Interest portrait generation method, server and storage medium based on knowledge graph

Abstract

The embodiments of the present invention relate to the technical field of big data mining, and specifically provide a method, server and storage medium for generating interest portraits based on knowledge graphs. The embodiments of the present invention can accurately capture the sequential logic of users browsing images within a target session cycle by obtaining a first page interactive image set sorted based on page response priority. Knowledge vector sets are obtained by performing knowledge extraction based on image content descriptions, and the rich information of the image can be converted into quantifiable and analyzable data forms. The extraction of interest knowledge entities with multiple feature depths can mine user interests from different levels and comprehensively reveal user interest entities and relationships. Finally, a knowledge feature enhancement strategy is used to generate an interest portrait, which accurately presents the user's interest features in the process of browsing images, helps to realize personalized recommendations, optimize search result sorting and other applications, improve user experience and improve the accuracy and effectiveness of platform content push.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of big data mining technology, and specifically, to a method, server and storage medium for generating interest portraits based on a knowledge graph. Background Art

[0002] In today's network environment, users browse a large number of page interaction images. However, traditional user interest analysis methods have many shortcomings. On the one hand, when collecting image data browsed by users, the page response priority is not considered, and the order of users' attention to different images cannot be accurately reflected, resulting in biased analysis results. On the other hand, the analysis of images only stays on the surface, and it is impossible to deeply explore the different levels of user interests behind the image content, and lacks a multi-dimensional analysis of interest entities and relationships. In addition, there is a lack of effective strategies to strengthen and integrate knowledge features related to user interests to generate accurate interest portraits, which makes it ineffective in applications such as personalized recommendations and search result optimization, and cannot accurately meet user needs, resulting in poor user experience. Summary of the invention

[0003] The embodiments of the present invention at least provide a method, a server and a storage medium for generating interest portraits based on a knowledge graph.

[0004] An embodiment of the present invention provides a method for generating an interest portrait based on a knowledge graph, which is applied to an interest portrait generation server, and the method includes: obtaining a first page interaction image set, wherein the first page interaction image set is obtained by sequentially arranging multiple page interaction images browsed by a first page subscriber within a target session period based on page response priority; performing knowledge extraction on the first page interaction image set according to the image content description of each page interaction image in the first page interaction image set to obtain a first interaction preference knowledge vector set; performing subscription interest knowledge entity extraction at multiple feature depths on the first interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the first page subscriber; wherein one feature depth corresponds to one subscription interest knowledge entity set, and each subscription interest knowledge entity set is used to characterize the interest entities and interest relationships of the first page interaction image set at the corresponding feature depth; performing interest portrait generation processing on the multiple subscription interest knowledge entity sets of the first page subscriber using a knowledge feature enhancement strategy to obtain a page subscription interest portrait, and the page subscription interest portrait is used to characterize the interest features of the first page subscriber in the process of browsing page interaction images.

[0005] An embodiment of the present invention also provides an interest portrait generation server, including a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the above method.

[0006] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. The computer program implements the above method when running.

[0007] The technical solution provided by the embodiment of the embodiment of the present invention may include the following beneficial effects: the embodiment of the present invention can accurately capture the sequential logic of users browsing images within the target session cycle by obtaining the first page interactive image set sorted based on the page response priority, which reflects the potential difference in users' attention to different page images. Knowledge extraction based on the image content description obtains a knowledge vector set, which can convert the rich information of the image into a quantifiable and analyzable data form. The interest knowledge entity extraction of multiple feature depths can mine user interests from different levels and comprehensively reveal user interest entities and relationships, covering both surface preferences and deep associations. Finally, the knowledge feature enhancement strategy is used to generate an interest portrait, which accurately presents the user's interest characteristics in the process of browsing images, which helps to realize personalized recommendations, optimize search result sorting and other applications, improve user experience and improve the accuracy and effectiveness of platform content push.

[0008] For a description of the effects of the above-mentioned interest portrait generation server and computer-readable storage medium, please refer to the description of the above-mentioned method.

[0009] In order to make the above-mentioned purposes, features and advantages of the embodiments of the present invention more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the embodiments of the present invention and are used together with the specification to illustrate the technical solutions of the embodiments of the present invention. It should be understood that the following drawings only illustrate certain embodiments of the embodiments of the present invention and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 It is a block diagram of an interest portrait generation server shown in an embodiment of the present invention.

[0012] Figure 2 It is a flow chart of a method for generating interest portraits based on a knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0013] Here, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the embodiments of the present invention. Instead, they are only examples of devices and methods consistent with some aspects of the embodiments of the present invention.

[0014] Figure 1 The structural diagram of the interest portrait generation server 10 provided in the embodiment of the present invention includes a processor 102, a memory 104, and a bus 106. Among them, the memory 104 is used to store execution instructions, including internal memory and external memory. The internal memory can also be understood as an internal memory, which is used to temporarily store the calculation data in the processor 102 and the data exchanged with the external memory such as the hard disk. The processor 102 exchanges data with the external memory through the internal memory. When the interest portrait generation server 10 is running, the processor 102 communicates with the memory 104 through the bus 106, so that the processor 102 executes the interest portrait generation method based on the knowledge graph of the embodiment of the present invention.

[0015] Please combine Figure 2 , Figure 2 It is a flow chart of a method for generating an interest portrait based on a knowledge graph provided in an embodiment of the present invention, which is applied to an interest portrait generation server. The method may exemplarily include the following S110-S170.

[0016] S110: Acquire a first page interaction image set, where the first page interaction image set is obtained by arranging multiple page interaction images browsed by a first page subscriber within a target session period in sequence based on page response priorities.

[0017] S130: performing knowledge extraction on the first page interaction image set according to the image content description of each page interaction image in the first page interaction image set to obtain a first interaction preference knowledge vector set.

[0018] S150: performing multiple feature depth subscription interest knowledge entity extraction on the first interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the first page subscribers.

[0019] Among them, one of the feature depths corresponds to one of the subscription interest knowledge entity sets, and each of the subscription interest knowledge entity sets is used to characterize the interest entities and interest relationships of the first page interaction image set at the corresponding feature depth.

[0020] S170: Using the knowledge feature enhancement strategy, interest portrait generation processing is performed on multiple subscription interest knowledge entity sets of the first page subscriber to obtain a page subscription interest portrait, and the page subscription interest portrait is used to characterize the interest characteristics of the first page subscriber in the process of browsing the page interaction image.

[0021] In the present application, an exemplary introduction of the interest portrait generation server when executing the technical solutions recorded in the above S110-S170 is as follows.

[0022] First, in S110, the interest portrait generation server obtains the first page interaction image set. The first page interaction image set in the present application is obtained by sequentially sorting multiple page interaction images browsed by the first page subscriber within the target session period based on the page response priority. Page response priority is an important concept, which can be determined based on multiple factors. For example, in an image browsing platform, if the page where some page interaction images are located loads faster, or these images are in a more prominent position in the page layout (such as home page recommendations, popular sections, etc.), then these images may have a higher page response priority. For example, within the target session period (for example, set to one hour), the first page subscriber browses 10 page interaction images, of which the three images recommended on the home page may be sorted first due to their higher priority, and then the remaining images are sorted in order according to the order in other page layouts, thereby obtaining the first page interaction image set. This sorting method based on page response priority helps to more reasonably reflect the potential order of attention of users to different images during browsing.

[0023] Then, enter S130, the interest portrait generation server performs knowledge extraction on the first page interactive image set according to the image content description of each page interactive image in the first page interactive image set, and obtains a first interactive preference knowledge vector set. The image content description contains rich information, such as the object type, color, scene, etc. in the image. Taking a page interactive image set containing landscape images as an example, the image content description may include information such as mountains (specifically, the name of a well-known mountain range), blue lakes, green trees, and clear skies in the image. The interest portrait generation server processes these content descriptions through a specific algorithm and converts them into knowledge elements. For example, the entity recognition algorithm in the natural language processing technology can be used to identify these object entities, and then these entities are mapped to corresponding knowledge concepts according to the predefined knowledge system, such as mountains belong to natural geographical entities, blue lakes belong to water landscapes, etc. These knowledge elements are combined to form a knowledge vector. For example, for each page interactive image, 10 knowledge elements can be extracted, and after sorting, a 10-dimensional first interactive preference knowledge vector set is obtained, and each vector represents the knowledge extraction result of a page interactive image.

[0024] Subsequently, in S150, the interest portrait generation server extracts subscription interest knowledge entities at multiple feature depths for the first interaction preference knowledge vector set, and obtains multiple subscription interest knowledge entity sets of the first page subscribers. Here, one feature depth corresponds to one subscription interest knowledge entity set, and each subscription interest knowledge entity set is used to characterize the interest entities and interest relationships of the first page interactive image set at the corresponding feature depth. Feature depth can be understood as different levels of mining of user interests. For example, at a shallow feature depth, it may simply extract the direct object entities in the image as interest entities, such as the mountains and lakes in the above-mentioned landscape images. At a deeper feature depth, the relationship between objects and deeper associations based on the knowledge graph may be mined. For example, the geographical proximity between mountains and lakes, and the ecosystem represented by this proximity (such as water from mountains flowing to lakes, nourishing the surrounding ecology). For example, if there are three feature depths, at the first feature depth, the extracted subscription interest knowledge entity set may mainly be obvious objects in the image; at the second feature depth, in addition to objects, it also includes the relationship between them; at the third feature depth, it includes complex relationships and related entities based on a broader knowledge graph.

[0025] Finally, in S170, the interest portrait generation server uses the knowledge feature reinforcement strategy to generate interest portraits for multiple subscription interest knowledge entity sets of the first page subscriber to obtain a page subscription interest portrait. The knowledge feature reinforcement strategy aims to highlight those knowledge features that have important representational significance for user interests. For example, a weight-based reinforcement method can be used to give higher weights to those knowledge entities that frequently appear in multiple page interactive images and have important associations at different feature depths. Taking an art image browsing scene as an example, if the works of a well-known painter (such as Van Gogh) appear in multiple images, and at different feature depths, Van Gogh's works are closely related to specific painting styles (such as Impressionism), color applications (such as bright contrasting colors), and the social and cultural background of the creation period, then in the interest portrait, Van Gogh and these related knowledge entities will be strengthened and have a higher weight. In this way, multiple subscription interest knowledge entity sets are combined to construct a page subscription interest portrait that fully reflects the interest characteristics of the first page subscriber in the process of browsing the page interactive image. This portrait can be used in a variety of application scenarios, such as personalized image recommendation, recommending images of artworks that users may be interested in based on their interest portraits; or in the sorting of image search results, placing images that highly match the user's interest portrait at a higher position, thereby improving the user's search experience and satisfaction with the platform.

