Book recommendation method and system based on knowledge graph

Abstract

The application belongs to the technical field of book recommendation, and particularly relates to a book recommendation method and system based on a knowledge graph. The method comprises the following steps: S1, acquiring book related data and book system user related data; S2, cleaning the book related data and the book system user related data; S3, constructing a knowledge ontology according to the cleaned book data, and generating a classified book knowledge graph; S4, constructing a knowledge ontology according to the book system user related data and user system use condition information, and generating a user knowledge ontology and a user portrait knowledge graph; and S5, generating a personalized book recommendation for a user according to the book knowledge graph, the user knowledge ontology and the user portrait knowledge graph, and combining a hybrid sequence recommendation algorithm. The application has the characteristics of combining a deep learning technology, using a knowledge graph to represent book knowledge information and user interest conditions, and being highly automated, dynamic and efficient.

Description

Technical Field

[0001] This invention belongs to the field of book recommendation technology, specifically relating to a book recommendation method and system based on knowledge graphs. Background Technology

[0002] Prior to 2012, recommender systems primarily used traditional methods, relying on algorithms such as matrix factorization to provide relatively high-quality services to users. The introduction of the AlexNet model then ignited a surge in deep learning, and its models have been widely applied in the field of recommender systems. Previous recommender systems mainly employed methods such as content-based recommendation, collaborative filtering recommendation, and knowledge-based recommendation systems.

[0003] Content-based recommendation uses machine learning to derive user interest data from examples of content feature descriptions. The system evaluates the characteristics of objects based on the user's historical data, learns the user's interests, and examines the degree of match between the user profile and the item to be predicted. The user profile model changes as the user's preferences evolve.

[0004] Collaborative filtering recommendations typically employ nearest neighbor technology. It calculates the distance between different users using their historical preference information, and then uses the weighted ratings of the target user's nearest neighbors to predict the target user's preference for specific products. The system then makes recommendations to the target user based on this preference level. Collaborative filtering is based on recommending content to users who have similar interests to a given user.

[0005] Knowledge-based recommendation systems rely on effective information to further coordinate the relationship between user needs and system recommendations, and assist the recommendation system in making relevant inferences based on the user's personal information. However, current mainstream recommendation systems combined with deep learning have many drawbacks. Deep learning itself follows an "end-to-end" design philosophy, emphasizing minimal prior representations and computational assumptions. It is also costly, has poor interpretability, low scalability, and lacks systematic inductive biases. These construction principles partially contradict the pairwise structured data reasoning required by recommendation system models. Furthermore, the user's subjective initiative leads to the sparsity and narrowness of data, making it difficult for deep learning models to abstract from fine-grained differences in hierarchical structure and capture the more general commonalities between labels and behaviors.

[0006] Therefore, it is essential to design a knowledge graph-based book recommendation method and system that combines deep learning technology, can represent the knowledge information of books and user interests using knowledge graphs, and has the characteristics of high automation, dynamism and efficiency.

[0007] For example, Chinese patent document CN201310412306.3 describes a book recommendation method and system based on social networks. The method includes: Step 1, extracting user interaction information with other users on social networks, constructing several interaction type friend groups for the user, and then classifying other users with successful interactions with the user into different interaction type friend groups according to their interaction type; Step 2, calculating the number of successful interactions between the user and each friend in each interaction type friend group, then selecting the top few friends with the highest number of successful interactions from each interaction type friend group, and finally recommending the most read books by these friends to the user. While this method can personalize book recommendations based on user interaction behavior on social networks, its drawbacks are that it still uses conventional content-based recommendation or collaborative filtering methods, failing to analyze users' potential interest characteristics, making it difficult to enhance the persuasiveness of the results and improve user satisfaction. Furthermore, an increase in the number of users and books can lead to a decrease in system performance. Summary of the Invention

[0008] This invention aims to overcome the problems of low automation, poor interpretability, low scalability, and poor model generalization in existing book recommendation systems. It provides a knowledge graph-based book recommendation method and system that combines deep learning technology, can represent the knowledge information of books and user interests using knowledge graphs, and has the characteristics of high automation, dynamism, and efficiency.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] The knowledge graph-based book recommendation method includes the following steps:

[0011] S1, obtain book-related data and book system user-related data;

[0012] S2, Clean the book-related data and the book system user-related data;

[0013] S3: Construct a knowledge ontology based on the cleaned book data and generate a book knowledge graph categorized by type;

[0014] S4. Construct a knowledge ontology based on user-related data of the library system and information on user usage of the system, and generate a user knowledge ontology and a user profile knowledge graph.

