CN105654126A

CN105654126A - Computing equipment, kernel matrix evaluation method and multi-kernel learning method

Info

Publication number: CN105654126A
Application number: CN201511009879.7A
Authority: CN
Inventors: 孙涛; 徐礼锋; 曹莉
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2015-12-29
Filing date: 2015-12-29
Publication date: 2016-06-08

Abstract

An embodiment of the invention provides computing equipment, a kernel matrix evaluation method and a multi-kernel learning method. The computing equipment comprises a central processing unit, a memory, a storage medium and a power supply. The central processing unit is used for constructing an initial kernel matrix by a kernel function according to the characteristics of a training sample, multiplexing the initial kernel matrix by the type label of the training sample for obtaining predicated values of the initial kernel matrix to N training samples, introducing the predicated values of the N training samples of the initial kernel matrix into a symbol function and performing transposition for obtaining a first N-dimensional column vector, multiplying the first N-dimensional column vector by the type label of the training sample for obtaining the kernel matrix, and performing summation on N elements included in the kernel matrix for obtaining discrimination value of the kernel matrix. Because the discrimination value is a number, the computing equipment, the kernel matrix evaluation method and the multi-kernel learning method have advantages of generating a quantitative evaluation result, effectively improving kernel matrix evaluation accuracy and reducing complexity in computing process. The computing equipment, the kernel matrix evaluation method and the multi-kernel learning method can be applied on large-scale data.

Description

A kind of computing equipment, nuclear matrix appraisal procedure and Multiple Kernel Learning method

Technical field

The present invention relates to Data Mining, in particular a kind of computing equipment, nuclear matrix appraisal procedure and Multiple Kernel Learning method.

Background technology

Data mining come from database technology cause mass data and people want to make full use of these data thus obtaining the highly desirable of potential valuable information, data mining is also commonly known as the Knowledge Discovery in data base, it it is automatic or convenient schema extraction, these model representatives are hidden in large database, knowledge in data warehouse or the storage of other bulk informations, it relates to multi-disciplinary field, therefore draw nourishment from from multiple subjects, including database technology, artificial intelligence, neutral net, statistics, pattern recognition, knowledge base system, knowledge is extracted, information retrieval, high-performance calculation and data visualization are many-sided, at present, in oil exploration, financial market, commercial distribution, production marketing, the fields such as medical care insurance, the achievement in research of data mining is obtained for and is widely applied.

Classification is as in Data Mining very important task, and its purpose is one classification function of association or disaggregated model, and wherein, linear classification is very simple also effectively a kind of grader form, for instance logistic regression, bayes method. If data are fairly linear dividing, linear classifier can obtain good classifying quality; Divide if data are not fairly linear, the classifying quality of linear classifier is often unsatisfactory, and Nonlinear Classification is inseparable for original linear data, through non-linear projection a to higher dimensional space, then in higher dimensional space, a classifying face is found, different types of data field separately. Such as, random forest, Nonlinear Support Vector Machines (English full name: SupportVectorMachine, English abbreviation: SVM) etc. Theoretical according to VC dimension (English full name: Vapnik-ChervonenkisDimension), as long as the Spatial Dimension projected to is sufficiently high, it becomes possible to different types of data can be divided completely.

Multiple Kernel Learning (English full name: MultipleKernelLearning, English abbreviation: MKL) it is a kind of effective ways solving Nonlinear Classification problem, Multiple Kernel Learning refers to and is incorporated in unified framework by kernel function by multiple nuclear matrix, by Optimization Solution, a kind of optimum linear between multiple nuclear matrix is found to combine.

Illustrate below in conjunction with the Multiple Kernel Learning method shown in Fig. 1 and Fig. 2, prior art provided:

The training classification process of Multiple Kernel Learning is as follows:

201, by kernel function, training sample is projected to higher dimensional space.

It is positioned at the training sample of the initial data of lower dimensional space, it is possible to be mapped to described higher dimensional space by kernel function.

Multiple different kernel function can be set.

As it is shown in figure 1, each feature for described training sample provides candidate's kernel function that M kind is different.

Assume that the training sample of initial data has P feature, then just total total P*M the first nuclear matrix.

Namely can determine that P*M the first nuclear matrix through step 201, through step 202 and step 203, P*M the first nuclear matrix is synthesized the second nuclear matrix K below.

202, the weighted sum that the second nuclear matrix K is multiple described first nuclear matrix is determined.

Wherein, the second nuclear matrixWherein, d_kFor initial core weight.

Described core weight d_kFor identifying the importance of nuclear matrix.

203, described second nuclear matrix K is substituted into support vector machines formula, train sorter model.

Described training sorter model is for different types of data field separately.

Visible, under many core frameworks, described training sample problem of representation in higher dimensional space transforms into nuclear matrix and core weight d_kSelect permeability, by the different characteristic of isomeric data is inputted respectively correspondence kernel function map, make training sample better be expressed in higher dimensional space, classification accuracy rate or precision of prediction can be significantly improved.

Wherein it is possible to adopt some quick SVM optimization methods, for instance Libsvm, SVMlight etc. Features different for training sample can be converted into different nuclear matrix by support vector machine, and each nuclear matrix has no core weight d_kCorresponding, thus can obtain, Multiple Kernel Learning is different from monokaryon study, and it forms the feature space of different IPs combination, is obtained the core weight d of each nuclear matrix further by adaptive optimization algorithm_k��

204, judge whether the sorter model obtained meets end condition, if the condition of being unsatisfactory for, sorter model is fixed, update core weight d_k��

Wherein, it is judged that whether the sorter model obtained meets the method for end condition optionally for degradation method under incisal plane, gradient.

205, use the core weight updated to re-construct nuclear matrix, train sorter model, so alternately reciprocal, until meeting end condition.

Visible, due to different kernel functions, the nuclear matrix constructed is relatively larger to the performance impact of training sorter model. When specific problem is not had any priori by people time, it is difficult to determine which kind of kernel function of selection and parameter are calculated. So-called core assessment, is that the resolving ability to nuclear matrix is estimated, thus selecting optimum kernel function.

Nuclear arrangement is currently popular a kind of core appraisal procedure. First it construct a perfect nuclear matrix, then measures the similarity of each nuclear matrix and perfect nuclear matrix, assesses the resolving ability of each nuclear matrix. Nuclear matrix K_uDiscriminative power definition as follows:

D i s (K_{u}, K_{i d e a l}) = < K_{u}^{*}, K_{i d e a l}^{*} >_{F}

\begin{matrix} s . t . & K_{u}^{*} = \frac{K_{u}}{\sqrt{< K_{u}, K_{u} >_{F}}}, \end{matrix}

K_{i d e a l}^{*} = \frac{K_{i d e a l}}{\sqrt{< K_{i d e a l}, K_{i d e a l} >_{F}}},

Wherein, K_ideal=yy^TBeing referred to as perfect nuclear matrix, it is to be formed by the transposition inner product of training sample label Yu its own.The similarity between two nuclear matrix is obtained by Frobenius norm calculation

< K_{u}^{*}, K_{i d e a}^{*} >_{F} = t r (K_{u}^{*}, K_{i d e a}^{*}) = Σ_{i, j} K_{u}^{*} (x_{i}, x_{j}), K_{i d e a}^{*} (x_{i}, x_{j}) .

WithCarry out regularization respectively, in order to allow nuclear matrix K_uDistinguish in the scope that force value drops on [-1,1]. When a nuclear matrix is more high with the similarity degree of perfect nuclear matrix, then the resolving ability that this nuclear matrix has is more high.

That the core that prior art provides is assessed disadvantageously, nuclear arrangement only provides result qualitatively, and a result quantified can not be specifically given. This assessment result inaccuracy, and it is difficult to the calculating carrying out formulating.

Summary of the invention

Embodiments provide a kind of nuclear matrix can being provided and quantify the computing equipment of assessment result, nuclear matrix appraisal procedure and Multiple Kernel Learning method.

Embodiment of the present invention first aspect provides a kind of computing equipment, including: central processing unit, memorizer, for storing storage medium and the power supply of application program and/or storage data;

Described central processing unit, for by the kernel function latent structure incipient nucleus matrix according to training sample, described kernel function is for the feature space by the training sample non-linear projection of lower dimensional space to higher-dimension, described kernel function is described training sample tolerance of similarity in described feature space, described incipient nucleus matrix is the real symmetric matrix of N*N, wherein, N is the number of described training sample, for natural number;

Described central processing unit, is additionally operable to described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample;

Described central processing unit, is additionally operable to bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector;

Described central processing unit, is additionally operable to a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix;

Described central processing unit, the N number of element summation being additionally operable to the to comprise described nuclear matrix force value that distinguishes to obtain nuclear matrix, described nuclear matrix distinguish that force value is used to determine whether to bring in support vector machines grader to train disaggregated model described nuclear matrix into.

Adopt the computing equipment shown in the present embodiment, can distinguish that nuclear matrix is estimated by force value according to the nuclear matrix of each nuclear matrix, distinguish that force value is numerical value because of described nuclear matrix, then nuclear matrix can be provided the assessment result of quantization by the computing equipment shown in the present embodiment, and effectively improve the accuracy rate of nuclear matrix assessment, reduce the complexity in calculating process, it is possible to be applied to large-scale data.

In conjunction with embodiment of the present invention first aspect, in the first implementation of embodiment of the present invention first aspect,

Described central processing unit, is additionally operable to that described incipient nucleus matrix carries out leading diagonal and processes.

In the present embodiment, the element on leading diagonal in described incipient nucleus matrix does not weigh the similarity between different elements, then the element on the leading diagonal in incipient nucleus matrix to classification be do not have helpful, for promoting the accuracy of classification, and reducing the complexity calculated, then described incipient nucleus matrix can be carried out leading diagonal process by computing equipment.

The first implementation of embodiment of the present invention first aspect or embodiment of the present invention first aspect, in the second implementation of embodiment of the present invention first aspect,

Described central processing unit, is additionally operable to described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample, particularly as follows:

Described central processing unit, it is additionally operable to the J of described incipient nucleus matrix is arranged the kind label being multiplied by described training sample to obtain the first distance and second distance, wherein, the j-th training sample of incipient nucleus matrix and the distance of all described training samples are shown in the J list of described incipient nucleus matrix, the J of described incipient nucleus matrix is classified as the either rank of described incipient nucleus matrix, the distance of all positive sample of the j-th training sample that described first distance is described incipient nucleus matrix and kind label, described second distance is the j-th training sample distance with all negative samples of kind label of described incipient nucleus matrix,

Described central processing unit, is additionally operable to calculate the difference of described first distance and described second distance and the predictive value of the j-th training sample that the difference of described first distance and described second distance is described incipient nucleus matrix.

In conjunction with the second implementation of embodiment of the present invention first aspect, in the third implementation of embodiment of the present invention first aspect,

Described central processing unit, is additionally operable to described first distance divided by the number of all positive sample of described kind label to obtain the first parameter;

Described central processing unit, is additionally operable to described second distance divided by the number of all negative samples of described kind label to obtain the second parameter;

Described central processing unit, is additionally operable to calculate the difference of described first parameter and described second parameter and the predictive value of the j-th training sample that the difference of described first parameter and described second parameter is described incipient nucleus matrix.

In the present embodiment, computing equipment is when considering positive sample and negative sample number is often different, then in the process calculating the predictive value of j-th training sample of described incipient nucleus matrix, to the weight of the described incipient nucleus Matrix Multiplication weight with positive sample and negative sample, thus effectively avoiding the erroneous judgement because being formed when having larger difference between positive and negative sample size.

In conjunction with the third implementation of the second implementation of embodiment of the present invention first aspect or embodiment of the present invention first aspect, in the 4th kind of implementation of embodiment of the present invention first aspect,

Described central processing unit, is additionally operable to bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector, specifically includes:

Described central processing unit, the predictive value being additionally operable to the j-th training sample by described incipient nucleus matrix brings described sign function into export result value;

Described central processing unit, is additionally operable to synthesize the output result value of N number of training sample of incipient nucleus matrix the vector of N dimension;

Described central processing unit, is additionally operable to the N of the described incipient nucleus matrix vector transposition tieed up to obtain a described N dimensional vector.

In conjunction with the 4th kind of implementation of embodiment of the present invention first aspect to embodiment of the present invention first aspect, in the 5th kind of implementation of embodiment of the present invention first aspect,

Described central processing unit, is additionally operable to a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix, specifically includes:

Described central processing unit, if being additionally operable to the predictive value kind label j-th label equal to described training sample of the j-th training sample of described incipient nucleus matrix, then the j-th element of described nuclear matrix is 1, wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample;

Or,

Described central processing unit, if the predictive value being additionally operable to the j-th training sample of described incipient nucleus matrix is not equal to the kind label j-th label of described training sample, then the j-th element of described nuclear matrix is-1.

Embodiment of the present invention second aspect provides a kind of nuclear matrix appraisal procedure, and described nuclear matrix appraisal procedure is applied to computing equipment, including:

By the kernel function latent structure incipient nucleus matrix according to training sample, described kernel function is for the feature space by the training sample non-linear projection of lower dimensional space to higher-dimension, described kernel function is described training sample tolerance of similarity in described feature space, described incipient nucleus matrix is the real symmetric matrix of N*N, wherein, N is the number of described training sample, for natural number;

By described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample;

Bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector;

A described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix;

N number of element summation force value that distinguishes to obtain nuclear matrix that described nuclear matrix is comprised, described nuclear matrix distinguish that force value is used to determine whether to bring in support vector machines grader to train disaggregated model described nuclear matrix into.

Adopt the nuclear matrix appraisal procedure shown in the present embodiment, can distinguish that nuclear matrix is estimated by force value according to the nuclear matrix of each nuclear matrix, distinguish that force value is numerical value because of described nuclear matrix, then nuclear matrix can be provided the assessment result of quantization by the nuclear matrix appraisal procedure shown in the present embodiment, and effectively improve the accuracy rate of nuclear matrix assessment, reduce the complexity in calculating process, it is possible to be applied to large-scale data.

In conjunction with embodiment of the present invention second aspect, in the first implementation of embodiment of the present invention second aspect,

Described by kernel function according to after the latent structure incipient nucleus matrix of training sample, described method also includes:

Described incipient nucleus matrix carries out leading diagonal process.

In the present embodiment, the element on leading diagonal in described incipient nucleus matrix does not weigh the similarity between different elements, then the element on the leading diagonal in incipient nucleus matrix to classification be do not have helpful, for promoting the accuracy of classification, and reducing the complexity calculated, then described incipient nucleus matrix can be carried out leading diagonal process by the nuclear matrix appraisal procedure shown in the present embodiment.

In conjunction with the first implementation of embodiment of the present invention second aspect or embodiment of the present invention second aspect, in the second implementation of embodiment of the present invention second aspect,

Described described incipient nucleus Matrix Multiplication is included to obtain the incipient nucleus matrix predictive value to N number of training sample with the kind label of training sample:

The J of described incipient nucleus matrix is arranged the kind label being multiplied by described training sample to obtain the first distance and second distance, wherein, the j-th training sample of incipient nucleus matrix and the distance of all described training samples are shown in the J list of described incipient nucleus matrix, the J of described incipient nucleus matrix is classified as the either rank of described incipient nucleus matrix, the distance of all positive sample of the j-th training sample that described first distance is described incipient nucleus matrix and kind label, described second distance is the j-th training sample distance with all negative samples of kind label of described incipient nucleus matrix,

Calculate the difference of described first distance and described second distance and the predictive value of the j-th training sample that the difference of described first distance and described second distance is described incipient nucleus matrix.

In conjunction with the second implementation of inventive embodiments second aspect, in the third implementation of embodiment of the present invention second aspect,

The described J by described incipient nucleus matrix arranges the kind label being multiplied by described training sample with after obtaining the first distance and second distance, and described method also includes:

By described first distance divided by the number of all positive sample of described kind label to obtain the first parameter;

By described second distance divided by the number of all negative samples of described kind label to obtain the second parameter;

Calculate described first parameter and the difference of described second parameter and the predictive value of the j-th training sample that the difference of described first parameter and described second parameter is described incipient nucleus matrix.

In the present embodiment, when considering positive sample and negative sample number is often different, then in the process calculating the predictive value of j-th training sample of described incipient nucleus matrix, to the weight of the described incipient nucleus Matrix Multiplication weight with positive sample and negative sample, thus effectively avoiding the erroneous judgement because being formed when having larger difference between positive and negative sample size.

In conjunction with the third implementation of the second implementation of inventive embodiments second aspect or embodiment of the present invention second aspect, in the 4th kind of implementation of embodiment of the present invention second aspect,

The described predictive value by described incipient nucleus matrix N training sample is brought sign function transposition into and is included to obtain a N dimensional vector:

The predictive value of the j-th training sample of described incipient nucleus matrix is brought described sign function into export result value;

The output result value of N number of training sample of incipient nucleus matrix is synthesized the vector of N dimension;

By the N of the described incipient nucleus matrix vector transposition tieed up to obtain a described N dimensional vector.

In conjunction with the method described in the 4th kind of any one of implementation of embodiment of the present invention second aspect to embodiment of the present invention second aspect, in the 5th kind of implementation of embodiment of the present invention second aspect,

Described the kind label that a described N dimensional vector is multiplied by described training sample is included to obtain nuclear matrix:

If the predictive value of the j-th training sample of described incipient nucleus matrix is equal to the kind label j-th label of described training sample, then the j-th element of described nuclear matrix is 1, wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample;

Or,

If the predictive value of the j-th training sample of described incipient nucleus matrix is not equal to the kind label j-th label of described training sample, then the j-th element of described nuclear matrix is-1.

The embodiment of the present invention third aspect provides a kind of Multiple Kernel Learning method based on nuclear matrix appraisal procedure, described Multiple Kernel Learning method is applied to computing equipment, described nuclear matrix appraisal procedure such as embodiment of the present invention second aspect to embodiment of the present invention second aspect the 5th kind of any one of implementation shown in, described Multiple Kernel Learning method includes:

M nuclear matrix of each latent structure to described training sample respectively according to kernel functions different for M, wherein, M is natural number;

M corresponding to each feature of described training sample described nuclear matrix distinguish force value, described M nuclear matrix is selected at least one target nuclear matrix;

Mean vector g will be in_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set, wherein, described mean vector g_ave=(1/N) ��_pg_p, g_pFor selected arbitrary described target nuclear matrix, described reference vector g_ref=o+ �� g_ave, o represents the diagonal vector of first quartile, and �� is a constant, and described reference vector g_refWith described mean vector g_avePerpendicular;

Described target nuclear matrix set is brought in support vector machines grader to train disaggregated model.

Multiple Kernel Learning method shown in the present embodiment improves the speed in Multiple Kernel Learning process, it is to avoid alternative optimization process in Multiple Kernel Learning process. Whole computing cost is nearly equivalent to running a SVM, and saves memory space, the Multiple Kernel Learning method shown in prior art, and the space complexity of one nuclear matrix of storage is O (N²), become quadratic relationship with number of samples. And in embodiments of the present invention, nuclear matrix is after binary core is assessed, it is thus only necessary to complexity O (N), linear with number of samples, the saving of memory headroom is very huge by this.

In conjunction with the embodiment of the present invention third aspect, in the first implementation of the embodiment of the present invention third aspect,

Described M corresponding to each feature of described training sample described nuclear matrix distinguish force value, described M nuclear matrix is selected at least one target nuclear matrix and includes:

From the described nuclear matrix of the M corresponding to each feature of described training sample, selected described nuclear matrix distinguishes that the maximum nuclear matrix of force value is described target nuclear matrix.

In conjunction with the first implementation of the embodiment of the present invention third aspect or the embodiment of the present invention third aspect, in the second implementation of the embodiment of the present invention third aspect,

Described will be in mean vector g_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set after, described method also includes:

Will be located in the first object nuclear matrix in described target nuclear matrix set to be multiplied by mutually with the second target nuclear matrix and obtain the 2nd N dimensional vector, described first object nuclear matrix and described second target nuclear matrix are target nuclear matrix described in any two differed in described target nuclear matrix set;

The N number of element comprised by described 2nd N dimensional vector is sued for peace to obtain redundant prediction value;

If described redundant prediction value is more than or equal to predetermined threshold value, then described first object nuclear matrix is deleted from described target nuclear matrix set.

Multiple Kernel Learning method shown in the present embodiment can assess the similarity between nuclear matrix, effectively avoids the redundancy that closely similar nuclear matrix superposition can bring, thus avoiding the over-fitting caused in data categorizing process.

The embodiment provided for a better understanding of the present invention, illustrates the embodiment of the present invention below in conjunction with concrete application scenarios:

It is applied in gastric cancer detection with the present embodiment below and illustrates:

Lymphatic metastasis detection (English full name: LymphNodeMetastasis, English abbreviation: LNM), preoperative to gastric cancer is extremely important by stages.

Traditionally, lymph node size is used as the reference index of lymph node diagnosis,

Later, clinical medicine found that it was inadequate for simply using lymph node size.

Because big lymph node is probably and is caused by inflammation, and the small lymph node that those are not concerned is likely to be real metastatic lymph node.

So, only rely only on the size of lymph node, it is possible to form erroneous judgement.

The combination of multinomial physical signs uses the inevitable choice just becoming lymph node diagnosis.

But, the physical signs of people has a lot, and which physical signs is to become clinical medical important topic with lymph node diagnosis is relevant.

Lymph node test experience in gastric cancer is example, and the data that this test uses come from DKFZ of Medical School of Peking University.

Every patient has 18 physical signs: sex, age, knub position, tumor maximum radius, tumor thickness, tumor pattern, serous coat is invaded, the invasion degree of depth, Bollman type (English full name: Borrmanntype), enhancement mode (English full name: enhancementpattern), number of lymph nodes, maximum lymph node size, lymph node numbers, maximum lymph node size, lymph node (English full name: station), the CT decay of lymph node, the minor axis of lymph node and draw ratio, lymph node is distributed, histological type (English full name: pathologicaltype) and serum cancer training antigen (English full name: CEAofserum).

18 physical signs each patient being had are as the feature input calculating equipment of training sample;

Each index M nuclear matrix of structure of 18 physical signs that each patient is had respectively by computing equipment according to M different kernel function;

Described incipient nucleus matrix is carried out leading diagonal and processes by computing equipment;

The J of described incipient nucleus matrix is arranged the kind label being multiplied by described training sample to obtain the first distance and second distance by computing equipment, wherein, the j-th training sample of incipient nucleus matrix and the distance of all described training samples are shown in the J list of described incipient nucleus matrix, the J of described incipient nucleus matrix is classified as the either rank of described incipient nucleus matrix, the distance of all positive sample of the j-th training sample that described first distance is described incipient nucleus matrix and kind label, described second distance is the j-th training sample distance with all negative samples of kind label of described incipient nucleus matrix,

Computing equipment calculates the difference of described first distance and described second distance and the predictive value of the j-th training sample that the difference of described first distance and described second distance is described incipient nucleus matrix;

Computing equipment brings the predictive value of the j-th training sample of described incipient nucleus matrix into described sign function to export result value;

The output result value of N number of training sample of incipient nucleus matrix is synthesized the vector of N dimension by computing equipment;

Computing equipment by the N of the described incipient nucleus matrix vector transposition tieed up to obtain a described N dimensional vector;

Computing equipment is if it is determined that the predictive value of j-th training sample of described incipient nucleus matrix is equal to the kind label j-th label of described training sample, the j-th element then determining described nuclear matrix is 1, wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample; Or, computing equipment is if it is determined that the predictive value of j-th training sample of described incipient nucleus matrix is not equal to the kind label j-th label of described training sample, it is determined that the j-th element of described nuclear matrix is-1;

The summation of N number of element that described nuclear matrix is comprised by computing equipment distinguishes force value with what obtain nuclear matrix;

Computing equipment from the described nuclear matrix of the M corresponding to each feature of described training sample selected described nuclear matrix distinguish that the maximum nuclear matrix of force value is described target nuclear matrix;

Computing equipment will be in mean vector g_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set, wherein, described mean vector g_ave=(1/N) ��_pg_p, g_pFor selected arbitrary described target nuclear matrix, described reference vector g_ref=o+ �� g_ave, o represents the diagonal vector of first quartile, and �� is a constant, and described reference vector g_refWith described mean vector g_avePerpendicular;

Described target nuclear matrix set is brought in support vector machines grader to train disaggregated model by computing equipment.

For the advantage shown in the embodiment of the present invention is described, illustrate below in conjunction with chart:

One, adopt the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment that the nicety of grading of the Nonlinear Support Vector Machines SVM shown in lymphatic metastasis nicety of grading and prior art, Multiple Kernel Learning MKL is compared to refer to shown in table 1.

Wherein, table 1 is the Nonlinear Support Vector Machines SVM shown in the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment, prior art and the Multiple Kernel Learning MKL comparison for the nicety of grading of 30% sample provided, 50% sample and 70% sample.

Table 1

	30% sample	50% sample	70% sample
				SVM-LN	65.3	65.4	69.7
MKL	74.1	79.6	83.7
				PMKL-L��	75.6	80.1	84.0

Visible, adopting the Multiple Kernel Learning method shown in the present embodiment, because distinguishing that nuclear matrix is estimated by force value according to the nuclear matrix of each nuclear matrix, and distinguishing that force value carries out Multiple Kernel Learning according to described nuclear matrix, thus effectively improving the precision of classification.

Two, adopt the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment to relatively the referring to shown in table 2 the lymphatic metastasis training time of the Multiple Kernel Learning MKL shown in lymphatic metastasis training time and prior art.

Wherein, table 2 is the comparison for the training time of 30% sample provided, 50% sample and 70% sample of the Multiple Kernel Learning MKL shown in the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment, prior art.

Table 2

	30% sample	50% sample	70% sample
				MKL	1356.5	3674.3	6226.7
PMKL-L��	206.7	487.8	513.4

Visible, the Multiple Kernel Learning method shown in the present embodiment of employing, because distinguishing that nuclear matrix is estimated by force value according to the nuclear matrix of each nuclear matrix, thus force value can be distinguished according to described nuclear matrix, the nuclear matrix being unprofitable to classification is deleted, thus greatly reducing amount of calculation, reduce the complexity of calculating, thus reducing the training time.

Three, adopt the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment that lymph node selects relatively the referring to shown in table 3 lymph node selection daughter nucleus number of the Multiple Kernel Learning MKL shown in daughter nucleus number and prior art.

Wherein, table 3 is the comparison selecting daughter nucleus number for 30% sample provided, 50% sample and 70% sample of the Multiple Kernel Learning MKL shown in the Multiple Kernel Learning method PMKL-L �� shown in the present embodiment, prior art.

Table 3

	30% sample	50% sample	70% sample
				MKL	19.6	20.6	22.1
PMKL-L��	5.3	5.5	5.4

Visible, the Multiple Kernel Learning method shown in the present embodiment of employing, because distinguishing that nuclear matrix is estimated by force value according to the nuclear matrix of each nuclear matrix, thus force value can be distinguished according to described nuclear matrix, the nuclear matrix being unprofitable to classification is deleted, thus greatly reducing the number selecting daughter nucleus, the number of the daughter nucleus because selecting is less, then reduce amount of calculation, and reduce the complexity of calculating, and then greatly reduce taking of computing equipment storage medium.

Present embodiments provide a kind of computing equipment, nuclear matrix appraisal procedure and Multiple Kernel Learning method, this computing equipment includes: central processing unit, memorizer, for storing storage medium and the power supply of application program and/or storage data, described central processing unit is for by the kernel function latent structure incipient nucleus matrix according to training sample, by described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample, bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector, a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix, the N number of element summation described nuclear matrix comprised distinguishes force value with what obtain nuclear matrix, distinguish that force value is numerical value described in cause, then nuclear matrix can be provided the assessment result of quantization by the present embodiment, and effectively improve the accuracy rate of nuclear matrix assessment, reduce the complexity in calculating process, can apply to large-scale data.

Accompanying drawing explanation

A kind of schematic flow sheet of the Multiple Kernel Learning method that Fig. 1 provides for prior art;

A kind of flow chart of steps of the Multiple Kernel Learning method that Fig. 2 provides for prior art;

Fig. 3 is a kind of embodiment flow chart of steps of nuclear matrix appraisal procedure provided by the present invention;

Fig. 4 is that nuclear matrix appraisal procedure provided by the present invention calculates the incipient nucleus matrix a kind of embodiment schematic diagram to the predictive value of N number of training sample;

Fig. 5 is that nuclear matrix appraisal procedure provided by the present invention calculates the incipient nucleus matrix another kind of embodiment schematic diagram to the predictive value of N number of training sample;

Fig. 6 is a kind of embodiment schematic diagram that nuclear matrix appraisal procedure provided by the present invention obtains nuclear matrix;

Fig. 7 is a kind of embodiment flow chart of steps of Multiple Kernel Learning method provided by the present invention;

Fig. 8 is a kind of schematic flow sheet of Multiple Kernel Learning method provided by the present invention;

Fig. 9 is a kind of embodiment schematic diagram of Multiple Kernel Learning method computing redundancy predictive value provided by the present invention;

Figure 10 is a kind of example structure schematic diagram of computing equipment provided by the present invention.

Detailed description of the invention

First the computing equipment that can apply the nuclear matrix appraisal procedure shown in the present embodiment is illustrated:

Existing big data prediction class business generally adopts linear classification algorithm, such as logistic regression, Linear SVM etc., calculates speed quickly, but nicety of grading is frequently not very satisfactory. Non-linear reach higher nicety of grading, but time complexity generally becomes with sample number square or cube relation, is very difficult to apply in the calculating of big data. And practical application there are many business of linearly inseparable, at this time linear classification algorithm cannot solving practical problems.

And the computing equipment that the present embodiment provides can be used in the business of linearly inseparable, thus data being carried out quickly, classify accurately, and can solve the problem that big data prediction class business and problem, the credit card evaluation for credit degree of such as financial field and prediction, the off-network user in predicting of carrier network, police field suspect prediction and meteorological disaster prediction, geological hazards prediction etc., purposes is widely.

Described computing equipment is not limited by the present embodiment, as long as described computing equipment can properly functioning big data platform, for instance, described computing equipment can be the electronic equipment that computer etc. have computing capability.

Described computing equipment can carry out the calculating of Nonlinear Support Vector Machines, described computing equipment can change into linear separability linearly inseparable, by a nonlinear mapping, the data characteristics in lower dimensional space is mapped in higher dimensional space, higher dimensional space is asked linear optimal Optimal Separating Hyperplane.

The idiographic flow that described computing equipment realizes Multiple Kernel Learning method is as follows:

First, computing equipment shown in the present embodiment is capable of the nuclear matrix appraisal procedure shown in Fig. 3, this nuclear matrix appraisal procedure is for being estimated the nuclear matrix in Multiple Kernel Learning flow process, to select the nuclear matrix of optimum, it is described in detail below in conjunction with the nuclear matrix appraisal procedure shown in Fig. 3, the present embodiment provided, the nuclear matrix appraisal procedure provided by the present embodiment can under the premise reducing computation complexity, it is achieved the selection to kernel function.

301, by the kernel function latent structure incipient nucleus matrix according to training sample.

Described kernel function is for the feature space by the training sample non-linear projection of lower dimensional space to higher-dimension, and described kernel function is described training sample tolerance of similarity in described feature space.

Wherein, illustrating of described incipient nucleus matrix also referring to the first nuclear matrix shown in Fig. 2 step 201, specifically can not repeat in the present embodiment.

Incipient nucleus matrix shown in the present embodiment is:

Concrete, described incipient nucleus matrix has reacted the described training sample position relationship at described higher dimensional space.

More specifically, described incipient nucleus matrix is the real symmetric matrix of N*N, and wherein, N is the number of described training sample, for natural number.

In described incipient nucleus matrix, the i-th row, jth row element a_ijRepresent the distance between i-th training sample and jth training sample.

302, described incipient nucleus matrix carries out leading diagonal process.

The element on leading diagonal in described incipient nucleus matrix is a_ii, wherein, the value of i is the natural number between 1 to N.

Wherein, a_iiRepresent the i-th training sample distance with self.

Visible, the element on leading diagonal in incipient nucleus matrix does not weigh the similarity between different elements, then the element on the leading diagonal in incipient nucleus matrix to classification be do not have helpful.

In the present embodiment, for promoting the accuracy of classification, and reduce the complexity calculated, then described incipient nucleus matrix can be carried out leading diagonal and process.

Concrete, the matrix after going leading diagonal to process is as follows:

Need it is clear that the step 302 in the present embodiment is optional step, do not limit.

303, the kind label of training sample is determined.

By shown on it can be seen that the present embodiment has N number of training sample.

The kind label being determined training sample by training sample is prior art, does not specifically repeat in the present embodiment.

Hereinafter lift concrete application scenarios the kind label how determining training sample is illustrative, need it is clear that the present embodiment is to determining that the mode of kind label is optional example, does not limit.

This application scene is based on the tracking scene of video:

In this application scene, utilize grader that the image sheet (i.e. training sample) of every two field picture of video sequence is labeled as target, one target generally has certain position in every two field picture, and the position that target is in every two field picture of video sequence can be defined as track.

If training sample leaves the right or normal track, mark is close, namely training sample and the distance of described track are less than default value, then this training sample is positive sample, if training sample leaves the right or normal track, mark is far, namely training sample and the distance of described track are more than default value, then this training sample is negative sample.

The column vector of the N dimension that the kind label of described training sample is the positive sample of described training sample and the negative sample of described training sample is formed.

304, the J of described incipient nucleus matrix is arranged the kind label being multiplied by described training sample to obtain the first distance and second distance.

Wherein, the j-th training sample of incipient nucleus matrix and the distance of all described training samples are shown in the J list of described incipient nucleus matrix.

The J of described incipient nucleus matrix is classified as the either rank of described incipient nucleus matrix, and namely the span of J is the arbitrary natural number in 1 to N.

Such as, if J value is 2, then the 2nd training sample of incipient nucleus matrix and the distance of all described training samples are shown in the 2nd list of described incipient nucleus matrix, e.g., and a₁₂Represent the distance between the 2nd training sample of described incipient nucleus matrix and the 1st training sample, a₃₂Represent the distance between the 2nd training sample of described incipient nucleus matrix and the 3rd training sample, a_N2Represent the 2nd distance between training sample and n-th training sample of described incipient nucleus matrix.

Showing example with Fig. 4, the y in Fig. 4 is the kind label of the described training sample shown in the present embodiment.

Need it is clear that the kind label shown in the present embodiment Fig. 4 is optional example, do not limit.

The distance of all positive sample of the j-th training sample that described first distance is described incipient nucleus matrix and kind label, described second distance is the j-th training sample distance with all negative samples of kind label of described incipient nucleus matrix.

305, the predictive value of the j-th training sample of incipient nucleus matrix is calculated.

The present embodiment can directly using the difference of described first distance and the described second distance predictive value as the j-th training sample of described incipient nucleus matrix.

Hereinafter lift concrete sample calculation to illustrate, show example with the 1st of the described incipient nucleus matrix shown in the broken box shown in Fig. 4 the row, namely illustrate for 1 with J value:

As shown in Figure 4, when needing the predictive value calculating the 1st training sample of described incipient nucleus matrix, then need to be multiplied by the 1st row of described incipient nucleus matrix the kind label of described training sample, the first distance and the second distance of the 1st row of described incipient nucleus matrix can be obtained, and calculate the predictive value of the 1st training sample of described incipient nucleus matrix.

Visible, if the 1st of described incipient nucleus matrix the training sample is bigger with the similarity of the positive sample of described kind label, then the result obtained be exactly on the occasion of.

If the 1st of described incipient nucleus matrix the training sample is bigger with the similarity of the negative sample of described kind label, then the result obtained is exactly negative value.

Optionally, for effectively reducing the erroneous judgement of categorizing process, then shown in Figure 5, step 305 further comprises:

By the J of described incipient nucleus matrix described first distance arranged number N+to obtain the first parameter divided by all positive sample of described kind label;

By the J of the described incipient nucleus matrix described second distance the arranged number N-to obtain the second parameter divided by all negative samples of described kind label;

Hereinafter lift concrete sample calculation to illustrate, show example with the 2nd of the described incipient nucleus matrix shown in the broken box shown in Fig. 5 the row, namely illustrate for 2 with J value:

As shown in Figure 5, when needing the predictive value calculating the 2nd training sample of described incipient nucleus matrix, then need to be multiplied by the 2nd row of described incipient nucleus matrix the kind label of described training sample, the first distance and the second distance of the 2nd row of described incipient nucleus matrix can be obtained;

Further, by the first distance of the 2nd of the incipient nucleus matrix row number N+to obtain the first parameter divided by all positive sample of described kind label;

By the second distance of the 2nd of the incipient nucleus matrix row number N-to obtain the second parameter divided by all negative samples of described kind label;

The predictive value of the 2nd training sample of described incipient nucleus matrix is the first parameter and the difference of described second parameter.

Visible, if the 2nd of described incipient nucleus matrix the training sample is bigger with the similarity of the positive sample of described kind label, then the result obtained be exactly on the occasion of.

If the 2nd of described incipient nucleus matrix the training sample is bigger with the similarity of the negative sample of described kind label, then the result obtained is exactly negative value.

When considering positive sample and negative sample number is often different, then in the process calculating the predictive value of j-th training sample of described incipient nucleus matrix, to the weight of the described incipient nucleus Matrix Multiplication weight with positive sample and negative sample, thus effectively avoiding the erroneous judgement because being formed when having larger difference between positive and negative sample size.

306, the predictive value of the j-th training sample of described incipient nucleus matrix is brought into described sign function to export result value.

By the described predictive value obtained via step 305 after mathematical symbol function sign seeks symbol, it is possible to obtain its output result.

If described predictive value be on the occasion of, then described output result is+1.

Such as, if the 1st of described incipient nucleus matrix the training sample output result value is 1, then illustrate the 1st training sample of described incipient nucleus matrix from positive sample distance closer to.

If described predictive value is negative value, then described output result is-1.

Such as, if the 1st of described incipient nucleus matrix the training sample output result value is-1, then illustrate the 1st training sample of described incipient nucleus matrix from negative sample distance closer to.

307, the output result value of N number of training sample of incipient nucleus matrix is synthesized the vector of N dimension.

The output result value of all training samples in described incipient nucleus matrix can be obtained via step 306, then the output result value of N number of training sample of incipient nucleus matrix can be synthesized the vector of N dimension.

308, by the N of the described incipient nucleus matrix vector transposition tieed up to obtain a described N dimensional vector.

Optionally, a described N dimensional vector can the predictive value transposition of N number of training sample of described incipient nucleus matrix as shown in Figure 5 to obtain, namely as shown in Figure 6, column vector N is a described N dimensional vector.

Optionally, a described N dimensional vector can the predictive value transposition of N number of training sample of described incipient nucleus matrix as shown in Figure 4 to obtain.

The present embodiment show example with Fig. 6 and carries out exemplary explanation, does not limit.

309, a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix.

Illustrating of the kind label of described training sample please refer to shown in step 303, does not specifically repeat in this step.

Concrete, if the predictive value of the j-th training sample of described incipient nucleus matrix is equal to the kind label j-th label of described training sample, then the j-th element of described nuclear matrix is 1.

Wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample.

For Fig. 6, the predictive value of first training sample in a described N dimensional vector N the 1st label equal to kind label y, then the 1st element of described nuclear matrix D is 1.

The predictive value of second training sample in a described N dimensional vector N is not equal to the 2nd label of kind label y, then the 2nd element of described nuclear matrix D is-1.

The like, until determining the value of all elements of described nuclear matrix.

310, the N number of element summation described nuclear matrix comprised distinguishes force value with what obtain nuclear matrix.

As shown in Figure 6, the N number of element summation comprised by described nuclear matrix D distinguishes force value with what obtain nuclear matrix.

Described nuclear matrix distinguish that force value is used to determine whether to bring in support vector machines grader to train disaggregated model described nuclear matrix into.

Wherein, nuclear matrix distinguish that force value is more big, represent that the resolving ability of this nuclear matrix is more high, nuclear matrix distinguish that force value is more little, represent that the resolving ability of this nuclear matrix is more low.

Described computing equipment can also realize the Multiple Kernel Learning method shown in Fig. 7 and Fig. 8 based on the nuclear matrix appraisal procedure shown in Fig. 3 to Fig. 6:

701, M nuclear matrix of each latent structure to described training sample respectively according to kernel functions different for M.

Wherein, M is natural number.

The detailed process of the latent structure nuclear matrix according to training sample is prior art, does not specifically repeat in the present embodiment.

702, the described nuclear matrix of the M corresponding to each feature of described training sample distinguish force value, described M nuclear matrix is selected at least one target nuclear matrix.

What can calculate each nuclear matrix in the described nuclear matrix of the M corresponding to each feature of described training sample via the nuclear matrix appraisal procedure shown in Fig. 3 to Fig. 6 distinguishes force value.

Concrete how to calculate each nuclear matrix distinguish please the referring to shown in Fig. 3 to Fig. 6 of force value, be not specifically described further in the present embodiment.

Nuclear matrix appraisal procedure shown in Fig. 3 to Fig. 6 is because only that 1 exports result with-1 two kind, so binary core assessment (as shown in Figure 9) can be referred to as.

In the present embodiment, target nuclear matrix can be selected from the M corresponding to each feature of described training sample described nuclear matrix;

Wherein, described target nuclear matrix be described nuclear matrix distinguish the nuclear matrix that force value is maximum.

Optionally, the M corresponding to each feature of described training sample described nuclear matrix can being ranked up, and then determine that target nuclear matrix is the nuclear matrix of S before sequence, wherein, S is the natural number more than 1.

Certainly, the quantity of described target nuclear matrix is not limited by the present embodiment, and the present embodiment carries out exemplary explanation with described target nuclear matrix for one.

703, mean vector g will be in_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set.

Visible, through step 702, the corresponding target nuclear matrix of each feature of described training sample.

Because described training sample has P feature, then can determine that P target nuclear matrix { K through step 702_1best,K_2best,K_3best��K_Pbest}��

Step 703 is for being undertaken integrating to form target nuclear matrix set by P target nuclear matrix.

Concrete, the present embodiment is by integrating P target nuclear matrix based on the integrated approach of direction sequencing (English full name: OrientationOrderingOO).

Wherein, direction sequencing is a kind of integrated approach, it is possible to the result of multiple graders, be integrated into final result.

More specifically, after step 702, each target nuclear matrix is defined as g_p, g_pFor arbitrary described target nuclear matrix selected in described target nuclear matrix set.

Wherein, if g_pIt is in first quartile, just represents pth target nuclear matrix and all samples are all identified correctly.

The mean vector g of all described target nuclear matrix_ave=(1/N) ��_pg_p��

According to described mean vector g_aveDetermine reference vector g_ref, wherein, reference vector g_refRepresent that the diagonal vector of first quartile is at described mean vector g_aveThe projection in plane opened.

Concrete, make described reference vector and the perpendicular g of described mean vector_ref��g_ave��

More specifically, described reference vector g_ref=o+ �� g_ave��

Wherein, o represents the diagonal vector of first quartile, and �� is a constant.

Then can draw ��=-og_ave/|g_ave|²��

�� is actually equivalent to one big torsion of distribution and those training sample places being averaged vector division mistake is tieed up.

If a target nuclear matrix is in g in the middle of the angle of described mean vector and described reference vector_p��qua(g_ave,g_ref), then this target nuclear matrix is exactly the nuclear matrix being of value to classification.

In the present embodiment, mean vector g will be in_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated in target nuclear matrix set, and filter out and be not at mean vector g_aveWith reference vector g_refAngle in the middle of target nuclear matrix, and then can by be unprofitable to classification target nuclear matrix delete.

Target nuclear matrix set is can determine that through step 703, and the target nuclear matrix being arranged in described target nuclear matrix set is all the nuclear matrix being of value to classification, then can be trained the step of disaggregated model according to the described target nuclear matrix being arranged in target nuclear matrix set.

Optionally, the Multiple Kernel Learning method shown in the present embodiment, the described target nuclear matrix that can further be pointed in target nuclear matrix set is screened, thus removing the target nuclear matrix of redundancy as much as possible.

In the present embodiment, the detailed process of the target nuclear matrix removing redundancy asks for an interview step 704 as follows to shown in step 706.

704, will be located in the first object nuclear matrix in described target nuclear matrix set to be multiplied by mutually with the second target nuclear matrix and obtain the 2nd N dimensional vector.

Concrete, if two the target nuclear matrix similarity system design being arranged in described target nuclear matrix set are big, then can cause the redundancy in categorizing process, and add the complexity of calculating, then in the present embodiment, different target nuclear matrix can be multiplied by mutually and determine that whether two target nuclear matrix similaritys are relatively larger.

In the present embodiment, described first object nuclear matrix and described second target nuclear matrix are target nuclear matrix described in any two differed in described target nuclear matrix set.

Concrete, described second target nuclear matrix is in described target nuclear matrix set except described first object nuclear matrix, either objective nuclear matrix.

Shown in Figure 9, described first object nuclear matrix K1 is multiplied by with described second target nuclear matrix K2 phase and obtains the 2nd N dimensional vector N.

705, the N number of element comprised by described 2nd N dimensional vector is sued for peace to obtain redundant prediction value.

If 706 described redundant prediction values are more than or equal to predetermined threshold value, then described first object nuclear matrix is deleted from described target nuclear matrix set.

Described predetermined threshold value is not limited by the present embodiment, as long as the redundant prediction value between two described target nuclear matrix is more than or equal to predetermined threshold value, redundancy occur in two described target nuclear matrix.

Visible, shown in step 704 to step 706, in described first object nuclear matrix and described target nuclear matrix set, when the redundant prediction value between all the other target nuclear matrix except self is respectively less than described predetermined threshold value, described first object nuclear matrix just can determine whether to stay in described target nuclear matrix.

If described first object nuclear matrix is with described target nuclear matrix set, when the redundant prediction value between either objective nuclear matrix except self is more than described predetermined threshold value, then described first object nuclear matrix will be deleted from described target nuclear matrix set.

In the present embodiment, deleting in the process of target nuclear matrix of redundancy in described target nuclear matrix set, the complexity of whole search only only has O (Nlog (N)).

707, described target nuclear matrix set is brought in support vector machines grader to train disaggregated model.

In the present embodiment, the detailed process that support vector machines grader trains disaggregated model according to target nuclear matrix set please refer to shown in prior art, does not specifically repeat in the present embodiment.

Multiple Kernel Learning method shown in the present embodiment improves the speed in Multiple Kernel Learning process, it is to avoid alternative optimization process in Multiple Kernel Learning process.Whole computing cost is nearly equivalent to running a SVM, and saves memory space, the Multiple Kernel Learning method shown in prior art, and the space complexity of one nuclear matrix of storage is O (N²), become quadratic relationship with number of samples. And in embodiments of the present invention, nuclear matrix is after binary core is assessed, it is thus only necessary to complexity O (N), linear with number of samples, the saving of memory headroom is very huge by this.

And the Multiple Kernel Learning method shown in the present embodiment can assess the similarity between nuclear matrix, effectively avoid the redundancy that closely similar nuclear matrix superposition can bring, thus avoiding the over-fitting caused in data categorizing process.

Below in conjunction with shown in Figure 10, the concrete structure of described computing equipment is described in detail:

Figure 10 is a kind of computing equipment structural representation that the embodiment of the present invention provides, this computing equipment 1000 can produce relatively larger difference because of configuration or performance difference, one or more central processing units (centralprocessingunits can be included, CPU) 1022 (such as, one or more processors) and memorizer 1032, the storage medium 1030 (such as one or more mass memory units) of one or more storage application programs 1042 or data 1044.

Wherein, memorizer 1032 and storage medium 1030 can be of short duration storage or persistently store. The program being stored in storage medium 1030 can include one or more modules (diagram does not mark), and each module can include a series of command operatings in computing equipment.

Further, central processing unit 1022 could be arranged to communicate with storage medium 1030, performs a series of command operatings in storage medium 1030 on computing equipment 1000.

Computing equipment 1000 can also include one or more power supplys 1026, one or more wired or wireless network interfaces 1050, one or more input/output interfaces 1058, and/or, one or more operating systems 1041, for instance WindowsServerTM, MacOSXTM, UnixTM, LinuxTM, FreeBSDTM etc.

In above-described embodiment, the step performed by computing equipment can based on the computing equipment structure shown in this Figure 10.

Described central processing unit 1022, for by the kernel function latent structure incipient nucleus matrix according to training sample, described kernel function is for the feature space by the training sample non-linear projection of lower dimensional space to higher-dimension, described kernel function is described training sample tolerance of similarity in described feature space, described incipient nucleus matrix is the real symmetric matrix of N*N, wherein, N is the number of described training sample, for natural number;

Described central processing unit 1022, is additionally operable to described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample;

Described central processing unit 1022, is additionally operable to bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector;

Described central processing unit 1022, is additionally operable to a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix;

Described central processing unit 1022, the N number of element summation being additionally operable to the to comprise described nuclear matrix force value that distinguishes to obtain nuclear matrix, described nuclear matrix distinguish that force value is used to determine whether to bring in support vector machines grader to train disaggregated model described nuclear matrix into.

Optionally, described central processing unit 1022, it is additionally operable to that described incipient nucleus matrix is carried out leading diagonal and processes.

Optionally, described central processing unit 1022, it is additionally operable to the J of described incipient nucleus matrix is arranged the kind label being multiplied by described training sample to obtain the first distance and second distance, wherein, the j-th training sample of incipient nucleus matrix and the distance of all described training samples are shown in the J list of described incipient nucleus matrix, the J of described incipient nucleus matrix is classified as the either rank of described incipient nucleus matrix, the distance of all positive sample of the j-th training sample that described first distance is described incipient nucleus matrix and kind label, described second distance is the j-th training sample distance with all negative samples of kind label of described incipient nucleus matrix,

Described central processing unit 1022, is additionally operable to calculate the difference of described first distance and described second distance and the predictive value of the j-th training sample that the difference of described first distance and described second distance is described incipient nucleus matrix.

Described central processing unit 1022, is additionally operable to described first distance divided by the number of all positive sample of described kind label to obtain the first parameter;

Described central processing unit 1022, is additionally operable to described second distance divided by the number of all negative samples of described kind label to obtain the second parameter;

Described central processing unit 1022, is additionally operable to calculate the difference of described first parameter and described second parameter and the predictive value of the j-th training sample that the difference of described first parameter and described second parameter is described incipient nucleus matrix.

Optionally, described central processing unit 1022, the predictive value being additionally operable to the j-th training sample by described incipient nucleus matrix brings described sign function into export result value;

Described central processing unit 1022, is additionally operable to synthesize the output result value of N number of training sample of incipient nucleus matrix the vector of N dimension;

Described central processing unit 1022, is additionally operable to the N of the described incipient nucleus matrix vector transposition tieed up to obtain a described N dimensional vector.

Optionally, described central processing unit 1022, if being additionally operable to the predictive value kind label j-th label equal to described training sample of the j-th training sample of described incipient nucleus matrix, the j-th element then determining described nuclear matrix is 1, wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample;

Or, described central processing unit 1022, if the predictive value being additionally operable to the j-th training sample of described incipient nucleus matrix is not equal to the kind label j-th label of described training sample, it is determined that the j-th element of described nuclear matrix is-1.

Wherein, the described central processing unit 1022 shown in the present embodiment specifically performs the detailed process of nuclear matrix appraisal procedure please refer to embodiment illustrated in fig. 3, does not specifically repeat in the present embodiment.

The computing equipment that the present embodiment provides can also perform Multiple Kernel Learning method, in the process performing Multiple Kernel Learning method,

Described central processing unit 1022, M nuclear matrix of each latent structure to described training sample respectively for the kernel function different according to M, wherein, M is natural number;

Described central processing unit 1022, what be additionally operable to the M corresponding to each feature of described training sample described nuclear matrix distinguishes force value, at least one target nuclear matrix selected in described M nuclear matrix;

Described central processing unit 1022, is additionally operable to be in mean vector g_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set, wherein, described mean vector g_ave=(1/N) ��_pg_p, g_pFor selected arbitrary described target nuclear matrix, described reference vector g_ref=o+ �� g_ave, o represents the diagonal vector of first quartile, and �� is a constant, and described reference vector g_refWith described mean vector g_avePerpendicular;

Described central processing unit 1022, is additionally operable to bring in support vector machines grader to train disaggregated model described target nuclear matrix set into.

Optionally, described central processing unit 1022, what be additionally operable to from the described nuclear matrix of the M corresponding to each feature of described training sample selected described nuclear matrix distinguishes that the maximum nuclear matrix of force value is described target nuclear matrix.

Optionally, described central processing unit 1022, the first object nuclear matrix being additionally operable to will be located in described target nuclear matrix set is multiplied by mutually with the second target nuclear matrix and obtains the 2nd N dimensional vector, and described first object nuclear matrix and described second target nuclear matrix are target nuclear matrix described in any two differed in described target nuclear matrix set;

Described central processing unit 1022, the N number of element being additionally operable to comprise described 2nd N dimensional vector is sued for peace to obtain redundant prediction value;

Described central processing unit 1022, if being additionally operable to described redundant prediction value more than or equal to predetermined threshold value, then deletes described first object nuclear matrix from described target nuclear matrix set.

Wherein, the described central processing unit 1022 shown in the present embodiment specifically performs the detailed process of Multiple Kernel Learning method please refer to embodiment illustrated in fig. 7, does not specifically repeat in the present embodiment.

Those skilled in the art is it can be understood that arrive, for convenience and simplicity of description, the system of foregoing description, the specific works process of device and unit, it is possible to reference to the corresponding process in preceding method embodiment, do not repeat them here.

In several embodiments provided herein, it should be understood that disclosed system, apparatus and method, it is possible to realize by another way. Such as, device embodiment described above is merely schematic, such as, the division of described unit, being only a kind of logic function to divide, actual can have other dividing mode when realizing, for instance multiple unit or assembly can in conjunction with or be desirably integrated into another system, or some features can ignore, or do not perform. Another point, shown or discussed coupling each other or direct-coupling or communication connection can be through INDIRECT COUPLING or the communication connection of some interfaces, device or unit, it is possible to be electrical, machinery or other form.

The described unit illustrated as separating component can be or may not be physically separate, and the parts shown as unit can be or may not be physical location, namely may be located at a place, or can also be distributed on multiple NE. Some or all of unit therein can be selected according to the actual needs to realize the purpose of the present embodiment scheme.

It addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, it is also possible to be that unit is individually physically present, it is also possible to two or more unit are integrated in a unit. Above-mentioned integrated unit both can adopt the form of hardware to realize, it would however also be possible to employ the form of SFU software functional unit realizes.

If described integrated unit is using the form realization of SFU software functional unit and as independent production marketing or use, it is possible to be stored in a computer read/write memory medium. Based on such understanding, part or all or part of of this technical scheme that prior art is contributed by technical scheme substantially in other words can embody with the form of software product, this computer software product is stored in a storage medium, including some instructions with so that a computer equipment (can be personal computer, server, or the network equipment etc.) perform all or part of step of method described in each embodiment of the present invention.And aforesaid storage medium includes: USB flash disk, portable hard drive, read only memory (ROM, Read-OnlyMemory), the various media that can store program code such as random access memory (RAM, RandomAccessMemory), magnetic disc or CD.

The above, above example only in order to technical scheme to be described, is not intended to limit; Although the present invention being described in detail with reference to previous embodiment, it will be understood by those within the art that: the technical scheme described in foregoing embodiments still can be modified by it, or wherein portion of techniques feature is carried out equivalent replacement; And these amendments or replacement, do not make the essence of appropriate technical solution depart from the spirit and scope of various embodiments of the present invention technical scheme.

Claims

1. a computing equipment, it is characterised in that including: central processing unit, memorizer, for storing application program and/or the storage storage medium of data and power supply;

2. computing equipment according to claim 1, it is characterised in that described central processing unit, is additionally operable to that described incipient nucleus matrix carries out leading diagonal and processes.

3. computing equipment according to claim 1 and 2, it is characterised in that described central processing unit, is additionally operable to described incipient nucleus Matrix Multiplication with the kind label of training sample to obtain the incipient nucleus matrix predictive value to N number of training sample, particularly as follows:

4. computing equipment according to claim 3, it is characterised in that described central processing unit, is additionally operable to described first distance divided by the number of all positive sample of described kind label to obtain the first parameter;

5. the computing equipment according to claim 3 or 4, it is characterised in that described central processing unit, is additionally operable to bring the predictive value of described incipient nucleus matrix N training sample into sign function transposition to obtain a N dimensional vector, specifically includes:

6. the computing equipment according to any one of claim 1 to 5, it is characterised in that described central processing unit, is additionally operable to a described N dimensional vector is multiplied by the kind label of described training sample to obtain nuclear matrix, specifically includes:

Described central processing unit, if being additionally operable to the predictive value kind label j-th label equal to described training sample of the j-th training sample of described incipient nucleus matrix, the j-th element then determining described nuclear matrix is 1, wherein, the j-th training sample of described incipient nucleus matrix is the arbitrary training sample in described incipient nucleus matrix N training sample;

Or,

Described central processing unit, if the predictive value being additionally operable to the j-th training sample of described incipient nucleus matrix is not equal to the kind label j-th label of described training sample, it is determined that the j-th element of described nuclear matrix is-1.

7. a nuclear matrix appraisal procedure, described nuclear matrix appraisal procedure is applied to computing equipment, it is characterised in that including:

8. method according to claim 7, it is characterised in that described by kernel function according to after the latent structure incipient nucleus matrix of training sample, described method also includes:

Described incipient nucleus matrix carries out leading diagonal process.

9. the method according to claim 7 or 8, it is characterised in that described described incipient nucleus Matrix Multiplication is included to obtain the incipient nucleus matrix predictive value to N number of training sample with the kind label of training sample:

10. method according to claim 9, it is characterised in that the described J by described incipient nucleus matrix arranges the kind label being multiplied by described training sample with after obtaining the first distance and second distance, and described method also includes:

11. the method according to claim 9 or 10, it is characterised in that the described predictive value by described incipient nucleus matrix N training sample is brought sign function transposition into and included to obtain a N dimensional vector:

12. according to the method described in any one of claim 7 to 11, it is characterised in that described the kind label that a described N dimensional vector is multiplied by described training sample is included to obtain nuclear matrix:

Or,

13. based on a Multiple Kernel Learning method for nuclear matrix appraisal procedure, described Multiple Kernel Learning method is applied to computing equipment, it is characterised in that described nuclear matrix appraisal procedure is such as shown in any one of claim 7 to 12, and described Multiple Kernel Learning method includes:

14. method according to claim 13, it is characterised in that described M corresponding to each feature of described training sample described nuclear matrix distinguish force value, described M nuclear matrix is selected at least one target nuclear matrix and includes:

15. the method according to claim 13 or 14, it is characterised in that described will be in mean vector g_aveWith reference vector g_refAngle in the middle of target nuclear matrix be incorporated into target nuclear matrix set after, described method also includes: