A method, device and computer equipment for constructing multiple extreme learning machine models
Through the multi-limit learning machine model and the nunclaw construction theory under the Bayesian framework, the problem of difficulty in optimizing the existing anomaly point detection regression model and the inability to characterize complex error distributions is solved, and a strong fit and robustness of complex distributions are achieved Performance improvement.
Patent Information
- Application Number
- CN202010929560.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-09-07
- Publication Date
- 2024-11-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing anomaly point detection regression model is difficult to optimize and cannot fully characterize the real complex error distribution.
Using the multi-limit learning machine model, by obtaining multiple single-limit learning machine models with different number of hidden layers, combining Bayesian framework and nunscrab construction theory, an infinitely mixed Gaussian distribution of error terms is constructed, and sparse Bayesian leaves are used. The SS learning and variational Bayesian method optimize model parameters.
A strong fitting ability to complex distributions is achieved, and the need to select the best number of hidden layer nodes and loss functions is avoided, and a robust prediction effect and a tight model parameter distribution is obtained.
Smart Images

Figure CN112070157B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and in particular to a method and device for constructing a multi-extreme learning machine model and a computer device. Background Art
[0002] In the era of big data, the amount of data is getting larger and larger, and it is relatively easier to obtain large-scale information on the research object. At the same time, due to factors such as human factors or instruments and equipment in the data collection process, there will inevitably be some anomalies in the data. These points will have a great impact on the performance of the model to be established. Therefore, it is necessary to develop a regression model that is robust to anomalies.
[0003] In the prior art, two methods are usually used to deal with regression problems with outliers: data preprocessing methods and robust regression models. There are two types of data preprocessing methods: data processing methods based on outlier detection and signal processing methods. The former only processes outliers, usually using statistics, distance-based, density-based models and soft computing techniques to detect outliers, and then deletes these points from the training data; the latter usually uses signal processing methods (such as wavelet transform) to preprocess all sample points in the original data to reduce the adverse effects of outliers or unknown noise. The principle of designing a robust loss function for a robust regression model is to ensure that the value of the loss function does not change too much with the increase of error. A common strategy is to add weights to different samples, such as Hurbe, Hampel, Logistic and Myriad weighted functions, and then derive a weighted loss function, such as a robust support vector machine based on the lncosh loss function and a robust least squares support vector machine based on the minimum absolute deviation loss function.
[0004] Although the above-mentioned regression model, which is more robust to outliers or unknown noise, can predict the results more accurately, the data preprocessing method cannot determine whether all outliers have been detected and preprocessed. The loss function is difficult to optimize on the one hand, and on the other hand, it cannot fully characterize the real complex error distribution. Summary of the invention
[0005] Based on this, the present invention provides a multi-extreme learning machine model construction method for the above-mentioned problems, aiming to solve the technical problems that the existing outlier detection regression model is difficult to optimize and cannot comprehensively characterize the real complex error distribution.
[0006] The present invention provides a method for constructing a multi-extreme learning machine model, which specifically includes:
[0007] Acquire multiple single extreme learning machine models with different numbers of hidden layer nodes, assign weight values to the single extreme learning machine models, and construct a multi-extreme learning machine model; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term;
[0008] In the Bayesian framework, the distribution weight of the error term is constructed according to the stick construction theory to obtain the infinite mixed Gaussian distribution of the error term;
[0009] Obtaining an output weight vector and a model weight vector given by a user, and obtaining a sparse prior distribution of the output weight vector and the model weight vector;
[0010] Obtaining a joint probability of data and a set of parameters in the multi-extreme learning machine model according to the sparse prior distribution of the error term mixed Gaussian distribution, the output weight vector, and the model weight vector; and
[0011] The parameters in the joint probability are inferred using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model.
[0012] Furthermore, under the Bayesian framework, the distribution weight of the error term is constructed according to the stick construction theory, and obtaining the infinite mixture Gaussian distribution of the error term specifically includes:
[0013] In the Bayesian framework, the indicator variable z is introduced ik , the mixed Gaussian distribution of the error term can be expressed as:
[0014]
[0015]
[0016] The z i ={z i1 , …, z iK}, represents the Gaussian probability distribution function, π={π1,…,π K}, π k >0 is the weight of each distribution, τ k is the precision parameter of the Gaussian distribution, represents the multinomial distribution, ε i represents the regression error term corresponding to the i-th sample, and K represents the number of Gaussian distributions in the mixed Gaussian distribution.
