Book recommendation system and method based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a book recommendation system and method based on artificial intelligence, and the system comprises a relation confidence calculation module which obtains the relation confidence of node embedding and edges; the target translation screening module is used for screening to obtain a target translation set; the cross-layer shortest path calculation module is used for calculating and obtaining the cost of the cross-layer shortest borrowed path; the cross-layer cost calculation module is used for calculating to obtain candidate cross-layer cost; the recommendation list generation module is used for generating a recommendation list; and the online updating module is used for updating the learnable parameters. According to the method, a knowledge graph multi-level association model is constructed, so that a user work layer demand is accurately mapped to a version and copy layer; meanwhile, real business constraints such as a borrowing state, authorization limitation and geographical allocation cost are fused, and through cross-layer path cost quantification and reachability gradient calculation, it is ensured that a recommendation result stably points to a target translation expected by a user and can be borrowed at present.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence-based book recommendation system and method. Background Technology

[0002] With the advancement of digital library construction and the diversification of collection resources, users' demand for accurate book recommendations is becoming increasingly urgent. Currently, book cataloging generally adopts a three-tiered hierarchical management system of works, editions, and copies. The same work often has multiple languages, multiple translators, and multiple revisions. User needs are mostly expressed at the work level, but collection management authorization and lending execution all occur at the edition and copy level. Electronic resources are limited by region and concurrent licenses, while paper resources are constrained by the distribution of branch libraries and their shelf status. This inherent misalignment between cataloging hierarchy and business processes requires the recommendation system to accurately map needs to available copies; otherwise, recommendation bias is likely to occur.

[0003] Data quality defects and heterogeneity of multi-source systems further exacerbate the problems; cataloging suffers from inconsistencies in scope, incomplete coverage of unified titles, misuse and reuse of identifiers, and difficulties in disambiguating author names. The collection data also needs to be overlaid with third-party metadata and multiple sets of thesaurus; at the same time, related systems such as academic schedules and inventory systems belong to different data sources, with frequent status changes and synchronization delays, resulting in uncertainty in the mapping path from work to version and from version to copy. The cost of obtaining target translations is increased, while the cost of obtaining non-target translations is reduced, and the mapping stability is greatly reduced.

[0004] The cumulative bias of traditional recommendation algorithms has perpetuated the problem. Collaborative filtering and content matching rely excessively on historical behavior and text features, naturally favoring popular versions and essentially choosing the lowest-cost path. However, authorization restrictions, changes in borrowable status, and geographical allocation dynamically alter the real-world costs. When there are no explicit constraints during the ranking stage, the algorithm will continuously tilt towards non-target or unavailable versions, resulting in cumulative bias. Existing systems have not effectively integrated business constraints and lack targeted optimization mechanisms, ultimately leading to recommendation results that often point to non-target or unavailable copies, making it difficult for users to quickly find borrowable books that meet their needs. Summary of the Invention

[0005] This invention provides an artificial intelligence-based book recommendation system and method, which solves the technical problems mentioned in the background.

[0006] This invention provides an artificial intelligence-based book recommendation system, comprising: The relation confidence calculation module is used to obtain the relation confidence of nodes and edges by performing message passing on the knowledge graph through a graph neural network; Node types include: work nodes, version nodes, and copy nodes; The target translation filtering module is used to generate user intent embeddings based on user search content, calculate the probability of works, language priors, and translator priors based on user intent embeddings and node embeddings, and combine knowledge graphs to filter and obtain a set of target translations. The cross-layer shortest path calculation module is used to obtain the edge cost by weighted summation based on edge relation confidence, authorization penalty, inversion of available indication, geographical allocation cost, and time period congestion. Based on this, the cost of the cross-layer shortest available path is calculated, and the minimum value is taken within the target translation set to obtain the target cost. The target costs of different works are then weighted and summarized according to the work probability. The cross-layer cost calculation module is used to calculate the shortest available path cost across layers for each copy in the candidate copy set based on the probability of the work, and obtain the candidate cross-layer cost by subtracting the candidate cross-layer cost from the weighted target cost. The recommendation list generation module is used to generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradients, and generate a recommendation list from the candidate replica set based on the joint score. The online update module is used to update the learnable parameters based on the user's selection of the recommended list, combined with joint scores and cross-level accessibility gradients.

[0007] Furthermore, the knowledge graph is denoted as ,in Represents a set of nodes. Represents the set of directed edges. This represents the set of relation types in a knowledge graph, where each node possesses the initial node features. The total number of layers in the graph neural network is L; for the Lth layer of the graph neural network... Layer-by-layer calculation to obtain the first... Layer node features The calculation formula is as follows: ,in This represents the set of neighboring nodes connected to node v under relation type r. express The normalization constant, Indicates the first The learnable weight matrix corresponding to relation type r in the layer. Indicates the first The learnable residual weight matrix of the layer, Indicates the first The feature vector of the neighbor node u of the layer, Indicates the first The feature vector of node v in layer , Represents a non-linear activation function, and the node features of the Lth layer. As a node embedding of node v.

