Reading recommendation method and system based on artificial intelligence

Abstract

The invention relates to the field of artificial intelligence and big data analysis, and discloses a reading recommendation method and system based on artificial intelligence, and the method comprises the steps: obtaining behavior data and reading data of a user; a diversified recommendation page is set, a diversified page prediction recommendation strategy is executed, and the interest direction of a new user is preliminarily predicted; setting a global recommendation model, and distributing parameters of the global recommendation model to each platform; executing a cross-platform federated learning strategy, and performing noise processing on the parameters by using a differential privacy algorithm; after training is completed, model updating parameters are uploaded, a secure multi-party computing strategy is executed, and a local recommendation model is updated; the average weight of the updated global recommendation model is calculated, and new global recommendation model parameters are generated; and executing a real-time dynamic feedback adjustment strategy according to the feedback of the user on the recommended content, and adjusting the personalized recommended content in real time. The method has remarkable advantages in the aspects of improving the recommendation effect, protecting the privacy of the user, enhancing the participation degree of the user and the like.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and big data analysis, and specifically to a reading recommendation method and system based on artificial intelligence. Background Art

[0002] The reading recommendation method based on artificial intelligence can provide personalized content recommendations through accurate user behavior analysis and data mining. This method can not only improve the user's reading experience, but also effectively increase the interaction between the user and the platform and extend the user's reading time. By analyzing the user's behavior data (such as page stay time, scrolling behavior, and click preference), the system can understand the user's interests and preferences, so as to recommend the most relevant content for them. In addition, using guiding questions to further optimize interest prediction makes the recommendation results more in line with the actual needs of users, thereby improving user satisfaction and loyalty.

[0003] Existing reading recommendation systems often rely on simple collaborative filtering or content-based recommendation algorithms. These methods often lack effective data support when facing new users, resulting in poor recommendation effects. In addition, many systems fail to make full use of cross-platform data resources, and there is a problem of data islands, which limits the learning and optimization capabilities of the model. At the same time, the issue of user privacy protection has become increasingly prominent, and traditional model training methods are prone to exposing users' personal information. Therefore, the existing technologies have obvious defects in the accuracy of personalized recommendation, the flexibility of cross-platform learning, and the effectiveness of privacy protection.

[0004] This solution overcomes multiple deficiencies of traditional recommendation systems by introducing a cross-platform federated learning strategy, and has significant advantages in improving the recommendation effect, protecting user privacy, and enhancing user participation. Summary of the Invention

[0005] The present invention provides a reading recommendation method and system based on artificial intelligence, which promotes the solution of the problems mentioned in the above background art.

[0006] In the first aspect, the present application provides a reading recommendation method based on artificial intelligence, adopting the following technical solution: A reading recommendation method based on artificial intelligence, comprising: Obtain the user's behavior data and reading data, where the behavior data includes page stay time, scrolling behavior, and click preference; determine whether the user is a new user; If the user is a new user, set a diversified recommendation page, execute a diversified page prediction recommendation strategy, calculate the interest weights of the new user for different categories of content, and initially predict the interest direction of the new user; Optimize the predicted interest direction using guiding questions and perform content recommendation; If the user is an old user, a global recommendation model is established based on the old user's historical behavior data and reading data combined with the current behavior data and reading data, and the parameters of the global recommendation model are distributed to each platform; Implement a cross-platform federated learning strategy, access local data on different platforms to independently train local recommendation models, use differential privacy algorithms, and perform noise processing on parameters during training; After training is completed, the model is uploaded to update parameters, a secure multi-party computing strategy is executed, the updated parameters of each platform are aggregated and calculated, and the local recommendation model is updated; Calculate the average weight of the updated global recommendation model and generate new global recommendation model parameters; Generate a personalized recommendation list, calculate the interest matching degree, and sort the recommended contents in the personalized recommendation list according to the interest matching degree; Based on user feedback on recommended content, long-term and short-term interest models are set to optimize recommended content, implement real-time dynamic feedback adjustment strategies, and adjust personalized recommended content in real time.

[0007] Through the AI-based reading recommendation method described in this paper, the user experience has been greatly improved. First, this method can accurately obtain the user's behavior data and reading data, including page dwell time, scrolling behavior and click preferences, so as to fully understand the user's reading habits and interests. This data-driven approach enables the recommendation system to make recommendations based on actual behavior rather than simple historical records or static information, ensuring the relevance and timeliness of the recommended content. Secondly, through different processing strategies for new and old users, the system can flexibly respond to user changes. The diversified recommendations and optimization of guided questions for new users can quickly establish the user's interest map, while the cross-platform federated learning strategy for old users ensures the user's consistent experience across different platforms and prevents the emergence of information islands. Finally, the implementation of the real-time dynamic feedback adjustment strategy enables the recommended content to be optimized according to the user's immediate response, improving user satisfaction and engagement. This series of advantages makes the entire recommendation system not only personalized, but also intelligent, adaptable and efficient, greatly improving the user's reading experience.

[0008] Preferably, the setting of diversified recommendation pages, executing diversified page prediction recommendation strategies, and calculating the interest weights of new users for different categories of content include: Get user behavior data D U and read data to determine whether the reading data is empty; When the user's reading data is empty, the user is a new user U1; Set the diversified recommendation page to contain different categories P, P = {p1, p2, ..., p n}, where pi It represents the i-th page within a page. n represents the total number of pages, and there is only one type of content within a page; Obtain the residence time T, scrolling behavior S, and click preference C of the new user U1 on page p i Execute a diversified page prediction and recommendation strategy, calculate the interest weight I(U1, p i ) of the new user U1 for different types of content, I(U1, p i ) = ω1·T(U1, p i ) + ω2·S(U1, p i ) + ω3·C(U1, p i ), where ω1, ω2, and ω3 are weight coefficients, and T(U1, p i ), S(U1, p i ), C(U1, p i ) are respectively the residence time, scrolling behavior, and click preference of the new user U on page p i ; According to the interest weight I(U1, p i ), predict the interest direction of the new user U1 Obtain the answer q to the guiding question by asking the new user U1 about their interest preferences through guiding questions U , and optimize the predicted interest direction ; Based on the optimized predicted interest direction, perform content recommendation.

