Enlightenment book recommendation system based on knowledge graph

Abstract

The invention discloses an enlightenment book recommendation system based on a knowledge graph, and relates to the technical field of intelligent recommendation. The invention discloses an enlightenment book recommendation system based on a knowledge graph. The enlightenment book recommendation system comprises a user information acquisition module, a knowledge graph construction module, a personalized model construction module, an enlightenment book recommendation module, a feedback optimization module and a visual interaction module. Through the association recommendation subunits, the system can accurately match enlightenment books suitable for the user according to the cognitive development stage and interest preference of the user, fuzzy association degree calculation ensures that the recommended books are highly matched with the cognitive level of the user, and through missing clustering analysis in the group recommendation unit, the recommendation efficiency of the user is improved. Data of other users in the same cognitive development stage of the user are utilized for analysis, it is ensured that the recommendation result is not only based on individual demands, but also considers group trends, and the diversity and adaptability of recommendation are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation technology, and in particular to an enlightenment book recommendation system based on knowledge graph. Background Art

[0002] With the rapid development of the Internet and the rise of big data, the number and types of online book resources are increasing. People need to spend a lot of time and energy to find books that interest them. Traditional book recommendation systems mainly rely on the rating matrix of users and books, and make recommendations for users by calculating similarities. However, with the growth of the number of users and books, traditional recommendation algorithms face the problems of sparse data and insufficient feature mining, making it difficult to meet the personalized needs of users.

[0003] As a semantic network, the knowledge graph connects entities and relationships in the form of triples and contains a large amount of knowledge. It is widely used in search engines, question-answering systems, and recommendation systems. In the field of book recommendations, the knowledge graph can help mine the potential associations between users and books, alleviate the problem of data sparsity, and improve the performance of the recommendation system.

[0004] In recent years, combining knowledge graphs with recommendation algorithms has become a research hotspot. By constructing a book knowledge graph, modeling entities and relationships in the book field, and combining it with a neural collaborative filtering model, we can fully explore the feature interactions between users and books and improve the accuracy and diversity of recommendations. At present, some studies have applied knowledge graphs to book recommendation systems. For example, an individual book recommendation model based on a knowledge graph is constructed by using the content characteristics of books, the user's historical behavior, and the associated information in the knowledge graph. By introducing neural collaborative filtering technology, users' potential interest preferences can be further explored and the performance of the recommendation system can be improved. However, existing methods still face certain challenges when dealing with large-scale knowledge graph data and complex user behavior characteristics. The recommendation results lack interpretability, and children's authentication is different from that of adults. Effective recommendations cannot be made for preschool children. Summary of the invention

[0005] The present invention proposes an enlightenment book recommendation system based on knowledge graph, which aims to construct an enlightenment book knowledge graph, integrate the user personalization model and the neural collaborative filtering algorithm, realize personalized enlightenment book recommendation, and meet the reading needs of users.

[0006] An enlightenment book recommendation system based on knowledge graph, including the following modules: A user information acquisition module, used to acquire, collect and process user information, including an initial information acquisition unit and a supplementary information acquisition unit; The initial information acquisition unit is used to acquire initial user information through the user's registration information when the user uses the device for the first time; The supplementary information acquisition unit updates the user information based on the interactive information generated by the user; The knowledge graph construction module is used to construct the enlightenment book knowledge graph, obtain enlightenment book information through the enlightenment book database, and generate the enlightenment book knowledge graph, including the book database construction unit, the knowledge graph generation unit and the graph update unit; the knowledge graph consists of a triple (h, r, t), where h represents the head entity, r represents the relationship, and t represents the tail entity; The personalized model building module is used to build a personalized model for users. It generates the user's feature information through the user's basic information. The feature information is stored in a personalized knowledge graph. The user's feature information includes the cognitive development stage. The module includes a feature information unit, a personalized graph generation unit, and a graph update unit. The book recommendation module is used to generate a book recommendation list, including an individual recommendation unit and a group recommendation unit. The individual recommendation unit includes an associated recommendation subunit and a collaborative filtering subunit. The association recommendation subunit generates a candidate book recommendation list through the fuzzy association between the enlightenment book knowledge graph and the personalized graph; The collaborative filtering subunit screens the candidate book recommendation list through the collaborative filtering algorithm to obtain an individual book recommendation list; The group recommendation unit uses the missing clustering algorithm to perform group analysis on the user's personalized model to obtain a group book recommendation list; Feedback optimization module, used to dynamically optimize book recommendations through user interaction information, including feedback mining unit and recommendation optimization unit; Feedback mining unit, used to perform data mining on user interaction information, update the personalized model and enlightenment book knowledge graph, and update the relationship weights of associated recommendations to optimize recommendation results; A recommendation optimization unit, used to weight the initial book recommendation list by interest according to the interest interaction information, so as to obtain the final recommended books; The visual interactive module is used to realize the interactive interface of the system, including the GUI interface, the demand acquisition unit and the usage recording unit. The GUI interface is used to present recommended book information, user operation options and interactive feedback. It obtains user needs, adjusts the personalized model, and records user operations to continuously optimize the user's basic information.

