Book recommendation method based on graph attention algorithm

Abstract

A book recommendation method based on a graph attention algorithm, by inputting a knowledge graph containing users, books and book attributes into a graph attention recommendation model constructed and trained, a book similarity scoring matrix is obtained for front-end visualization. The graph attention recommendation model comprises an input layer, an information propagation layer, an information aggregation generation layer and a recommendation result generation layer, wherein: the input layer randomly generates a vector e of each node as an initialization representation according to the input knowledge graph, the information propagation layer calculates the information propagation received by each node from surrounding nodes, the information aggregation layer aggregates the propagation information and node information to generate next-hop information, and the recommendation result generation layer scores and generates a book similarity scoring matrix according to the training result. The present application significantly improves the analysis efficiency and performance of book recommendation by mining high-order paths of the knowledge graph through the graph attention method, using an improved knowledge graph propagation algorithm, an aggregation method and a knowledge graph construction for the library collection.

Description

Technical Field

[0001] This invention relates to a technology in the field of knowledge graphs, specifically a book recommendation method based on graph attention algorithms. Background Technology

[0002] Knowledge graph recommendation algorithms were developed to solve the cold start problem, but they have the following shortcomings: First, traditional algorithms do not distinguish between book nodes and user nodes, and do not make full use of the characteristics of knowledge graphs; second, the relationship between users and their neighboring book nodes is simple and lacks differentiation, making them more like GCNs than GANs; finally, the use of graph attention networks and collaborative filtering in combination brings high time costs. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as cold start, excessively long training time, lack of effective encoding of user-book relationships, inability to guarantee the capture of long-range connectivity, inability to interpret high-order modeling, and excessive cost. It proposes a book recommendation method based on graph attention algorithms. This method utilizes graph attention and random walk techniques to mine high-order paths in the knowledge graph and employs improved knowledge graph propagation algorithms, aggregation methods, and knowledge graph construction tailored to library collections. This significantly improves the analytical efficiency and performance of book recommendation.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a book recommendation method based on a graph attention algorithm, using the aforementioned system, comprising:

[0006] Step 1) Data Import: Clean and import the user-book borrowing dataset and the book information dataset. Users need to provide methods for converting book category information and ISBNs, which will be used to convert all book categories into ISBN information and segment them.

[0007] Step 2) Constructing a knowledge graph: First, the introduction information is segmented and cleaned into keyword information. Then, the relationships of all book information, including the keyword information, are statistically analyzed, with the numbers incrementing from 0, to construct a triple of (book ID, relationship, book attribute). Then, a similar process is performed on the user-book dataset to obtain a triple of (user ID, borrowing relationship, book ID), thus completing the construction of the knowledge graph.

[0008] Step 3) Input the final knowledge graph containing users, books, and book attributes to construct and train the graph attention recommendation model, and obtain the book similarity scoring matrix for front-end visualization.

[0009] The graph attention recommendation model includes an input layer, an information propagation layer, an information aggregation and generation layer, and a recommendation result generation layer. The input layer randomly generates a vector e for each node as an initial representation based on the input knowledge graph. The information propagation layer calculates the information propagated by each node to its surrounding nodes. The information aggregation layer aggregates the propagated information and node information to generate next-hop information. The recommendation result generation layer scores the data based on the training results and generates a book similarity scoring matrix.

[0010] The graph attention recommendation model described above uses the following loss function during training: Where: T represents the set of knowledge graph triples, h, r, t represent the head entity, relation, and tail entity with triple relations, f represents any random entity other than these, and Wr∈R k ×d It is the mapping matrix corresponding to relation r, e∈R d It is the corresponding vector, and F is the activation function, which defaults to F(x) = -lnσ(x). Technical effect

[0011] This invention captures long-range connectivity in a library's book knowledge graph using a graph attention recommendation model, enabling the extraction of information that traditional recommendation algorithms cannot uncover and reducing training time. By adding randomness to the propagation-aggregation process of book information to user information, it achieves faster training time than other recommendation algorithms, realizing better performance with less than half the training time. The knowledge graph construction process incorporates more diverse ISBN information, ranging from single-classification information to multi-level classification structures, expressing book popularity and completeness through randomness. Compared to other recommendation algorithms, this approach results in a more complete knowledge graph construction, capable of extracting information hidden in attributes such as book category and description. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention;

[0013] Figure 2 This is a schematic diagram of the system of the present invention;

[0014] Figure 3 The recall rate results are shown in the graph.

