A classification method and device for generalized feature space data stream

Abstract

The present invention discloses a classification method and apparatus for generalized feature space data streams. When a new generalized feature space data stream arrives, the method performs feature selection based on the MRMR algorithm. The currently selected features and currently stored features are classified into three categories: active, new, and silent. The information contained in the data is then compressed into parameters associated with each feature. When the active parameter is not null, the output of an ER Rule classifier constructed based on the active feature is used as a compact representation of the corresponding feature for learning knowledge related to the new feature. Finally, after the learning process is completed, a subsequent feature selection process based on the activation frequency and weight of each feature is performed, if necessary, to meet storage and computing requirements. This method solves the learning problem of generalized feature space data streams and employs different strategies to address generalization issues in different situations.

Description

Technical Field

[0001] The present invention relates to the field of computer machine learning, and in particular to a classification method and device for generalized feature space data streams. Background Art

[0002] In recent years, data stream mining has become a hot topic in research. Instances in data streams are acquired individually or in batches. Existing learning algorithms can incrementally learn the information from data streams in real time, reducing computational and storage requirements while also not relying on the assumption that samples are independent and identically distributed. However, most data stream research assumes a fixed feature space. In many real-world applications, the features of data streams acquired over time vary, such as document topic classification, environmental quality monitoring, and job applicant quality prediction. Consequently, some recent research has focused on learning data streams with dynamic feature spaces and proposed corresponding algorithms, such as trapezoidal data streams and feature evolution streams with or without overlapping cycles. However, these algorithms impose strict assumptions on the dynamic nature of the feature space, which some real-world problems do not conform to. For example, human activity recognition based on various smart devices has widespread applications and is a burgeoning research topic. Data arrives continuously during the operation of each device, naturally forming a data stream problem, and the features collected by different devices vary. Devices may cease operation at any time due to insufficient battery or memory and can resume operation upon recovery. Furthermore, new devices can be added at any stage if necessary. Therefore, the characteristics of the data stream obtained at each stage may be different, and the appearance and exit of the characteristics do not follow any fixed pattern.

[0003] Data similar to the above problem is called a data stream with a generalized feature space. Its main characteristics are that existing and new features may appear or disappear randomly at each stage, and the number of instances in each stage may also vary. Currently, only a few papers have addressed very similar issues. Masud et al. proposed a feature set homogenization technique to address the feature evolution problem. This technique expands the model and the instance to be tested into the union of their feature spaces, with the value of the expanded features set to 0. This approach is unreasonable for some features, as 0 can sometimes have practical significance. Beyazit et al. proposed an online learning from diverse feature spaces (OLVF) algorithm, which classifies samples and feature spaces separately. However, OLVF implicitly assumes that the features of newly arriving data must contain existing features. Recently, He et al. proposed the generative learning from arbitrary feature streams (GLSC) algorithm to mine data streams with erratic feature spaces. GLSC uses a generative graph model to establish relationships between the features of the current instance and existing features, but it assumes that there must be linear correlation between features.

[0004] Without any assumptions about the changes in the feature space, learning from a generalized feature space data stream mainly includes four issues that need to be considered: (1) the feature space of each stage is random, so it is impossible to know which features will appear in the new stage; (2) identifying the specific problem to be solved and determining the specific strategy for leveraging existing knowledge and learning new knowledge; (3) as new features are continuously introduced in the data stream, the dimension of the global feature space will become infinite; (4) the proposed method needs to meet the constraints of data stream mining, such as processing each sample only once (no historical samples are stored) and limited storage space. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a classification method for generalized feature space data stream, which includes:

[0006] First, feature selection is performed on the data stream arriving at stage t. Then, the features selected at stage t and the features stored in the global feature space at stage t-1 are classified into a new feature set, an active feature set, and a silent feature set, respectively. The new feature set is the features introduced at stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear at stage t.

