A customized perspective multi-domain campus element knowledge graph collaborative representation method

Abstract

The present invention discloses a customized perspective multi-domain campus element knowledge graph collaborative representation method, which belongs to the field of artificial intelligence technology, and includes the following steps: First, according to the highly semantic features of the campus element knowledge graph, the RotatE knowledge graph embedding method is used to obtain the semantic features of the initial knowledge graph of a specific perspective. Then, a graph attention neural network of joint information is constructed to extract the representation vectors of the semantic and structural features of the joint initial graph. Finally, the multi-domain knowledge graph is processed by a collective entity alignment method, and the alignment accuracy is improved by combining additional information and using a variety of metrics, thereby realizing the collaborative representation of the knowledge graph. The present invention adopts the above-mentioned customized perspective multi-domain campus element knowledge graph collaborative representation method, which can eliminate information islands in the campus system, promote the collaborative representation of multi-domain campus elements, and provide information support for a variety of downstream applications based on big data.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a customized perspective multi-domain campus element knowledge graph collaborative representation method. Background Art

[0002] Customized education has become an important trend in the evolution of current and future education methods. The realization of customized education requires information support from all perspectives and all time and space to optimize education evaluation and decision-making. In the current informatization process, the development and deployment of information systems are usually aimed at meeting the needs and functions of specific fields, resulting in the focus on the completion of independent tasks and the realization of goals in the design and operation of campus systems, while ignoring the coordination and data interoperability between systems. This has led to the formation of a large number of scattered and isolated information systems on campus, namely the so-called information islands. These information islands not only hinder the integration and comprehensive analysis of data, but also lead to increasingly prominent problems such as repeated information entry, data inconsistency, and low information utilization. How to effectively link and integrate these information islands and build an education ecosystem based on big data has become an important challenge in the current development of customized education.

[0003] As a multi-relation heterogeneous graph, knowledge graph has been widely used in social networks, intelligent search, recommendation systems, question-answering systems and other fields. Because of its highly semantic and rich connectivity characteristics, it is highly consistent with the data in the field of education, so it has been widely used in the field of education. Yongjun Wei published the article "Research on the Construction and Application of Knowledge Graph of Digital Resources in Vocational Colleges" in "Journal of Computers" in 2023. By analyzing the characteristics and current status of digital resources in vocational colleges, he proposed a way to use knowledge graphs to represent digital resources in vocational colleges. ZhiYao Wu published "Construction and Application of a Major-Specific Knowledge Graph Based on Big Data in Education" in the journal "International Journal of Emerging Technologies in Learning (iJET)" in 2022. The construction and application of English professional knowledge graphs based on big data were studied to achieve structured utilization of teaching resources and optimization of teaching quality. The above research explored the application possibilities of knowledge graphs in the representation of campus elements and their application level in downstream tasks. Summary of the invention

[0004] The purpose of this invention is to provide a customized perspective multi-domain campus element knowledge graph collaborative representation method, focusing on how to associate information islands represented as knowledge graphs under the premise of unified expressiveness and realize the collaborative representation of multi-domain campus elements.

[0005] To achieve the above object, the present invention provides a customized perspective multi-domain campus element knowledge graph collaborative representation method, comprising the following steps:

[0006] S1. For the user-customized multi-domain initial knowledge graph, the semantic features of the initial knowledge graph are extracted through the RotatE algorithm, and its structural features are extracted by combining the improved graph attention network GAT. The two types of features are fused and used as the input of subsequent modules;

[0007] S2, introduce K-means clustering algorithm and multiple similarity measurement methods to perform collective entity alignment on multi-domain knowledge graphs and achieve effective association of cross-domain elements;

[0008] S3. The Bootstrapping training method is introduced to improve the performance of the model in the absence of prior information. The node degree comparison method is used to replace the triple elements to achieve a unified representation of cross-domain elements.

[0009] Preferably, the specific operation of S1 is:

[0010] S11. For all initial knowledge graphs IKG, use the RotatE method to convert the nodes in IKG into low-dimensional embedding vectors that reflect the semantic relationship of their triples. , let a set of relation triples be ,in Representing the head entity, relation and tail entity in a set of triples respectively, the model of RotatE is expressed as:

[0011] (1)

[0012] in, is the scoring function between the head entity and the tail entity embedding. The higher it is, the better the model performance is. for norm, is the Hadamard product, , , Head entity node 、Relationship Edge , tail entity node The initial low-dimensional embedding vector of

[0013] S12. Obtain rich structural features in IKG, design GAT that integrates features, and realize knowledge graph representation learning;

[0014] S13. Considering that some nodes in the knowledge graph will be difficult to distinguish their structural features and semantic features after the neural network iteration due to structural problems, the acquired feature vectors are processed to make the features more recognizable;

[0015] S14. In order to make the final representation of the node able to be well used for the subsequent entity alignment task, a suitable loss function is defined to reasonably train the final representation of the node.

[0016] Preferably, the specific operation of S12 is:

[0017] S121. To obtain the structural features of a node in the knowledge graph, we first need to obtain the influence factor between the node and other adjacent nodes:

[0018] (2)

[0019] in, For the The node and The influence factor between nodes, is the LeakyReLU activation function, is a learnable linear transformation vector The transpose of and are the learnable linear transformation matrices corresponding to the source node and neighbor node, respectively. For the The first nodes, To find the function of the sum of the first-order neighbor nodes, here , For the The first nodes, for The initial low-dimensional embedding vector of for The initial low-dimensional embedding vector of

[0020] S122. Use the softmax function to normalize the influence factor of each node's neighbors on themselves, and obtain the neighbor's attention coefficient to the node:

[0021] (3)

[0022] in, For the The node and The attention coefficient between nodes, For Node The set consisting of itself and its first-order neighbor nodes;

