An approximate query processing method based on incremental learning to solve concept drift

Abstract

The application provides an approximate query processing method for solving concept drift based on incremental learning, comprising: obtaining a query sample set, dividing the query sample set to obtain a plurality of query sample subsets; uniformly sampling each query sample subset to construct a corresponding training set; training two BNNs (Bayesian neural networks) using each training set to construct an initial posterior mixed Gaussian distribution model; using an EM algorithm to approximate the initial posterior mixed Gaussian distribution model to obtain an estimated posterior mixed Gaussian distribution model; when a new query sample data set arrives in the database, the estimated posterior mixed Gaussian distribution model is updated; and according to the final estimated posterior mixed Gaussian distribution model and a user's query statement, a query result of the user is returned, so that the application can well cope with the query of newly inserted data, avoid forgetting the old data, and save the storage resource overhead.

Description

Technical Field

[0001] This invention belongs to the field of approximate query algorithms, and in particular relates to an approximate query processing method based on incremental learning to solve concept drift. Background Technology

[0002] Currently, databases play a fundamental role in various computer applications, storing massive amounts of critical data. Furthermore, with the internet increasingly permeating people's lives, more and more people are generating data anytime, anywhere, leading to a rapid expansion of data volume. How to efficiently manage and analyze such massive amounts of data is one of the major challenges facing databases today. Query processing has always been a focus of attention. Faced with huge amounts of data and complex conditions, accurate results take a long time to be provided, making it difficult to achieve fast interactive responses. Excessive response time is unacceptable for certain users and applications. If the requirement for accurate results is slightly relaxed, allowing for a certain amount of error in the query results, the speed of query processing can usually be effectively improved. This fact forms the premise for the development of approximate query technology.

[0003] With the development of artificial intelligence, machine learning methods have been gradually applied to the field of approximate query. Unlike traditional methods, it builds a model by learning from data or sampled data, and uses the model to approximate the answer to the user query. Traditional query requires maintaining a large number of samples or pre-computation to maintain the accuracy of the query, which leads to the consumption of space resources. However, machine learning-based methods abandon this concept. They do not need to maintain samples or spend resources on pre-computation. The learned data can be discarded, and only a small model needs to be maintained, which reduces the consumption of memory resources.

[0004] Currently, AQP based on machine learning has made significant progress in terms of time, accuracy, and performance. However, it still faces many challenges. Experts and scholars are optimizing query performance, time, and memory overhead, but they often overlook the diversity of datasets. There are static local datasets as well as constantly growing data streams. In big data environments such as the Industrial Internet, time-series data such as temperature changes, water temperature detection, and pH values ​​are characterized by continuous updates, irregularity, and rapid updates. Therefore, they are prone to significant differences in distribution from the original data, leading to concept drift. The different distribution of new data from the original data can cause offline models to become unsuitable. One solution is to periodically retrain the model to cope with data updates, but this approach leads to significant resource and cost overhead. Existing models lack good adaptability to large amounts of new data. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides an approximate query processing method based on incremental learning to solve concept drift. It employs the continuous learning method of Bayesian Learning Networks (CBLNs), enabling the network to allocate additional resources to learn the distribution of newly inserted data, avoiding "catastrophic forgetting" of previously learned tasks, and eliminating the need to access past data. This allows the ML model to adapt to changes in data distribution. By using Bayesian neural network methods, the amount of resources used to learn different data is optimized, avoiding exponential growth of parameters. This method can effectively handle queries with newly inserted data, avoids forgetting old data, and avoids re-accessing the original data, saving storage resource overhead.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0007] S1: Obtain the query sample set, where each query sample in the query sample set includes: multiple attribute information;

[0008] S2: Divide the query sample set into multiple query sample subsets based on the first attribute of the query sample in the query sample set;

[0009] S3: For each subset of query samples, uniform sampling is performed to construct a corresponding training set; and two BNNs Bayesian neural networks are trained using each training set to construct a BNNs density estimation model for the second attribute of the query samples in the training set and a BNNs regression model from the second attribute to the first attribute.