[0009] By adopting the methods of diversified recommendation pages and guiding questions, it greatly facilitates users to find the content they are interested in when they first come into contact. This strategy can not only effectively judge users' interest preferences but also reduce the selection difficulties that new users may encounter on the platform through rich category selections. By obtaining the residence time, scrolling behavior, and click preference on different pages, the system can establish a preliminary interest weight model for new users, which lays a solid foundation for subsequent personalized recommendations. In addition, when the reading data of users is empty, the system can quickly identify them as new users and provide targeted recommendations. This timeliness and adaptability are incomparable to traditional recommendation methods. The application of guiding questions can further optimize interest prediction through interaction with users, making the recommended content more in line with users' real needs. This method not only improves users' initial experience but also guarantees users' subsequent activity and retention rate.

[0010] Preferably, the execution of the cross-platform federated learning strategy, independently training local recommendation models by accessing local data on different platforms, includes: Obtain the behavior data and reading data of users, and judge whether the reading data is empty; When the user's reading data is not empty, the user is an old user U2; It is assumed that there are N different platforms, and each platform maintains local user data D i2 ; According to the historical behavior data and reading data of the old user U2, combined with the current behavior data and reading data, a global recommendation model f is set θ , and the global recommendation model is a shared model trained with cross-platform data of all old users U2. θ represents the parameters of the global recommendation model, and the predicted score of the interests of the old user U2 is calculated Distribute the parameters θ of the global recommendation model to each platform, execute the cross-platform federated learning strategy, and independently train the local recommendation model by accessing the local data on different platforms. The local recommendation model is a data training model for a single platform; On the platform , obtain the user's real reading feedback Calculate the loss function Calculate the local recommendation model parameters for the t-th round of training where θ t is the initial parameter for the t-th round of training of the global recommendation model, η is the learning rate, is the platform on the gradient; Use the differential privacy algorithm to process the parameters with noise during the training process to obtain the updated local recommendation model parameters where, is Gaussian noise.

[0011] By adopting the cross-platform federated learning strategy, the behaviors of users on different platforms can be effectively integrated to form a global recommendation model. This strategy makes full use of the historical behavior data and current behavior data of old users, ensuring the accuracy and personalization of the recommended content. In addition, by using the differential privacy algorithm, the system ensures the security of user privacy during data training, avoiding the risk of data leakage, which is particularly important in the context of high attention to data security today. The recommendation strategy for old users also ensures the continuous optimization and improvement of the recommendation model by obtaining the user's feedback in real time and calculating the loss function. This dynamic adjustment mechanism can not only quickly respond to the changing needs of users, but also continuously improve the intelligence level of the recommendation system. Through this integrated strategy, users can enjoy a consistent and personalized reading experience among different platforms, further enhancing user loyalty and activity.

[0012] Preferably, after the training is completed, the model updates the parameters and executes the secure multi-party computing strategy to aggregate and calculate the updated parameters of each platform, including: Sending the updated local recommendation model parameters to the server, executing the secure multi-party computing strategy, and calculating the average weight ω of the updated global recommendation model i1 , Splitting into multiple parts and sending the amounts s k to multiple computing parties M, and encrypting the multiple amounts s k ; Aggregating and calculating the updated global recommendation model parameters θ of each platform g , Sending the updated global recommendation model parameters θ g to each platform to update the local recommendation model of each platform.

[0013] By implementing the secure multi-party computing strategy, the parameter update and aggregation calculation between different platforms can be smoothly carried out on the premise of protecting user privacy. By sending the updated local recommendation model parameters to the server, the system can effectively calculate the average weight of the updated global recommendation model, thereby improving the accuracy and efficiency of recommendations. In addition, the encryption of parameters ensures that even during data transmission, user information will not be leaked, enhancing data security. The distributed computing method in this process not only improves the computing efficiency of the system but also supports the joint learning and optimization of multi-party data, enabling the recommendation models of each platform to utilize the behavior data of different user groups and accelerating the iterative update of the algorithm. This method effectively solves the data island problem faced by traditional centralized recommendation systems, realizes true cross-platform collaboration and user experience unification, and further improves the intelligent level of the entire recommendation system.

[0014] Preferably, for generating the personalized recommendation list, calculating the interest matching degree, and sorting the recommended content in the personalized recommendation list according to the interest matching degree, it includes: Setting the interest vector V of user U u , V u ={v u1 , v u2 , …, v uG}, where G represents the dimension of the interest vector, and v ua represents the preference degree of user U for the interest dimension a; Setting the feature vector V of the category content P p , V p ={v p1 , v p2 , …, v pG}, where v paRepresents the eigenvalue of category content P on the interest dimension i; Calculate the norm of the user interest vector Calculate the norm of the category content feature vector Calculate the interest matching degree of user U to category content P For all category contents P' = {p'1, p'2,..., p' n′} in the predicted interest direction, sort them in descending order according to the interest matching degree W(U, P); Show the top q contents in the sorted list of interest matching degree W(U, P) to the user as personalized recommended content.

[0015] By calculating the interest matching degree, the user's interest vector is effectively compared with the content's feature vector, thus generating recommended content that best meets the user's needs. This method ensures the relevance of the recommended content, enabling users to find interesting articles or books faster. In addition, by calculating the norm of the interest matching degree, the system can quantify the user's preference degree for different category contents, providing strong data support for personalized recommendation. By sorting all category contents according to the interest matching degree, users can easily obtain the content that best matches their current interests. This efficient recommendation mechanism significantly improves the user experience and satisfaction. The real-time update of the personalized recommendation list combined with the user feedback mechanism enables the system to flexibly respond to changes in user preferences, ensuring that the recommended content always remains consistent with the user's needs. This personalized recommendation strategy based on dynamic calculation brings a more intuitive and convenient reading experience to users, and also enhances the user stickiness of the platform.