[0026] During the entire process, the interest portrait generation server starts with obtaining the first page interaction image set sorted according to the page response priority, gradually extracts the knowledge of image content description and interest knowledge entities with different feature depths, and finally uses the knowledge feature reinforcement strategy to generate page subscription interest portraits, forming a complete and coherent technical solution, which provides an effective technical means for accurately grasping the interest characteristics of users when browsing page interaction images.

[0027] In actual application, in S110, the setting of the target session period is flexible and reasonable. For example, for a news image browsing platform, the target session period may be shorter, such as 15-30 minutes, because news images are time-sensitive, and the images browsed by users in a shorter time can better reflect their immediate interests. For an artwork image browsing platform, the target session period can be longer, such as 2-3 hours, because users may spend more time browsing in depth when appreciating artworks, and a longer session period can more comprehensively cover the user's browsing behavior. In addition, the determination of page response priority depends not only on the page layout and loading speed, but may also be related to the business logic of the platform. For example, in the image browsing section of an e-commerce platform, if certain product images are related to ongoing promotional activities, then these images may be given a higher page response priority, even if they are not recommended on the homepage or the loading speed is not the fastest.

[0028] In S130, the processing method of image content description is also relatively complex. In addition to the entity recognition algorithm in the natural language processing technology mentioned above, the image classification algorithm can also be involved. For example, for some complex images, the image classification algorithm can be used to classify the image into a broad category (such as landscape, character, architecture, etc.), and then more detailed entity recognition and knowledge element extraction can be performed based on this category. Moreover, the construction of a knowledge system is crucial for converting image content description into knowledge elements. This knowledge system can be constructed based on industry standards, expert knowledge, and a large amount of data statistical analysis. For example, in the knowledge system of architectural images, it not only includes the types of buildings (such as classical buildings, modern buildings), but also includes architectural styles (such as Gothic, Baroque), architectural functions (such as residential, church) and other multi-dimensional knowledge concepts, so that the architectural information in the image can be accurately extracted into corresponding knowledge elements, thereby forming an accurate first interactive preference knowledge vector set.

[0029] In S150, the division of feature depth needs to be based on specific application scenarios and the requirements for the level of detail of user interest mining. In addition to the three feature depth examples mentioned above, more feature depths may be set in certain specific scenarios. For example, in a medical image browsing platform, for medical researchers, 5-6 feature depths may need to be set to deeply explore their interests in different disease images, treatment methods images, etc. At a shallow feature depth, it may only identify the organ or tissue type in the image; at a medium depth, it will involve the relationship between the appearance characteristics of the disease and the organ tissue; at a deeper feature depth, the cause of the disease, the treatment plan and the complex relationship between various elements in the image will be mined. The construction of the subscription interest knowledge entity set at each feature depth needs to be combined with a professional medical knowledge graph to ensure that the extracted interest entities and interest relationships are accurate.

[0030] In S170, in addition to the weight-based method, the knowledge feature reinforcement strategy can also adopt a cluster analysis-based method. For example, multiple subscription interest knowledge entity sets of the first page subscriber are clustered, similar knowledge entities are clustered into one category, and then the clustering results are used to determine which categories of knowledge entities are more important in the user's interest portrait. Taking tourism image browsing as an example, if the clustering results show that users have obvious clustering of the three categories of knowledge entities, namely beach tourism, mountain tourism, and historical and cultural city tourism, then the knowledge features of these categories can be strengthened according to the proportion and degree of association of these three categories in the user's browsing images, thereby constructing a page subscription interest portrait that is more in line with the user's actual interests. At the same time, in order to improve the accuracy of the interest portrait, a time decay factor can also be introduced. For example, if certain knowledge entities appear in earlier browsing records but rarely appear in recent browsing, the weights of these knowledge entities will gradually decrease over time to better reflect the dynamic changes in user interests.

[0031] Through this series of operations, the interest portrait generation server establishes a close connection between processing page interaction images and user interest features, providing a solid technical foundation for various applications based on user interests.

[0032] In some other exemplary application scenarios, the association and synergy between the various steps can be explored in depth. The first page interactive image set obtained in S110 provides basic data for subsequent steps. Its sorting method (based on page response priority) affects the subsequent analysis direction of user interests. For example, if some high-priority images are about science and technology topics during the sorting process, and users stay for a long time when browsing these images, then the subsequent knowledge extraction and interest portrait generation process will be more centered around science and technology-related knowledge elements and interest entities.

[0033] The knowledge extraction in S130 is the key bridge between image content and user interests. By converting the image content description into the first interactive preference knowledge vector set, various elements in the image can be presented in a form more suitable for computer processing and analysis. This knowledge vector set not only contains the basic information of the image, but also incorporates semantic information based on the knowledge system, providing rich materials for further mining the user's subscription interest knowledge entities.

[0034] S150 conducts in-depth mining based on S130. The setting of multiple feature depths makes the analysis of user interests more comprehensive and detailed. Subscription interest knowledge entity sets at different feature depths complement each other and depict user interest characteristics from different levels and angles. For example, entity sets with shallower feature depths may reflect the user's superficial interests, while entity sets with deeper feature depths may reveal the user's potential interests related to a broader knowledge background. These different levels of interest entity sets together construct a three-dimensional user interest model.

[0035] Finally, S170 integrates the multiple subscription interest knowledge entity sets obtained in the previous steps and generates the final page subscription interest portrait through the knowledge feature enhancement strategy. In this process, the knowledge feature enhancement strategy comprehensively considers factors such as the importance and relevance of each subscription interest knowledge entity set and the dynamic changes of user interests. For example, through weight adjustment and cluster analysis, those knowledge entities that have key representational significance for user interests are highlighted, while those interest elements that may appear by chance or are outdated are weakened, so as to obtain an accurate, dynamic and comprehensive page subscription interest portrait, providing users with more personalized and accurate services, such as personalized image recommendation, search result optimization, etc.

[0036] In short, when the interest portrait generation server executes this application, each step works closely together to gradually build a page subscription interest portrait that can accurately reflect the user's interest characteristics from the original page interaction image data, which has important application value in image browsing, recommendation, search and other fields.

[0037] In actual application scenarios, this application has strong adaptability and scalability. Taking the social media platform as an example, when a user browses dynamics containing pictures, the interest portrait generation server can generate a unique page subscription interest portrait for each user according to the above steps. For those dynamics that contain a large number of picture sharing, such as travel photography sharing, food picture sharing, etc., the server can accurately capture the user's interest in different types of pictures. For example, a user may frequently browse and like pictures of exotic buildings over a period of time. Through this technical solution, the server can accurately reflect the user's interest in exotic architectural culture in the interest portrait, and then push more relevant tourist destination pictures, architectural culture introduction pictures, etc. to the user in subsequent recommendations.

[0038] Taking the browsing of image resources in the online education platform as an example, there may be various subject-related image materials on the platform, such as scientific experiment pictures, historical relics pictures, etc. The interest portrait generation server can generate accurate interest portraits based on the students' browsing behavior of these images within the target session cycle. If a student often browses cell structure images in biology and shows concern about the relationship between cell function and related diseases at different feature depths, then in the generated interest portrait, the cell research-related knowledge entities under the biology subject will be highlighted. This helps the online education platform provide students with more personalized learning resource recommendations, such as pushing more pictures of cutting-edge results in cell disease research or related online course picture introductions.

[0039] From the perspective of data processing and algorithm optimization, the present application also has certain advantages. In S130, in order to improve the efficiency and accuracy of knowledge extraction, a distributed computing framework can be used to process a large number of page interactive image content descriptions. For example, using the ApacheSpark framework, the data of the image content description is distributed to multiple computing nodes for parallel processing, which can greatly shorten the time for knowledge extraction. At the same time, in terms of algorithm models, pre-trained models in deep learning can be used to assist in the processing of image content descriptions. For example, a pre-trained convolutional neural network model (such as ResNet) is used for image classification, and then combined with a pre-trained model in natural language processing (such as BERT) for entity recognition and knowledge element mapping. In this way, the learning results of the pre-trained model on large-scale data can be used to improve the accuracy of image content description processing.

[0040] In S150, in order to better extract subscription interest knowledge entities with deep features, graph database technology can be used to store and manage knowledge graphs. For example, using the Neo4j graph database, each subscription interest knowledge entity set is stored in the form of a graph, which is convenient for querying and mining the relationship between entities at different feature depths. In addition, an algorithm based on a graph convolutional neural network (GCN) can be used to analyze the knowledge graph in the graph database to mine deeper interest relationships.

[0041] In S170, in order to optimize the knowledge feature reinforcement strategy, a reinforcement learning algorithm may be used. For example, the accuracy of the interest portrait (e.g., measured by the user's feedback on the recommended image) is used as a reward signal, and the weight and other parameters in the knowledge feature reinforcement strategy are continuously adjusted through the reinforcement learning algorithm, so that the generated interest portrait can better adapt to the actual interest changes of the user.

[0042] In summary, the embodiment of the present invention can accurately capture the sequential logic of users browsing images within the target session cycle by obtaining the first page interactive image set sorted based on the page response priority, which reflects the potential difference in users' attention to different page images. Knowledge extraction based on the image content description obtains a knowledge vector set, which can convert the rich information of the image into a quantifiable and analyzable data form. The extraction of interest knowledge entities with multiple feature depths can explore user interests from different levels and comprehensively reveal user interest entities and relationships, covering both surface preferences and deep associations. Finally, the knowledge feature enhancement strategy is used to generate an interest portrait, which accurately presents the user's interest characteristics in the process of browsing images, which helps to realize personalized recommendations, optimize search result sorting and other applications, improve user experience and improve the accuracy and effectiveness of platform content push.

[0043] In some other optional application scenarios, on a digital office platform, users can browse various office-related images, such as project flow charts, office software interface diagrams, business meeting scene diagrams, etc.

[0044] As for the first interactive preference knowledge vector set, for a project flowchart image, 12 knowledge elements are set to be used to describe it. For example: [Number of departments involved: 3, number of process steps: 8, main color: blue (quantized as 1), whether there is a decision node: yes (1), whether external cooperation is involved: no (0), software tool identification used: Visio (1), whether it is a key process: yes (1), whether there is a timeline: no (0), document relevance: high (3), personnel role identification: project manager (1), whether there is a risk warning: yes (1), process type: linear (1)]. After such knowledge extraction is performed on multiple office-related images browsed by the user during the target session cycle, the first interactive preference knowledge vector set is formed.

[0045] At the first feature depth, the set of subscription interest knowledge entities that may be extracted is {project process: 5 (involved in 5 images), Visio software: 3, business meeting: 2, key process: 4}. Here we mainly focus on direct office elements in the image, such as office process type and related software entities.