[0023] S123, process the attention coefficient and node embedding through R independent attention heads, observe the node features from multiple perspectives, and obtain a vector containing the node structure information:

[0024] (4)

[0025] in, Contains nodes The feature vector of structural information, and In the The attention heads win the attention coefficient and the linear transformation matrix of the neighboring nodes. A vector concatenation operator that concatenates the numbers To serial number A vector of

[0026] S124, in order to make the next layer network retain the context information of the node, the structure information vector and the feature vector of the current layer are input into the next layer network in parallel:

[0027] (5)

[0028] in, for In GAT The structural information characteristics of the layer, and For the graph attention network The learnable parameters in the layer are hour, for , is the vector parallel operator;

[0029] The specific operation of S13 is:

[0030] S131. In order to obtain good feature resolution, the attention mechanism is used to fuse the semantic information vector and the structural information vector to obtain the combined feature:

[0031] (6)

[0032] in, Is a node The combined representation of and are the attention weight coefficients corresponding to the structural information features and semantic information features, for and The common linear transformation matrix, is the number of network layers of GAT, is the feature information vector output from the final layer of GAT;

[0033] S132, and Using softmax processing and The linear transformation of is:

[0034] (7)

[0035] Similarly:

[0036] (8)

[0037] S133, making the final representation of the node include multi-dimensional semantic and structural information, concatenating the semantic representation, structural representation and combined representation to obtain the final representation vector:

[0038] (9)

[0039] in, Is a node The final representation of

[0040] The specific operation of S14 is as follows:

[0041] S141, maximize the similarity between the final representation vectors of the node pairs belonging to the same alignment seed tuple, and obtain the positive example loss function:

[0042] (10)

[0043] in, is the set of all alignment seed tuples The base number, is the binomial coefficient, For the Group alignment seed tuple, the alignment seed is the prior knowledge used to train the node representation for subsequent entity alignment tasks, is the cosine similarity function, which is defined as , and belong to the same alignment seed tuple, i.e. , , for The final representation of for The final representation of

[0044] S142. Consider minimizing the similarity between the final representation vectors of node pairs that do not belong to the same alignment seed tuple, and obtain the negative example loss function:

[0045] (11)

[0046] in, For the removal The random negative sample nodes extracted from the aligned seed tuple outside are ;

[0047] S143. Combine and normalize the positive example loss function and the negative example loss function to obtain the final loss function:

[0048] (12).

[0049] Preferably, the specific operation of S2 is:

[0050] S21, using K-means clustering method to process node representation in blocks to identify potential matching node combinations, and then only calculate the similarity of nodes in the same cluster, thereby effectively reducing the amount of calculation;

[0051] S22, after the blocks are divided, a plurality of similarity measures are applied to the node pairs in the block clusters, and then the similarities are comprehensively compared to align the nodes;

[0052] S23. The Gale-Shapley algorithm is introduced in the entity alignment decision. According to the requirements of the multi-graph entity alignment task, the GS algorithm is improved to obtain the node groups representing the same real entity in different knowledge graphs, which are called corresponding entity groups. , and the predicted corresponding entity group set is .

[0053] Preferably, the specific operation of S21 is:

[0054] S211, first randomly select node representations as cluster centers, For the number of clusters required, the following distance function should be minimized during the clustering process:

[0055] (13)

[0056] in, For the The node representation set contained in a cluster, For the The node representation as the cluster center in the cluster, For Node The final representation of

[0057] S212, the cluster center is updated as follows:

[0058] (14)

[0059] After continuously updating the cluster center, The value of no longer changes or changes less than a certain threshold, and the block processing is completed.

[0060] Preferably, the specific operation of S22 is:

[0061] S221. First, the representations of the nodes in the block clusters are compared using cosine similarity to obtain the cosine similarity of the representations of each node:

[0062] (15)

[0063] in Is a node The final representation of

[0064] S222. There is a large amount of text information in the knowledge graph. In order to achieve the association of multi-domain elements, extremely high entity alignment accuracy is required. The additional information of node name is jointly measured by the dice coefficient and Levenshtein similarity to enhance the accuracy of entity alignment.

[0065] S2221. The node name string is regarded as a string. The dice coefficient focuses on calculating the similarity between the characters contained in the node name. The dice coefficient between the node names is:

[0066] (16)

[0067] in For Node The node name character sequence; For Node The node name character sequence;

[0068] S2222, Levenshtein distance refers to the number of times characters are inserted, deleted, and replaced in the process of changing a string into another string. This number is regarded as , Levenshtein distance is expressed in the form of similarity to obtain Levenshtein similarity, which focuses on calculating the similarity of the positions of the same characters in the string between node names:

[0069] (17)

[0070] in The function is designed to obtain the maximum value of the elements within the brackets;

[0071] S2223. Using a single similarity to compare node names will result in poor entity alignment due to the limitations of a single indicator. Therefore, a joint similarity combining the dice coefficient and the Levenshtein similarity is used in similarity measurement, which is calculated as follows:

[0072] (18)

[0073] S223, the similarity obtained by comparing the node representation vector and the similarity obtained by comparing the node name are expressed in matrix form, and the similarity is obtained respectively. and In order to reasonably use the similarity matrix for entity alignment tasks, combined with and Get the final similarity matrix , which is calculated as follows:

[0074] (19)

[0075] in, is a hyperparameter used to adjust the impact of additional node name information on the entity alignment task.