[0007] When the active feature set is empty, the dynamic evidence reasoning rule classifier is constructed using the features selected in stage t to obtain the accuracy acc t and classifier parameters par t ;

[0008] When the active feature set is not empty, the activation frequency of each feature in the active feature set is updated, and the classifier parameter par of stage t-1 is used. t-1 Build a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty;

[0009] If the active feature set is not empty and the new feature set is not empty, the initial fusion result is first merged with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters;

[0010] If the active feature set is not empty and the new feature set is empty, use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t ;

[0011] According to the accuracy acc t and the classifier parameter par t , update the global feature space of stage t.

[0012] Furthermore, if the number of features contained in the global feature space of the updated stage t is greater than the maximum number of features that the system can bear, and the number of features contained in the silent feature set is greater than the number of features that need to be deleted, then the feature selection of stage t+1 will only be performed in the silent feature set. Otherwise, the active feature set will also participate in the feature selection process.

[0013] Furthermore, the active feature set also participates in the feature selection process, specifically deleting a certain number of features with low activation frequencies. When some features have the same activation frequency, the features with smaller weights are deleted.

[0014] Furthermore, the feature selection is specifically as follows: when the number of features in the data stream is greater than the maximum number of features that the system can bear, the maximum relevance minimum redundancy algorithm is used to select the required features.

[0015] Furthermore, the dynamic evidence reasoning rule classifier (dynamic ERRule classifier) first calculates feature reliability through feature reliability calculation. Then, through repeated steps of evidence acquisition, evidence activation and combination, parameter optimization, and optimal result update, it determines the number of reference values, feature reference values, and evidence weights. The optimal result for the number of reference values is the vector with the highest test set accuracy and the lowest mean squared error. When the dynamic evidence reasoning rule classifier is updated, the feature reference values and evidence weights are updated first.

[0016] Furthermore, for the data stream arriving at stage 0, feature selection is performed first, and then a dynamic evidence reasoning rule classifier is constructed based on the selected features to obtain the accuracy acc0 and classifier parameters par0 of stage 0, and update the global feature space of stage 0.

[0017] The present invention also proposes a classification device for generalized feature space data stream, the device comprising:

[0018] Initialization module, for the data stream arriving at stage 0, first performs feature selection, then builds a dynamic evidence reasoning rule classifier based on the selected features, obtains the accuracy acc0 and classifier parameters par0 of stage 0, and updates the global feature space of stage 0;

[0019] The question classification module first performs feature selection on the data stream arriving at stage t, and then classifies the features selected at stage t and the features stored in the global feature space at stage t-1 into a new feature set, an active feature set, and a silent feature set, respectively. The new feature set is the features introduced at stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear at stage t.

[0020] In the problem I processing module, when the active feature set is empty, the dynamic evidence reasoning rule classifier is constructed using the features selected in stage t, and the accuracy acc is obtained. t and classifier parameters par t ;

[0021] Problem II and Problem III preprocessing modules, when the active feature set is not empty, update the activation frequency of each feature in the active feature set and adjust the activation frequency according to the classifier parameter par in stage t-1 t-1 Build a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty;

[0022] Problem II processing module, if the active feature set is not empty and the new feature set is not empty, then first merge the initial fusion result with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters;

[0023] Problem III processing module, if the active feature set is not empty and the new feature set is empty, use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t ;

[0024] Update module, according to the accuracy acc tand the classifier parameter par t , update the global feature space of stage t.

[0025] The present invention further includes a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above method when executed by a processor.

[0026] The present invention solves the classification problem of naturally formed data streams through the proposed method, proposes three types of features to divide the existing feature space, solves the learning problem of generalized feature space data streams, and uses the classification advantages of the dynamic evidence reasoning rule classifier and the parameters contained therein to associate with each feature separately, adopting different strategies to solve the generalization problem in different situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0028] Figure 1 is the generalized feature space data flow.

[0029] Figure 2 For the generalization learning problem.

[0030] Figure 3 The overall framework diagram of a classification method for generalized feature space data stream. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0032] The following first defines the issues and symbols that may be involved in the embodiment, as shown in Table 1.

