CN117743684A - Book recommendation method based on graph attention algorithm - Google Patents
Book recommendation method based on graph attention algorithm Download PDFInfo
- Publication number
- CN117743684A CN117743684A CN202311737128.1A CN202311737128A CN117743684A CN 117743684 A CN117743684 A CN 117743684A CN 202311737128 A CN202311737128 A CN 202311737128A CN 117743684 A CN117743684 A CN 117743684A
- Authority
- CN
- China
- Prior art keywords
- book
- information
- node
- user
- propagation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 239000013598 vector Substances 0.000 claims abstract description 32
- 239000011159 matrix material Substances 0.000 claims abstract description 29
- 238000012549 training Methods 0.000 claims abstract description 15
- 238000004220 aggregation Methods 0.000 claims abstract description 13
- 230000002776 aggregation Effects 0.000 claims abstract description 10
- 238000010276 construction Methods 0.000 claims abstract description 8
- 238000012800 visualization Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 13
- 238000010586 diagram Methods 0.000 claims description 8
- 230000000644 propagated effect Effects 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000004140 cleaning Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 5
- 238000012216 screening Methods 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 230000005540 biological transmission Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000010365 information processing Effects 0.000 claims description 3
- 230000000903 blocking effect Effects 0.000 claims description 2
- 230000008676 import Effects 0.000 claims description 2
- 239000013589 supplement Substances 0.000 claims description 2
- 238000005516 engineering process Methods 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 239000003086 colorant Substances 0.000 description 2
- 238000011426 transformation method Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013079 data visualisation Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000005295 random walk Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A book recommendation method based on a graph attention algorithm is characterized in that a knowledge graph containing user, book and book attributes is input into a graph attention recommendation model obtained through construction and training, and a book similarity scoring matrix is obtained and used for front-end visualization. The graphical attention recommendation model comprises: an input layer, an information propagation layer, an information aggregation generation layer, and a recommendation 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 information propagation of surrounding nodes received by each node, the information aggregation layer aggregates the propagation information and the node information to generate next-hop information, and the recommendation result generation layer scores according to training results and generates a book similarity scoring matrix. The invention excavates the high-order path of the knowledge graph through the graph attention method, adopts the improved knowledge graph propagation algorithm and the aggregation method and aims at knowledge graph construction of the library, thereby remarkably improving the analysis efficiency and performance of book recommendation.
Description
Technical Field
The invention relates to a technology in the field of knowledge maps, in particular to a book recommendation method based on a graph attention algorithm.
Background
The knowledge graph recommendation algorithm is designed to solve the cold start problem, but has the following disadvantages: firstly, the traditional algorithm does not distinguish book nodes and user nodes and does not fully utilize the characteristics of a knowledge graph; secondly, the node relation between the user and the adjacent books is single, and the user lacks of the difference, and is more similar to GCN instead of GAN; finally, the mix of the graph-based network and collaborative filtering results in high time-consuming costs.
Disclosure of Invention
Aiming at the problems of cold start, overlong training time, no good coding of user-book relation and the defects of incapability of capturing long-range connectivity, incapability of explaining high-order modeling and overlarge expense in the prior art, the invention provides a book recommendation method based on a graph attention algorithm, which is based on a graph attention technology and a random walk technology, and adopts an improved knowledge graph propagation algorithm, an improved aggregation method and knowledge graph construction aiming at library books to excavate a high-order path of a knowledge graph by the graph attention method, thereby remarkably improving the analysis efficiency and performance of book recommendation.
The invention is realized by the following technical scheme:
the invention relates to a book recommendation method based on a graph attention algorithm and based on the system, which comprises the following steps:
step 1) data import: the user-book borrowing data set and the book information data set are cleaned and imported, and the user is required to give book classification information and ISBN conversion method for converting all book classifications into ISBN information and cutting.
Step 2) constructing a knowledge graph: firstly, dividing words and cleaning profile information into keyword information, then counting all book information relations containing the keyword information, increasing numbers from 0, constructing (book id, relation, book attribute) triples, and then similarly processing a user-book data set to obtain (user id, borrowing relation, book id) triples, thus completing knowledge graph construction.
And 3) inputting and constructing the knowledge graph finally containing the user, the book and the book attribute, and training an obtained graph annotation meaning recommendation model to obtain a book similarity scoring matrix for front-end visualization.
