A multi-modal heterogeneous admission data integration method based on knowledge graph

Through the multimode heterogeneous admission data integration method based on knowledge graphs, the graph neural network is used to process the college entrance examination admission data of multiple provinces and cities, and the timeliness and correctness of traditional manual processing methods is solved, and efficient and accurate data integration processing is achieved.

CN118014065BActive Publication Date: 2024-11-22XINJIANG ZEZHI INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410140244.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-30
Publication Date
2024-11-22
Estimated Expiration
2044-01-30

AI Technical Summary

Technical Problem

Traditional manual processing methods are difficult to effectively handle the integration of college entrance examination admission data in many provinces and cities, resulting in poor timeliness and correctness. The exhaustive replacement method of existing products has high uncertainty and labor costs, and the risk of errors is relatively high.

Method used

A multimodal heterogeneous admission data integration method based on knowledge graph is adopted, and a data unified model and integration processing model are constructed by collecting and optimizing triplets, and a graph neural network is used to perform data integration processing to achieve data unity and integration.

Benefits of technology

It reduces the uncertainty of results, improves the automatic data integration processing capability, reduces the hassle and error risk of manual judgment, improves the accuracy rate, and is suitable for data processing in the college entrance examination mode in different provinces and cities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118014065B_ABST
    Figure CN118014065B_ABST
Patent Text Reader

Abstract

The present invention provides a multi-mode heterogeneous admission data integration method based on knowledge graph, which relates to the field of data integration and processing technology, and the steps include S1: collecting basic data related to the college entrance examination in various provinces and cities, extracting triples according to the basic data and optimizing and supplementing the triples; S2: constructing a neo4j graph database according to the optimized triples, and generating a unified data model based on the knowledge graph; S3: collecting admission data files, integrated records, and recruitment rule schemes of various provinces and cities over the years, and training the mapping rules of data integration based on the college entrance examination mode information of various provinces and cities, and generating an integrated processing model; S4: connecting the output end of the data unified model with the input end of the integrated processing model, constructing a combination model, and using the combination model to perform unified integrated processing output of data. The present invention decomposes the original problem into two small problems by constructing a unified data model and an integrated processing model, which is more suitable for computer processing, reduces human errors, and has a higher accuracy rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data integration and processing, and specifically to a multi-mode heterogeneous admission data integration method based on knowledge graph. Background Art

[0002] As the beginning of a closed loop of two-way selection and connection between candidates and colleges, the standardized processing of college admission data is an integrated task that needs to be done within a limited time, targeting different batches, and facing many differences. Its timeliness and accuracy greatly affect the normal admission of candidates, the allocation of college resources, and the development of subsequent work such as data statistical analysis. Therefore, it is more valuable to study at the moment when the old and new college entrance examination models are changing and the plans and policies of various provinces and cities are different.

[0003] The integrated processing of admission data involves many provinces and cities, and the college entrance examination models, plans, and recruitment rules of each province and city are different. The traditional manual processing method has been exhausted to cope with the data integration dilemma of the surge in the number of candidates, frequent model changes, and complex plan types. Existing products on the market are also mainly based on exhaustive replacement. Too many variables lead to considerable uncertainty in the generated results, and users are required to confirm the generated results twice. The labor cost is too high and the risk of error is high. In the context of informatization construction, artificial intelligence and machine learning technologies are becoming more and more mature. Compared with traditional manual processing solutions and current exhaustive products on the market, emerging knowledge graphs and neural networks have better basic conditions for structural data analysis, semantic association reasoning, and dynamic update and expansion, so they have great potential in the task of multi-modal heterogeneous admission data integration.

