A graph model mobile application classification method based on word frequency of description text

Abstract

The present invention proposes a graph model mobile application classification method based on the frequency of descriptive text, which uses a word segmentation tool to segment the description text of all applications into sentences and perform word segmentation on each sentence, and filter stop words and low-frequency words on the word segmentation results. Record all the single phrases that have appeared and the categories to which the applications belong; take the word segmentation results of each text in the training corpus, take categories, words and applications as nodes, and use the relevant weight ratios between words and categories, words and applications, words and words, and categories and words as corresponding edge weights to construct an undirected graph; use a two-layer graph convolution model algorithm to implement iterative operations on node weight vectors, and use a softmax activation function to complete the classification and prediction of mobile applications. The present invention can improve the accuracy of mobile application classification and is more helpful for mobile application stores to provide services to users.

Description

Technical Field

[0001] The present invention relates to a mobile application classification method, in particular to a mobile application classification method based on a graph model describing text word frequency. Background Art

[0002] With the increasing popularity of mobile devices, the number of mobile applications has exploded. In order to facilitate users to download and use, various mobile application stores have emerged on the Internet, such as Wandoujia, Xiaomi App Market, Tencent Store in China, and App Store, Google Play, etc. in foreign countries. These application stores provide consumers with mobile application downloads and corresponding follow-up services in two main ways: (1) Users search by entering keywords, and the application store finds and returns relevant mobile applications based on the keywords; (2) The application store displays mobile applications that may attract users in the form of homepage recommendations based on the user's historical browsing and download records. Both methods rely on the pre-classification of mobile applications. A good classification system and accurate classification results of applications can greatly improve the efficiency of search services and recommendation services. By locating user needs to certain specific application classification clusters and further accurately selecting mobile applications from them, the user experience can be effectively improved.

[0003] There have been some research results on the classification of mobile applications. They mainly transform the classification of mobile applications into a text classification problem. Most of their classification methods rely on user comments on applications (such as text feedback and star ratings), application names, description information, etc., and use text processing, topic models, machine learning and other methods to extract mobile application topics and further complete the classification of mobile applications. These methods have achieved certain results, but there are also some problems that have not been considered. (1) User comments are often disordered and chaotic, containing spam comments or just repeated star ratings, which makes it difficult to filter and process useful information from them; (2) The quality of mobile application description information varies, and the text length span is large. Previous text representation technologies are often not enough to accurately represent the text content of mobile applications; (3) In the mobile application description text, not all words contribute equally to the classification of mobile applications. Summary of the invention

[0004] The purpose of the present invention is to propose a mobile application classification method based on a graph model describing the word frequency of text.

[0005] The technical solution to achieve the purpose of the present invention is: a mobile application classification method based on a graph model describing the word frequency of text, and the specific steps are as follows:

[0006] Step 1: Use the word segmentation tool to split the description text of all applications into sentences and perform word segmentation on each sentence. Filter the word segmentation results for stop words and low-frequency words. Record all the single phrases that appear and the categories to which the applications belong.

[0007] Step 2: Take the word segmentation results of each text in the training corpus, take categories, words and applications as nodes, and take the correlation weight ratios between words and categories, words and applications, words and words, and categories and words as corresponding edge weights to construct an undirected graph;

[0008] Step 3: Use a two-layer graph convolutional model algorithm to implement iterative operations on node weight vectors, and use the softmax activation function to complete the classification and prediction of mobile applications.

[0009] Furthermore, in step 1, the word segmentation tool is used to segment the description texts of all applications into sentences, and each sentence is segmented. The stop words and low-frequency words are filtered out of the word segmentation results, and all the single phrases that have appeared and the categories to which the applications belong are recorded. The specific process is as follows:

[0010] Step 1.1: Corpus segmentation and low-frequency word statistics: Use the LTP tool of Harbin Institute of Technology to separate the mobile application description texts to be classified into sentences, further perform word segmentation and part-of-speech tagging on the sentence-segmented results, and remove the stop words contained in the stop word list published by Baidu Natural Language Laboratory;

[0011] Step 1.2: Filter low-frequency words: Count the global frequency and the frequency of each word in the word segmentation results in a single category, remove the words that appear too few times in proportion, save the results again, and calculate the length of each training corpus after filtering, remove the corpus with shorter length, and thus obtain the text corpus that can fully support the feature requirements of model training;

[0012] Step 1.3: Category labeling: Record all the individual phrases that have appeared and the categories to which they belong. The results of the segmentation of each training corpus are separated by spaces and saved as a json file with its category label.