[0033] Table 1 Glossary

[0034]

[0035] Use n t and d t Indicates the number of instances and features contained in the data arriving at stage t, and represents the data stream sequence as {(X t ,Y t )|t=1,2,…,T}, where is a (n t *d t )-dimensional matrix, Represents the category label, which is an n t dimensional vector; F i (i=1,2,…,d t ) is a representation feature f i The value of n t dimensional vector; {(x k ,y k )|k=1,2,…,n t} represents a specific instance in stage t. It is also assumed that the data obtained in the same stage have no missing values, that is, each feature of each instance has a specific value.

[0036] Feature space refers to a feature set. At stage t, all the features that have appeared so far are called the global feature space. It can be divided into three categories: (1) active features, which are with U t-1 The intersection features, that is, These features have been introduced and stored in the previous stage; (2) new features, as the name suggests, are just t Features that have been introduced or that have been removed and reappeared. (3) The silent feature is defined as silent = U t-1 -active, i.e., they are features that appeared in the previous phase but have not reappeared in the current phase. For example, Figure 1 In Stage 7, U7={f1,f2,f3,f4,f5,f6,f7,f8,f9,f 10},active={f1,f2,f3,f5,f7,f8,f9},new={f 10},silent={f4,f6}.

[0037] Since silent features are not included in the data of the current stage, it is unrealistic to update their knowledge. Therefore, there are two types of features left to consider in the learning task: active and new. Depending on whether they are empty, we can get the following Figure 2 The four generalization scenarios shown are: (1) In the first quadrant, both active and new are empty, which is impossible because X t (2) The scene shown in the second quadrant is Figure 1 The corresponding Stage 5 in the table indicates that the current stage only has new features. (3) Figure 1Stage 2, Stage 4 and Stage 7 in the diagram correspond to the situation in the third quadrant, indicating that the features of the current stage contain both existing features and new features. (4) Finally, Figure 1 The remaining Stage 3 and Stage 6 can be generalized to the scenario in the fourth quadrant, which means there are no new features but some old features randomly reappear.

[0038] The dynamic ER Rule classifier is a supervised classifier used to solve the problem of applying the ERRule classifier in the absence of prior knowledge. This method learns parameters based on the nature of the data, which can reduce the subjectivity of the original method. There are three steps in total.

[0039] Assume that there are K sample data in a classification problem, and there are N categories to which they actually belong Θ={y1,y2,…,y n ,…,y N},y n is the nth class, Θ is the identification framework, each data consists of M features and the category label to which it belongs x={x1,x2,…,x M ,y},x i Refers to the value of the i-th feature of the sample. The training set data is represented as (K S ≤K,S i is feature x i )

[0040]

[0041] Step 1: Characteristic reliability calculation

[0042] The reliability of evidence indicates the classification ability of the corresponding feature that generated it, feature x i The more reliable it is, the more samples that can be directly judged by it alone. Therefore, feature x i The reliability of is defined as:

[0043]

[0044] Among them, Q i Indicates that only based on feature x i The value of can directly determine the total number of samples of the category. Then r i =1, that is, x i is the most reliable feature.

[0045] Step 2: Single randomized experiment

[0046] A single random experiment of the dynamic ER Rule classifier includes four steps: evidence acquisition, evidence activation and combination, parameter optimization, and optimal result update.

[0047] (1) Evidence acquisition

[0048] First, determine each feature x i Reference value The number of reference values for each feature J i The random value is an integer between 2 and 9, J={J1,J2,…,J M}Save the reference value quantity of all features. Then, A i The other values in are determined using an arithmetic progression.

[0049] Secondly, for a particular eigenvalue Its reference value A can be obtained by the following formula i Similarity distribution of where α i,j Represents eigenvalues With reference value Matching degree:

[0050]

[0051]

[0052]

[0053] Next, calculate all sample pairs (x k ,y k ) Regarding the matching degree of each feature reference value, we can get feature x i The projection matrix. Each element a in this matrix n,j It means the category is y n And its eigenvalue With reference value The sum of the matching degrees of the matched sample pairs.

