Object prediction model construction method and prediction method in multi-center small-sample scenario

Abstract

The present invention relates to an object prediction model construction method and prediction method in a multi-center small-sample scenario. On the basis of the idea of transfer learning, prediction is directly performed on a training set of a new node by using knowledge of trained nodes, and the difference between the new node and the trained nodes is reflected by means of an incorrectly predicted sample, which difference is used as supplementary knowledge, thus quickly acquiring model knowledge of the new node, and avoiding training the new node from the very beginning each time; and finally, parallel integration of incremental sub-classifiers is realized by using a ridge regression method, thereby greatly reducing deployment time and costs. The generalization of a model is ensured by means of historical knowledge sharing and a knowledge discard mechanism, and therefore with the support of sufficient shared knowledge, even for a node with a small sample size, a relatively good classification effect can be achieved. An improvement is performed on the basis of a zero-order TSK fuzzy system having high interpretability, all of the model parameters can be embodied in the form of a rule, and high semantic transparency is achieved; thus, a user can be assisted in analyzing and understanding data in practical applications.

Description

Target prediction model construction method and prediction method in multi-center small sample scenario Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a target prediction model construction method and a prediction method in a multi-center small sample scenario. Background Art

[0002] With the rise of computing power and the advent of the big data era, machine learning has become a powerful tool for data analysis, making significant contributions to improving the quality of human life. However, due to data privacy concerns, data barriers exist between different institutions and departments, making it difficult to share massive amounts of data. Deploying and jointly training data across multiple centers presents challenges. To address this, Google has proposed federated learning technology, which effectively protects user privacy by storing and training each user's data locally and interacting only with a central server for model updates. While numerous algorithms exist in this field, most are based on deep learning networks. Current deep learning methods have numerous parameters, large models, a high risk of data leakage, and are difficult to deploy. While classification accuracy can reach very high levels, most are "black box" models lacking semantic interpretability. This makes them inadequate for high-risk scenarios and unsuitable for organizations requiring high security and interpretability. Traditional machine learning algorithms, such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Naive Bayes (NB), Latent Dirichlet Allocation (LDA), Classification and Regression Tree (CART), etc., are also prone to low accuracy for multi-center data due to data drift and small sample problems. They are not as effective as deep learning classification and cannot guarantee the privacy of multi-center data.

[0003] Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides a target prediction model construction method and prediction method in a multi-center small sample scenario.

[0005] A method for constructing a target prediction model in a multi-center small sample scenario includes the following steps:

[0006] S1: Input the features of the training samples into a multi-center training dataset with m data centers, and use multiple sub-classifiers and a historical knowledge classifier to predict the training samples respectively. The historical knowledge classifier is a classifier that integrates the rules of all sub-classifiers. The integration of all prediction results has a total of m+1 dimensions, where each dimension represents the prediction result of a classifier. If it is greater than 0, it is judged as a positive class, and if it is less than 0, it is a negative class.

[0007] S2: Normalize the prediction result vector to between [-1,1];

[0008] S3: Using the normalized prediction vector as the prediction feature and the training labels corresponding to the training samples, the ridge regression method is used to train the weights of each sub-classifier;

[0009] S4: Combine the prediction vector and the trained classifier weight to obtain the comprehensive prediction result;

[0010] S5: Classify the comprehensive prediction results to obtain the same results as the training labels.

[0011] Preferably, the objective function of the weight training and the solution formula of the weight parameters are: P g =(λI+P T P) -1 P T Y,

[0012] Among them, P is the prediction vector, W is the weight vector of the sub-classifier to be trained, Y is the true value label of the corresponding training sample, λ is the regularization parameter of the structural risk term, P g are the consequent parameters calculated by the ridge regression method, and I is the identity matrix.

[0013] Preferably, the training of the first sub-classifier includes two stages: a basic stage and a learning stage.

[0014] Preferably, the basic stage includes: obtaining the initial fuzzy rule parameters c, P by means of the classic 0th order TSK g , δ and misclassified sample information errors, using the Gaussian kernel function as the membership function; where c and δ are antecedent parameters, c is the cluster center obtained by FCM clustering of the input data Data, δ is the membership matrix U of the Data to the cluster center obtained by FCM clustering of the input data Data, and the membership function kernel width obtained by c.

[0015] Preferably, the learning phase includes: firstly g Execute knowledge discarding processing, then organize knowledge to obtain the best knowledge combination, and finally calculate the consequent knowledge corresponding to the current fuzzy rule based on the input data to obtain the prediction result.