[0010] S4: Construct an initial posterior Gaussian mixture model based on the BNNs density estimation model of the second attribute of all query samples in the training set and the BNNs regression model from the second attribute to the first attribute.

[0011] S5: The EM algorithm is used to approximate the initial posterior Gaussian mixture model to obtain the estimated posterior Gaussian mixture model.

[0012] S6: When a new query sample dataset arrives in the database, the estimated posterior mixture Gaussian distribution model is updated according to steps S2 to S5.

[0013] S7: Return the user's query results based on the final estimated posterior mixture Gaussian distribution model and the user's query statement.

[0014] Preferably, the step of partitioning the query sample set according to the first attribute of the query samples in the query sample set includes: partitioning the query sample set according to the value range of the first attribute of the query samples in the query sample set; dividing the query samples in the query sample set into K query sample subsets according to the value of the first attribute, denoted as D = {D1, D2, ..., D...} k,..D K}, D k This represents the k-th subset of query samples.

[0015] Preferably, step S3 includes:

[0016] S31: For the query sample subset D k The training set S is obtained by uniform sampling. k Using training set S k Train a Bayesian neural network (BNN) to construct a BNN density estimation model D for the second attribute of the query samples in the training set. k (x), Model D k The input to (x) is the second attribute of the query sample in the training set, and the model D k The output of (x) is the density estimate of the second attribute of the query sample in the training set;

[0017] S32: Using the training set S k Train another Bayesian neural network (BNN) to construct a BNN regression model R that maps the second attribute of query samples in the training set to the first attribute. k (x), Model R k The input to (x) is the second attribute of the query sample in the training set, and the model R k The output of (x) is the first attribute of the query sample in the training set.

[0018] Preferably, the initial posterior mixture Gaussian distribution model includes:

[0019]

[0020] α = 1 / K

[0021] b k =(D k (x),R k (x))

[0022] Among them, b 1:K Let b represent the initial posterior mixture Gaussian distribution model. k Let α represent the initial posterior Gaussian distribution model corresponding to the k-th training set, and let α represent the weight parameters.

[0023] Preferably, step S5 includes:

[0024] S51: Sample N observation data points from the posterior mixture distribution of the initial posterior mixture Gaussian distribution model;

[0025] S52: In the E-step of the EM algorithm, calculate the sub-model b based on the parameters of the initial posterior Gaussian mixture model. k Response to observed data:

[0026]

[0027] Where, γ jk Representing the sub-model b k The response degree to the j-th observation data, μ k and Representing the sub-model b k Parameters; α k Representing the sub-model b k Weight parameters;

[0028] S53: In the M-step of the EM algorithm, when calculating the new iteration, sub-model b k Parameters:

[0029]

[0030]

[0031]

[0032] in, and Indicates the updated sub-model b k The parameters, Indicates the updated sub-model b k Weight parameters;

[0033] S54: Repeat steps S52 and S53 until convergence, yielding an approximate posterior Gaussian mixture distribution model:

[0034]

[0035]

[0036] in, This represents an approximate posterior mixture Gaussian distribution model;

[0037] S55: For the posterior Gaussian mixture distribution model Merge similar sub-models and perform posterior mixture Gaussian distribution model. Redundant sub-models are removed to obtain the estimated posterior Gaussian mixture distribution model b. 1:n .

[0038] Preferably, step S55 includes:

[0039] S551: Set a threshold When the model The sub-model in Weight parameters When it is less than t1, then the view split model These are redundant components, which will be used in the sub-model. delete;

[0040] S552: After executing step S551, for the model Any two remaining sub-models and Calculation of the sub-model and Mahalanobis distance Set the threshold t2 = 3, if Then the sub-model and By merging, we obtain the estimated posterior mixture Gaussian distribution model.