[0054] Then, the category of the input sample is known to be y n hour, Considered as a reference value The reliability is expressed as

[0055]

[0056] Finally, the descriptive attribute x can be obtained by the following formula i With class y n The confidence matrix of the relationship between It is an element in the belief matrix, which represents the attribute value of the hypothetical input sample Take reference value When , the sample is considered to belong to class y n Reliability:

[0057]

[0058] Thus, the attribute reference value can be obtained The corresponding weight is Reliability is r i Evidence The evidence The initial weight Set to r i Equal values, and then optimize the weights.

[0059] (2) Evidence activation and combination

[0060] For the sample if The value is in the interval Inside, then the attribute will activate the reference value and Two corresponding adjacent pieces of evidence and By the attribute Evidence obtained i for:

[0061] e i ={(y n , p n，i ), n=1, 2, ..., N}

[0062]

[0063] Here, p n,i Indicates the attribute value of the sample Activation Evidence and When , the sample is considered to belong to class y n The reliability, α i,j and α i,j+1 Calculated according to the above formula. Similarly, evidence e i The weight w i Available through and The weighted sum of is calculated as:

[0064]

[0065] Thus, we can obtain M pieces of evidence e1, e2, ..., e about M attributes. M, and then use ER Rule (inference evidence rule) to combine them and get the following fusion result R(x k ), the sample x k Belong to The category y corresponding to the maximum value in n .

[0066] R(x k )={(y n , p n，e(M) ), n=1, 2, ..., N}

[0067] At this point, the initial ER Rule classifier is constructed. Through the above calculation process, the fusion result of each sample can be obtained, and then their categories can be determined.

[0068] (3) Parameter optimization

[0069] The initial classifier involves some initial parameters, including the initial reference value of each attribute and the initial weight of each evidence, which need to be optimized using appropriate methods to improve the accuracy and other performance of the classifier. Indicates the set of parameters that need to be optimized, and the remaining parameters and They are all fixed endpoint values and need to meet the following constraints:

[0070]

[0071]

[0072] The mean square error (MSE) between the fusion result and the actual category is used as the objective function for parameter optimization:

[0073]

[0074] Among them, d E Represents the fusion result and sample x k The Euclidean distance between the actual categories. Reference vector V k It is directly obtained from the actual category of the sample. For example, for a three-category problem, if the sample x k Actually belongs to the second category, then V k =(0,1,0).

[0075] A variety of optimization algorithms can be used to solve the above optimization problem, such as genetic algorithm, ant colony optimization algorithm, particle swarm optimization algorithm, etc. Different optimization algorithms have different application ideas. The present invention uses a relatively common swarm intelligence optimization algorithm, particle swarm optimization algorithm (PSO).

[0076] (4) Optimal result update

[0077] After completing a single randomized experiment, compare the current experimental results with the previous results to determine whether the results have improved. The objective function of this step is to maximize the test set accuracy, which is the most concerned indicator in balanced dataset classification problems.

[0078]

[0079] The denominator is the total number of samples in the test set, and the numerator is the number of samples in the test set whose predicted category is the same as the actual category.

[0080] The result update method is: if the test set accuracy of this experiment is higher than that of the previous experiment, then this experiment is more ideal, and the stored optimal reference value quantity experimental results need to be updated; otherwise, the currently stored optimal results remain unchanged.

[0081] Step 3: Determination of optimal results

[0082] Step 2 is repeated multiple times to achieve high-precision results, resulting in multiple sets of results. The maximum number of iterations is denoted by iter. Determining the optimal attribute reference value requires a trade-off between computational time complexity and classification accuracy. In this paper, the vector J with the highest test set accuracy and the lowest MSE is selected as the optimal result.