[0046] At the second feature depth, in addition to the above entities, the relationships between entities are also included, such as {Project Process-Visio Software-Process Design: 3 (indicates that there are 3 images reflecting this relationship), Business Meeting-Personnel Role-Decision Making: 2}. This subscription interest knowledge entity set begins to reflect a deeper level of relationship.

[0047] At the third feature depth, based on a broader office knowledge graph, there may be {project process-key process-company strategy-goal achievement: 1, business meeting-decision making-resource allocation-project progress: 1}, etc. This set reflects more complex relationships and associated entities.

[0048] The knowledge entity sets of subscription interests at different feature depths are integrated, and the knowledge feature reinforcement strategy is adopted to assign weights according to factors such as the frequency of entity occurrence and the closeness of association. For example, the project process and its related knowledge entities have high weights due to their frequent appearance and close association at multiple feature depths, such as project process: 0.7, Visio software: 0.5, while some relatively isolated or less frequent entities have lower weights, such as company strategy: 0.2. In this way, a page subscription interest portrait is obtained, such as [project process: 0.7, Visio software: 0.5, business meeting: 0.3, key process: 0.6, process design: 0.4, decision making: 0.3, company strategy: 0.2, project progress: 0.1]. This portrait comprehensively represents the interest characteristics of the first page subscriber in the process of browsing the interactive image of the digital office page.

[0049] On an online education platform, users can browse images related to various subjects, such as diagrams of physics experimental devices, portraits of historical figures, and images of mathematical functions.

[0050] Taking a physics experiment device diagram as an example, 10 knowledge elements are set to describe it: [Subject: Physics (1), Experiment Type: Electricity (1), Number of Device Components: 5, Is it a Classic Experiment: Yes (1), Color Identification: Red (1), Is there a Danger Warning: No (0), Associated Scientist: Ohm (1), Experiment Difficulty: Medium (2), Data Display Method: Instrument (1), Is it a Demonstration Experiment: Yes (1)]. After such knowledge extraction is performed on multiple education-related images browsed by users, a first interactive preference knowledge vector set is formed.

[0051] At the first feature depth, an exemplary set of subscription interest knowledge entities is {physics: 4 (involved in 4 images), history: 2, electrical experiments: 3, classical experiments: 3, Newton (historical figure): 1}. Here we mainly focus on entities such as direct disciplines and experimental types in the image.

[0052] At the second feature depth, in addition to the above entities, the relationships between entities are also included, such as {Physics-Electrical Experiment-Ohm's Law: 2 (indicates that there are 2 images reflecting this relationship), History-Newton-Universal Gravitation: 1}. This subscription interest knowledge entity set begins to reflect a deeper level of relationship.

[0053] At the third characteristic depth, based on a broader educational knowledge graph, there may be {Physics-Electrical Experiment-Ohm-Circuit Theory Development: 1, History-Newton-Scientific Changes-Knowledge System Changes: 1}, etc. This set embodies more complex relationships and associated entities.

[0054] Comprehensive subscription interest knowledge entity sets at different feature depths, adopt knowledge feature reinforcement strategy, and assign weights according to factors such as the frequency of entity occurrence and the closeness of association. For example, physics and related knowledge entities have higher weights due to their frequent appearance and close association at multiple feature depths, such as physics: 0.8, electrical experiments: 0.6, while some relatively isolated or less frequent entities have lower weights, such as scientific changes: 0.2. In this way, a page subscription interest portrait is obtained, such as [physics: 0.8, history: 0.3, electrical experiments: 0.6, classical experiments: 0.5, Newton: 0.4, Ohm's law: 0.5, circuit theory development: 0.3, scientific changes: 0.2, knowledge system changes: 0.1]. This portrait comprehensively represents the interest characteristics of the first page subscriber in the process of browsing the interactive images of the online education page.

[0055] To facilitate the consistency of the overall solution, the description of the subsequent embodiments will continue with the application scenarios of digital office and online education as examples.

[0056] In some optional embodiments, the number of feature depths is Q, and any one of the Q feature depths is represented as a p-th level feature depth. The subscription interest knowledge entity extraction includes knowledge entity mining and entity relationship construction, where p is a positive integer and p is not greater than Q; the subscription interest knowledge entity extraction of multiple feature depths on the first interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the first page subscriber includes: in response to p being equal to 1, setting the first interaction preference knowledge vector set to the first level subscription interest knowledge entity set corresponding to the first level feature depth; in response to p being less than Q, iterating p to p+1, performing knowledge entity mining of the p+1-th level feature depth on the first interaction preference knowledge vector set to obtain the p+1-th level interest entity set, where p+1 is an integer greater than 1 and not greater than Q; performing entity relationship construction on the p+1-th level interest entity set to obtain the p+1-th level subscription interest knowledge entity set corresponding to the p+1-th level feature depth; in response to p being equal to Q, obtaining Q subscription interest knowledge entity sets corresponding to the Q feature depths respectively.

[0057] In the next step, the feature depth is used to indicate the number of entities divided into the same feature cluster during the knowledge entity mining process of the corresponding feature depth, and the p-th level feature depth is used to indicate that the number of entities divided into the same feature cluster during the knowledge entity mining process of the p-th level feature depth is p; the knowledge entity mining of the first interaction preference knowledge vector set at the p+1-th level feature depth is performed to obtain the p+1-th level interest entity set, including: based on the indication information of the p+1-th level feature depth, the first interaction preference knowledge vector set is summarized into knowledge entities through the influence weights of each entity in the first interaction preference knowledge vector set to obtain the p+1-th level interest entity set; wherein the p+1-th level interest entity set contains at least one feature cluster, and each feature cluster contains p+1 entities with a weighted order relationship.

[0058] Furthermore, the entity relationship construction of the p+1th level interest entity set to obtain the p+1th level subscription interest knowledge entity set corresponding to the p+1th level feature depth includes: associating knowledge features of entities in each feature cluster in the p+1th level interest entity set to obtain the p+1th level associated feature entity set; the p+1th level associated feature entity set includes at least one associated feature entity combination, each associated feature entity combination is obtained by associating entities in a feature cluster in the p+1th level interest entity set; knowledge extraction is performed on the associated feature entity combinations in the p+1th level associated feature entity set to obtain the p+1th level subscription interest knowledge entity set corresponding to the p+1th level feature depth.

[0059] In the digital office scenario, we first clarify that the number of feature depths is Q, and each feature depth here has its specific meaning. When refining subscription interest knowledge entities, it involves knowledge entity mining and entity relationship construction.

[0060] When p is equal to 1, the first interaction preference knowledge vector set is directly set to the first-level subscription interest knowledge entity set corresponding to the first-level feature depth. For example, in the knowledge vector set related to the project flowchart image mentioned in the previous digital office scenario, if this setting is followed, the initial entities directly related to office elements constitute the first-level subscription interest knowledge entity set, such as {project process: 5 (involved in 5 images), Visio software: 3, business meeting: 2, key process: 4}, which is a set of elements directly extracted from the knowledge vector set, is the first-level case.

[0061] When p is less than Q, an iteration operation will be performed. For example, when p=1, p should be iterated to p+1, i.e., 2, and the knowledge entity mining of the first interactive preference knowledge vector set at the second level of feature depth is performed. In this process, based on the indication information of the second level of feature depth, the knowledge entity of the first interactive preference knowledge vector set is summarized by the influence weight of each entity in the first interactive preference knowledge vector set. Taking the digital office scene as an example, if entities such as project process and Visio software have been determined at the first level of feature depth, the influence weight of each entity should be considered at the second level of feature depth. For example, the weight of the project process is high, and the weight of Visio software is relatively low. By summarizing these entities according to these weights, entities that are closely related to the project process and have high weights may be summarized together to form feature clusters, and the second level of interest entity set is obtained. Here, the second level of interest entity set contains at least one feature cluster, and each feature cluster contains two entities with a weight sequence relationship. For example, the feature cluster {project process-Visio software} is one of the possible situations.

[0062] Then, the entity relationship of the second-level interest entity set is constructed. First, the entities in each feature cluster in the second-level interest entity set are associated with knowledge features to obtain the second-level associated feature entity set. For example, for the feature cluster {project process-Visio software}, through association, a second-level associated feature entity set with a clearer relationship such as {project process-use-Visio software} may be obtained. This set includes at least one associated feature entity combination. Then, the associated feature entity combination in the second-level associated feature entity set is refined to obtain the second-level subscription interest knowledge entity set corresponding to the second-level feature depth.

[0063] This iteration is repeated until p is equal to Q, and then Q subscription interest knowledge entity sets corresponding to the Q feature depths are obtained.

[0064] The same process is followed in the online education scenario. When p is equal to 1, for the knowledge vector set related to the physical experiment device diagram mentioned above, a knowledge vector set such as {Subject: Physics (1), Experiment Type: Electricity (1), Number of Device Components: 5, Is it a Classical Experiment: Yes (1), Color Identification: Red (1), Is there a Danger Warning: No (0), Associated Scientist: Ohm (1), Experiment Difficulty: Medium (2), Data Display Method: Instrument (1), Is it a Demonstration Experiment: Yes (1)} constitutes the first-level subscription interest knowledge entity set, for example, {Physics: 4 (involved in 4 images), History: 2, Electrical Experiments: 3, Classical Experiments: 3, Newton (historical figure): 1}.

[0065] When p is less than Q, iterate. For example, in the process from p=1 to p=2, take physics-related entities as an example, and summarize physics-related entities according to the influence weight of the entities. For example, the weight of physics subject entities is higher, and the weight of electrical experiments is second, which may form a feature cluster such as {physics-electrical experiments} to obtain the second-level interest entity set. Then construct entity relationships, associate the entities in the feature cluster with knowledge features, and obtain the second-level associated feature entity set such as {physics-inclusion-electrical experiments}, and then perform knowledge extraction to obtain the second-level subscription interest knowledge entity set corresponding to the second-level feature depth.

[0066] The feature depth is used to indicate the concept of the number of entities that are classified into the same feature cluster during the knowledge entity mining process of the corresponding feature depth. In the digital office scenario, if the feature depth is level 2, then the number of entities that are classified into the same feature cluster during the knowledge entity mining process is 2. For example, the feature cluster {Project Process-Visio Software} is 2 entities. In the online education scenario, if the feature depth is level 3, then the number of entities in the feature cluster formed during the mining process is 3. For example, a feature cluster such as {Physics-Electrical Experiment-Ohm} meets the requirements of level 3 feature depth.

[0067] In this way, in both scenarios, multiple subscription interest knowledge entity sets can be accurately constructed step by step from the first interaction preference knowledge vector set, and then these subscription interest knowledge entity sets are processed using the knowledge feature enhancement strategy to obtain the page subscription interest portrait.