[0016] Preferably, the long-term interest and short-term interest models are set to optimize the recommended content, and a real-time dynamic feedback adjustment strategy is executed to adjust the personalized recommended content in real time, including: Calculate the feedback score W(U, P) of the user for the personalized recommended content, W(U, P) = α·T(U, P) + β·S(U, P) + γ·C(U, P), where α, β, and γ are parameters affecting the adjustment behavior; Obtain the set Y1(U) = {y1, y2,..., y n′} of the user U's historical reading content, where y i4 represents the articles or books read by the user, and n' represents the total number of n' articles or books read by the user; Calculate the historical residence time weight according to the user's historical reading content set Y(U) where ω t is the time decay factor, and T(U, Y, t) represents the residence time of the user reading content y i4 at time t; Calculate the historical reading category preference vector F(U), Based on the user's historical behavior data, a long-term interest model L(U) representing the user's long-term and stable interest preferences is set up. where Z1 is a normalization factor; Obtain the set of the user U's current reading content Y2(U) = {y t-1 , y t-2 , …, y t-n″}, where y i4-1 represents the article or book that the user is currently reading, and n′ represents the number of articles or books that the user is currently reading. Calculate the weighted score R(U, y i4-1 ) of the current reading data, R(U, y i4-1 ) = α·T(U, y i4-1 ) + β·S(U, y i4-1 ) + γ·C(U, y i4-1 ), where α, β, and γ are parameters for adjusting the behavior influence. Based on the weighted score of the user's current reading data, a short-term interest model L′(U) representing the user's recent interest preferences is set up. where Z2 is a normalization factor; Perform weighted fusion on the long-term interest model and the short-term interest model, and calculate the final personalized interest model L * (U) = λL(U) + (1 - λ)L′(U), where λ ∈ [0, 1] is a weight coefficient; When λ takes a larger value, the influence of the long-term interest is stronger, which is suitable for users with stable interests; When λ takes a smaller value, the influence of the short-term interest is stronger, which is suitable for users with faster-changing interests; Optimize and generate personalized recommended content according to the long-term interest model and the short-term interest model, monitor the user's real-time reading data, and adjust the personalized recommended content in real time, L * (U)′ = L * (U) + δ·W(U, P), where δ is the update step length for controlling the adjustment amplitude.

[0017] By setting up long-term and short-term interest models, this method can effectively capture the changes and stability of users' interests. The design of this two-layer interest model enables the recommendation system to perfectly integrate the long-term stable interest preferences and the recent dynamic interest manifestations based on the users' historical behavior data and current reading behavior. When the long-term interest of the user is strong, the system can provide recommended content that better matches the user's core interests; when the user's interests are more changeable, the system can quickly adapt and provide the latest recommendations. By continuously monitoring the users' reading data and dynamically adjusting the personalized recommendation content, this flexible adjustment mechanism enables users to experience a higher level of personalized service during use and reduces the possibility of information overload. In addition, the feedback score and weighted fusion mechanisms in this method ensure that the recommendation effect of the system is a reaction based on actual user behavior, greatly improving the accuracy and satisfaction of the recommendations. Overall, this optimization strategy that combines long-term and short-term interests enables the recommendation system to maintain a high level of service quality even when the users' needs change, thereby enhancing user loyalty and activity.

[0018] In a second aspect, the present application provides a reading recommendation system based on artificial intelligence, adopting the following technical solutions: A reading recommendation system based on artificial intelligence, comprising: Data collection module: Collects the behavior data of users on the platform, including page stay time, scrolling behavior, click preferences, etc., and records the reading historical data of users; User classification module: Judges whether the user is a new user or an old user according to whether the reading data of the user is empty, and extracts the behavior characteristics of new users and old users; Recommendation model construction module: Sets diverse recommendation pages for new users, calculates the interest weights through the behavior data of users, optimizes the predicted interest direction by using guiding questions, establishes a global recommendation model based on the historical behavior data and current data of old users, and distributes the parameters to each local platform; Cross-platform federated learning module: Independently trains local recommendation models on each different platform, executes cross-platform federated learning strategies, and applies differential privacy algorithms during the training process to add noise to the model parameters; Model update and aggregation module: Uploads the updated local model parameters to the server, performs secure multi-party computing to aggregate and calculate the global model parameters, sends the updated global recommendation model parameters to each platform, and updates the local recommendation models; Recommended content generation module: Calculates the matching degree between the interest vector of the user and the category content feature vector, generates personalized recommended content, sorts the personalized recommendation list according to the interest matching degree, and presents it to the user; Real-time Feedback and Adjustment Module: Collects user feedback on recommended content, calculates feedback scores, and dynamically adjusts the user's long-term and short-term interest models based on the user's historical reading and current behavior data, and adjusts the recommended content in real time; Monitoring and Analysis Module: Monitors the user's reading behavior in real time and adjusts personalized recommendations in a timely manner; User Interface Module: Designs a user-friendly interface, displays personalized recommended content, and provides a user feedback entry.

[0019] The present invention has the following beneficial effects: 1. For the method of reading recommendation based on artificial intelligence, through the implementation of a method of reading recommendation based on artificial intelligence, the user experience has been significantly improved. By comprehensively obtaining the user's behavior data and reading data, including page stay time, scrolling behavior, and click preferences, this method can deeply understand the user's real reading habits and interests. Different from the traditional static recommendation method, this data-driven strategy ensures the relevance and timeliness of the recommended content. In addition, the different processing strategies for new users and old users enable the system to flexibly respond to changes in user needs. For new users, the system provides diverse recommendation pages and optimizes the user's interest prediction through guiding questions, thereby quickly establishing the user's interest map. For old users, the cross-platform federated learning strategy ensures the consistency of the user experience across different platforms and avoids the emergence of information silos. More importantly, the implementation of the real-time dynamic feedback adjustment strategy enables the recommended content to be optimized according to the user's immediate reaction, significantly improving the user's satisfaction and engagement. This comprehensive approach gives the recommendation system significant advantages in personalization, intelligence, and efficiency, greatly enhancing the user's reading experience and promoting user activity and retention.