[0083] As mentioned above, the information contained in the data is compressed into the parameters of the ER Rule classifier. In general, there are four types of parameters in the dynamic ER Rule classifier: feature f i Reliability i , reference value number J, feature f i Reference values and evidence The weight w i , these parameters are related to each feature f i It is worth noting that J determines the structure of the ER Rule classifier and has an important impact on the time complexity and accuracy. For the purpose of subsequent explanation, J is defined as a structural parameter, except for the reliability r i The remaining parameters are defined as general parameters.

[0084] For feature f i Reliability i, which is the characteristic f i The essential attribute of represents the classification ability of the information source, which is completely calculated based on the newly arrived data stream to reflect the essential characteristics of the acquired instance. Therefore, during the data stream learning process, the optimization of reliability will not be considered.

[0085] When building an ER Rule classifier, optimizing general parameters is almost essential to achieve higher accuracy. However, the structural parameters of each feature influence the structure of the ER Rule classifier and, in particular, determine each feature's reference value. To ensure accuracy, changing the structural parameters also requires optimizing the general parameters. Therefore, updating the structural parameters essentially amounts to rebuilding a new ER Rule classifier. Therefore, when updating the ER Rule classifier, updating the general parameters is prioritized, with the structural parameters being the last option. This approach minimizes the optimization effort and time complexity.

[0086] Finally, it is worth noting that during the parameter optimization step, the first and last reference values for each feature remain fixed. Clearly, these two special reference values constitute the value range of a feature, which is the most direct information a dataset provides about each feature. Furthermore, these two reference values also determine the search range for other adjustable reference values. Therefore, in the proposed strategy described in the following embodiments, the range of each feature will be updated at each stage to improve the usability of the classification method for a generalized feature space data stream proposed by the present invention.

[0087] Figure 2 The three generalization problems are defined as Problem I (active is empty, new is not empty), Problem II (active and new are not empty), and Problem III (active is not empty, new is empty).

[0088] Figure 3 It is a general framework for the classification method of generalized feature space data stream. t It is a matrix used to store the relevant parameters of the ER Rule classifier and the activeNum of each feature in the t stage, par t It is updated at each stage to preserve the latest knowledge. activeNum refers to the number of times an existing feature appears repeatedly in subsequent stages. This method considers the reappearance of an existing feature as its activation. Therefore, activeNum refers to the activation frequency of a feature, and activeNum = 0 indicates the first appearance of the feature.

[0089] The classification method of the generalized feature space data stream is specifically as follows:

[0090] For the data stream (X0, Y0) arriving at stage 0, feature selection is performed first, and then a dynamic evidence reasoning rule classifier is constructed based on the selected features to obtain the accuracy acc0 and classifier parameters par0 of stage 0, and update the global feature space of stage 0.

[0091] For the original data of stage 0

[0092] For the data stream arriving at stage t (X t ,Y t ), the processing steps are as follows:

[0093] The first step is to select features and update (X t ,Y t ), then the features selected in stage t and the global feature space U in stage t-1 are t-1 The features stored in are divided into corresponding categories: new feature set, active feature set and silent feature set. The new feature set is the features introduced in stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear in stage t.

[0094] In the second step, when the active feature set is empty, we need to solve problem I and use the features selected in stage t to build a dynamic evidence reasoning rule classifier to obtain the accuracy acc t and classifier parameters par t .

[0095] Problem I: Only new features

[0096] For Question I, active=null,silent=U t-1 The current stage only has new features, that is, |new| = d t ,In this case, the same scheme as stage 0 is adopted, that is, a new dynamic ERRule classifier is directly constructed using the current data, that is, the parameters of the new features are completely relearned.

[0097] The third step is to update the activation frequency activeNum of each feature in the active feature set when the active feature set is not empty, and to adjust the activation frequency according to the classifier parameter par of stage t-1. t-1 Construct a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty.

[0098] Update activeNum: At stage t, if one is already in U t-1The features in X t If it reappears in par, the feature is considered to be activated. Accordingly, its activeNum will be t-1 Add 1 to the specific value in , and each feature in the active set needs to complete this process.