[0068] In digital office scenarios, interest portraits constructed through such detailed feature depth division and knowledge entity extraction can more accurately reflect users' interests in office-related image content. For example, for users who often browse images related to project processes and have multiple associations with project processes at different feature depths, the weights of project processes and their related entities will be accurately reflected in the interest portraits, which will help provide users with more accurate office resource recommendations on the digital office platform, such as recommending specific project process management tools or related project experience sharing documents. At the same time, in terms of office process optimization, if it is known that users have a deep interest in certain process elements, targeted process improvement suggestions can also be made.

[0069] In the online education scenario, this technical solution can deeply explore students' interests in knowledge of different disciplines. Taking physics as an example, if students often browse images related to electrical experiments and show a variety of related interests in electrical experiments, physics disciplines, scientists, etc. at different feature depths, then the weights of related knowledge entities such as physics and electrical experiments in the interest portrait will be reasonably reflected. This helps online education platforms provide students with more personalized learning resource recommendations, such as recommending extended learning materials for electrical experiments or introductions to the research results of relevant scientists, etc., to improve students' learning outcomes.

[0070] With this design, through clear feature depth division and corresponding knowledge entity mining and entity relationship construction methods, users' interests in page interactive images can be deeply mined in digital office and online education scenarios. The subscription interest knowledge entity set is accurately constructed, and a more accurate page subscription interest portrait is generated. This helps to provide users with more personalized resource recommendations in different scenarios, improve user experience, and provide valuable reference for office and education-related business optimization, thereby improving office efficiency and education quality.

[0071] In an alternative technical solution, the first page interactive image set is subjected to knowledge extraction based on the image content description of each page interactive image in the first page interactive image set to obtain a first interactive preference knowledge vector set, including: obtaining a category relationship vector between X content category tags and a plurality of image descriptors for each of the content category tags; the X content category tags include: one or more of content push annotations, page interactive events, subscription feedback comments, community graphs between page subscribers, content update nodes, and content heat values; determining a global image content description based on the category relationship vectors between the X content category tags and the plurality of image descriptors for each of the content category tags; any image content description in the global image content description is obtained by splicing any image descriptor included in each of the content category tags based on the category relationship vectors between the X content category tags; generating interactive preference quantified knowledge of each image content description in the global image content description based on the number of the global image content descriptions; and converting the image content description of each page interactive image in the first page interactive image set into interactive preference quantified knowledge of the corresponding image content description to obtain a first interactive preference knowledge vector set.

[0072] In a digital office scenario, the process of performing knowledge extraction based on the image content description of each page interaction image in the first page interaction image set to obtain a first interaction preference knowledge vector set is as follows.

[0073] First, obtain the category relationship vectors between X content category labels and multiple image descriptors for each content category label. In the digital office scenario, taking content push annotation as an example, content push annotation may include different types of annotations such as office document push and office software update push, and there are certain category relationship vectors between these different types. For each content category label, there are multiple image descriptors. For example, under the content category label of page interaction event, the image descriptor may include the image of the mouse clicking on the office document, the image of the keyboard input information to the office software interface, etc.

[0074] The X content category tags include one or more of content push annotations, page interaction events, subscription feedback comments, community graphs between page subscribers, content update nodes, and content heat value. For example, the content heat value reflects the popularity of office image content on the platform, which may be determined based on factors such as the number of views and the number of likes. The community graph between page subscribers describes the association between office users, such as which users are in the same department or project group, and their interaction behaviors with office images may be similar.

[0075] The global image content description is determined based on the category relationship vectors between X content category tags and multiple image descriptors of each content category tag. For example, in a digital office scenario, for a project flowchart image, the construction of its global image content description may be based on project process related push in content push annotation, click operation on flowchart nodes in page interaction events, comments on process optimization in subscription feedback comments, information related to project team members in the community map between page subscribers, information about process version updates in content update nodes, and the popularity of the flowchart image in content heat value, etc. Any image content description in the global image content description is obtained by splicing any image descriptor included in each content category tag based on the category relationship vector between X content category tags. For example, the specific push descriptor in the content push annotation related to the project flowchart image, the click coordinate descriptor in the page interaction event, etc. are spliced ​​according to the category relationship vector.

[0076] Then, based on the number of global image content descriptions, the interaction preference quantification knowledge of each image content description in the global image content description is generated. For example, if the number of global image content descriptions is 10, each image content description is converted into interaction preference quantification knowledge according to certain quantification rules. For example, in the global image content description of an office document image, the content push annotation-related part is quantified to a higher value, such as 8, due to frequent pushes and certain user interactions; while in the page interaction event, if there is only an occasional mouse hover operation, it is quantified to a lower value, such as 2.

[0077] Finally, the image content descriptions of each page interactive image in the first page interactive image set are converted into the interactive preference quantitative knowledge of the corresponding image content descriptions to obtain the first interactive preference knowledge vector set. In the digital office scenario, for multiple office images browsed by users, such as business meeting scene images, office software interface images, etc., their image content descriptions are converted into interactive preference quantitative knowledge in the above manner to form the first interactive preference knowledge vector set.

[0078] In the online education scenario, the same process is followed. Taking content push annotation as an example, in online education, there may be different types of annotations such as course material push and learning task push, and there are category relationship vectors between these annotations. In terms of the image descriptor of each content category label, under the content category label of page interaction event, the image descriptor may include the mouse clicking on the mathematical function image to enlarge the view, the dwell time on the portrait of a historical figure, etc.

[0079] For subscription feedback comments in X content category tags, in the online education scenario, they may be students' comments on teaching image content, such as questions and suggestions. The community map between page subscribers may reflect the relationship between students in learning groups and classes, and their interactive behaviors on educational images may be similar. Content update nodes may involve version updates of course image materials. The content heat value reflects the popularity of educational images on the platform. For example, a physics experiment device diagram has a high content heat value because it is related to important knowledge points and explained in detail.

[0080] When determining the global image content description, for example, for a biological cell structure image, its global image content description will integrate the cell structure knowledge point push in the content push annotation, the click-to-view operation of different areas of the image in the page interaction event, the students' questions about cell structure and function in the subscription feedback comments, the interaction of the same group of students on the image in the community map between page subscribers, the update information about the cell structure image in the content update node, and the popularity of the image in the content heat value. Based on the category relationship vector between X content category labels, the image content descriptions in the global image content description of the image are obtained by splicing any image descriptor included in each content category label.

[0081] Generate the quantitative knowledge of interaction preference for each image content description based on the number of global image content descriptions. For example, if the number of global image content descriptions is 8, for a global image content description of a historical figure portrait image, if the content push annotation related part is closely related to the course focus and is pushed frequently, the quantification is 7; if there are many mouse clicks to view the deeds of the person in the page interaction event, the quantification is 5.

[0082] Finally, the image content descriptions of each page interactive image in the first page interactive image set are converted into the interactive preference quantitative knowledge of the corresponding image content description to obtain the first interactive preference knowledge vector set. For many education-related images browsed by users on the online education platform, such as physics experiment device diagrams, historical figure portraits, etc., they are converted in this way to form the first interactive preference knowledge vector set.

[0083] It can be seen that in the scenarios of digital office and online education, the global image content description is constructed by integrating the relationship between multiple content category labels and image descriptors, and then the interaction preference quantitative knowledge is generated to obtain the first interaction preference knowledge vector set. This helps to more comprehensively and accurately convert the image content description into a quantifiable and analyzable data form. In digital office, it can better grasp the user's interest preference for office images and provide a basis for office resource recommendations; in online education, it can deeply understand the students' interest in educational images, which is conducive to personalized learning resource push and teaching optimization, and improve user experience and business results.

[0084] In some other embodiments, any page interaction image in the first page interaction image set is represented as a first page interaction image; the method further includes: obtaining a category relationship vector between X content category tags, the X content category tags including: content push annotation, page interaction event, subscription feedback comment, community map between page subscription users, content update node, content heat value or one or more thereof; obtaining an image descriptor of the first page interaction image under each of the content category tags; and integrating description features of the image descriptors of the first page interaction image under each of the content category tags through the category relationship vectors between the X content category tags to obtain an image content description of the first page interaction image.

[0085] In a digital office scenario, for any first page interactive image in the first page interactive image set, the following is a process of operating according to the technical solution of this embodiment.

[0086] First, the category relationship vectors between X content category tags are obtained, where the X content category tags include one or more of content push annotations, page interaction events, subscription feedback comments, community graphs between page subscription users, content update nodes, and content heat values.

[0087] In terms of content push annotation, taking the digital office scenario as an example, different types of annotations may include office software function update push, specific project document push, etc. There are certain category relationship vectors between these annotations. For example, office software function update push may be associated with specific project document push, because the new function may be applied to the processing of project documents, and this association is reflected in the category relationship vector.

[0088] Page interaction events include operations such as clicking, sliding, zooming in and out of office images with the mouse, and inputting information into interfaces related to office images with the keyboard. There are also relationships between these operations. For example, a mouse click may trigger a subsequent keyboard input operation. This logical relationship is also reflected in the category relationship vector.

[0089] Subscription feedback comments include office users' evaluations and suggestions on the content related to the office images they browse. For example, users may make optimization suggestions for an office process image, or express satisfaction with the presentation of a business meeting image. These comments are related to other content category tags. For example, positive comments may affect the push labeling of content or increase the heat value of content.

[0090] The community graph between page subscribers describes the relationship network between office users. For example, within an enterprise, employees in different departments may have different interaction patterns with office images, and employees in the same department or the same project team may have similar image interaction behaviors due to work relevance. This relationship is reflected in the community graph and has an interactive relationship with other content category tags.

[0091] The content update node indicates the update of office image content, such as version update of office document image, process adjustment of office flowchart, etc. It is closely related to content push annotation, and new content update may trigger new content push annotation. At the same time, the content update node is also related to the community map between page subscribers, because different user groups may have different reactions to different types of content updates.

[0092] The content heat value reflects the popularity of office images on the platform, which may be determined based on factors such as views, likes, and shares. It interacts with other content category tags. For example, office images with high content heat values ​​may trigger more subscription feedback and comments, and may also affect the content push annotation strategy.

[0093] Next, the image descriptor of the first page interactive image under each content category label is obtained. For the image descriptor under the content push label, taking the office software function update push as an example, if the first page interactive image is an office software interface image, then the image descriptor may include information such as the display area and display mode of the update function on the interface. For the image descriptor under the page interactive event, if it is a mouse click operation, the image descriptor may include information such as the coordinate position of the click and the number of clicks.

[0094] For the image descriptor under the subscription feedback comment, if a user makes a comment about process simplification on an office flowchart image, the image descriptor may include the comment's keywords (such as "process simplification"), the comment's sentiment (such as positive or negative), etc. For the image descriptor under the community graph between page subscription users, if the first page interaction image is a business meeting scene graph, the image descriptor may include information such as the department and position of the users participating in the meeting to reflect their position in the community graph.

