Science and technology service resource classification method, system and computer-readable storage medium based on evolutionary neural network

Abstract

The present invention relates to a classification method for science and technology service resources based on an evolutionary neural network. This method combines a novel optimization algorithm with an extreme learning machine network to optimize the parameters of both the random hidden layer and the output layer simultaneously. While improving the training efficiency of the science and technology service resource classification model, it enhances the classification accuracy of text data and can automatically classify and integrate the collected science and technology service resources according to industry divisions. It also includes a corresponding system, specifically including: a dataset construction module that extracts different types of resource data, classifies them by industry category and adds labels to construct classification training and test datasets; a data preprocessing module that cleans the data in the dataset and performs Chinese word segmentation, stop word removal, and Word2Vec vectorization processing to obtain the feature vectors and vocabulary of all data; a network evolution module that constructs an evolutionary neural network based on the Banyan Tree Growth Optimization (BTGO) algorithm and ELM, and applies a novel update operator to optimize the parameters of the hidden layer and the output layer; a resource classification module that uses the evolved neural network model to evaluate the model for the text data in the science and technology service resource test set and perform automatic resource classification on unknown data.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data processing, in particular to the technical field of science and technology services in modern service industries, and specifically relates to a method and system for classifying science and technology service resources based on an evolutionary neural network and a computer-readable storage medium thereof. Background Art

[0002] While providing comprehensive services for various scientific and technological innovation activities, the science and technology service industry has also greatly promoted economic growth and scientific and technological progress and has become an important part of the modern service industry. As the basic element of science and technology services, effective integration of science and technology service resources can improve the resource utilization rate and the achievement transformation efficiency. Science and technology resources include types such as talent information, intellectual property rights, and patented technologies, and are mainly composed of text data. How to quickly and effectively extract features and classify massive text information is a problem that current machine learning models need to solve.

[0003] Currently, classification algorithms for text data include decision trees, support vector machines, K-nearest neighbor algorithms, neural networks, etc. These methods are simple to implement, but there are often problems with insufficient accuracy for science and technology resource data with high dimensions and large sparsity. The Extreme Learning Machine (ELM) is a new type of feedforward neural network training algorithm that uses randomly set weights and biases for the input layer and hidden layer and calculates the weights of the output layer by combining the generalized inverse matrix. ELM has good data fitting and generalization capabilities, and at the same time, due to its simple implementation and fast training speed, it is widely used in tasks such as classification and fitting.

[0004] Since ELM introduces random hidden layer weights, it reduces the training complexity during training, but its random parameter settings often make it difficult to ensure stability and reliability in practical applications. Therefore, in many studies, the random parameters of ELM are optimized based on meta-heuristic optimization algorithms, and the network performance is globally improved after multiple iterative solutions. For example, particle swarm optimization, genetic algorithms, differential evolution, etc. are used to search for the weights and structures of the hidden layer and input layer to form an evolutionary neural network method. However, these evolutionary neural network methods only focus on the optimization of random parameters and do not effectively utilize and optimize the output layer parameters. Therefore, there is a need to improve and study new evolutionary neural network methods.

[0005] Currently, for the method of processing science and technology service resource data, Chinese Patent Application No. "CN201410260379.X" discloses "A method for obtaining and pushing scientific and technological intelligence based on text classification and image deep mining"; Chinese Patent Application No. "CN201811127718.1" discloses "An intelligent sorting method for agricultural science and technology services based on big data"; Chinese Patent Application No. "CN202110554334.3" discloses "A method for classifying scientific and technological literature based on a knowledge graph"; Chinese Patent Application No. "CN201510846339.8" discloses "A neural network image classification method based on an improved adaptive genetic algorithm"; Chinese Patent Application No. "CN201610012847.0" discloses "A neural network evolution method based on multi-local search"; Chinese Patent Application No. "CN201510714025.2" discloses "An implementation method of an adaptive learning neural network based on an evolutionary algorithm". In the disclosed inventions or documents, no instance of a scientific and technological service resource data classification method using the evolutionary neural network method based on the present invention is mentioned. Summary of the Invention

[0006] The present invention provides a classification method, system and computer-readable storage medium for scientific and technological service resources based on an evolutionary neural network, which uses an evolutionary neural network based on a new optimization algorithm to classify massive scientific and technological resource text data, improving the training efficiency of the scientific and technological service resource classification model while enhancing the classification accuracy of the text data. For the above purposes, the technical solutions adopted by the present invention are as follows:

[0007] The classification method for scientific and technological service resources based on an evolutionary neural network is mainly characterized in that the method includes the following steps:

[0008] (1) Extract different types of scientific and technological service resource data, classify each data according to the classification of the scientific and technological service industry and add labels to construct a classification training and test data set;

[0009] (2) Perform Chinese word segmentation and stop word removal processing on the text data of the data set, and use the Word2Vec method for text vectorization to obtain the feature vectors and word libraries of each data for subsequent model evolution and evaluation;

[0010] (3) Construct an evolutionary neural network based on the Banyan Tree Growth Optimization BTGO algorithm and ELM, and at the same time optimize the random parameters and output layer parameters of the neural network to improve the network classification accuracy while enhancing the convergence speed, and obtain the best network model for resource classification;

[0011] (4) Use the finally obtained neural network model to evaluate the text data of the test data set of the scientific and technological service resource data, and perform automated resource classification on unknown data.