[0013] Furthermore, in step 2, the word segmentation results of each text in the training corpus are taken, and the category C, word W and application D are used as nodes, and the correlation weight ratios between words and categories, words and applications, words and words, and categories and words are used as corresponding edge weights to construct an undirected graph. The specific method is:

[0014] Step 2.1, combined with the TF-KAI algorithm, calculate the weight ratio between each word and each category and application according to the frequency of the word in the category and the frequency of the word in the application corpus, and use it as the edge weight of the corresponding node in the undirected graph;

[0015] (1) For the edge weight calculation between word nodes and category nodes, a calculation method similar to TF-IDF is used to measure the correspondence between words and each category. The formula is as follows:

[0016]

[0017] in, represents the edge weight between word node i and category node c, and n(i,c) represents the number of times word i appears in the application text marked as category c, while n(c) represents the total number of words in all application texts under category c, n represents the total number of categories, and freq(i,c) represents the number of categories in which word i appears in the text;

[0018] (2) For the edge weight calculation between word nodes and application nodes, the cross entropy calculation method is used, and the formula is as follows:

[0019]

[0020]

[0021] Among them, E(i) represents the cross entropy of word i, represents the probability of word i appearing in category c, and n(i,all) represents the number of times word i appears in the description text of all categories;

[0022] Normalizing the cross entropy of word i, we get:

[0023]

[0024] Among them, NE(i) represents the normalized entropy value, E max represents the maximum value in E(i), and E min Indicates the minimum value;

[0025] Combining the normalized entropy value with the TF-KAI algorithm, we get:

[0026]

[0027] in, represents the edge weight between word node i and application node d, n(i,d) represents the number of times word i appears in the description text of application d, and n(i,D) represents the number of applications containing word i in the description text;

[0028] Step 2.2, calculate the weight between the application and the category, and use it as the edge weight between the application node and the category node in the undirected graph. That is, for the edge weight between the application and the category, add the weight ratio of the words contained in the description text of each mobile application in the training corpus to each category. The formula is as follows:

[0029]

[0030] in, represents the edge weight between application node d and category node c, t i represents the set of words i in application d;

[0031] Step 2.3, use the PMI method to measure the connection information between words, and use it as the edge weight PMI (i, j) between word nodes in the undirected graph. That is, for the edge weight between word nodes, use the PMI method to measure the related information between words. The formula is as follows:

[0032]

[0033] In the formula, PMI(i,j) represents the edge weight between word nodes, p(i,j) is the frequency of word i and word j appearing at the same time, p(i) is the frequency of word i appearing, and p(j) is the frequency of word j appearing. The calculation formula is:

[0034]

[0035]

[0036] Among them, #W(i) represents the number of sliding windows containing word i in the corpus, #W(i,j) represents the number of sliding windows containing both word i and word j in the corpus, and #W represents the total number of sliding windows in the corpus. The size of the sliding window can be adjusted according to the length of the corpus.

[0037] Step 2.4, combining the above edge weight information, we get the adjacency matrix A, the formula is as follows:

[0038]

[0039] Step 2.5: In order to alleviate the inherent problems of the graph model such as model non-convergence, a self-loop structure is added to determine the final adjacency matrix A′=A+I.