[0041] Preferably, the sub-model and Mahalanobis distance include:

[0042]

[0043] in, and Representation Model The i-th sub-model The parameters, and Representation Model The j-th sub-model The parameters.

[0044] Preferably, the sub-model and The merger includes:

[0045]

[0046]

[0047]

[0048] Where μ and σ 2 Representation of the sub-model and The parameters of the merged model, where α represents the sub-model. and The weights of the merged model.

[0049] Preferably, step S6 includes: whenever a new query sample dataset D is available in the database... new Upon arrival; use the same rules as in step S2 to transfer D new Divided into K * A subset of query samples were obtained. The training set is obtained by sampling each subset of query samples using a uniform sampling method. b′ is obtained by training a new BNNs model on each training set. k =(D k ′(x),R k ′(x));

[0050] Assign b′ to each sub-model k and the original model b old =b 1:n ,by The initial weights constitute the initial posterior mixture distribution model b. 1:a a = K * +1:

[0051]

[0052]

[0053] The updated estimated posterior mixture Gaussian distribution model b is obtained by performing step S5. 1:m .

[0054] The present invention has at least the following beneficial effects

[0055] This invention employs the continuous learning method of Bayesian Learning Networks (CBLNs), enabling the network to allocate additional resources to learn the distribution of newly inserted data and avoiding "catastrophic forgetting" of previously learned tasks. Furthermore, it eliminates the need to access past data, allowing the ML model to adapt to changes in data distribution. By using Bayesian neural network methods, the amount of resources used to learn different data is optimized, avoiding exponential growth of parameters. This method effectively handles queries for newly inserted data, avoids forgetting old data, and avoids re-accessing the original data, thus saving storage resource overhead. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0057] Figure 2 This is a schematic diagram illustrating the synthesis of the posterior mixture Gaussian distribution model of the present invention;

[0058] Figure 3 This is a schematic diagram of the incremental learning process model update of the present invention. Detailed Implementation

[0059] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0060] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0061] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0062] Please see Figure 1 This invention provides an approximate query processing method based on incremental learning to solve concept drift, comprising:

[0063] S1: Obtain the query sample set, where each query sample in the query sample set includes: multiple attribute information;

[0064] To facilitate understanding, this embodiment provides a representation of the query sample set, where multiple query samples are stored in a table format in the database, as shown in Table 1:

[0065] Table 1 Query Sample Set

[0066] Sample 1 10.4 321 2019.10.01 A12001 S22120 Sample 2 11.34 200 22019.11.23 A13024 S23014 Sample 3 12.56 356 2019.12.4 B32012 S45104 Sample 4 56.23 1043 2020.01.25 A32012 S34102 Sample 5 68.34 1145 2020.02.23 B13013 S35102 … … … … … …

[0067] In Table 1, the attribute information of the query sample includes: price, quantity, sales date, employee number, and product ID. The query sample is first processed through word embedding and then stored in the database.

[0068] S2: Divide the query sample set into multiple query sample subsets based on the first attribute of the query sample in the query sample set;

[0069] Preferably, the step of partitioning the query sample set according to the first attribute of the query samples in the query sample set includes: partitioning the query sample set according to the value range of the first attribute of the query samples in the query sample set; dividing the query samples in the query sample set into K query sample subsets according to the value of the first attribute, denoted as D = {D1, D2, ..., D...} k ,..D K}, D k This represents the k-th subset of query samples.

[0070] The present invention will be further illustrated using Table 1 as an example. In Table 1, the first attribute of the query sample is represented by any one of price, quantity, sales date, employee number, and product ID. Taking the product quantity as an example, it can be seen from the table that the product quantity will vary for different time periods. In this embodiment, the query sample will be divided according to the range of product quantity values. The samples with a value range of 1-500 will be divided into a query sample subset, which includes sample 1, sample 2, and sample 3. The samples with a value range of 1000-1500 will be divided into a query sample subset, which includes sample 4 and sample 5. This is only a further illustration of the present invention. In the actual training process, the number of samples is very large. Those skilled in the art can divide the query sample according to the actual situation to obtain multiple query sample subsets, which will not be elaborated further here.