The drawing meaning force recommendation model comprises the following steps: an input layer, an information propagation layer, an information aggregation generation layer, and a recommendation 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 information propagation of surrounding nodes received by each node, the information aggregation layer aggregates the propagation information and the node information to generate next-hop information, and the recommendation result generation layer scores according to training results and generates a book similarity scoring matrix.
The loss function adopted in the training of the graph annotating force recommendation model is thatWherein: t represents a knowledge graph triplet set, h, R and T represent head entities, relations and tail entities with triplet relations, f represents any random entity except the random entity, and Wr epsilon R k ×d Is a mapping matrix corresponding to the relation R, e E R d Is the corresponding vector, F is the activation function, default toF(x)＝-lnσ(x)。
Technical effects
According to the method, the long-range connectivity in the knowledge graph of the library book is captured through the graph annotation meaning recommendation model, so that the information which cannot be mined by the traditional recommendation algorithm can be mined, and the training time is reduced. By adding randomness to the propagation-aggregation process of book information to user information, the system has faster training time than other recommendation algorithms, and better performance can be achieved with less than half of model training time. In the knowledge graph construction process, more diversified ISBN information is introduced, and information expression of whether books are popular or not and information expression of whether books are complete or not due to randomness are from single classification information to multi-level classification structure.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a schematic diagram of a system according to the present invention;
FIG. 3 is a graph of recall results;
FIG. 4 is a graph of NDCG results;
FIG. 5 is a schematic diagram of a recommended model structure;
FIG. 6 is a flow chart of an embodiment;
FIG. 7 is a flow chart of step 2;
FIG. 8 is a flow chart of step 3;
FIG. 9 is a flow chart of step 4;
FIG. 10 is a schematic diagram of a book-centric knowledge graph structure;
Detailed Description
As shown in fig. 1 and fig. 6, this embodiment relates to a book recommendation method based on a graph attention algorithm, and after a graph attention force recommendation model based on a knowledge graph shown in fig. 4 is constructed and trained, each candidate book is scored and recommended for an item with highest score according to a book knowledge graph shown in fig. 10, which is constructed according to book information to be processed, specifically including:
step 1) borrowing information and book information processing, which specifically comprises the following steps:
1.1 All book relations are first extracted, all book relations are marked as integers which are increased from 0, and a triple relation exists between each book and the attribute of each book, namely (book id, relation, book attribute), wherein the relation refers to an author, a publisher, a book category and the like, such as (towards the sunset, the author, the roux).
The book category refers to: the book ISBN number corresponds to a classification which is expanded in the form of a tree diagram, which conforms to the data structure characteristics of ISBN. Each time the book category names in the book category blocking module are clicked, books containing the category in the recommended books are screened, and the book recommendation module only displays the recommendation of the screened books.
1.2 Data cleaning is carried out on book attribute information, then special data is needed to process book category data, each book category data corresponds to one ISBN number, books lacking ISBN numbers are needed to be converted into ISBN numbers, the ISBN numbers are split according to the tree structure of the ISBN to obtain primary, secondary and tertiary ISBN numbers which are used as book information supplement.
1.3 Extracting key word information in the brief introduction of the book by taking other attribute information of the existing book as key words, and replacing the key word information in the brief introduction with the brief introduction to be used as a book information triplet.
1.4 The book hot degree information is added, the borrowed times of each book can be obtained through book borrowing data, then a threshold value n is selected, books with borrowed times larger than n are regarded as hot books, books with properties of hot and cold are regarded as cold data, books with borrowed times smaller than n are regarded as cold, and the ratio of the hot and cold books can be set to be close to 1: an integer of 19.
1.5 Numbering all books, wherein the number is an increasing integer sequence with the minimum value of 0, the books are numbered after being ordered randomly, the number is an increasing integer sequence after the maximum number of the books, the books are in one-to-one correspondence, and finally, the user is numbered, and the number is an increasing integer sequence after the maximum number of the books.
Step 2) knowledge graph construction, which specifically comprises the following steps: as shown in fig. 7, a knowledge graph (book id, relation group, book id) is built by combining triples (book id, relation group, book attribute) and (user id, borrowing relation, book id), all relation groups are encoded, and a matrix W corresponding to each relation is obtained i The method comprises the steps of carrying out a first treatment on the surface of the And integrating all triples to sort and encode all user nodes, book nodes and attribute nodes, and converting all triples into entity vectors and k-dimensional relation vectors corresponding to d dimensions.
Step 3) calculating the knowledge graph propagation information, namely calculating the propagation information propagated to the book nodes firstly, and then calculating the propagation information propagated to the user nodes, and carrying out knowledge graph update by information propagation and information aggregation of the book nodes and the user nodes, wherein the knowledge graph update method specifically comprises the following steps as shown in fig. 8:
3.1 According to a relation space transformation method (TransR) for knowledge representation learning, calculating the intensity of the propagation information propagated from the neighbor nodes of the book node to the book node, specifically: propagation information strength pi from neighbor node to book node 1 ＝(W r e t ) T g(W r e h +e r ) Wherein: pi 1 ∈R，e h ∈R d Book vector of first vector, e r ∈R k Is the tail vector book attribute vector, matrix W r ∈R k×d Is a mapping matrix corresponding to the relation r, g is an activation function tanh. Thus, each book node has the propagation information of all triples.
Preferably, the intensity of the propagated information propagated to the book nodes is exponentially normalized.
3.2 Calculating the propagation information intensity from book node to user node, namely randomly proceeding by book node, multiplying the propagation information intensity by adjacent node information after the attribute node possessed by each book has a simple frequency relation f for the user, and summing, specifically: propagation information strength from book node to user nodeWherein: pi 2 E, R, J is the set of all borrowed books of the user node, I j Is the collection of all attribute nodes connected with book nodes, pi 1 E R is the propagation information intensity from the neighbor node to the book node in the step 3.1, the superscript i is the numbering sequence of the neighbor node relative to the user, and P is the random function representing f in the t E N iteration process i The probability of 1 otherwise 0.
For example, when a user borrows five books, where the authors of both books are stun, then the frequency of stun to the user is f=2/5.
Preferably, the strength of the propagation information propagated to the user nodes is exponentially normalized.
3.3 Respectively calculating the propagation information of each book node and user node, specifically: and (3) transmitting the transmission information intensity pi from the neighbor node of the book node obtained in the step 3.1 to the book node 1 Summing to obtain propagation information of book nodesAnd (3) the strength pi of the information transmitted from the book node obtained in the step 3.2 to the user node 2 Summing to obtain propagation information of the user node>J represents the set of all properties of the book and the set of all borrowed books of the user node, respectively, wherein: e, e Nu ∈R d Propagation information representing user nodes e Ni ∈R d Representing the propagation information of book nodes, i representing the ordering of the corresponding set.
Step 4) aggregating the information transmitted to the book node and the information transmitted to the user node, and calculating a loss function, as shown in fig. 9, including:
4.1 Information of the input node itself and information transmitted to the node are summed by a Bi-Interaction method, specifically: node vector of t-th hop of book nodeNode vector of t-th hop of book node +.>Wherein: />e Nu ∈R d Propagation information representing user nodes e Ni ∈R d Propagation information representing book nodes +.>Representing element multiplication, W 1 ∈R d×d And W is 2 ∈R d×d Representing the corresponding weight matrix respectively, no weights can be set as the diagonal matrix I. And aggregating the node information and the propagation information of the previous hop, and updating the node information forming the next hop. The user node and the book node are aggregated and processed serially, and the circulation times are equal.
4.2 Calculating a loss functionWherein: t represents a knowledge graph triplet set, h, R and T represent head entities, relations and tail entities with triplet relations, f represents any random entity except the random entity, and Wr epsilon R k×d Is a mapping matrix corresponding to the relation R, e E R d Is the corresponding vector, F is the activation function, default to F (x) = -lnσ (x).
4.3 Repeating the propagation-aggregation operation between the step 3.1 and the step 4.2 for a plurality of times until the loss function is smaller than the set threshold value, wherein the threshold value can be set to be 0.003, and updating the node representation information of the new book node and the user node once each time.
Step 5) book recommendation prediction, specifically comprising:
5.1 Combining the book node obtained by the last N updating iterations with the user node to form complete node information, so that errors caused by only considering one node updating can be avoided.