[0012] Preferably, in step (1), the extracted resource data types include types such as talent information, intellectual property, and patented technologies, including metadata of different industries regularly extracted from the science and technology service resource library in actual operation, and the training, validation, and test sets are divided according to a ratio to meet the actual resource integration needs.

[0013] Preferably, in step (2), after data cleaning, Chinese word segmentation, and stop word removal processing on the data set constructed in step (1), the Word2Vec model pre-trained on a large batch of science and technology text data sets is used to perform text vectorization processing on it, convert it into dense feature vectors, and by sorting out all text data, the feature vectors and One-hot encoded labels of the data set are obtained for subsequent model evolution and evaluation.

[0014] Preferably, in step (3), the banyan tree growth optimization algorithm is applied to encode and optimize all parameters of the ELM network, and the global optimization strategy of BTGO is used to evolve both the random parameters and the output layer parameters simultaneously to achieve the optimal prediction accuracy and training speed at the same time.

[0015] Preferably, the initialization optimization process in step (3) specifically includes the following steps:

[0016] (3.1) Initialize the parameters of the banyan tree growth optimization BTGO algorithm and the ElM network parameters, set the population size NP of BTGO, the maximum number of iterations Gmax, the number of banyan tree branches m, the growth stage threshold T growth and the regeneration stage threshold T recover ; set the number of hidden layer nodes of the ELM network;

[0017] (3.2) Set the fitness function of the optimization algorithm, that is, the classification error rate of the neural network determined by x i on the test set, and is represented by the following formula:

[0018]

[0019] where TP and TN are the number of positive classes predicted as positive classes and the number of negative classes predicted as negative classes respectively, and FP and FN are the number of negative classes predicted as positive classes and the number of positive classes predicted as negative classes respectively;

[0020] (3.3) BTGO adopts a real number coding framework, and each branch and leaf individual represents a set of neural network parameters W, which consists of the hidden layer weight W hidden and the output layer weight W output of the neural network. The i-th individual can be expressed as:

[0021] x i =[Whidden ; W output

[0022] = [x i，1 , x i，2 , …, x i，h , x i，h+1 , …, x i，D (2)

[0023] where h is the sum of the number of weights and biases in the hidden layer of the neural network, and D is the sum of the number of all weights;

[0024] (3.4) Randomly initialize 80% of the individuals in the BTGO population. The set of serial numbers of this part of the individuals is denoted as I (i ∈ I), that is, the values of all dimensions of the individual x i (i ∈ I) are random values within the range of [-1, 1], and its calculation is shown in Equation (3):

[0025] x ij = x min，j + r × (x max，j - x min，j ) (3)

[0026] where x min，j and x max，j represent the minimum and maximum values of the solution space respectively, and r is a random value within the range of [0, 1];

[0027] (3.5) For the remaining 20% of the individuals in step (3.4), the set of their serial numbers is denoted as I * , indicating that the positions of the hidden layer weight part are randomly generated, and the output layer weight part is calculated by the generalized inverse MPP formula of ELM, that is, the first h bits of the individual x i (i ∈ I * ) are calculated by Equation (3), and the calculation of the remaining bits is shown in Equation (4), and the complete calculation is shown in Equation (5).

[0028]

[0029]

[0030] where H is the output matrix of the hidden layer, X is the input matrix, β is the output layer weight, and β m represents the m-th bit of the output layer weight β calculated by the generalized inverse;

[0031] (3.6) Apply the parameters x i saved in the individuals of the initial population to construct the network model o i , and predict the labels of the test set

[0032] ​

[0033] where \(g()\) is the activation function of the neural network, specifically the sigmoid function, to compare the predicted label and the true label \(y\) to calculate the classification error rate, and calculate the fitness value of the individual according to Equation (1) in step (3.2);

[0034] (3.7) Initialize the position \(P\) of the BTGO aerial roots root and take it as the individual's historical optimal position. Divide the branch group according to the fitness ranking of the individuals in the population, and divide them into \(m\) subgroups in ascending order. The individual with the best fitness in each branch group is used as the optimal position of the subgroup, denoted as \(P\) trunk ;

[0035] (3.8) The individual with the smallest adaptive fitness value in all subgroups is regarded as the current best position \(P\) best , and update the fitness value of the best position.