[0071] S3: For each subset of query samples, uniform sampling is performed to construct a corresponding training set; and two BNNs Bayesian neural networks are trained using each training set to construct a BNNs density estimation model for the second attribute of the query samples in the training set and a BNNs regression model from the second attribute to the first attribute.

[0072] In this embodiment, each query sample set D is first processed. k The training set S is obtained by uniform sampling. k In this embodiment, the first attribute and the second attribute belong to two different attributes in the query sample. In this embodiment, the first attribute is denoted as y and the second attribute is denoted as x. Then, through the training set S k The density estimation model D of the BNNs for the second attribute x is obtained by training the BNNs Bayesian neural network. k (x); through the training set S k Training another BNN (Bayesian Neural Network) yields a BNN regression model R that maps the second attribute x to the first attribute y. k(x), Model R k The input to (x) is the second attribute x of the query sample in the training set, and the model R k The output of (x) is the first attribute y of the query sample in the training set.

[0073] Preferably, step S3 includes:

[0074] S31: For the query sample subset D k The training set S is obtained by uniform sampling. k Using training set S k A BNN (Bayesian Neural Network) is trained to construct a BNN density estimation model D for the second attribute of query samples in the training set. k (x), Model D k The input to (x) is the second attribute of the query sample in the training set, and the model D k The output of (x) is the density estimate of the second attribute of the query sample in the training set;

[0075] S32: Using the training set S k Train another Bayesian neural network (BNN) to construct a BNN regression model R that maps the second attribute of the query samples in the training set to the first attribute. k (x), Model R k The input to (x) is the second attribute of the query sample in the training set, and the model R k The output of (x) is the first attribute of the query sample in the training set.

[0076] Please see Figure 2 S4: Construct an initial posterior mixture Gaussian distribution model based on the BNNs density estimation model of the second attribute of all query samples in the training set and the BNNs regression model from the second attribute to the first attribute.

[0077] Preferably, the initial posterior mixture Gaussian distribution model includes:

[0078]

[0079] α = 1 / K

[0080] b k =(D k (x),R k (x))

[0081] Among them, b 1:K Let b represent the initial posterior mixture Gaussian distribution model. k Let represent the initial posterior Gaussian distribution model corresponding to the k-th training set, and α represent the weight parameters.

[0082] Please see Figure 3S5: The estimated posterior Gaussian mixture model is obtained by approximating the initial posterior Gaussian mixture model using the EM algorithm.

[0083] Preferably, step S5 includes:

[0084] S51: Sample N observation data points from the posterior mixture distribution of the initial posterior mixture Gaussian distribution model;

[0085] S52: In the E-step of the EM algorithm, calculate the sub-model b based on the parameters of the initial posterior Gaussian mixture model. k Response to observed data:

[0086]

[0087] Where, γ jk Representing the sub-model b k The response degree to the j-th observation data, μ k and Representing the sub-model b k Parameters; α k Representing the sub-model b k Weight parameters;

[0088] S53: In the M-step of the EM algorithm, when calculating the new iteration, sub-model b k Parameters:

[0089]

[0090]

[0091]

[0092] in, and Indicates the updated sub-model b k The parameters, Indicates the updated sub-model b k Weight parameters;

[0093] S54: Repeat steps S52 and S53 until convergence, yielding an approximate posterior Gaussian mixture distribution model:

[0094]

[0095]

[0096] in, This represents an approximate posterior mixture Gaussian distribution model;

[0097] S55: For the posterior Gaussian mixture distribution model Merge similar sub-models and perform posterior mixture Gaussian distribution model. Redundant sub-models are removed to obtain the estimated posterior Gaussian mixture distribution model b. 1:n .