[0007] As a preferred technical solution of the present invention, the group recommendation unit of the book recommendation module includes: Randomly obtain several personalized models at the same cognitive development stage as the user in the personalized model database as group analysis samples, perform attribute composition analysis on the user's personalized model samples, and obtain the composition degree of each attribute of the user's personalized model samples. The composition degree is obtained by the proportion of the value of the corresponding attribute in the group analysis sample. After obtaining the composition degree of all attributes, normalize them so that the sum of the composition degrees is 1; The attributes are grouped according to the set number of groups so that the sum of the composition degrees of each group of attributes is close. The attributes of different groups are subjected to missing cluster analysis. The missing cluster analysis removes the attributes of this group and then clusters them to obtain the clustering results of the group analysis samples under multiple groups of missing attributes. Obtain reading statistics of all personalized models of the same type as the user's personalized model sample in each group of clustering results, and calculate the recommendation value of the enlightenment book corresponding to the reading record based on the credibility of the cluster analysis and the composition of the group of attributes. The credibility is obtained through the intra-cluster similarity of the user's personalized model sample in the cluster category, expressed in percentage form; The formula for calculating the recommendation value of the mth enlightenment book is: ; where x represents the xth group of attributes, θ x represents the composition degree of the x-th group of attributes, γ x represents the credibility of the x-th group of cluster analysis results, i represents the i-th reading statistics of the m-th enlightenment book in a group of cluster analysis results, and O represents the time of reading statistics; Repeat the above process three times to avoid falling into the local optimal solution, calculate the average recommendation value of the results of the three group analyses, sort the average recommendation values ​​of the enlightenment books, and take the top-ranked books as the group book recommendation list.

[0008] As a preferred technical solution of the present invention, the association recommendation subunit in the individual recommendation unit generates a candidate book recommendation list by calculating the association between the personalized model in the personalized knowledge graph and each enlightenment book in the enlightenment book knowledge graph; Specifically, the enlightenment books are stage-limited according to the user's cognitive development stage, and the fuzzy correlation is calculated for the enlightenment books in the stage-limited stage; the semantic similarity α between the current user's personalized model and the personalized model of the same cognitive development stage is obtained. The semantic similarity is only obtained through the user's personal feature information, and the value range is (0,1). In all personalized models of the same cognitive development stage, reading restrictions are made based on whether the book for which the fuzzy correlation is currently calculated has been read. A number of personalized model samples j are randomly extracted from the personalized models with reading restrictions, and the sum of the semantic similarities of all extracted personalized model samples is M Q Satisfies the following restrictions: β1j<M Q≤β2j, where the value range of β1 and β2 is (0,1). The two are fuzzy adjustment parameters. The difference between the two determines the randomness of the sample. The larger the difference, the higher the randomness of the sample. For the result, the local optimal solution can be avoided. The larger the absolute value of β1 and β2, the stronger the specific pairing ability of the sample. The formula for calculating the fuzzy correlation of the nth enlightenment book is: ; Where T a 、T b and T c They represent the proportion of types, keywords and themes of enlightenment books in the stage limit, respectively. j represents the semantic similarity of the jth user, R j represents the reading value of the jth user, obtained by reading time and understanding level, I j Indicates the proportion of types in the reading records of the jth user; Sort the relevance of enlightenment books and obtain the top-ranked books as a candidate book recommendation list.

