CN109951327A - A kind of network failure data synthesis method based on Bayesian mixture models - Google Patents

Abstract

The invention discloses a kind of network failure data synthesis method based on Bayesian mixture models, this method is used to solve in existing network failure predication due to the defect for the estimated performance decline that fault data is less and generates, the present invention is firstly for the collected Network data set with lack of balance characteristic, indicate that the minority class fault data in the data set is distributed with Bayesian mixture models, then complete model parameter estimation, minority class fault sample is generated using trained model, failure and non-faulting these two types data is made to reach balanced.Using method of the invention, it is possible to accurately hold the Network data set characteristic with lack of balance characteristic, network failure predictablity rate is effectively promoted.

Description

A kind of network failure data synthesis method based on Bayesian mixture models

Technical field

The present invention relates to a kind of network failure data synthesis method based on Bayesian mixture models, belongs to lack of balance data Processing technology field.

Background technique

With the development of internet technology, more and more users begin to use disparate networks business.Network operator It is that user provides higher quality the streaming media video service more stable with transmission in effort.Due to the generation of network failure, It is easy to cause user experience quality to decline.In other words, it if operator accurately can prejudge network failure in advance, and takes and arranges It applies and solves the problems, such as to be likely to occur in network, then user experience can be promoted effectively.Therefore, the prediction of the failure of user It is most important for network operator with timely processing.

In systems in practice, network failure data ratio shared in the whole network data set that system is collected into is opposite Smaller, in other words, the probability that network failure generates will be well below the normal probability of network.Therefore, Network data set has Lack of balance characteristic.Lack of balance data set refers to that a kind of data in data set are obviously few more many than other class data.Herein, The data volume of network failure (minority class sample) will be far less than the data volume of network normal (most class samples).For such Situation, for two traditional classifiers when handling lack of balance data, the classifier that usually training obtains has preference, so that most classes Prediction has very high accuracy, and then accuracy is very low for minority class.In the method for processing lack of balance data set, usually Method based on sampling, by changing the distribution of data set, so that lack of balance data set becomes balanced data set.

Most of existing methods are non-equal to handle by the way of directly generating new minority class sample from available sample The data that weigh, such as Synthetic Minority Oversampling Technique (SMOTE) method.These modes compare Intuitively, but since it is there is no the distribution character for deeply excavating minority class sample, the sample generated is not only not necessarily helped In classification, reaction often is played to classification, to find out its cause, the new minority class sample generated is not representative, thus Network failure can not be preferably applied to predict.

Summary of the invention

Object of the present invention is to be to solve the defects of existing network failure data processing, propose a kind of based on pattra leaves The network failure data synthesis method of this mixed model, this method is using Bayesian mixture models for describing network failure data The distribution situation of (minority class sample) is distributed firstly, describing existing fault data with Bayesian mixture models, then carries out parameter Estimation, it is final to generate new fault data using the model for estimating parameter, so that lack of balance data become relatively Weighing apparatus.

The technical scheme is that a kind of network failure data synthesis method based on Bayesian mixture models, the party Method includes the following steps:

Step 1: set collected Network data set asWherein x_nBe made of six attributes, respectively lose Packet rate, terminal downloads rate, propagation delay time, shake, video transmission quality, end-user experience scoring；The data set is corresponding Tag set isy_n=0 or 1, i.e. X correspond to two class labels, wherein y_n=0 is the normal class label of network, y_n=1 Class is network failure class label, the data amount check due to the data amount check of the normal class of network far more than network failure class, definition y_n=1 corresponding x_nComposed collection is combined into minority classWhereinFor minority class sample, N^almFor minority Class number of samples, and y_i=0 corresponding x_iComposed collection is combined into most classesWhereinFor most class samples This, N^majFor most class numbers of samples；

Step 2: Bayesian mixture models being selected to indicate X^almDistribution, probability-distribution function expression formula includes:

Wherein, K is to be mixed into score, π_j(V)、μ_j、Λ_jAnd ν_jRespectively indicate weight, the mean value, association of j-th of blending constituent Variance matrix and freedom degree parameter；For the probability density function of t distribution, indicate are as follows:

Wherein N () and Gam () respectively represent Gaussian distribution function and Gamma distribution function, u_njFor with x_nWith The relevant hidden variable of j-th of blending constituent, weight π_j(V) meetIts expression formula are as follows:

Variable V in above formula_jObey Beta distribution, i.e. p (V_j)=Beta (V_j| 1, α), α is the hyper parameter of Beta distribution, In addition, μ_j,Λ_jObey joint Gaussian-Wishart distribution, the i.e. product of Gaussian Profile and Wishart distribution, N () W ():

p(μ_j,Λ_j)=N (μ_j|m_j,λ_jΛ_j)W(Λ_j|W_j,ρ_j)

WhereinFor the hyper parameter of joint Gaussian-Wishart distribution, m_jFor 6 DOF column arrow Amount, λ_jAnd ρ_jIt is scalar, W_jFor the matrix of one (6 × 6)；Introduce a hidden variableWherein z_nIt indicates currently Data x_nIt is to be generated by which ingredient in t mixed model, works as x_nIt is z when being generated by j-th of blending constituent_nj=1, it is based on The above, the hyper parameter of entire model are as follows:

Step 3: utilizing X^almParameter Estimation is carried out to mixed model, specific as follows:

3-1) generate N^almEqually distributed random integers on a obedience [1, K] section count each integer on the section and occur Probability；That is, if producing N_jA integer j, then δ_j=N_j/N^alm；For eachCorresponding hidden variable z_nIt is initial It is distributed as

z_nFor K n dimensional vector n, in each dimension z_njValue on (j=1 ..., K) is { 0,1 }；

3-2) set hyper parameterThe initial value of α；For all j (j=1 ..., K), m_j =0, λ_j=1, ρ_jTake the arbitrary number between 3~20, W_j=10I, I are unit matrix, ν_jTake the arbitrary number between 1~100, α Take the arbitrary number between 1~10；In addition, the number of iterations counting variable k=1；

3-3) update hidden variableDistribution, that is,Its hyper parameterMore new formula are as follows:

Wherein

It is calculated in iteration for the first timeWhen,

3-4) update stochastic variableDistribution, that is, Corresponding hyper parameterMore new formula it is as follows:

Wherein,

3-5) update stochastic variableDistribution, that is,Corresponding hyper parameterMore new formula are as follows:

3-6) update hidden variableDistribution

Wherein

In above formula In, the calculation formula of items expectation<>is as follows:

Wherein Γ () is the gamma function of standard, and Γ () ' is the derivative of standard gamma function；In addition,<u_nj> calculation method respectively in step 3-3) and step 3-4) provide；

3-7) update freedom degree parameterThat is, solution contains ν as follows_jEquation:

Common numerical computation method, such as Newton method are selected, this non trivial solution ν is obtained_j；

3-8) calculate the likelihood value LIK after current iteration_itr, itr is current the number of iterations:

3-9) the difference △ LIK=LIK after calculating current iteration with the likelihood value after last iteration_itr-LIK_itr-1；Such as Otherwise fruit △ LIK≤δ is gone to step (3-3) then parameter estimation procedure terminates, the value of itr increases by 1, is continued next time Iteration；The value range of threshold value δ is 10^-5~10^-4；

Step 4: using the Bayesian mixture models estimated, generating new Network data set (X^alm) ', if needing to generate Data volume be N ', comprising:

It 4-1) is randomly generated between one 0 to 1 and obeys equally distributed random number ε；

Obedience 4-2) is randomly generatedDistribution

4-3) calculate

Obedience 4-4) is randomly generatedDistribution

4-5) utilization estimatesIf ε ∈ [0, π₁], then it generates an obedience t and is distributed t (μ₁,Λ₁,ν₁) Sample；IfIt then generates an obedience t and is distributed t (μ_k,Λ_k, ν_k) sample；IfIt then generates an obedience t and is distributed t (μ_K,Λ_K,ν_K) sample；

(4-1)~(4-5) N ' that 4-6) repeats the above steps is secondary, obtains (X^alm) ', final network failure data set isTotal data set after synthesis is

The invention has the following advantages:

1. the present invention by generate network failure data, well solved occur in network failure prediction task it is non- The not accurate enough problem of classification, the prediction for the data that weigh.

2. the present invention has modeled the distribution of network failure data using Bayesian mixture models, the data are held well Characteristic, compared with traditional method, the new more representative and classificatory area of network failure data caused by the present invention Indexing.

3. Bayesian mixture models designed by the present invention can determine optimal mould according to minority class data adaptive Type structure.