[0076] Preferably, the specific operation of S23 is:

[0077] S231. Create an empty dictionary , used to record the aligned node pairs between different graphs. The dictionary key is the graph number tuple sorted by size, and the value is an empty list. Set a threshold , used to determine whether the node meets the alignment conditions and set the loop counter Initialized to 0, all nodes in the IKG are uniformly numbered, from 1 to the total number of nodes in all IKGs , Indicates the number Nodes;

[0078] S232, searching for aligned node pairs according to node similarities between different graphs;

[0079] S2321. For each pair of nodes , calculate their similarity in turn ;

[0080] S2322, obtain node pair Corresponding knowledge graph combination ;

[0081] S2323, if the similarity Greater than threshold , then further determine whether the node pair already exists in the dictionary of In the key corresponding item;

[0082] S2324. If the node pair does not exist, initialize the maximum similarity value and , respectively calculate and update and The maximum similarity of the correlation;

[0083] S2325, if the current similarity If the maximum similarity value exceeds the existing one, Remove the include and Node pairs, and the new node pairs join in middle;

[0084] S233, merging node pairs representing the same entity to construct an aligned entity group, when When there are still key values ​​that are not an empty set, continue to execute steps S2331 to S2333;

[0085] S2331: From Pick any key ,from Randomly select a pair of nodes in , add its elements to the Aligned entity groups and from Delete the node pair ;

[0086] S2332: When Any one of them is not Key , The node pairs that exist in the current alignment entity group Continue this step while they share at least one common element; group these entities with the current alignment Node pairs that share at least one common element are added to in, and from Delete these node pairs;

[0087] S2333: Align entity groups Add to collection , and make .

[0088] Preferably, the specific operation of S3 is:

[0089] S31. Introduce the Bootstrapping method in entity alignment. By iteratively utilizing high-confidence alignment results, the alignment ability of the model is gradually enhanced, thereby improving the overall alignment effect.

[0090] S32. The corresponding entity groups obtained through entity alignment can be used as link points for aggregated IKG to achieve collaborative representation of multi-domain campus elements and construct a triple library of AKG to achieve collaborative representation of all discrete information.

[0091] Preferably, the specific operation of S31 is:

[0092] S311, using the initial alignment seed to perform an alignment task, resampling a high confidence node group that is not the original alignment seed from the result of the task, adding it to the alignment seed, and performing initial alignment;

[0093] S312, retraining the representation learning model using a new alignment seed containing a high confidence alignment node group;

[0094] S313, using the node representation obtained by the retraining model to perform entity alignment, clearing the node group of the alignment seed added in the previous training, resampling the alignment node group with high confidence in this training, and adding them to the alignment seed;

[0095] S314. Repeat steps S312 to S313 to perform iterative updates until the alignment results converge, and then stop the iteration.

[0096] Preferably, the specific operation of S32 is:

[0097] S321, compare the node degree in the corresponding entity group with the weighted sum of the node degrees of its neighbor nodes to determine the unified expression of the node, and the weighted sum is calculated as follows:

[0098] (20)

[0099] in, is a function used to calculate the node degree, u for The neighbor nodes of is any node in the corresponding entity group;

[0100] S322: In a series of semantically aligned triples, the expressiveness of the relationship edge is judged by its frequency of occurrence in the corresponding IKG, thereby determining the unified expression of the edge; in the knowledge graph, the frequently occurring relationship edges remain consistent in different contexts, showing high semantic stability;

[0101] S323, after obtaining the unified expression of nodes and the unified expression of edges, performing triple node replacement and triple elimination respectively;

[0102] S3231, corresponding to the entity group The middle node is a set of triplets consisting of the head entity or the tail entity, which is defined as ,Will Replace all head or tail entities in with The corresponding unified node expression completes the triple node replacement, namely:

[0103] (twenty one)

[0104] in, represents a triple, After node replacement , is the difference operator, represents the triplet obtained by replacing node B with node C in triplet A. To unify node expressions;

[0105] S3232: If after the node replacement of the triples, there are two or more triples with the same head entity and tail entity, select the triple with the most expressive relationship edge to retain, and remove the rest of such triples, that is:

[0106] (twenty two)

[0107] in, for The set of triples with the same head and tail entities in , After the relationship edge removal operation , To unify the edge expression;

[0108] S3233: Will be retained in all Add the triples in to the triples of AKG to get AKG, that is,

[0109] (twenty three)

[0110] in, is a triplet set of AKG, is the number of campus element entities represented by all isolated knowledge graph information, is a set concatenation operator, which is used to concatenate To serial number A collection of For the The triple set of IKGs, The number of IKG.

[0111] Therefore, the present invention adopts the above-mentioned customized perspective multi-domain campus element knowledge graph collaborative representation method, which can eliminate information islands in the campus system, promote the collaborative representation of multi-domain campus elements, and provide information support for various downstream applications based on big data.

[0112] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is a representation learning framework diagram for obtaining isolated knowledge graph representation in an embodiment of a customized visual threshold multi-domain campus element knowledge graph collaborative representation method of the present invention;

[0114] Figure 2 It is a multi-domain knowledge graph collective entity alignment framework diagram of an embodiment of a customized perspective multi-domain campus element knowledge graph collaborative representation method of the present invention;

[0115] Figure 3 It is an attention neural network framework diagram for acquiring node structure information in an embodiment of a customized visual threshold multi-domain campus element knowledge graph collaborative representation method of the present invention. DETAILED DESCRIPTION

[0116] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0117] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0118] Embodiment 1

[0119] The present invention provides a customized perspective multi-domain campus element knowledge graph collaborative representation method, comprising the following steps:

[0120] S1. For the user-customized multi-domain initial knowledge graph, the semantic features of the initial knowledge graph are extracted by the RotatE algorithm, and its structural features are extracted by combining the improved graph attention network GAT. The two types of features are fused and used as the input of subsequent modules, such as Figure 1 As shown, the specific operations are:

[0121] S11. For all initial knowledge graphs IKG, use the RotatE method to convert the nodes in IKG into low-dimensional embedding vectors that reflect the semantic relationship of their triples. , let a set of relation triples be ,in Representing the head entity, relation and tail entity in a set of triples respectively, the model of RotatE is expressed as:

[0122] (1)

[0123] in, is the scoring function between the head entity and the tail entity embedding. The higher it is, the better the model performance is. for norm, is the Hadamard product, , , Head entity node 、Relationship Edge , tail entity node The initial low-dimensional embedding vector of

[0124] S12. In order to obtain the rich structural features in IKG, a GAT with fusion features is designed to realize knowledge graph representation learning, such as Figure 3 As shown, the specific operations are:

[0125] S121. To obtain the structural features of a node in the knowledge graph, we first need to obtain the influence factor between the node and other adjacent nodes:

[0126] (2)

[0127] in, For the The node and The influence factor between nodes, is the LeakyReLU activation function, is a learnable linear transformation vector The transpose of and are the learnable linear transformation matrices corresponding to the source node and neighbor node, respectively. For the The first nodes, To find the function of the sum of the first-order neighbor nodes, here , For the The first nodes, for The initial low-dimensional embedding vector of for The initial low-dimensional embedding vector of

[0128] S122. Use the softmax function to normalize the influence factor of each node's neighbors on themselves, and obtain the neighbor's attention coefficient to the node:

[0129] (3)

[0130] in, For the The node and The attention coefficient between nodes, For Node The set consisting of itself and its first-order neighbor nodes;

[0131] S123, process the attention coefficient and node embedding through R independent attention heads, observe the node features from multiple perspectives, and obtain a vector containing the node structure information:

[0132] (4)

[0133] in, Contains nodes The feature vector of structural information, and In the The attention heads win the attention coefficient and the linear transformation matrix of the neighboring nodes. A vector concatenation operator that concatenates the numbers To serial number A vector of

[0134] S124, in order to make the next layer network retain the context information of the node, the structure information vector and the feature vector of the current layer are input into the next layer network in parallel:

[0135] (5)

[0136] in, for In GAT The structural information characteristics of the layer, and For the graph attention network The learnable parameters in the layer are hour, for , is the vector parallel operator;

[0137] S13. Considering that some nodes in the knowledge graph will be difficult to distinguish their structural features and semantic features after the neural network iteration due to structural problems, in order to make the features more recognizable, the acquired feature vectors are processed. The specific operations are as follows:

[0138] S131. In order to obtain good feature resolution, the attention mechanism is used to fuse the semantic information vector and the structural information vector to obtain the combined feature:

[0139] (6)

[0140] in, Is a node The combined representation of and are the attention weight coefficients corresponding to the structural information features and semantic information features, for and The common linear transformation matrix, is the number of network layers of GAT, is the feature information vector output from the final layer of GAT;

[0141] S132, and Using softmax processing and The linear transformation of is:

[0142] (7)

[0143] Similarly:

[0144] (8)

[0145] S133, making the final representation of the node include multi-dimensional semantic and structural information, concatenating the semantic representation, structural representation and combined representation to obtain the final representation vector:

[0146] (9)

[0147] in, Is a node The final representation of

[0148] S14. In order to make the final representation of the node well used for the subsequent entity alignment task, a suitable loss function is defined to reasonably train the final representation of the node. The specific operation is as follows:

[0149] S141, maximize the similarity between the final representation vectors of the node pairs belonging to the same alignment seed tuple, and obtain the positive example loss function:

[0150] (10)

[0151] in, is the set of all alignment seed tuples The base number, is the binomial coefficient, For the Group alignment seed tuple, the alignment seed is the prior knowledge used to train the node representation for subsequent entity alignment tasks, is the cosine similarity function, which is defined as , and belong to the same alignment seed tuple, i.e. , , for The final representation of for The final representation of

[0152] S142. Consider minimizing the similarity between the final representation vectors of node pairs that do not belong to the same alignment seed tuple, and obtain the negative example loss function:

[0153] (11)

[0154] in, For the removal The random negative sample nodes extracted from the aligned seed tuple outside are ;

[0155] S143. Combine and normalize the positive example loss function and the negative example loss function to obtain the final loss function:

[0156] (12).

[0157] S2, introduces K-means clustering algorithm and a variety of similarity measurement methods to perform collective entity alignment on multi-domain knowledge graphs and achieve effective association of cross-domain elements, such as Figure 2 As shown, the specific operations are:

[0158] S21. Entity alignment aims to identify different nodes representing the same entity from knowledge graphs of different domains. It searches for aligned node combinations by matching similarities between nodes. However, since there is no alignment relationship between most nodes, directly calculating the similarity between all nodes will lead to sparse matching problems. To this end, the K-means clustering method is used to block the node representation to identify potential matching node combinations, and then only the similarity of nodes in the same cluster is calculated, thereby effectively reducing the amount of calculation. The specific operation is as follows:

[0159] S211, first randomly select node representations as cluster centers, is the number of clusters required. During the clustering process, the following distance function should be minimized:

[0160] (13)

[0161] in, For the The node representation set contained in a cluster, For the The node representation as the cluster center in the cluster, For Node The final representation of

[0162] S212, the cluster center is updated as follows:

[0163] (14)

[0164] After continuously updating the cluster center, The value of no longer changes or changes less than a certain threshold, and the block processing is completed.