[0036] Preferably, the complete processing of optimizing the network model in step (3) specifically includes the following steps:

[0037] (3.9) Denote \(count\) i as the number of times the individual \(i\) has not been updated. If the fitness value of the current individual \(i\) is smaller than that of its aerial root individual , then set the value of \(count\) i to 0, otherwise increment its value by 1;

[0038] (3.10) Execute different update operators of BTGO according to the value of \(count\) i to generate new individuals of the banyan population. Among them, for the first \(h\) bits of the individual \(i\in I\) * , they are updated by the update operator, and the remaining bits are calculated by Equation (5): When the number of times the individual \(i\) has not been updated \(count\) i is lower than \(T\) growth , with a probability of \(pr\), select to execute the aerial root operator or the multi-trunk operator; when it is greater than \(T\) growth and less than \(T\) recover , execute the adjustment operator to generate a new population;

[0039] (3.11) Execute the rooting operator or partial generalized inverse to update the individual: When the random probability of the individual is less than \(pr\), execute the rooting operator to update as follows:

[0040]

[0041] where \(r\) is a random number between \([-1, 1]\), and \(F\) is a scaling factor. The partial positions of the individuals satisfying the conditions in the above step (3.10) are updated by MPP;

[0042] (3.12) Execute multi-trunk operator or partial MPP update: When the random probability of individual i is greater than pr, execute the multi-trunk operator, and its update method is as shown in formula (8):

[0043]

[0044] Where r is a random number in the range of [-1, 1], F is the scaling factor, is the best position of the trunk group where the i-th individual is located. Similarly, for individuals that meet the conditions, MPP is used to update some positions;

[0045] (3.13) Execute the adjustment operator or update the partial position in MPP mode: When the individual count i Cumulatively higher than T growth And less than T recover When BTGO executes the adjustment operator, it randomly selects a subgroup C from the m trunk subgroups. r , individual i moves closer to the optimal position of the subgroup, which is expressed as follows:

[0046]

[0047] Where s is the adjustment factor and is set to a fixed value;

[0048] (3.14) Execute the rebirth operator or update the individual partial position in MPP mode: When the individual count i Cumulative greater than T recover When , individual i turns to execute the rebirth operator, and the position of the individual is randomly initialized again, and its update is as follows:

[0049] x ij =x min,j +r×(x max,j -x min,j )(10)

[0050] If the individual i∈I * , then the first h bits are reinitialized by equation (10), and the remaining bits are calculated by equation (4);

[0051] (3.15) Calculate the new population individual x obtained by the above update operator i Build the corresponding network model o i , and predict the label and calculate the fitness value according to step (3.5);

[0052] (3.16) Based on the spatial distance between individuals and each trunk population, they are re-divided into the new trunk subgroup with the closest distance. The distance and the new subgroup number are calculated as follows:

[0053]

[0054]

[0055] (3.17) Compare x i with 's fitness value. If x i has a smaller fitness, then update the corresponding position of the aerial root to the position of x i , and clear the number of times count i of this individual to 0, otherwise increase the number by 1;

[0056] (3.18) Update the best positions P trunk and fitness values of m sub - populations, and update the optimal individual P best of the entire population

[0057] (3.19) Determine whether the current termination condition is reached. If the iteration reaches the maximum number of iterations, return to step (3.9) to continue execution, otherwise output the optimal individual in the current dataset and the classification accuracy of its corresponding classifier.

[0058] Preferably, in step (4), use the neural network model in the above step (3) to classify the unlabeled and unclassified new resource data, and evaluate the model and integrate and optimize the resources in combination with the test dataset.

[0059] More preferably, the specific steps of step (4) are as follows:

[0060] Construct a neural network model o best using the best parameters obtained in step (3). The parameters of the neural network model are composed of the parameters represented by the optimal individual P best , and it is expressed as follows:

[0061]

[0062] where X is the input vector, and are the determined best neural network hidden - layer weights and output - layer weights respectively;

[0063] Select the text data in the science and technology service resource test set to evaluate the model to obtain the accuracy of the final model, and perform the following formula (13) to automatically classify the unknown data X and obtain the prediction result Y:

[0064]

[0065] The science and technology service resource classification system based on an evolutionary neural network for implementing the above method is mainly characterized in that the system includes:

[0066] A dataset construction module, which is used to extract different types of technology service resource data, classify each piece of data according to the technology service industry classification and add labels, and construct a classification training and test dataset;

[0067] A data preprocessing module, connected to the dataset construction module, which is used to process the acquired dataset by clearing missing, incorrect and redundant data, perform Chinese word segmentation and stop word removal on the text data therein, and at the same time use the Word2Vec method for text vectorization to obtain the feature vectors and vocabulary of each data, for subsequent model evolution and evaluation;

[0068] A network evolution module, connected to the data preprocessing module, which is used to construct an evolutionary neural network based on the banyan tree growth optimization BTGO algorithm and ELM, and at the same time optimize the random parameters and output layer parameters of the neural network to improve the network classification accuracy while enhancing the convergence speed, and obtain the best network model for resource classification; and

[0069] A resource classification module, connected to the network evolution module, which is used to evaluate the model for the text data of the technology service resource test set based on the best network model and perform automatic resource classification on unknown data.