[0098] Preferably, step S55 includes:

[0099] S551: Set a threshold When the model The sub-model in Weight parameters When it is less than t1, then the view split model These are redundant components, which will be used in the sub-model. delete;

[0100] S552: For the model Any two sub-models and Calculation of the sub-model and Mahalanobis distance Set the threshold t2 = 3, if Then the sub-model and By merging, we obtain the estimated posterior mixture Gaussian distribution model.

[0101] Preferably, the sub-model and Mahalanobis distance include:

[0102]

[0103] in, and Representation Model The i-th sub-model The parameters, and Representation Model The j-th sub-model The parameters.

[0104] Preferably, the sub-model and The merger includes:

[0105]

[0106]

[0107]

[0108] Where μ and σ 2 Representation of the sub-model and The parameters of the merged model, where α represents the sub-model. and The weights of the merged model.

[0109] S6: When a new query sample dataset arrives in the database, the estimated posterior mixture Gaussian distribution model is updated according to steps S2 to S5.

[0110] Preferably, step S6 includes: whenever a new query sample dataset D is available in the database... new Upon arrival; use the same rules as in step S2 to transfer D new Divided into K * A subset of query samples were obtained. The training set is obtained by sampling each subset of query samples using a uniform sampling method. b′ is obtained by training a new BNNs model on each training set. k =(D k ′(x),R k ′(x));

[0111] Assign b′ to each sub-model k and the original model b old =b 1:n ,by The initial weights constitute the initial posterior mixture distribution model b. 1:a a = K * +1:

[0112]

[0113]

[0114] The updated estimated posterior Gaussian mixture model b is obtained by performing the same approximation rule as in step S5. 1:m In step S5, b 1:a See it as b 1:K The sub-model b in step S5 k See as b′ k or b old Those skilled in the art should understand the equivalent substitution relationship here, and the final result is b. 1:m Equivalent to b in step S5 1:n b 1:m For the updated b 1:n .

[0115] S7: Return the user's query results based on the estimated posterior mixture Gaussian distribution model and the user's query statement.

[0116] In this embodiment, the user's query statement can be represented as:

[0117] SELECT AF(y) FROM tbl

[0118] WHERE x BETWEEN low AND high

[0119] The user inputs the above query statement, and the SQL parses out the key fields corresponding to the query statement. Among them, AF represents the aggregate query type, x is the query condition (second attribute), low and high represent the value range of x, and y represents the query result attribute.

[0120] An approximate result is obtained by calculating the expression AF(y) = Query(R(x), D(x)), where Query is a formula used for different calculations based on different AF values. For example, if AF = SUM, then the approximate result is obtained by: This will give you an approximate result and return the user's query result.

[0121] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An approximate query processing method based on incremental learning to solve concept drift, characterized in that, include: S1: Obtain the query sample set, wherein each query sample in the query sample set includes: multiple attribute information; the attribute information of the query sample includes: price, quantity, sales date, employee number, and product ID; S2: Divide the query sample set into multiple query sample subsets based on the first attribute of the query sample in the query sample set; S3: For each subset of query samples, uniform sampling is performed to construct a corresponding training set; and two BNNs Bayesian neural networks are trained using each training set to construct a BNNs density estimation model for the second attribute of the query samples in the training set and a BNNs regression model from the second attribute to the first attribute. Step S3 includes: S31: For the query sample subset The training set is obtained by uniform sampling. Using the training set Train a Bayesian neural network (BNNs) to construct a BNNs density estimation model for the second attribute of query samples in the training set. ,Model The input is the second attribute of the query sample in the training set, and the model... The output is a density estimate of the second attribute of the query sample in the training set; S32: Utilizing the training set Train another Bayesian neural network (BNN) to construct a BNN regression model that maps the second attribute of query samples in the training set to the first attribute. ,Model The input is the second attribute of the query sample in the training set, and the model... The output is the first attribute of the query sample in the training set; S4: Construct an initial posterior Gaussian mixture model based on the BNNs density estimation model of the second attribute of all query samples in the training set and the BNNs regression model from the second attribute to the first attribute. The initial posterior mixture Gaussian distribution model includes: in, This represents the initial posterior mixture Gaussian distribution model. Indicates the first The initial posterior Gaussian distribution model corresponding to each training set. This represents the weight parameter; K represents the number of query sample subsets. S5: The EM algorithm is used to approximate the initial posterior Gaussian mixture model to obtain the estimated posterior Gaussian mixture model. S6: When a new query sample dataset arrives in the database, the estimated posterior mixture Gaussian distribution model is updated according to steps S2~S5; S7: Return the user's query results based on the final estimated posterior mixture Gaussian distribution model and the user's query statement.