[0016] Preferably, the knowledge discarding process includes: first, determining the knowledge discarding ratio r1 and the knowledge confusion feature ratio r2, randomly selecting the rule to be discarded based on r1, and randomly selecting the feature column lose_columns to be processed for knowledge confusion based on r2; for each rule to be discarded, randomly determining whether to perform knowledge forgetting or knowledge confusion; for the rule to be forgotten, directly convert c and P g Leave it empty; for the rules that need to be processed with knowledge confusion, replace the lose_columns column of the rule with the corresponding column values ​​of other rules.

[0017] Preferably, the knowledge arrangement includes: discarding the processed c and P g and the error classification sample information errors as candidate knowledge, from which the best knowledge combination is selected as the preliminary knowledge of the current stage. The error classification sample information errors includes the corresponding c and P g The sample features and sample labels.

[0018] Preferably, the subsequent knowledge corresponding to the current fuzzy rule is calculated in combination with the input data Data to obtain the prediction result, specifically including: calculating the membership function kernel width δ in combination with the input data Data and c in the preliminary knowledge; using the antecedent parameters c and δ and the historical migration knowledge of the corresponding consequent parameters, through the ridge regression method, taking the empirical risk term, the structural risk term and the historical consequent parameter migration knowledge regularization term as the objective function to calculate the consequent knowledge corresponding to the current fuzzy rule, the objective function expression and the consequent parameter solution formula are as follows:

[0019] Among them, X g is the input vector of the training sample after Gaussian kernel mapping, P g-history Knowledge of historical consequent parameter transfer.

[0020] Preferably, after each sub-classifier is trained, the generated fuzzy rules are updated and added to the historical fuzzy rule base so as to be continued to be used by subsequent sub-classifiers.

[0021] A target prediction method in a multi-center small sample scenario, comprising the following steps:

[0022] S1: Build a prediction model using the target prediction model construction method in the multi-center small sample scenario as described above;

[0023] S2: Inputting features of the current target corresponding to the features of the training sample into the prediction model;

[0024] S3: The prediction model obtains a prediction result of the current target based on the features.

[0025] The above technical solution of the present invention has the following advantages over the prior art:

[0026] 1. The target prediction model construction method in the multi-center small sample scenario described in the present invention combines the idea of ​​transfer learning to directly use the knowledge of trained nodes to predict the training set of new nodes, and uses prediction error samples to reflect the difference between new nodes and trained nodes, which is used as supplementary knowledge to quickly obtain model knowledge of new nodes, avoiding training new nodes from scratch every time. Finally, the ridge regression method can be used to realize the parallel integration of incremental sub-classifiers, which greatly saves deployment time and cost.

[0027] 2. The target prediction model construction method in the multi-center small sample scenario described in the present invention ensures the generalization of the model through sharing historical knowledge and knowledge discarding mechanism. With the support of sufficient shared knowledge, better classification effects can be achieved for nodes with small sample sizes.

[0028] 3. The target prediction model construction method in the multi-center small sample scenario is improved based on the highly interpretable zero-order TSK fuzzy system. All model parameters can be expressed in the form of rules, with a high degree of semantic transparency, which can better help users analyze and understand data in practical applications.

[0029] 4. The target prediction model construction method in the multi-center small sample scenario described in the present invention effectively ensures that samples between different nodes will not be leaked to each other, and shows better classification performance in the experiment. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.

[0031] FIG1 is the overall model architecture of the target prediction model construction method in a multi-center small sample scenario of the present invention.

[0032] FIG2 is a diagram showing the training details of the first subclassifier of the target prediction model construction method in a multi-center small sample scenario of the present invention.

[0033] FIG3 is a collaborative training process of sub-classifiers of the target prediction model construction method in a multi-center small sample scenario of the present invention.

[0034] FIG4 is an experimental flow chart of the target prediction model construction method in a multi-center small sample scenario of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0036] As shown in FIG1 , the present invention provides a method for constructing a target prediction model in a multi-center small sample scenario, comprising the following steps:

[0037] S1: Input the features of the training samples in a multi-center training dataset with m data centers, and use multiple sub-classifiers and a historical knowledge classifier to predict the training samples respectively. The historical knowledge classifier is a classifier that integrates the rules of all sub-classifiers; all the prediction results are integrated into a total of m+1 dimensions, each dimension represents the prediction result of a classifier, and if it is greater than 0, it is judged as a positive class, and if it is less than 0, it is a negative class.