[0009] As a preferred technical solution of the present invention, the collaborative filtering subunit in the book recommendation module performs neural collaborative filtering on the enlightenment books in the candidate book recommendation list through the individual book recommendation model to obtain the book recommendation list; The individual book recommendation model includes a cross compression sub-unit, a book knowledge graph embedding sub-unit, and a unit book recommendation sub-unit; The cross compression subunit is used to connect the book recommendation subunit and the book knowledge graph embedding subunit, and realize the feature interaction between the entity vector in the knowledge graph embedding subunit and the book vector in the book recommendation subunit by constructing a cross matrix; The book knowledge graph embedding subunit uses a multi-layer perceptron to transform entities and relationships in the enlightenment book knowledge graph into low-dimensional dense entity vectors; The book recommendation subunit is used to use a multi-layer perceptron to personalize the user feature vector in the graph; the user feature vector and the book feature vector are input into the neural collaborative filtering network, and finally the probability of the user interacting with the candidate books is predicted to obtain an individual book recommendation list.

[0010] As a preferred technical solution of the present invention, the recommendation optimization unit includes: Integrate the individual book recommendation list and the group book recommendation list as a book recommendation list, display the book recommendation list to the target user children on a visual interface, and obtain the children's interest interaction information on different enlightenment books in the book recommendation list. The interest interaction information is obtained through image analysis algorithms and / or user parent feedback information; The interest weight is obtained through interest interaction information, and the enlightenment books with interest weight are weighted to obtain the final recommended books.

[0011] As a preferred technical solution of the present invention, the construction of the subunit; The cross compression subunit performs explicit interaction operations on the book vector and entity vector in the d-dimensional feature space to obtain a d*d dimensional cross feature matrix, maps the matrix back to the d-dimensional feature space, outputs the book vector and entity vector of the next layer, completes the compression operation, and obtains the book feature vector v L , the formula is: e l+1 =C(v l ,e l )[e],v l+1 =C(v l ,e l )[v], where v l represents the book vector of the lth layer, e l Represents the entity vector of the lth layer, [e] and [v] are used to distinguish the entity vector output and the book vector output in the cross compression subunit; The book knowledge graph embedding subunit uses an L-layer multi-layer perceptron to extract the relationship feature vector r between entities L , input the entity vector into the cross compression subunit of the L layer to obtain the entity feature vector: h L =E v~s(h) [C L (v,h)[e]], where s(h) represents the set of books v associated with book entity h; the relationship feature vector and the entity feature vector are concatenated and sent to the K-layer multi-layer perceptron to obtain the predicted entity feature vector t ^ , use the normalized inner product function to calculate the score of the triplet of the predicted entity: score(h,r,t)=f KG (t,t ^ );where f KG represents the normalized inner product function, t represents the tail entity of the triple, and t ^ Represents the predicted entity feature vector; The book recommendation subunit uses a multi-layer perceptron with L layers to extract the user feature vector u L , the extracted user feature vector and book feature vector are input into the neural collaborative filtering network, and the user features and book features are interacted through generalized matrix decomposition and multi-layer perceptron to output recommended books; Specifically, the element product of the user feature vector and the book feature vector is calculated, and generalized matrix decomposition is performed: , where the two elements represent the corresponding eigenvectors of the generalized matrix decomposition; Input the user feature vector and the book feature vector into the multi-layer perceptron to obtain the output y of the Lth layer MLP ; The final output of the computational model is: , where ρ represents the weight vector.

[0012] As a preferred technical solution of the present invention, the loss function of the individual book recommendation model consists of three parts: a book recommendation module loss function, a book knowledge graph embedding loss function and a regularization term. The book recommendation module loss function is a loss value for calculating the actual click-through rate and predicted click-through rate of the model, and the regularization term is used to prevent overfitting.

[0013] As a preferred technical solution of the present invention, the construction of the knowledge graph includes a pattern layer and a data layer; The model layer determines the entity concepts of the domain, the classes corresponding to the entity concepts, the attributes and attribute value ranges of the entities, and the attribute description features by analyzing the relevant knowledge of the domain; establishes a local library for the corresponding domain to build a knowledge graph; The data layer obtains data through the database, cleans the data, obtains the nodes and relationships of the data, and uses the graph database to store the knowledge graph.

[0014] As a preferred technical solution of the present invention, the personalized model is constructed by the user's characteristic information, and the characteristic information is obtained by processing the user information. The personalized model includes the user's personal characteristic information, reading records and statistics, book search and browsing records and statistics; wherein the user's personal characteristic information includes the user's cognitive development stage and family profile, and the personalized model is used to construct a personalized knowledge graph. The user's cognitive development stage is obtained by inputting or observing children's behavior, language ability, interests and hobbies.

