CN108596272A

CN108596272A - A kind of semisupervised classification machine learning new method based on figure

Info

Publication number: CN108596272A
Application number: CN201810437033.0A
Authority: CN
Inventors: 刘建峰
Original assignee: Chongqing Three Gorges University
Current assignee: Chongqing Three Gorges University
Priority date: 2018-05-09
Filing date: 2018-05-09
Publication date: 2018-09-28

Abstract

The present invention proposes a kind of semisupervised classification new method based on figure, gives the sample classification method including step S001 to step S005；And the active sample labeling method including step D000, D001, D002 is provided, training sample set is pre-processed, supervised learning and unsupervised learning advantage are combined, there is training sample to prepare simple, the high advantage of training precision；It is proposed in this paper to pre-process active sample labeling method in training set and sorting technique based on figure makes training precision higher compared to general semi-supervised learning method, the label difficulty of data sample is reduced.

Description

A kind of semisupervised classification machine learning new method based on figure

Technical field

The present invention relates to a kind of machine learning methods, and in particular to a kind of semisupervised classification machine learning side based on figure Method.

Background technology

Problem concerning study is divided into three classes in machine learning：Supervised learning, semi-supervised learning and unsupervised learning；Reality The class label that sample is obtained in data is the thing of a time and effort consuming, and exemplar will improve the classification capacity of grader. Method based on figure is a kind of important, effective semi-supervised learning method, and existing semi-supervised learning classification is correct Rate and two evaluation index performances of standard error are poor.

Invention content

The semisupervised classification machine learning new method based on figure that the purpose of the present invention is to provide a kind of so that grader exists It is improved in two index performances of classification accuracy rate and standard error.

To achieve the above object, the semisupervised classification machine learning new method based on figure that the present invention provides a kind of, mainly Include the following steps：

Step 1：Divide training set；Training set X=L ∪ U={ x₁,…x_l,x_l+1,…x_l+u, L={ x₁,…x_lIt is few Amount has exemplar, U={ x_l+1,…x_l+uIt is a large amount of unlabeled exemplars；Marker samples accounting

Step 2：Construct asymmetric sparse weight matrix；Structural map G=(V, E), V indicate each sample point, E representative samples Side between this point, side e ∈ E are determining by w (e), w (e)=w_ij, w_ijIndicate sample point x_iAnd x_jBetween similitude；

According to linear property and sparsity, a sample x is given_i, as all base vectors and x_iWhen from same class, it can obtain Obtain the sparse expression of the sample.T_kIt indicates to remove x_kExcept row sample matrixIndicate sparse decomposition coefficients.Wherein T_k= [x₁,…,x_k-1,x_k+1,…,x_n],

Asymmetric weight matrix is：

Step 3：Solve asymmetric weight matrix；Give a sample x_i, weight system is gone by following optimization problem solving Number：

To be converted into the following problem of optimization：

min||q||₁S.t.Pq=x_k,q_i>=0, i=1,2 ..., k-1, k+1 ..., n

Wherein P=[G_k I_d]∈R^d×(d+n-1), q=[a e]^T, can be solved by linear programming problem, result is：

w_ii=0,

The weight matrix is unsymmetrical matrix；

Step 4：Objective function；The object function formed using loss function and regular terms, expression formula are as follows：

Wherein C is loss function, | | f | |²It is the penalty term in the spaces PKHS,Indicate epidemiological features inside data,F=[f (x in formula₁),f(x₂),…,f(x_l+u)], the optimization above problem can obtain：

In formula, L is Laplce, is indicated using sparse weight, L=D-W ∈ R^n×n, W is sample sparse matrix.Above formula It is represented by：

Regular terms A can be calculated by input sample；

Step 5, judgement sample classification；Symbol function pair sample class is defined to make decisions

In above-mentioned semisupervised classification learning method, marker samples accounting is to decide classifier performance in training set sample； Marker samples accounting is bigger, and nicety of grading is higher；But the label of sample is not easy to obtain, and reduces semi-supervised learning efficiency and property Energy；Further, the present invention is to promote semi-supervised learning efficiency and performance, it is proposed that a kind of active sample labeling method；

It is implemented as follows：

General sample M={ m are defined first_i, N={ n_i, i=1,2 ..., the symmetry distance between t：

In formula, symmetry distance is smaller, and two sample distributions of characterization are more similar, change to symmetry distance in conjunction with Bayesian probability It writes：

In formula, p (c_i| it is x) that sample x belongs to classification c_iPosterior probability, test probability and prior probability be corrected, more Close to truth, p (c_i| it x) is calculated by Bayesian formula, calculates such as following formula in detail：

And i, j≤s, s are classification number.This formula representative sample x belongs to the i-th class and belongs to the difference of the probability of jth class, letter It is denoted as：L_i,j(x)；Difference gets over large sample generic and is easy for determining, conversely, sample is at the fuzzy boundary of comparison.

Select the class probability difference L of unmarked sample_i,j(x) it is less than within the scope of predetermined threshold value δ, assert that the sample is in class Between other i, j on boundary relatively fuzzyyer, there is maximum boundary information content；

Selection is subjected to classification calibration with maximum boundary information content sample by the 3 initialization graders trained；Institute The 3 initialization graders stated are to randomly choose 2/3rds total sample trainings by NB Algorithm to obtain；

Further, predetermined threshold value δ=2.5 × 10^-2It is optimal.

Compared with existing semisupervised classification machine learning method, the present invention provides a kind of semisupervised classifications based on figure Machine learning new method has the advantages that：

The present invention proposes a kind of semisupervised classification new method based on figure, combines supervised learning and unsupervised learning is excellent Point there is training sample to prepare simple, the high advantage of training precision；Compared to general semi-supervised learning method, set forth herein Pre-process active sample labeling method in training set and sorting technique based on figure makes training precision higher, to data sample Label difficulty reduce.

Description of the drawings

Fig. 1, the semisupervised classification new method flow chart based on figure.

Fig. 2, active sample labeling method flow diagram.

Specific implementation mode

Embodiment 1

According to flow described in Fig. 1, the realization of the semisupervised classification new method based on figure is completed, divides training set first, it will Sample is according to having label and without labeling two major classes in training set, and calculates its marker samples accounting；Prepare marker samples to account for Training set than 10%；

To reduce sample labeling difficulty and promoting nicety of grading, using active sample labeling method to the training set of division into Row and processing；Calculate the class probability difference L of all unlabeled exemplars_i,j(x):

And choose all samples for meeting class probability difference and being less than predetermined threshold value δ；And 3 using training completion are initial Change grader to demarcate selected sample, sample after calibration has been divided into exemplar collection, improves training and be concentrated with label Sample accounting；

Predetermined threshold value δ=2.5 × 10 for comparing^-2；

According to the asymmetric sparse weight matrix of training set sample architecture；To passing through solving-optimizing problem Solve asymmetric weight matrix；

Its target function value is calculated according to waiting for that judgement sample inputs Regular terms A can be calculated by input sample in formula, be specifically calculated as：

A=P-PQ+Q

P=X^TX(λI+X^TX)^-1

Symbolization functionIt is exported, class prediction is carried out to treat judgement sample.

Claims

1. a kind of semisupervised classification new method based on figure, includes the following steps：

Step S001：Divide training set；

Training set X=L ∪ U={ x₁,…x_l,x_l+1,…x_l+u, L={ x₁,…x_lIt is to have exemplar on a small quantity, U={ x_l+1,… x_l+uIt is a large amount of unlabeled exemplars；Marker samples accounting

Step S002：Construct asymmetric sparse weight matrix；

In formula,J=1,2 ..., k-1, k+1 ..., n indicates sample x_kBy the sparse decomposition system of remaining sample linear expression Number；

Step S003：Solve asymmetric weight matrix；Pass through solving-optimizing problemIt solves asymmetric Weight matrix；

Step S004：Calculating target function

Step S005：Symbolization functionJudgement sample classification.

2. a kind of semisupervised classification new method based on figure according to claim 1, it is characterised in that：In the step S001 is divided in training set, and using active sample labeling method, label has maximum boundary information content sample, is carried out to sample pre- Processing, includes the following steps：

Step D000：It is calculated as the class probability difference L of all unlabeled exemplars_i,j(x)；The class probability mathematic interpolation is public Formula is：

Step D001：Choose class probability difference L_i,j(x) all unlabeled exemplars for being less than predetermined threshold value δ are used as with maximum Boundary information amount sample；

Step D002：Have maximum boundary information content sample by training 3 initialization graders of completion to carry out classification by choosing Calibration, exemplar collection has been divided by sample after calibration.

3. a kind of semisupervised classification new method based on figure according to claim 2, it is characterised in that：The predetermined threshold value δ=2.5 × 10^-2。