[0015] S5 generates personalized book recommendations for users based on the book knowledge graph, user knowledge ontology, and user profile knowledge graph, combined with a hybrid sequence recommendation algorithm.

[0016] Preferably, step S1 includes the following steps:

[0017] S11 uses automated testing methods to simulate how users access data.

[0018] The book-related data includes the book title, author, publisher, distributor, subtitle, publication year, translator, ISBN, synopsis, binding, price, and the book's category.

[0019] The user-related data of the book system includes user ID, comments, likes, user ratings, click-through rate, and activity level;

[0020] S12, Based on the data involved in step S11, extract the user's interaction information with the book.

[0021] Preferably, step S2 includes the following steps:

[0022] Book-related data cleaning:

[0023] S21, fill in the missing translator information for Chinese books in the book content related data;

[0024] S22, translate the relevant English book introductions and author names into Chinese;

[0025] S23, delete books that are missing page numbers or key price information;

[0026] S24, Clean up books with duplicate information in different book categories;

[0027] S25, perform data visualization analysis to identify outliers;

[0028] Library system user-related data:

[0029] S26, Clean up abnormal user data.

[0030] Preferably, step S3 includes the following steps:

[0031] S31. Logically define the cleaned book data and construct the schema layer in the form of triples to form a knowledge ontology knowledge base containing book data.

[0032] The triplet form includes entity-relation-entity or entity-attribute-value;

[0033] S32, use an NLP model to extract information from the book's synopsis; information extraction includes entity extraction, relation extraction, and attribute extraction.

[0034] S33, Based on the extracted book information and existing book information, knowledge fusion is performed to generate a basic knowledge graph of book entities;

[0035] S34, Professional teachers and experts review the generated book entity basic knowledge graph and determine whether the book entity basic knowledge graph meets the standards;

[0036] S35. If the book entity basic knowledge graph meets the standard, the final book knowledge graph system is generated; otherwise, return to step S32.

[0037] Preferably, step S4 includes the following steps:

[0038] S41, analyze and process user information, and construct a user knowledge ontology with different weights and feature factors;

[0039] S42, Generate a knowledge graph based on user profiles based on user knowledge ontology and book data related to user interactions.

[0040] In generating the knowledge graph for user profiles, the transD translation technique and dynamic transformation matrix method are used to map user features and book features to two different feature spaces.

[0041] Preferably, step S5 includes the following steps:

[0042] S51, by performing a weighted Deepwalk random walk on the knowledge graph, a user-book sequence is constructed; after the random walk, each specific entity generates a data sequence;

[0043] S52 uses word2vec to generate an embedding vector for each entity;

[0044] S53, for users who click frequently, uses the deep learning technology framework transformer model to perform time series perception, deeply mines user features from time-related embeddings, and introduces nested modules to form a deep neural network to ensure the model's representation ability;

[0045] S54 introduces time information and uses a new type of position encoding to calculate parameters:

[0046]

[0047] In the formula, k is the index number of the book; V k For user vectors; TPos(V) kThe index number k is the encoded value of the interval between user clicks on the same book, which is the sum of the index number k and the user click time interval; i is the current user index number; n is the total number of users;

[0048] Each transformer model contains two layers: one layer uses a multi-head attention mechanism and the other layer uses a feedforward neural network to represent different subspace location information and enable the different subspace location information to interact between different dimensions.

[0049] Two additional feedforward neural networks and a ReLU activation function are added as the output function at the end of the transformer model.

[0050] S55, based on the inner product of the embedding vectors, calculate the normalized similarity pi between the book vector and (h,r) on the sequence diffusion set formed by the first step of the random walk in step S51. The specific calculation formula is as follows:

[0051]

[0052] Where (h,r) are two sets of vectors representing the relationship between users and book browsing; v is the book vector, v T Let v be the transpose of v; R be the vector of all relations that construct the knowledge graph; R i Let h be the relation vector of the i-th knowledge graph; h is the vector of all users; h i Let r be the vector of the i-th user; r be the relation vector in the triplet of the knowledge graph; and S be the triplet of the knowledge graph. It represents the triple vectors from the first to the uth knowledge graph, i.e., traversing all triple vectors on the knowledge graph;

[0053] Based on the normalized similarity, the (t) of the first layer diffusion set is weighted and summed, and the result is used as the output of this layer. This step is repeated H times.

