, an input unit 1230, a display unit 1240, a sensor 1250, an audio circuit 1260, a wireless fidelity (WiFi) module 1270, a processor 1280, and a power supply 1290. The input unit 1230 may include a touch panel 1231 and other input devices 1232, the display unit 1240 may include a display panel 1241, and the audio circuit 1260 may include a speaker 1261 and a microphone 1262. It is understood that Figure 12 The structure of the smartphone shown in the figure does not constitute a limitation on the smartphone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0200] The memory 1220 can be used to store software programs and modules. The processor 1280 executes the various functional applications and data processing of the smartphone by running the software programs and modules stored in the memory 1220. The memory 1220 may mainly include a program storage area and a data storage area. The program storage area may store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created based on the use of the smartphone (such as audio data, a phone book, etc.). In addition, the memory 1220 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0201] Processor 1280 is the control center of the smartphone, connecting all components of the smartphone using various interfaces and circuits. It executes software programs and / or modules stored in memory 1220 and accesses data stored in memory 1220 to perform various smartphone functions and process data. Optionally, processor 1280 may include one or more processing units. Preferably, processor 1780 integrates an application processor and a modem processor. The application processor primarily handles the operating system, user interface, and application programs, while the modem processor primarily handles wireless communications. It is understood that the modem processor may not be integrated into processor 1280.

[0202] In this embodiment, the processor 1280 in the smartphone can execute the content recommendation method provided in each embodiment of the present application.

[0203] The computer device provided in the embodiment of the present application may also be a server, see Figure 13 As shown, Figure 13The structural diagram of the server 1300 provided in the embodiment of the present application, the server 1300 may have relatively large differences due to different configurations or performances, and may include one or more processors, such as a central processing unit (CPU) 1322, and a memory 1332, one or more storage media 1330 (such as one or more massive storage devices) for storing application programs 1342 or data 1344. Among them, the memory 1332 and the storage medium 1330 can be temporary storage or persistent storage. The program stored in the storage medium 1330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the server. Furthermore, the central processing unit 1322 can be configured to communicate with the storage medium 1330 to execute a series of instruction operations in the storage medium 1330 on the server 1300.

[0204] The server 1300 may also include one or more power supplies 1326, one or more wired or wireless network interfaces 1350, one or more input and output interfaces 1358, and / or one or more operating systems 1341, such as Windows Server 2003. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM etc.

[0205] In this embodiment, the central processing unit 1322 in the server 1300 can execute the content recommendation method provided in each embodiment of the present application.

[0206] According to one aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium is used to store a computer program, and the computer program is used to execute the content recommendation method described in the above embodiments.

[0207] According to one aspect of the present application, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the methods provided in various optional implementations of the above-described embodiments.

[0208] The descriptions of the processes or structures corresponding to the above figures have different emphases. For parts that are not described in detail in a certain process or structure, please refer to the relevant descriptions of other processes or structures.

[0209] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0210] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0211] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0212] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0213] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a terminal, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store computer programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0214] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0215] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, ordinary technical members in this field should understand that they can still modify the technical solutions recorded in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A content recommendation method, characterized in that: The method comprises: Obtain multiple candidate contents; Performing recommendation prediction on the plurality of candidate contents by using a target recommendation model to obtain a recommendation prediction value for each of the plurality of candidate contents; Performing interest exploration on the plurality of candidate contents respectively through an interest exploration model, obtaining a first predicted probability value for each candidate content, the first predicted probability value being used to reflect the probability that the corresponding candidate content is new interest content of the interactive object, the new interest content being content whose satisfaction reaches a satisfaction threshold and to which the classification belongs does not belong to the historical interest classification set of the interactive object, the interest exploration model being trained based on a plurality of sample contents of the interactive object and sample labels corresponding to the plurality of sample contents, the sample labels being used to indicate the interaction between the interactive object and the corresponding sample content; For each candidate content, fusing the first prediction probability value and the recommendation prediction value to obtain a fused prediction value, where the fused prediction value is positively correlated with the first prediction probability value; The candidate content whose fusion prediction value meets the preset recommendation condition among the multiple candidate contents is determined as the recommendation result.

2. The method according to claim 1, characterized in that The fusing the first predicted probability value and the recommended predicted value to obtain a fused predicted value includes: fusing the first predicted probability value and the recommended predicted value using a fusion calculation formula to obtain a fused predicted value; Alternatively, the first prediction probability value and the recommended prediction value are fused using a result fusion layer to obtain a fused prediction value.

3. The method according to claim 2, characterized in that The fusion calculation formula is a power multiplication calculation formula, and the fusion calculation formula is used to fuse the first prediction probability value and the recommended prediction value to obtain a fusion prediction value, including: Using the power multiplication calculation formula, the first predicted probability value is multiplied by the recommended prediction value in a power multiplication manner to obtain the fused prediction value. In the power multiplication calculation formula, the exponent of the first predicted probability value is the weight of the first predicted probability value.