[0165] S22, after the blocks are divided, multiple similarity measures are used for the node pairs in the block clusters, and then these similarities are comprehensively compared to align these nodes. The specific operation is as follows:

[0166] S221. First, the representations of the nodes in the block clusters are compared using cosine similarity to obtain the cosine similarity of the representations of each node:

[0167] (15)

[0168] in Is a node The final representation of

[0169] S222. There is a large amount of text information in the knowledge graph. In order to achieve the association of multi-domain elements, extremely high entity alignment accuracy is required. To this end, the dice coefficient and Levenshtein similarity are used to jointly measure the additional information of node name to enhance the accuracy of entity alignment;

[0170] S2221. The node name string is regarded as a string. The dice coefficient focuses on calculating the similarity between the characters contained in the node name. The dice coefficient between the node names is:

[0171] (16)

[0172] in For Node The node name character sequence; For Node The node name character sequence;

[0173] S2222, Levenshtein distance refers to the number of times characters are inserted, deleted, and replaced in the process of changing a string into another string. This number is regarded as Levenshtein similarity can be obtained by expressing the Levenshtein distance in the form of similarity, which focuses on calculating the similarity of the positions of the same characters in the string between node names:

[0174] (17)

[0175] in The function is designed to obtain the maximum value of the elements within the brackets;

[0176] S2223. Using a single similarity to compare node names will result in poor entity alignment due to the limitations of a single indicator. Therefore, a joint similarity combining the dice coefficient and the Levenshtein similarity is used in similarity measurement, which is calculated as follows:

[0177] (18).

[0178] S223, the similarity obtained by comparing the node representation vector and the similarity obtained by comparing the node name are expressed in matrix form, and the similarity is obtained respectively. and In order to reasonably use the similarity matrix for entity alignment tasks, combined with and Get the final similarity matrix , which is calculated as follows:

[0179] (19)

[0180] in, is a hyperparameter used to adjust the impact of additional node name information on the entity alignment task.

[0181] S23. In the entity alignment task, the objects of similarity matching are divided into source nodes and target nodes. Usually, the entity alignment task searches for the target node with the highest similarity through the source node. However, in order to achieve entity alignment in multi-domain knowledge graphs, unlike traditional entity alignment, it is necessary to consider the potential collective dependency problem caused by the simultaneous alignment of multiple graphs and multiple nodes. The root cause of this problem is that the similarity between the source node and the target node is only a local maximum, not a global maximum. Therefore, this problem can be regarded as a stable matching problem. To solve this problem, the Gale-Shapley algorithm is introduced in the entity alignment decision, which ensures that each node can find a stable match. According to the needs of the multi-graph entity alignment task, the GS algorithm is improved. The goal of this algorithm is to obtain a group of nodes that represent the same real entity in different knowledge graphs, which are called corresponding entity groups. , and the predicted corresponding entity group set is , and its specific operations are:

[0182] S231. Create an empty dictionary , used to record the aligned node pairs between different graphs. The dictionary key is a tuple of graph numbers sorted by size, and the value is an empty list. Set a threshold , used to determine whether the node meets the alignment conditions and set the loop counter Initialized to 0. All nodes in the IKG are numbered uniformly, from 1 to the total number of nodes in all IKGs. . Indicates the number Nodes;

[0183] S232, searching for aligned node pairs according to the node similarities between different graphs, the specific operation is:

[0184] S2321. For each pair of nodes , calculate their similarity in turn ;

[0185] S2322, obtain node pair Corresponding knowledge graph combination ;

[0186] S2323, if the similarity Greater than threshold , then further determine whether the node pair already exists in the dictionary of In the key corresponding item;

[0187] S2324. If the node pair does not exist, initialize the maximum similarity value and , respectively calculate and update and The maximum similarity of the correlation;

[0188] S2325, if the current similarity If the maximum similarity value exceeds the existing one, Remove the include and Node pairs, and the new node pairs join in middle.

[0189] S233, merging node pairs representing the same entity to construct an aligned entity group, when When there are still key values ​​that are not an empty set, continue to execute steps S2331 to S2333;

[0190] S2331: From Pick any key ,from Randomly select a pair of nodes in , add its elements to the Aligned entity groups and from Delete the node pair ;

[0191] S2332: When Any one of them is not Key , The node pairs that exist in the current alignment entity group Continue this step while they share at least one common element; group these entities with the current alignment Node pairs that share at least one common element are added to in, and from Delete these node pairs;

[0192] S2333: Align entity groups Add to collection , and make .

[0193] S3, introduces the Bootstrapping training method to improve the performance of the model in the absence of prior information, and uses the node degree comparison method to replace the triple elements to achieve a unified representation of cross-domain elements. The specific operations are as follows:

[0194] S31. The IKG for graph aggregation originates from isolated graphs of different times and systems, and usually faces the cold start problem due to the lack of alignment seeds as prior knowledge. If the entity alignment task is performed directly, the alignment effect is often poor due to too few alignment seeds for reference. To this end, the Bootstrapping method is introduced in entity alignment. By iteratively utilizing high-confidence alignment results, the alignment ability of the model is gradually enhanced, thereby improving the overall alignment effect. The specific operations are as follows:

[0195] S311, using the initial alignment seed to perform an alignment task, resampling a high confidence node group that is not the original alignment seed from the result of the task, adding it to the alignment seed, and performing initial alignment.

[0196] S312: Retrain the representation learning model using a new alignment seed containing a high-confidence alignment node group.

[0197] S313, using the node representation obtained by the retraining model to perform entity alignment, clearing the node group of the alignment seed added in the previous training, resampling the alignment node group with high confidence in this training, and adding them to the alignment seed.

[0198] S314, repeat steps S312 to S313, perform iterative updates, and stop iteration until the alignment results converge. By using Bootstrapping, the alignment performance of multiple isolated graph entity alignment can be effectively improved, which effectively improves the automation and accuracy of this semi-supervised learning.