[0008] Furthermore, each directed edge , represented as Where u represents the starting node of the directed edge, v represents the ending node of the directed edge, and r represents the relation type corresponding to the directed edge. The relation confidence of the directed edge is calculated based on the node embeddings of the starting and ending nodes. The calculation formula is as follows: ,in This represents the learnable bilinear parameter matrix of the correspondence type r. This represents the transpose of the embedding vector of node u. This represents the embedding vector of node v. This represents a non-linear activation function.

[0009] Furthermore, the user's search content is encoded by a neural network encoder to obtain the user intent embedding, the node embedding of the node type being the work node is obtained, and the work probability, language prior and translator prior are calculated by combining the user intent embedding. Probability of works The calculation formula is as follows: ,in This represents the transpose of the user intent embedding. Node embedding represents the nodes of the work. This represents the learnable weight matrix corresponding to the matching of works, and softmax represents the softmax normalization function. Language a priori The calculation formula is as follows: ,in The learnable weight matrix corresponding to language matching; Translator's a priori knowledge The calculation formula is as follows: ,in This represents the learnable weight matrix corresponding to the translator matching; Obtain the language set corresponding to the largest language prior, obtain the translator set corresponding to the largest translator prior, and in the knowledge graph, for each work node, select the replica node that meets the following conditions: the work node corresponding to the replica node has a directed edge pointing to the version node, the language of the version node belongs to the language set, and the translator of the version node belongs to the translator set. Finally, the selected copy nodes will form the target translation set for the corresponding work node.

[0010] Furthermore, the authorization penalty, the inverted value of the current time replica borrowable indication, the geographical allocation cost, and the time period congestion degree associated with each directed edge in the knowledge graph are obtained. The edge cost of each directed edge is obtained by weighted summing the negative logarithm of the relation confidence of the edge by the learnable weight vector, the authorization penalty, the inverted value of the borrowable indication, the geographical allocation cost, and the time period congestion degree. For each work node and each copy node, determine all legal paths from the work node to the copy node with the end copy in a borrowable state, calculate the cumulative edge cost of all directed edges on each legal path, and select the minimum value among the cumulative values ​​as the cross-layer shortest borrowable path cost from the work node to the copy node. For each work node, the minimum value is selected from the cross-layer shortest available path cost corresponding to all copy nodes included in its target translation set, and this value is taken as the target cost of that work node. Multiply the target cost of each work node by the corresponding work probability, and sum all the multiplication results to obtain the weighted target cost.

[0011] Furthermore, the replicas currently in a borrowable state are identified as the candidate replica set; For each replica in the candidate replica set, the shortest available path cost across all work nodes corresponding to that replica is multiplied by the work probability of the corresponding work node, and all multiplication results are summed to obtain the candidate cross-layer cost of that replica. For each replica's candidate cross-layer cost, the weighted target cost is subtracted from the candidate cross-layer cost to obtain the cross-layer accessibility gradient of that replica.

[0012] Furthermore, the personalized preference probability is calculated for each replica node, and the personalized preference probability of replica node i is... The calculation formula is as follows: ,in This represents the transpose of the user intent embedding. This represents the learnable weight matrix corresponding to personalized preference matching. This indicates the node embedding of replica node i. This represents a non-linear activation function.

[0013] Furthermore, a joint score is calculated for each replica node, and the joint score of replica node i is... The calculation formula is as follows: ,in and Let represent the first learnable positive scalar and the second learnable positive scalar, respectively. This represents the cross-level borrowable reachability gradient corresponding to replica node i; The joint scores of all replicas in the candidate replica set are sorted in descending order, and the top N replicas after sorting are selected to obtain the recommended list.

[0014] Furthermore, the user's selection results in the recommended list are obtained, and the copy selected by the user is marked as the target copy; The objective function is constructed based on the joint score and the cross-level accessibility gradient. The calculation formula is as follows: ,in Represents the set of candidate replicas. Indicates the replica node selected by the user. The combined score This represents the joint score of any replica node j in the candidate replica set. Indicates the replica node selected by the user. Cross-level accessibility gradient, represents a nonnegative scalar, and exp represents an exponential function with the natural constant as the base; Set a positive scalar as the learning rate, calculate the gradient of the objective function with respect to the learnable parameters, and subtract the product of the learning rate and the corresponding gradient from each parameter to obtain the updated learnable parameters.