2. The approximate query processing method based on incremental learning to solve concept drift according to claim 1, characterized in that, The step of partitioning the query sample set according to the first attribute of the query samples in the query sample set includes: partitioning the query sample set according to the value range of the first attribute of the query samples in the query sample set; dividing the query samples in the query sample set into K query sample subsets according to the value of the first attribute, denoted as K subsets. , Indicates the first A subset of query samples.

3. The approximate query processing method based on incremental learning to solve concept drift according to claim 1, characterized in that, Step S5 includes: S51: Sample N observation data points from the posterior mixture distribution of the initial posterior mixture Gaussian distribution model; S52: In the EM algorithm Based on the parameters of the initial posterior mixture Gaussian distribution model, the sub-model is calculated. Response to observed data: in, Representation of the sub-model For the first The response of each observation data point and Representation of the sub-model Parameters; Representation of the sub-model Weight parameters; S53: In the EM algorithm When calculating a new round of iterations, the sub-models Parameters: in, and Indicates the updated sub-model The parameters, Indicates the updated sub-model Weight parameters; S54: Repeat steps S52 and S53 until convergence, yielding an approximate posterior Gaussian mixture distribution model: in, This represents an approximate posterior mixture Gaussian distribution model; Indicates the model after convergence The first in Individual model; Representation of the sub-model Weight parameters; S55: For the posterior Gaussian mixture distribution model Merge similar sub-models and perform posterior mixture Gaussian distribution model. Redundant sub-models are removed to obtain the estimated posterior mixture Gaussian distribution model. .

4. The approximate query processing method based on incremental learning to solve concept drift according to claim 3, characterized in that, Step S55 includes: S551: Set a threshold When the model The sub-model in Weight parameters Less than Then, the view split model These are redundant components, which will be used in the sub-model. delete; S552: After executing step S551, for the model Any two remaining sub-models and Calculation of the sub-model and Mahalanobis distance Set threshold ,like Then the sub-model and By merging, we obtain the estimated posterior mixture Gaussian distribution model. , .

5. The approximate query processing method based on incremental learning to solve concept drift according to claim 4, characterized in that, The sub-model and Mahalanobis distance include: in, and Representation Model The Middle Individual Model The parameters, and Representation Model The Middle Individual Model The parameters.

6. The approximate query processing method based on incremental learning to solve concept drift according to claim 4, characterized in that, The sub-model and The merger includes: in, and Representation of the sub-model and The parameters of the merged model Representation of the sub-model and The weights of the merged model.

7. The approximate query processing method based on incremental learning to solve concept drift according to claim 1, characterized in that, Step S6 includes: whenever a new query sample dataset is available in the database. Upon arrival; use the same rules as in step S2 to... Divided into A subset of query samples were obtained. The training set is obtained by sampling each subset of query samples using a uniform sampling method. New BNN models are obtained by training each training set. ; Assign each sub-model and the original model ,by The initial weights constitute the initial posterior mixture distribution model. , : The updated estimated posterior mixture Gaussian distribution model is obtained by performing step S5.

Images