4. The method according to claim 1, wherein The training method of the interest exploration model includes: Obtaining a plurality of sample contents and sample labels corresponding to the plurality of sample contents; Performing interest exploration on the plurality of sample contents respectively through the initial network model to obtain a second predicted probability value for each of the sample contents; For each of the sample contents, determining a degree of difference between the second predicted probability value of the sample content and the sample label of the sample content; Determining a sample training weight for each of the sample contents, wherein the sample training weight of the new interest sample content among the plurality of sample contents is greater than the sample training weight of the historical interest sample content, the new interest sample content being sample content whose satisfaction reaches the satisfaction threshold and whose category does not belong to the historical interest category set, the historical interest category set including the category to which the historical interest sample content belongs; Based on the sample training weight of each of the sample contents, weighted summation is performed on the difference degrees corresponding to the plurality of sample contents to obtain a target loss function; The target loss function is used to optimize the model parameters of the initial network model to obtain the interest exploration model.

5. The method according to claim 4, characterized in that Determining the sample training weight of each sample content includes: Obtaining an initial weight for each of the sample contents; For each of the sample contents, if it is determined that the sample content belongs to the new interest sample content, the initial weight of the sample content is enhanced to obtain a sample training weight of the sample content; If it is determined that the sample content belongs to the historical interest sample content, the initial weight of the sample content is weakened or kept unchanged to obtain a sample training weight of the sample content.

6. The method according to claim 5, characterized in that The step of enhancing the initial weight of the sample content to obtain a sample training weight of the sample content includes: Based on the interactive object's satisfaction with the sample content, the initial weight of the sample content is enhanced to obtain a sample training weight of the sample content. The enhanced sample training weight is positively correlated with the interactive object's satisfaction with the sample content.

7. The method according to claim 5, characterized in that The determining that the sample content belongs to new interest sample content includes: Obtaining the physical duration of the sample content and the duration of the interactive object's stay on the sample content, and obtaining the historical interest classification set, wherein the duration of the interactive object's stay on the sample content is used to reflect the interactive object's satisfaction with the sample content; Performing an average duration prediction based on the physical duration of the sample content to obtain an average dwelling duration of the sample content, wherein the average dwelling duration of the sample content is used to reflect the satisfaction threshold; If the duration of the interaction object's stay on the sample content reaches the average duration, and the category to which the sample content belongs does not exist in the historical interest category set, it is determined that the sample content belongs to new interest sample content.

8. The method according to claim 7, characterized in that The performing of average duration prediction based on the physical duration of the sample content to obtain the average residence time of the sample content includes: The average duration of the physical duration of the sample content is predicted by fitting the obtained satisfaction curve to obtain the average stay duration of the sample content.

9. The method according to claim 8, characterized in that The method further comprises: Performing duration statistics on the plurality of sample contents to obtain a plurality of data points, wherein each of the plurality of data points is used to reflect an average duration of stay under a corresponding physical duration; According to the satisfaction function, curve fitting is performed on the multiple data points to obtain the satisfaction curve, the satisfaction function is a function that calculates the average stay time based on the physical time, and the satisfaction function satisfies at least the following conditions: the values ​​of the physical time and the average stay time in the satisfaction function are respectively greater than zero; the average stay time in the satisfaction function increases monotonically relative to the physical time; when the physical time in the satisfaction function approaches infinity, the average time calculated based on the satisfaction function and the physical time approaches a constant greater than zero.

10. The method according to claim 1, characterized in that The step of determining, as a recommendation result, a candidate content whose fusion prediction value satisfies a preset recommendation condition among the plurality of candidate contents, includes: Sorting the plurality of candidate contents in descending order of the fusion prediction value; determining the candidate contents ranked in the top K as the recommendation results, wherein the preset recommendation condition is that the candidate contents ranked in the top K of the fusion prediction value; Alternatively, a candidate content whose fused prediction value reaches a prediction value threshold among the plurality of candidate contents is determined as the recommendation result, and the preset recommendation condition is that the fused prediction value reaches the prediction value threshold.

11. A content recommendation device, characterized in that: The device includes an acquisition unit, a recommendation prediction unit, an interest exploration unit, a fusion unit, and a determination unit: The acquisition unit is configured to acquire a plurality of candidate contents; The recommendation prediction unit is configured to perform recommendation prediction on the plurality of candidate contents using a target recommendation model to obtain a recommendation prediction value for each of the plurality of candidate contents; The interest exploration unit is configured to perform interest exploration on each of the plurality of candidate contents through an interest exploration model to obtain a first predicted probability value for each of the candidate contents, wherein the first predicted probability value is used to reflect the probability that the corresponding candidate content is a new interest content of the interactive object, wherein the new interest content is content whose satisfaction reaches a satisfaction threshold and whose classification does not belong to the historical interest classification set of the interactive object, and the interest exploration model is trained based on a plurality of sample contents of the interactive object and sample labels corresponding to the plurality of sample contents, respectively, and wherein the sample labels are used to indicate the interaction between the interactive object and the corresponding sample content; The fusion unit is configured to fuse the first prediction probability value and the recommendation prediction value for each candidate content to obtain a fused prediction value, where the fused prediction value is positively correlated with the first prediction probability value; The determining unit is configured to determine, among the plurality of candidate contents, a candidate content whose fusion prediction value satisfies a preset recommendation condition as a recommendation result.

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 configured to execute the method according to any one of claims 1 to 10 according to 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, which, when executed by a processor, causes the processor to perform the method according to any one of claims 1 to 10.

14. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 10 when executed by a processor.

Images