[0020] 2. The method of reading recommendation based on artificial intelligence can effectively improve the user experience when new users first come into contact with the platform through the strategies of diversified recommendation pages and guiding questions for new users. This strategy not only helps the system quickly judge the user's interest preferences, but also reduces the selection confusion that new users may encounter on the platform by providing a rich variety of category options. When the user's reading data is empty, the system can quickly identify the user as a new user and provide targeted recommendations. This timeliness and adaptability are impossible to achieve by traditional recommendation systems. By obtaining the user's stay time, scrolling behavior, and click preferences on different pages, the system can establish a preliminary interest weight model for the user, which lays a good foundation for subsequent personalized recommendations. The application of guiding questions further optimizes the prediction of the user's interests, making the recommended content more in line with the user's real needs. This strategy not only improves the user's initial experience, but also ensures the user's activity and retention rate, thereby enhancing the attractiveness and competitiveness of the platform. Through these measures, new users can find the content they are interested in faster, improving their satisfaction and loyalty to the platform.

[0021] 3. The method of reading recommendation based on artificial intelligence can effectively integrate the user's behavior data on different platforms to form a global recommendation model by implementing a cross-platform federated learning strategy for old users. This strategy makes full use of the old user's historical behavior and current activity data to ensure the accuracy and personalization of the recommended content. At the same time, the differential privacy algorithm is used to ensure the user's privacy security during the data training process, preventing potential data leakage risks, which is particularly important in the context of increasing attention to data protection. In addition, the recommendation strategy for old users continuously optimizes and improves the recommendation model by collecting the user's feedback in real time and calculating the loss function. The dynamic adjustment mechanism enables the system to quickly respond to changes in user needs and continuously improve the intelligence level of the recommendation system. In this way, users can enjoy a consistent and personalized reading experience across different platforms, which not only enhances the user's loyalty but also improves the user activity of the platform. In summary, this comprehensive data processing and optimization strategy ensures the efficiency and intelligence of the recommendation system when facing old users, further improving the user's satisfaction and usage experience.

[0022] 4. The method of reading recommendation based on artificial intelligence enables the parameter update and aggregation calculation between different platforms to proceed smoothly while ensuring user privacy through the implementation of secure multi-party computing strategies. This process allows the updated local recommendation model parameters to be securely transmitted to the server, effectively calculating the average weights of the updated global recommendation model and enhancing the accuracy and efficiency of the recommendation. Encryption processing ensures that user information will not be leaked even during data transmission, enhancing data security. By splitting the updated parameters and sending them to multiple computing parties for encryption processing, the system realizes data sharing while avoiding potential privacy risks. This distributed computing method not only improves the system's computing efficiency but also supports collaborative learning between different platforms, enabling the recommendation models of each platform to be optimized using extensive user behavior data. This method effectively solves the data silo problem faced by traditional centralized recommendation systems, achieves true cross-platform collaboration and the unification of user experience, thereby enhancing the intelligence level of the entire recommendation system. In short, the implementation of secure multi-party computing strategies not only guarantees the security of user data but also creates conditions for the continuous optimization of the recommendation system, promoting the further improvement of user experience.

[0023] 5. The method of reading recommendation based on artificial intelligence can generate a personalized recommendation list by calculating the user interest matching degree, effectively comparing the user's interest vector with the feature vector of the content to ensure the relevance of the recommended content. This method enables users to find interesting articles or books faster, enhancing the overall usage satisfaction. By quantifying the user's preference degree for different categories of content, the system can provide strong data support for personalized recommendations. Sorting all categories of content according to the interest matching degree allows users to easily obtain the content that best meets their needs, reducing the time consumption when browsing on the platform. The combination of real-time updating of the recommendation list and the user feedback mechanism enables the system to flexibly respond to changes in user preferences, ensuring that the recommended content always aligns with user needs. This efficient recommendation mechanism significantly improves the user experience and promotes user activity. The personalized recommendation list not only provides users with an intuitive reading experience but also increases the user stickiness of the platform, enhancing users' trust and loyalty to the platform. In summary, the generation strategy of the personalized recommendation list provides users with personalized and accurate content recommendations, further enhancing user satisfaction and participation.

[0024] 6. The method of reading recommendation based on artificial intelligence can effectively capture the interest changes and stability of users by setting long-term and short-term interest models, so as to provide more accurate recommendation services. The design of this double-layer interest model enables the recommendation system to perfectly integrate long-term interest preferences and recent dynamic interest manifestations based on the user's historical behavior data and current reading behavior. When the user's long-term interest is strong, the system can provide recommended content that conforms to the user's core interests; when the user's interests change rapidly, the system can quickly adapt and provide the latest recommendations. By real-time monitoring the user's reading data, the system can dynamically adjust personalized recommended content. This flexible adjustment mechanism ensures the continuous optimization of the user experience and reduces the risk of information overload. In addition, the feedback score and weighted fusion mechanism in this method ensure that the recommendation effect is based on the response of actual user behavior, greatly improving the accuracy and satisfaction of the recommendation. Overall, the combined optimization strategy of long-term and short-term interest models not only allows the system to adapt to changes in user needs, but also finds a balance between the diversity and accuracy of recommended content, thereby enhancing user loyalty and activity and improving the overall effect of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic flow diagram of the method of the present invention.