[0015] Figure 4 Image of NDCG results;

[0016] Figure 5 Recommendation model structure diagram;

[0017] Figure 6 This is a flowchart of an implementation example;

[0018] Figure 7This is the flowchart for step 2;

[0019] Figure 8 This is the flowchart for step 3;

[0020] Figure 9 This is the flowchart for step 4;

[0021] Figure 10 A schematic diagram of a knowledge graph structure centered on books; Detailed Implementation

[0022] like Figure 1 and Figure 6 As shown, this embodiment relates to a book recommendation method based on a graph attention algorithm, which involves constructing and training a graph attention algorithm as follows: Figure 4 After implementing the knowledge graph-based graph attention recommendation model shown, the resulting model is constructed based on the book information to be processed. Figure 10 The book knowledge graph shown scores each candidate book and recommends the highest-scoring item, specifically including:

[0023] Step 1) Processing borrowing and book information, specifically including:

[0024] 1.1) First, extract all book relations and mark all book relations as integers starting from 0 and increasing. There is a triple relation between each book and its attribute, namely (book id, relation, book attribute), where the relation refers to author, publisher, book category, etc., for example (Dawn Blossoms Plucked at Dusk, author, Lu Xun).

[0025] The book category refers to the classification corresponding to the book's ISBN number. This classification is presented in a tree diagram, which conforms to the data structure characteristics of ISBNs. Each time a book category name is clicked in the book category segmentation module, books containing that category will be filtered in the recommended books. The book recommendation module will only display the recommendations of the filtered books.

[0026] 1.2) Data cleaning is performed on the book attribute information. Then, special data processing is required for the book category data. Each book category corresponds to an ISBN number. For books that are missing an ISBN number, the book category needs to be converted into an ISBN number. The ISBN number is then split according to the tree structure of the ISBN to obtain the first-level, second-level, and third-level ISBN numbers as supplementary book information.

[0027] 1.3) Extract keyword information from the book synopsis by using other attribute information of the existing book as keywords, and then replace the synopsis with the keyword information in the synopsis as the book information triple.

[0028] 1.4) Add information on book popularity. The number of times each book has been borrowed can be obtained from the book borrowing data. Then, select a threshold n. Books with more than n borrowing times are considered popular books and have the attribute "popular". Books with less than n are considered unpopular data and are considered "unpopular". n can be set to an integer that makes the ratio of popular to unpopular books close to 1:19.

[0029] 1.5) Number all book information using an increasing integer sequence with a minimum value of 0. Book information and numbers are in one-to-one correspondence. After random sorting, number all books using an increasing integer sequence after the maximum book information number. Books and numbers are in one-to-one correspondence. Finally, number users using an increasing integer sequence after the maximum book number.

[0030] Step 2) Knowledge graph construction, specifically including: Figure 7 As shown, a knowledge graph (book ID, relation group, book attribute) is constructed by merging triples (book ID, borrowing relation, book ID). All relation groups are encoded to obtain the matrix W corresponding to each relation. i Integrate all triples, sort and encode all user nodes, book nodes, and attribute nodes, and transform all triples into corresponding d-dimensional entity vectors and k-dimensional relation vectors.

[0031] Step 3) Knowledge graph propagation information calculation: First, calculate the propagation information to book nodes, then calculate the propagation information to user nodes. The knowledge graph is updated based on the information propagation and aggregation between book nodes and user nodes. Specifically, as follows... Figure 8 As shown, it includes:

[0032] 3.1) Calculate the propagation information strength from the neighboring nodes of the book node to the book node using the relational space transformation method (TransR) for knowledge representation learning. Specifically: the propagation information strength from the neighboring node to the book node is π1 = (W r e t ) T g(W r e h +e r ), where: π1∈R, e h ∈R d First vector of book vector, e r ∈R k It is the tail vector of book attributes, matrix W r ∈R k×d Let r be the mapping matrix corresponding to relation r, and g be the activation function tanh. This gives us the propagation information of all triples for each book node.