[0099] Extracting existing knowledge of active features: As mentioned above, active features have appeared in the previous stage, so in par t-1 There is relevant knowledge that can be utilized. Represents X t The value of the active feature in , where Indicates the number of active features. First, from par t-1 Extract the parameters of active features from the Matching ER Rule classifier C a ; Then, based on C a and X can be calculated t The initial prediction result predict, at the same time, you can also get X t The initial accuracy acc t _raw and the fusion result of each instance.

[0100] Step 4: If the active feature set is not empty and the new feature set is not empty, we will proceed to Problem II, which is to combine the initial fusion result with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters.

[0101] Problem II: Both new and active features

[0102] For Question II, silent=U t-1 -active, In the process of solving this problem, we need to pay attention to how to notify the learning of active features and new features. Obviously, X t It consists of two parts, and in, is the number of new features, satisfying Since the ER Rule classifier C introduced above aOnly based on active features, then The output results can reflect the knowledge they contain, and the data format of the output results is better than the original data. More standardized.

[0103] Step 1: To inherit the learned knowledge, the results will be output As Then, fusion is combined with Merge, X t will be replaced by

[0104] Step 2: Based on Construct a new dynamic ER Rule classifier. The trained parameters consist of two parts, one related to the compact representation fusion and the other related to the new features. The former will not be used again, and the latter will be stored in the matrix par t Therefore, in essence, the information of the new feature is learned with the help of the compressed information of the active feature.

[0105] Step 3: Finally, update par according to the value range of the active feature in the current stage t-1 The first and last reference values of the active features stored in . For example, suppose feature f1 is activated at stage t, par t-1 The first and last reference values of f1 stored in are 0 and 5 respectively, that is, the range of f1 is [0,5]. t The range of f1 is [0,10], par t The last reference value of f1 will become 10 and the first value will remain unchanged; if X t The range of f1 is [-1,4], par t The first reference value of f1 in will become -1. In short, these two special reference values will be replaced by the union of the current range and the old range, which can ensure the generalization of the classifier.

[0106] Step 5: If the active feature set is not empty and the new feature set is empty, we need to solve problem III and use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t .

[0107] Problem III: Only active features

[0108] For Problem III, new = null, Obviously, the features of the current stage are composed only of active features, that is, |active| = d t In this case, we need to consider how to utilize and update existing knowledge. Obviously, the extracted existing knowledge is not consistent with the entire dataset X. t Related, because Therefore, the ER Rule classifier C a and its output is exactly the same as X t Match. If C a Calculated acc t If _raw is larger than minAcc, it means that the distribution of the current data is similar to that of the previous stage, and there is no need to update the classifier. Otherwise, a certain strategy is needed to update the existing knowledge.

[0109] Step 1: In this problem, the output fusion is still used as a compact representation. That is, the dataset X t Replace with

[0110] Step 2: Based on Train a new dynamic ER Rule classifier to get the updated accuracy acc t If acc t Less than acc t _raw, it means that the current stage of compact representation is not as effective as the previous related parameters, then acc t will be equal to acc t _raw.

[0111] Step 3: Update X according to the principles described in Step 3 of Problem II t The first and last reference values of all features in .

[0112] Step 6: According to the accuracy acc t and the classifier parameter par t , update the global feature space of stage t.

[0113] After the learning process is completed, the accuracy rate acc can be obtained t and the updated parameter par t , and update the global feature set U according to the parameters selected in step 1 t If necessary, subsequent feature selection will be performed to deal with the hazards brought by the increasing number of features.

[0114] In general, the main idea of this method is to use the output of the classifier constructed by the active feature as the compressed information of the response feature in the current data.

[0115] As data streams continue to arrive, the number of global features will gradually increase, putting greater pressure on system storage and computing loads. At the same time, some redundant or irrelevant features may be introduced. In addition, too many features may cause overfitting. Therefore, it is necessary to develop feature selection strategies in generalized feature space data stream learning to retain as many features as possible. Let totalF represent the maximum number of features the system can handle, such as Figure 3 As shown in Figure 3, the proposed feature selection strategy is executed when the data stream just arrives and when the learning is finished.

