Sparse data identification and recovery neural network model based on adaptive resonance theory

Abstract

The invention relates to the field of data processing, in particular to a sparse data recognition and recovery neural network model based on an adaptive resonance theory, which comprises an ARTWD network and an ARTDWD network, the ARTWD network is used for classifying sparse data with missing attributes to obtain a possible unique category of the sparse data, the ARTDWD network is used for recovering missing attribute values on the basis of the ARTWD network, and the ARTWD network is used for classifying the sparse data with the missing attributes to obtain the unique category of the sparse data with the missing attributes. The method solves the problem that the clustering precision and accuracy of a machine learning clustering algorithm are sharply reduced when the machine learning clustering algorithm faces sparse data with missing attributes, and further has the advantages of being high in operation speed, high in robustness, high in accuracy, accurate in data recovery and the like.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and specifically to a sparse data recognition and recovery neural network model based on the adaptive resonance theory. Background Art

[0002] In real life, due to physical, material, legal, moral, ethical and other reasons, the experimental data of a certain type of problem is very few and difficult to obtain. For example, the data of soil samples on other planets, personal privacy data (especially information related to minors), and relevant data of medical examinations, etc. Usually, it is quite difficult to build a model for this type of data. Due to insufficient research on the physical laws describing the data relationship in this type of system model, incomplete information on external factors affecting the data, and insufficient measurement accuracy and measurement frequency, etc., in this case, it is almost impossible to build a model using traditional analytical modeling methods. However, currently, there are several unique machine learning methods that have achieved varying degrees of success in training models with a small amount of data. For machine learning methods, to improve the accuracy of the model, expanding the sample is considered the most desirable, effective, and valuable method. However, in the case of those rare and almost unique data, new information must be carefully weighed before adding it to the data set, because even adding a sample with significantly different values from other samples is very likely to cause imbalance in the model, and then have a negative impact. Therefore, when learning such data, a program that can evaluate the correlation between the sample and the existing data should be developed, and then decide whether such a sample can be incorporated into the data set. For this reason, the present invention proposes a sparse data recognition and recovery neural network model based on the adaptive resonance theory. Summary of the Invention

[0003] The purpose of the present invention is to provide a sparse data recognition and recovery neural network model based on the adaptive resonance theory to solve the above problems. The present invention is achieved through the following technical solutions:

[0004] The present invention is a sparse data recognition and recovery neural network model based on the adaptive resonance theory, including an ART_WD network for classifying sparse data and an ART_DWD network for recovering sparse data:

[0005] The data recognition process of the ART_WD network is as follows: Step 1, the input sample X=(x1, x2,..., x m ) enters the data preprocessing module I, and in module I, the data is normalized or random noise is removed to obtain U=(u1, u2,..., u m ); Step 2, the vector U=(u1, u2,..., u m ) enters the recognition module II, and through ascending weight w ijIt is calculated that Y=(y1,y2,…,y k ), and through the formula:

[0006]

[0007] The winning neuron is selected therefrom; Step 3, after the winning neuron is calculated in Step 2, further the U=(u1,u2,…,u m ) obtained by the data preprocessing module I is calculated with the descending weight w ji corresponding to the winning neuron to obtain the similarity RN; Step 4, the similarity RN calculated by the model control module III is input into the model conversion module IV, and RN and Y=(y1,y2,…,y k ) are compared with the similarity threshold P ρ . According to the size relationship among the similarity RN, the similarity threshold P ρ and the neuron Y=(y1,y2,…,y k ), as well as the number k of the remaining neurons, the recognition result of the model is obtained, triggering the corresponding resonance unit S1, reset unit S2 and adaptation unit S3;

[0008] I. The data recovery process of the ART_DWD network is as follows. Step 1, after the input sample X=(x1,x2,…,x m ) is normalized and the random noise is removed to obtain U=(u1,u2,…,u m ), the number q of missing attributes is compared with the suitability threshold P a . If q is equal to 0, then U=(u1,u2,…,u m ) performs the above data recognition process; if q<P a , then U=(u1,u2,…,u m ) is recovered through the ART_DWD network; if q>P a , it is determined that the corresponding sample is not suitable for data recovery; Step 2, the positions of the missing attributes in U=(u1,u2,…,u m ) are set to 0, and the corresponding positions in the ascending weight w ij and the descending weight w ji are set to 0. Then the vector U=(u1,u2,…,u m ) enters the recognition module II, and Y=(y1,y2,…,y ij ) is calculated by calculating with the ascending weight w k and through the formula:

[0009]

[0010] Select the winning neuron from them; Step 3, after calculating the winning neuron in Step 2, further calculate the similarity RN by calculating the U=(u1, u2, …, u m ) and the descending weight w ji corresponding to the winning neuron; Step 4, input the similarity RN calculated by the model control module III into the model conversion module IV, and compare RN with Y=(y1, y2, …, y k ) and the similarity threshold P ρ . According to the size relationship between the similarity RN, the similarity threshold P ρ and the neuron Y=(y1, y2, …, y k ), as well as the number k of remaining neurons, obtain the recognition result of the model, and trigger the corresponding reset unit S2 and resonance unit S1; Step 5, after the sparse data with missing attributes resonates, the ART_DWD model will perform data recovery according to its belonging category.

[0011] Preferably: The normalization process of the input sample X=(x1, x2, …, x m ) in Step 1 of the ART_WD network and the ART_DWD network is specifically as follows,

[0012] Preferably: The calculation process in Step 2 of the ART_WD network and the ART_DWD network and the selection formula for the winning neuron are respectively:

[0013]

[0014] where the ascending weight w ij is the centroid vector calculated according to all the normalized vectors in the category Y j , and the formula is:

[0015] where S represents the number of vectors belonging to category j.

[0016] Preferably: In the similarity RN calculation process in Step 3 of the ART_WD network and the ART_DWD network, each input sample will be stored in the descending weight w ji of the winning neuron of the belonging category, that is:

[0017]

[0018] Belong to the corresponding category: where S is the number of vectors previously assigned to category j, and the descending weight connecting the winning neuron contains a total of S, that is, m×S. According to the descending weight w jiFirst, calculate the single impact r of the samples included in the descending weight on the input sample l , and then calculate the similarity RN by pairwise average method:

[0019]

[0020] Preferably: the trigger condition of the adaptation unit S3 in step 4 of the ART_WD network is that the value of any neuron in Y is greater than the similarity threshold P ρ Or when the winning neuron y j* The value is less than the similarity threshold P ρ And the similarity RN is greater than the similarity threshold P ρ And k = 1, the adaptation unit S3 will be triggered, and the recognition module II will create a new category Y k+1 , and neuron Y k+1 The ascending weight and descending weight of the connection are defined as:

[0021] w ik+1 =u i ,

[0022] S=1.

[0023] Preferably: the triggering conditions of the resonance unit S1 and the reset unit S2 in step 4 of the ART_WD network and the ART_DWD network are respectively the winning neuron y j* The value is less than the similarity threshold P ρ And the similarity RN is less than the similarity threshold P ρ , the resonance unit S1 will be triggered, and the recognition module II will classify the input sample into the category of the winning neuron, and update the ascending weight and descending weight. The update formula is:

[0024]

[0025] Winning neuron y j* The value is greater than the similarity threshold P ρ Or when the winning neuron y j* The value is less than the similarity threshold P ρ And the similarity RN is greater than the similarity threshold P ρ When k is greater than 1, the reset unit S2 will be triggered, and the recognition module II will reset the neuron, k = k-1, y j* will be temporarily removed from Y.

[0026] Preferably: after the input sample of the ART_DWD network resonates, the attribute value of the input sample X will be restored through the category j* to which it belongs, and the formula is:

[0027]

[0028] where S is the number of samples belonging to class j*, is the attribute value at position g of the complete vector X belonging to class j*. After data recovery, this temporary model is destroyed.

[0029] Preferably, both the ART_WD network and the ART_DWD network include a preprocessing module I; an identification module II; a control module III; a model conversion module IV, and the model conversion module receives the similarity RN calculated by the control module and the output result Y=(y1, y2,..., y k ), and compares it with the similarity threshold P ρ for comparison, and inputs it into different units according to the comparison result: a resonance unit S1; a reset unit S2; an adaptation unit S3. The ART_DWD network further includes an input detection module ① for comparing the number of missing attributes with the suitability threshold P ρ ; a data recovery module ② for the data after sparse data resonance.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] When the network model ART_WD based on the adaptive resonance theory of the present invention clusters sparse data with missing attributes, the accuracy rate is greatly improved compared with other machine learning algorithms;