[0095] For the image descriptor under the content update node, if it is a version update of an office document image, the image descriptor may include the specific content of the update, the time of the update, etc. For the image descriptor under the content heat value, it may include specific information such as the number of views and the number of likes.

[0096] Then, through the category relationship vectors between the X content category labels, the image descriptors of the first page interactive image under each content category label are integrated to obtain the image content description of the first page interactive image. For example, for an office flowchart image, through the category relationship vector between the content push annotation and the page interactive event, the image descriptor of the flowchart version update push under the content push annotation (such as the updated process link) and the image descriptor of the mouse clicking on the specific link of the flowchart under the page interactive event (such as the name and coordinates of the clicked link) are integrated. Combined with the image descriptor of flowchart optimization under the subscription feedback comment (such as the optimization suggestion keywords in the comment), the user information image descriptor of participating in flowchart-related meetings under the community map between page subscription users (such as the user's department and position), the image descriptor of the flowchart update time under the content update node, and the image descriptor of the number of views of the flowchart image under the content heat value, the image content description of the office flowchart image is obtained. This image content description is not just a simple information listing, but an organic integration based on the relationship between the various content category labels, so as to more comprehensively and accurately reflect the content characteristics of the first page interactive image.

[0097] In the online education scenario, content push annotations may include different types of annotations such as course material push, learning task push, etc. For example, for mathematics, there may be push of mathematical formula derivation materials, and there are category relationship vectors between these annotations. For example, the push of mathematical formula derivation materials may be associated with the push of specific mathematical learning tasks, because formula derivation helps to complete the learning task.

[0098] Page interaction events cover various operations that students perform on educational images, such as clicking on a math function image to view function characteristics, sliding the mouse on a historical figure portrait to view more information, etc. There is a logical relationship between these operations and they are related to other content category tags. For example, frequently clicking on a math function image to view function characteristics may trigger more course material pushes (related to content push annotations) and may also affect the content heat value.

[0099] Subscription feedback comments contain students' questions and insights about educational image content. For example, students may have questions about the experimental principle in a physical experiment setup diagram and raise them in comments, which are related to other content category tags. For example, a large number of questions about the principle of a physical experiment setup diagram may trigger the content update node, that is, update the relevant explanation content of the physical experiment setup diagram.

[0100] The community graph between page subscribers describes the relationships between students, such as study groups, classes, etc. For example, students in a study group may have similar interactive behaviors on educational images, and they may have common concerns about the same historical figure portrait. This relationship interacts with other content category tags. For example, the common concern of a study group on a historical figure portrait may increase the content heat value of the image.

[0101] The content update node indicates the update of educational image content, such as the knowledge point update of biological cell structure images, supplementary information of historical event images, etc. It is closely related to content push annotations, and new content updates will lead to new content push annotations. At the same time, it is also related to subscription feedback comments, and students' questions and insights may prompt the occurrence of content update nodes.

[0102] The content heat value reflects the popularity of educational images on the platform, which is determined based on factors such as views, likes, and shares. It interacts with other content category tags. For example, educational images with high content heat values ​​may attract more subscriptions, feedback, and comments, and also affect the strategy of content push annotation.

[0103] Get the image descriptor of the first page interactive image under each content category label. Taking the image descriptor under the content push label as an example, if the first page interactive image is a physical experiment device diagram, the image descriptor may include information such as the associated area of ​​the pushed experimental principle explanation on the image and the presentation method of the explanation. For the image descriptor under the page interactive event, if the mouse clicks on a component in the physical experiment device diagram, the image descriptor may include information such as the name of the clicked component and the coordinate position of the click.

[0104] For the image descriptors under subscription feedback comments, if a student raises a question about the contribution of a historical figure, the image descriptor may include the keywords of the question (such as "contribution of the figure"), the emotional tendency of the question (such as curiosity), etc. For the image descriptors under the community graph between page subscribers, if the first page interaction image is an image of a mathematical function discussed by a study group, the image descriptor may include the name of the study group to which the students participating in the discussion belong, the role of the students in the group, etc.

[0105] For the image descriptor under the content update node, if it is a knowledge point update of a biological cell structure image, the image descriptor may include information such as the specific knowledge point updated and the time of update. For the image descriptor under the content heat value, it may include specific information such as the number of views and the number of likes.

[0106] Through the category relationship vectors between X content category labels, the image descriptors under each content category label of the first page interactive image are integrated with description features to obtain the image content description of the first page interactive image. For example, for a biological cell structure image, through the category relationship vector between the content push annotation and the page interactive event, the image descriptor of the cell structure knowledge point push under the content push annotation (such as the annotation position of the pushed knowledge point on the image) and the image descriptor of the mouse clicking on the specific part of the cell structure image under the page interactive event (such as the name and coordinates of the clicked cell structure) are integrated. Combined with the question image descriptor about the cell structure knowledge point under the subscription feedback comment (such as the knowledge point keyword in the question), the student information image descriptor participating in the cell structure-related discussion under the community map between page subscription users (such as the study group and role to which the student belongs), the update time image descriptor of the cell structure image under the content update node, and the image descriptor of the number of views of the cell structure image under the content heat value, etc., the image content description of the biological cell structure image is obtained.

[0107] Therefore, in the scenarios of digital office and online education, the image content description of the interactive image on the first page is constructed by integrating the relationship between multiple content category labels and the corresponding image descriptors. This process can comprehensively and deeply mine various information related to the image content, avoiding the limitations of single-dimensional description. In digital office, it helps to accurately grasp the connotation of office images and provide a more accurate basis for office process optimization, resource push, etc.; in online education, it can better understand students' needs and concerns for educational images, thereby achieving more personalized learning resource recommendations and teaching improvements, and improving user experience and business efficiency.

[0108] In an optional design idea, the interest portrait generation process is implemented by calling a deep convolution algorithm with a knowledge feature enhancement function; the number of feature depths is Q, and the deep convolution algorithm includes an interest portrait generation branch and Q knowledge feature enhancement branches, and one subscription interest knowledge entity set corresponds to one knowledge feature enhancement branch; Q is a positive integer; the use of the knowledge feature enhancement strategy to perform interest portrait generation processing on multiple subscription interest knowledge entity sets of the first page subscriber to obtain a page subscription interest portrait includes: respectively passing the Q subscription interest knowledge entity sets of the first page subscriber to the corresponding knowledge feature enhancement branches to mine interest entities and interest relationships, and obtaining Q interaction preference knowledge enhancement vectors, each interaction preference knowledge enhancement vector originating from a subscription interest knowledge entity set; the Q interaction preference knowledge enhancement vectors are fully connected and then imported into the interest portrait generation branch to perform interest portrait generation processing to obtain a page subscription interest portrait.

[0109] In a digital office scenario, when operating according to the technical solution of this embodiment, since the interest portrait generation process is implemented by calling a deep convolution algorithm with a knowledge feature enhancement function, and the number of feature depths is Q, the deep convolution algorithm includes an interest portrait generation branch and Q knowledge feature enhancement branches, and each subscribed interest knowledge entity set corresponds to a knowledge feature enhancement branch.

[0110] Taking the project management image in digital office as an example, set Q=3, then there are 3 feature depths. For the page interaction image related to project management browsed by the first page subscription user, 3 subscription interest knowledge entity sets are obtained through the previous steps. When these 3 subscription interest knowledge entity sets are respectively passed to the corresponding knowledge feature enhancement branches, each branch will mine interest entities and interest relationships.

[0111] In the first knowledge feature enhancement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to the project process, such as project process steps, key decision points in the process, etc. This branch will dig out deeper interest entities and relationships, such as the time dependency between process steps, the impact of key decision points on the overall progress of the project, etc., to obtain an interactive preference knowledge enhancement vector. This vector may be a feature vector containing multiple elements, such as [weight of time dependency between process steps: 0.6, impact of key decision points on progress: 0.8, ranking of importance of process links: 0.7], etc. It reflects the result of this subscription interest knowledge entity set after being processed by the knowledge feature enhancement branch.

[0112] In the second knowledge feature enhancement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to project team members, such as member roles, collaboration relationships between members, etc. The interest entities and relationships mined by this branch may include the collaboration frequency of different roles in specific tasks, the efficiency of information transfer between members, etc. The obtained interaction preference knowledge enhancement vector may be a feature vector of elements such as [member collaboration frequency weight: 0.7, information transfer efficiency quantification value: 0.6, importance of collaboration between different roles: 0.8].

[0113] In the third knowledge feature enhancement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to project resources, such as funds, equipment, etc. The interest entities and relationships mined by this branch may include the association between capital investment and project stage, the impact of equipment use on project efficiency, etc. The obtained interactive preference knowledge enhancement vector may be a feature vector of elements such as [correlation between capital investment and project stage: 0.8, impact value of equipment use efficiency: 0.7, rationality of resource allocation: 0.6].

[0114] Then, these three interactive preference knowledge reinforcement vectors are fully connected. Full connection means that a comprehensive connection relationship is established between the elements of these vectors to integrate the information they contain. The results of the full connection processing are then imported into the interest portrait generation branch for interest portrait generation processing. In the interest portrait generation branch, a page subscription interest portrait that can fully characterize the interest characteristics of the first page subscriber in the process of browsing the interactive image of the project management page will be constructed based on these integrated information. This page subscription interest portrait may be a feature vector that integrates multiple interest information such as project process, team members, project resources, etc., such as [comprehensive interest in project process: 0.7, team member collaboration interest: 0.6, project resource attention: 0.8] and other elements. It accurately reflects the user's interest characteristics in project management related images in the digital office scenario.

[0115] In the online education scenario, for example, Q=3, which is the interactive images of pages related to subject knowledge browsed by students on the online education platform, such as experimental images of physics and character images of history.

[0116] In the first knowledge feature reinforcement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to the physics subject knowledge system, such as physical concepts, physical laws, etc. The interest entities and relationships mined by this branch may include the logical deduction relationship between physical concepts, the application of physical laws in different experiments, etc. The obtained interactive preference knowledge reinforcement vector may be a feature vector of elements such as [concept logical deduction weight: 0.7, law experimental application frequency: 0.8, concept importance ranking: 0.6].

[0117] In the second knowledge feature reinforcement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to historical events in the history discipline, such as the time of the event, the characters in the event, etc. The interest entities and relationships mined by this branch may include the impact of the sequence of events on historical development, the role of characters in the event, etc. The obtained interactive preference knowledge reinforcement vector may be a feature vector of elements such as [event time sequence influence: 0.6, character role weight: 0.7, event relevance: 0.8].