[0054] The formula for calculating the weighted sum of (t) of the first layer diffusion set is as follows:

[0055]

[0056] Weighted summation involves multiplying the newly obtained normalized vector pi with the knowledge graph triple vector t in a matrix manner, h i Let r be the vector of the i-th user; i Let t be the i-th relation vector. i Let i be the vector of the i-th book;

[0057] S56. Repeat steps S51 to S55 until the deep learning model transformer tends to converge.

[0058] This invention also provides a knowledge graph-based book recommendation system, including:

[0059] The data acquisition module is used to acquire book-related data and book system user-related data;

[0060] The data cleaning module is used to clean book-related data and user data of the book system;

[0061] The book knowledge graph generation module is used to construct knowledge ontology based on the cleaned book data and generate book knowledge graphs categorized by type.

[0062] The user knowledge ontology and user profile knowledge graph generation module is used to construct a knowledge ontology and generate a user knowledge ontology and user profile knowledge graph based on user-related data of the library system and information on user usage of the system.

[0063] The book recommendation generation module is used to generate personalized book recommendations for users based on book knowledge graphs, user knowledge ontology, and user profile knowledge graphs, combined with a hybrid sequence recommendation algorithm.

[0064] Compared with the prior art, the beneficial effects of this invention are: (1) This invention has high extensibility: the recommendation system has the excellent characteristics of "configuration software" and strong general adaptability. When it is necessary to build a recommendation system, such as a book recommendation system, the appropriate functional components can be selected according to business needs and the recommendation system can be built in the form of configuration; (2) This invention has interpretability: in the process of suggestion formation, users do not need to passively accept the recommendation results. In the process of recommendation, not only are biased estimated recommendation results given, but also the reasons for recommendation can be given to enhance the persuasiveness of the results and improve user satisfaction; (3) This invention has strong generalization: the model can extract the hidden features of the data, such as the hidden features of the user's personal information, the books associated with the borrowed books, etc., and reasonably match the relevant recommended books; (4) This invention has low coupling: it reduces the dimensions of large-scale data and can perform unified semantic representation of multi-source heterogeneous data. Attached Figure Description

[0065] Figure 1 This is a flowchart of a knowledge graph-based book recommendation method in this invention;

[0066] Figure 2 This is a flowchart illustrating the process of cleaning book-related data and book system user-related data in this invention.

[0067] Figure 3 This is a flowchart illustrating the process of constructing a knowledge ontology based on cleaned book data and generating a categorized book knowledge graph in this invention.

[0068] Figure 4 This is a schematic diagram of the position encoding parameter calculation process in this invention. Detailed Implementation

[0069] To more clearly illustrate the embodiments of the present invention, specific implementation methods will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0070] Example:

[0071] like Figure 1 As shown, this invention provides a book recommendation method based on knowledge graphs, including the following steps:

[0072] S1, obtain book-related data and book system user-related data;

[0073] S2, Clean the book-related data and the book system user-related data;

[0074] S3: Construct a knowledge ontology based on the cleaned book data and generate a book knowledge graph categorized by type;

[0075] S4. Construct a knowledge ontology based on user-related data of the library system and information on user usage of the system, and generate a user knowledge ontology and a user profile knowledge graph.

[0076] S5 generates personalized book recommendations for users based on the book knowledge graph, user knowledge ontology, and user profile knowledge graph, combined with a hybrid sequence recommendation algorithm.

[0077] Step S1 specifically includes the following steps:

[0078] S11 uses automated testing methods to simulate how users access data.

[0079] The book-related data includes the book title, author, publisher, distributor, subtitle, publication year, translator, ISBN, synopsis, binding, price, and the book's category.

[0080] The user-related data of the book system includes user ID, comments, likes, user ratings, click-through rate, and activity level;

[0081] S12, Based on the data involved in step S11, extract the user's interaction information with the book.

[0082] Step S2 specifically includes the following steps:

[0083] Book-related data cleaning:

[0084] S21, fill in the missing translator information for Chinese books in the book content related data;

[0085] S22, translate the relevant English book introductions and author names into Chinese;

[0086] S23, delete books that are missing page numbers or key price information;

[0087] S24, Clean up books with duplicate information in different book categories;

[0088] S25, perform data visualization analysis to identify outliers;

[0089] Library system user-related data:

[0090] S26, Clean up abnormal user data.