[0116] For the original data (X0, Y0), when the number of features X0 (i.e., d0) is greater than totalF, an effective feature selection algorithm is needed to select seleNum valuable features. This embodiment adopts the MRMR algorithm, which is widely used in many fields. The main idea of this algorithm is to select features with the greatest correlation with the class labels and the least redundancy between related features.

[0117] For the data of stage t (X t ,Y t ), the feature selection strategy consists of two parts. On the one hand, in learning (X t ,Y t ) before, if X t The number of features in (ie d t ) is greater than 3, then use MRMR to select the first seleNum features seleNum=min(round(d t *p),totalF), where p∈(0,1] is the proportion of features that will be selected from the current data, round(d t *p) indicates the t *p is rounded off.

[0118] On the other hand, after the learning process is completed and the global feature space (universal) is updated, if the total feature number of the global feature space |universal| is greater than totalF, then the features contained in universal except new will be used for subsequent feature selection, and the features to be deleted are defined as delNum = |universal|-totalF. If the total feature number of the silent feature set silent| is greater than delNum, subsequent feature selection will only be performed in the silent set, otherwise, the active set will also participate in the selection process. The basic idea is to delete a certain number of features with smaller activeNum, because the smaller the activeNum, the fewer times the feature has been activated so far; when some features have the same activeNum, their weights will be further compared, and the features with smaller weights will be deleted because the weight of each feature represents its relative importance in the classification process.

[0119] Obviously, once feature selection is performed before learning, the initial model building and ER Rule-GFS learning strategy will be performed only on the basis of the selected features. At the same time, if subsequent feature selection is performed, the universal needs to be updated again before entering the next stage.

[0120] The pseudo code of a classification method for generalized feature space data stream based on ER Rule (ER Rule-GFS) is shown in Table 2, which also explains its implementation steps.

[0121] Table 2 Algorithm 1

[0122]

[0123]

[0124] The present invention proposes a classification device for generalized feature space data stream, comprising:

[0125] Initialization module, for the data stream arriving at stage 0, first performs feature selection, then builds a dynamic evidence reasoning rule classifier based on the selected features, obtains the accuracy acc0 and classifier parameters par0 of stage 0, and updates the global feature space of stage 0;

[0126] The question classification module first performs feature selection on the data stream arriving at stage t, and then classifies the features selected at stage t and the features stored in the global feature space at stage t-1 into a new feature set, an active feature set, and a silent feature set, respectively. The new feature set is the features introduced at stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear at stage t.

[0127] In the problem I processing module, when the active feature set is empty, the dynamic evidence reasoning rule classifier is constructed using the features selected in stage t, and the accuracy acc is obtained. t and classifier parameters par t ;

[0128] Problem II and Problem III preprocessing modules, when the active feature set is not empty, update the activation frequency of each feature in the active feature set and adjust the activation frequency according to the classifier parameter par in stage t-1 t-1 Build a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty;

[0129] Problem II processing module, if the active feature set is not empty and the new feature set is not empty, then first merge the initial fusion result with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters;

[0130] Problem III processing module, if the active feature set is not empty and the new feature set is empty, use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t ;

[0131] Update module, according to the accuracy acc t and the classifier parameter par t , update the global feature space of stage t.

[0132] While embodiments of the present invention have been described above, these embodiments do not exhaustively describe all details and do not limit the invention to the specific embodiments described. Numerous modifications and variations are possible based on the above description. These embodiments are selected and described in detail herein to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to utilize the present invention and its modifications and uses. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A classification method for generalized feature space data stream, characterized in that: The method comprises: First, feature selection is performed on the data stream arriving at stage t. Then, the features selected at stage t and the features stored in the global feature space at stage t-1 are classified into a new feature set, an active feature set, and a silent feature set, respectively. The new feature set is the features introduced at stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear at stage t. When the active feature set is empty, the dynamic evidence reasoning rule classifier is constructed using the features selected in stage t to obtain the accuracy acc t and classifier parameters par t ; When the active feature set is not empty, the activation frequency of each feature in the active feature set is updated, and the classifier parameter par in stage t-1 is used. t-1 Build a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty; If the active feature set is not empty and the new feature set is not empty, the initial fusion result is first merged with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters; If the active feature set is not empty and the new feature set is empty, use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t ; According to the accuracy acc t and the classifier parameter par t , update the global feature space of stage t.