[0118] In the third knowledge feature enhancement branch, for example, the corresponding subscription interest knowledge entity set contains entities related to learning resources, such as textbook versions, online course lecturers, etc. The interest entities and relationships mined by this branch may include the matching degree between textbook versions and knowledge point coverage, the influence of lecturer teaching style on learning effect, etc. The obtained interactive preference knowledge enhancement vector may be a feature vector of elements such as [textbook version matching degree: 0.8, lecturer teaching style influence value: 0.7, learning resource importance: 0.6].

[0119] After fully connected, these three interactive preference knowledge reinforcement vectors are imported into the interest portrait generation branch for interest portrait generation. The page subscription interest portrait constructed in the interest portrait generation branch may be a feature vector that integrates multiple interest information such as physics knowledge, history knowledge, and learning resources. For example, a vector of elements such as [physics interest: 0.7, history attention: 0.6, learning resource preference: 0.8] accurately reflects the student's interest characteristics in subject knowledge-related images in online education scenarios.

[0120] In summary, in the scenarios of digital office and online education, the knowledge feature reinforcement branch in the deep convolution algorithm is used to mine the interest entities and relationships of different subscription interest knowledge entity sets, obtain the interactive preference knowledge reinforcement vector, and then generate the page subscription interest portrait through full connection processing and interest portrait generation branch. This process can deeply mine user interests and comprehensively and accurately build interest portraits by combining various information. In digital office, it helps to accurately push office resources and optimize project management; in online education, it can realize personalized learning resource recommendations and improve learning effects.

[0121] In an exemplary technical solution, any subscription interest knowledge entity set among the Q subscription interest knowledge entity sets is represented as the pth subscription interest knowledge entity set, and the Q interaction preference knowledge reinforcement vectors include the pth interaction preference knowledge reinforcement vector derived from the pth subscription interest knowledge entity set; the pth subscription interest knowledge entity set corresponds to the pth knowledge feature reinforcement branch among the Q knowledge feature reinforcement branches, and the pth knowledge feature reinforcement branch includes a residual processing sub-branch and multiple secondary knowledge feature reinforcement layers with a cascade relationship; p is a positive integer not greater than Q; the pth subscription interest knowledge entity set is passed into the pth knowledge feature reinforcement branch to mine interest entities and interest relationships to obtain the pth interaction preference knowledge reinforcement vector, including: inputting the pth subscription interest knowledge entity set into the residual processing sub-branch for residual processing to obtain the subscription interest residual knowledge of the pth subscription interest knowledge entity set; inputting the subscription interest residual knowledge of the pth subscription interest knowledge entity set into the multiple secondary knowledge feature reinforcement layers with a cascade relationship for inter-class feature reinforcement processing to obtain the pth interaction preference knowledge reinforcement vector.

[0122] In the digital office scenario, operations are carried out according to this technical solution. For any p-th subscription interest knowledge entity set among the Q subscription interest knowledge entity sets, take the case related to office document management in digital office as an example. For example, when Q=3, when p=1, the first subscription interest knowledge entity set may contain entities related to the type of office documents, document access rights, etc.

[0123] This first subscription interest knowledge entity set corresponds to the first knowledge feature reinforcement branch among the Q knowledge feature reinforcement branches, and the first knowledge feature reinforcement branch includes a residual processing sub-branch and multiple secondary knowledge feature reinforcement layers with a cascade relationship.

[0124] First, the first subscription interest knowledge entity set is input into the residual processing sub-branch for residual processing. For example, the first subscription interest knowledge entity set is represented in the form of a feature vector as [office document type: document A, access rights: partial access, document creation time: specific time and other elements]. Residual processing may adjust these elements to supplement or correct possible errors or incomplete information. For example, after residual processing, it is found that the actual access rights of document A should be full access, and there is a certain deviation in the document creation time. After correction, the subscription interest residual knowledge of the first subscription interest knowledge entity set is obtained. The feature vector of this residual knowledge may become [office document type: document A, access rights: full access, document creation time: corrected time and other elements].

[0125] This subscription interest residual knowledge is then input into multiple secondary knowledge feature reinforcement layers with cascading relationships for inter-class feature reinforcement processing. These secondary knowledge feature reinforcement layers gradually strengthen the features of different categories. For example, the first secondary knowledge feature reinforcement layer may strengthen the relationship between office document type and access rights, so that in the feature vector, when the office document type is document A, the access right is full access. This relationship is strengthened, and its weight may be increased from the initial 0.5 to 0.7. The second secondary knowledge feature reinforcement layer may strengthen the relationship between document creation time and other factors, such as the relationship between document creation time and document version update, which is reflected in the feature vector as the correlation between document creation time and document version update is increased from 0.3 to 0.5, etc. After being processed by multiple secondary knowledge feature reinforcement layers, the first interactive preference knowledge reinforcement vector is obtained. This vector may be [office document type and access right relationship weight: 0.7, document creation time and version update correlation: 0.5, other enhanced relationship weights and other elements].

[0126] When p=2, for example, the second subscription interest knowledge entity set is related to office staff collaboration, including staff departments, collaboration tasks and other entities. Similarly, it is first input into the corresponding residual processing sub-branch of the second knowledge feature enhancement branch. For example, the feature vector of the second subscription interest knowledge entity set is [staff department: Department B, collaboration task: Task X, collaboration frequency: 3 times a week, etc.]. The residual processing sub-branch may find that there is a statistical error in the collaboration frequency, correct it to 5 times a week, and obtain the feature vector of the subscription interest residual knowledge as [staff department: Department B, collaboration task: Task X, collaboration frequency: 5 times a week, etc.].

[0127] It is then input into multiple secondary knowledge feature reinforcement layers with cascading relationships for inter-class feature reinforcement processing. The first secondary knowledge feature reinforcement layer may strengthen the relationship between personnel departments and collaborative tasks. For example, in the feature vector, the matching weight between department B and task X is increased from 0.6 to 0.8. The second secondary knowledge feature reinforcement layer may strengthen the relationship between collaboration frequency and task progress. For example, the weight of the impact of collaboration frequency 5 times a week on task progress is increased from 0.4 to 0.6. Finally, the second interactive preference knowledge reinforcement vector is obtained, such as [department and task matching weight: 0.8, collaboration frequency impact weight on task progress: 0.6, other enhanced relationship weights and other elements].

[0128] In the online education scenario, the operation is performed in a similar manner. For example, when Q=3 and p=1, the first subscription interest knowledge entity set may be related to subject knowledge, such as the content of the subject chapter, the difficulty of the knowledge point, and other entities. Taking mathematics as an example, the feature vector of the first subscription interest knowledge entity set may be [chapter content: function chapter, difficulty of knowledge point: medium, number of examples associated with knowledge point: 5, etc.].

[0129] It is input into the residual processing sub-branch of the corresponding first knowledge feature reinforcement branch for residual processing. For example, if the evaluation of the difficulty of a knowledge point is found to be biased, it is corrected to be relatively difficult, and the feature vector of the residual knowledge of subscription interest is obtained as [section content: function chapter, difficulty of knowledge point: relatively difficult, number of examples associated with knowledge point: 5, etc.].

[0130] Then it is input into multiple secondary knowledge feature reinforcement layers with cascading relationships for inter-class feature reinforcement processing. The first secondary knowledge feature reinforcement layer may strengthen the relationship between chapter content and the difficulty of knowledge points, such as increasing the association weight between function chapters and more difficult knowledge points from 0.4 to 0.6. The second secondary knowledge feature reinforcement layer may strengthen the relationship between the difficulty of knowledge points and the number of examples, such as increasing the fitness weight between more difficult knowledge points and 5 examples from 0.3 to 0.5. Finally, the first interactive preference knowledge reinforcement vector is obtained, such as [the association weight between chapter content and the difficulty of knowledge points: 0.6, the fitness weight between the difficulty of knowledge points and the number of examples: 0.5, and other elements such as the weight of other strengthened relationships].

[0131] When p=2, for example, the second subscription interest knowledge entity set is related to educational resources, such as online course lecturers, course duration, etc. The feature vector of the second subscription interest knowledge entity set is [lecturer: lecturer A, course duration: 60 minutes, course evaluation level: good, etc.].

[0132] First, input it into the residual processing sub-branch of the second knowledge feature enhancement branch for residual processing. For example, if it is found that there is a misjudgment of the course evaluation level, it is corrected to excellent, and the characteristic vector of the subscription interest residual knowledge is obtained as [lecturer: lecturer A, course duration: 60 minutes, course evaluation level: excellent, etc.].

[0133] Then it is input into multiple secondary knowledge feature reinforcement layers with cascading relationships for inter-class feature reinforcement processing. The first secondary knowledge feature reinforcement layer may strengthen the relationship between the lecturer and the course evaluation grade, such as increasing the association weight between lecturer A and the excellent evaluation grade from 0.5 to 0.7. The second secondary knowledge feature reinforcement layer may strengthen the relationship between course duration and evaluation grade, such as increasing the association weight between a 60-minute course duration and an excellent evaluation grade from 0.4 to 0.6. Finally, the second interactive preference knowledge reinforcement vector is obtained, such as [lecturer and evaluation grade association weight: 0.7, course duration and evaluation grade association weight: 0.6, other elements such as enhanced relationship weights].

[0134] In the scenarios of digital office and online education, the interactive preference knowledge reinforcement vector is obtained by processing the subscription interest knowledge entity set through the residual processing sub-branch and the cascaded secondary knowledge feature reinforcement layer. This helps to correct possible entity set information errors and strengthen the relationship between entities. In digital office, it can improve the accuracy of office resource management and collaboration efficiency, and in online education, it can improve the accuracy of learning resource matching and learning effects.

[0135] Based on the above optional design ideas, the method also includes: obtaining a second page interaction image set, which is obtained by sequentially sorting multiple page interaction images browsed by the second page subscriber within the target session period based on the page response priority; performing knowledge extraction on the second page interaction image set according to the image content description of each page interaction image in the second page interaction image set to obtain a second interaction preference knowledge vector set; performing subscription interest knowledge entity extraction with multiple feature depths on the second interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the second page subscriber; calling the original deep convolution algorithm to perform interest portrait generation processing on the multiple subscription interest knowledge entity sets of the second page subscriber to obtain a page subscription interest training portrait, which is used to characterize the interest characteristics of the second page subscriber in the process of browsing page interaction images; debugging the original deep convolution algorithm based on the page subscription interest training portrait and the second page subscriber's prior debugging annotations to obtain the deep convolution algorithm with knowledge feature enhancement function.

[0136] In the next step, the original deep convolution algorithm is debugged based on the page subscription interest training portrait and the prior debugging annotations of the second page subscription user to obtain the deep convolution algorithm with the knowledge feature enhancement function, including: determining the algorithm training error of the original deep convolution algorithm based on the page subscription interest training portrait and the prior debugging annotations of the second page subscription user; optimizing the algorithm variable weights of the original deep convolution algorithm towards the goal of converging the algorithm training error to obtain the deep convolution algorithm with the knowledge feature enhancement function; wherein, when the second page subscription user is updated as an active page subscription user in the targeted push server, the prior debugging annotation of the second page subscription user is a targeted push annotation; when the second page subscription user is not updated as an active page subscription user in the targeted push server, the prior debugging annotation of the second page subscription user is a non-targeted push annotation.