[0015] The present invention has the following advantages: Through associating recommendation sub-units, the system of the present invention can accurately match enlightenment books suitable for users according to the users' cognitive development stage and interest preferences. The fuzzy correlation calculation ensures that the recommended books are highly consistent with the users' cognitive level, avoiding the recommendation of age-inappropriate content. In the group recommendation unit, missing cluster analysis is performed by using the data of other users at the same cognitive development stage as the user to ensure that the recommendation results are not only based on individual needs, but also take group trends into consideration, thereby avoiding the recommendation limitations caused by over-personalization and enhancing the diversity and adaptability of recommendations.

[0016] The present invention combines the advantages of individual recommendation and group recommendation. It can not only provide accurate recommendations based on the user's personalized model, but also supplement the recommended content through group data, thereby improving the coverage and accuracy of the recommendation, especially in the segmentation and adaptation of the cognitive development stage, and can provide suitable enlightenment books for children at different stages; through real-time analysis of user interest interaction and feedback, the recommendation system can continuously optimize the recommended content to adapt to changes in user cognition and interests, thereby improving the timeliness and accuracy of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of the structure of an enlightenment book recommendation system based on a knowledge graph adopted in an embodiment of the present invention.

[0018] Figure 2 It is a schematic diagram of the enlightenment book knowledge graph in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] An enlightenment book recommendation system based on knowledge graph, see Figure 1 and Figure 2 , including the following modules: Module S1: User information acquisition module, used to acquire, collect and process user information, including an initial information acquisition unit and a supplementary information acquisition unit; The initial information acquisition unit is used to acquire initial user information through the user's registration information when the user uses the device for the first time; The supplementary information acquisition unit updates the user information based on the interactive information generated by the user; Module S2: Knowledge graph construction module, used to construct the enlightenment book knowledge graph, obtain enlightenment book information through the enlightenment book database, and generate the enlightenment book knowledge graph, including the book database construction unit, the knowledge graph generation unit and the graph update unit; the knowledge graph consists of a triple (h, r, t), where h represents the head entity, r represents the relationship, and t represents the tail entity; The construction of the knowledge graph includes a model layer and a data layer; The model layer determines the entity concepts of the domain, the classes corresponding to the entity concepts, the attributes and attribute value ranges of the entities, and the attribute description features by analyzing the relevant knowledge of the domain; establishes a local library for the corresponding domain to build a knowledge graph; The data layer obtains data through the database, cleans the data, obtains the nodes and relationships of the data, and uses the graph database to store the knowledge graph.

[0021] Module S3: Personalized model building module, used to build a personalized model of the user, generate the user's feature information through the user's basic information, and the feature information is stored in a personalized knowledge graph. The user's feature information includes the cognitive development stage. The module includes a feature information unit, a personalized graph generation unit, and a graph update unit; The personalized model is constructed by the user's characteristic information, which is obtained by processing the user information. The personalized model includes the user's personal characteristic information, reading records and statistics, book search and browsing records and statistics; the user's personal characteristic information includes the user's cognitive development stage and a brief family situation. The personalized model is used to form a personalized knowledge graph. The user's cognitive development stage is obtained by inputting or observing children's behavior, language ability, interests and hobbies.

[0022] Module S4: book recommendation module, used to generate a book recommendation list, including an individual recommendation unit and a group recommendation unit, the individual recommendation unit includes an associated recommendation subunit and a collaborative filtering subunit; The association recommendation subunit generates a candidate book recommendation list through the fuzzy association between the enlightenment book knowledge graph and the personalized graph; The associated recommendation subunit in the individual recommendation unit generates a candidate book recommendation list by calculating the association between the personalized model in the personalized knowledge graph and each enlightenment book in the enlightenment book knowledge graph; Specifically, the enlightenment books are stage-limited according to the user's cognitive development stage, and the fuzzy correlation is calculated for the enlightenment books in the stage-limited stage; the semantic similarity α between the current user's personalized model and the personalized model of the same cognitive development stage is obtained. The semantic similarity is only obtained through the user's personal feature information, and the value range is (0,1). In all personalized models of the same cognitive development stage, reading restrictions are made based on whether the book for which the fuzzy correlation is currently calculated has been read. A number of personalized model samples j are randomly extracted from the personalized models with reading restrictions, and the sum of the semantic similarities of all extracted personalized model samples is M Q Satisfies the following restrictions: β1j<M Q ≤β2j, where the value range of β1 and β2 is (0,1). The two are fuzzy adjustment parameters. The difference between the two determines the randomness of the sample. The larger the difference, the higher the randomness of the sample. For the result, the local optimal solution can be avoided. The larger the absolute value of β1 and β2, the stronger the specific pairing ability of the sample. The formula for calculating the fuzzy correlation of the nth enlightenment book is: ; Where T a , T b and T cThey represent the proportion of types, keywords and themes of enlightenment books in the stage limit, respectively. j represents the semantic similarity of the jth user, R j represents the reading value of the jth user, obtained by reading time and understanding level, I j Indicates the proportion of types in the reading records of the jth user; Sort the relevance of enlightenment books and obtain the top-ranked books as a candidate book recommendation list.

[0023] The collaborative filtering subunit screens the candidate book recommendation list through the collaborative filtering algorithm to obtain an individual book recommendation list; The collaborative filtering subunit in the book recommendation module performs neural collaborative filtering on the enlightenment books in the candidate book recommendation list through the individual book recommendation model to obtain a book recommendation list; The individual book recommendation model includes a cross compression sub-unit, a book knowledge graph embedding sub-unit, and a unit book recommendation sub-unit; The cross compression subunit is used to connect the book recommendation subunit and the book knowledge graph embedding subunit, and realize the feature interaction between the entity vector in the knowledge graph embedding subunit and the book vector in the book recommendation subunit by constructing a cross matrix; The book knowledge graph embedding subunit uses a multi-layer perceptron to transform entities and relationships in the enlightenment book knowledge graph into low-dimensional dense entity vectors; The book recommendation subunit is used to use a multi-layer perceptron to personalize the user feature vector in the graph; the user feature vector and the book feature vector are input into the neural collaborative filtering network, and finally the probability of the user interacting with the candidate books is predicted to obtain an individual book recommendation list.

[0024] Construction of said subunit; The cross compression subunit performs explicit interaction operations on the book vector and entity vector in the d-dimensional feature space to obtain a d*d dimensional cross feature matrix, maps the matrix back to the d-dimensional feature space, outputs the book vector and entity vector of the next layer, completes the compression operation, and obtains the book feature vector v L , the formula is: e l+1 =C(v l ,e l )[e],v l+1 =C(v l ,e l )[v], where v l represents the book vector of the lth layer, e l Represents the entity vector of the lth layer, [e] and [v] are used to distinguish the entity vector output and the book vector output in the cross compression subunit; The book knowledge graph embedding subunit uses an L-layer multi-layer perceptron to extract the relationship feature vector r between entities L , input the entity vector into the cross compression subunit of the L layer to obtain the entity feature vector: h L =E v~s(h) [C L (v,h)[e]], where s(h) represents the set of books v associated with book entity h; the relationship feature vector and the entity feature vector are concatenated and sent to the K-layer multi-layer perceptron to obtain the predicted entity feature vector t ^ , use the normalized inner product function to calculate the score of the triplet of the predicted entity: score(h,r,t)=f KG (t,t ^ );where f KG represents the normalized inner product function, t represents the tail entity of the triple, and t ^ Represents the predicted entity feature vector; The book recommendation subunit uses a multi-layer perceptron with L layers to extract the user feature vector u L , the extracted user feature vector and book feature vector are input into the neural collaborative filtering network, and the user features and book features are interacted through generalized matrix decomposition and multi-layer perceptron to output recommended books; Specifically, the element product of the user feature vector and the book feature vector is calculated, and generalized matrix decomposition is performed: , where the two elements represent the corresponding eigenvectors of the generalized matrix decomposition; Input the user feature vector and the book feature vector into the multi-layer perceptron to obtain the output y of the Lth layer MLP ; The final output of the computational model is: , where ρ represents the weight vector.

[0025] The loss function of the individual book recommendation model consists of three parts: the book recommendation module loss function, the book knowledge graph embedding loss function and the regularization term. The book recommendation module loss function is the loss value for calculating the actual click-through rate and predicted click-through rate of the model, and the regularization term is used to prevent overfitting.

[0026] The group recommendation unit uses the missing clustering algorithm to perform group analysis on the user's personalized model to obtain a group book recommendation list; The group recommendation unit of the book recommendation module includes: Randomly obtain several personalized models at the same cognitive development stage as the user in the personalized model database as group analysis samples, perform attribute composition analysis on the user's personalized model samples, and obtain the composition degree of each attribute of the user's personalized model samples. The composition degree is obtained by the proportion of the value of the corresponding attribute in the group analysis sample. After obtaining the composition degree of all attributes, normalize them so that the sum of the composition degrees is 1; The attributes are grouped according to the set number of groups so that the sum of the composition degrees of each group of attributes is close. The attributes of different groups are subjected to missing cluster analysis. The missing cluster analysis removes the attributes of this group and then clusters them to obtain the clustering results of the group analysis samples under multiple groups of missing attributes. The grouping method is to determine the number of set groups. The number of set groups has multiple values, including (3, 4, 5). If the number of set groups is 5, the sum of the composition degree of each group of attributes is close to 20%. The attributes are sorted according to the composition degree and the greedy algorithm is used to group the attributes.