[0091] In recommendation systems, abnormal data in user data includes missing values, error values, and duplicate values. Error values ​​refer to situations where some users click on a certain book more than 500 times, resulting in an excessively high click-through rate, or repeatedly commenting on the same book.

[0092] like Figure 3 As shown, step S3 includes the following steps:

[0093] S31. Logically define the cleaned book data and construct the schema layer in the form of triples to form a knowledge ontology knowledge base containing book data.

[0094] The triplet form includes entity-relation-entity or entity-attribute-value;

[0095] S32, use an NLP model to extract information from the book's synopsis; information extraction includes entity extraction, relation extraction, and attribute extraction.

[0096] S33, Based on the extracted book information and existing book information, knowledge fusion is performed to generate a basic knowledge graph of book entities;

[0097] S34, Professional teachers and experts review the generated book entity basic knowledge graph and determine whether the book entity basic knowledge graph meets the standards;

[0098] S35. If the book entity basic knowledge graph meets the standard, the final book knowledge graph system is generated; otherwise, return to step S32.

[0099] Step S4 includes the following steps:

[0100] S41, analyze and process user information, and construct a user knowledge ontology with different weights and feature factors;

[0101] S42, Generate a knowledge graph based on user profiles based on user knowledge ontology and book data related to user interactions.

[0102] It should be noted that, in order to ensure low computational cost and high matching accuracy when generating the knowledge graph of user profiles, transD is used as the main translation technology. The dynamic transformation matrix method is used to map user features and book features to two different feature spaces.

[0103] Step S5 includes the following steps:

[0104] S51, by performing a weighted Deepwalk random walk on the knowledge graph, a user-book sequence is constructed; after the random walk, each specific entity generates a data sequence;

[0105] S52 uses word2vec to generate an embedding vector for each entity;

[0106] Among them, word2vec, or word vector product, is a technique used to map words to a vector space;

[0107] S53, for users who click frequently, uses the deep learning technology framework transformer model to perform time series perception, deeply mines user features from time-related embeddings, and introduces nested modules to form a deep neural network to ensure the model's representation ability;

[0108] Among them, transformer is a type of converter;

[0109] S54, in order to better capture time-dependent user interest information, not only are the books browsed by users annotated, but time information will also be introduced, and new location coding parameters will be used to calculate parameters:

[0110]

[0111] In the formula, k is the index number of the book; V k For user vectors; TPos(V) k The index number k is the encoded value of the interval between user clicks on the same book, which is the sum of the index number k and the user click time interval; i is the current user index number; n is the total number of users;

[0112] Each transformer model contains two layers: one layer uses a multi-head attention mechanism and the other layer uses a feedforward neural network to represent different subspace location information and enable the different subspace location information to interact between different dimensions.

[0113] Two additional feedforward neural networks and a ReLU activation function are added as the output function at the end of the transformer model.

[0114] S55, based on the inner product of the embedding vectors, calculate the normalized similarity pi between the book vector and (h,r) on the sequence diffusion set formed by the first step of the random walk in step S51. The specific calculation formula is as follows:

[0115]

[0116] Where (h,r) are two sets of vectors representing the relationship between users and book browsing; v is the book vector, v T Let v be the transpose of v; R be the vector of all relations that construct the knowledge graph; R i Let h be the relation vector of the i-th knowledge graph; h is the vector of all users; h i Let r be the vector of the i-th user; r be the relation vector in the triplet of the knowledge graph; and S be the triplet of the knowledge graph. It represents the triple vectors from the first to the uth knowledge graph, i.e., traversing all triple vectors on the knowledge graph;

[0117] Based on the normalized similarity, the (t) of the first layer diffusion set is weighted and summed, and the result is used as the output of this layer. This step is repeated H times.

[0118] The formula for calculating the weighted sum of (t) of the first layer diffusion set is as follows:

[0119]

[0120] Weighted summation involves multiplying the newly obtained normalized vector pi with the knowledge graph triple vector t in a matrix manner, h i Let r be the vector of the i-th user; i Let t be the i-th relation vector. i Let i be the vector of the i-th book;

[0121] S56. Repeat steps S51 to S55 until the deep learning model transformer tends to converge.