[0137] In the digital office scenario, the second page interactive image set is first obtained. This set is obtained by sorting multiple page interactive images browsed by the second page subscriber within the target session cycle based on the page response priority. For example, in digital office, the second page subscriber browses various images on the office platform within a working day (target session cycle), such as office software interface images, project progress charts, business meeting scene images, etc. According to the page response priority, for example, the progress charts of some important projects may have a higher page response priority due to their key impact on the business, and will be sorted first, thereby forming the second page interactive image set.

[0138] Then, based on the image content description of each page interaction image in the second page interaction image set, the set is subjected to knowledge extraction to obtain the second interaction preference knowledge vector set. Taking the office software interface image as an example, the image content description may include information such as functional modules, operation buttons, and displayed data on the interface. Through a specific knowledge extraction method, this information is converted into knowledge elements and then constructed into the second interaction preference knowledge vector set. For example, the knowledge elements of an office software interface image may include [functional module A: exists (1), operation button B: display (1), data type: report (1), etc.], and multiple such image knowledge elements constitute the second interaction preference knowledge vector set.

[0139] Then, the second interactive preference knowledge vector set is subjected to multiple feature depths of subscription interest knowledge entity extraction to obtain multiple subscription interest knowledge entity sets of the second page subscribers. For example, three feature depths are set. At the first feature depth, an entity set directly related to the surface content of the office image may be extracted, such as {office software interface: 3 (indicates that it appears in 3 images), project progress chart: 2, business meeting scene: 1}. At the second feature depth, in addition to these entities, it also includes the relationship between entities, such as {office software interface-function module-operation button: 2 (indicates that there are 2 images reflecting this relationship), project progress chart-project stage-timeline: 1}. At the third feature depth, based on a broader office knowledge graph, there may be {office software interface-function module-data processing-work efficiency: 1, project progress chart-project stage-resource allocation-project goal: 1}, etc.

[0140] Then, the original deep convolution algorithm is called to generate interest portraits for multiple subscription interest knowledge entity sets of the second page subscribers to obtain a page subscription interest training portrait. This portrait is used to characterize the interest characteristics of the second page subscribers in the process of browsing the page interaction image. For example, the page subscription interest training portrait may be a feature vector [Office software interface interest: 0.6, project progress chart attention: 0.4, business meeting scene relevance: 0.3, function module and operation button relationship weight: 0.5, project stage and resource allocation relevance: 0.4, etc.], which integrates interest characteristics at different levels.

[0141] Based on the page subscription interest training portrait and the prior debugging annotations of the second page subscriber, the original deep convolution algorithm is debugged to obtain a deep convolution algorithm with knowledge feature enhancement function. When the second page subscriber is updated as an active page subscriber in the directional push server, the prior debugging annotation of the second page subscriber is a directional push annotation. For example, if the user is a member of an important project team (active page subscriber), the directional push annotation may be related to the project-related office resource push preference. Based on the page subscription interest training portrait and the directional push annotation, the algorithm training error of the original deep convolution algorithm is determined. For example, the portrait shows that the user's attention to the project progress chart is 0.4, but the directional push annotation indicates that the user should pay more attention to the project progress chart (for example, the ideal attention is 0.7), which means there is an algorithm training error. Towards the goal of converging this algorithm training error, the algorithm variable weights of the original deep convolution algorithm are optimized. For example, by adjusting the weights of the feature extraction layer related to the project progress chart, the algorithm's ability to capture the interest features of this type of image is improved, and finally a deep convolution algorithm with knowledge feature enhancement function is obtained.

[0142] When the second page subscriber is not updated as an active page subscriber in the targeted push server, the prior debugging annotation of the second page subscriber is a non-targeted push annotation. For example, for ordinary office workers (non-active page subscribers), non-targeted push annotations may be based on general office image browsing habits. Similarly, based on the page subscription interest training portrait and non-targeted push annotations, the algorithm training error is determined and the algorithm variable weights are optimized. For example, the portrait shows that the relevance to the business meeting scene is 0.3, and the non-targeted push annotation indicates that the average relevance of ordinary office workers to the business meeting scene should be 0.2, which has a certain error. The algorithm is optimized by adjusting the algorithm variable weights related to the business meeting scene.

[0143] In the online education scenario, similarly, the second page interactive image set is obtained. For example, within a semester (target session cycle), the second page subscribers (students) browse subject knowledge images on the online education platform, such as physics experiment device diagrams, mathematical function images, historical figure portraits, etc. According to the page response priority, images related to the current course focus may have a higher priority, and the second page interactive image set is organized.

[0144] The second interactive preference knowledge vector set is obtained by performing knowledge extraction based on the image content description. Taking the physical experiment device diagram as an example, the image content description may contain information such as the components of the experiment device, the experiment type, and the data display. After being converted into knowledge elements, they form the second interactive preference knowledge vector set, such as [experimental device component A: existence (1), experiment type: electricity (1), data display method: instrument (1), etc.].

[0145] Subscription interest knowledge entity extraction with multiple feature depths is performed to obtain multiple subscription interest knowledge entity sets. For example, with three feature depths, at the first feature depth, there may be {physics: 4 (indicates that it is involved in 4 images), mathematics: 2, history: 1}. At the second feature depth, such as {physics-experimental device-experimental principle: 2 (indicates that there are 2 images that reflect this relationship), mathematics-function-image property: 1}. At the third feature depth, {physics-experimental device-physical concept-scientific exploration: 1, mathematics-function-variable relationship-mathematical thinking: 1}, etc.

[0146] The original deep convolution algorithm is called to obtain the page subscription interest training portrait, for example [interest in physics: 0.7, attention in mathematics: 0.3, relevance to history: 0.2, weight of the relationship between experimental apparatus and experimental principle: 0.6, relevance between function and image property: 0.5, etc.].

[0147] The original deep convolution algorithm is debugged based on the page subscription interest training portrait and prior debugging annotations. When the student is an active user (such as actively participating in a specific subject learning plan), the prior debugging annotation is a directional push annotation. For example, the portrait shows that the interest in physics is 0.7, but the directional push annotation indicates that the student's interest in physics should reach 0.9 in a special learning stage. There is an algorithm training error. By optimizing the algorithm variable weights, such as adjusting the weights related to the extraction of physics knowledge, a deep convolution algorithm with knowledge feature enhancement function is obtained. When the student is an inactive user, the prior debugging annotation is a non-directional push annotation, and the error is also determined and the algorithm is optimized based on the portrait and annotation.

[0148] It can be understood that in the scenarios of digital office and online education, training portraits are obtained by processing the image data related to the second page subscribers, and then the algorithm is debugged in combination with prior debugging annotations. This helps to improve the ability of the deep convolution algorithm to capture user interest characteristics. In digital office, office resources can be pushed to different users more accurately, and in online education, more personalized learning resource recommendations can be achieved, improving user experience and business efficiency.

[0149] On the basis of S110 - S170 , the method further includes: based on the page subscription interest portrait, performing targeted information push to the first page subscriber.

[0150] Exemplarily, the first page subscriber is any page subscriber among the authorized page subscribers of the directional push server. Based on this, the directional information push to the first page subscriber based on the page subscription interest portrait includes: if the page subscription interest portrait indicates that the interest characteristics of the first page subscriber in the process of browsing the page interaction image meet the directional push conditions, then the first page subscriber is updated to an active page subscriber in the directional push server and interest subscription message push is performed based on the interest characteristics of the first page subscriber in the process of browsing the page interaction image; or, the directional push strategy for the first page subscriber is counted in the directional push server; or, the directional push event link of the first page subscriber is cached in the directional push server.

[0151] In the digital office scenario, operations are carried out according to this technical solution. First, let's review the previous S110-S170 process. In this process, for the first page subscriber, a page subscription interest portrait is obtained after a series of operations. This page subscription interest portrait comprehensively reflects the interest characteristics of the first page subscriber when browsing the page interaction images related to digital office.

[0152] For example, in an enterprise office environment, the first page subscriber may be an employee within the enterprise who browses various work-related images on the office platform, such as office process diagrams, project management Gantt charts, office software operation interface diagrams, etc. The page subscription interest portrait generated by the previous process may be a feature vector containing multiple elements, such as [office process attention: 0.8, project management interest: 0.6, office software operation proficiency: 0.7, specific office document type preference: 0.5, etc.].

[0153] Based on this page subscription interest portrait, targeted information is pushed to the first page subscriber. If this page subscription interest portrait indicates that the interest characteristics of the first page subscriber in the process of browsing the page interactive image meet the targeted push conditions, then in the targeted push server, the first page subscriber is updated to an active page subscriber, and interest subscription messages are pushed based on his interest characteristics. For example, if the office process attention in the portrait reaches 0.8, it indicates that the user is very interested in office process-related content. At this time, the user is marked as an active page subscriber in the targeted push server. Then, based on his interest characteristics, interest subscription messages such as training materials for office process optimization and the latest office process templates are pushed.

[0154] Alternatively, the targeted push strategy for the user who subscribes to the first page is counted in the targeted push server. For example, according to the office software operation proficiency of 0.7 in the page subscription interest portrait, it can be known that the user has a certain foundation in office software operation but still has room for improvement. When the targeted push server counts the targeted push strategy for the user, it may consider pushing some tutorials on how to use advanced office software functions, or a recommended strategy for efficiency improvement tools related to the user's commonly used office software.

[0155] Alternatively, the targeted push event link of the first page subscriber is cached in the targeted push server. For example, when the page subscription interest profile shows that the user's interest in project management is 0.6, there may be a targeted push event related to project management, such as an online lecture by a project management expert. The targeted push server caches the targeted push event link of this online lecture so that it can push this link to the user at an appropriate time (such as based on the user's online time, recent work schedule, etc.).

[0156] In the online education scenario, the first page subscriber is a student who studies on an online education platform. The interactive images of the page browsed by the student may be subject knowledge explanation diagrams, experimental operation diagrams, historical event context diagrams, etc. The generated page subscription interest portrait may be [Subject knowledge interest distribution: [Mathematics: 0.7, Physics: 0.5, History: 0.4], Experimental operation attention: 0.6, Historical event exploration desire: 0.3, etc.].

[0157] If the page subscription interest profile indicates that the student's interest characteristics meet the conditions for targeted push, for example, the interest in mathematics knowledge reaches 0.7, the targeted push server will update the student as an active page subscriber and push interest subscription messages based on his interest characteristics, such as pushing mathematics competition training materials, mathematics problem analysis courses, etc.