[0004] Therefore, it is necessary to provide an iterative multi-modal and heterogeneous admission data integration solution based on knowledge graph technology and build a systematic technology cluster for it.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-modal heterogeneous admission data integration method based on knowledge graph to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A multi-mode heterogeneous admission data integration method based on knowledge graph includes the following steps:

[0009] S1: Collect basic data related to the college entrance examination in various provinces and cities, extract triples based on the basic data, and optimize and supplement the triples;

[0010] S2: Build a neo4j graph database based on the optimized triples and generate a unified data model based on the knowledge graph;

[0011] S3: Collect the admission data files, integrated records, and recruitment rules of various provinces and cities over the years, and train the mapping rules of data integration based on the college entrance examination model information of various provinces and cities to generate an integrated processing model;

[0012] S4: Connect the output end of the data unification model with the input end of the integrated processing model, build a combined model, and use the combined model to perform unified integrated processing and output of the data.

[0013] Preferably, the basic data include the recruitment plans, majors, subjects, and batches of the institutions to be processed in various provinces and cities, and the admission data files include the original MDB files of recruitment in previous years and the original FDB files of recruitment in the current year.

[0014] Preferably, the specific steps of optimizing and supplementing the triples in step S1 are:

[0015] S101: Collect and organize the existing triples in the knowledge graph, including entity and relationship information, and randomly initialize a vector representation for each entity and relationship;

[0016] S102: define a positive triplet (h, r, t) and a negative triplet (h′, r′, t′), where the negative triplet (h′, r′, t′) is obtained by replacing entities or relations in the positive triplet with other random entities or relations, where h is the head entity, t is the tail entity, h′ is the replaced head entity, and t′ is the replaced tail entity;

[0017] S103: Randomly select a positive example triplet and several negative example triplet from the existing triples, and use the distance metric function optimization algorithm to minimize the distance difference between the positive example and the negative example. The calculation formula is as follows:

[0018] d pos =d(h,r,t)=|h+rt| 1

[0019] d neg =d(h′,r′,t′)=|h′+r′-t′| 1

[0020]

[0021] Where d pos Represents the distance metric of the positive triples, d neg represents the distance metric of negative triples, represents the gradient of the distance metric function, S represents the set of all positive triples, N represents the set of all negative triples, [X]+ Indicates a positive value, and γ is the preset boundary hyperparameter;

[0022] S104: using a back propagation algorithm to calculate the gradient of the distance metric function on the vector representation, generating a loss function of the distance metric function, and updating the vector representation of the entity and the relationship according to the loss function;

[0023] S105: Steps S102 to S104 are repeatedly executed until a predetermined number of iterations or convergence conditions are reached, and the negative example triples that meet the conditions are classified as completion objects, thereby completing the optimization and supplementation of the triples.

[0024] Preferably, a vector is randomly initialized for each entity and relationship by drawing random values ​​from a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0025] Preferably, the generation logic of the data unified model in step S2 is:

[0026] S201: construct a graph structure according to the optimized triples, and use entities and relationships as nodes of the graph and triples as edges of the graph;

[0027] S202: Build a neo4j graph database to store triples, use a graph neural network to train the triples, and configure a training function at each layer;

[0028] S203: Calculate and update the gradient of the training function through the back propagation algorithm, and then update the parameters of the data unification model through the stochastic gradient descent algorithm until the loss of the data unification model tends to be stable.

[0029] Preferably, the training function includes a message passing aggregation function for extracting information from neighbor nodes, a neighbor feature aggregation function for fusing information features of neighbor nodes and central nodes, and an update function for updating the feature representation of the central node. The specific calculation formula is as follows:

[0030] Message passing aggregation functions:

[0031] Neighbor feature aggregation function:

[0032] Update function:

[0033] In the formula represents the aggregation result of node i at layer l, N(i) represents the set of neighbor nodes of node i, represents the feature representation of node i at layer l, c ij represents the normalization constant, W (l) represents the weight matrix of the lth layer, and σ represents the activation function.

[0034] Preferably, the generation logic of the integrated processing model in step S3 is:

[0035] S301: Introduce neural network for training, and build an integrated processing model of three-layer unidirectional network based on two neural network layers connected to one fully connected layer;

[0036] S302: Initialize the model parameters of the integrated processing model, set the hidden neurons, learning rate, activation function and number of iterations of two neural network layers and one fully connected layer;

[0037] S303: Using the admission data file, integration records, and the target data dictionary for the year as inputs to the integrated processing model, the mapping rules for data integration of each province and city are trained to complete the optimization iteration of the integrated processing model.

[0038] Preferably, if the provincial and municipal college entrance examination mode of the current year is unchanged compared with the college entrance examination mode of last year, the neural network used is a one-way neural network, with last year's original recruitment MDB file, last year's integrated record and the current year's target data dictionary as input;

[0039] If the provincial or municipal college entrance examination model for the current year has changed compared to last year's model, the referenced neural network will be a bidirectional neural network, with the original FDB file of recruitment for that year, the recruitment rules and regulations for that year, and the target data dictionary for that year as input.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes a modular multi-mode heterogeneous data integration mode decoupled into "unification + integration" by constructing a data unification model and an integrated processing model and combining the two, and decomposes the original complex problem into small-scale problems of "data unification" and "data integration". By decomposing the complex problem, the uncertainty of the results caused by multiple variables is reduced, making the two small-scale problems more suitable for computer processing rather than judgment by users. This not only improves the ability to automatically integrate and process data, but also reduces the trouble of users manually making secondary judgments on the results, thereby reducing human errors and having a higher accuracy rate, which is convenient for colleges and universities to process and integrate the filing and admission data of colleges and universities under different college entrance examination modes in different provinces and cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of the present invention;

[0042] Figure 2 Schematic diagram of the process of step S1 in the present invention;

[0043] Figure 3 Schematic diagram of the process of step S2 in the present invention;

[0044] Figure 4 Schematic diagram of the process of step S3 in the present invention;

[0045] Figure 5 Schematic diagram of the knowledge graph in the present invention;

[0046] Figure 6 Schematic diagram of the architecture of the knowledge graph in the present invention. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0048] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0049] Embodiment 1:

[0050] See also Figure 1 , the present invention provides a technical solution:

[0051] A multi-mode heterogeneous admission data integration method based on knowledge graph includes the following steps:

[0052] S1: Collect basic data related to the college entrance examination in various provinces and cities, including the recruitment plans, majors, subjects, and batches of the institutions to be processed in various provinces and cities. Extract triples based on the basic data and optimize and supplement the triples. The extracted triples are as follows:

[0053] (Institutions, orientation, provinces and cities);

[0054] (province, city, recruitment, plan);

[0055] (province, setting, provincial code);

[0056] (plan, cover, professional);

[0057] (major, orientation, batch);

[0058] (professional, orientation, subject);

[0059] (Professional, mapping, national standard code)

[0060] Taking the TransH model as an example, the mathematical formula involved in embedding entity relationships based on the knowledge graph is as follows:

[0061] h r =h- <h,w r > ·w r +w r ,

[0062] t r =t- <t,w r > ·w r +w r ,

[0063] where h r and t r It means that through the relationship w r The embedding vector obtained by mapping the head entity h and the tail entity t can be used to perform other tasks such as triple reasoning completion and knowledge graph analysis applications.

[0064] By extracting triples based on knowledge graphs, information can be organized and retrieved in a structured manner, allowing computers to better understand and utilize knowledge. Given partial triples, it is also possible to predict missing triples or infer new relationships, thereby improving the ability to understand and apply knowledge and improving the effectiveness of data integration.

[0065] S2: Build a neo4j graph database based on the optimized triples, enter the above triples into the database and flexibly call the graph neural network as needed to generate a unified data model based on the knowledge graph. Compared with traditional relational databases, the neo4j graph database is more focused on storing and processing graph data structures, and can better represent complex relationships and network structures between entities.

[0066] S3: Collect the admission data files, integrated records, and recruitment rules of various provinces and cities over the years, and generate an integrated processing model based on the mapping rules of the training data integration based on the college entrance examination model information of each province and city. The admission data files include the original MDB files of previous years' recruitment and the original FDB files of the current year's recruitment.

[0067] S4: Connect the output end of the data unification model with the input end of the integrated processing model, build a combined model, and use the combined model to perform unified integrated processing and output of the data.

[0068] See also Figure 2 In the present invention, the specific steps of optimizing and supplementing the triples in step S1 are:

[0069] S101: Collect and organize the existing triples in the knowledge graph, including information such as entities, relationships, and attributes, and randomly initialize a vector representation for each entity and relationship.

[0070] S102: define a positive triple (h, r, t) and a negative triple (h′, r′, t′), wherein the negative triple (h′, r′, t′) is obtained by replacing entities or relations in the positive triple with other random entities or relations, wherein h is the head entity, t is the tail entity, h′ is the replaced head entity, and t′ is the replaced tail entity. During replacement, the head entity and the tail entity are respectively replaced with another random entity different from the original head entity and the tail entity, and the relation is replaced with another random relation different from the original relation. Taking the triple (college, facing, province and city) mentioned in this embodiment as an example, if it is defined as a positive triple, the negative triple can be (province and city, covering, provincial code) or (plan, recruitment, batch), etc. It can be seen that the richer the input target data dictionary, the more elements in the entity set and the relationship set in the triple, and the more positive triples and negative triples that can be extracted.

[0071] S103: Randomly select a positive example triplet and several negative example triplet from the existing triples, and use the distance metric function optimization algorithm to minimize the distance difference between the positive example and the negative example. The calculation formula is as follows:

[0072] d pos =d(h,r,t)=|h+rt| 1

[0073] d neg =d(h′,r′,t′)=|h′+r′-t′| 1

[0074]

[0075] Where d pos Represents the distance metric of the positive triples, d neg represents the distance metric of negative triples, represents the gradient of the distance metric function, S represents the set of all positive triples, N represents the set of all negative triples, [X] + It indicates a positive value, and γ is the preset boundary hyperparameter.

[0076] S104: According to the gradient of the distance metric function to the vector representation, the back propagation algorithm is used to calculate, generate the loss function of the distance metric function, and update the vector representation of the entity and relationship according to the loss function. Since the gradient descent method has good convergence and parallelism, it is not only easy to implement but also can process multiple data in parallel. Therefore, the gradient descent method is generally used to update the gradient of the vector representation, that is, the gradient of the vector representation is continuously updated by reducing the value of the loss function through the gradient descent method, thereby reducing the distance difference between the positive example triples and the negative example triples, so that the constructed data unified model is more accurate.

[0077] S105: Steps S102 to S104 are repeatedly executed until a predetermined number of iterations or convergence conditions are reached, and the negative example triples that meet the conditions are classified as completion objects, thereby completing the optimization and supplementation of the triples.

[0078] See also Figure 3 In the present invention, the generation logic of the data unified model in step S2 is:

[0079] S201: construct a graph structure according to the optimized triples, and use entities and relationships as nodes of the graph and triples as edges of the graph.

[0080] S202: Construct a neo4j graph database to store triples, use a graph neural network to train triples, and configure training functions at each layer, including a message passing aggregation function for extracting information from neighbor nodes, a neighbor feature aggregation function for fusing information features of neighbor nodes and central nodes, and an update function for updating the feature representation of the central node. The graph neural network can adopt a GCN neural network model, which has excellent local structure sensitivity, sharing, applicability and scalability, and can effectively capture the local structure information of nodes in the graph. Through the information transmission of neighbor nodes, the relationship between the node and the node directly connected to it can be better understood, and the GCN neural network can share weight parameters, effectively reducing the number of parameters of the model, reducing the complexity of the model, making it easier to expand, and more suitable for processing large quantities of data from various provinces and cities in the present invention. The specific calculation formulas of the three training functions are as follows:

[0081] Message passing aggregation functions:

[0082] Neighbor feature aggregation function:

[0083] Update function:

[0084] In the formula represents the aggregation result of node i at layer l, N(i) represents the set of neighbor nodes of node i, represents the feature representation of node i at layer l, c ij represents the normalization constant, W (l) Represents the weight matrix of the lth layer, σ represents the activation function. The above three functions work together on each layer of GCN. Through the process of information transmission and feature updating, the representation of nodes in the graph is gradually and iteratively learned.