[0040] Furthermore, in step 3, a two-layer graph convolution model algorithm is used to implement the iterative operation of the node weight vector, and a softmax activation function is used to complete the classification and prediction of mobile applications. The specific method is as follows:

[0041] Step 3.1, build a two-layer graph convolution model to implement iterative operations on node weight vectors;

[0042] As a multi-layer neural network, the graph convolution network (GCN) runs directly on the constructed undirected heterogeneous graph and continuously updates the node vector according to the properties and contact information of the node neighborhood. The update formula is as follows:

[0043]

[0044] Among them, H (l) represents the hidden layer vector of the node in layer l; W (l) represents the weight matrix of the lth layer. The weight matrix is ​​an N-order matrix randomly generated at the beginning of training; ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x); Represents the normalized adjacency matrix of the adjacency matrix A′, and the calculation formula is:

[0045]

[0046] Where P is the degree matrix of the adjacency matrix A′, where the main diagonal elements of the matrix P are The remaining elements are 0, P ii represents the element in the i-th row and i-th column of the P matrix, A′ ij Represents the element in the i-th row and j-th column of the A′ matrix; calculates the normalized adjacency matrix of A′ After that, randomly generate the weight matrix W (0) And share the weight parameters in multiple layers, namely W (1) =W (0) , and randomly generate corresponding feature vectors for all nodes, using Represents the feature vector of node v, where m is the dimension of the feature vector, and its matrix representation is Let the hidden layer vector H of the first layer be (0) =X, since the parameters of the weight matrix are shared in the process layer, the training process only changes the feature vector corresponding to each node, that is, the hidden layer vector parameters are gradually iterated as the training progresses;

[0047] Step 3.2, after the two-layer graph convolution structure, add a fully connected layer and use the softmax activation function to complete the classification and prediction of all mobile applications. The formula is:

[0048]

[0049] in, Where K represents the number of all applications, n represents the number of all categories, represents the category distribution probability predicted by the model for mobile application d, ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x);

[0050] Step 3.3, construct a training loss function, train the model in a supervised learning manner, calculate the loss function and regress it until the number of iterations reaches a preset upper limit or the model error is less than a certain threshold.

[0051] Furthermore, the loss function is as follows:

[0052]

[0053] Among them, D train represents the training data set; Y ij ∈{0,1} is used as an indicator variable to represent the label of mobile application i corresponding to category j. If the true label of application i is j, then Y ij =1, otherwise Y ij =0; Z ij is the distribution of mobile application categories predicted by the model, representing the probability that mobile application i is predicted as category j in the model prediction result, corresponding to the element in the i-th row and j-th column in the Z matrix; η represents the custom L2 regularization parameter; Θ represents the model parameter. Here, due to the weight matrix W (l) and the undirected graph node edge weight matrix are relatively fixed, and the main changing model parameters are the characteristic vectors of each node, namely H (l) .

[0054] A graph model mobile application classification system based on description text word frequency performs graph model mobile application classification based on the above method.

[0055] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above method is implemented to classify graph model mobile applications.

[0056] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above method for classifying graph model mobile applications.

[0057] Compared with the prior art, the present invention has the following significant advantages: based on the Text-GCN model, more types of node information are introduced in the construction of the undirected graph, the frequency information of words in categories and application texts is deeply utilized, and the weights between nodes are set in combination with methods such as PMI and TF-KAI. Compared with other methods, the present invention can improve the accuracy of mobile application classification and is more helpful for mobile application stores to provide services to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flowchart of the text classification method based on et-gcn;

[0059] Figure 2 It is a schematic diagram of the overall network model of et-gcn;

[0060] Figure 3 It is a schematic diagram of the precision, regression rate, and F1 rate of the test data under different methods. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0062] like Figure 1 As shown in FIG. 1 , the mobile application classification method based on the graph model describing the word frequency of text (the text classification method based on et-gcn) includes the following steps:

[0063] In step 1, use the word segmentation tool to split the description text of all applications into sentences, and then segment each sentence. Then filter the stop words and low-frequency words of the word segmentation results, and record all the single phrases that have appeared and the categories to which the applications belong. The specific process is as follows:

[0064] Step 1.1: Word segmentation and low-frequency word statistics of the corpus. Use the LTP tool of Harbin Institute of Technology to separate the mobile application description texts to be classified into sentences, and further perform word segmentation and part-of-speech tagging on the sentence-segmented results. Compare with the stop word list published by Baidu Natural Language Laboratory and remove the stop words contained in it.