[0158] When counting the targeted push strategies for the student in the targeted push server, based on the experimental operation attention of 0.6, a strategy can be formulated to push more experimental extension courses, experimental operation skills explanation videos, etc.

[0159] At the same time, the directional push event link is cached in the directional push server. For example, when the historical event exploration desire is 0.3, if there is a directional push event such as a historical and cultural lecture, the link of the lecture is cached for subsequent push.

[0160] In the digital office and online education scenarios, targeted information is pushed to users based on their interest profiles. This helps to accurately meet user interest needs, improve employee office efficiency in digital office, and improve student learning outcomes in online education. The accuracy and timeliness of resource push are optimized by updating qualified users to active users, counting push strategies, and caching push event links.

[0161] Furthermore, a readable storage medium is provided, on which a program is stored, and the program implements the above method when executed by a processor.

[0162] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described in the above exemplary embodiment can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, 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 through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

Claims

1. A method for generating interest portraits based on knowledge graph, characterized in that: The method is applied to an interest portrait generation server, and the method includes: Acquire a first page interaction image set, where the first page interaction image set is obtained by sequentially arranging a plurality of page interaction images browsed by a first page subscriber within a target session period based on a page response priority; According to the image content description of each page interaction image in the first page interaction image set, knowledge is extracted from the first page interaction image set to obtain a first interaction preference knowledge vector set; The first interaction preference knowledge vector set is subjected to subscription interest knowledge entity extraction at multiple feature depths to obtain multiple subscription interest knowledge entity sets of the first page subscribers; wherein one feature depth corresponds to one subscription interest knowledge entity set, and each subscription interest knowledge entity set is used to characterize the interest entities and interest relationships of the first page interaction image set at the corresponding feature depth; Using a knowledge feature enhancement strategy, interest portrait generation processing is performed on multiple subscription interest knowledge entity sets of the first page subscriber to obtain a page subscription interest portrait, wherein the page subscription interest portrait is used to characterize the interest features of the first page subscriber in the process of browsing the page interaction image; The number of the feature depths is Q, any one of the Q feature depths is represented as a p-th level feature depth, the subscription interest knowledge entity extraction includes knowledge entity mining and entity relationship construction, p is a positive integer and p is not greater than Q; The extracting of subscription interest knowledge entities with multiple feature depths from the first interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the first page subscribers includes: In response to p being equal to 1, setting the first interaction preference knowledge vector set to the first-level subscription interest knowledge entity set corresponding to the first-level feature depth; In response to p being less than Q, iterating p to p+1, performing knowledge entity mining of the p+1th level feature depth on the first interaction preference knowledge vector set to obtain the p+1th level interest entity set, where p+1 is an integer greater than 1 and not greater than Q; Build entity relationships for the p+1th level interest entity set to obtain the p+1th level subscription interest knowledge entity set corresponding to the p+1th level feature depth; In response to p being equal to Q, obtaining Q subscription interest knowledge entity sets corresponding to the Q feature depths respectively; The feature depth is used to indicate the number of entities classified into the same feature cluster during the knowledge entity mining process of the corresponding feature depth, and the p-th level feature depth is used to indicate that the number of entities classified into the same feature cluster during the knowledge entity mining process of the p-th level feature depth is p.

2. The method according to claim 1, characterized in that The performing knowledge entity mining of the p+1th level feature depth on the first interaction preference knowledge vector set to obtain the p+1th level interest entity set includes: Based on the indication information of the p+1th level feature depth, the first interaction preference knowledge vector set is summarized into knowledge entities by the influence weight of each entity in the first interaction preference knowledge vector set to obtain the p+1th level interest entity set; Among them, the p+1th level interest entity set contains at least one feature cluster, and each feature cluster contains p+1 entities with a weighted order relationship; The entity relationship is constructed for the p+1th level interest entity set to obtain the p+1th level subscription interest knowledge entity set corresponding to the p+1th level feature depth, including: Associating the entities in each feature cluster in the p+1th level interest entity set with knowledge features to obtain a p+1th level associated feature entity set; The p+1th level associated feature entity set includes at least one associated feature entity combination, each associated feature entity combination is obtained by associating entities in a feature cluster in the p+1th level interest entity set; Knowledge extraction is performed on the associated feature entity combinations in the p+1th level associated feature entity set to obtain the p+1th level subscription interest knowledge entity set corresponding to the p+1th level feature depth.

3. The method according to any one of claims 1 to 2, characterized in that: The step of performing knowledge extraction on the first page interaction image set according to the image content description of each page interaction image in the first page interaction image set to obtain a first interaction preference knowledge vector set includes: Obtaining a category relationship vector between X content category tags and a plurality of image descriptors of each of the content category tags; the X content category tags include: one or more of content push annotations, page interaction events, subscription feedback comments, community graphs between page subscription users, content update nodes, and content heat values; Determine a global image content description based on the category relationship vectors between the X content category labels and a plurality of image descriptors of each of the content category labels; any image content description in the global image content description is obtained by performing a splicing operation on any image descriptor included in each of the content category labels based on the category relationship vectors between the X content category labels; generating, according to the number of the global image content descriptions, quantitative knowledge of interaction preferences of each image content description in the global image content descriptions; The image content description of each page interaction image in the first page interaction image set is converted into interaction preference quantified knowledge of the corresponding image content description to obtain a first interaction preference knowledge vector set.

4. The method according to any one of claims 1 to 2, characterized in that: Any page interaction image in the first page interaction image set is represented as a first page interaction image; and the method further includes: Obtaining a category relationship vector between X content category tags, where the X content category tags include: one or more of content push annotation, page interaction event, subscription feedback comment, community graph between page subscription users, content update node, and content heat value; Obtain an image descriptor of the first page interactive image under each of the content category tags; The image content description of the first page interactive image is obtained by integrating description features of the image descriptors under each of the content category labels through the category relationship vectors between the X content category labels.

5. The method according to claim 1, characterized in that The interest portrait generation process is implemented by calling a deep convolution algorithm with a knowledge feature enhancement function; the number of feature depths is Q, the deep convolution algorithm includes an interest portrait generation branch and Q knowledge feature enhancement branches, and one subscription interest knowledge entity set corresponds to one knowledge feature enhancement branch; Q is a positive integer; The method of using the knowledge feature enhancement strategy to generate interest portraits for the plurality of subscription interest knowledge entity sets of the first page subscriber to obtain the page subscription interest portraits includes: The Q subscription interest knowledge entity sets of the first page subscribers are respectively transferred to the corresponding knowledge feature enhancement branches to mine interest entities and interest relationships, and Q interaction preference knowledge enhancement vectors are obtained, each of which is derived from a subscription interest knowledge entity set; The Q interaction preference knowledge reinforcement vectors are fully connected and then introduced into the interest portrait generation branch to perform interest portrait generation processing to obtain a page subscription interest portrait.

6. The method according to claim 5, characterized in that Any subscription interest knowledge entity set among the Q subscription interest knowledge entity sets is represented as a p-th subscription interest knowledge entity set, and the Q interaction preference knowledge reinforcement vectors include a p-th interaction preference knowledge reinforcement vector derived from the p-th subscription interest knowledge entity set; The p-th subscription interest knowledge entity set corresponds to the p-th knowledge feature reinforcement branch among the Q knowledge feature reinforcement branches, and the p-th knowledge feature reinforcement branch includes a residual processing sub-branch and a plurality of secondary knowledge feature reinforcement layers having a cascade relationship; p is a positive integer not greater than Q; The p-th subscription interest knowledge entity set is transferred to the p-th knowledge feature enhancement branch to mine interest entities and interest relationships, and the p-th interaction preference knowledge enhancement vector is obtained, including: Inputting the p-th subscription interest knowledge entity set into the residual processing sub-branch for residual processing to obtain subscription interest residual knowledge of the p-th subscription interest knowledge entity set; The subscription interest residual knowledge of the p-th subscription interest knowledge entity set is input into the multiple secondary knowledge feature enhancement layers with cascade relationship to perform inter-class feature enhancement processing to obtain the p-th interaction preference knowledge enhancement vector.

7. The method according to claim 5, characterized in that The method further comprises: Acquire a second page interaction image set, where the second page interaction image set is obtained by sequentially arranging a plurality of page interaction images browsed by a second page subscriber within a target session period based on a page response priority; According to the image content description of each page interaction image in the second page interaction image set, knowledge extraction is performed on the second page interaction image set to obtain a second interaction preference knowledge vector set; Performing multiple feature depth subscription interest knowledge entity extraction on the second interaction preference knowledge vector set to obtain multiple subscription interest knowledge entity sets of the second page subscribers; Calling the original deep convolution algorithm to perform interest portrait generation processing on multiple subscription interest knowledge entity sets of the second page subscriber to obtain a page subscription interest training portrait, wherein the page subscription interest training portrait is used to characterize the interest characteristics of the second page subscriber in the process of browsing the page interaction image; Based on the page subscription interest training portrait and the prior debugging annotations of the second page subscription user, debugging the original deep convolution algorithm to obtain the deep convolution algorithm with knowledge feature enhancement function; The method of debugging the original deep convolution algorithm based on the page subscription interest training portrait and the prior debugging annotation of the second page subscription user to obtain the deep convolution algorithm with the knowledge feature enhancement function includes: Determining an algorithm training error of the original deep convolution algorithm based on the page subscription interest training portrait and the prior debugging annotations of the second page subscription user; Towards the goal of converging the algorithm training error, the algorithm variable weights of the original deep convolution algorithm are optimized to obtain the deep convolution algorithm with knowledge feature enhancement function; Among them, when the second page subscriber is updated as an active page subscriber in the targeted push server, the prior debugging annotation of the second page subscriber is a targeted push annotation; when the second page subscriber is not updated as an active page subscriber in the targeted push server, the prior debugging annotation of the second page subscriber is a non-targeted push annotation.

8. The method according to claim 1, characterized in that The method further comprises: Based on the page subscription interest portrait, push targeted information to the first page subscriber; The first page subscriber is any page subscriber among authorized page subscribers of the directional push server; and the directional information push to the first page subscriber based on the page subscription interest portrait includes: If the page subscription interest portrait indicates that the interest characteristics of the first page subscriber in the process of browsing the page interactive image meet the directional push condition, then the first page subscriber is updated as an active page subscriber in the directional push server and an interest subscription message is pushed according to the interest characteristics of the first page subscriber in the process of browsing the page interactive image; Alternatively, statistics are collected in the directional push server on the directional push strategy for the subscribers of the first page; Alternatively, the directional push event link of the first page subscriber is cached in the directional push server.

9. An interest portrait generation server, characterized in that: It comprises a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program implements the method according to any one of claims 1 to 8 when running.

Images