[0033] Preferably, the intensity of the propagated information transmitted to the book node is exponentially normalized.

[0034] 3.2) Calculate the propagation information strength from book nodes to user nodes. This is done by randomly selecting book nodes, where each book's attribute nodes have a simple frequency relationship f with the user, and then multiplying the information from adjacent nodes by the propagation information strength and summing the results. Specifically, this is: Propagation Information Strength from Book Nodes to User Nodes. Where: π2∈R, J is the set of all borrowed books of the user node, I j It is the set of all attribute nodes connected to the book node, π1∈R is the propagation strength of the neighbor node to the book node in step 3.1, the superscript i is its number order relative to the user, and P is a random function representing that in the t∈N-th iteration, f i The probability is 1 if it is true and 0 otherwise.

[0035] For example, if a user borrows five books, two of which are by Lu Xun, then the frequency of Lu Xun for the user is f = 2 / 5.

[0036] Preferably, the intensity of the propagated information transmitted to user nodes is exponentially normalized.

[0037] 3.3) Calculate the propagation information for each book node and user node separately. Specifically, sum the propagation information intensity π1 of the book node's neighbor nodes obtained in step 3.1 to obtain the propagation information of the book node. The propagation information of the user node is obtained by summing the propagation information intensity π2 obtained from the book node in step 3.2. J represents the set of all attributes of a book and the set of all borrowed books of a user node, respectively, where: e Nu ∈R d Information propagated by user nodes, e Ni ∈R d This represents the propagation information of the book node, where i represents the sorting of the corresponding set.

[0038] Step 4) Aggregate the information propagated to the book nodes and the information propagated to the user nodes, and calculate the loss function, as follows: Figure 9 As shown, it includes:

[0039] 4.1) Summing is performed using the Bi-Interaction method, based on the information of the input node itself and the information propagated to the node. Specifically, this is the node vector of the t-th hop of the book node. The node vector of the t-th hop of the book node in: e Nu ∈R dThe propagation information representing user nodes, e Ni ∈R d Information disseminated representing book nodes, Represents element-wise multiplication, W1∈R d×d and W2∈R d×d These represent the corresponding weight matrices; if there are no weights, they can be set to a diagonal matrix I. The node information and propagation information of the previous hop are aggregated to update and form the node information for the next hop. User nodes and book nodes are aggregated and processed serially, with the same number of iterations.

[0040] 4.2) Calculate the loss function Where: T represents the set of knowledge graph triples, h, r, t represent the head entity, relation, and tail entity with triple relations, f represents any random entity other than these, and Wr∈R k×d It is the mapping matrix corresponding to relation r, e∈R d It is the corresponding vector, and F is the activation function, which defaults to F(x) = -lnσ(x).

[0041] 4.3) Repeat the propagation-aggregation operation between steps 3.1 and 4.2 multiple times until the loss function is less than the set threshold, which can be set to 0.003. Update the node representation information of the new book node and user node once each time.

[0042] Step 5) Book recommendation prediction, specifically including:

[0043] 5.1) Merge the book nodes and user nodes obtained from the last N update iterations to form complete node information. This can avoid the error caused by considering only one node update.

[0044] The N update iterations are preferably 100.

[0045] 5.2) Calculate the similarity scores between all complete node information and user nodes to form a book-user similarity matrix Y. Specifically, this involves multiplying the elements of two node vectors once and then summing the results to form an m×n dimensional book-user similarity matrix Y. The elements in the matrix Y... ij This is the rating given by the i-th user to the j-th book, specifically: Where: u i ∈R d v represents the vector of the i-th user. j ∈R d Let represent the vector of the j-th book, m be the total number of users, n be the total number of books, and * represent vector multiplication.

[0046] 5.3) For any user i's book borrowing prediction, in the book-user similarity matrix Y, the user's corresponding row number i, the scores in that row are sorted by size, and the books corresponding to the top of the sort are the books recommended to the user by the recommendation system.