The N update iterations are preferably 100.
5.2 Calculating similarity scores of all complete node information and user nodes to form a book-user similarity matrix Y, wherein the similarity scores are specifically as follows: multiplying the vector elements of the two nodes once and summing to form a book-user similarity matrix Y with m multiplied by n dimensions, wherein the elements Y in the matrix ij The scoring of the ith user on the jth book is specifically as follows:wherein: u (u) i ∈R d Representing the ith user vector, v j ∈R d Representing the j-th book vector, m is the total number of users, n is the total number of books, and the representing vectors are multiplied.
5.3 For book borrowing prediction of any user i, the number i of lines corresponding to the user in the book-user similarity matrix Y, and the scoring in the lines are ranked in size, and books corresponding to the ranked book in front are the books recommended to the user by the recommendation system.
Preferably, the user-book similarity matrix generated according to the algorithm and containing the input user borrowing information, book attribute information and is simultaneously input into the visualization system for user book recommendation.
As shown in fig. 2, the book recommendation system for implementing the method according to the present embodiment includes: the book recommendation system comprises a book recommendation module, a book category switching module, a book information display module, a book other information selection module and a server side which are positioned at the front end, wherein: the book recommendation module performs map layout processing according to the knowledge map information, displays the knowledge map to a user, the book type switching module performs screening processing according to the book type information to obtain book results of specified types, the book information display module performs display processing according to the library-collected book information, books are displayed in a list form, the other book information selection module performs screening processing according to the knowledge map information to obtain book results of the specified book information, and the server performs filtering processing according to the book information transmitted by the front end and returns data connected with the transmitted book information in the knowledge map after the filtering of the front end.
The book recommendation module is realized through a diagram layout view, dots in the diagram represent nodes in the knowledge graph, and colors of different dots represent different types of nodes; the dot lines represent relationships, and the colors through the different lines represent different types of relationships. The book recommendation system comprises two book nodes and a user node, wherein the two book nodes are connected with the user node, one book is borrowed by the user, and the other book is recommended to the user by the system, so that the books recommended by the system are displayed, and the reason of the system recommendation is also given.
The book information display means that: the complete information of the book is presented in the form of a list, both containing the properties of the book itself and including a score as to whether the user recommends his borrowing. The list classifies two types-a book list borrowed by the current user and a book list recommended to the user by the system, and can be switched through a module type button
The other information selection of the book is as follows: and displaying other categories of the books in a word cloud form, wherein the other categories are book attributes except ISBN and all attributes except the book categories. The screened category is selected through clicking word cloud interaction, and the book recommending module and the book category switching module only display books with the attribute, so that books are recommended selectively.
As shown in fig. 2, in this embodiment, a front-end and back-end interactive software architecture including a web front-end page and a back-end server is used as a visualization system, a front-end language such as html, CSS, javaScript is used to write a design web page, and data visualization echarts.js and d3.Js are used to perform visualization processing vector diagrams; the back end uses a Web server framework Django based on python to carry out front-back end transfer, wherein the front end transfers parameters to a server through an HTTP request method, the server receives data to be converted into json format and transits to a background function interface to be processed, and asynchronous page transfer is carried out through ajax.
Through concrete practical experiment, adopt Shanghai library collection data set, contain this library collection information data and library books respectively and borrow data, borrow information and book information processing, construct complete collection books knowledge map that contains user, books and book information, high borrow user 36712, general borrow books 60626 originally, according to 7:1:2, dividing the training set, the verification set and the test set in proportion; the method is realized under the specific environment settings of python3.6 and torch1.6, the embedding size is set to be 64, an Adam optimizer is used, the number of samples input to a model at one time is 512, the learning rate is 0.005, the number of epoch training times is 500, three knowledge graph recommendation algorithms of CKE [1], rippleNet [2] and KGAT [3] are respectively selected for comparison, the result is shown in the following table, the obtained recall rate is 0.1325, and the ndcg value is 0.1013, as shown in fig. 3 and 4.
| Recall | Ndcg | |
| CKE | 0.0743 | 0.0755 |
| RippleNet | 0.0913 | 0.0843 |
| KGAT | 0.1211 | 0.0944 |
| The invention is that | 0.1325 | 0.1013 |
The total time consumption of every 100 epcoh training is 1.3 hours, and the total time consumption of the KGAT algorithm is 2.2 hours. The invention is superior to the performance of other recommended algorithms in recall, ndcg value and training time.