[0026] Figure 2 It is a schematic diagram of the system module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] Example 1. Refer to Figure 1 , a method of reading recommendation based on artificial intelligence, including: Obtain the user's behavior data and reading data, where the behavior data includes page stay time, scrolling behavior, and click preference; determine whether the user is a new user; If the user is a new user, set a diversified recommendation page, execute a diversified page prediction recommendation strategy, calculate the interest weights of the new user for different categories of content, and initially predict the interest direction of the new user; Optimize the predicted interest direction using guiding questions and perform content recommendation; If the user is an old user, based on the historical behavior data and reading data of the old user, combined with the current behavior data and reading data, a global recommendation model is established, and the parameters of the global recommendation model are distributed to each platform; Execute the cross-platform federated learning strategy, independently train the local recommendation model by accessing local data on different platforms, and use the differential privacy algorithm to process noise for the parameters during the training process; After training is completed, upload the model to update the parameters, execute the secure multi-party computing strategy, aggregate and calculate the updated parameters of each platform, and update the local recommendation model; Calculate the average weight of the updated global recommendation model to generate new global recommendation model parameters; Generate a personalized recommendation list, calculate the interest matching degree, and sort the recommended content in the personalized recommendation list according to the interest matching degree; Based on the user's feedback on the recommended content, set up long-term and short-term interest models to optimize the recommended content, execute the real-time dynamic feedback adjustment strategy, and adjust the personalized recommended content in real time.

[0029] By adopting a reading recommendation method based on artificial intelligence, the user experience is significantly improved. This method first comprehensively obtains the user's behavior data and reading data, including page stay time, scrolling behavior, and click preferences, and deeply analyzes the user's reading habits and interests. This data-driven approach is different from traditional static recommendation systems, as it can update the user's interest profile in real time to ensure the relevance and timeliness of the recommended content. Especially in the processing strategies for new users and old users, this method demonstrates a high degree of flexibility and adaptability. For new users, the system can quickly identify the user's interest preferences and provide personalized recommended content by means of diversified recommendation pages and guiding questions. In contrast, for old users, a cross-platform federated learning strategy is adopted to ensure a consistent experience across different platforms and avoid user loss due to information isolation. More importantly, the implementation of the real-time dynamic feedback adjustment strategy enables the recommended content to be optimized according to the user's immediate behavior, enhancing the user's satisfaction and engagement. In summary, this comprehensive method has obvious advantages in personalization, intelligence, and efficiency, greatly enhancing the user's reading experience, promoting user activity and retention, and creating greater value for the platform.

[0030] Set up diversified recommendation pages, execute the diversified page prediction recommendation strategy, and calculate the interest weights of new users for different categories of content, including: Obtain the user's behavior data D U And reading data, and determine whether the reading data is empty; When the user's reading data is empty, the user is a new user U1; Set the diversified recommendation page to contain different categories P, P = {p1, p2, …, p n}, where p i represents the i-th page in the page, n represents a total of n pages, and there is only one type of content in one page; Obtain the residence time T, scrolling behavior S, and click preference C of the new user U1 on the p i page, execute the diversified page prediction recommendation strategy, and calculate the interest weight I(U1, p i ) of the new user U1 for different types of content. I(U1, p i ) = ω1·T(U1, p i ) + ω2·S(U1, p i ) + ω3·C(U1, p i ), where ω1, ω2, and ω3 are weight coefficients, and T(U1, p i ), S(U1, p i ), C(U1, p i ) are the residence time, scrolling behavior, and click preference of the new user U on the page p i respectively; According to the interest weight I(U1, p i ), predict the interest direction of the new user U1 Obtain the answer q U to the guided question by asking the new user U1 about their interest preferences, and optimize the predicted interest direction ; According to the optimized predicted interest direction, perform content recommendation.

[0031] By designing a diversified recommendation page and guided questions for new users, this method effectively improves the user experience when new users first contact the platform. This strategy not only quickly helps the system identify the user's interest preferences but also reduces the choice confusion that users may face by providing content selections of different categories. When the user's reading data is empty, the system can quickly determine that it is a new user and provide targeted recommendations, thereby increasing the new user's sense of participation. By analyzing the residence time, scrolling behavior, and click preference of new users on different pages, the system can establish a preliminary interest weight model, laying a good foundation for subsequent personalized recommendations. In addition, the use of guided questions further optimizes the interest prediction, making the recommended content more in line with the user's needs. This strategy for new users not only improves their initial use experience but also provides strong support for subsequent activity and retention rates. While effectively reducing the user churn rate, it enhances the user's trust and loyalty to the platform, enabling new users to integrate into the platform faster and enjoy personalized services.

[0032] Execute a cross - platform federated learning strategy to independently train local recommendation models by accessing local data on different platforms, including: obtaining the user's behavior data and reading data, and determining whether the reading data is empty; When the user's reading data is not empty, this user is an old user U2; Suppose there are N different platforms, and each platform maintains local user data D i2 ; Based on the historical behavior data and reading data of the old user U2, combined with the current behavior data and reading data, set the global recommendation model as f θ , where the global recommendation model is a shared model trained with cross - platform data of all old users U2, θ represents the parameters of the global recommendation model, and calculate the predicted score of the interests of the old user U2 Distribute the parameters θ of the global recommendation model to each platform, execute the cross - platform federated learning strategy, and independently train local recommendation models by accessing local data on different platforms. The local recommendation model is a data - training model for a single platform; On the platform , obtain the user's real reading feedback Calculate the loss function Calculate the local recommendation model parameters for the t - th round of training where θ t is the initial parameter for the t - th round of training of the global recommendation model, η is the learning rate, is the gradient on the platform ; Use the differential privacy algorithm to process noise on the parameters during training to obtain the updated local recommendation model parameters where, is Gaussian noise.

[0033] By implementing a cross-platform federated learning strategy, this method demonstrates significant advantages in optimizing recommendations for old users. This strategy effectively integrates the behavioral data of users on different platforms to form a global recommendation model, ensuring the accuracy and personalization of recommended content. Compared with traditional centralized recommendation systems, by leveraging the historical behavioral data and current activity information of old users, the system can provide content that better meets the needs of users. At the same time, the differential privacy algorithm is adopted to protect user privacy, optimizing the recommendation effect while ensuring data security, which is particularly important in the current environment where data protection is increasingly emphasized. The recommendation strategy for old users continuously optimizes the recommendation model by real-time monitoring of user feedback and calculating the loss function. This dynamic adjustment mechanism enables the system to quickly respond to changes in user needs and continuously improve the overall intelligence level of the system. Ultimately, users enjoy a consistent and personalized reading experience across different platforms, thereby enhancing user loyalty and usage frequency. This series of measures not only increases the user activity on the platform but also provides a foundation for creating a good user ecosystem.