[0038] S2: Normalize the prediction result vector to [-1, 1]. This operation can ensure the uniformity of dimensions.

[0039] S3: Using the normalized prediction vector as the prediction feature and the training labels corresponding to the training samples, the ridge regression method is used to train the weights of each sub-classifier. The objective function of the weight training and the solution formula of the weight parameters are as follows: P g =(λI+P T P) -1 P T Y,

[0040] Among them, P is the prediction vector, W is the weight vector of the sub-classifier to be trained, Y is the true value label of the corresponding training sample, λ is the regularization parameter of the structural risk term, P g is the consequent parameter calculated by the ridge regression method, and I is the identity matrix. By using the ridge regression method with the above formula, the solution of the weight W can be quickly obtained.

[0041] S4: Calculate the comprehensive prediction result by combining the prediction vector and the trained classifier weights.

[0042] S5: Classify the prediction results. If it is greater than 0, it is a positive class, otherwise it is a negative class.

[0043] In a specific embodiment, the first sub-classifier ensures the rapid training of subsequent sub-classifiers. The first sub-classifier should be trained in a data center with relatively sufficient data. As shown in Figure 2, the training of the first sub-classifier includes two phases: the basic phase and the learning phase.

[0044] In the basic stage, the initial fuzzy rule parameters c, P are obtained by using the first-order form and the zero-order form of the classic zero-order TSK fuzzy system (Takagi-Sugeno-Kang, TSK). g , δ, and the misclassified sample information errors, using the Gaussian kernel function as the membership function. Among them, c and δ are the antecedent parameters, c is the cluster center obtained by the fuzzy C-means algorithm (FCM clustering) of the input data Data, δ is the membership matrix U of the Data to the cluster center obtained by FCM clustering, and c is the membership function kernel width.

[0045] During the learning phase, the following steps are included:

[0046] S1: First, c and P g Perform knowledge discard processing,

[0047] S2: Then organize the knowledge to get the best knowledge combination,

[0048] S3: Finally, the consequent knowledge corresponding to the current fuzzy rule is calculated based on the input data to obtain the prediction result.

[0049] In S1, the knowledge discarding mechanism simulates the forgetting characteristics of the human brain, including knowledge forgetting and knowledge confusion, corresponding to the two situations of complete knowledge forgetting and partial knowledge feature forgetting. First, the knowledge discarding ratio r1 and the knowledge confusion feature ratio r2 are determined. Based on r1, the rules to be discarded are randomly selected, and based on r2, the feature columns lose_columns to be processed for knowledge confusion are randomly selected. For each rule to be discarded, it is randomly determined whether to forget or confusion the knowledge. The rule to be forgotten directly replaces c and P g Just leave it blank; for the rules that need to be processed for knowledge confusion, replace the lose_columns column of the rule with the corresponding column values ​​of other rules to achieve the effect of knowledge confusion.

[0050] In S2, the processed c and P are discarded g And the error classification sample information errors are taken as candidate knowledge, from which the best knowledge combination is selected as the preliminary knowledge of the current stage. The error classification sample information errors includes the corresponding c and P g The distance calculation involved in knowledge organization is the distance between the centers of the center set composed of the feature columns of the antecedent knowledge c and the errors. The selection of knowledge follows two principles: high reliability and low similarity.

[0051] First, to measure the credibility of candidate centers, the number of training samples within the neighborhood of each candidate center is calculated and then the values ​​are mapped to [0, 1]. The more training samples in the neighborhood, the higher the credibility of the candidate center. The neighborhood radius is the average of the sum of the Euclidean distances between all candidate centers. Furthermore, the similarity between each candidate center and the representative center set is estimated by calculating the Euclidean distance between the candidate center and the nearest center in the representative center set, with an attempt to ensure that the similarity between the selected centers is low. Similarly, the similarity values ​​are also mapped to [0, 1]. The importance of each center is reflected by calculating the sum of the mapped reliability and similarity indicators. Based on this, the final optimal center combination is selected. Finally, these centers and the corresponding consequent knowledge are selected as the optimal knowledge combination.