[0085] S203: Calculate and update the gradient of the training function through the back propagation algorithm, and then update the parameters of the data unification model through the stochastic gradient descent algorithm until the loss of the data unification model tends to be stable.

[0086] See also Figure 4 In the present invention, the generation logic of the integrated processing model in step S3 is:

[0087] S301: First, compare the college entrance examination model of the target province or city in the current year with that of last year, and then introduce a neural network for training. Taking the BP neural network as an example, it has excellent adaptability, parallelism, learning and generalization capabilities, and can learn and approximate complex nonlinear mapping relationships. The training and reasoning processes can be highly parallelized, so the processing capability on large-scale data sets is relatively strong, and it is suitable for distributed computing environments. If the college entrance examination model of the province or city in the current year is unchanged compared with that of last year, the neural network referenced is a unidirectional BP neural network, with last year's original MDB file of recruitment, last year's integrated record and that year's target data dictionary as input. If the college entrance examination model of the province or city in the current year is changed compared with that of last year, the neural network referenced is a bidirectional BP neural network, with that year's original FDB file of recruitment, that year's recruitment rules and plans, and that year's target data dictionary as input. Finally, a three-layer unidirectional network integrated processing model is constructed based on two neural network layers connected to one fully connected layer.

[0088] When using a unidirectional BP neural network, the forward propagation of the unidirectional network is:

[0089] Input signal transmission: a l =σ(w l ·a l-1 +b l )

[0090] Among them, l is the current number of layers in the network, w l is the weight matrix connecting the l-1th and lth layers, b l is the bias matrix, a l-1 is the output of the l-1th layer, σ is the activation function;

[0091] Output layer transfer: y = σ (w L ·a L-1 +b L )

[0092] Among them, w L is the weight matrix connecting the penultimate layer and the output layer, b L is the bias matrix of the output layer, a L-1 is the output of the penultimate layer.

[0093] When a bidirectional BP neural network is used, its forward propagation is divided into forward propagation and backward propagation processes, which are specifically expressed as follows:

[0094] Forward pass: a l =σ(w l ·a l-1 +U l ·a′ l +b l )

[0095] Similar to the symbolic representation of the input signal transmission in the forward propagation of a unidirectional network, additionally, U l represents the backward weight connecting the lth layer and the l+1th layer, a′ l is the input of the backward pass of layer l;

[0096] Backward pass: a′ l =σ(V l ·a l+1 +b′ l )

[0097] Among them, V l represents the backward weight connecting the l+1th layer and the lth layer, b′ l is the backward bias term of the lth layer, a l+1 is the output of the l+1th layer;

[0098] Output layer transfer: y = σ (w L ·a L-1 +U L ·a′ L +b L )

[0099] Similar to the symbolic representation of the input signal transmission in the forward propagation of a unidirectional network, additionally, U L represents the backward weight connecting the output layer and the penultimate layer, a′ L is the input to the backward pass of the output layer.

[0100] The forward propagation process of a unidirectional network is irreversible and is suitable for supervised model training under the condition of complete historical data. The forward propagation process of a bidirectional network involves two transmission processes, forward and backward. The training takes longer but the feature extraction is more detailed, making it more suitable for supervised model training without historical data support.

[0101] S302: Initialize the model parameters of the integrated processing model, set the hidden neurons, learning rate, activation function and number of iterations of two neural network layers and one fully connected layer, set the number of hidden neurons in the first and second neural network layers to 10, the number of hidden neurons in the fully connected layer to 24, the learning rate to 0.1, select sigmoid and tanh as activation functions, and finally set the number of iterations to 1000.