[0034] After training is completed, upload the model to update the parameters, execute the secure multi-party computing strategy, and aggregate and calculate the updated parameters of each platform, including: The updated local recommendation model parameters are sent to the server, and the secure multi-party computing strategy is executed to calculate the average weight ω of the updated global recommendation model i1 , Split into multiple copies and send them as s k to multiple computing parties M, and encrypt the multiple copies of s k . Aggregate and calculate the updated global recommendation model parameters θ of each platform g , Send the updated global recommendation model parameters θ g to each platform to update the local recommendation model of each platform.

[0035] By implementing the secure multi-party computing strategy, this method has successfully achieved parameter updates and aggregated calculations between different platforms while ensuring the security of user privacy. During this process, the updated local recommendation model parameters are securely transmitted to the server, and then the average weight of the updated global recommendation model is calculated, significantly improving the accuracy and efficiency of recommendations. Encryption processing ensures the security of user information during data transmission and prevents potential data leakage risks. By splitting the updated parameters and sending them to multiple computing parties, the system realizes data sharing while reducing the possibility of privacy leakage. This distributed computing method not only improves computing efficiency but also promotes collaborative learning between different platforms, enabling the recommendation models of each platform to be optimized using a wider range of user behavior data. Thus, the secure multi-party computing strategy effectively solves the data silo problem faced by traditional centralized recommendation systems and achieves cross-platform collaboration and unified user experience. All of this ultimately enhances the intelligence level of the entire recommendation system and provides users with a safer and better user experience.

[0036] Generate a personalized recommendation list, calculate the interest matching degree, and sort the recommended content in the personalized recommendation list according to the interest matching degree, including: Set the interest vector V of user U u , V u = {v u1 , v u2 , …, v uG}, where G represents the dimension of the interest vector, and v ua represents the preference degree of user U for the interest dimension a; Set the feature vector V of category content P p , V p = {v p1 , v p2 , …, v pG}, where v pa represents the eigenvalue of category content P in the interest dimension i; Calculate the norm of the user interest vector Calculate the norm of the category content feature vector Calculate the interest matching degree of user U for category content P Sort all the category content P′ = {p′1, p′2, …, p′ n′} in the predicted interest direction in descending order according to the interest matching degree W(U, P); Present the top q contents of the interest matching degree W(U, P) sorted list to the user as personalized recommendation content.

[0037] By precisely calculating the user interest matching degree, this method can generate a personalized recommendation list, effectively compare the user's interest vector with the content feature vector, so as to ensure a high degree of relevance of the recommended content. This strategy enables users to quickly find articles or books of interest, significantly improving the usage satisfaction. By quantifying the user's preferences for different categories of content, the system provides strong data support for personalized recommendations. Sorting all categories of content according to the interest matching degree allows users to easily obtain the content that best meets their needs, reducing the time cost during browsing. In addition, the combination of the real-time updated recommendation list and the user feedback mechanism enables the system to flexibly respond to changes in user preferences, ensuring that the recommended content always remains consistent with user needs. This efficient recommendation mechanism not only improves the user experience, but also promotes user activity and enhances the user stickiness of the platform. All in all, the generation strategy of the personalized recommendation list provides users with intuitive and accurate content recommendations, further improving user satisfaction and participation, and strengthening users' trust in the platform.

[0038] Set long-term and short-term interest models to optimize recommended content, implement a real-time dynamic feedback adjustment strategy, and adjust personalized recommended content in real time, including: Calculate the feedback score W(U, P) of the user for the personalized recommended content, W(U, P) = α·T(U, P) + β·S(U, P) + γ·C(U, P), where α, β, and γ are parameters that adjust the influence of behavior; Obtain the set Y1(U) = {y1, y2,..., y n′} of the user U's historical reading content, where y i4 represents the articles or books read by the user, and n′ represents the total number of n′ articles or books read by the user; According to the set Y(U) of the user's historical reading content, calculate the historical residence time weight where ω t is the time decay factor, and T(U, Y, t) represents the residence time of the user reading content y i4 at time t; Calculate the historical reading category preference vector F(U), According to the user's historical behavior data, set a long-term interest model L(U) representing the user's long-term and stable interest preferences, where Z1 is the normalization factor; Obtain the set Y2(U) = {y t-1 , y t-2 ,..., y t-n″} of the user U's current reading content, where y i4-1 represents the articles or books that the user is currently reading, and n′' represents the number of n′' articles or books that the user is currently reading; Calculate the weighted score R(U, y i4-1 ), R(U, y i4-1 ) = α·T(U, y i4-1 ) + β·S(U, y i4-1 ) + γ·C(U, y i4-1 ), where α, β, and γ are parameters for adjusting the influence of behavior; Based on the weighted score of the user's current reading data, set a short-term interest model L′(U) representing the user's recent interest preferences, where Z2 is a normalization factor; Perform weighted fusion of the long-term interest model and the short-term interest model, and calculate the final personalized interest model L * (U) = λL(U) + (1 - λ)L′(U), where λ ∈ [0, 1] is a weight coefficient; When λ takes a larger value, the influence of long-term interest is stronger, which is suitable for users with stable interests; When λ takes a smaller value, the influence of short-term interest is stronger, which is suitable for users with faster-changing interests; Optimize and generate personalized recommendation content based on the long-term interest model and the short-term interest model, monitor the user's real-time reading data, and adjust the personalized recommendation content in real time, L * (U)′ = L * (U) + δ·W(U, P), where δ is the update step size for controlling the adjustment amplitude.