[0017] Reconstruct the distribution weight π according to the stick construction theory:
[0018]
[0019]
[0020] Said is the probability density function of the Beta distribution, δ represents the parameter in the Beta distribution, represents the probability density function of the Gamma distribution, e0, f0 are the parameters of the Gamma distribution, θ k represents the proportion of the remaining stick length intercepted for the kth time, and θ represents the length of all θ k The constructed vector;
[0021] Obtain an infinite mixture of Gaussian distributions for the error terms.
[0022] Furthermore, the sparse prior distribution of the output weight vector and the model weight vector is:
[0023] Get the output weight vector β given by the user m , model weight vector ω g , the expressions of the sparse prior distribution of the output weight vector and the model weight vector are respectively:
[0024]
[0025]
[0026] The j0, k0, g0, and h0 are the prior parameters of the Gamma distribution. η g They represent the precision parameters of Gaussian distribution respectively.
[0027] Furthermore, the joint probability of the data and the parameter set in the multi-extreme learning machine model is: assuming that the data is T, all parameter sets in the model are Θ, and the joint probability is:
[0028]
[0029] The p(T) represents the conditional distribution of the data which can be derived from the error distribution.
[0030] Furthermore, the posterior distribution of the parameters in the multi-extreme learning machine model specifically includes:
[0031] Parameter Z = {z1, ..., z N The posterior distribution of} can be expressed as
[0032]
[0033] Said
[0034] The parameter τ k , δ and The posterior distribution follows the Gamma distribution, with parameter θ kThe posterior distribution follows the Beta distribution;
[0035] The posterior distribution of the parameter ω is a mean μ ω and covariance∑ ω The multidimensional Gaussian distribution of is expressed as:
[0036]
[0037] The covariance matrix
[0038] Mean
[0039] The parameter η g The posterior distribution of is a parameter and Gamma distribution,
[0040]
[0041] Said
[0042] The posterior distribution of the parameter β is a multidimensional Gaussian distribution, expressed as:
[0043]
[0044] The covariance matrix
[0045] Mean
[0046] Another object of an embodiment of the present invention is to provide a device for constructing a multi-extreme learning machine model, the device specifically comprising:
[0047] A multi-extreme learning machine model building module is used to obtain multiple single extreme learning machine models with different numbers of hidden layer nodes, and assign weight values to the single extreme learning machine models to build a multi-extreme learning machine model; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term;
[0048] An error term distribution construction module is used to construct the distribution weight of the error term according to the stick construction theory in a Bayesian framework to obtain an infinite mixed Gaussian distribution of the error term;
[0049] A sparse prior module, used to obtain an output weight vector and a model weight vector given by a user, and obtain a sparse prior distribution of the output weight vector and the model weight vector;
[0050] A joint probability acquisition module, used to obtain the joint probability of the data and the parameter set in the multi-extreme learning machine model according to the sparse prior distribution of the error term mixed Gaussian distribution, the output weight vector and the model weight vector;
[0051] The parameter posterior distribution module is used to infer the parameters in the joint probability using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model.
[0052] Another object of an embodiment of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the steps of the above method when executing the computer program.
[0053] Beneficial effects:
[0054] The present invention expands a single extreme learning machine model (ELM) into a multiple extreme learning machine model (MELM), uses an infinite mixture of Gaussian distributions (IMoG) to fit the regression error of the model MELM, adopts a sparse Bayesian learning method to optimize the output weights and model weights in the model MELM, and adopts a variational Bayesian method to optimize all parameters in the model. The MELM-IMoG model does not need to determine the optimal number of hidden layer nodes, and can comprehensively utilize the prediction capabilities of ELM models with different numbers of hidden layer nodes, thereby obtaining a good prediction effect; the super strong fitting ability for complex distributions does not require the selection of the best loss function or the assumption of a suitable single error distribution in advance to obtain super strong robust performance, and they can also avoid predetermining the number of single Gaussian distributions in MoG; the sparse Bayesian learning method allows us to obtain a relatively sparse model parameter, and then obtain a compact prediction model, and the probability distribution of all parameters in the model can be automatically obtained under the condition of given model parameter initial values.