[0102] S303: Take the admission data file, integration record, and target data dictionary of the year as the input of the integrated processing model, train the mapping rules for the data integration of each province and city, and complete the optimization iteration of the integrated processing model. After the optimization is completed, the data integration processing work of the year can be carried out based on this integrated processing model. The admission database results of the colleges and universities to be processed in each province and city are input into the unified model of the knowledge graph to standardize the dictionary format of colleges, majors, batches and special plans, and then the standardized results and the data rules of the province and city are delivered to the data integration processing model. After specifying the output format, the processing model will uniformly process the standardized data of each province and city and output them in a standardized manner.

[0103] Embodiment 2:

[0104] In the present invention, the GCN neural network can be used as an encoder in combination with a decoder to complete the model training in addition to being a construction subject in Example 1. Let the input feature be X (a feature matrix representing a node in the graph, where each row corresponds to a feature vector of a node), and the reconstructed feature be The mathematical formulas involved after combining the decoder are as follows:

[0105] Encoder part: in represents the feature representation of node i at layer l, represents the set of neighbor nodes of node i, W (l) represents the weight matrix of the lth layer, and σ represents the activation function.

[0106] Decoder part: in represents the output feature representation of the final encoder, and U represents the weight matrix of the decoder.

[0107] By combining GCN neural network and graph autoencoder, feature extraction, reconstruction and generation of graph data can be achieved. GCN neural network can extract meaningful node representations from the graph, while graph autoencoder can use global structural information to reconstruct features, thereby improving the robustness and expressiveness of the model.

[0108] In summary, the present invention first extracts some triples from the basic data related to the college entrance examination in various provinces and cities, such as the recruitment plans, majors, subjects, batches, etc. of science colleges in various provinces and cities, and uses knowledge graphs to optimize and predict triples. By organizing and retrieving information in a structured manner, computers can better understand and utilize knowledge, improve their understanding and application capabilities of knowledge, and improve the effect of data integration. The optimized triples are then used as input to construct a unified data model using the GCN graph neural network, and the model is continuously optimized and iterated through the back propagation algorithm and the gradient descent method, so that the constructed unified data model is more accurate. At the same time, an integrated processing model is constructed based on two different neural networks, and the college entrance examination model of the target province and city in the current year is compared with the college entrance examination model of last year, and a more suitable neural network model is selected according to the comparison results. The forward propagation of the unidirectional BP neural network The process is irreversible and is suitable for supervised model training under the condition of complete data from previous years. The forward propagation process of the bidirectional BP neural network has two transmission processes, forward and backward. The training takes longer but the feature extraction is more detailed. It is more suitable for supervised model training without historical data support. By using different neural network models in different situations, the accuracy of the integrated processing model can be further improved. Finally, the data unification model is combined with the integrated processing model, and the input end of the data unification model is used as the input end of the integrated processing model. The original complex problem is decomposed into small-scale problems of "data unification" and "data integration". The computer has fewer variables when processing each small problem, and the results are more accurate, thereby reducing the trouble of users manually making secondary judgments on the results, reducing human errors, and having a higher accuracy rate, which is convenient for colleges and universities to process and integrate the filing and admission data of colleges and universities under different college entrance examination models in different provinces and cities.

[0109] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0110] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by computer software, electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0111] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0112] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A multi-modal heterogeneous admission data integration method based on knowledge graph, characterized in that: The steps include: S1: Collect basic data related to the college entrance examination in various provinces and cities, extract triples based on the basic data, and optimize and supplement the triples; S2: Build a neo4j graph database based on the optimized triples and generate a unified data model based on the knowledge graph; The generation logic of the data unified model in step S2 is: S201: construct a graph structure according to the optimized triples, and use entities and relationships as nodes of the graph and triples as edges of the graph; S202: Build a neo4j graph database to store triples, use a graph neural network to train the triples, and configure a training function at each layer; S203: Calculate and update the gradient of the training function through the back propagation algorithm, and then update the parameters of the data unification model through the stochastic gradient descent algorithm until the loss of the data unification model tends to be stable; The training function includes a message passing aggregation function for extracting information from neighbor nodes, a neighbor feature aggregation function for fusing information features of neighbor nodes and central nodes, and an update function for updating the feature representation of the central node. The specific calculation formula is as follows: Message passing aggregation functions: Neighbor feature aggregation function: Update function: In the formula represents the aggregation result of node i at layer l, N(i) represents the set of neighbor nodes of node i, represents the feature representation of node i at layer l, c ij represents the normalization constant, W (l) represents the weight matrix of the lth layer, σ represents the activation function; S3: Collect the admission data files, integrated records, and recruitment rules of various provinces and cities over the years, and train the mapping rules of data integration based on the college entrance examination model information of various provinces and cities to generate an integrated processing model; The generation logic of the integrated processing model in step S3 is: S301: Introduce neural network for training, and build an integrated processing model of three-layer unidirectional network based on two neural network layers connected to one fully connected layer; S302: Initialize the model parameters of the integrated processing model, set the hidden neurons, learning rate, activation function and number of iterations of two neural network layers and one fully connected layer; S303: Using the admission data file, integration records, and the target data dictionary of the year as inputs to the integrated processing model, the mapping rules for data integration of each province and city are trained to complete the optimization iteration of the integrated processing model; S4: Connect the output end of the data unification model with the input end of the integrated processing model, build a combined model, and use the combined model to perform unified integrated processing and output of the data.

2. According to claim 1, a multi-mode heterogeneous admission data integration method based on knowledge graph is characterized by: The basic data include the recruitment plans, majors, subjects and batches of the institutions to be processed in various provinces and cities, and the admission data files include the original MDB files of recruitment in previous years and the original FDB files of recruitment in the current year.

3. According to claim 2, a multi-mode heterogeneous admission data integration method based on knowledge graph is characterized by: The specific steps of optimizing and supplementing the triples in step S1 are: S101: Collect and organize the existing triples in the knowledge graph, including entity and relationship information, and randomly initialize a vector representation for each entity and relationship; S102: define a positive triplet (h, r, t) and a negative triplet (h′, r′, t′), where the negative triplet (h′, r′, t′) is obtained by replacing entities or relations in the positive triplet with other random entities or relations, where h is the head entity, t is the tail entity, h′ is the replaced head entity, and t′ is the replaced tail entity; S103: Randomly select a positive example triplet and several negative example triplet from the existing triples, and use the distance metric function optimization algorithm to minimize the distance difference between the positive example and the negative example. The calculation formula is as follows: d pos =d(h,r,t)=|h+r-t|1 d neg =d(h′,r′,t′)=|h′+r′-t′|1 Where d pos Represents the distance metric of the positive triples, d neg represents the distance metric of negative triples, represents the gradient of the distance metric function, S represents the set of all positive triples, N represents the set of all negative triples, [X] + Indicates a positive value, and γ is the preset boundary hyperparameter; S104: using a back propagation algorithm to calculate the gradient of the distance metric function on the vector representation, generating a loss function of the distance metric function, and updating the vector representation of the entity and the relationship according to the loss function; S105: Steps S102 to S104 are repeatedly executed until a predetermined number of iterations or convergence conditions are reached, and the negative example triples that meet the conditions are classified as completion objects, thereby completing the optimization and supplementation of the triples.

4. According to claim 3, a multi-mode heterogeneous admission data integration method based on knowledge graph is characterized by: A vector is randomly initialized for each entity and relation by drawing random values ​​from a standard normal distribution with mean 0 and standard deviation 1.

5. According to claim 1, a multi-mode heterogeneous admission data integration method based on knowledge graph is characterized by: If the provincial or municipal college entrance examination model for the current year remains unchanged compared to last year's model, the neural network used is a one-way neural network, with last year's original recruitment MDB file, last year's integrated records, and this year's target data dictionary as input; If the provincial or municipal college entrance examination model for the current year has changed compared to last year's model, the referenced neural network will be a bidirectional neural network, with the original FDB file of recruitment for that year, the recruitment rules and regulations for that year, and the target data dictionary for that year as input.