[0052] In S3, the membership function kernel width δ is calculated by combining the input data Data and the preliminary knowledge c. At this point, using the antecedent parameters c and δ and the historical migration knowledge of the corresponding consequent parameters, through the ridge regression method, the empirical risk term, the structural risk term, and the historical consequent parameter migration knowledge regularization term are used as the objective function to quickly calculate the consequent knowledge corresponding to the current fuzzy rule. The objective function expression and the consequent parameter solution formula are as follows:

[0053] Among them, X g is the input vector of the training sample after Gaussian kernel mapping, P g-history The knowledge of historical consequent parameters is transferred. Thus, with the help of historical knowledge and knowledge of misclassified samples, further generalization of knowledge is achieved, and more robust new knowledge is continuously obtained.

[0054] In a specific embodiment, the collaborative training process of multiple sub-classifiers is shown in Figure 3. After training, each sub-classifier updates the generated fuzzy rules and adds them to the historical fuzzy rule base for continued use by subsequent sub-classifiers. For sub-classifiers other than the first sub-classifier, the basic stage training is no longer required. Instead, the knowledge of the current historical fuzzy rule base is directly used to predict the data samples of the center. After obtaining the error information of the misclassified samples, the learning stage training can be carried out just like the first sub-classifier. This greatly saves training time and better utilizes historical knowledge, avoiding the situation where the basic stage cannot obtain good knowledge in small sample scenarios.

[0055] The present invention also provides a target prediction method in a multi-center small sample scenario, the steps comprising:

[0056] S1: Build a prediction model using the target prediction model construction method in the multi-center small sample scenario as described above;

[0057] S2: Inputting features of the current target corresponding to the features of the training sample into the prediction model;

[0058] S3: The prediction model obtains a prediction result of the current target based on the features.

[0059] The target prediction model construction method and prediction method in a multi-center small sample scenario of the present invention can be applied to scenarios such as traffic situation prediction and disease auxiliary diagnosis.

[0060] When traffic is congested, if enough traffic police are not dispatched in advance to clear the traffic, traffic accidents and violations will occur, posing a serious threat to people's lives and property. Specifically, when predicting traffic conditions, different cities are used as data centers. The traffic congestion time, congestion location, etc. of each data center will be different due to various reasons, but there are similar trends. Each city collects its traffic data over the years as training features and jointly uses the target prediction model construction method under the multi-center small sample scenario described in this application to construct a congestion prediction model, so that traffic data from different data centers can be shared in an abstract fuzzy rule without leaking to other data centers. Each center can input current traffic data as a feature into the congestion prediction model to obtain a prediction result of the current congestion situation. The prediction result draws on the congestion feature trends of other centers to more reasonably judge which time periods are more likely to have congestion peaks, so that traffic police can be dispatched in a timely manner to ensure traffic and ensure people's normal passage and driving safety.

[0061] In the medical field, doctors are usually unable to provide a large amount of labeled data for artificial intelligence model training. In most cases, they can only obtain a small amount of labeled samples, and the prediction results obtained from small sample training are usually relatively poor. For medical data from different hospitals, although the data distribution may be different due to regional and other reasons, there are still similar trends, which have great guiding significance for the training of each other's models. However, for the sake of patient privacy, medical data are generally not disclosed to the outside world. Therefore, although multiple hospitals can have a large amount of data, they cannot be used interchangeably with each other. The target prediction model construction method in a multi-center small sample scenario described in this application can extract the abstract fuzzy rules of each hospital. Under the premise of ensuring patient privacy, each hospital can develop a more reasonable prediction model under the guidance of the fuzzy rules of other hospitals, which becomes an auxiliary means of doctor's diagnosis, greatly reduces the workload of doctors, improves the efficiency of patients' medical treatment, and makes full use of the data of different hospitals.

[0062] This application uses the personal-key-indicators-of-heart-disease dataset on the Kaggle website (https: / / www.kaggle.com / datasets / kamilpytlak / personal-key-indicators-of-heart-disease) as the experimental dataset. Due to the excessive amount of data in the original dataset, in order to simulate the small sample node scenario of a multi-center dataset, the experiment selected some samples from it for testing. First, based on the racial feature columns in the dataset, whites, blacks, and Asians were taken as the three centers to create a dataset. The feature column information of the selected data samples was complete, and the character feature columns were replaced by discrete numbers for subsequent calculations. Each sample finally selected included 16 feature columns and 1 label column. In the experiment, the training set and the test set were divided in a ratio of 4:1, and five-fold cross-validation was used to determine the parameters. The data features were normalized and the features were mapped to the range of [-1,1]. The following table records the specific division of the dataset:

[0063] Among them, the positive class represents sample data of people with heart disease, and the negative class represents sample data of people without heart disease.