In conclusion, the method has better performance than an approximate recommendation algorithm, and can obtain better accuracy and recall under the top80 recommendation standard, thereby obtaining better performance improvement.
The foregoing embodiments may be partially modified in numerous ways by those skilled in the art without departing from the principles and spirit of the invention, the scope of which is defined in the claims and not by the foregoing embodiments, and all such implementations are within the scope of the invention.
Claims (7)
1. The book recommendation method based on the graph attention algorithm is characterized by comprising the following steps of:
step 1) data import: cleaning and importing a user-book borrowing data set and a book information data set, wherein the user is required to give book classification information and an ISBN conversion method for converting all book classifications into ISBN information and cutting;
step 2) constructing a knowledge graph: firstly, dividing words and cleaning profile information into keyword information, counting all book information relations including the keyword information, increasing numbers from 0, constructing (book id, relation, book attribute) triples, and then performing similar processing on a user-book data set to obtain (user id, borrowing relation, book id) triples, so as to complete knowledge graph construction;
and 3) inputting and constructing the knowledge graph finally containing the user, the book and the book attribute, and training an obtained graph annotation meaning recommendation model to obtain a book similarity scoring matrix for front-end visualization.
2. The book recommendation method based on the attention drawing algorithm as claimed in claim 1, wherein the attention drawing recommendation model comprises: an input layer, an information propagation layer, an information aggregation generation layer, and a recommendation 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 information propagation of surrounding nodes received by each node, the information aggregation layer aggregates the propagation information and the node information to generate next-hop information, and the recommendation result generation layer scores according to training results and generates a book similarity scoring matrix.
3. The book recommendation method based on the attention drawing algorithm as claimed in claim 1 or 2, wherein the attention drawing force recommendation model is trained by adopting a loss function ofWherein: t represents a knowledge graph triplet set, h, R and T represent head entities, relations and tail entities with triplet relations, f represents any random entity except the random entity, and Wr epsilon R k×d Is a mapping matrix corresponding to the relation R, e E R d Is the corresponding vector, F is the activation function, default to F (x) = -lnσ (x).
4. A book recommendation method based on a graph attention algorithm according to any one of claims 1 to 3, characterized by comprising:
step 1) borrowing information and book information processing, which specifically comprises the following steps:
1.1 Firstly, all book relations are extracted, all book relations are marked as integers which are increased from 0, and a triplet relation exists between each book and the attribute thereof, namely (book id, relation and book attribute), wherein the relation refers to author, publishing agency and book category;
1.2 Data cleaning is carried out on book attribute information, then special data is needed to process book category data, each book category data corresponds to one ISBN number, book categories of books lacking ISBN numbers are needed to be converted into ISBN numbers, the ISBN numbers are split according to a tree structure of the ISBN to obtain primary, secondary and tertiary ISBN numbers which are used as book information supplement;
1.3 Extracting key word information in the brief introduction of the book by taking other attribute information of the existing book as key words, and replacing the key word information in the brief introduction with the brief introduction to be used as a book information triplet;
1.4 Adding book popularity information, obtaining borrowed times of each book through book borrowing data, and then selecting a threshold value n, wherein books with the borrowed times larger than n are regarded as hot books;
1.5 Numbering all books, wherein the number is an increasing integer sequence with the minimum value of 0, the books are numbered after being ordered randomly, the number is an increasing integer sequence after the maximum number of the books, the books are numbered after being numbered one by one, and finally, the user is numbered, and the number is an increasing integer sequence after the maximum number of the books;
step 2) knowledge graph construction, which specifically comprises the following steps: combining and constructing a knowledge graph (book id, relation group, book id) by using triples (book id, relation group, book attribute) and (user id, borrowing relation, book id), and coding all relation groups to obtain a matrix W corresponding to each relation i The method comprises the steps of carrying out a first treatment on the surface of the Integrating all triples to sort and encode all user nodes, book nodes and attribute nodes, and converting all triples into entity vectors and k-dimensional relation vectors corresponding to d dimensions;