[0055] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 A flowchart of a method for constructing a multi-extreme learning machine model provided by the present invention;
[0057] Figure 2 A schematic diagram of the structure of a device for building a multi-extreme learning machine model provided by the present invention;
[0058] Figure 3 The sparse results of the MELM-IMoG output weight vector β and the model weight vector ω on the benchmark data set provided by the embodiment of the present invention;
[0059] Figure 4The sparse results of the MELM-IMoG output weight vector β and the model weight vector ω in the traffic flow prediction provided by the embodiment of the present invention. DETAILED DESCRIPTION
[0060] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0061] like Figure 1 As shown, in one embodiment, a method for constructing a multi-extreme learning machine model is proposed, which specifically includes the following steps:
[0062] Step 101, obtain multiple single learning machine models of the sample with different numbers of hidden points, assign weight values to the single learning machine models, and construct a sample multi-extreme learning machine model; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term.
[0063] In the embodiment of the present invention, it is assumed that there are G different numbers of hidden layer nodes, namely, M1, M2, ..., M G , corresponding to the hidden layer output matrices H1, H2, ..., H with different sizes G For the convenience of calculation, let H1, H2, …, H G Converted to The matrix is as shown in formula (1).
[0064]
[0065] in represents the zero matrix, Represents the output matrix after the g-th transformation.
[0066] The multi-extreme learning machine model combines many single ELM models with different numbers of hidden layer nodes by assigning different weights, which is expressed as formula (2):
[0067]
[0068] Among them, T i is the target of the i-th sample, ω∈R G×1 represents the weight vector of all hidden layer output matrices (model weight vector), is the output weight vector, is a matrix The i-th row of i is the error term. The matrix form of MELM is as shown in formula (3):
[0069]
[0070] where T,ω∈R G×1 , ω g represents the weight of the g-th output matrix, and N represents the number of training samples. In the framework of MELM, the output weight vector β and the model weight vector ω can be learned simultaneously with the training data. Given the sparse Bayesian prior of the output weight vector β and the model weight vector ω, the fitting ability of the model is increased, and there is no need to select the best loss function or assume a suitable error distribution.
[0071] Step 102, under the Bayesian framework, construct the distribution weight of the error term according to the stick construction theory to obtain the infinite mixture Gaussian distribution of the error term.
[0072] In the embodiment of the present invention, it is assumed that the error term ε of MELM i Obeying MoG, the probability of error is expressed as formula (4):
[0073]
[0074] where K is the number of Gaussian distributions, π k >0 is the weight of each distribution, and τ k is the precision parameter of the Gaussian distribution, represents the Gaussian probability distribution function, and p(·) represents the probability distribution of a parameter. In the Bayesian framework, the indicator variable z is introduced ik , the above formula can be rewritten into the following two-layer generative model formulas (5) and (6):
[0075]
[0076]
[0077] where z i ={z i1 , …, z iK},π={π1,…,π K}, represents the multinomial distribution. To complete the Bayesian modeling, assume that τ k Subject to the Gamma prior with parameters a0, b0, π follows the Dirichlet distribution:
[0078]
[0079]
[0080] in represents the probability density function of the Gamma distribution, represents the probability density function of the Dirichlet distribution, and c0 is the parameter of the Dirichlet distribution.
[0081] In the embodiment of the present invention, the distribution weight π is reconstructed according to the stick construction theory, and the original mixed Gaussian distribution is extended to an infinite mixed Gaussian (IMoG):
[0082]
[0083]
[0084] in is the probability density function of the Beta distribution, δ represents the parameter in the Beta distribution, represents the probability density function of the Gamma distribution, e0, f0 are the parameters of the Gamma distribution, θ k represents the proportion of the remaining stick length intercepted for the kth time, and θ represents the length of all θ k The constructed vector. From this we can get:
[0085]
[0086] Step 103: Obtain the output weight vector and the model weight vector given by the user, and obtain the sparse prior distribution of the output weight vector and the model weight vector.
[0087] Step 104, obtaining the joint probability of the data and the parameter set in the multi-extreme learning machine model according to the mixed Gaussian distribution of the error term, the output weight vector and the sparse prior distribution of the model weight vector.
[0088] In the embodiment of the present invention, in order to obtain the sparse output weight vector and the weight vector of the hidden layer output matrix, the parameter β is given as m and ω g The following sparse prior distribution,
[0089]
[0090]
[0091] Among them, j0, k0, g0, and h0 are all prior parameters of the Gamma distribution. η g They represent the precision parameters of Gaussian distribution respectively.
[0092] In summary, the prior distribution of all parameters in the model is given. The joint probability of the data T and the set of all parameters Θ in the model is:
[0093]
[0094] Here, p(T) represents the conditional distribution of the data that can be derived from the error distribution.