[0032] The logical neuron D of the present invention can greatly reduce the amount of calculation and reduce the training time compared with the ART-2 algorithm;

[0033] The ART_DWD network model for data recovery of sparse data with missing attributes of the present invention can effectively integrate the ART_WD clustering algorithm, and the data recovery accuracy rate is very high, reaching 90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the overall algorithm flowchart of the present invention;

[0035] Figure 2 is the algorithm flowchart of the original ART-2 network model;

[0036] Figure 3 is the detailed flowchart of the similarity RN calculation;

[0037] Figure 4 is the discrimination mechanism of the logical neuron D;

[0038] Figure 5 is the confusion matrix diagram of the test results of the ART-DWD model for training super-small samples and hiding the attribute x4 in the test set in Example 2;

[0039] Figure 6 Test result graph of the ART-DWD model that hides the attribute x3 in the test set for the training of the hyper-kid sample in Example 2; where Figure 6 (a) is the difference between the true value and the predicted value, Figure 6 (b) is the error rate. Detailed implementation manners

[0040] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] Example 1

[0042] Please refer to Figures 1-4 , the present invention is a sparse data recognition and recovery neural network model based on the adaptive resonance theory, including an ART_WD network for classifying sparse data and an ART_DWD network for recovering sparse data: the data recognition process of the ART_WD network is as follows. Step 1, the input sample X=(x1, x2,..., x m ) enters the data preprocessing module I, and the data is normalized or random noise is removed in module I to obtain U=(u1, u2,..., u m ); Step 2, the vector U=(u1, u2,..., u m ) enters the recognition module II, and Y=(y1, y2,..., y ij ) is calculated by calculating with the ascending weight w k , and through the formula:

[0043]

[0044] The winning neuron is selected from among them; Step 3, after the winning neuron is calculated in Step 2, further calculate the similarity RN by calculating U=(u1, u2,..., u m ) obtained by the data preprocessing module I and the descending weight w ji corresponding to the winning neuron; Step 4, the similarity RN calculated by the model control module III is input into the model conversion module IV, and RN and Y=(y1, y2,..., y k ) are compared with the similarity threshold P ρ , and according to the similarity RN, the similarity threshold P ρ and the neuron Y=(y1, y2,..., y kBased on the size relationship of (), and the remaining number of neurons k, the recognition result of the model is obtained, triggering the corresponding resonance unit S1, reset unit S2, and adaptation unit S3.

[0045] The data recovery process of the ART_DWD network is as follows: Step 1, for the input sample X = (x1, x2,..., x m ), after normalization and removal of random noise to obtain U = (u1, u2,..., u m ), the number of missing attributes q is compared with the suitability threshold P a . If q equals 0, then U = (u1, u2,..., u m ) undergoes the above data recognition process; if q < P a , then U = (u1, u2,..., u m ) is recovered through the ART_DWD network; if q > P a , it is determined that the corresponding sample is not suitable for data recovery; Step 2, set the positions of the missing attributes in U = (u1, u2,..., u m ) to 0, and set the corresponding positions in the ascending weight w ij and the descending weight w ji to 0. Then the vector U = (u1, u2,..., u m ) enters the recognition module II, and Y = (y1, y2,..., y ij ) is calculated by multiplying with the ascending weight w k . And through the formula:

[0046]

[0047] The winning neuron is selected from them; Step 3, after calculating the winning neuron in Step 2, further calculate the similarity RN between U = (u1, u2,..., u m ) obtained from the data preprocessing module I and the descending weight w ji corresponding to the winning neuron; Step 4, the similarity RN calculated by the model control module III is input into the model conversion module IV, and RN and Y = (y1, y2,..., y k ) are compared with the similarity threshold P ρ . Based on the size relationship of the similarity RN, similarity threshold P ρ and the neuron Y = (y1, y2,..., y k ), and the remaining number of neurons k, the recognition result of the model is obtained, triggering the corresponding reset unit S2 and resonance unit S1; Step 5, when the sparse data of the missing attributes resonates, the ART_DWD model will perform data recovery according to its category.

[0048] In this embodiment: For the input sample X = (x1, x2, …, x m ) in the ART_WD network and ART_DWD network, the normalization process is specifically as follows:

[0049] In this embodiment: The calculation process and the formula for selecting the winning neuron in step 2 of the ART_WD network and ART_DWD network are respectively:

[0050]

[0051] Among them, the ascending weight w ij is the centroid vector calculated based on all the normalized vectors within the category Y j , and the formula is:

[0052] U l ∈Y j , where S represents the number of vectors belonging to category j.

[0053] In this embodiment: For the calculation process of the similarity RN in step 3 of the ART_WD network and ART_DWD network, each input sample will be stored in the descending weight w ji of the winning neuron of the category it belongs to, that is:

[0054]

[0055] Belong to the corresponding category: Among them, S is the number of vectors previously assigned to category j. The descending weights connected to the winning neuron altogether contain S, that is, m×S. According to the descending weight w ji connected to the winning neuron, first calculate the single influence r l of the sample contained in the descending weight on this input sample, and then calculate the similarity RN through the pairwise averaging method:

[0056]

[0057] In this embodiment: The triggering condition of the adaptation unit S3 in step 4 of the ART_WD network is that the value of any neuron in Y is greater than the similarity threshold P ρ or when the value of the winning neuron y j* is less than the similarity threshold P ρ and the similarity RN is greater than the similarity threshold P ρ and k = 1, then the adaptation unit S3 will be triggered, and the recognition module II will create a new category Y k+1 , and the ascending weight and descending weight connected to the neuron Y k+1 are respectively defined as:

[0058] wik+1 =u i ,

[0059] S=1.

[0060] In this embodiment: the triggering conditions of the resonance unit S1 and the reset unit S2 in step 4 of the ART_WD network and the ART_DWD network are respectively the winning neuron y j* The value is less than the similarity threshold P ρ And the similarity RN is less than the similarity threshold P ρ , the resonance unit S1 will be triggered, and the recognition module II will classify the input sample into the category of the winning neuron, and update the ascending weight and descending weight. The update formula is:

[0061]

[0062] Winning neuron y j* The value is greater than the similarity threshold P ρ Or when the winning neuron y j* The value is less than the similarity threshold P ρ And the similarity RN is greater than the similarity threshold P ρ When k is greater than 1, the reset unit S2 will be triggered, and the recognition module II will reset the neuron, k = k-1, y j* will be temporarily removed from Y.

[0063] In this embodiment: after the input sample of the ART_DWD network resonates, the attribute value of the input sample X will be restored through the category j* to which it belongs, and the formula is:

[0064]

[0065] Where S is the number of samples belonging to category j*, It is the attribute value of position g of the complete vector X belonging to category j*. After data recovery, this temporary model is destroyed.

[0066] In this embodiment, the ART_WD network and the ART_DWD network both include a preprocessing module I; an identification module II; a control module III; and a model conversion module IV, and the model conversion module receives the similarity RN calculated by the control module and the output result Y=(y1, y2, ..., y k ), and the similarity threshold P ρ The comparison is performed and input into different units according to the comparison results: resonance unit S1; reset unit S2; adaptation unit S3. The ART_DWD network also includes comparing the number of missing attributes with the suitability threshold P aInput detection module ① for comparison; data recovery module ② after sparse data resonance.

[0067] Furthermore, the similarity threshold P ρ is not a fixed value nor a hyperparameter. It is a vector calculated in advance based on the training set. Taking the iris dataset as an example, since there are three types of iris flowers in total, the similarity thresholds P for these three types need to be calculated separately in advance. ρ , and the calculation steps are as follows:

[0068] Step 1, since the labels of the training set are known, first calculate the center of each class of samples according to the category;

[0069] Step 2, calculate the maximum distance from the samples in each category to the class center;

[0070] Step 3, multiply the maximum value of each class by 1.15 as the similarity threshold P ρ ;

[0071] Step 4, because the ART_WD network model adaptively generates categories. Taking the iris dataset as an example, it is very likely to generate other categories during training, exceeding three categories. Therefore, when constructing the similarity threshold P ρ , in addition to calculating and determining the values for the known categories, the same number of thresholds as the number of known categories still need to be supplemented at the back. The supplemented thresholds take the minimum value of the known similarity threshold P ρ , so that when making a judgment in the logical neuron D, the similarity threshold P will be dynamically selected according to the category of the winning neuron. ρ .

[0072] In addition, in addition to the similarity threshold P ρ that needs attention, the reset process of the reset unit also needs to be seriously considered. In the present invention, when the logical neuron D determines that the winning neuron should be reset, the model conversion module will temporarily remove this neuron from Y, and then return to Step 2; once again select the minimum value from the remaining neurons and calculate its similarity RN; then still input it into the logical neuron D and determine again whether this neuron should be reset, resonate or adapt.

[0073] One of the most notable modules of the present invention is the control module. The calculation process of the similarity RN takes into account both the influence of a single sample of this category and the influence of the whole category on this sample. The calculated similarity RN represents the average value of the distances between the input sample and the samples within the class pairwise, and Y = (y1, y2,..., y k ) exactly represents the distance from the input sample to the center of the samples in this class. These two unique calculation methods can clearly show y j*<RN, thus effectively reducing the computational process of the logical neuron D calculating the similarity RN multiple times during the discrimination process.

[0074] Example 2

[0075] This example tested a dataset of zinc content in the blood of a child patient. The dataset contains 240 groups of samples. At the same time as testing the child patient's blood, the zinc content in the drinking water in the child patient's apartment was also detected, and their height and weight were recorded. Since children lack ways to introduce metals into their bodies, such as eating non-specific foods, harmful production conditions, and bad habits, etc., this group of people was selected as the target group.

[0076] In this dataset, each subject was assigned a corresponding identification number, and all data are anonymous. The sample categories are 18 residential areas related to the water intake points of the subjects' drinking water supplies. Therefore, each data vector contains four categorical features: x1 (weight, unit: kg), x2 (height, unit: cm), x3 (zinc content in blood, unit: mg / ml), and x4 (zinc content in drinking water, unit: mg / ml), as well as a coded category of the residential area.

[0077] The dataset segment (sequence numbers 1 - 5) used for testing is shown in Table 1.

[0078] Table 1 Survey data of zinc content in children's blood

[0079]

[0080] The experimental plan is as follows. A specific-sized subsample is randomly selected from the overall sample for the experiment in this example. Among them, 85% of the experimental samples are used as the training set, and 15% are used as the test set. The selection of the test data vectors is random. Then, an ART-WD model is constructed using the training set according to the method of the present invention, and its accuracy on the training set is analyzed. Subsequently, based on the trained ART-WD model, a dynamic ART-DWD model is constructed by removing a certain attribute from the vectors of the test set, and its ability to classify and recover the test set vectors is tested. Since the proposed model not only performs classification tasks but also is responsible for recovering missing attributes, two model accuracy evaluation criteria are proposed. Criterion 1 is the percentage of correct classification, and criterion 2 is the accuracy of attribute recovery.

[0081] The accuracy assessment of the ART-WD model is only based on criterion 1, while the assessment of the ART-DWD model combines two criteria and introduces a deviation threshold for criterion 2. If a sample is assigned to the correct category by the model and the calculated standard deviation of the missing attribute values does not exceed the specified threshold, the sample is considered to be correctly identified. In the private experiment of this case, the allowed deviation threshold is set to 15%. To test the performance of the model on datasets of different scales, a series of experiments were conducted using subsamples of ultra-small (50 rows), small (100 rows), medium (150 rows), large (200 rows), and full (240 rows) scales respectively.

[0082] First, the experimental results on the ultra-small subsample are presented: 7 samples were randomly selected as the test set, and the remaining 43 samples were used as the training set, and the attribute x4 in the test set was hidden. By training on the ultra-small subsample, the test accuracy of the ART-DWD model reached 100%, and the confusion matrix of the test results is as Figure 5 shown, and the detailed test results are shown in Table 2.

[0083] Table 2

[0084]

[0085] By analyzing the accuracy of the attribute value recovery of the ART-DWD model in Table 2, it can be seen that the sample numbered 49 was assigned to the correct category. However, the calculation error of its missing attribute value is 1% higher than the acceptable threshold of 15%. After calculation, it can be obtained that the recovery accuracy of the missing attribute value reached 90%.

[0086] Repeat the above experiment. This time, the attribute x3 was hidden, and the experimental results of the vector recovery and classification of the ART-DWD model are as Figure 6 shown. It can be seen from Figure 6 that the sample numbered 13 was correctly classified by the algorithm. However, the calculation error of its missing attribute value exceeded the acceptable threshold by 4%. The average accuracy of attribute recovery is 90%.

[0087] To further verify the performance of the model in the field of weakly defined data, the method of the present invention (ART-WD) was compared with some classical machine learning algorithms, including ART-2, multi-layer perceptron (MLP), logistic regression (LR), random forest (RF), and k-nearest neighbor (k-NN) method. The experiment was divided into 5 groups and experiments were conducted on subsamples of small scale (100 rows), medium scale (150 rows), large scale (200 rows), and full scale (240 rows) respectively. It can be seen from Table 3 that the performance of the ART-WD model far exceeds the other 5 models, especially under the condition of small samples, the ART-WD has significant advantages.

[0088] Table 3 Classification Accuracy of Different Models for Complete Data

[0089]

[0090] To further verify the classification ability of the model for weakly defined data with missing attributes, the classification accuracy of ART-DWD was also compared with other basic classification methods. First, the attribute value x3 (zinc content in blood) was removed from the dataset, and then tests were conducted on samples of different scales in turn. As can be seen from Table 4, when one attribute was removed, the accuracy of all classical models decreased, while the accuracy of ART-DWD remained very high. The analysis believes that this is because ART-DWD has extremely strong classification ability in dealing with samples with missing attributes and can cluster weakly defined data well.

[0091] Table 4 Classification Accuracy of Different Models in the Case of Missing Attribute x3

[0092]

[0093] Similarly, the attribute value x4 (zinc content in drinking water) was removed from the dataset, and tests were conducted on samples of different scales in turn. The results are shown in Table 5.

[0094] Table 5 Classification Accuracy of Different Models in the Case of Missing Attribute x4

[0095]

[0096] The experimental results show that regardless of which attribute is removed, the classification performance of the ART-DWD model remains very high. For other models, the performance of the model is more dependent on the missing attribute value. For example, the multi-layer perceptron is less affected when removing attribute x3 than when removing attribute x4.

[0097] The experimental results show that the ART-WD model performs excellently in the classification task. Specifically, its classification accuracy is significantly better than that of existing classical classification algorithms. Its powerful classification ability is mainly due to the design of the model's adaptive resonance threshold and the efficient decision-making mechanism. In addition, the ART-DWD model performs excellently in dealing with samples with missing attributes. The experimental results show that the ART-DWD model can not only efficiently cluster samples with missing attributes but also has excellent attribute value recovery ability. Under the experimental settings of different missing attributes, the ART-DWD model shows stable and excellent clustering effects and attribute value recovery effects, verifying its practicability and robustness in dealing with incomplete data.

[0098] The above content is a further detailed description of the present invention in combination with specific embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as falling within the protection scope determined by the claims submitted for the present invention.

Claims

1. A sparse data recognition and recovery neural network model based on the adaptive resonance theory, including an ART_WD network for classifying sparse data and an ART_DWD network for recovering the attribute values of sparse data, characterized in that: The data recognition process of the ART_WD network is as follows: Step 1, input the sample X=(x1, x2, …, x m ) into the data preprocessing module I, and perform data normalization or remove random noise in module I to obtain U=(u1, u2, …, u m ); Step 2, the vector U=(u1, u2, …, u m ) enters the recognition module II, and through calculation with the ascending weights w ij , obtain Y=(y1, y2, …, y k ), and through the formula: Select the winning neuron from them; Step 3, after calculating the winning neuron in Step 2, further calculate the similarity RN by calculating U=(u1, u2, …, u m ) and the descending weights w ji corresponding to the winning neuron; Step 4, input the similarity RN calculated by the model control module III into the model conversion module IV, and compare RN with Y=(y1, y2, …, y k ) and the similarity threshold P ρ . According to the size relationship among the similarity RN, the similarity threshold P ρ and the neuron Y=(y1, y2, …, y k ), as well as the number k of remaining neurons, obtain the recognition result of the model, and trigger the corresponding resonance unit S1, reset unit S2 and adaptation unit S3; The data recovery process of the ART_DWD network is as follows. Step 1, after the input sample X=(x1, x2, …, x m ) is normalized and the random noise is removed to obtain U=(u1, u2, …, u m ), the number q of missing attributes is compared with the suitability threshold P a . If q is equal to 0, then U=(u1, u2, …, u m ) undergoes the above data identification process; if q < P a , then U=(u1, u2, …, u m ) is recovered through the ART_DWD network; if q > P a , it is determined that the corresponding sample is not suitable for data recovery; Step 2, set the positions of the missing attributes in U=(u1, u2, …, u m ) to 0, and set the corresponding positions in the ascending weight w ij and the descending weight w ji to 0. Then the vector U=(u1, u2, …, u m ) enters the recognition module II, and Y=(y1, y2, …, y ij ) is calculated by multiplying with the ascending weight w k , and through the formula: Select the winning neuron from them; Step 3, after calculating the winning neuron in Step 2, further calculate the similarity RN by calculating U=(u1, u2, …, u m ) and the descending weights w ji corresponding to the winning neuron; Step 4, input the similarity RN calculated by the model control module III into the model conversion module IV, and compare RN and Y=(y1, y2, …, y k ) with the similarity threshold P ρ . According to the size relationship between the similarity RN, the similarity threshold P ρ and the neuron Y=(y1, y2, …, y k ), as well as the number k of remaining neurons, obtain the recognition result of the model, and trigger the corresponding resonance unit S1 and reset unit S2; Step 5, when the sparse data with missing attributes resonates, the ART_DWD model will perform data recovery according to its category.

2. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, wherein: The normalization process of the input sample X = (x1, x2, …, x m ) in steps 1 of the ART_WD network and the ART_DWD network is specifically as follows:

3. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: The calculation processes in step 2 of the ART_WD network and the ART_DWD network and the selection formula for the winning neuron are respectively: where the ascending weight w ij is the centroid vector calculated from all the normalized vectors within category Y j and the formula is: U l ∈Y j , where S represents the number of vectors belonging to class j.

4. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: In the similarity RN calculation process of step 3 of the ART_WD network and the ART_DWD network, each input sample will be stored in the descending weights w of the winning neuron of the category to which it belongs, i.e.: ji That is: Belong to the corresponding category: where S is the number of vectors previously assigned to category j. There are a total of S descending weights connecting the winning neuron, that is, m×S. According to the descending weights w connected to the winning neuron ji , first calculate the single influence r of the samples contained in the descending weights on this input sample l , and then calculate the similarity RN by the pairwise averaging method:

5. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: The trigger condition of the adaptation unit S3 in step 4 of the ART_WD network is that the value of any neuron in Y is greater than the similarity threshold P ρ Or when the winning neuron y j* has a value less than the similarity threshold P ρ and the similarity RN is greater than the similarity threshold P ρ and k = 1, the adaptation unit S3 will be triggered, and the recognition module II will create a new category Y k+1 , and the ascending weight and descending weight connected to the neuron Y k+1 are respectively defined as: wik + 1 = ui, S＝1。 6. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: The triggering conditions of the resonance unit S1 and the reset unit S2 in steps 4 of the ART_WD network and the ART_DWD network are that the value of the winning neuron y j* is less than the similarity threshold P ρ and the similarity RN is less than the similarity threshold P ρ , then the resonance unit S1 will be triggered, and the recognition module II will classify the input sample into the category where the winning neuron is located, and update the ascending weight and the descending weight. The update formula is as follows: The winning neuron y j* has a value greater than the similarity threshold P ρ Or when the winning neuron y j* has a value less than the similarity threshold P ρ and the similarity RN is greater than the similarity threshold P ρ and when k is greater than 1, the reset unit S2 will be triggered, and the recognition module II will reset the neurons, k = k - 1, and y j* will be temporarily removed from Y.

7. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: After resonance occurs in the input samples of the ART_DWD network, for the input sample X, its attribute values will be recovered through its belonging class j*, and the formula is: where S is the number of samples belonging to class j*, is the attribute value of position g of the complete vector X belonging to class j*. After data recovery, this temporary model is destroyed.

8. The sparse data recognition and recovery neural network model based on the adaptive resonance theory according to claim 1, characterized in that: Both the ART_WD network and the ART_DWD network include a preprocessing module Ⅰ; an identification module Ⅱ; a control module Ⅲ; a model conversion module Ⅳ, and the model conversion module receives the similarity RN calculated by the control module and the output result Y=(y1, y2, …, y k ), and compares it with the similarity threshold P ρ . According to the comparison result, it is input into different units: a resonance unit S1; a reset unit S2; an adaptation unit S3. The ART_DWD network further includes an input detection module ① that compares the number of missing attributes with the suitability threshold P a ; and a data recovery module ② for the sparse data after resonance.

Images