[0064] The specific experimental steps are shown in Figure 4, including:

[0065] S1: Data preprocessing. According to race, samples with complete feature information were selected from three centers: Caucasians, blacks, and Asians to create data sets, and character feature columns were replaced with discrete numbers for subsequent calculations. Each sample finally selected included 16 feature columns and 1 label column, where the label column was used to calculate the weights of each sub-classifier or to verify the accuracy of the target prediction model construction method in the multi-center small sample scenario of this application. In the experiment, the training set and test set were randomly divided for each center in a ratio of 4:1, and the data features were normalized to map the features to the range of [-1,1].

[0066] S2: Create a multi-center dataset. Based on the processed data, save the training set and test set for each center in .mat format. Additionally, save the training set and test set for all centers in .mat format to facilitate training the weights of each sub-classifier.

[0067] S3: Model training and parameter range adjustment. Since the network will have different effect deviations for specific data sets, grid optimization is required. The main function is to obtain the best parameters for a specific network for a specific data set. The optimal parameters are found from the parameter optimization range in a five-fold cross-validation manner. The final optimal hyperparameters are obtained by continuously updating the current optimal parameters until the end of the iteration. In this process, the model results are observed through multiple experiments, and the performance of the model on the training set and test set is combined to determine whether the model has achieved the best fit, and gradually determine the optimal optimization range of the model parameters. In the above data set, the relevant parameters of the algorithm during optimization are as follows:

[0068] S4: Retrain. Use grid optimization to find the optimal hyperparameters and retrain the model.

[0069] S5: Model testing. Use the test set to test the trained model and observe whether the test results meet expectations.

[0070] S6: Evaluation indicators. The test results of the test set have been obtained through the above steps. Multiple evaluation indicators are calculated using the formula. The evaluation indicators used in this experiment are accuracy, precision, recall, F1 score, and AUC.

[0071] This example compares the target prediction model construction method in the multi-center small sample scenario provided by this application with the existing classic traditional machine learning algorithms, including Linear Discriminant Analysis (LDA), K-Nearest Neighbor (KNN), Naive Bayes ( The first-order and zero-order forms of the three algorithms are Bayes (NB), Classification and Regression Tree (CART), Support Vector Machines (SVM), and Takagi-Sugeno-Kang Fuzzy System (TSK-FS). The following table records the means and variances of the five classification evaluation metrics for ten experiments of these algorithms (all results are rounded to four decimal places for easier visualization):

[0072] Among them, AUC (Area Under Curve) is defined as the area under the ROC curve. The AUC value is usually used as the evaluation criterion of the model because the ROC curve often cannot clearly indicate which classifier is more effective. As a numerical value, the classifier with a larger AUC is more effective. As can be seen from the table, in terms of the three indicators of accuracy, precision and AUC, the target prediction model construction method for multi-center small sample scenarios provided by this application has better performance than other algorithms, and is comparable to other traditional algorithms in terms of recall rate and F1 score. Therefore, from the perspective of classification performance, the target prediction model construction method for multi-center small sample scenarios provided by this application has better performance than other traditional algorithms.

[0073] In addition, considering the interpretability of the algorithm, the rule complexity comparison of the three fuzzy-based systems in the experiment, 1-TSK-FS, 0-TSK-FS, and the target prediction model construction method in the multi-center small sample scenario provided by this application, is shown in the following table:

[0074] It can be seen from the above table that although 0-TSK-FS has a lower rule complexity than the target prediction model construction method for multi-center small sample scenarios provided by this application, the classification performance of the target prediction model construction method for multi-center small sample scenarios provided by this application is significantly better than 0-TSK-FS. However, 1-TSK-FS is not as good as the target prediction model construction method for multi-center small sample scenarios provided by this application in terms of classification performance and rule complexity. In summary, the classification performance and interpretability of the target prediction model construction method for multi-center small sample scenarios provided by this application are better than these traditional algorithms.

[0075] During formal model training, the cross-validation parameter optimization range is continuously adjusted by observing each training result to avoid overfitting caused by overly complex parameter models and improve model generalization. After multiple training sessions, a stable, high-performing model is gradually established. In practical applications, as long as a subset of institutions has sufficient sample data, institutions with smaller sample sizes can also quickly and effectively deploy local models without infringing on the data privacy of other institutions. This allows for the production of local models with superior classification, generalization, and interpretability, enabling more efficient joint training across multiple institutions.