[0095] Step 105: Infer the parameters in the joint probability using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model.
[0096] In the embodiment of the present invention, in order to estimate all parameters in the joint probability p(T, Θ), the present patent adopts the variational Bayes method (VB) and lets q(·) represent the posterior distribution of the parameters.
[0097] Parameter Z = {z1, ..., z N The posterior distribution of} can be expressed as
[0098]
[0099] in
[0100] Parameter τ k The posterior of follows a Gamma distribution,
[0101]
[0102] in
[0103] Parameter θ k The posterior of is Beta distributed,
[0104]
[0105] in
[0106] The posterior q(δ) of parameter δ is a Gamma distribution,
[0107]
[0108] The parameter e δ =e0+K-1,
[0109] The posterior distribution of parameter ω is a mean μ ω and covariance∑ ω The multidimensional Gaussian distribution of
[0110]
[0111] The covariance matrix Mean
[0112] Parameter η g The posterior distribution of is a parameter and Gamma distribution,
[0113]
[0114] in
[0115] The posterior distribution of parameter β is a multidimensional Gaussian distribution,
[0116]
[0117] The covariance matrix Mean
[0118] parameter The posterior distribution of is Gamma distribution,
[0119]
[0120] in
[0121] The embodiment of the present invention expands a single extreme learning machine model (ELM) into a multiple extreme learning machine model (MELM), uses an infinite mixture of Gaussian distributions (IMoG) to fit the regression error of the model MELM, adopts a sparse Bayesian learning method to optimize the output weights and model weights in the model MELM, and adopts a variational Bayesian method to optimize all parameters in the model. The MELM-IMoG model does not need to determine the optimal number of hidden layer nodes, and can comprehensively utilize the prediction capabilities of ELM models with different numbers of hidden layer nodes, thereby obtaining a good prediction effect; the super strong fitting ability for complex distributions does not require the selection of the best loss function or the assumption of a suitable single error distribution in advance to obtain super strong robust performance, and they can also avoid predetermining the number of single Gaussian distributions in MoG; the use of a sparse Bayesian learning method allows us to obtain a relatively sparse model parameter, and then obtain a compact prediction model, and the probability distribution of all parameters in the model can be automatically obtained when the initial value of the model parameter is given.
[0122] like Figure 2 As shown, the present invention provides a structural schematic diagram of a device for building a multi-extreme learning machine model.
[0123] A multi-extreme learning machine model building module 210 is used to obtain multiple single learning machine models of the sample with different numbers of hidden points, and assign weight values to the single learning machine models to build a multi-extreme learning machine model of the sample; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term;
[0124] The error term distribution construction module 220 is used to construct the distribution weight of the error term according to the stick construction theory in a Bayesian framework to obtain an infinite mixture Gaussian distribution of the error term;
[0125] A sparse prior module 230 is used to obtain an output weight vector and a model weight vector given by a user, and obtain a sparse prior distribution of the output weight vector and the model weight vector;
[0126] A joint probability acquisition module 240, for obtaining the joint probability of the data and the parameter set in the multi-extreme learning machine model according to the sparse prior distribution of the error term mixed Gaussian distribution, the output weight vector and the model weight vector;
[0127] The parameter posterior distribution module 250 is used to infer the parameters in the joint probability using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model.
[0128] In an embodiment of the present invention, a computer device is further provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, steps of the methods of the present invention are performed.
[0129] Example
[0130] Experimental setting: In the embodiment of the present invention, three activation functions (Sigmoid, Fourier and Gaussian functions) are used to construct the MELM-IMoG model proposed in this patent. The number of hidden layer nodes in MELM-IMoG is {10, 20, 40, 60, 80, 100, 120, 140, 160, 180, 200, 250, 300}, the proposed model has a total of 39 hidden layer output matrices, and the following 5 comparative models are used, namely ELM, WELM (weighted ELM), ORELM (outlier robust ELM), ELM-T (ELM based on t-distribution error assumption) and MELM-G (MELM based on Gaussian noise assumption).
[0131] Experimental data: The eight benchmark datasets are shown in Table 1. The number of training samples and test samples in each data set is shown in Table 1. The output in the training sample is added with outliers, and the proportion of outliers is 30%, while the output in the test sample is noise-free. The error indicators used are mean absolute error (MAE), root mean square error (RMSE), and normalized mean absolute percentage error (NMAPE).