step 3) calculating knowledge graph propagation information, namely calculating propagation information propagated to book nodes firstly, and then calculating propagation information propagated to user nodes, and carrying out knowledge graph update by information propagation and information aggregation of the book nodes and the user nodes, wherein the method specifically comprises the following steps:
3.1 According to the relation space conversion method of knowledge representation learning, calculating the transmission information intensity transmitted from the neighbor node of the book node to the book node, specifically: propagation information strength pi from neighbor node to book node 1 ＝(W r e t ) T g(W r e h +e r ) Wherein: pi 1 ∈R，e h ∈R d Book vector of first vector, e r ∈R k Is the tail vector book attribute vector, matrix W r ∈R k×d Is a mapping matrix corresponding to the relation r, g is activationObtaining the function tanh, so as to obtain the propagation information of all the triples of each book node;
3.2 Calculating the propagation information intensity from book node to user node, namely randomly proceeding by book node, multiplying the propagation information intensity by adjacent node information after the attribute node possessed by each book has a simple frequency relation f for the user, and summing, specifically: propagation information strength from book node to user nodeWherein: pi 2 E, R, J is the set of all borrowed books of the user node, I j Is the collection of all attribute nodes connected with book nodes, pi 1 E R is the propagation information intensity from the neighbor node to the book node in the step 3.1, the superscript i is the numbering sequence of the neighbor node relative to the user, and P is the random function representing f in the t E N iteration process i The probability of (1) is 1 otherwise 0,
3.3 Respectively calculating the propagation information of each book node and user node, specifically: and (3) transmitting the transmission information intensity pi from the neighbor node of the book node obtained in the step 3.1 to the book node 1 Summing to obtain propagation information of book nodesAnd (3) the strength pi of the information transmitted from the book node obtained in the step 3.2 to the user node 2 Summing to obtain propagation information of the user node>Respectively representing a set of all properties of the book and a set of all borrowed books of the user node, wherein: e, e Nu ∈R d Propagation information representing user nodes e Ni ∈R d Representing the propagation information of book nodes, wherein i represents the ordering of the corresponding set;
step 4) aggregating the information transmitted to the book node and the information transmitted to the user node, and calculating a loss function, wherein the method specifically comprises the following steps:
4.1 Information of the input node itself and information transmitted to the node are summed by a Bi-Interaction method, specifically: node vector of t-th hop of book nodeNode vector of t-th hop of book node +.>Wherein: />e Nu ∈R d Propagation information representing user nodes e Ni ∈R d Propagation information representing book nodes +.>Representing element multiplication, W 1 ∈R d×d And W is 2 ∈R d×d Representing corresponding weight matrixes respectively, wherein no weight can be set as a diagonal matrix I, node information and propagation information of the previous hop are aggregated, node information of the next hop is updated and formed, user nodes and book nodes are aggregated and processed in series, and the circulation times are equal;
4.2 Calculating a loss functionWherein: t represents a knowledge graph triplet set, h, R and T represent head entities, relations and tail entities with triplet relations, f represents any random entity except the random entity, and Wr epsilon R k×d Is a mapping matrix corresponding to the relation R, e E R d Is the corresponding vector, F is the activation function, default to F (x) = -lnσ (x);
4.3 Repeating the propagation-aggregation operation between the step 3.1 and the step 4.2 for a plurality of times until the loss function is smaller than a set threshold value, and repeatedly updating and iterating the node representation information of the new book node and the user node once each time;
step 5) book recommendation prediction, specifically comprising:
5.1 Combining book nodes and user nodes obtained by the last N updating iterations to form complete node information, so that errors caused by only considering one node updating can be avoided;
5.2 Calculating similarity scores of all complete node information and user nodes to form a book-user similarity matrix Y, wherein the similarity scores are specifically as follows: multiplying the vector elements of the two nodes once and summing to form a book-user similarity matrix Y with m multiplied by n dimensions, wherein the elements Y in the matrix ij The scoring of the ith user on the jth book is specifically as follows:wherein: u (u) i ∈R d Representing the ith user vector, v j ∈R d Representing the j-th book vector, m is the total number of users, n is the total number of books, and the representing vectors are multiplied;
6.3 For book borrowing prediction of any user i, the number i of lines corresponding to the user in the book-user similarity matrix Y, and the scoring in the lines are ranked in size, and books corresponding to the ranked book in front are the books recommended to the user by the recommendation system.