[0122] The main feature changes during image processing in this invention are as follows: Figure 3As shown. In the foreground and background separation part of the image, after image enhancement and Grabcut processing, the original input image is separated into foreground and background. However, the separated image retains the shadow parts of the original image, i.e., the white parts in the figure. Then, the corresponding color gamut is converted and the components cr and a are extracted. Among them, the component cr handles the information of the shadow parts very well, that is, the shadow area is effectively dissolved in the component cr, but it cannot completely capture the information of the lesion area; the component a effectively captures the information of the lesion area, but it cannot dissolve the shadow noise, that is, the shadow area is preserved in the component a. Next, adaptive equalization is performed on the two component channels to make the feature information of the image more obvious than before processing. Then, image thresholding is performed to remove invalid information. The information of the two channels after thresholding is fused. The fused information effectively dissolves the shadow noise and also captures the lesion area information very well. Finally, masking is performed to extract the segmented lesion area.

[0123] like Figure 4 The diagram illustrates the process of calculating the location encoding parameters according to this invention. Specifically, this invention encodes the temporal and location information of the sequence of books visited by each user and the sequence of candidate books (i.e., unvisited but similar books), and then passes this information into multiple transformer modules (the number of transformer layers is selected based on the amount of data; this invention uses six layers). Each transformer module includes a connection between two multi-head attention layers and a normalization layer. After the computation is complete, the information of the candidate books is concatenated with that of the visited books and then fed into a DNN (i.e., a feedforward neural network) to obtain the final baseline prediction probability for reference.

[0124] This invention also provides a knowledge graph-based book recommendation system, including:

[0125] The data acquisition module is used to acquire book-related data and book system user-related data;

[0126] The data cleaning module is used to clean book-related data and user data of the book system;

[0127] The book knowledge graph generation module is used to construct knowledge ontology based on the cleaned book data and generate book knowledge graphs categorized by type.

[0128] The user knowledge ontology and user profile knowledge graph generation module is used to construct a knowledge ontology and generate a user knowledge ontology and user profile knowledge graph based on user-related data of the library system and information on user usage of the system.

[0129] The book recommendation generation module is used to generate personalized book recommendations for users based on book knowledge graphs, user knowledge ontology, and user profile knowledge graphs, combined with a hybrid sequence recommendation algorithm.

[0130] This invention features high scalability: the recommender system possesses the excellent characteristics of "configuration software," with strong general adaptability. When building a recommender system, such as a book recommender system, appropriate functional components can be selected according to business needs, and the system can be constructed in a configurable manner. This invention also features interpretability: during the suggestion formation process, users do not need to passively accept recommendation results. The system not only provides biased estimates of recommendation results but also provides reasons for the recommendations, enhancing the persuasiveness of the results and improving user satisfaction. Furthermore, this invention exhibits strong generalization: the model can extract hidden features from the data, such as users' personal information and hidden features of books associated with borrowed books, and reasonably match relevant recommended books. Finally, this invention features low coupling: it reduces the dimensionality of large-scale data and can provide a unified semantic representation for multi-source heterogeneous data.

[0131] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention. For those skilled in the art, there may be changes in specific implementation methods based on the ideas provided by the present invention, and these changes should also be considered within the scope of protection of the present invention.

Claims

1. A book recommendation method based on knowledge graphs, characterized in that, Includes the following steps: S1, obtain book-related data and book system user-related data; S2, Clean the book-related data and the book system user-related data; S3: Construct a knowledge ontology based on the cleaned book data and generate a book knowledge graph categorized by type; S4. Construct a knowledge ontology based on user-related data of the library system and information on user usage of the system, and generate a user knowledge ontology and a user profile knowledge graph. S5 generates personalized book recommendations for users based on book knowledge graphs, user knowledge ontology, and user profile knowledge graphs, combined with a hybrid sequence recommendation algorithm. The hybrid sequence recommendation algorithm includes knowledge embedding, time-series awareness, and information aggregation prediction. Step S3 includes the following steps: S31. Logically define the cleaned book data and construct the schema layer in the form of triples to form a knowledge ontology knowledge base containing book data. The triplet form includes entity-relation-entity or entity-attribute-value; S32, use an NLP model to extract information from the book's synopsis; information extraction includes entity extraction, relation extraction, and attribute extraction. S33, Based on the extracted book information and existing book information, knowledge fusion is performed to generate a basic knowledge graph of book entities; S34, Professional teachers and experts review the generated book entity basic knowledge graph and determine whether the book entity basic knowledge graph meets the standards; S35. If the book entity basic knowledge graph meets the standard, the final book knowledge graph system is generated; otherwise, return to step S32. Step S4 includes the following steps: S41, analyze and process user information, and construct a user knowledge ontology with different weights and feature factors; S42, Generate a knowledge graph based on user profiles based on user knowledge ontology and book data related to user interactions; In generating the knowledge graph for user profiles, the transD translation technique and the dynamic transformation matrix method are used to map user features and book features to two different feature spaces. Step S5 includes the following steps: S51, by performing a weighted Deepwalk random walk on the knowledge graph, a user-book sequence is constructed; after the random walk, each specific entity generates a data sequence; S52 uses word2vec to generate an embedding vector for each entity; S53, for users who click frequently, uses the deep learning technology framework transformer model to perform time series perception, deeply mines user features from time-related embeddings, and introduces nested modules to form a deep neural network to ensure the model's representation ability; S54 introduces time information and uses a new type of position encoding to calculate parameters: In the formula, k is the index number of the book; V k For user vectors; TPos(V) k The index number k is the encoded value of the interval between user clicks on the same book, which is the sum of the index number k and the user click time interval; i is the current user index number; n is the total number of users; Each transformer model contains two layers: one layer uses a multi-head attention mechanism and the other layer uses a feedforward neural network to represent different subspace location information and enable the different subspace location information to interact between different dimensions. Two additional feedforward neural networks and a ReLU activation function are added as the output function at the end of the transformer model. S55, based on the inner product of the embedding vectors, calculate the normalized similarity pi between the book vector and (h,r) on the sequence diffusion set formed by the first step of the random walk in step S51. The specific calculation formula is as follows: Where (h,r) are two sets of vectors representing the relationship between users and book browsing; v is the book vector. Let v be the transpose of v; R is the vector of all relations that construct the knowledge graph; Let h be the relation vector of the i-th knowledge graph; h is the vector of all users; h i Let r be the vector of the i-th user; r be the relation vector in the triplet of the knowledge graph; and S be the triplet of the knowledge graph. It represents the triple vectors from the first to the uth knowledge graph, i.e., traversing all triple vectors on the knowledge graph; Based on the normalized similarity, the (t) of the first layer diffusion set is weighted and summed, and the result is used as the output of this layer. This step is repeated H times. The formula for calculating the weighted sum of (t) of the first layer diffusion set is as follows: Weighted summation involves multiplying the newly obtained normalized vector pi with the knowledge graph triple vector t in a matrix manner, h i Let i be the vector of the i-th user; For the i-th relation vector, Let i be the vector of the i-th book; S56. Repeat steps S51 to S55 until the deep learning model transformer tends to converge.

2. The book recommendation method based on knowledge graphs according to claim 1, characterized in that, Step S1 includes the following steps: S11 uses automated testing methods to simulate how users access data. The book-related data includes the book title, author, publisher, distributor, subtitle, publication year, translator, ISBN, synopsis, binding, price, and the book's category. The user-related data of the book system includes user ID, comments, likes, user ratings, click-through rate, and activity level; S12, Based on the data involved in step S11, extract the user's interaction information with the book.

3. The book recommendation method based on knowledge graphs according to claim 2, characterized in that, Step S2 includes the following steps: Book-related data cleaning: S21, fill in the missing translator information for Chinese books in the book content related data; S22, translate the relevant English book introductions and author names into Chinese; S23, delete books that are missing page numbers or key price information; S24, Clean up books with duplicate information in different book categories; S25, perform data visualization analysis to identify outliers; Library system user-related data: S26, Clean up abnormal user data.

4. A knowledge graph-based book recommendation system, used to implement the knowledge graph-based book recommendation method according to any one of claims 1-3, characterized in that, The knowledge graph-based book recommendation system includes: The data acquisition module is used to acquire book-related data and book system user-related data; The data cleaning module is used to clean book-related data and user data of the book system; The book knowledge graph generation module is used to construct knowledge ontology based on the cleaned book data and generate book knowledge graphs categorized by type. The user knowledge ontology and user profile knowledge graph generation module is used to construct a knowledge ontology and generate a user knowledge ontology and user profile knowledge graph based on user-related data of the library system and information on user usage of the system. The book recommendation generation module is used to generate personalized book recommendations for users based on book knowledge graphs, user knowledge ontology, and user profile knowledge graphs, combined with a hybrid sequence recommendation algorithm.

Images