[0027] Obtain reading statistics of all personalized models of the same type as the user's personalized model sample in each group of clustering results, and calculate the recommendation value of the enlightenment book corresponding to the reading record based on the credibility of the cluster analysis and the composition of the group of attributes. The credibility is obtained through the intra-cluster similarity of the user's personalized model sample in the cluster category, expressed in percentage form; The formula for calculating the recommendation value of the mth enlightenment book is: ; where x represents the xth group of attributes, θ x represents the composition degree of the x-th group of attributes, γ x represents the credibility of the x-th group of cluster analysis results, i represents the i-th reading statistics of the m-th enlightenment book in a group of cluster analysis results, and O represents the time of reading statistics; Repeat the above process three times to avoid falling into the local optimal solution, calculate the average recommendation value of the results of the three group analyses, sort the average recommendation values ​​of the enlightenment books, and take the top-ranked books as the group book recommendation list.

[0028] Module S5: Feedback optimization module, used to dynamically optimize book recommendations through user interaction information, including feedback mining unit and recommendation optimization unit; Feedback mining unit, used to perform data mining on user interaction information, update the personalized model and enlightenment book knowledge graph, and update the relationship weights of associated recommendations to optimize recommendation results; A recommendation optimization unit, used to weight the initial book recommendation list by interest according to the interest interaction information, so as to obtain the final recommended books; The recommendation optimization unit comprises: Integrate the individual book recommendation list and the group book recommendation list as a book recommendation list, display the book recommendation list to the target user children on a visual interface, and obtain the children's interest interaction information on different enlightenment books in the book recommendation list. The interest interaction information is obtained through image analysis algorithms and / or user parent feedback information; The interest weight is obtained through interest interaction information, and the enlightenment books with interest weight are weighted to obtain the final recommended books.

[0029] Module S6: Visual interaction module, used to realize the interactive interface of the system, including GUI interface, demand acquisition unit and usage recording unit. It obtains user needs, adjusts the personalized model, records user operations, and continuously optimizes user basic information.

[0030] The GUI interface is responsible for displaying the recommended list of enlightenment books, detailed information of recommended books, and user feedback options, and provides interactive methods such as clicking and sliding, so that users can browse and select recommended books more intuitively. The demand acquisition unit records the user's demand input through the GUI interface, such as reading interests and preferences, while the usage recording unit records the user's operation behavior on the GUI interface, including book selection, reading feedback, etc. The introduction of the GUI interface improves the user experience of the system and makes the interaction more intuitive and friendly.

[0031] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An enlightenment book recommendation system based on knowledge graph, It is characterized in that Includes the following modules: A user information acquisition module, used to acquire, collect and process user information, including an initial information acquisition unit and a supplementary information acquisition unit; The initial information acquisition unit is used to acquire initial user information through the user's registration information when the user uses the device for the first time; The supplementary information acquisition unit updates the user information based on the interactive information generated by the user; The knowledge graph construction module is used to construct the enlightenment book knowledge graph, obtain enlightenment book information through the enlightenment book database, and generate the enlightenment book knowledge graph, including the book database construction unit, the knowledge graph generation unit and the graph update unit; the knowledge graph consists of a triple (h, r, t), where h represents the head entity, r represents the relationship, and t represents the tail entity; The personalized model building module is used to build a personalized model for users. It generates the user's feature information through the user's basic information. The feature information is stored in a personalized knowledge graph. The user's feature information includes the cognitive development stage. The module includes a feature information unit, a personalized graph generation unit, and a graph update unit. The book recommendation module is used to generate a book recommendation list, including an individual recommendation unit and a group recommendation unit. The individual recommendation unit includes an associated recommendation subunit and a collaborative filtering subunit. The association recommendation subunit generates a candidate book recommendation list through the fuzzy association between the enlightenment book knowledge graph and the personalized graph; The collaborative filtering subunit screens the candidate book recommendation list through the collaborative filtering algorithm to obtain an individual book recommendation list; The group recommendation unit uses the missing clustering algorithm to perform group analysis on the user's personalized model to obtain a group book recommendation list; Feedback optimization module, used to dynamically optimize book recommendations through user interaction information, including feedback mining unit and recommendation optimization unit; Feedback mining unit, used to perform data mining on user interaction information, update the personalized model and enlightenment book knowledge graph, and update the relationship weights of associated recommendations to optimize recommendation results; A recommendation optimization unit, used to weight the initial book recommendation list by interest according to the interest interaction information, so as to obtain the final recommended books; The visual interactive module is used to realize the interactive interface of the system, including the GUI interface, the demand acquisition unit and the usage recording unit. The GUI interface is used to present recommended book information, user operation options and interactive feedback. It obtains user needs, adjusts the personalized model, and records user operations to continuously optimize the user's basic information.

2. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The group recommendation unit of the book recommendation module includes: Randomly obtain several personalized models at the same cognitive development stage as the user in the personalized model database as group analysis samples, perform attribute composition analysis on the user's personalized model samples, and obtain the composition degree of each attribute of the user's personalized model samples. The composition degree is obtained by the proportion of the value of the corresponding attribute in the group analysis sample. After obtaining the composition degree of all attributes, normalize them so that the sum of the composition degrees is 1; The attributes are grouped according to the set number of groups so that the sum of the composition degrees of each group of attributes is close. The attributes of different groups are subjected to missing cluster analysis. The missing cluster analysis removes the attributes of this group and then clusters them to obtain the clustering results of the group analysis samples under multiple groups of missing attributes. Obtain reading statistics of all personalized models of the same type as the user's personalized model sample in each group of clustering results, and calculate the recommendation value of the enlightenment book corresponding to the reading record based on the credibility of the cluster analysis and the composition of the group of attributes. The credibility is obtained through the intra-cluster similarity of the user's personalized model sample in the cluster category, expressed in percentage form; The formula for calculating the recommendation value of the mth enlightenment book is: ; where x represents the xth group of attributes, θ x represents the composition degree of the x-th group of attributes, γ x represents the credibility of the x-th group of cluster analysis results, i represents the i-th reading statistics of the m-th enlightenment book in a group of cluster analysis results, and O represents the time of reading statistics; Repeat the above process three times to avoid falling into the local optimal solution, calculate the average recommendation value of the results of the three group analyses, sort the average recommendation values ​​of the enlightenment books, and take the top-ranked books as the group book recommendation list.

3. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The associated recommendation subunit in the individual recommendation unit generates a candidate book recommendation list by calculating the association between the personalized model in the personalized knowledge graph and each enlightenment book in the enlightenment book knowledge graph; Specifically, the enlightenment books are stage-limited according to the user's cognitive development stage, and the fuzzy correlation is calculated for the enlightenment books in the stage-limited stage; the semantic similarity α between the current user's personalized model and the personalized model of the same cognitive development stage is obtained. The semantic similarity is only obtained through the user's personal feature information, and the value range is (0,1). In all personalized models of the same cognitive development stage, reading restrictions are made based on whether the book for which the fuzzy correlation is currently calculated has been read. A number of personalized model samples j are randomly extracted from the personalized models with reading restrictions, and the sum of the semantic similarities of all extracted personalized model samples is M Q Satisfy the following restrictions: 1 j<M Q ≤β 2 j, where β 1 and β 2 The value range of is (0,1), and the two are fuzzy adjustment parameters. The difference between the two determines the randomness of the sample. The larger the difference, the higher the randomness of the sample. For the result, it can avoid the local optimal solution. 1 and β 2 The larger the absolute value of is, the stronger the specific pairing ability of the sample is; The formula for calculating the fuzzy correlation of the nth enlightenment book is: ； Where T a , T b and T c They represent the proportion of types, keywords and themes of enlightenment books in the stage limit, respectively. j represents the semantic similarity of the jth user, R j represents the reading value of the jth user, obtained by reading time and understanding level, I j Indicates the proportion of types in the reading records of the jth user; Sort the relevance of enlightenment books and obtain the top-ranked books as a candidate book recommendation list.

4. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The collaborative filtering subunit in the book recommendation module performs neural collaborative filtering on the enlightenment books in the candidate book recommendation list through the individual book recommendation model to obtain a book recommendation list; The individual book recommendation model includes a cross compression sub-unit, a book knowledge graph embedding sub-unit, and a unit book recommendation sub-unit; The cross compression subunit is used to connect the book recommendation subunit and the book knowledge graph embedding subunit, and realize the feature interaction between the entity vector in the knowledge graph embedding subunit and the book vector in the book recommendation subunit by constructing a cross matrix; The book knowledge graph embedding subunit uses a multi-layer perceptron to transform entities and relationships in the enlightenment book knowledge graph into low-dimensional dense entity vectors; The book recommendation subunit is used to use a multi-layer perceptron to personalize the user feature vector in the graph; the user feature vector and the book feature vector are input into the neural collaborative filtering network, and finally the probability of the user interacting with the candidate books is predicted to obtain an individual book recommendation list.

5. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The recommendation optimization unit comprises: Integrate the individual book recommendation list and the group book recommendation list as a book recommendation list, display the book recommendation list to the target user children on a visual interface, and obtain the children's interest interaction information on different enlightenment books in the book recommendation list. The interest interaction information is obtained through image analysis algorithms and / or user parent feedback information; The interest weight is obtained through interest interaction information, and the enlightenment books with interest weight are weighted to obtain the final recommended books.

6. Based on the knowledge graph-based enlightenment book recommendation system described in claim 4, It is characterized in that Construction of said subunit; The cross compression subunit performs explicit interaction operations on the book vector and entity vector in the d-dimensional feature space to obtain a d*d dimensional cross feature matrix, maps the matrix back to the d-dimensional feature space, outputs the book vector and entity vector of the next layer, completes the compression operation, and obtains the book feature vector v L , the formula is: e l+1 =C(v l ,e l )[e],v l+1 =C(v l ,e l )[v], where v l represents the book vector of the lth layer, e l Represents the entity vector of the lth layer, [e] and [v] are used to distinguish the entity vector output and the book vector output in the cross compression subunit; The book knowledge graph embedding subunit uses an L-layer multi-layer perceptron to extract the relationship feature vector r between entities L , input the entity vector into the cross compression subunit of the L layer to obtain the entity feature vector: h L =E v~s(h) [C L (v,h)[e]], where s(h) represents the set of books v associated with book entity h; the relationship feature vector and the entity feature vector are concatenated and sent to the K-layer multi-layer perceptron to obtain the predicted entity feature vector t ^ , use the normalized inner product function to calculate the score of the triplet of the predicted entity: score(h,r,t)=f KG (t,t ^ );where f KG represents the normalized inner product function, t represents the tail entity of the triple, and t ^ Represents the predicted entity feature vector; The book recommendation subunit uses a multi-layer perceptron with L layers to extract the user feature vector u L , the extracted user feature vector and book feature vector are input into the neural collaborative filtering network, and the user features and book features are interacted through generalized matrix decomposition and multi-layer perceptron to output recommended books; Specifically, the element product of the user feature vector and the book feature vector is calculated, and generalized matrix decomposition is performed: , where the two elements represent the corresponding eigenvectors of the generalized matrix decomposition; Input the user feature vector and the book feature vector into the multi-layer perceptron to obtain the output y of the Lth layer MLP ; The final output of the computational model is: , where ρ represents the weight vector.

7. Based on the knowledge graph-based enlightenment book recommendation system described in claim 4, It is characterized in that The loss function of the individual book recommendation model consists of three parts: the book recommendation module loss function, the book knowledge graph embedding loss function and the regularization term. The book recommendation module loss function is the loss value for calculating the actual click-through rate and predicted click-through rate of the model, and the regularization term is used to prevent overfitting.

8. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The construction of the knowledge graph includes a model layer and a data layer; The model layer determines the entity concepts of the domain, the classes corresponding to the entity concepts, the attributes and attribute value ranges of the entities, and the attribute description features by analyzing the relevant knowledge of the domain. Establish a local library in the corresponding field to build a knowledge graph; The data layer obtains data through the database, cleans the data, obtains the nodes and relationships of the data, and uses the graph database to store the knowledge graph.

9. Based on the knowledge graph-based enlightenment book recommendation system described in claim 1, It is characterized in that The personalized model is constructed by the user's characteristic information, which is obtained by processing the user information. The personalized model includes the user's personal characteristic information, reading records and statistics, book search and browsing records and statistics; the user's personal characteristic information includes the user's cognitive development stage and a brief family situation. The personalized model is used to form a personalized knowledge graph. The user's cognitive development stage is obtained by inputting or observing children's behavior, language ability, interests and hobbies.

Images