[0199] S32. The corresponding entity groups obtained through entity alignment can be used as link points for aggregated IKG to achieve collaborative representation of multi-domain campus elements. To this end, a triple library of AKG is constructed to achieve collaborative representation of all discrete information. Specifically, it is necessary to perform element replacement operations on the triples in IKG, and add the replaced triples to the triple set of AKG. The element replacement operation of the triples first needs to determine the unified node expression and edge expression of each corresponding entity group. The unified node expression is determined by finding the most expressive node in the corresponding entity group, while the unified edge expression is achieved by finding the most expressive relationship edge in the semantically aligned triple set. Finally, the most expressive nodes and relationships found are used as unified expressions to replace the elements of the triples, thereby completing the construction of AKG. The specific operations are:

[0200] S321. The expressiveness of a node is determined by the degree of the node and the degree of its neighboring nodes to determine the unified expression of the node. The degree of a node reflects its connectivity, that is, the number of edges directly connected to the node, which means that the degree of a node determines the degree of its direct association with other nodes. At the same time, the degree of a node is positively correlated with the ability of information to propagate between nodes and its scope of influence. The degree of a neighboring node reflects the closeness between a node and other non-neighboring nodes. If a neighboring node has a higher degree, it indicates that the node is in a highly connected area in the graph and has a higher centrality and importance. Therefore, it is believed that nodes with a high degree and whose neighboring nodes also have a high degree have stronger information dissemination and acquisition capabilities, and their own information influence range is larger. Such nodes are regarded as representative nodes in the information set and have good expressiveness. Based on this, the unified expression of the node is determined by comparing the weighted sum of the node degree and the degree of its neighboring nodes in the corresponding entity group. The weighted sum is calculated as follows:

[0201] (20)

[0202] in, is a function used to calculate the node degree, u for The neighbor nodes of is any node in the corresponding entity group.

[0203] S322: In a series of semantically aligned triples, the expressiveness of the relationship edge can be judged by the frequency of occurrence in the corresponding IKG, thereby determining the unified expression of the edge. In the knowledge graph, frequently occurring relationship edges remain consistent in different contexts, showing high semantic stability. This indicates that these relationships can accurately reflect the true semantic connection between entities in different contexts, thereby reducing the risk of mismatching. Therefore, relationship edges with high frequency of occurrence have higher representativeness and credibility in the knowledge graph and are suitable as the unified expression of edges.

[0204] S323, after obtaining the unified expression of nodes and the unified expression of edges, triple node replacement and triple elimination are performed respectively, and the specific operations are as follows:

[0205] S3231, corresponding to the entity group The middle node is a set of triplets consisting of the head entity or the tail entity, which is defined as .Will Replace all head or tail entities in with The corresponding unified node expression can complete the triple node replacement. That is:

[0206] (twenty one)

[0207] in, represents a triple, After node replacement , is the difference operator, represents the triplet obtained by replacing node B with node C in triplet A. To unify the node expression.

[0208] S3232: Furthermore, if after the nodes of the triples are replaced, there are two or more triples with the same head entity and tail entity, the triple with the most expressive relationship edge is selected and retained, and the rest of such triples are eliminated. That is:

[0209] (twenty two)

[0210] in, for The set of triples with the same head and tail entities in , After the relationship edge removal operation , To unify the edge expression.

[0211] S3233: Will be retained in all Add the triples in to the triples of AKG to get AKG. That is:

[0212] (twenty three)

[0213] in, is a triplet set of AKG, is the number of campus element entities represented by all isolated knowledge graph information, is a set concatenation operator, which is used to concatenate To serial number A collection of For the The triple set of IKGs, The number of IKG, through The construction of AKG realizes the information aggregation and association of multi-domain and diverse campus elements.

[0214] Therefore, the present invention adopts the above-mentioned customized perspective multi-domain campus element knowledge graph collaborative representation method, which can eliminate information islands in the campus system, promote the collaborative representation of multi-domain campus elements, and provide information support for various downstream applications based on big data.

[0215] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A customized perspective multi-domain campus element knowledge graph collaborative representation method, characterized by: The following steps are involved: S1. There are information islands in different domains of interpersonal relationships, subject knowledge, personal abilities, and campus facilities in the campus information system. They are expressed in the form of knowledge graphs. The multi-domain initial knowledge graphs formed by the information of the domains required for different customized educational needs are used as input. For the user-customized multi-domain initial knowledge graphs, the nodes in the knowledge graphs are converted into low-dimensional embedding vectors reflecting the semantic relationship of their triples through the RotatE algorithm, and the improved graph attention network is combined to extract and fuse structural features. After fusion, the representation vectors of all nodes in all initial knowledge graphs are output, and the feature vectors are processed to enhance their recognizability. Finally, a suitable loss function is defined to train the node representation, and the node representation is used as the input of subsequent modules. S2, introduce K-means clustering algorithm and multiple similarity measurement methods to perform collective entity alignment on multi-domain knowledge graphs and achieve effective association of cross-domain elements; S21, using K-means clustering method to process node representation in blocks to identify potential matching node combinations, and then only calculate the similarity of nodes in the same cluster, thereby effectively reducing the amount of calculation; S22, after the blocks are divided, a plurality of similarity measures are applied to the node pairs in the block clusters, and then the similarities are comprehensively compared to align the nodes; S23. The Gale-Shapley algorithm is introduced in the entity alignment decision. The GS algorithm is improved according to the requirements of the multi-graph entity alignment task. The traditional GS algorithm only performs the best match for objects in two domains. The improved GS algorithm expands the targeted domain to a number greater than 2 and with no upper limit, and does not limit the number of objects in the matching group, and allows the objects in the matching group to achieve stable matching, which adapts to the matching problem in the entity alignment of multi-domain knowledge graphs. By introducing a dictionary M and a similarity threshold B that record multi-graph alignment node pairs, it ensures that only high-similarity node pairs are considered for alignment. In addition, the algorithm maintains the maximum similarity value of each node to avoid confusion caused by multiple alignments, and constructs a unified aligned entity group by merging aligned entity pairs of shared nodes, achieving consistent and effective entity alignment across multiple domains, improving the accuracy and robustness of alignment in a multi-domain environment, and obtaining node groups that represent the same real entity in different knowledge graphs, which are called corresponding entity groups. , and the predicted corresponding entity group set is ; S3. Introduce the Bootstrapping training method, and regard the node combination with high similarity obtained by the similarity measurement in S2 as a pseudo-alignment seed with high confidence. The pseudo-alignment seed and the real alignment seed are used together as the prior knowledge to retrain the model, and the node representation vector obtained after retraining is entity aligned again to obtain the predicted aligned entity group. The predicted aligned entity group associates the knowledge graphs between different domains, so that the information between them is organically combined, thereby realizing the collaborative representation of knowledge graphs in multiple different domains.

2. According to claim 1, a customized perspective multi-domain campus element knowledge graph collaborative representation method is characterized by: The specific operation of S1 is: S11. For all initial knowledge graphs IKG, use the RotatE method to convert the nodes in IKG into low-dimensional embedding vectors that reflect the semantic relationship of their triples. , let a set of relation triples be ,in Representing the head entity, relation and tail entity in a set of triples respectively, the model of RotatE is expressed as: （1）; in, is the scoring function between the head entity and the tail entity embedding. The higher it is, the better the model performance is. for norm, is the Hadamard product, , , Head entity node 、Relationship Edge , tail entity node The initial low-dimensional embedding vector of S12. Obtain rich structural features in IKG, design GAT that integrates features, and realize knowledge graph representation learning; S13. Considering that some nodes in the knowledge graph will be difficult to distinguish their structural features and semantic features after the neural network iteration due to structural problems, the acquired feature vectors are processed to make the features more recognizable; S14. In order to make the final representation of the node able to be well used for the subsequent entity alignment task, a suitable loss function is defined to reasonably train the final representation of the node.

3. According to claim 2, a customized perspective multi-domain campus element knowledge graph collaborative representation method is characterized by: The specific operation of S12 is: S121. To obtain the structural features of a node in the knowledge graph, we first need to obtain the influence factor between the node and other adjacent nodes: （2）； in, For the The node and The influence factor between nodes, is the LeakyReLU activation function, is a learnable linear transformation vector The transpose of and are the learnable linear transformation matrices corresponding to the source node and neighbor node, respectively. For the The first nodes, To find the function of the sum of the first-order neighbor nodes, here , For the The first nodes, for The initial low-dimensional embedding vector of for The initial low-dimensional embedding vector of S122. Use the softmax function to normalize the influence factor of each node's neighbors on themselves, and obtain the neighbor's attention coefficient to the node: （3）； in, For the The node and The attention coefficient between nodes, For Node The set consisting of itself and its first-order neighbor nodes; S123, process the attention coefficient and node embedding through R independent attention heads, observe the node features from multiple perspectives, and obtain a vector containing the node structure information: （4）； in, Contains nodes The feature vector of structural information, and In the The attention heads win the attention coefficients and the linear transformation matrix of neighbor nodes. A vector concatenation operator that concatenates the numbers To serial number A vector of S124, in order to make the next layer network retain the context information of the node, the structure information vector and the feature vector of the current layer are input into the next layer network in parallel: （5）； in, for In GAT The structural information characteristics of the layer, and For the graph attention network The learnable parameters in the layer are hour, for , is the vector parallel operator; The specific operation of S13 is: S131. In order to obtain good feature resolution, the attention mechanism is used to fuse the semantic information vector and the structural information vector to obtain the combined feature: （6）； in, Is a node The combined representation of and are the attention weight coefficients corresponding to the structural information features and semantic information features, for and The common linear transformation matrix, is the number of network layers of GAT, is the feature information vector output from the final layer of GAT; S132, and Using softmax processing and The linear transformation of is: （7）； Similarly: （8）； S133, making the final representation of the node include multi-dimensional semantic and structural information, concatenating the semantic representation, structural representation and combined representation to obtain the final representation vector: （9）； in, Is a node The final representation of The specific operation of S14 is: S141, maximize the similarity between the final representation vectors of the node pairs belonging to the same alignment seed tuple, and obtain the positive example loss function: （10）； in, is the set of all alignment seed tuples The base number, is the binomial coefficient, For the Group alignment seed tuple, the alignment seed is the prior knowledge used to train the node representation for subsequent entity alignment tasks, is the cosine similarity function, which is defined as , and belong to the same alignment seed tuple, i.e. , , for The final representation of for The final representation of S142. Consider minimizing the similarity between the final representation vectors of node pairs that do not belong to the same alignment seed tuple, and obtain the negative example loss function: （11）； in, To remove The random negative sample nodes extracted from the aligned seed tuple outside are ; S143. Combine and normalize the positive example loss function and the negative example loss function to obtain the final loss function: （12）。 4. According to claim 3, a customized perspective multi-domain campus element knowledge graph collaborative representation method is characterized by: The specific operation of S21 is: S211, first randomly select node representations as cluster centers, For the number of clusters required, the following distance function should be minimized during the clustering process: （13）； in, For the The node representation set contained in a cluster, For the The node representation as the cluster center in the cluster, For Node The final representation of S212, the cluster center is updated as follows: （14）； After continuously updating the cluster center, until The value of no longer changes or changes less than a certain threshold, and the block processing is completed.

5. According to claim 4, a customized perspective multi-domain campus element knowledge graph collaborative representation method is characterized by: The specific operation of S22 is: S221. First, the representations of the nodes in the block clusters are compared using cosine similarity to obtain the cosine similarity of the representations of each node: （15）； in Is a node The final representation of S222. There is a large amount of text information in the knowledge graph. In order to achieve the association of multi-domain elements, extremely high entity alignment accuracy is required. The additional information of node name is jointly measured by the dice coefficient and Levenshtein similarity to enhance the accuracy of entity alignment. S2221. The node name string is regarded as a string. The dice coefficient focuses on calculating the similarity between the characters contained in the node name. The dice coefficient between the node names is: （16）； in For Node A sequence of node names, For Node The node name character sequence; S2222, Levenshtein distance refers to the number of times characters are inserted, deleted, and replaced in the process of changing from one string to another. This number is regarded as , Levenshtein distance is expressed in the form of similarity to obtain Levenshtein similarity, which focuses on calculating the similarity of the positions of the same characters in the string between node names: （17）； in The function is designed to obtain the maximum value of the elements within the brackets; S2223. Using a single similarity to compare node names will result in poor entity alignment due to the limitations of a single indicator. Therefore, a joint similarity combining the dice coefficient and the Levenshtein similarity is used in similarity measurement, which is calculated as follows: （18）； S223, the similarity obtained by comparing the node representation vector and the similarity obtained by comparing the node name are expressed in matrix form, and the similarity is obtained respectively. and In order to reasonably use the similarity matrix for entity alignment tasks, combined with and Get the final similarity matrix , which is calculated as follows: （19）； in, is a hyperparameter used to adjust the impact of additional node name information on the entity alignment task.

6. According to claim 5, a customized perspective multi-domain campus element knowledge graph collaborative representation method is characterized by: The specific operation of S23 is: S231. Create an empty dictionary , used to record the aligned node pairs between different graphs. The dictionary key is the graph number tuple sorted by size, and the value is an empty list. Set a threshold , used to determine whether the node meets the alignment conditions and set the loop counter Initialized to 0, all nodes in the IKG are uniformly numbered, from 1 to the total number of nodes in all IKGs , Indicates the number Nodes; S232, searching for aligned node pairs according to node similarities between different graphs; S2321. For each pair of nodes , calculate their similarity in turn ; S2322, obtain node pair Corresponding knowledge graph combination ; S2323, if the similarity Greater than threshold , then further determine whether the node pair already exists in the dictionary of In the key corresponding item; S2324. If the node pair does not exist, initialize the maximum similarity value and , respectively calculate and update and The maximum similarity of the correlation; S2325, if the current similarity If the maximum similarity value exceeds the existing one, Remove the include and Node pairs, and the new node pairs join in middle; S233, merging node pairs representing the same entity to construct an aligned entity group, when When there are still key values ​​that are not an empty set, continue to execute steps S2331 to S2333; S2331: From Pick any key ,from Randomly select a pair of nodes in , add its elements to the Aligned entity groups and from Delete the node pair ; S2332: When Any one of them is not Key , The node pairs that exist in the current alignment entity group Continue this step while they share at least one common element; group these entities with the current alignment Node pairs that share at least one common element are added to in, and from Delete these node pairs; S2333: Align entity groups Add to collection , and make .

7. The customized perspective multi-domain campus element knowledge graph collaborative representation method according to claim 6 is characterized by: The specific operation of S3 is: S31. Introduce the Bootstrapping method in entity alignment. By iteratively utilizing high-confidence alignment results, the alignment ability of the model is gradually enhanced, thereby improving the overall alignment effect. S32. The corresponding entity groups obtained through entity alignment can be used as link points for aggregated IKG to achieve collaborative representation of multi-domain campus elements and construct a triple library of AKG to achieve collaborative representation of all discrete information.

8. The customized perspective multi-domain campus element knowledge graph collaborative representation method according to claim 7 is characterized by: The specific operation of S31 is: S311, using the initial alignment seed to perform an alignment task, resampling a high confidence node group that is not the original alignment seed from the result of the task, adding it to the alignment seed, and performing initial alignment; S312, retraining the representation learning model using a new alignment seed containing a high confidence alignment node group; S313, using the node representation obtained by the retraining model to perform entity alignment, clearing the node group of the alignment seed added in the previous training, resampling the alignment node group with high confidence in this training, and adding them to the alignment seed; S314. Repeat steps S312 to S313 to perform iterative updates until the alignment results converge, and then stop the iteration.

9. The customized perspective multi-domain campus element knowledge graph collaborative representation method according to claim 8 is characterized by: The specific operation of S32 is: S321, compare the node degree in the corresponding entity group with the weighted sum of the node degrees of its neighbor nodes to determine the unified expression of the node, and the weighted sum is calculated as follows: （20）； in, is the function used to calculate the node degree, u for The neighbor nodes of is any node in the corresponding entity group; S322: In a series of semantically aligned triples, the expressiveness of the relationship edge is judged by its frequency of occurrence in the corresponding IKG, thereby determining the unified expression of the edge; in the knowledge graph, the frequently occurring relationship edges remain consistent in different contexts, showing high semantic stability; S323, after obtaining the unified expression of nodes and the unified expression of edges, performing triple node replacement and triple elimination respectively; S3231, corresponding to the entity group The middle node is a set of triplets consisting of the head entity or the tail entity, which is defined as ,Will Replace all head or tail entities in with The corresponding unified node expression completes the triple node replacement, namely: （21）； in, represents a triple, After node replacement , is the difference operator, represents the triplet obtained by replacing node B with node C in triplet A. To unify node expressions; S3232: If after the node replacement of the triples, there are two or more triples with the same head entity and tail entity, select the triple with the most expressive relationship edge to retain, and remove the rest of such triples, that is: （22）； in, for The set of triples with the same head and tail entities in , After the relationship edge removal operation , To unify the edge expression; S3233: Will be retained in all Add the triples in to the triples of AKG to get AKG, that is, （23）； in, is a triplet set of AKG, is the number of campus element entities represented by all isolated knowledge graph information, is a set concatenation operator, which is used to concatenate To serial number A collection of For the The triple set of IKGs, The number of IKG.

Images