[0076] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for constructing a target prediction model in a multi-center small sample scenario, characterized in that: The following steps are involved: S1: Input the features of the training samples in a multi-center training data set with m data centers, use multiple sub-classifiers and a historical knowledge classifier to predict the training samples respectively, and the historical knowledge classifier is a classifier that integrates the rules of all sub-classifiers; integrate all the prediction results to have a total of m+1 dimensions, each dimension represents the prediction result of a classifier, and if it is greater than 0, it is judged as a positive class, and if it is less than 0, it is a negative class; S2: Normalize the prediction result vector to between [-1,1]; S3: Using the normalized prediction vector as the prediction feature and the training labels corresponding to the training samples, the ridge regression method is used to train the weight of each sub-classifier; S4: Combine the prediction vector and the trained classifier weight to obtain a comprehensive prediction result; S5: classifying the comprehensive prediction results to obtain the same results as the training labels; The training of the first sub-classifier includes a basic stage and a learning stage; the training of other sub-classifiers is based on the training of the first sub-classifier, including a learning stage; the basic stage includes: obtaining the initial fuzzy rule parameters c, P by using the classic 0th order TSK g , δ and misclassified sample information errors, and use Gaussian kernel function as the membership function; where c and δ are antecedent parameters, c is the cluster center obtained by FCM clustering of the input data Data, δ is the membership matrix U of the cluster center obtained by FCM clustering of the input data Data and c as well as the kernel width of the membership function.

2. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 1, characterized in that: The objective function and solution formula of weight parameters for weight training using ridge regression method are as follows: P g =(λI+P T P) -1 P T Y, Among them, P is the prediction vector, W is the weight vector of the sub-classifier to be trained, Y is the true value label of the corresponding training sample, λ is the regularization parameter of the structural risk term, and P g is the consequent parameter calculated by the ridge regression method, and I is the identity matrix.

3. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 1, characterized in that: The learning phase includes: first, g Execute knowledge discarding processing, then organize knowledge to get the best knowledge combination, and finally combine the input data to calculate the consequent knowledge corresponding to the current fuzzy rule to get the prediction result.

4. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 3, characterized in that: The knowledge discarding process includes: first, determining the knowledge discarding ratio r1 and the knowledge confusion feature ratio r2, randomly selecting the rule to be discarded according to r1, and randomly selecting the feature column lose_columns to be processed according to r2; for each rule to be discarded, randomly determining whether to forget knowledge or to be confused; the rule to be forgotten directly replaces c and P g Leave it blank; for the rules that need to be processed with knowledge confusion, replace the lose_columns column of the rule with the corresponding column values ​​of other rules.

5. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 4, characterized in that: The knowledge arrangement includes: discarding the processed c and P g and the error classification sample information errors are taken as candidate knowledge, and the best knowledge combination is selected as the preliminary knowledge of the current stage. The error classification sample information errors includes the corresponding knowledge of c and P g The sample features and sample labels.

6. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 5, characterized in that: Combine the input data Data to calculate the consequent knowledge corresponding to the current fuzzy rule, and obtain the prediction results, which specifically include: combining the input data Data and the preliminary knowledge c to calculate the membership function kernel width δ; using the antecedent parameters c and δ and the historical migration knowledge of the corresponding consequent parameters, through the ridge regression method, the empirical risk term, the structural risk term and the historical consequent parameter migration knowledge regularization term are used as the objective function to calculate the consequent knowledge corresponding to the current fuzzy rule. The objective function expression and the consequent parameter solution formula are as follows: Among them, X g is the input vector of the training sample after Gaussian kernel mapping, P g-history Knowledge of historical consequent parameter migration.

7. The method for constructing a target prediction model in a multi-center small sample scenario according to claim 1, characterized in that: After each sub-classifier is trained, the generated fuzzy rules are updated and added to the historical fuzzy rule base for continued use by subsequent sub-classifiers.

8. A target prediction method in a multi-center small sample scenario, characterized in that the steps include: S1: constructing a prediction model using the target prediction model construction method in a multi-center small sample scenario as described in any one of claims 1 to 7; S2: Inputting the features of the current target corresponding to the features of the training sample into the prediction model; S3: The prediction model obtains a prediction result of the current target according to the features.

Images