[0070] The computer-readable storage medium, the main feature of which is that a computer program is stored thereon, and the computer program can be executed by a processor to implement each step of the above-mentioned technology service resource classification method based on an evolutionary neural network.

[0071] The technology service resource data processing method, system and its computer-readable storage medium based on the optimization algorithm of the present invention have the following beneficial effects:

[0072] (1) The present invention extracts technology service resource texts from different industries for training and testing, and uses the new meta-heuristic optimization method BTGO to solve the problem of slow convergence speed of traditional neural networks, effectively improving the model training speed and quality.

[0073] (2) Using the neural network evolution method of the BTGO algorithm, effectively optimize the network parameters to improve the resource classification accuracy and enhance the operation efficiency of machine learning algorithms.

[0074] (3) The present invention is simple to implement and easy to expand, and this method can be conveniently extended to other resource aggregation problems in text classification. Description of the Drawings

[0075] Figure 1 It is a structural diagram of the technology service resource classification system based on an evolutionary neural network of the present invention.

[0076] Figure 2This is the flowchart of step 2 of the technology service resource classification method based on the evolutionary neural network of the present invention.

[0077] Figure 3 This is the flowchart of the initialization stage of step 3 of the technology service resource classification method based on the evolutionary neural network of the present invention.

[0078] Figure 4 This is the complete flowchart of step 3 of the technology service resource classification method based on the evolutionary neural network of the present invention for network optimization processing. Detailed implementation manners

[0079] To more clearly illustrate the technical solutions and advantages of the present invention, the following part will be described in detail in combination with corresponding examples and diagrams. It should be noted that the specific examples described here are only used to explain the present invention and are not used to limit the present invention.

[0080] The system structure diagram of the present invention is as Figure 1 shown. The following will be further described in combination with preferred examples and drawings to clarify the application of the evolutionary neural network method based on BTGO in the classification of technology service resources:

[0081] The technology service resource classification method based on the evolutionary neural network specifically includes the following steps:

[0082] Step 1: S1 dataset construction module, extract different types of technology service resource data, classify each data according to the technology service industry and add labels to construct a classification training and test dataset;

[0083] (1.1) Extract metadata of different industries from the technology service resource library in actual operation, including types such as talent information, intellectual property, and patented technologies;

[0084] (1.2) Classify each data according to the technology service industry and add labels, and divide the training and test datasets according to a ratio as the original sample data;

[0085] Step 2: S2 data preprocessing module, perform data preprocessing on the original dataset, that is, clean missing, incorrect, and redundant data, perform Chinese word segmentation, remove stop words, and text vectorization processing to obtain the feature vectors and vocabulary of the resource data. The flowchart of the S2 module is as Figure 2 shown.

[0086] (2.1) Perform data cleaning on the original dataset in step 1 above, clear the data rows with missing resource names or resource description information, and at the same time delete the data rows with redundant items and incorrect labels;

[0087] (2.2) Perform Chinese word segmentation and stop word removal on the cleaned dataset, and use the Word2Vec model pre-trained on an extremely large-scale scientific and technological text dataset to vectorize the text and convert it into dense text feature vectors;

[0088] (2.3) Organize all text data to obtain the feature vectors and One-hot encoded labels of the dataset for subsequent model evolution and evaluation;

[0089] Step 3: S3 network evolution module. Based on the Banyan Tree Growth Optimization (BTGO) algorithm and ELM, construct an evolutionary neural network, and simultaneously optimize the random parameters and output layer parameters of the neural network to obtain the best network model for resource classification. The flow chart of the initialization stage of the S3 module and the complete flow chart of network evolution are shown in Figure 3 and Figure 4 respectively.