[0047] Preferably, the user's borrowing information, book attribute information, and user-book similarity matrix generated by the algorithm are simultaneously input into the visualization system for user book recommendation.

[0048] like Figure 2 The diagram shows a book recommendation system implementing the above method in this embodiment. It includes: a front-end book recommendation module, a book category switching module, a book information display module, and a book other information selection module, as well as a server. Specifically: the book recommendation module performs graph layout processing based on knowledge graph information to display the knowledge graph to the user; the book category switching module performs filtering processing based on book category information to obtain results for books of a specified category; the book information display module performs display processing based on library collection information to display books in a list format; the book other information selection module performs filtering processing based on knowledge graph information to obtain results for books with specified information; and the server performs filtering processing based on the book information received from the front-end and returns the filtered data in the knowledge graph linked to the received book information to the front-end.

[0049] The book recommendation module is implemented using a graph layout view. In the graph, dots represent nodes in the knowledge graph, with different colors indicating different node types. Lines connecting dots represent relationships, with different line colors indicating different relationship types. It needs to include two types of book nodes connected to user nodes: books borrowed by the user and books recommended by the system. This displays both the system-recommended books and the reasons for the recommendations.

[0050] The aforementioned book information display refers to presenting complete book information in a list format, including both the book's own attributes and a rating indicating whether the user recommends borrowing it. This list is categorized into two types: a list of books currently borrowed by the user and a list of books recommended by the system, which can be switched using module buttons.

[0051] The aforementioned selection of other book information refers to displaying other book categories in the form of a word cloud. These other categories include all book attributes except for ISBN and book category. By clicking on the word cloud to interactively select a category, the book recommendation module and book category switching module will only display books containing that attribute, thus achieving selective book recommendations.

[0052] like Figure 2As shown, this embodiment uses a front-end and back-end interactive software architecture, including a web front-end page and a back-end server, as the visualization system. Front-end languages ​​such as HTML, CSS, and JavaScript are used to design web pages, and data visualization tools echarts.js and d3.js are used to visualize vector graphics. The back-end uses the Python-based web server framework Django for front-end and back-end communication. The front-end sends parameters to the server via HTTP requests, and the server receives the data, converts it to JSON format, and then forwards it to the back-end function interface for processing. Asynchronous page value transfer is achieved via AJAX.

[0053] Through specific practical experiments, the Shanghai Library collection dataset was used, which includes the library's collection information data and library book borrowing data. Borrowing information and book information were processed to construct a complete knowledge graph of the collection books containing users, books and book information. There are 36,712 users with high borrowing frequency and a total of 60,626 books borrowed. The training set, validation set and test set were divided according to the ratio of 7:1:2. The above method was implemented under the specific environment settings of Python 3.6 and Torch 1.6. The embedding size was set to 64, the Adam optimizer was used, the number of samples input to the model at one time was 512, the learning rate was 0.005, and the number of training epochs was 500. Three knowledge graph recommendation algorithms, CKE[1], RippleNet[2] and KGAT[3], were selected for comparison. The results are shown in the table below. The recall rate was 0.1325 and the ndcg value was 0.1013. Figure 3 and Figure 4 As shown. CKE 0.0743 0.0755 RippleNet 0.0913 0.0843 KGAT 0.1211 0.0944 This invention 0.1325 0.1013

[0054] The total training time for each 100 epcohs in this invention is 1.3 hours, while the total training time for the KGAT algorithm is 2.2 hours. This invention outperforms other recommendation algorithms in terms of recall, ndcg value, and training time.

[0055] In summary, this invention has better performance than approximate recommendation algorithms, achieving better accuracy and recall under the top 80 recommendation standard, resulting in a greater performance improvement.

[0056] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A book recommendation method based on graph attention algorithm, characterized in that, include: Step a) Data import: Clean and import the user-book borrowing dataset and the book information dataset. Based on the book classification information and ISBN conversion method provided by the users, convert all book categories into ISBN information and cut them. Step b) Constructing a knowledge graph: First, the introduction information is segmented and cleaned into keyword information. Then, the relationships of all book information, including the keyword information, are statistically analyzed, with the numbers incremented from 0, to construct a triple of (book ID, relationship, book attribute). Then, the user-book dataset is processed to obtain a triple of (user ID, borrowing relationship, book ID), thus completing the construction of the knowledge graph. Step c) The final knowledge graph containing users, books, and book attributes is input to construct and train a graph attention recommendation model, and the resulting book similarity scoring matrix is ​​used for front-end visualization. The book recommendation method specifically includes: Step 1) Processing borrowing and book information, specifically including: 1.1) First, extract all book relationships and mark all book relationships as integers starting from 0 and increasing. There is a triple relationship between each book and its attribute, namely (book id, relationship, book attribute), where the relationship refers to author, publisher, and book category; 1.2) Clean the book attribute information and then perform special data processing on the book category data, with each book category corresponding to an ISBN number; For books lacking an ISBN number: convert the book category into an ISBN number and then split it according to the ISBN tree structure, using the first-level, second-level, and third-level ISBN numbers as supplementary book information; 1.3) Extract keyword information from the book synopsis by using other attribute information of the existing book as keywords, and then replace the synopsis with the keyword information in the synopsis as the book information triple; 1.4) Add information on book popularity. By using book borrowing data, we can obtain the number of times each book has been borrowed. Then, we select a threshold n. Books with a borrowing count greater than n are considered popular books. 1.5) Number all book information using an incrementing integer sequence with a minimum value of 0. Book information and numbers are in one-to-one correspondence. After random sorting, number all books using an incrementing integer sequence following the maximum book information number. Books and numbers are in one-to-one correspondence. Finally, number users using an incrementing integer sequence following the maximum book number. Step 2) Knowledge graph construction, specifically including: constructing a knowledge graph (book ID, relation group, book attribute) by merging triples (book ID, borrowing relation, book ID), encoding all relation groups, and obtaining the matrix W corresponding to each relation. i Integrate all triples, sort and encode all user nodes, book nodes, and attribute nodes, and transform all triples into corresponding d-dimensional entity vectors and k-dimensional relation vectors; Step 3) Knowledge graph propagation information calculation: First, calculate the propagation information to book nodes, then calculate the propagation information to user nodes. The knowledge graph is updated based on the information propagation and aggregation between book nodes and user nodes. Specifically, this includes: 3.1) Calculate the propagation information strength from the neighbor nodes of the book node to the book node using the relational space transformation method of knowledge representation learning. Specifically, this involves calculating the propagation information strength from the neighbor nodes to the book node. Where: π1∈R, e h ∈R d First vector of book vector, e r ∈R k It is the tail vector of book attributes, matrix W r ∈R k×d is the mapping matrix corresponding to relation r, and g is the activation function tanh. This way, we can obtain the propagation information of all triples for each book node. 3.2) Calculate the propagation information strength from book nodes to user nodes. This is done by randomly selecting book nodes, where each book's attribute nodes have a simple frequency relationship f with the user, and then multiplying the information from adjacent nodes by the propagation information strength and summing the results. Specifically, this is: Propagation Information Strength from Book Nodes to User Nodes. Where: π2∈R, J is the set of all borrowed books of the user node, I j It is the set of all attribute nodes connected to the book node, π1∈R is the propagation strength of the neighbor node to the book node in step 3.1, the superscript i is its number order relative to the user, and P is a random function representing that in the t∈N-th iteration, f i The probability is 1 if it is true and 0 otherwise. 3.3) Calculate the propagation information for each book node and user node separately. Specifically, sum the propagation information intensity π1 of the book node's neighbor nodes obtained in step 3.1 to obtain the propagation information of the book node. The propagation information of the user node is obtained by summing the propagation information intensity π2 obtained from the book node in step 3.