[0132] Table 1 Description of eight benchmark datasets
[0133]
[0134] The traffic flow data used was collected on Los Angeles freeways from March 1 to July 7, 2012. The time span between two observations is 5 minutes. There are 2016 samples in the dataset, of which the first 1513 samples are used as training samples and the remaining 503 samples are used to test the performance of different prediction models. Here, the traffic flow prediction of 10 minutes ahead and 15 minutes ahead is studied.
[0135] Comparison of regression results of different regression models on benchmark datasets: The regression results of different regression models on eight benchmark datasets are shown in Table 2. The sparse results of MELM-IMoG output weight vector β and model weight vector ω on the benchmark datasets are shown in Table 2. Figure 3 shown.
[0136] Table 2 Regression errors of different regression models
[0137]
[0138] By comparing MELM-IMoG and MELM-G, we found that the error assumption of infinite mixed Gaussian distribution can increase the robustness of the model. Both the ELM and MELM-G models assume that the error is Gaussian distributed. By comparing their regression results, it can be found that the performance of MELM-G is better than ELM. The main reason is that the former combines the regression performance of multiple ELM models and can more accurately describe the complex relationship between input and output when fitting complex regression tasks. By comparing ELM and three robust models WELM, ORELM and ELM-T, we found that the robust performance of ORELM and ELM-T is better than the other two models, and is also stronger than the robustness of the model MELM-G. As can be seen from Table 2, when there are outliers in the data, the model MELM-IMoG proposed in this patent has the smallest regression error and the strongest robustness among all models, while the traditional ELM model has the worst regression performance.
[0139] Figure 3 The sparsity of the output weight vector β and the model weight vector ω in the presence of outliers is shown. When the weight value is close to 0, it means that the corresponding output matrix or hidden node has little effect on the final result. Figure 1 It can be seen that the sparse weight prior does produce a sparse solution, which can help us obtain a more compact regression model and remove the influence of redundant information on the regression model.
[0140] In summary, MELM-IMoG has better regression performance in the presence of outliers. Given the infinite mixture distribution of the error term assumption, the regression model will be more robust to outliers than a single error distribution assumption. The proposed model also utilizes the multivariate regression results generated by multiple ELMs with different numbers of hidden nodes. Therefore, they can eliminate the need for the process of selecting the best hidden nodes.
[0141] Comparison of results of different regression models in traffic flow prediction: The traffic flow prediction results are shown in Table 3. Figure 4 The 15-minute-ahead traffic flow prediction results of different models are shown.
[0142] Table 3 Wind speed prediction results of different models 60 minutes in advance
[0143]
[0144]
[0145] According to Table 3, the model MELM-IMoG has the smallest error index among the three models, so its prediction performance is the best among all models. In addition, by comparing with MELM-G, it can be found that the infinite mixture Gaussian distribution in MELM-IMoG can fully fit the complex traffic flow prediction error, thereby obtaining better prediction accuracy. It can also be seen from Table 3 that compared with the linear model ARIMA, the prediction performance of most nonlinear models is better than that of linear models. Figure 4 The figure shows the sparse results of the MELM-IMoG output weight vector β and the model weight vector ω in traffic flow prediction. It can be seen from the figure that the output weights and model weights in MELM-IMoG have good sparsity.
[0146] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
[0147] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.
[0148] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.
[0149] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0150] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A method for constructing a multi-extreme learning machine model, characterized in that: Specifically include: Acquire multiple single extreme learning machine models with different numbers of hidden layer nodes, assign weight values to the single extreme learning machine models, and construct a multi-extreme learning machine model; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term; In the Bayesian framework, the distribution weight of the error term is constructed according to the stick construction theory to obtain the infinite mixed Gaussian distribution of the error term; Obtaining an output weight vector and a model weight vector given by a user, and obtaining a sparse prior distribution of the output weight vector and the model weight vector; Obtaining a joint probability of data and a set of parameters in the multi-extreme learning machine model according to the sparse prior distribution of the error term mixed Gaussian distribution, the output weight vector, and the model weight vector; and The parameters in the joint probability are inferred using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model; based on the traffic flow data of the historical time period, the multi-extreme learning machine model is used to make a prediction to obtain the traffic flow of the predicted time period; Wherein, under the Bayesian framework, the distribution weight of the error term is constructed according to the stick construction theory, and obtaining the infinite mixture Gaussian distribution of the error term specifically includes: In the Bayesian framework, the indicator variable z is introduced ik , the mixed Gaussian distribution of the error term can be expressed as: The z i ={z i1 , …, z iK }, represents the Gaussian probability distribution function, π={π1,…,π K }, π k >0 is the weight of each distribution, τ k is the precision parameter of the Gaussian distribution, represents the multinomial distribution, ε i represents the regression error term corresponding to the i-th sample, and K represents the number of Gaussian distributions in the mixed Gaussian distribution; Reconstruct the distribution weight π according to the stick construction theory: Said is the probability density function of the Beta distribution, δ represents the parameter in the Beta distribution, represents the probability density function of the Gamma distribution, e0, f0 are the parameters of the Gamma distribution, θθ k represents the proportion of the remaining stick length intercepted for the kth time, and θθ represents the length of all θθ k The constructed vector; Obtain an infinite mixture of Gaussian distributions for the error terms.