[0090] (3.1) Initialize the parameters of the Banyan Tree Growth Optimization BTGO algorithm and the parameters of the ElM network. Set the population size NP, the maximum number of iterations Gmax, the number of banyan tree branches m, and the growth stage threshold T growth , and the regeneration stage threshold T recover ; set the number of hidden layer nodes of the ELM network;

[0091] (3.2) Set the fitness function of the optimization algorithm, that is, the classification error rate of the neural network determined by x i on the test set, which can be expressed as follows:

[0092]

[0093] where TP and TN are the numbers of correctly predicting positive classes as positive classes and correctly predicting negative classes as negative classes respectively, and FP and FN are the numbers of incorrectly predicting negative classes as positive classes and incorrectly predicting positive classes as negative classes respectively.

[0094] (3.3) BTGO adopts a real-number coding framework. Each branch individual represents a set of neural network parameters W, which consists of the hidden layer weights W hidden and the output layer weights W output . The i-th individual can be expressed as:

[0095] x i = [W hidden ; W output

[0096] = [x i，1 , x i，2 , …, x i，h , x i，h+1 , …, xi，D (2)

[0097] Among them, h is the sum of the weights and biases of the neural network hidden layer, and D is the sum of all the weights;

[0098] (3.4) Randomly initialize 80% of the individuals in the BTGO population. The set of serial numbers of this part of the individuals is denoted as I (i ∈ I), that is, the values of all dimensions of the individual x i (i ∈ I) are random values within the range of [-1, 1], and its calculation is shown in Equation (3):

[0099] x ij = x min，j + r × (x max，j - x min，j ) (3)

[0100] Among them, x min，j and x max，j respectively represent the minimum and maximum values of the solution space, and r is a random value within the range of [0, 1].

[0101] (3.5) For the remaining 20% of the individuals in step (3.4), the set of their serial numbers is denoted as I * , indicating that the positions of the hidden layer weight part are randomly generated, and the output layer weight part is calculated by the generalized inverse (Moore - Penrose Pseudoinverse, MPP) formula of ELM. That is, for the individual x i (i ∈ I * ), the first h bits are calculated by Equation (3), and the calculation of the remaining bits is shown in Equation (4), and the complete calculation is shown in Equation (5).

[0102]

[0103]

[0104] Among them, H is the hidden layer output matrix, X is the input matrix, β is the output layer weight, and β m represents the m - th bit of the output layer weight β calculated by the generalized inverse;

[0105] (3.6) Apply the parameters x i saved in the individuals of the initial population to construct the network model o i , and predict the labels of the test set

[0106]

[0107] Among them, g() is the activation function of the neural network, generally the sigmoid function, and compare the predicted labels Calculate the classification error rate with the true label y, and calculate the fitness value of the individual according to Equation (1) in step (3.2).

[0108] (3.7) Initialize the position p of the BTGO aerial roots root And take it as the individual's historical optimal position. Divide the branch groups (subgroups) according to the fitness ranking of the individuals in the population, and divide them into m subgroups in ascending order. The individual with the best fitness in each branch group is used as the optimal position of the subgroup, denoted as P trunk ;

[0109] (3.8) The individual with the smallest fitness value in all subgroups is regarded as the current best position p best , and update the fitness value of the best position;

[0110] (3.9) Denote count i as the number of times the individual i has not been updated. If the fitness value of the current individual i is smaller than that of its aerial root individual , then set the value of count i to 0, otherwise increment its value by 1.

[0111] (3.10) Execute different update operators of BTGO according to the value of count i to generate new individuals of the banyan population. Among them, for the first h bits of the individual i∈I * , they are updated by the update operator, and the remaining bits are calculated by Equation (5): When the number of times the individual i has not been updated counti is lower than T growth , with a probability of pr, choose to execute the aerial root operator or the multi-trunk operator; when it is greater than T growth and less than T recover , execute the adjustment operator to generate a new population.

[0112] (3.11) Execute the rooting operator or the Moore-Penrose pseudoinverse (MPP) to update the individual: When the random probability of the individual is less than pr, execute the rooting operator to update as follows:

[0113]

[0114] where r is a random number in the range [-1, 1], and F is the scaling factor. Update the partial positions of the individuals that meet the conditions in (3.9) using MPP.

[0115] (3.12) Execute the multi-trunk operator or update in a partial MPP manner: When the random probability of the individual i is greater than pr, execute the multi-trunk operator, and its update method is as shown in Equation (8):

[0116]

[0117] where r is a random number in the range [-1, 1], and F is the scaling factor, is the best position in the trunk group where the i-th individual is located. Similarly, for individuals that meet the conditions, MPP is used to update some positions.

[0118] (3.13) Execute the adjustment operator or update the partial position in MPP mode: when the individual counti accumulation is higher than T growth And less than T recover When BTGO executes the adjustment operator, it randomly selects a subgroup C from the m trunk subgroups. r , individual i moves closer to the optimal position of the subgroup, which is expressed as follows:

[0119]

[0120] Here, s is the adjustment factor and is set to a fixed value.