2. The classification method of generalized feature space data stream according to claim 1, characterized in that: If the number of features contained in the global feature space of the updated stage t is greater than the maximum number of features that the system can bear, and the number of features contained in the silent feature set is greater than the number of features that need to be deleted, then the feature selection of stage t+1 will only be performed in the silent feature set. Otherwise, the active feature set will also participate in the feature selection process.

3. The classification method of generalized feature space data stream according to claim 2, characterized in that: The active feature set also participates in the feature selection process, specifically deleting a certain number of features with low activation frequencies. When some features have the same activation frequency, the features with smaller weights are deleted.

4. The classification method for generalized feature space data stream according to any one of claims 1 to 3, characterized in that: Specifically, the feature selection is to use a maximum relevance minimum redundancy algorithm to select required features when the number of features in the data stream is greater than the maximum number of features that the system can bear.

5. The classification method for generalized feature space data stream according to any one of claims 1 to 3, characterized in that: The dynamic evidence inference rule classifier first obtains the reliability of the feature through feature reliability calculation, and then obtains the number of reference values, the reference value of the feature, and the weight of the evidence by repeating the four steps of evidence acquisition, evidence activation and combination, parameter optimization, and optimal result update. The number of reference values uses the vector with the maximum test set accuracy and the minimum mean square error as the optimal result.

6. The classification method of generalized feature space data stream according to claim 5, characterized in that: When the dynamic evidence reasoning rule classifier is updated, the reference value of the feature and the weight of the evidence are updated first.

7. The classification method for generalized feature space data stream according to any one of claims 1 to 3, characterized in that: For the data stream arriving at stage 0, feature selection is performed first, and then a dynamic evidence reasoning rule classifier is constructed based on the selected features to obtain the accuracy acc0 and classifier parameters par0 of stage 0, and update the global feature space of stage 0.

8. A classification device for generalized feature space data stream, characterized in that: The device comprises: Initialization module, for the data stream arriving at stage 0, first performs feature selection, then builds a dynamic evidence reasoning rule classifier based on the selected features, obtains the accuracy acc0 and classifier parameters par0 of stage 0, and updates the global feature space of stage 0; The question classification module first performs feature selection on the data stream arriving at stage t, and then classifies the features selected at stage t and the features stored in the global feature space at stage t-1 into a new feature set, an active feature set, and a silent feature set, respectively. The new feature set is the features introduced at stage t, the active feature set is the features introduced and stored before stage t, and the silent feature set is the features that appeared before stage t but did not reappear at stage t. In the problem I processing module, when the active feature set is empty, the dynamic evidence reasoning rule classifier is constructed using the features selected in stage t, and the accuracy acc is obtained. t and classifier parameters par t ; Problem II and Problem III preprocessing modules, when the active feature set is not empty, update the activation frequency of each feature in the active feature set and adjust the activation frequency according to the classifier parameter par in stage t-1 t-1 Build a dynamic evidence reasoning rule classifier that matches the active feature set, obtain the initial accuracy and initial fusion results of the data stream in stage t, and then determine whether the new feature set is empty; Problem II processing module, if the active feature set is not empty and the new feature set is not empty, then first merge the initial fusion result with the new feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t , the classifier parameter par t Store the parameters related to the new features in the trained parameters; Problem III processing module, if the active feature set is not empty and the new feature set is empty, use the initial fusion result to replace the active feature set to obtain Then based on Construct a dynamic evidence reasoning rule classifier and get the accuracy acc t and classifier parameters par t ; Update module, according to the accuracy acc t and the classifier parameter par t , update the global feature space of stage t.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images