Detailed description of the invention

Fig. 1 is flow chart of the method for the present invention.

Fig. 2 is that the present invention is fitted the distribution map after artificially generated sample with Bayesian mixture models.

Fig. 3 is Bayesian mixture models iterative process likelihood value change curve of the present invention.

Fig. 4 is Kmeans-SMOTE method, the G value comparison of GMM oversampler method and the method for the present invention.

Specific embodiment

The invention is described in further detail with reference to the accompanying drawings of the specification.

As shown in Figure 1, the present invention provides a kind of network failure data synthesis method based on Bayesian mixture models, it should Method includes the following steps:

Step 1: set collected Network data set asWherein x_nBe made of six attributes, respectively lose Packet rate, terminal downloads rate, propagation delay time, shake, video transmission quality, end-user experience scoring；The data set is corresponding Tag set isy_n=0 or 1, i.e. X correspond to two class labels, wherein y_n=0 is the normal class label of network, y_n=1 Class is network failure class label, the data amount check due to the data amount check of the normal class of network far more than network failure class, definition y_n=1 corresponding x_nComposed collection is combined into minority classWhereinFor minority class sample, N^almFor minority Class number of samples, and y_i=0 corresponding x_iComposed collection is combined into most classesWhereinFor most class samples This, N^majFor most class numbers of samples；

Step 2: Bayesian mixture models being selected to indicate X^almDistribution, probability-distribution function expression formula includes:

Wherein, K is to be mixed into score, π_j(V),μ_j,Λ_j,ν_jRespectively indicate the weight of j-th of blending constituent, mean value, association Variance matrix and freedom degree parameter.For the probability density function of t distribution, can indicate are as follows:

Wherein N () and Gam () respectively represent Gaussian distribution function and Gamma distribution function, u_njFor with x_nWith The relevant hidden variable of j-th of blending constituent.Weight π_j(V) meetIts expression formula are as follows:

Variable V in above formula_jObey Beta distribution, i.e. p (V_j)=Beta (V_j| 1, α), α is the hyper parameter of Beta distribution. In addition, μ_j,Λ_jObey joint Gaussian-Wishart distribution (the i.e. product of Gaussian Profile and Wishart distribution, N () W ()):

p(μ_j,Λ_j)=N (μ_j|m_j,λ_jΛ_j)W(Λ_j|W_j,ρ_j)

WhereinFor the hyper parameter of joint Gaussian-Wishart distribution.m_jFor 6 DOF column arrow Amount, λ_jAnd ρ_jIt is scalar, W_jFor the matrix of one (6 × 6).Also need to introduce a hidden variableWherein z_nInstruction Current data x_nIt is to be generated by which ingredient in t mixed model.Work as x_nIt is z when being generated by j-th of blending constituent_nj= 1.Based on the above, the hyper parameter of entire model are as follows:

Step 3: utilizing X^almParameter Estimation is carried out to mixed model, specific as follows:

(3-1) generates N^almEqually distributed random integers on a obedience [1, K] section, count each integer on the section and go out Existing probability；That is, if producing N_jA integer j, then δ_j=N_j/N^alm；For eachCorresponding hidden variable z_nJust Begin to be distributed are as follows:

It should be noted that z_nFor K n dimensional vector n, in each dimension z_njValue on (j=1 ..., K) is { 0,1 }；

(3-2) sets hyper parameterThe initial value of α；For all j (j=1 ..., K), m_j =0, λ_j=1, ρ_jThe arbitrary number between 3~20, W can be taken_j=10I, I are unit matrix, ν_jIt can take between 1~100 Arbitrary number, α can take the arbitrary number between 1~10；In addition, the number of iterations counting variable k=1；

(3-3) updates hidden variableDistribution, that is,Its hyper parameterMore new formula are as follows:

Wherein:

It is calculated in iteration for the first timeWhen,

(3-4) updates stochastic variableDistribution, that is, Corresponding hyper parameterMore new formula it is as follows:

Wherein,

(3-5) updates stochastic variableDistribution, that is,Corresponding hyper parameterMore new formula are as follows:

(3-6) updates hidden variableDistribution

Wherein:

In above formula In, the calculation formula of items expectation<>is as follows:

Wherein Γ () is the gamma function of standard, and Γ () ' is the derivative of standard gamma function；In addition,<u_nj> calculation method provided respectively in step (3-3) and step (3-4)；

(3-7) updates freedom degree parameterThat is, solution contains ν as follows_jEquation:

Common numerical computation method, such as Newton method can be selected, this non trivial solution ν is rapidly obtained_j；

(3-8) calculates the likelihood value LIK after current iteration_itr, itr is current the number of iterations:

Difference △ LIK=LIK after (3-9) calculating current iteration with the likelihood value after last iteration_itr-LIK_itr-1；Such as Otherwise fruit △ LIK≤δ is gone to step (3-3) then parameter estimation procedure terminates, the value of itr increases by 1, is continued next time Iteration；The value range of threshold value δ is 10^-5~10^-4。

Step 4: using the Bayesian mixture models estimated, generating new Network data set (X^aim) ', if needing to generate Data volume be N ', comprising:

(4-1), which is randomly generated between one 0 to 1, obeys equally distributed random number ε；

Obedience is randomly generated in (4-2)Distribution

(4-3) is calculated

Obedience is randomly generated in (4-4)Distribution

What (4-5) utilization estimatedIf ε ∈ [0, π₁], then it generates an obedience t and is distributed t (μ₁,Λ₁,ν₁) Sample；IfK=2 ..., K-1 then generates an obedience t and is distributed t (μ_k,Λ_k,ν_k) Sample；IfIt then generates an obedience t and is distributed t (μ_K,Λ_K,ν_K) sample；

(4-6) (4-1)~(4-5) N ' that repeats the above steps is secondary, obtains (X^alm) ', final network failure data set isTotal data set after synthesis is

Performance compares:

The Clustering Effect of Bayesian mixture models (DPMM) is tested first.Thinking is as follows: using certain for coming from multiple clusters And the unknown several samples of cluster label, unsupervised study is carried out with DPMM clustering algorithm, finally by cluster result and originally The label of sample, which compares, shows classifying quality.

In this experiment, 1000 three-dimensional samples, the iteration time of experiment are generated using three single Gauss models Number is 200.The probability distribution of samples points completed after fitting with the Bayesian mixture models that the present invention designs is as shown in Figure 2.It is wherein correct The sample size of classification is 942, and the accuracy of fitting has reached 94.2%.What Fig. 3 was indicated is the class in 200 iterative process The variation line chart of other quantity, can the score K that is mixed into of the therefrom model in figure be generally in the fluctuation of 3 surroundings, and it is big It is final convergent when about iterating to 160 times.The experimental results showed that can be according to described sample based on Bayesian mixture models Originally model structure is automatically determined.

Then for the network data that certain network operator provides, the method for the present invention is subjected to replication experiment.Use this Inventive method synthesizes new sample and is added in minority class, finally makes new data set relative equilibrium, then using simple shellfish This classifier of leaf is trained modeling to new data set as base classifier, is then tested using test data set.Make It is compared with traditional Kmeans-SMOTE method and GMM oversampler method.Test data set uses initial data, surveys In examination data the ratio of minority class and most classes we select 1:30,1:60 and 1:89 to be trained test, the result of experiment It is as shown in Figure 4: from fig. 4, it can be seen that relative to Kmeans-SMOTE algorithm and GMM oversampler method, the method for the present invention DPMM G value improve 16% and 4.8% respectively.Therefore, the method for the present invention produces the classification prediction effect of unequalizing network data Effective promotion is given birth to.

Claims (1)

1. a kind of network failure data synthesis method based on Bayesian mixture models, which is characterized in that the method walks as follows It is rapid:

Step 1: set collected Network data set asWherein x_nBe made of six attributes, respectively packet loss, Terminal downloads rate, propagation delay time, shake, video transmission quality, end-user experience scoring；The corresponding tally set of the data set It is combined intoy_n=0 or 1, i.e. X correspond to two class labels, wherein y_n=0 is the normal class label of network, y_n=1 class is net Network failure classes label defines y since the data amount check of the normal class of network is far more than the data amount check of network failure class_n=1 pair The x answered_nComposed collection is combined into minority classWhereinFor minority class sample, N^almFor minority class sample number Mesh, and y_i=0 corresponding x_iComposed collection is combined into most classesWhereinFor most class samples, N^maj For most class numbers of samples；

Step 2: Bayesian mixture models being selected to indicate X^almDistribution, probability-distribution function expression formula includes:

Wherein, K is to be mixed into score, π_j(V)、μ_j、Λ_jAnd ν_jRespectively indicate weight, the mean value, covariance of j-th of blending constituent Matrix and freedom degree parameter；For the probability density function of t distribution, indicate are as follows:

Wherein N () and Gam () respectively represent Gaussian distribution function and Gamma distribution function, u_njFor with x_nWith j-th The relevant hidden variable of blending constituent, weight π_j(V) meetIts expression formula are as follows:

Variable V in above formula_jObey Beta distribution, i.e. p (V_j)=Beta (V_j| 1, α), α is the hyper parameter of Beta distribution, in addition, μ_j,Λ_jJoint Gaussian-Wishart distribution, the i.e. product of Gaussian Profile and Wishart distribution are obeyed, N () W ():

p(μ_j,Λ_j)=N (μ_j|m_j,λ_jΛ_j)W(Λ_j|W_j,ρ_j)

WhereinFor the hyper parameter of joint Gaussian-Wishart distribution, m_jFor sextuple column vector, λ_j And ρ_jIt is scalar, W_jFor the matrix of one (6 × 6)；Introduce a hidden variableWherein z_nIndicate current data x_n It is to be generated by which ingredient in t mixed model, works as x_nIt is z when being generated by j-th of blending constituent_nj=1, it is based on the above institute It states, the hyper parameter of entire model are as follows:

Step 3: utilizing X^almParameter Estimation is carried out to mixed model, specific as follows:

3-1) generate N^almEqually distributed random integers on a obedience [1, K] section, count each integer on the section occur it is general Rate；That is, if producing N_jA integer j, then δ_j=N_j/N^alm；For eachCorresponding hidden variable z_nInitial distribution For

z_nFor K n dimensional vector n, in each dimension z_njValue on (j=1 ..., K) is { 0,1 }；

3-2) set hyper parameterThe initial value of α；For all j, j=1 ..., K, m_j=0, λ_j =1, ρ_jTake the arbitrary number between 3~20, W_j=10I, I are unit matrix, ν_jTaking the arbitrary number between 1~100, α takes 1~ Arbitrary number between 10；In addition, the number of iterations counting variable k=1；

3-3) update hidden variableDistribution, that is,Its hyper parameterMore new formula are as follows:

Wherein

It is calculated in iteration for the first timeWhen,

3-4) update stochastic variableDistribution, that is, Corresponding hyper parameterMore new formula it is as follows:

Wherein,

3-5) update stochastic variableDistribution, that is,Corresponding hyper parameterMore new formula are as follows:

3-6) update hidden variableDistribution

Wherein

In above formula, the calculation formula of items expectation<>is as follows:

Wherein Γ () is the gamma function of standard, and Γ () ' is the derivative of standard gamma function；In addition,WithCalculation method respectively in step 3-3) and step 3-4) provide；

3-7) update freedom degree parameterThat is, solution contains ν as follows_jEquation:

Common numerical computation method, such as Newton method are selected, this non trivial solution ν is obtained_j；

3-8) calculate the likelihood value LIK after current iteration_itr, itr is current the number of iterations:

3-9) the difference △ LIK=LIK after calculating current iteration with the likelihood value after last iteration_itr-LIK_itr-1；If △ Otherwise LIK≤δ is gone to step (3-3) then parameter estimation procedure terminates, the value of itr increases by 1, continue next time repeatedly Generation；The value range of threshold value δ is 10^-5~10^-4；

Step 4: using the Bayesian mixture models estimated, generating new Network data set (X^alm) ', if the number for needing to generate It is N ' according to amount, comprising:

It 4-1) is randomly generated between one 0 to 1 and obeys equally distributed random number ε；

Obedience 4-2) is randomly generatedDistribution

4-3) calculate

Obedience 4-4) is randomly generatedDistribution

4-5) utilization estimatesIf ε ∈ [0, π₁], then it generates an obedience t and is distributed t (μ₁,Λ₁,ν₁) sample； IfIt then generates an obedience t and is distributed t (μ_k,Λ_k,ν_k) sample This；IfIt then generates an obedience t and is distributed t (μ_K,Λ_K,ν_K) sample；

(4-1)~(4-5) N ' that 4-6) repeats the above steps is secondary, obtains (X^alm) ', final network failure data set isTotal data set after synthesis is