2. J represents the set of all attributes of a book and the set of all borrowed books of a user node, respectively, where: e Nu ∈R d The propagation information representing user nodes, e Ni ∈R d The propagation information represents the book node, and i represents the sorting of the corresponding set; Step 4) Aggregate the information propagated to the book nodes and the information propagated to the user nodes, and calculate the loss function, specifically including: 4.1) Summing is performed using the Bi-Interaction method, based on the information of the input node itself and the information propagated to the node. Specifically, this is the node vector of the t-th hop of the book node. The node vector of the t-th hop of the book node ,in: ∈R d e Nu ∈R d The propagation information representing user nodes, e Ni ∈R d Information disseminated representing book nodes, Represents element-wise multiplication, W1∈R d×d and W2∈R d×d These represent the corresponding weight matrices. If there are no weights, they are set as a diagonal matrix I. The node information and propagation information of the previous hop are aggregated and updated to form the node information of the next hop. User nodes and book nodes are aggregated and processed serially, with the same number of loops. 4.2) Calculate the loss function Where: T represents the set of knowledge graph triples, h, r, t represent the head entity, relation, and tail entity with triple relations, f represents any random entity other than these, and Wr∈R k×d It is the mapping matrix corresponding to relation r, e∈R d It represents the corresponding vector, and F is the activation function, which defaults to 0. ; 4.3) Repeat the propagation-aggregation operation between steps 3.1 and 4.2 multiple times until the loss function is less than the set threshold. Update the node representation information of the new book node and user node once each time. Step 5) Book recommendation prediction, specifically including: 5.1) Merge the book nodes and user nodes obtained from the last N update iterations to form complete node information; 5.2) Calculate the similarity scores between all complete node information and user nodes to form a book-user similarity matrix Y. Specifically, this involves multiplying the elements of two node vectors once and then summing them to form an m×n dimensional book-user similarity matrix Y. The elements in the matrix Y... ij This is the rating given by the i-th user to the j-th book, specifically: , where: u i ∈R d v represents the vector of the i-th user. j ∈R d Let represent the vector of the j-th book, m be the total number of users, n be the total number of books, and * represent vector multiplication; 6.3) For any user i's book borrowing prediction, in the book-user similarity matrix Y, the user's corresponding row number i, the scores in that row are sorted by size, and the books corresponding to the top of the sort are the books recommended to the user by the recommendation system.

2. The book recommendation method based on graph attention algorithm according to claim 1, characterized in that, The graph attention recommendation model includes an input layer, an information propagation layer, an information aggregation and generation layer, and a recommendation result generation layer. The input layer randomly generates a vector e for each node as an initial representation based on the input knowledge graph. The information propagation layer calculates the information propagated by each node to its surrounding nodes. The information aggregation layer aggregates the propagated information and node information to generate next-hop information. The recommendation result generation layer scores the data based on the training results and generates a book similarity scoring matrix.

3. The book recommendation method based on graph attention algorithm according to claim 1 or 2, characterized in that, The graph attention recommendation model described above uses the following loss function during training: Where: T represents the set of knowledge graph triples, h, r, t represent the head entity, relation, and tail entity with triple relations, f represents any random entity other than these, and Wr∈R k ×d It is the mapping matrix corresponding to relation r, e∈R d It represents the corresponding vector, and F is the activation function, which defaults to 0. .

4. The book recommendation method based on graph attention algorithm according to claim 1, characterized in that, The book category refers to the classification corresponding to the book's ISBN number. This classification is presented in a tree diagram, which conforms to the data structure characteristics of ISBNs. Each time a book category name is clicked in the book category segmentation module, books containing that category will be filtered in the recommended books. The book recommendation module will only display the recommendations of the filtered books.

5. The book recommendation method based on graph attention algorithm according to claim 1, characterized in that, The system inputs user borrowing information, book attribute information, and a user-book similarity matrix generated by the algorithm into a visualization system for user book recommendations.

6. A book recommendation system based on graph attention algorithm that implements the method of any one of claims 1-5, characterized in that, include: The system comprises a front-end module for book recommendation, a module for switching book categories, a module for displaying book information, a module for selecting other book information, and a server-side component. Specifically: the book recommendation module uses knowledge graph information to perform graph layout processing and displays the knowledge graph to the user; the book category switching module filters books based on category information to obtain results for the specified category; the book information display module displays books in a list format based on the library's collection information; the other book information selection module filters books based on knowledge graph information to obtain results for the specified book information; and the server-side filters the book information received from the front-end and returns the filtered data to the front-end, which is then linked to the received book information in the knowledge graph.

Images