2. The method for constructing a multi-extreme learning machine model according to claim 1, characterized in that: The sparse prior distribution of the output weight vector and the model weight vector is: Get the output weight vector β given by the user m , model weight vector ω g , the expressions of the sparse prior distribution of the output weight vector and the model weight vector are respectively: The j0, k0, g0, and h0 are the prior parameters of the Gamma distribution. They represent the precision parameters of Gaussian distribution respectively.
3. The method for constructing a multi-extreme learning machine model according to claim 2, characterized in that: The joint probability of the data and the parameter set in the multi-extreme learning machine model is: Assume that the data is T, the set of all parameters in the model is Θ, and the joint probability is: The p(T) represents the conditional distribution of the data which can be derived from the error distribution.
4. The method for constructing a multi-extreme learning machine model according to claim 3, characterized in that: The posterior distribution of the parameters in the multi-extreme learning machine model specifically includes: Parameter Z = {z1, ..., z N The posterior distribution of} can be expressed as: Said The parameter τ k , δ and The posterior distribution follows the Gamma distribution, with parameter θ k The posterior distribution follows the Beta distribution; The posterior distribution of the parameter ω is a mean μ ω and covariance∑ ω The multidimensional Gaussian distribution of is expressed as: Covariance matrix Mean The parameter η g The posterior distribution of is a parameter and Gamma distribution, Said The posterior distribution of the parameter β is a multidimensional Gaussian distribution, expressed as: The covariance matrix Mean 5. A device for constructing a multi-extreme learning machine model, characterized in that: The device specifically comprises: A multi-extreme learning machine model building module obtains multiple single extreme learning machine models with different numbers of hidden layer nodes, and assigns weight values to the single extreme learning machine models to build a multi-extreme learning machine model; the multi-extreme learning machine model includes an output value vector, a model weight vector and an error term; An error term distribution construction module is used to construct the distribution weight of the error term according to the stick construction theory in a Bayesian framework to obtain an infinite mixed Gaussian distribution of the error term; A sparse prior module, used to obtain an output weight vector and a model weight vector given by a user, and obtain a sparse prior distribution of the output weight vector and the model weight vector; A joint probability acquisition module, used to obtain the joint probability of the data and the parameter set in the multi-extreme learning machine model according to the sparse prior distribution of the error term mixed Gaussian distribution, the output weight vector and the model weight vector; A parameter posterior distribution module is used to infer the parameters in the joint probability using a variational Bayesian method to obtain the posterior distribution of the parameters of the multi-extreme learning machine model; based on the traffic flow data of the historical time period, the multi-extreme learning machine model is used to make a prediction to obtain the traffic flow of the predicted time period; The error term distribution building module is specifically used to achieve: In the Bayesian framework, the indicator variable z is introduced ik , the mixed Gaussian distribution of the error term can be expressed as: The z i ={z i1 , …, z iK }, represents the Gaussian probability distribution function, π={π1,…,π K }, π k >0 is the weight of each distribution, τ k is the precision parameter of the Gaussian distribution, represents the multinomial distribution, ε i represents the regression error term corresponding to the i-th sample, and K represents the number of Gaussian distributions in the mixed Gaussian distribution; Reconstruct the distribution weight π according to the stick construction theory: Said is the probability density function of the Beta distribution, δ represents the parameter in the Beta distribution, represents the probability density function of the Gamma distribution, e0, f0 are the parameters of the Gamma distribution, θθ k represents the proportion of the remaining stick length intercepted for the kth time, and θθ represents the length of all θθ k The constructed vector; Obtain an infinite mixture of Gaussian distributions for the error terms.
6. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the method according to any one of claims 1 to 4 when executing the computer program.