[0039] By setting the long-term interest and short-term interest models, this method can effectively capture the changes and stability of the user's interests, thereby providing more accurate recommendation services. The design of this two-layer interest model enables the recommendation system to combine the user's historical behavior data with the current reading behavior, perfectly integrating the long-term interest preferences and the recent dynamic interest performance. When the user's long-term interest is strong, the system can recommend content that conforms to the user's core interests; when the user's interests change rapidly, the system can quickly adapt and provide the latest recommendations. By real-time monitoring the user's reading data, the system dynamically adjusts the personalized recommendation content to ensure the continuous optimization of the user experience and reduce the risk of information overload. In addition, the feedback score and weighted fusion mechanism in this method ensure that the recommendation effect is based on the response of the actual user behavior, further improving the accuracy and satisfaction of the recommendation. Overall, the combined optimization strategy of the long-term and short-term interest models enables the recommendation system to maintain a high service quality when the user's needs change, enhance the user's loyalty and activity, and improve the overall effect of the recommendation system.

[0040] Example 2, refer to Figure 2 , a reading recommendation system based on artificial intelligence, including: Data collection module: Collects users' behavioral data on the platform, including page stay time, scrolling behavior, click preferences, etc., and records users' reading history data; User classification module: Determines whether a user is a new user or an old user based on whether the user's reading data is empty, and extracts the behavioral characteristics of new users and old users; Recommendation model construction module: Sets diverse recommendation pages for new users, calculates interest weights through users' behavioral data, optimizes the predicted interest direction using guiding questions, builds a global recommendation model based on old users' historical behavior data and current data, and distributes the parameters to each local platform; Cross-platform federated learning module: Independently trains local recommendation models on various different platforms, executes cross-platform federated learning strategies, and applies differential privacy algorithms during the training process to add noise to the model parameters; Model update and aggregation module: Uploads the updated local model parameters to the server, performs secure multi-party computing to aggregate and calculate the global model parameters, sends the updated global recommendation model parameters to each platform, and updates the local recommendation models; Recommended content generation module: Calculates the matching degree between the user's interest vector and the category content feature vector, generates personalized recommended content, sorts the personalized recommendation list according to the interest matching degree, and presents it to the user; Real-time feedback and adjustment module: Collects users' feedback on the recommended content, calculates the feedback score, dynamically adjusts the user's long-term and short-term interest models based on the user's historical reading and current behavioral data, and adjusts the recommended content in real time; Monitoring and analysis module: Monitors users' reading behaviors in real time and adjusts personalized recommendations in a timely manner; User interface module: Designs a user-friendly interface, displays personalized recommended content, and provides a user feedback entry.

[0041] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0042] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A reading recommendation method based on artificial intelligence, characterized in that, including: Obtain the user's behavior data and reading data, where the behavior data includes page stay time, scrolling behavior, and click preference; Determine whether the user is a new user; If the user is a new user, set a diversified recommendation page, execute a diversified page prediction recommendation strategy, calculate the interest weights of the new user for different categories of content, and preliminarily predict the new user's interest direction; Optimize the predicted interest direction using guiding questions and perform content recommendation; If the user is an old user, establish a global recommendation model based on the old user's historical behavior data and reading data, combined with the current behavior data and reading data, and distribute the parameters of the global recommendation model to each platform; Execute a cross-platform federated learning strategy, independently train local recommendation models by accessing local data on different platforms, and use differential privacy algorithms to add noise to the parameters during the training process; After training is completed, upload the model to update the parameters, execute a secure multi-party computing strategy, aggregately calculate the updated parameters of each platform, and update the local recommendation model; Calculate the average weight of the updated global recommendation model to generate new global recommendation model parameters; Generate a personalized recommendation list, calculate the interest matching degree, and sort the recommended content in the personalized recommendation list according to the interest matching degree; Set long-term and short-term interest models based on the user's feedback on the recommended content to optimize the recommended content, execute a real-time dynamic feedback adjustment strategy, and adjust the personalized recommendation content in real time.

2. The reading recommendation method based on artificial intelligence according to claim 1, wherein, The above-mentioned setting of the diversified recommendation page executes the diversified page prediction and recommendation strategy, and calculates the interest weights of new users for different categories of content, including: obtaining the user's behavior data D U and reading data, and determining whether the reading data is empty; When the user's reading data is empty, the user is a new user U1; Set the diversified recommendation page to contain different categories P, P = {p1, p2,..., p n}, where p i represents the i-th page in the page, n represents the total number of n pages, and there is only one type of content in a page; Obtain the residence time T, scrolling behavior S, and click preference C of the new user U1 on page p i Execute a diversified page prediction and recommendation strategy, and calculate the interest weight I(U1, p i ) of the new user U1 for different categories of content. I(U1, p i ) = ω1·T(U1, p i ) + ω2·S(U1, p i ) + ω3·C(U1, p i ), where ω1, ω2, and ω3 are weight coefficients, and T(U1, p i ), S(U1, p i ), and C(U1, p i ) are respectively the residence time, scrolling behavior, and click preference of the new user U on page p i . According to the interest weight I(U1, p i ), predict the interest direction of the new user U1 Obtain the answer q to the guiding question by asking the new user U1 about their interest preferences through guiding questions U , and optimize the predicted interest direction ​ Perform content recommendation based on the optimized predicted interest direction.

3. The reading recommendation method based on artificial intelligence according to claim 1, characterized in that The execution of the cross-platform federated learning strategy to independently train local recommendation models by accessing local data on different platforms includes: Obtain the user's behavior data and reading data, and determine whether the reading data is empty; When the user's reading data is not empty, the user is an old user U2; Suppose there are a total of N different platforms, and each platform maintains local user data D i2 ; Based on the historical behavior data and reading data of the old user U2, combined with the current behavior data and reading data, set the global recommendation model as f θ , the global recommendation model is a shared model trained with cross-platform data of all old users U2, θ represents the parameters of the global recommendation model, and the predicted score of the interests of the old user U2 is calculated Distribute the parameters θ of the global recommendation model to each platform, execute a cross-platform federated learning strategy, and independently train local recommendation models by accessing local data on different platforms. The local recommendation model is a data training model for a single platform; On the platform obtain the user's true reading feedback calculate the loss function Calculate the local recommendation model parameters for the t-th round of training where θ t is the initial parameter for the t-th round of training of the global recommendation model, η is the learning rate, is the platform gradient; Use the differential privacy algorithm to add noise to the parameters during the training process to obtain the updated local recommendation model parameters Among them, is Gaussian noise.

4. The reading recommendation method based on artificial intelligence according to claim 3, characterized in that The training completion and uploading of the model to update the parameters, and the execution of the secure multi-party computing strategy to aggregately calculate the updated parameters of each platform includes: Send the updated local recommendation model parameters to the server, execute the secure multi-party computing strategy, and calculate the average weight ω of the updated global recommendation model i1 , Split into multiple shares and send them as s k to multiple computing parties M, and encrypt the multiple shares s k for aggregation calculation of the updated global recommendation model parameters θ of each platform g , ​ Send the updated global recommendation model parameters θ g to each platform to update the local recommendation models of each platform.

5. The reading recommendation method based on artificial intelligence according to claim 2, wherein The generation of the personalized recommendation list, the calculation of the interest matching degree, and the sorting of the recommended content in the personalized recommendation list according to the interest matching degree includes: Set the interest vector V of user U u , V u = {v u1 , v u2 , …, v uG}, where G represents the dimension of the interest vector, and v ua represents the preference degree of user U for the interest dimension a; Set the feature vector V of category content P p , V p = {v p1 , v p2 , …, v pG}, where v pa represents the eigenvalue of category content P on the interest dimension i; Calculate the norm of the user interest vector Calculate the norm of the category content feature vector Calculate the interest matching degree of user U to category content P For all predicted interest directions, category content P' = {p'1, p'2,..., p' n′} is sorted in descending order according to the size of the interest matching degree W(U, P); Present the top q contents of the sorted list of interest matching degrees W(U,P) as personalized recommendation content to the user.

6. The reading recommendation method based on artificial intelligence according to claim 5, wherein The setting of long-term and short-term interest models to optimize the recommended content, and the execution of the real-time dynamic feedback adjustment strategy to adjust the personalized recommendation content in real time includes: Calculate the feedback score W(U,P) of the user for the personalized recommendation content, W(U,P)=α·T(U,P)+β·S(U,P)+γ·C(U,P), where α, β, and γ are parameters that adjust the behavior impact; Obtain the set of historical reading contents of user U, Y1(U) = {y1, y2, …, y n′}, where y i4 represents the articles or books read by the user, and n′ represents the total number of n′ articles or books read by the user; Calculate the historical residence time weight according to the user's historical reading content set Y(U). Among them, ω t is the time decay factor, and T(U, Y, t) represents the residence time of the user reading the content y i4 within the time t; Calculate the historical reading category preference vector F(U). Set a long-term interest model L(U) representing the user's long-term and stable interest preferences based on the user's historical behavior data. where Z1 is a normalization factor. Obtain the current reading content set Y2(U) = {y t-1 , y t-2 , …, y t-n″} of user U, where y i4-1 represents the article or book that the user is currently reading, and n′ represents the number of articles or books that the user is currently reading; Calculate the weighted score R(U, y i4-1 ), R(U, y i4-1 ) = α · T(U, y i4-1 ) + β · S(U, y i4-1 ) + γ · C(U, y i4-1 ), where α, β, and γ are parameters for adjusting the influence of behavior; Set a short-term interest model \(L'(U)\) representing the user's recent interest preferences according to the weighted score of the user's current reading data. where \(Z2\) is a normalization factor; Fuse the long-term interest model and the short-term interest model with weights to calculate the final personalized interest model L * (U) = λL(U) + (1 - λ)L'(U), where λ ∈ [0, 1] is the weight coefficient; When λ takes a larger value, the long-term interest has a stronger impact, which is suitable for users with stable interests; When λ takes a smaller value, the short-term interest has a stronger impact, which is suitable for users with rapidly changing interests; Optimize and generate personalized recommendation content according to the long-term interest model and the short-term interest model, monitor the user's real-time reading data, and adjust the personalized recommendation content in real time, L * (U)' = L * (U) + δ·W(U, P), where δ is the update step size that controls the adjustment amplitude.

7. A system for reading recommendation based on artificial intelligence, which is applied to a method for reading recommendation based on artificial intelligence according to any one of claims 1-6, characterized in that, including: Data Acquisition Module: Collects users' behavioral data on the platform, including page stay time, scrolling behavior, click preferences, etc., and records users' reading history data; User Classification Module: Determines whether a user is a new user or an old user based on whether the user's reading data is empty, and extracts the behavioral characteristics of new users and old users; Recommendation Model Construction Module: Sets diverse recommendation pages for new users, calculates interest weights through users' behavioral data, optimizes the predicted interest direction using guiding questions, builds a global recommendation model based on old users' historical behavioral data and current data, and distributes the parameters to each local platform; Cross-Platform Federated Learning Module: Independently trains local recommendation models on each different platform, executes cross-platform federated learning strategies, and applies differential privacy algorithms during the training process to add noise to the model parameters; Model Update and Aggregation Module: Uploads the updated local model parameters to the server, performs secure multi-party computation to aggregate and calculate the global model parameters, sends the updated global recommendation model parameters to each platform, and updates the local recommendation models; Recommended Content Generation Module: Calculates the matching degree between the user's interest vector and the category content feature vector, generates personalized recommended content, sorts the personalized recommendation list according to the interest matching degree, and presents it to the user; Real-Time Feedback and Adjustment Module: Collects users' feedback on the recommended content, calculates the feedback score, dynamically adjusts the user's long-term and short-term interest models based on the user's historical reading and current behavioral data, and adjusts the recommended content in real time; Monitoring and Analysis Module: Monitors users' reading behaviors in real time and adjusts personalized recommendations in a timely manner; User Interface Module: Designs a user-friendly interface, displays personalized recommended content, and provides a user feedback entry.

Images