[0015] This invention provides an artificial intelligence-based book recommendation method, comprising the following steps: Step S201: Message passing is performed on the knowledge graph using a graph neural network to obtain node embeddings and edge relationship confidence. Step S202: Generate user intent embedding based on user search content; calculate work probability, language prior and translator prior based on user intent embedding and node embedding; and obtain target translation set by combining knowledge graph filtering. Step S203: Obtain the edge cost by weighted summation based on edge relation confidence, authorization penalty, inversion of borrowable indication, geographical allocation cost, and time period congestion. Calculate the shortest borrowable path cost across layers based on this cost, take the minimum value within the target translation set to obtain the target cost, and weight and summarize the target costs of different works based on the work probability. Step S204: Based on the work probability, the shortest available path cost across layers for each copy in the candidate copy set is weighted and summarized to obtain the candidate cross-layer cost. The candidate cross-layer cost is subtracted from the weighted target cost to obtain the cross-layer available accessibility gradient. Step S205: Generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradient, and generate a recommendation list in the candidate replica set based on the joint score; Step S206: Based on the user's selection of the recommended list, update the learnable parameters by combining the joint score and cross-level accessibility gradient.

[0016] The beneficial effects of this invention are as follows: By constructing a multi-level association model of a knowledge graph, this invention accurately maps user work layer requirements to version and copy layers, effectively solving the recommendation bias problem caused by cataloging layering and business process misalignment; at the same time, it integrates real-world business constraints such as borrowability status, authorization restrictions, and geographical allocation costs, and ensures that the recommendation results stably point to the user's desired target translation and are currently available for borrowing through cross-layer path cost quantification and accessibility gradient calculation; relying on the closed-loop optimization mechanism of user selection feedback, it dynamically updates learnable parameters, continuously corrects the cumulative effect of algorithm bias, and continuously improves the accuracy and adaptability of recommendations; ultimately, it allows users to quickly obtain book resources that match their personalized preferences and are actually available for borrowing, greatly improving the practicality and user satisfaction of the recommendation system, and adapting to the complex collection environment where multiple translations and versions coexist. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the book recommendation system based on artificial intelligence according to the present invention; Figure 2 This is a flowchart of the book recommendation method based on artificial intelligence of the present invention.

[0018] In the diagram: Relationship confidence calculation module 101, target translation screening module 102, cross-layer shortest path calculation module 103, cross-layer cost calculation module 104, recommendation list generation module 105, and online update module 106. Detailed Implementation

[0019] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0020] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] like Figures 1-2 As shown, an artificial intelligence-based book recommendation system includes: The relation confidence calculation module 101 is used to obtain the relation confidence of node embeddings and edges by performing message passing on the knowledge graph through a graph neural network. Node types include: work nodes, version nodes, and copy nodes; The target translation filtering module 102 is used to generate user intent embedding based on user search content, calculate the work probability, language prior and translator prior based on user intent embedding and node embedding, and combine knowledge graph to filter and obtain the target translation set. The cross-layer shortest path calculation module 103 is used to obtain the edge cost by weighted summation based on the edge relation confidence, authorization penalty, inversion of the borrowable indication, geographical allocation cost, and time period congestion. Based on this, the cross-layer shortest borrowable path cost is calculated, and the minimum value is taken in the target translation set to obtain the target cost. The target costs of different works are weighted and summarized according to the work probability. The cross-layer cost calculation module 104 is used to perform a weighted summation of the shortest available cross-layer path cost for each copy in the candidate copy set based on the work probability to obtain the candidate cross-layer cost, and to subtract the candidate cross-layer cost from the weighted target cost to obtain the cross-layer available accessibility gradient. The recommendation list generation module 105 is used to generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradients, and generate a recommendation list in the candidate replica set based on the joint score. The online update module 106 is used to update the learnable parameters based on the user's selection of the recommended list, combined with joint scores and cross-level accessibility gradients.

[0022] In one embodiment of the present invention, the knowledge graph is denoted as... ,in Represents a set of nodes. Represents the set of directed edges. This represents the set of relation types in a knowledge graph, where each node possesses the initial node features. The total number of layers in the graph neural network is L; for the Lth layer of the graph neural network... Layer-by-layer calculation to obtain the first... Layer node features The calculation formula is as follows: ,in This represents the set of neighboring nodes connected to node v under relation type r. express The normalization constant, Indicates the first The learnable weight matrix corresponding to relation type r in the layer. Indicates the first The learnable residual weight matrix of the layer, Indicates the first The feature vector of the neighbor node u of the layer, Indicates the first The feature vector of node v in layer , Represents a non-linear activation function, and the node features of the Lth layer. As a node embedding of node v.

[0023] It should be noted that a knowledge graph is a structured data carrier composed of nodes and directed edges, used to organize book-related entities and the relationships between them. Node types include work nodes, version nodes, and copy nodes. Work nodes represent the content ontology of a book, without distinguishing between specific publication formats and media. Version nodes represent different publication formats of the same work, including differentiated attributes such as language, translator, and publication information. Copy nodes represent the specific library entity corresponding to the version node, including actual borrowing-related attributes such as location and availability. Directed edges are used to connect different types of nodes, with each edge corresponding to a specific relationship type, reflecting the relationship logic between nodes. This invention uses a graph neural network to fuse the initial features of nodes with the structural and semantic relationship information in the knowledge graph, generating a more comprehensive node representation. The initial node features only contain the node's own attribute information and lack information about its relationships with other nodes. Through the message passing mechanism of the graph neural network, the features of neighboring nodes under different relationships of the target node can be aggregated, allowing the node to embed itself while simultaneously carrying its own attributes and graph structure and semantic relationship information. This makes subsequent recommendation-related calculations more aligned with the actual scenario of book recommendation, improving the rationality and reliability of the recommendation system and helping users quickly obtain book resources that meet their needs and are available for borrowing.

[0024] It should be noted that the relation type set represents the set of all possible relation types in the knowledge graph, used to distinguish different edge semantics and ensure that messages are aggregated within the corresponding semantic channels. For example, a work has a certain translation, and a certain version has a copy available for borrowing in a certain branch library. The neighbor set represents the set of all nodes directly connected to the target node (the node whose embedding is currently being calculated) under a certain relation. The initial node features represent a unified dimensional representation obtained by processing fields such as title, keywords, author, language, translator, publication time, library holdings and borrowing status. For example, the text field is first segmented and stop word removed and converted into fixed-length numerical values. Then, discrete fields such as author, language, and translator are converted into trainable fixed-length numerical values. Then, the publication time, library holdings and borrowing status are quantified and concatenated with the aforementioned vector. Finally, a vector of a unified length (e.g., 512 dimensions) is obtained through linear transformation. The initial node features can be obtained through Sentence-BERT (a semantic representation model based on BERT), which will not be elaborated here.

[0025] It should be noted that the total number of layers L in the graph neural network is a preset parameter, usually set to 2 to 4 layers, preferably 3 layers; the normalization constant is used to adjust the scale when aggregating features of neighboring nodes, avoiding fluctuations in feature values ​​due to differences in the number of neighbors, and the normalization constant is set to the number of neighboring nodes; both the learnable weight matrix and the learnable residual weight matrix can be initialized using Xavier, that is, initial values ​​are randomly selected from a specific uniform distribution according to the dimensions of the input and output features, ensuring that the variance of the input and output of each layer is consistent; the non-linear activation function is set to the ReLU activation function; the node embedding represents the final feature representation of the node after aggregation calculation by the graph neural network, containing the structural and semantic information of the knowledge graph, and the node embedding is represented by a vector of a custom dimension, preferably 256. Therefore, according to the above, the learnable weight matrix and the learnable residual weight matrix of the first layer are both designed to be 256×512, and the learnable weight matrix and the learnable residual weight matrix of the second and third layers are both designed to be 256×256.

[0026] In one embodiment of the present invention, each directed edge , represented as Where u represents the starting node of the directed edge, v represents the ending node of the directed edge, and r represents the relation type corresponding to the directed edge. The relation confidence of the directed edge is calculated based on the node embeddings of the starting and ending nodes. The calculation formula is as follows: ,in This represents the learnable bilinear parameter matrix of the correspondence type r. This represents the transpose of the embedding vector of node u. This represents the embedding vector of node v. This represents a non-linear activation function.

[0027] It should be noted that relation confidence is used to quantify the structural and semantic credibility of directed edges. Its value ranges from 0 to 1. The closer the value is to 1, the more reasonable the structure and the better the semantic fit of the relation represented by the edge. Otherwise, it indicates that the structural rationality and semantic fit of the edge are lower. The non-linear activation function is set to the Sigmoid activation function. The node embedding is 256×1, so the transpose vector is 1×256. The learnable bilinear parameter matrix is ​​set to 256×256. The result of multiplying the three is a 1×1 relation confidence (scalar). Similarly, the learnable bilinear parameter matrix can be initialized by Xavier. Calculating the relation confidence of directed edges allows the cost of subsequent edges to carry both structural and semantic information, avoiding recommendation bias caused by ambiguity in relation confidence, thereby improving the accuracy of subsequent cross-layer path calculation.

[0028] In one embodiment of the present invention, the user's search content is encoded by a neural network encoder to obtain the user intent embedding, the node embedding of the node type being the work node is obtained, and the work probability, language prior and translator prior are calculated by combining the user intent embedding. Probability of works The calculation formula is as follows: ,in This represents the transpose of the user intent embedding. Node embedding represents the nodes of the work. This represents the learnable weight matrix corresponding to the matching of works, and softmax represents the softmax normalization function. Language a priori The calculation formula is as follows: ,in The learnable weight matrix corresponding to language matching; Translator's a priori knowledge The calculation formula is as follows: ,in This represents the learnable weight matrix corresponding to the translator matching; Obtain the language set corresponding to the largest language prior, obtain the translator set corresponding to the largest translator prior, and in the knowledge graph, for each work node, select the replica node that meets the following conditions: the work node corresponding to the replica node has a directed edge pointing to the version node, the language of the version node belongs to the language set, and the translator of the version node belongs to the translator set. Finally, the selected copy nodes will form the target translation set for the corresponding work node.

[0029] It should be noted that the neural network encoder is used to convert user-retrieved content into a vector representation, which can also be achieved through Sentence-BERT conversion; the user intent embedding is represented by a vector with a custom dimension, preferably set to 256; the work probability represents the probability distribution of the fit between each work node and the user intent; the language prior represents the probability distribution of the user's preference for different languages; the translator prior represents the probability distribution of the user's preference for different translators; the learnable weight matrix corresponding to work matching is set to 256×256; the learnable weight matrix corresponding to language matching is set to 1×256; the learnable weight matrix corresponding to translator matching is set to 1×256; and the language set corresponding to the maximum language prior represents the language with the highest probability in the language prior. The set includes one or more languages; the translator set corresponding to the maximum translator prior is the set of one or more translators with the highest probability in the translator prior; the target translation set represents the set of copy nodes that match the user's preferred language, translator, and have a work-version association; the language set and translator set are selected by prioritizing the language and translator with the highest probability. If multiple languages ​​(or translators) have the same highest prior probability, all languages ​​(or translators) with the same probability are included in the corresponding set to ensure that the user's core preferences are not missed. For example, if the language prior probability of Chinese and English is 0.4 and they are tied for first place, and the probabilities of other languages ​​are lower than this value, then the language set corresponding to the maximum language prior contains two elements: Chinese and English; if only Chinese has the highest probability, then the set contains only one element: Chinese.

[0030] It should be noted that by accurately capturing user search intent through neural network encoder, and combining the work probability generated by work node embedding, language prior and translator prior, the core user preferences are quantified; and based on these preferences, the target translation set is selected to ensure that the copies in the set not only match the user intent, but also meet the language and translator requirements, providing a precise range for subsequent cross-layer path calculation and cost weighting.

[0031] In one embodiment of the present invention, the authorization penalty, the inverted value of the current time replica borrowable indication, the geographical allocation cost, and the time period congestion degree associated with each directed edge in the knowledge graph are obtained. The edge cost of each directed edge is obtained by weighted summing the negative logarithm of the relation confidence of the edge by the learnable weight vector, the authorization penalty, the inverted value of the borrowable indication, the geographical allocation cost, and the time period congestion degree. For each work node and each copy node, determine all legal paths from the work node to the copy node with the end copy in a borrowable state, calculate the cumulative edge cost of all directed edges on each legal path, and select the minimum value among the cumulative values ​​as the cross-layer shortest borrowable path cost from the work node to the copy node. For each work node, the minimum value is selected from the cross-layer shortest available path cost corresponding to all copy nodes included in its target translation set, and this value is taken as the target cost of that work node. Multiply the target cost of each work node by the corresponding work probability, and sum all the multiplication results to obtain the weighted target cost.

[0032] It should be noted that the learnable weight vector represents a trainable parameter vector used to balance the importance of each component of the edge cost, with a dimension equal to the number of cost components of the edge cost, i.e., a five-dimensional vector; the authorization penalty represents a quantified value of the authorization restriction on the edge-related resource, with a larger value for stricter restrictions, and its value is equal to the quantified value corresponding to the authorization restriction level. The authorization restriction levels are divided into no restriction, light restriction, medium restriction, and heavy restriction, with corresponding quantified values ​​of 0, 0.3, 0.7, and 1, respectively; the borrowable indicator represents a binary identifier indicating whether the edge-related copy is currently borrowable, with a value of 1 indicating borrowable and a value of 0 indicating not borrowable; the judgment timing is at the current moment of edge cost calculation, and the status of the copy's collection system is queried in real time to ensure the accuracy of the borrowable status. The geographical allocation cost represents a quantified value of the distance and time cost of obtaining copies across branches or locations. This value is equal to the straight-line distance between the user's location and the copy's location divided by 10 kilometers, rounded to one decimal place. If it's within the same branch, the value is 0. The time-slot congestion represents a quantified value of the density of borrowing demand during the current time slot. Higher congestion results in a larger value, equal to the current number of borrowing requests divided by the historical maximum number of borrowing requests during that time slot, rounded to two decimal places, and ranging from 0 to 1. The edge cost represents the combined edge access cost, integrating structural costs and real-world constraints. The negative logarithm of the edge relationship confidence is calculated using the natural logarithm. If the edge relationship confidence is 0, the negative logarithm is a fixed value of 5.

[0033] It should be noted that a valid path refers to a path from a work node to a copy node where the final copy is available for borrowing. This path must follow the hierarchical association rule of "work node → version node → copy node," and cannot directly connect across intermediate nodes. Furthermore, the relationship type of each edge in the path must conform to the preset semantics of the knowledge graph, such as the correspondence between work and version, or the inclusion relationship between version and copy. The shortest borrowable path cost across layers represents the minimum cumulative path cost among all valid paths from a work node to a copy node. The target cost represents the minimum shortest borrowable path cost across layers corresponding to each work node within its target translation set. Filtering valid paths and solving for the shortest borrowable path cost ensures that cost quantification closely matches the actual borrowing scenario. The target cost is selected based on the target translation set and weighted by work probability, allowing the global target cost to fully align with the uncertainty of user intent, thus improving the accessibility and rationality of the recommendation system.

[0034] In one embodiment of the present invention, a set of candidate replicas is determined from replicas that are currently available for borrowing. For each replica in the candidate replica set, the shortest available path cost across all work nodes corresponding to that replica is multiplied by the work probability of the corresponding work node, and all multiplication results are summed to obtain the candidate cross-layer cost of that replica. For each replica's candidate cross-layer cost, the weighted target cost is subtracted from the candidate cross-layer cost to obtain the cross-layer accessibility gradient of that replica.

[0035] It should be noted that a copy currently in a borrowable status refers to a physical copy of the collection that is not checked out, is undamaged, and meets the borrowing rules at the current time of cost calculation. The borrowability status is determined based on real-time data from the collection management system, including whether the copy has been checked out, whether it is damaged or reserved, and whether it is within the borrowing open hours. The timing of the determination is consistent with the time of candidate cross-level cost calculation to ensure the timeliness of the status. The candidate cross-level cost represents the expected cost obtained by weighting the path costs of all works corresponding to the candidate copy according to the probability of the work. The cross-level borrowability accessibility gradient represents the cost difference between the candidate copy and the target translation in terms of accessibility, and is the core indicator for measuring the suitability of the candidate copy.

[0036] It should be noted that by screening currently available copies to construct a candidate set, the recommended objects are ensured to be realistically feasible. The cross-layer cost of candidates is summarized by combining the probability of the works, which fully considers the uncertainty of user intentions and makes the cost quantification more in line with actual needs. In addition, the accessibility gradient is calculated by the difference with the target cost, which transforms the suitability of candidate copies into an intuitive and comparable indicator. This enables the recommendation system to screen copies that both match user intentions and have high accessibility.

[0037] In one embodiment of the present invention, a personalized preference probability is calculated for each replica node, wherein the personalized preference probability of replica node i is... The calculation formula is as follows: ,in This represents the transpose of the user intent embedding. This represents the learnable weight matrix corresponding to personalized preference matching. This indicates the node embedding of replica node i. This represents a non-linear activation function.

[0038] It should be noted that the learnable weight matrix corresponding to personalized preference matching is set to 256×256; the non-linear activation function is set to the Sigmoid activation function; the personalized preference probability is used to quantify the user's preference for candidate copies, and its value ranges from 0 to 1. The larger the value, the higher the preference. Similarly, the calculation of personalized preference probability fully explores the user's personalized needs, avoids the subjectivity and one-sidedness of preference judgment, and makes the preference quantification results semantically reasonable, thereby selecting book copies that are more in line with the user's personal needs.

[0039] In one embodiment of the present invention, a joint score is calculated for each replica node, wherein the joint score of replica node i is... The calculation formula is as follows: ,in and Let represent the first learnable positive scalar and the second learnable positive scalar, respectively. This represents the cross-level borrowable reachability gradient corresponding to replica node i; The joint scores of all replicas in the candidate replica set are sorted in descending order, and the top N replicas after sorting are selected to obtain the recommended list.

[0040] It should be noted that the first learnable positive scalar is used to adjust the weight of the contribution of personalized preference probability to the joint score; the second learnable positive scalar is used to adjust the weight of the contribution of cross-level accessibility gradient to the joint score; both are learnable parameters, and their initial values ​​are both set to 1, which are then adaptively adjusted based on the recommendation effect during model training; the joint score is used to comprehensively reflect the preference suitability and accessibility of candidate replicas, and the higher the score, the stronger the suitability; the output item number parameter N is a preset parameter, preferably set to 10; in addition, if the personalized preference probability is zero, the result of the natural logarithm is taken as a fixed value of -10 to avoid calculation errors and weaken the score proportion of that replica.

[0041] In one embodiment of the present invention, the user's selection result in the recommended list is obtained, and the copy selected by the user is marked as the target copy; The objective function is constructed based on the joint score and the cross-level accessibility gradient. The calculation formula is as follows: ,in Represents the set of candidate replicas. Indicates the replica node selected by the user. The combined score This represents the joint score of any replica node j in the candidate replica set. Indicates the replica node selected by the user. Cross-level accessibility gradient, represents a nonnegative scalar, and exp represents an exponential function with the natural constant as the base; Set a positive scalar as the learning rate, calculate the gradient of the objective function with respect to the learnable parameters, and subtract the product of the learning rate and the corresponding gradient from each parameter to obtain the updated learnable parameters.

[0042] It should be noted that the objective function integrates the ranking loss and the regularization term as an optimization metric to guide the direction of parameter updates; the learning rate is a positive scalar that controls the step size of parameter updates, affecting training stability and convergence speed. The initial value can be set to 0.01, employing a fixed learning rate strategy. If convergence is slow in the later stages of training, it can be adjusted to 0.001; the learnable parameters encompass all learnable parameters of this invention; the gradient of the objective function with respect to the learnable parameters represents the degree and direction of the influence of parameter changes on the objective function, calculated using the backpropagation algorithm; this invention constructs a supervision signal through user selection results, ensuring that parameter updates align with actual usage feedback. The objective function integrates the ranking loss and the regularization term, guaranteeing the accuracy of recommendation ranking while avoiding model overfitting. Gradient descent is used to update the learnable parameters in real time, achieving online adaptive optimization of the model. This allows the system to continuously learn changes in user preferences, thereby continuously improving the accuracy and adaptability of subsequent recommendations.

[0043] It should be noted that the ranking loss is achieved by taking the negative logarithm of the ratio of the joint rating index of the selected replica to the sum of the joint rating indices of all candidate replicas. Specifically, it amplifies the rating difference between the selected replica and other candidate replicas. The higher the rating of the selected replica and the greater the difference between it and other replicas, the smaller the loss value and the better the model. Conversely, the loss value increases, driving parameter updates to correct this bias. The regularization term is the square of the accessibility gradient of the selected replica. Its function is to limit the parameters from overfitting to highly accessible replicas and balance preference adaptation with accessibility. It not only fits the current user's choice but also meets the recommendation needs of different users and different scenarios, improving the long-term stability of the system.

[0044] In one embodiment of the present invention, the book recommendation method based on artificial intelligence includes the following steps: Step S201: Message passing is performed on the knowledge graph using a graph neural network to obtain node embeddings and edge relationship confidence. Step S202: Generate user intent embedding based on user search content; calculate work probability, language prior and translator prior based on user intent embedding and node embedding; and obtain target translation set by combining knowledge graph filtering. Step S203: Obtain the edge cost by weighted summation based on edge relation confidence, authorization penalty, inversion of borrowable indication, geographical allocation cost, and time period congestion. Calculate the shortest borrowable path cost across layers based on this cost, take the minimum value within the target translation set to obtain the target cost, and weight and summarize the target costs of different works based on the work probability. Step S204: Based on the work probability, the shortest available path cost across layers for each copy in the candidate copy set is weighted and summarized to obtain the candidate cross-layer cost. The candidate cross-layer cost is subtracted from the weighted target cost to obtain the cross-layer available accessibility gradient. Step S205: Generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradient, and generate a recommendation list in the candidate replica set based on the joint score; Step S206: Based on the user's selection of the recommended list, update the learnable parameters by combining the joint score and cross-level accessibility gradient.

[0045] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0046] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. An artificial intelligence-based book recommendation system, characterized in that, include: The relation confidence calculation module is used to obtain the relation confidence of nodes and edges by performing message passing on the knowledge graph through a graph neural network; Node types include: work nodes, version nodes, and copy nodes; The target translation filtering module is used to generate user intent embeddings based on user search content, calculate the probability of works, language priors, and translator priors based on user intent embeddings and node embeddings, and combine knowledge graphs to filter and obtain a set of target translations. The cross-layer shortest path calculation module is used to obtain the edge cost by weighted summation based on edge relation confidence, authorization penalty, inversion of available indication, geographical allocation cost, and time period congestion. Based on this, the cost of the cross-layer shortest available path is calculated, and the minimum value is taken within the target translation set to obtain the target cost. The target costs of different works are then weighted and summarized according to the work probability. The cross-layer cost calculation module is used to calculate the shortest available path cost across layers for each copy in the candidate copy set based on the probability of the work, and obtain the candidate cross-layer cost by subtracting the candidate cross-layer cost from the weighted target cost. The recommendation list generation module is used to generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradients, and generate a recommendation list from the candidate replica set based on the joint score. The online update module is used to update the learnable parameters based on the user's selection of the recommended list, combined with joint scores and cross-level accessibility gradients.

2. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Let the knowledge graph be denoted as ,in Represents a set of nodes. Represents a set of directed edges. This represents the set of relation types in a knowledge graph, where each node possesses the initial node features. The total number of layers in the graph neural network is L; for the Lth layer of the graph neural network... Layer-by-layer calculation to obtain the first... Layer node features The calculation formula is as follows: ,in This represents the set of neighboring nodes connected to node v under relation type r. express The normalization constant, Indicates the first The learnable weight matrix corresponding to relation type r in the layer. Indicates the first The learnable residual weight matrix of the layer, Indicates the first The feature vector of the neighbor node u of the layer, Indicates the first The feature vector of node v in layer , Represents a non-linear activation function, and the node features of the Lth layer. As a node embedding of node v.

3. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Each directed edge , represented as Where u represents the starting node of the directed edge, v represents the ending node of the directed edge, and r represents the relation type corresponding to the directed edge. The relation confidence of the directed edge is calculated based on the node embeddings of the starting and ending nodes. The calculation formula is as follows: ,in This represents the learnable bilinear parameter matrix of the correspondence type r. This represents the transpose of the embedding vector of node u. This represents the embedding vector of node v. This represents a non-linear activation function.

4. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, The user's search content is encoded by a neural network encoder to obtain the user intent embedding, the node embedding of the node type is obtained, and the work probability, language prior and translator prior are calculated by combining the user intent embedding. Probability of works The calculation formula is as follows: ,in The transpose vector representing the user intent embedding. This indicates the node embedding of the work's nodes. This represents the learnable weight matrix corresponding to the matching of works, and softmax represents the softmax normalization function. Language a priori The calculation formula is as follows: ,in The learnable weight matrix corresponding to language matching; Translator's a priori knowledge The calculation formula is as follows: ,in This represents the learnable weight matrix corresponding to the translator matching; Obtain the language set corresponding to the largest language prior, obtain the translator set corresponding to the largest translator prior, and in the knowledge graph, for each work node, select the replica node that meets the following conditions: the work node corresponding to the replica node has a directed edge pointing to the version node, the language of the version node belongs to the language set, and the translator of the version node belongs to the translator set. Finally, the selected copy nodes will form the target translation set for the corresponding work node.

5. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Obtain the authorization penalty, the inverted value of the current time replica borrowable indication, the geo-allocation cost, and the time-period congestion degree associated with each directed edge in the knowledge graph. Then, perform a weighted summation of the negative logarithm of the relation confidence of the edge with the learnable weight vector, the authorization penalty, the inverted value of the borrowable indication, the geo-allocation cost, and the time-period congestion degree to obtain the edge cost of each directed edge. For each work node and each copy node, determine all legal paths from the work node to the copy node with the end copy in a borrowable state, calculate the cumulative edge cost of all directed edges on each legal path, and select the minimum value among the cumulative values ​​as the cross-layer shortest borrowable path cost from the work node to the copy node. For each work node, the minimum value is selected from the cross-layer shortest available path cost corresponding to all copy nodes included in its target translation set, and this value is taken as the target cost of that work node. Multiply the target cost of each work node by the corresponding work probability, and sum all the multiplication results to obtain the weighted target cost.

6. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Determine the replicas currently in a borrowable state as the candidate replica set; For each replica in the candidate replica set, the shortest available path cost across all work nodes corresponding to that replica is multiplied by the work probability of the corresponding work node, and all multiplication results are summed to obtain the candidate cross-layer cost of that replica. For each replica's candidate cross-layer cost, the weighted target cost is subtracted from the candidate cross-layer cost to obtain the cross-layer accessibility gradient of that replica.

7. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Calculate the personalized preference probability for each replica node, where the personalized preference probability for replica node i is... The calculation formula is as follows: ,in The transpose vector representing the user intent embedding. This represents the learnable weight matrix corresponding to personalized preference matching. This indicates the node embedding of replica node i. This represents a non-linear activation function.

8. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Calculate the joint score for each replica node, and the joint score for replica node i. The calculation formula is as follows: ,in and Let represent the first learnable positive scalar and the second learnable positive scalar, respectively. This represents the cross-level borrowable reachability gradient corresponding to replica node i; The joint scores of all replicas in the candidate replica set are sorted in descending order, and the top N replicas after sorting are selected to obtain the recommended list.

9. The book recommendation system based on artificial intelligence according to claim 1, characterized in that, Retrieve the user's selections from the recommended list and mark the copy selected by the user as the target copy; The objective function is constructed based on the joint score and the cross-level accessibility gradient. The calculation formula is as follows: ,in Represents the set of candidate replicas. Indicates the replica node selected by the user. The combined score This represents the joint score of any replica node j in the candidate replica set. Indicates the replica node selected by the user. Cross-level accessibility gradient, represents a nonnegative scalar, and exp represents an exponential function with the natural constant as the base; Set a positive scalar as the learning rate, calculate the gradient of the objective function with respect to the learnable parameters, and subtract the product of the learning rate and the corresponding gradient from each parameter to obtain the updated learnable parameters.

10. A book recommendation method based on artificial intelligence, characterized in that, Implementing the AI-based book recommendation system as described in any one of claims 1 to 9 includes the following steps: Step S201: Message passing is performed on the knowledge graph using a graph neural network to obtain node embeddings and edge relationship confidence. Step S202: Generate user intent embedding based on user search content; calculate work probability, language prior and translator prior based on user intent embedding and node embedding; and obtain target translation set by combining knowledge graph filtering. Step S203: Obtain the edge cost by weighted summation based on edge relation confidence, authorization penalty, inversion of borrowable indication, geographical allocation cost, and time period congestion. Calculate the shortest borrowable path cost across layers based on this cost, take the minimum value within the target translation set to obtain the target cost, and weight and summarize the target costs of different works based on the work probability. Step S204: Based on the work probability, the shortest available path cost across layers for each copy in the candidate copy set is weighted and summarized to obtain the candidate cross-layer cost. The candidate cross-layer cost is subtracted from the weighted target cost to obtain the cross-layer available accessibility gradient. Step S205: Generate personalized preference probabilities based on user intent embedding and node embedding, obtain a joint score by combining cross-level accessibility gradient, and generate a recommendation list in the candidate replica set based on the joint score; Step S206: Based on the user's selection of the recommended list, update the learnable parameters by combining the joint score and cross-level accessibility gradient.

Images