[0121] (3.14) Execute the rebirth operator or update the individual partial position in MPP mode: when the individual counti is greater than T recover When , individual i turns to execute the rebirth operator, and the position of the individual is randomly initialized again, and its update is as follows:

[0122] x ij =x min，j +r×(x max，j -x min，j ) (10)

[0123] If the individual i∈I * , then the first h bits are reinitialized by equation (10), and the remaining bits are calculated by equation (4).

[0124] (3.15) Calculate the new population individual x obtained by the above update operator i Build the corresponding network model o i , predict the label and calculate the fitness value according to the steps (3.5);

[0125] (3.16) Based on the spatial distance between individuals and each trunk population, they are re-divided into the new trunk subgroup with the closest distance. The distance and the new subgroup number are calculated as follows:

[0126]

[0127]

[0128] (3.17) Compare x i and The fitness value of x i With a smaller fitness, the corresponding aerial root position is updated is x iposition, and set the count of the number of times the individual has not been updated to zero, otherwise increment the count by 1. i Clear it to zero, otherwise increment the count by 1.

[0129] (3.18) Update the best positions P of the m sub - populations trunk and fitness values, and update the optimal individual P of the entire population best and its fitness.

[0130] (3.19) Determine whether the method has reached the termination condition. If the iteration reaches the maximum number of iterations, return to step (3.8) and continue to execute. Otherwise, output the optimal individual in the current dataset and the classification accuracy of the corresponding classifier.

[0131] Step 4: S4 Resource Classification Module. Use the best parameters obtained in the above steps to construct a neural network model o best , and the model parameters are composed of the parameters represented by the optimal individual P best as follows:

[0132]

[0133] where X is the input vector, and are the determined best weights of the hidden layer and output layer of the neural network respectively. Select the text data in the science and technology service resource test set for model evaluation to obtain the accuracy of the final model. Execute the following formula (13) to automatically classify the unknown data X and obtain the prediction result Y:

[0134]

[0135] The science and technology service resource classification system based on an evolutionary neural network for implementing the above - described method, wherein the system includes:

[0136] A dataset construction module for extracting science and technology service resource data of different types, classifying each data according to the classification of the science and technology service industry and adding labels, and constructing a classification training and test dataset;

[0137] A data pre - processing module connected to the dataset construction module for clearing missing, incorrect, and redundant data from the obtained dataset, performing Chinese word segmentation and stop - word removal on the text data therein, and using the Word2Vec method for text vectorization to obtain the feature vectors and word libraries of each data for subsequent model evolution and evaluation;

[0138] A network evolution module, connected to the data preprocessing module, is used to construct an evolutionary neural network based on the banyan tree growth optimization BTGO algorithm and ELM. Meanwhile, it optimizes the random parameters and output layer parameters of the neural network to improve the network classification accuracy and enhance the convergence speed, so as to obtain the best network model for resource classification; and

[0139] A resource classification module, connected to the network evolution module, is used to evaluate the model for the text data of the science and technology service resource test set based on the best network model, and perform automated resource classification on unknown data.

[0140] The computer-readable storage medium, on which a computer program is stored, and the computer program can be executed by a processor to implement each step of the above-mentioned science and technology service resource classification method based on an evolutionary neural network.

[0141] Adopting the science and technology service resource data processing method, system and its computer-readable storage medium based on the optimization algorithm of the present invention has the following beneficial effects:

[0142] (1) The present invention extracts and trains and tests the text of science and technology service resources in different industries, and uses the new meta-heuristic optimization method BTGO to solve the problem of slow convergence speed of traditional neural networks, effectively improving the model training speed and quality.

[0143] (2) Using the neural network evolution method of the BTGO algorithm can effectively optimize network parameters to improve resource classification accuracy and enhance the operation efficiency of machine learning algorithms.

[0144] (3) The present invention is simple to implement and easy to expand, and this method can be conveniently extended to other resource aggregation problems in text classification.

[0145] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A classification method for scientific and technological service resources based on an evolutionary neural network, characterized in that, The method described above includes the following steps: (1) Extract data of different types of technology service resources, classify each piece of data according to the classification of the technology service industry and add labels, and construct a classification training and test data set; (2) Perform Chinese word segmentation and stop word removal on the text data of the data set, and use the Word2Vec method for text vectorization to obtain the feature vectors and word libraries of each data, which are used for subsequent model evolution and evaluation; (3) Based on the Banyan Tree Growth Optimization BTGO algorithm and ELM, construct an evolutionary neural network, and at the same time optimize the random parameters and output layer parameters of the neural network to improve the network classification accuracy and enhance the convergence speed, and obtain the best network model for resource classification; (4) Use the final neural network model obtained above to evaluate the text data of the test data set of the technology service resource data, and perform automatic resource classification on unknown data.

2. The classification method of technology service resources based on an evolutionary neural network according to claim 1, wherein In the step (1), the types of resource data extracted include talent information, intellectual property, patent technology, and different industry metadata regularly extracted from the technology service resource library in actual operation. The training, validation, and test sets are divided according to a certain proportion to meet the actual resource integration needs.

3. The method for classifying science and technology service resources based on an evolutionary neural network according to claim 2, wherein In the step (2), after data cleaning, Chinese word segmentation, and stop word removal on the data set constructed in the step (1), use the Word2Vec model pre-trained on a large batch of technology text data sets to perform text vectorization processing on it, convert it into dense feature vectors, and obtain the feature vectors and One-hot encoded labels of the data set by sorting all text data, which are used for subsequent model evolution and evaluation.

4. The method for classifying technology service resources based on an evolutionary neural network according to claim 3, wherein The step (3) is specifically as follows: Apply the banyan tree growth optimization algorithm to encode and optimize all parameters of the ELM network, and use the global optimization strategy of BTGO to evolve the random parameters and output layer parameters at the same time to achieve the optimal prediction accuracy and training speed.

5. The classification method of science and technology service resources based on an evolutionary neural network according to claim 4, wherein The initialization optimization process of the step (3) specifically includes the following steps: (3.1) Initialize the parameters of the Banyan Tree Growth Optimization BTGO algorithm and the parameters of the ElM network, and set the population size NP, the maximum number of iterations Gmax, the number of banyan tree branches m, and the growth stage threshold T of BTGO growth , and the regeneration stage threshold T recover ; Set the number of hidden layer nodes of the ELM network; (3.2) Set the fitness function of the optimization algorithm, that is, the classification error rate of the neural network determined by x i on the test set, which is expressed by the following formula: Among them, TP and TN are the number of positive classes predicted as positive classes and the number of negative classes predicted as negative classes respectively, and FP and FN are the number of negative classes predicted as positive classes and the number of positive classes predicted as negative classes respectively; (3.3) BTGO adopts a real-number coding framework, and each branch and leaf individual represents a set of neural network parameters W, which consists of the hidden layer weights W hidden and the output layer weights W output The two parts together. The i-th individual can be represented as: x i = [W hidden ; W output ​ = [x i,1 , x i,2 ,..., x i,h , x i,h+1 ,..., x i,D (2) Among them, h is the sum of the weights and biases of the hidden layer of the neural network, and D is the sum of all weights; (3.4) Randomly initialize 80% of the individuals in the BTGO population. The set of serial numbers of this part of the individuals is denoted as I (i ∈ I), that is, for individual x i (i ∈ I), the values on all dimensions are random values within the range of [-1, 1], and its calculation is shown in Equation (3): x ij = x min,j + r × (x max,j - x min,j ) (3) where x min,j and x max,j represent the minimum and maximum values of the solution space respectively, and r is a random value in the range of [0, 1]; (3.5) For the remaining 20% of the individuals in step (3.4), the set of their serial numbers is denoted as I * , indicating that the positions of the hidden layer weight part are randomly generated, and the output layer weight part is calculated by the generalized inverse MPP formula of ELM, that is, for individual x i (i ∈ I * ), the first h bits are calculated by formula (3), and the calculation of the remaining bits is shown in formula (4), and the complete calculation is shown in formula (5): where H is the output matrix of the hidden layer, X is the input matrix, β is the weight of the output layer, and β m represents the m-th bit of the weight β of the output layer obtained by the generalized inverse calculation; (3.6) Construct a network model o with the parameter x saved by the individuals in the initial population. i Build a network model o i and predict the labels of the test set where g() is the activation function of the neural network, specifically the sigmoid function, to compare the predicted label and the true label y to calculate the classification error rate, and calculate the fitness value of the individual according to Equation (1) in step (3.2); (3.7) Initialize the position P of the BTGO aerial roots root And take it as the individual historical optimal position. Divide the branches into m subgroups according to the fitness ranking of individuals in the population, and divide them into m subgroups in ascending order. The individual with the best fitness in each branch group is used as the optimal position of the subgroup, denoted as P trunk ; (3.8) The individual with the smallest adaptability value among all subgroups is regarded as the current best position P best , and update the fitness value of the best position.

6. The method for classifying science and technology service resources based on an evolutionary neural network according to claim 5, wherein The complete network model optimization process of the step (3) specifically includes the following steps: (3.9) Record count i as the number of times individual i has not been updated. If the fitness value of the current individual i is smaller than that of its aerial root individual , then set the value of count i to 0, otherwise increment its value by 1; (3.10) Perform different BTGO update operators according to the value of count i to generate new individuals of the banyan tree population. For individual i ∈ I * , the first h bits are updated by the update operator, and the remaining bits are calculated by Equation (5): the number of times individual i has not been updated, count i is lower than T growth . When it is lower than T, with a probability of pr, either the aerial root operator or the multi-trunk operator is selected for execution; when it is greater than T growth is less than T recover , the adjustment operator is executed to generate a new population; (3.11) Execute the rooting operator or partial generalized inverse to update the individual: When the random probability of the individual is less than pr, execute the rooting operator to update as follows: Among them, r is a random number between [-1, 1], and F is a scaling factor. When the conditions in the above step (3.10) are met, the MPP update is performed on some positions of the individual; (3.12) Execute the multi-trunk operator or partial MPP method to update: When the random probability of individual i is greater than pr, execute the multi-trunk operator, and its update method is as shown in Equation (8): where r is a random number within the range of [-1, 1], F is the scaling factor, is the best position in the trunk population where the i-th individual is located. Similarly, for individuals that meet the conditions, MPP is used to update part of their positions; (3.13) Execute the adjustment operator or update some positions in MPP mode: When the cumulative count of an individual i is higher than T growth and lower than T recover BTGO executes the adjustment operator, that is, randomly selects a subgroup c from m trunk subgroups r , and individual i approaches the optimal position of this subgroup, which is expressed as follows: Among them, s is a regulation factor and is set to a fixed value; (3.14) Execute the regeneration operator or update the partial position of the individual in MPP mode: When the cumulative count i of individual is greater than T recover , individual i turns to execute the regeneration operator, and at this time the position of this individual is randomly re-initialized, and its update is as follows: x ij = x min,j + r × (x max,j - x min,j )(10) If the individual \(i\in I\) * , the first \(h\) bits are re-initialized by Equation (10), and the remaining bits are calculated by Equation (4); (3.15) Calculate the new population individual x obtained by the above update operator i Construct the corresponding network model o i , and predict the label according to step (3.5) and calculate the fitness value; (3.16) According to the spatial distance between the individual and each trunk population, re-divide it into the new trunk subgroup with the closest distance, and calculate its distance and the new subgroup serial number as follows: (3.17) Compare x i with 's fitness value. If x i has a smaller fitness, update the corresponding position of the aerial root to the position of x i , and clear the number of times count i of this individual to 0. Otherwise, increment the count by 1; (3.18) Update the best positions P of m sub - populations trunk and fitness values, and update the optimal individual P of the entire population best and its fitness; (3.19) Determine whether the termination condition is reached currently. If the maximum number of iterations is reached, return to step (3.9) and continue to execute. Otherwise, output the optimal individual in the current dataset and the classification accuracy of its corresponding classifier.

7. The classification method of technology service resources based on an evolutionary neural network according to claim 6, wherein In step (4), use the neural network model in step (3) to classify the unlabeled and unclassified new resource data, and evaluate the model and integrate and optimize the resources in combination with the test dataset.

8. The classification method of technology service resources based on an evolutionary neural network according to claim 7, wherein The specific step (4) is as follows: Construct a neural network model using the optimal parameters obtained in step (3). best , and the parameters of the neural network model are composed of the parameters represented by the optimal individual P best , which are represented as follows: where X is the input vector, and are the determined optimal neural network hidden layer weights and output layer weights, respectively; Select the text data in the science and technology service resource test set for model evaluation to obtain the accuracy of the final model. Execute the following formula (13) to automatically classify the unknown data X and obtain the prediction result Y:

9. A technology service resource classification system based on an evolutionary neural network for implementing the method according to any one of claims 1 to 8, characterized in that, The system includes: A dataset construction module, which is used to extract different types of science and technology service resource data, classify each data according to the science and technology service industry classification, add labels, and construct a classification training and test dataset; A data preprocessing module, connected to the dataset construction module, which is used to process the obtained dataset to remove missing, incorrect, and redundant data, perform Chinese word segmentation and stop word removal on the text data therein, and use the Word2Vec method for text vectorization to obtain the feature vectors and vocabulary of each data for subsequent model evolution and evaluation; A network evolution module, connected to the data preprocessing module, which is used to construct an evolutionary neural network based on the banyan tree growth optimization BTGO algorithm and ELM, and optimize the random parameters and output layer parameters of the neural network simultaneously to improve the network classification accuracy and enhance the convergence speed, and obtain the best network model for resource classification; and A resource classification module, connected to the network evolution module, which is used to evaluate the model of the text data in the science and technology service resource test set based on the best network model, and perform automatic resource classification on the unknown data.

10. A computer-readable storage medium, characterized in that, It stores a computer program, and the computer program can be executed by a processor to implement each step of the method for classifying science and technology service resources based on an evolutionary neural network described in any one of claims 1 to 8.

Images