[0065] Step 1.2: Filter low-frequency words. Count the global frequency of each word in the word segmentation results and the frequency of occurrence in a single category. Remove words that appear too few times according to a certain ratio. Save the results again, and calculate the length of each training corpus after filtering. Remove the corpus with smaller lengths to obtain text corpus that can fully support the model training feature requirements. Then record all the single phrases that have appeared and the categories to which the application belongs. The results of each training corpus after word segmentation are separated by spaces and saved as a json file with its classification label.

[0066] In step 2, the word segmentation results of each text in the training corpus are obtained, and an undirected graph is constructed with three types of nodes: word, application, and category, such as Figure 2 As shown, the specific steps are as follows:

[0067] Construct an undirected weighted heterogeneous network graph: G = (V, E), where category C, word W and application D are nodes V, and the weight ratios of the correlation between words and categories, words and applications, words and words, and categories and words are the corresponding edge weights. Specifically, it is represented by an N-order matrix, where N is equal to the sum of the total number of all categories, words and applications, and the elements in the matrix are the weight relationships between the corresponding rows and columns, that is, the edge weights between nodes.

[0068] Step 2.1: For the edge weight calculation between word nodes and category nodes, a calculation method similar to TF-IDF is used to measure the correspondence between words and each category. The formula is as follows:

[0069]

[0070] in, represents the edge weight between word node i and category node c, n(i,c) represents the number of times word i appears in the application text marked as category c, and n(c) represents the total number of words in all application texts under category c. n represents the total number of categories, and freq(i,c) represents the number of categories in which word i appears in the text.

[0071] For the edge weight calculation between word nodes and application nodes, the cross entropy calculation method is used, and the formula is as follows:

[0072]

[0073]

[0074] Among them, E(i) represents the cross entropy of word i, represents the probability of word i appearing in category c, and n(i,all) represents the number of times word i appears in the description text of all categories.

[0075] Normalizing the cross entropy of word i, we get:

[0076]

[0077] Among them, NE(i) represents the normalized entropy value, E max represents the maximum value in E(i), and E min Indicates the minimum value;

[0078] Combining the normalized entropy value with the TF-KAI algorithm, we get:

[0079]

[0080] in, represents the edge weight between word node i and application node d, n(i,d) represents the number of times word i appears in the description text of application d, and n(i,D) represents the number of applications containing word i in the description text;

[0081] Step 2.2: For the edge weights between applications and categories, add the weights of the words contained in the description text of each mobile application in the training corpus and each category. The formula is as follows:

[0082]

[0083] in, represents the edge weight between application node d and category node c, t i represents the set of words i in application d;

[0084] Step 2.3: For the edge weights between word nodes, use the PMI method to measure the relevant information between words. The formula is as follows:

[0085]

[0086] In the formula, PMI(i,j) represents the edge weight between word nodes, p(i,j) is the frequency of word i and word j appearing at the same time, p(i) is the frequency of word i appearing, and p(j) is the frequency of word j appearing. The calculation formula is:

[0087]

[0088]

[0089] Among them, #W(i) represents the number of sliding windows containing word i in the corpus, #W(i,j) represents the number of sliding windows containing both word i and word j in the corpus, and #W represents the total number of sliding windows in the corpus. The size of the sliding window can be adjusted according to the length of the corpus.

[0090] Step 2.4, combining the above edge weight information, construct a complete undirected graph structure, and obtain the N-order adjacency matrix A, the formula is as follows:

[0091]

[0092] Step 2.5: In order to alleviate the inherent problems of the graph model such as model non-convergence, a self-loop structure is added to determine the final adjacency matrix A′=A+I, where I is the unit matrix.

[0093] In step 3, a two-layer graph convolution model algorithm is used to implement the iterative operation of the node weight vector, and the softmax activation function is used to complete the classification and prediction of mobile applications. The specific method is as follows:

[0094] Step 3.1, build a two-layer graph convolution model to implement iterative operations on node weight vectors;

[0095] As a multi-layer neural network, the graph convolutional network runs directly on the constructed undirected heterogeneous graph and continuously updates the node vector according to the properties and contact information of the node neighborhood. The update formula is as follows:

[0096]

[0097] Among them, H (l) represents the hidden layer vector of the node in layer l; W (l) represents the weight matrix of the lth layer. The weight matrix is ​​an N-order matrix randomly generated at the beginning of training; ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x); Represents the normalized adjacency matrix of the adjacency matrix A′, and the calculation formula is:

[0098]

[0099] Where P is the degree matrix of the adjacency matrix A′, where the main diagonal elements of the matrix P are The remaining elements are 0, P ii represents the element in the i-th row and i-th column of the P matrix, A′ ij Represents the element in the i-th row and j-th column of the A′ matrix; calculates the normalized adjacency matrix of A′ After that, randomly generate the weight matrix W (0) And share the weight parameters in multiple layers, namely W (1) =W (0) , and randomly generate corresponding feature vectors for all nodes, using Represents the feature vector of node v, where m is the dimension of the feature vector, and its matrix representation is Let the hidden layer vector H of the first layer be (0) =X, since the parameters of the weight matrix are shared in the process layer, the training process only changes the feature vector corresponding to each node, that is, the hidden layer vector parameters are gradually iterated as the training progresses;

[0100] Step 3.2, after the two-layer graph convolution structure, add a fully connected layer and use the softmax activation function to complete the classification and prediction of all mobile applications. The formula is:

[0101]

[0102] in, Where K represents the number of all applications, n represents the number of all categories, represents the category distribution probability predicted by the model for mobile application d, ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x);

[0103] Step 3.3, construct a training loss function, train the model in a supervised learning manner, calculate the loss function and regress it until the number of iterations reaches a preset upper limit or the model error is less than a certain threshold. The loss function is as follows:

[0104]

[0105] Among them, D train represents the training data set; Y ij ∈{0,1} is used as an indicator variable to represent the label of mobile application i corresponding to category j. If the true label of application i is j, then Y ij =1, otherwise Y ij =0; Z ij is the distribution of mobile application categories predicted by the model, representing the probability that mobile application i is predicted to be category j in the model prediction result, corresponding to the element in the i-th row and j-th column in the Z matrix; η represents the custom L2 regularization parameter; Θ represents the model parameter.

[0106] After the model prediction results are normalized, the type with the largest distribution probability is the result of automatic classification after model training. It is compared with the actual classification label, and the current overall loss value is calculated according to the loss function calculation formula. By continuously iterating the calculation model, the classification model continues to be trained under the current result to narrow the gap with the actual classification label.

[0107] The present invention also proposes a graph model mobile application classification system based on description text word frequency, and performs graph model mobile application classification based on the above method.

[0108] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above method is implemented to classify graph model mobile applications.

[0109] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above method for classifying graph model mobile applications.

[0110] Example

[0111] In order to better verify that the mobile application classification model et-gcn based on text-gcn proposed in this invention has greater advantages than other common text classification algorithms, experiments with multiple common machine learning classifiers are conducted using the same experimental data. The experiment selected 6 text classification methods, including Naive Bayes, decision tree (J48), multilayer perceptron, LSTM, BERT, and text-gcn, and uniformly adopted the ten-fold cross validation method for training and learning.

[0112] The final experimental results based on 6 text classifiers are as follows Figure 3 As shown in the figure. Overall, the classification of common classifiers is not ideal. The classification models of Text-GCN and BERT are relatively good, among which Text-GCN achieves the highest classification accuracy of 70.05%. The improved classification model based on et-GCN proposed in the present invention can achieve an accuracy of 75.13%, which is better than these classification models.

[0113] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all the deformations and improvements of the technical features in the above embodiments are not included, which belong to the protection scope of this application. Therefore, the protection scope of the patent of this application shall be based on the attached claims.

Claims

1. A method for classifying mobile applications based on a graph model describing the frequency of words in text, characterized in that: The specific steps are as follows: Step 1: Use the word segmentation tool to split the description text of all applications into sentences and perform word segmentation on each sentence. Perform stop word and low-frequency word filtering on the word segmentation results, and record all the single phrases that have appeared and the category to which the application belongs. Step 2: Take the word segmentation results of each text in the training corpus, take categories, words and applications as nodes, and take the correlation weight ratios between words and categories, words and applications, words and words, and categories and words as corresponding edge weights to construct an undirected graph; Step 3: Use a two-layer graph convolutional model algorithm to implement iterative operations on node weight vectors, and use a softmax activation function to complete the classification and prediction of mobile applications. In step 2, take the word segmentation results of each text in the training corpus, take category C, word W and application D as nodes, take the correlation weight ratios between word and category, word and application, word and word, category and word as corresponding edge weights, and construct an undirected graph. The specific method is: Step 2.1, combined with the TF-KAI algorithm, calculate the weight ratio between each word and each category and application according to the frequency of the word in the category and the frequency of the word in the application corpus, and use it as the edge weight of the corresponding node in the undirected graph; (1) For the edge weight calculation between word nodes and category nodes, a calculation method similar to TF-IDF is used to measure the correspondence between words and each category. The formula is as follows: in, represents the edge weight between word node i and category node c, and n(i,c) represents the number of times word i appears in the application text marked as category c, while n(c) represents the total number of words in all application texts under category c, n represents the total number of categories, and freq(i,c) represents the number of categories in which word i appears in the text; (2) For the edge weight calculation between word nodes and application nodes, the cross entropy calculation method is used, and the formula is as follows: Among them, E(i) represents the cross entropy of word i, represents the probability of word i appearing in category c, and n(i,all) represents the number of times word i appears in the description text of all categories; Normalizing the cross entropy of word i, we get: Among them, NE(i) represents the normalized entropy value, E max represents the maximum value in E(i), and E min Indicates the minimum value; Combining the normalized entropy value with the TF-KAI algorithm, we get: in, represents the edge weight between word node i and application node d, n(i,d) represents the number of times word i appears in the description text of application d, and n(i,D) represents the number of applications containing word i in the description text; Step 2.2, calculate the weight between the application and the category, and use it as the edge weight between the application node and the category node in the undirected graph. That is, for the edge weight between the application and the category, add the weight ratio of the words contained in the description text of each mobile application in the training corpus to each category. The formula is as follows: in, represents the edge weight between application node d and category node c, t i represents the set of word i in application d; Step 2.3, use the PMI method to measure the connection information between words, and use it as the edge weight PMI (i, j) between word nodes in the undirected graph. That is, for the edge weight between word nodes, use the PMI method to measure the related information between words. The formula is as follows: In the formula, PMI(i,j) represents the edge weight between word nodes, p(i,j) is the frequency of word i and word j appearing at the same time, p(i) is the frequency of word i appearing, and p(j) is the frequency of word j appearing. The calculation formula is: Among them, #W(i) represents the number of sliding windows containing word i in the corpus, #W(i,j) represents the number of sliding windows containing both word i and word j in the corpus, and #W represents the total number of sliding windows in the corpus. The size of the sliding window can be adjusted according to the length of the corpus. Step 2.4, combining the above edge weight information, we get the adjacency matrix A, the formula is as follows: Step 2.5: In order to alleviate the inherent problems of the graph model such as model non-convergence, a self-loop structure is added to determine the final adjacency matrix A′=A+I.

2. The method for classifying mobile applications using a graph model based on the frequency of descriptive text words according to claim 1, characterized in that: In step 1, use the word segmentation tool to split the description text of all applications into sentences, and then segment each sentence. Then filter the stop words and low-frequency words of the word segmentation results, and record all the single phrases that have appeared and the categories to which the applications belong. The specific process is as follows: Step 1.1: Corpus segmentation and low-frequency word statistics: Use the LTP tool of Harbin Institute of Technology to separate the mobile application description texts to be classified into sentences, further perform word segmentation and part-of-speech tagging on the sentence-segmented results, and remove the stop words contained in the stop word list published by Baidu Natural Language Laboratory; Step 1.2: Filter low-frequency words: Count the global frequency and the frequency of each word in the word segmentation results in a single category, remove the words that appear too few times in proportion, save the results again, and calculate the length of each training corpus after filtering, remove the corpus with shorter length, and thus obtain the text corpus that can fully support the feature requirements of model training; Step 1.3: Category labeling: Record all the individual phrases that have appeared and the categories to which they belong. The results of the segmentation of each training corpus are separated by spaces and saved as a json file with its category label.

3. The method for classifying mobile applications using a graph model based on the frequency of descriptive text words according to claim 1, characterized in that: In step 3, a two-layer graph convolutional model algorithm is used to implement the iterative operation of the node weight vector, and a softmax activation function is used to complete the classification and prediction of mobile applications. The specific method is as follows: Step 3.1, build a two-layer graph convolution model to implement iterative operations on node weight vectors; As a multi-layer neural network, the graph convolutional network runs directly on the constructed undirected heterogeneous graph and continuously updates the node vector according to the properties and contact information of the node neighborhood. The update formula is as follows: Among them, H (l) represents the hidden layer vector of the node in layer l; W (l) represents the weight matrix of the lth layer. The weight matrix is ​​an N-order matrix randomly generated at the beginning of training; ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x); Represents the normalized adjacency matrix of the adjacency matrix A′, and the calculation formula is: Where P is the degree matrix of the adjacency matrix A′, where the main diagonal elements of the matrix P are ii =∑ j A′ ij , the rest of the elements are 0, P ii represents the element in the i-th row and i-th column of the P matrix, A′ ij Represents the element in the i-th row and j-th column of the A′ matrix; calculates the normalized adjacency matrix of A′ After that, randomly generate the weight matrix W (0) And share the weight parameters in multiple layers, namely W (1) =W (0) , and randomly generate corresponding feature vectors for all nodes, using Represents the feature vector of node v, where m is the dimension of the feature vector, and its matrix representation is Let the hidden layer vector H of the first layer be (0) =X, since the parameters of the weight matrix are shared in the process layer, the training process only changes the feature vector corresponding to each node, that is, the hidden layer vector parameters are gradually iterated as the training progresses; Step 3.2, after the two-layer graph convolution structure, add a fully connected layer and use the softmax activation function to complete the classification and prediction of all mobile applications. The formula is: in, Where K represents the number of all applications, n represents the number of all categories, represents the category distribution probability predicted by the model for mobile application d, ρ represents the activation function Relu, and its formula is ρ(x)=max(0,x); Step 3.3, construct a training loss function, train the model in a supervised learning manner, calculate the loss function and regress it until the number of iterations reaches a preset upper limit or the model error is less than a certain threshold.

4. The method for classifying mobile applications using a graph model based on the frequency of descriptive text words according to claim 3 is characterized in that: The loss function is as follows: Among them, D train represents the training data set; Y ij ∈{0,1} is used as an indicator variable to represent the label of mobile application i corresponding to category j. If the true label of application i is j, then Y ij =1, otherwise Y ij =0; Z ij is the distribution of mobile application categories predicted by the model, representing the probability that mobile application i is predicted to be category j in the model prediction result, corresponding to the element in the i-th row and j-th column in the Z matrix; η represents the custom L2 regularization parameter; Θ represents the model parameter.

5. A mobile application classification system based on a graph model describing the frequency of words in text, characterized in that: Based on the method described in any one of claims 1 to 4, graph model mobile application classification is performed.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method described in any one of claims 1 to 4 is implemented to classify graph model mobile applications.

7. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method described in any one of claims 1 to 4 is implemented to perform graph model mobile application classification.

Images