5. The book recommendation method based on the graph attention algorithm as set forth in claim 4, wherein the book category means: the classification corresponding to the book ISBN number is developed in the form of a tree diagram, which accords with the data structure characteristics of ISBN; each time the book category names in the book category blocking module are clicked, books containing the category in the recommended books are screened, and the book recommendation module only displays the recommendation of the screened books.
6. The book recommendation method based on the attention drawing algorithm of claim 4, wherein the user-book similarity matrix generated according to the algorithm, including the input user borrowing information, book attribute information and the input algorithm, is simultaneously input into the visualization system for user book recommendation.
7. A book recommendation system based on a graph attention algorithm implementing the method of any one of claims 1-6, comprising: the book recommendation system comprises a book recommendation module, a book category switching module, a book information display module, a book other information selection module and a server side which are positioned at the front end, wherein: the book recommendation module performs map layout processing according to the knowledge map information, displays the knowledge map to a user, the book type switching module performs screening processing according to the book type information to obtain book results of specified types, the book information display module performs display processing according to the library-collected book information, books are displayed in a list form, the other book information selection module performs screening processing according to the knowledge map information to obtain book results of the specified book information, and the server performs filtering processing according to the book information transmitted by the front end and returns data connected with the transmitted book information in the knowledge map after the filtering of the front end.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311737128.1A CN117743684A (en) | 2023-12-18 | 2023-12-18 | Book recommendation method based on graph attention algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311737128.1A CN117743684A (en) | 2023-12-18 | 2023-12-18 | Book recommendation method based on graph attention algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117743684A true CN117743684A (en) | 2024-03-22 |
Family
ID=90260307
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311737128.1A Pending CN117743684A (en) | 2023-12-18 | 2023-12-18 | Book recommendation method based on graph attention algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117743684A (en) |
-
2023
- 2023-12-18 CN CN202311737128.1A patent/CN117743684A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110263280B (en) | Multi-view-based dynamic link prediction depth model and application | |
| CN114117220A (en) | Deep reinforcement learning interactive recommendation system and method based on knowledge enhancement | |
| CN112613602A (en) | Recommendation method and system based on knowledge-aware hypergraph neural network | |
| CN106991127B (en) | Knowledge subject short text hierarchical classification method based on topological feature expansion | |
| CN103325061B (en) | A kind of community discovery method and system | |
| CN110175628A (en) | A kind of compression algorithm based on automatic search with the neural networks pruning of knowledge distillation | |
| CN113486190B (en) | Multi-mode knowledge representation method integrating entity image information and entity category information | |
| CN110674407A (en) | Hybrid recommendation method based on graph convolution neural network | |
| CN112487200B (en) | Improved deep recommendation method containing multi-side information and multi-task learning | |
| CN106997553A (en) | Multi-objective optimization-based commodity combination mode mining method | |
| CN103942571A (en) | Graphic image sorting method based on genetic programming algorithm | |
| CN112487020A (en) | Method and system for converting graph of SQL to text into natural language statement | |
| CN114065048A (en) | Article recommendation method based on multi-different-pattern neural network | |
| CN110781940A (en) | Fuzzy mathematics-based community discovery information processing method and system | |
| US20170004417A1 (en) | Method of generating features optimal to a dataset and classifier | |
| CN114254093A (en) | Multi-space knowledge enhanced knowledge graph question-answering method and system | |
| CN112884045A (en) | Classification method of random edge deletion embedded model based on multiple visual angles | |
| CN116166910A (en) | Social media account vermicelli water army detection method, system, equipment and medium | |
| Sun et al. | Graph force learning | |
| CN103914527A (en) | Graphic image recognition and matching method based on genetic programming algorithms of novel coding modes | |
| Hao et al. | Deep graph clustering with enhanced feature representations for community detection | |
| CN107451617A (en) | One kind figure transduction semisupervised classification method | |
| CN118153679A (en) | Semi-supervised entity alignment method based on multi-hop attention mechanism | |
| CN112561599A (en) | Click rate prediction method based on attention network learning and fusing domain feature interaction | |
| CN112905906A (en) | Recommendation method and system fusing local collaboration and feature intersection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |