CN101782976A - Automatic selection method for machine learning in cloud computing environment - Google Patents

Abstract

The invention relates to an automatic selection method for machine learning in cloud computer environment. By using a cloud computing platform, a user can automatically and intelligently build a machine learning mathematic model which meets actual problems without building the operation environment of machine learning, selecting a machine to learn algorithm and even adjusting complicated machine learning functions and accompanying parameters but only using a Web method to upload sample data. Through the method, the use of machine learning is free from the environmental constraints and displays the advantages of the cloud computing platform, so that the machine learning model building is transparent to the user, so as to best reduce the use threshold of machine learning. The automatic selection method for machine learning in cloud computing environment solves the disadvantages of the unpredictability of model building selection, the manual experience of parameter adjustment, the difficulties of common users and the like when machine learning is applied in actual life.

Description

A kind of automatic selection method for machine learning in cloud computing environment

Technical field

The present invention is the autonomous system of selection of a kind of machine learning based on cloud computing environment.By using cloud computing platform, make the user need not to build the running environment of machine learning, also need not to learn machine learning algorithm, more need not adjust the machine learning function and the parameter thereof of numerous and complicated, only need under cloud computing platform, use the Web mode to upload training data and prediction test data, and determine to comprise information seldom such as usable range, expectation territory, just can obtain needed multiple machine learning model and specific descriptions, so that solve practical problems.

Background technology

Machine learning is the another important application after expert system application, artificial intelligence application, also is a kind of core research topic of artificial intelligence simultaneously.Its objective is to make computing machine can simulate or test human learning behavior, thereby acquire knowledge or technical ability can constantly be improved performance according to new information simultaneously.The ability of machine learning is very important feature, and H.A.Simon thinks that study is the adaptations that system does, and the system that makes finishes same next time or obtaining better to finish effect during similar task.R.s.Michalski thinks that study is the expression of constructing or revising for the experience things.The people that are engaged in DEVELOPMENT OF EXPERT SYSTEM think that then study is obtaining of knowledge.These viewpoints emphasize particularly on different fields, and first kind of viewpoint emphasized the external behavior effect learnt, then emphasize the internal procedure learnt for second kind, and the third mainly is from the point of view of practicability of knowledge engineering.

The Research of Machine Learning method is to use for reference the understanding to the self-teaching mechanism of the mankind own such as physiology, psychology, cognitive science, foundation is to the computation model or the cognitive model of human learning process, thereby form the various theories of learning and learning method, set up the learning system with application-specific of oriented mission.These goals in research influence each other and mutually promote.Since 1980 since Ka Neiji-Mei Long university holds first machine scientific seminar, machine learning development is very fast, has become one of central topic.

And the history development procedure of machine learning is divided into four-stage: (1) the mid-50 is to the ardent period of the mid-1960s; (2) the mid-1960s is to the calm period of the mid-1970s; (3) the mid-1970s is to the recovery period of the mid-80; Beginning in (4) 1986 years then is the latest stage of machine learning.And this period at present, outstanding feature is that machine learning has developed into an emerging frontier branch of science, has merged various learning method, and range of application is also increasing, and relevant academic activities are very active.

Machine learning develops into present stage, uses very extensively, and a lot of outstanding algorithms that have been born may be summarized to be basically based on the study of symbol with based on is-not symbol study, just connectionist learning.And the former is based on the study of symbol, generally comprise rote learning, instruct formula study, the study of example formula, matching test study, based on study of explaining or the like.

Wherein comparatively common algorithm has: decision Tree algorithms, genetic algorithm, Bayesian statistics algorithm, artificial neural network algorithm, algorithm of support vector machine, association rule algorithm or the like.Carry the MBM of these common algorithms in the method for this paper design, and used the EM algorithm to come parameter is carried out maximal possibility estimation.

But be to use machine learning techniques to handle specific tasks, mainly face three problems: (1) is at a certain specific tasks the time, setting up machine learning model wastes time and energy, because the distinctiveness of specific tasks details, be difficult to directly use for reference the system model that other have built, need select according to the personal experience.(2) even certain subtask, correctly selected relatively to meet the machine learning algorithm of objective fact essence, how its complicated parameter is provided with also is a problem that must solve, need rule of thumb or the long computing of subscriber computer obtains, single user's computing power (3) user that is difficult to deal with problems fast need learn and use concrete machine learning software, the machine learning algorithm numerous and complicated, autonomous learning need spend the plenty of time, simultaneously certain some algorithm of user's autonomous learning differ also that suitable user surely need solve each run into task.

And the cloud computing technology of emerging appearance but can solve above problem, makes convenient being applied in the reality of machine learning, the faster and better creation of value.

Cloud computing is a kind of novel computation model that proposes on the basis of development such as distributed system, grid computing, is a kind of method of emerging shared architecture, and what it was faced is ultra-large distributed environment, and core provides data storage and network service.This is a kind of payment and use pattern of serving of referring to, refer to by network with as required, the mode of easily expansion obtains required service.It is relevant with software, internet that this service can be IT, also can be other service arbitrarily.Cloud computing provides the most reliable, safest data storage center, the user again concern of data lose, trouble such as poisoning intrusion, cloud computing has simultaneously reached minimum to the equipment requirements of user side." cloud " mentioned in the cloud computing is that some can self and the virtual computational resource of management, is generally some large server clusters, comprises calculation server, storage server, broadband resource or the like.Cloud computing puts together all computational resources by various clouds are provided, and realizes management automatically by software, need not artificial participation.It is loaded down with trivial details details worry that this feasible supplier of application need not, and can be absorbed in the business of oneself more, helps innovating and reducing cost.Be applied in the machine learning, cloud computing can set up machine learning model and correlation module is selected for the user, and the achievement that makes the user enjoy machine learning techniques fast solves problem.

Existing cloud computing platform is based on Theoretical Calculation and stores service substantially, finds no the special cloud computing method of setting up for machine learning, and the present invention provides a kind of feasible implementation method in conjunction with the advantage and the characteristic of machine learning techniques and cloud computing technology.

Summary of the invention

Technical matters: the purpose of this invention is to provide is a kind of automatic selection method for machine learning in cloud computing environment.By using cloud computing platform, solved the problem of machine learning modeling inconvenience, provide a kind of and provide conveniently method in conjunction with cloud computing skill and machine learning techniques processing realistic problem.Thereby it is worried that the user be need not to loaded down with trivial details details, can be absorbed in the business of oneself more, helps innovating and reducing cost.

Technical scheme: the present invention makes the use of machine learning break away from the constraint of environment, the cloud computing platform computing power and the transparency have efficiently been given full play to, farthest reduced the use threshold of machine learning, make the user need not to comform in the multimachine device learning method and seek the suitable machine learning method by experiment repeatedly, shortcomings such as the artificial experience that is difficult to predictability, parameter adjustment that modeling is selected, domestic consumer's difficulty of learning have been solved when the practical application machine learning.

The present invention seeks to set up the method for the cloud computing platform that the machine learning service is provided.Under cloud computing platform, carry out system constructing by following three aspects: the various machine learning clouds of setting up a large amount of computing machines compositions that exist with the cloud form on the one hand, comprise decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud, make the cloud computing platform acquiescence carry common machine learning algorithm; Generally estimate cloud, method discovery cloud, EM algorithm support cloud, valuation functions cloud, calculate cloud, machine learning algorithm expansion cloud by initial modeling cloud, the search volume formed by computer cluster equally on the other hand, thereby embody the advantage of cloud, calculate the suitable parameters of the machine learning use that domestic consumer is difficult to or need calculate for a long time by a large amount of computational resources; Last aspect is that cloud computing platform and user carry out mutual necessary module, comprises the Web interactive interface, and machine learning input/output module and cloud administration module are in order to support the operation of cloud computing platform.

Step 1) is under the uniform dispatching of cloud administration module, at first by the Web interactive interface, obtain the required rough description of dealing with problems of user, comprise the problem kind, big class under promptly selecting, from expert system, cognitive simulation, planning and problem solving, data mining, the network information service, pattern recognition, fault diagnosis, natural language understanding, robot and game, other classification, select

Step 2) enable initial modeling cloud,, enter different subclass interfaces, fill in corresponding more detailed information by the big class that the user in the step 1) provides, comprise carry out that sample is uploaded, selected method for expressing, determines the interpretation of result method, usable range, expectation territory,

Step 3) startup method is found cloud, compares according to the historical typical example of information of same that the user provides, and determines because of taking the machine learning algorithm of which kind of or which kind; This cloud module is accompanied by the subsequent step operation, thereby adjusts constantly according to each stage result of calculation,

The information that step 4) is imported step 2 kind of user then, input machine learning input/output module, must unitize, after the datumization, carry out operations such as missing values processing, noise data processing, data scrubbing, data integration, data conversion, data reduction successively, so that obtain the intermediate result that general algorithm can use

Step 5) starts the valuation functions cloud, according to the user in step 2) information of input sets up valuation functions, the quality that machine learning is separated judges and prepares, thereby the specific algorithm performance is predicted,

Step 6) is called the EM algorithm simultaneously and is supported cloud, and solution space is carried out maximal possibility estimation, calculates optimum solution or more excellent approximate location in solution space of separating, and increases search efficiency,

After step 7) arrives this step, illustrate that preliminary work finishes, be about to carry out the training process of machine learning, automatic judgement by above step, call one or several concrete machine learning cloud modules respectively and learn, comprise decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud; As User Defined machine learning algorithm expand cloud, then preferentially call machine learning algorithm and expand cloud,

Step 8) is calculated through above step, selects one or several algorithm clouds, with its startup, passes through the Web interactive interface simultaneously to field feedback, comprises the step of calculating operation, the intermediate result that obtains, and current optimum solution changes,

Step 9) supports in the process that iterates of cloud at the EM algorithm, constantly turn back to step 6, step 7 is calculated, judge whether to reach end condition simultaneously, if reach end condition then jump procedure 10, otherwise the performance prediction algorithm that uses step 5 to formulate is judged the outstanding degree of separating, this step needs a large amount of computational resources, thereby need utilize the calculating advantage of cloud computing, must calculate outstanding separating as far as possible

When step 10) satisfies at end condition, as arriving computing time, there be not more excellent separating or the iteration end of algorithm own, result of calculation is converted to information by the machine learning input/output module with readability, return the client by the Web interactive interface again, and provide detailed data to download, preserve the machine learning result simultaneously, so that reuse, avoid double counting.

One, architecture

Whole proposal has comprised that cloud is generally estimated in decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud, initial modeling cloud, search volume, method finds that cloud, EM algorithm support cloud, valuation functions cloud, calculating cloud, machine learning algorithm expansion cloud and Web interactive interface module, machine learning input/output module, cloud administration module.Mutual relationship as shown in Figure 2.

Provide the explanation of concrete module below:

Decision Tree algorithms cloud: decision Tree algorithms modeling service that the major function of this cloud provides and prediction service.Decision tree be a kind ofly be used to classify, the method for the forecasting type modeling of cluster, prediction, the thought of employing is " dividing and rule ", thus it is divided into some subclass with the search volume and sets up decision tree.This algorithm is one of most widely used induction algorithm at present, is a kind of algorithm that approaches discrete function, based on example, is commonly used to as sorter.And basic decision Tree algorithms can be described as a kind of greedy algorithm, and the algorithm of existing large-scale application all is that the improvement and the function of basic decision tree algorithm strengthened.This algorithm cloud comprises following common improved decision Tree algorithms equally: C4.5 method, CART method, SLIQ method, SPRINT method etc.

The C4.5 algorithm has increased the processing to continuous type attribute, property value vacancy, has improved the ability of cutting out of decision tree equally.The technology that the CART algorithm adopts a kind of two minutes recurrence to cut apart is divided into two subclass with sample, makes each non-leaf node of decision tree that two branches all be arranged, thereby obtains binary decision tree simple for structure.The SLIQ algorithm has mainly adopted the breadth First algorithm to claim decision tree; The SPRINT algorithm has well solved the restriction of memory size, has handled the ultra-large training set that other algorithms can not be suitable for, and effectively deep layer decision tree.

The function of decision Tree algorithms cloud provides basic decision Tree algorithms modeling training service and prediction service on the one hand, and the characteristic according to the input data of intelligence selects concrete improved decision Tree algorithms to carry out calculation process on the other hand.

Genetic algorithm cloud: genetic algorithm modeling service that the major function of this cloud provides and prediction service.Genetic algorithm is to carry out evolutionary process by " survival of the fittest " rule in simulating nature circle and the algorithm that designs.Bagley and Rosengerg have at first proposed the notion of genetic algorithm in the PhD dissertation at them in 1967.The monograph of Holland publication in 1975 has been established the theoretical foundation of genetic algorithm.Nowadays genetic algorithm has not only provided arthmetic statement clearly, and Set upOne The result of a little quantitative test has obtained widely should in various fieldsWhen being similar to production task planning, exploited in communication, TSP problem, knapsack problem and Flame Image Process and signal Processing etc., can adopt this algorithm cloud with, the problem description that proposes as the user.

The fundamental purpose of genetic algorithm cloud is to use genetic algorithm thought, and the simulation of evolving is gradually tried to achieve finally and separated to user's target.Carry out iteration by initial solution, constantly from old separating, produce new explanation according to certain rule, and expect new separate than old separate outstanding.If new explanation is high more by the value that valuation functions calculates, the chance that it keeps is also just big more.Genetic algorithm is only used the coded representation problem, and the value that obtains with valuation functions is foundation, does not require clear and definite analytical expression, therefore can solve the non-linear optimizing problem of arbitrary height.And easily combine, get its strong point, obtain more excellent effect with other algorithms.

Bayesian statistics algorithm cloud: Bayes Modeling service that the major function of this cloud provides and prediction service.Bayes is that statistic algorithm can be predicted may concern between the class members, belongs to the probable value of certain class as given certain sample number.Bayesian algorithm belongs to the certain kinds probability by the calculating sample and classifies mainly based on Bayes law.Compare with other method, Bayes can be specially adapted to the situation that sample is difficult to obtain in conjunction with sample information and prior probability.Owing to need to calculate prior probability,, be fit to small scale machine study more simultaneously along with increasing of sample makes rise appreciably computing time.When being similar to aspect such as information recovery and diagnosis, economic field, classification automatically, production quality control when user's problem description, can adopt this algorithm cloud.

The fundamental purpose of Bayesian statistics algorithm cloud is to use Bayes statistical method, calculates the probability of the object that belongs to a certain class, has the class of maximum probability and is class under this object.The object of its processing can be that disperse, continuous, also can be mixed type.Based on bayes method, common have naive Bayesian method and a Bayesian network method.This algorithm cloud comprises common bayes method method for establishing model and the expansion of constantly being upgraded.

Artificial neural network algorithm cloud: artificial neural network modeling service that the major function of this cloud provides and prediction service.Artificial neural network utilizes the intelligency activity of computer technology simulation human brain, simulates the structure and the information conduction pattern of biological neural network, and expresses with mathematical form.Artificial neural network is a basic technology in the current intelligent science and technology, and the symbolic reasoning mechanism of the connection mechanism of employing and artificial intelligence becomes two big camps of intelligence science and technology side by side.The anatomical physiology feature of artificial Neural Network Simulation human brain is with many parallel simple neurons, under certain topological structure connects, accept external information, stimulate mutually simultaneously, thereby reach distributed store, associative memory, the feedback refinement, the black box mapping, the weights balance is dynamically approached, record is deposited in holography, the effect of fault-tolerant anti-mistake.Simultaneously because the imictron interconnection, when quantity reaches certain rank, can form powerful self study, self-adaptation, self-organization, self diagnosis, self-reparing capability, by constantly feeding back between node, can simulate the reasoning from logic of human brain to a certain extent, scope therefore has a wide range of applications.Particularly, when approximation of function and credit risk assessment, can preferentially adopt this algorithm cloud in pattern-recognition.

The fundamental purpose of artificial neural network algorithm cloud is used artificial neural network technology, adopts the method for simulation cerebral nerve network, manual construction a kind of neural network that can realize certain function.This algorithm cloud can produce human brain neural network's mathematical model, forms a kind of model of setting up based on imitating cerebral nerve network structure and function.This model is interconnected by a large amount of simple components and neuron, and a kind of complex network of formation has the non-linear of height, can carry out complicated logical operation and nonlinear relationship and realize.Usable range is very extensive.

Algorithm of support vector machine cloud: support vector machine modeling service that the major function of this cloud provides and prediction service.Support vector machine (Support Vector Machine) is a kind of new mode identification method that grows up on the basis of Statistical Learning Theory in recent years, shows many distinctive advantages in solving small sample, non-linear and higher-dimension pattern recognition problem.The main application fields of support vector machine has pattern-recognition, approximation of function and probability density to estimate or the like that this algorithm cloud can be preferentially adopted in these fields.

The fundamental purpose of algorithm of support vector machine cloud is, sets up a kind of model, and sample vector is mapped to high latitude space, and structure optimal classification face obtains linear optimal decision function in higher dimensional space.The over-fitting of inhibition function is measured in interval by the control lineoid, simultaneously by having used kernel function to solve problem of dimension cleverly, and the directly related sample dimension of complexity of having avoided learning method to calculate.Optimal classification lineoid wherein makes classifying distance big as far as possible guaranteeing that sample does not have under the misclassification situation, and guarantees in the empiric risk minimum, makes fiducial range minimum in the boundary of generalization, thereby guarantees the real risk minimum.

Association rule algorithm cloud: correlation rule modeling service that the major function of this cloud provides and prediction service.The association rule algorithm purpose is to solve a class problem of excavating the relevance between the collection on the big transaction data set (TDS).The correlation rule analysis is a big generic task in the machine learning.It originates from the analysis to dichotomic variable, expresses two relations between the dichotomic variable with the mode of rule, and the relation between a plurality of dichotomic variable.Certainly, later development also makes correlation rule not only be confined to dichotomic variable, also can analyze many classified variables and continuous variable.So correlation rule can be regarded as between the situational variables and concerns, and this relation table is reached the method for the rule of be very easy to explaining.The correlation rule analytical approach distributes to data and does not do any requirement, and the result of gained is complete in data, without any the subjectivity supposition, has reflected the essence of data objectively, and very strong cogency is arranged.The result that correlation rule obtains data analysis can be regarded as in the data between variable regular summary.Therefore correlation rule is after proposing, obtained a large amount of application in all trades and professions, particularly in the machine learning modeling in the extremely huge fields such as astronomical meteorological biology of market analysis, prestige assessment, commodity price analysis, intrusion detection and quantity of information, can preferentially adopt this algorithm cloud.

The fundamental purpose of association rule algorithm cloud is that the model of setting up model can solve following problem: the pattern of the incidence relation between different objects or the description of form; Improve related rate calculated and reduce storage space; Association analysis in mass data etc.

Initial modeling cloud: this cloud is uploaded sample, basic representation method, is determined that interpretation of result method, usable range, expectation territory etc. carry out initialization what the user provided, obtains initial model.

Cloud is generally estimated in the search volume: the major function of this cloud provides feasible solution and outstanding estimation position of separating, both obtained the hunting zone that is complementary with problem description in the possible space, get rid of the outstanding space of separating of to be born as much as possible, thereby improve search efficiency, reduce calculated amount.

Method is found cloud: the major function of this cloud is to select the suitable machine learning algorithm to set up model.By generally estimating and initial calculation, predictability is selected certain or certain several machine algorithm cloud.

The EM algorithm supports cloud: support vector machine modeling service that the major function of this cloud provides and prediction service.A lot of algorithms all will carry out the parameter estimation of model in the machine learning, just will carry out maximum likelihood estimation or maximum posteriori likelihood and estimate.When but the variable in the model was the Direct observation variable, maximum likelihood or maximum posteriori likelihood were obvious.But when some variable is hidden, carry out maximum likelihood and estimate that just very complicated being difficult to directly obtains.Exist under the situation of latent variable, model parameter is carried out estimation approach to be had a variety ofly, and a kind of popular maximum likelihood method of estimation is the Expectation-Maxi2mization algorithm, abbreviates the EM algorithm usually as.It is not directly the posteriority of complexity to be distributed to maximize or carry out analog computation, but adds some potential data on the basis of observed data, calculates and finish a series of simple maximizations or simulation thereby simplify.The EM algorithm be a kind of from non-complete data the maximum likelihood method of estimation of solving model parameter.Non-complete data generally is divided into two kinds of situations: a kind of is because the restriction or the mistake of observation process itself, as human error, the fragmentary data that obtains such as be difficult to measure; A kind of to be that the likelihood function of parameter is directly optimized very difficult, and introduce extra parameter, as parameter implicit or that lose.So to its optimization method is that definition raw observation data add excessive data composition " complete data ", the raw observation data just become " fragmentary data " naturally.

The valuation functions cloud: the major function of this cloud be reflection when setting up machine learning model with the degree that conforms to of target, and to the assessment of established model.According to the characteristic and the historical experience regulation valuation functions of each algorithm, whether good by inspection performance on training data on the other hand on the one hand, independently testing on the test data again.Test data wherein is to break away from the modelling algorithm, only participates in prediction and judges.

Calculate cloud: the major function of this cloud is to give full play to the advantage of cloud computing, under ultra-large distributed environment, utilizes the calculated performance, data storage and the network service that provide to come the magnanimity computing that machine learning needs is calculated.The calculating advantage of parallel computation, Distributed Calculation and grid computing has been given full play in cloud computing, and calculation services can well be provided.

Machine learning algorithm expands cloud: the major function of this cloud provide machine learning algorithm can't satisfy user's needs the time, be User Defined or platform itself interface from the upgrading reservation.Machine learning algorithm expands cloud according to the new learning algorithm of certain rule structure on the one hand, and this cloud is responsible for getting in touch other cloud and module on the other hand, thus the use that the new learning algorithm of feasible structure can be complete.

Web interactive interface module: the major function of this cloud provides interactive interface.Cloud computing support the user at an arbitrary position, use various terminals to obtain application service.Institute's requested resource is from " cloud ", rather than fixing tangible entity.Be applied in somewhere operation in " cloud ", but in fact the user need not to understand, also do not worry using the particular location of operation, that is to say at the user it is transparent.Only need a notebook or a mobile phone, just can realize all that we need, even comprise the task that supercomputing is such by the network service.Therefore, be best mode alternately by the Web interface, the user needn't be concerned about operation and the computing that carry out on the backstage, only needs to be concerned about the information of input and the result of output.This module and processing and user's interaction problems.

Machine learning input/output module: support vector machine modeling service that the major function of this cloud provides and prediction service.For the mathematical modeling of machine learning provides feasible input sample and parametric description, in fact also comprised the pre-service work of numerous and complicated.As much as possible that different recording form, different custom, different time weak point, diverse location, different data set is right, the data centralization of different ill-formalnesses, integration, cleaning.Usually to unitize, datumization, format conversion is carried out operations such as missing values processing, noise data processing, data scrubbing, data integration, data conversion, data reduction

The cloud administration module: the major function of this cloud is startup, execution and the monitor state of each module of management.Cloud computing is because its ultra-large property, generally have hundreds of thousands of station servers, large enterprise even have the hundreds of thousands station server, and at user transparent, this all needs a large amount of bookkeepings, control the ruly operation of each module, scheduling and allocating task are rationally utilized storage, calculating, bandwidth resources.

Two, method flow

1, builds and operational scheme

1. the user at first installs and starts the cloud administration module, increases successively by administration module that decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud, initial modeling cloud, search volume estimate generally that cloud, method find that cloud, EM algorithm support cloud, valuation functions cloud, calculate cloud, machine learning algorithm expands cloud then and Web interactive interface module, machine learning input/output module.

2. start Web interactive interface module, wait for user's use.When the user set up the machine learning model request by the submission of Web interactive interface, the cloud administration module will start and call other cloud module, the mathematical modeling that carries out machine learning.

3. call initial modeling cloud, upload sample, basic representation method, determine that interpretation of result method, usable range, expectation territory etc. carry out initialization, obtain initial model what the user provided.

4. operation method is found cloud, compares according to the historical typical example of information of same that the user provides, and determines because of taking the machine learning algorithm of which kind of or which kind.This cloud module is accompanied by the subsequent step operation, thereby adjusts constantly according to each stage result of calculation.

5. the cloud administration module is by input machine learning input/output module, with the data of user by Web interactive interface input unitize, after the datumization, carry out operations such as missing values processing, noise data processing, data scrubbing, data integration, data conversion, data reduction successively, so that obtain the intermediate result that general algorithm can use.

6. starting the valuation functions cloud prepares to judging the quality that machine learning is separated.This step mainly is to formulate the specific algorithm of estimated performance, whether uses verification methods such as cross validation and leaving-one method, bootstrap method.

7. call the EM algorithm and support cloud, solution space is carried out maximal possibility estimation, calculate optimum solution or more excellent approximate location in solution space of separating, increase search efficiency.

The EM algorithm basic principle can be expressed as follows: can observed data be y, complete data x=(y, z), z is a hidden variable, the expression missing data, θ is a model parameter.θ about the posteriority distribution p (θ | y) very complicated, be difficult to carry out various different statistical computations.If z is known, then may obtain a simple interpolation posteriority distribution p about θ (θ | y, z), utilize p (θ | y, simplicity z) can be carried out various statistical computations.Then, can the supposition of z be conducted a survey and improve again, thereby with the maximization of a complexity or the problem reduction of sampling.

As can be seen, the EM algorithm is a kind of alternative manner, is mainly used in to ask the posteriority mode of distribution.

The specific implementation step is as follows: suppose that y is the non-complete observation data collection of obeying a certain distribution, and there is a complete data collection x=(y, z), then the density function of x is: p (x| θ)=p (y, z| θ)=p (z|y, θ) p (y| θ) therefrom as can be seen, density function p (x| θ) be by marginal density function p (θ | y), hypothesis, parameter θ initial estimate and the hidden variable z of hidden variable z and the decision of the relation between the observational variable y.

8. after preliminary work is finished, carry out the modeling process of machine learning, automatic judgement by above step, call one or several concrete machine learning cloud modules respectively and learn, comprise decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud.As User Defined machine learning algorithm expand cloud, then preferentially call machine learning algorithm and expand cloud.

2, machine learning modeling flow process

1. decision Tree algorithms modeling

Decision tree can be regarded a tree-shaped forecast model as, and it comes classified instance by example is aligned to certain leaf node from root node, and leaf node is the classification under the example, shown in Fig. 3 decision tree basic configuration figure.The key problem of decision tree is to select the beta pruning of Split Attribute and decision tree.The algorithm of decision tree has a lot, and ID3, C4.5, CART or the like are arranged.These algorithms all adopt top-down greedy algorithm, and the best attribute of each node selection sort effect is 2 or a plurality of child node with node splitting, continue this process up to this tree classification based training collection exactly, or all properties are used all.As wherein classification regression tree (CART) is a kind of classification and regression algorithm in the machine learning.If training sample set L={X ₁, X ₂, X ₃... X _n, Y}, wherein, X _i(i=1,2,3 ..., n) be called attribute vector; Y is called label vector or categorization vector.When Y is orderly quantitative value, be called regression tree; When Y is discrete value, be called classification tree.At the root node place of tree, search problem collection (data acquisition space) finds the optimum division variable and the corresponding division threshold value that make that the non-purity of data set descends maximum in the child node of future generation.

Non-here purity index is weighed with the Gini index, and it is defined as:

Wherein, i (t) is the Gini index of node t, and p (i/t) is illustrated in the shared ratio of sample that belongs to the i class among the node t, and p (j/t) is the shared ratio of sample that belongs to the j class among the node t.Divide variable and divide threshold value with this root node t ₁Split into t ₂And t ₃If, at certain node ti place, the remarkable reduction of further non-purity can not be arranged again, then this node ti becomes leaf node, otherwise continues to seek its optimum division variable and divide threshold value and divide.For classification problem, in leaf node, have only a class, this class is just as the class under the leaf node so, if there is the sample in a plurality of classes to exist in the node, determines classification under the node according to that maximum class of sample in the leaf node; For regression problem, then get the mean value of its quantitative value.Clearly, the very big undue fitting data of tree possibility, but less tree possibly can't catch important structure again.The best size of tree is the adjustment parameter of controlling models complicacy, and it should be by the selection of data adaptive.A kind of desirable strategy is to increase a bigger tree t ₀, only when reaching minimum node size (such as 3), just stop fission process.The method of utilizing beta pruning strategy and 5 foldings or 10 folding cross validations to combine is then pruned this tree, thereby some noises and interfering data are got rid of, and obtains optimal tree.Thereby set up the mathematical model of decision tree.

2. genetic algorithm modeling

For little space, the classical method of exhaustion is just enough; And, then need to use special artificial intelligence technology to large space.Genetic algorithm (Genetic Algorithm) is a kind of in these technology, it be an analoglike biological evolution process and produce by the global optimizing algorithm of selecting operator, hybridization operator and three basic operators of mutation operator to form.It by selecting operator to select the good male parent of proterties, hybridizes computing by the hybridization operator from an initial family, and mutation operator carries out a little variation, the random search model space under certain rule of probability control.Evolution generation upon generation of reaches the requirement of setting up to the error functional value of finally separating correspondence.

The t time iteration, genetic algorithm is kept a potential colony that separates

Each separates x ₁ ^tThe evaluation function evaluation evaluation of using the valuation functions cloud to obtain.Then by selecting more suitable individuality (t+1 iteration) to form a new colony.The member of new colony carries out conversion by hybridization and variation, forms new separating.Cross combination the feature of two parent chromosomes (promptly waiting to ask the binary coding string of parameter), formed two similar offsprings by exchange parent corresponding segment.For example parent chromosome is (a ₁, b ₁, c ₁, d ₁, e ₁) and (a ₂, b ₂, c ₂, d ₂, e ₂), behind second gene, to hybridize, the offspring of generation is (a ₁, b ₁, c ₂, d ₂, e ₂) and (a ₂, b ₂, c ₁, d ₁, e ₁).The purpose of hybridization operator is to carry out message exchange between separating in that difference is potential.Variation is to change one or more genes (bit in the chromosome) on the selected chromosome randomly by the probability that equals aberration rate with.Some extra variabilities are introduced in being intended that to colony of mutation operator.Modeling process is shown in Fig. 4 genetic algorithm basic process figure.This process has been set up the genetic algorithm mathematical model thus.

3. Bayesian statistics modeling

Bayes is that the modeling of Bayes statistical method is a kind of method for classifying modes under the situation of known prior probability and class conditional probability.The classification results for the treatment of the branch sample of its processing depends on all of sample in each class field.If

Training sample set is divided into the M class, is designated as C={c ₁, c ₂..., c _t..., c _M, the prior probability of every class is P (c _i), i=1,2 ..., M.When sample set is very big, can think P (c _i)=c _iSample number/total sample number.Treat branch sample X for one, it belongs to c _iThe class conditional probability of class is P (X/c _i), then according to the Bayes theorem,

Can obtain c _iPosterior probability P (the c of class _i/ X)=P (X/c _i) P (c _i)/P (X) P (ci/X).If P (c _i/ X)=MaxjP (c _j/ X), and i=1,2 ..., M, j=1,2 ..., M then has X ∈ c _i, maximum a posteriori probability decision rule that Here it is.The Bayes sorting technique is proved relatively fully in theory, also is very widely on using.Overall probability distribution and the probability distribution function of all kinds of samples (or density function) usually are ignorant.In order to obtain them, just require sample enough big.In addition, when being used for text classification, the Bayes method requires the descriptor of expression text separate, and such condition general being difficult in actual text satisfied, so this method often is difficult to reach theoretic maximal value on effect.The prior probability by being based upon statistical and this method of class conditional probability can be set up Bayesian statistical model.

4. artificial neural network algorithm modeling

Artificial neural network (Artificial Neural Network. is called for short ANN) is the neural network that can realize certain function of manual construction on the basis that the mankind understand its cerebral nerve network understanding just.It is the human brain neural network's that theorizes a mathematical model, is based on imitation cerebral nerve network structure and function and a kind of information handling system of setting up.It is actually by a large amount of simple components and interconnects the complex network that forms, and has the non-linear of height, can carry out the system that complicated logical operation and nonlinear relationship realize.

Artificial neural network is organized by layer, and each layer is made up of a plurality of artificial neurons, does not have connecting line to connect between them, and layer with layer between be connected by connecting line.Artificial neural network can have individual layer, also multilayer can be arranged, and at present commonly used have an individual layer, two layers and three layers.

Artificial neural network is divided into two kinds of frame modes according to artificial neuron's data flow mode: forward direction type and feedback-type.If the artificial neuron metadata less than feedback, is not referred to as the forward direction type from being input to the output uniflux,, then be referred to as feedback-type if feedback (no matter feeding back to this neuron or other neuron of same layer) is arranged.

The artificial neuron is the base unit of artificial neural network, and the artificial neuron can have multiple model, but has a kind of horizontal substantially type the most common, and it is composed as follows:

I. input: an artificial neuron can have a plurality of inputs.

II. output: an artificial neuron can only have an output.

III. inner structure: will import addition with totalizer, and add deviate then, and calculate it with activation function then, the neuronic output of result calculated conduct

A. totalizer: will import linear, additive, exactly, be that the product with input and corresponding weight value sums up.

B. deviate: the value that totalizer produced is subjected to external disturbance and influence and produces deviation through regular meeting, therefore needs a deviation to adjust, and generally is used for θ k to represent k neuronic deviate.

C. activation function: be used for limiting the scope of neuron output value, generally-1～+ 1 or 0～1.

Activation function commonly used has Logistic, Simoid etc.

Link to each other with connecting line between the artificial neuron, every connecting line all has weights, and as mentioned above, the inner totalizer of the target nerve of connecting line unit can use these weights when summing up.Use ω _IjRepresent the weights of i neuron to connecting line between j neuron.

Artificial neural network has learning functionality, and this study is trained it with real data sample exactly.A data sample has the input and output data, with of the input of input data as artificial neural network, the relatively output of artificial neural network and the output of sample then, by adjusting the parameter (being the weights and the neuronic deviation of connecting line) in the artificial neural network, make that both difference is 0 or within the acceptable scope.

Trained artificial neural network has certain judgement and inferential capability, and can carry out certain prediction and decision-making.Reflections propagate model (BP, Back Propagation) is the modal a kind of model of artificial neural network, has the application more than half of surpassing to adopt this model.It is a multilayer forward direction type structure, is made up of following three parts:

I. input layer: have only one deck, form by m neuron, receive extraneous m input xi (i=1,2 ..., m), each input links to each other with a neuron.The basic neuron of neuron right and wrong of this one deck does not have inner structure, and the value of its output is exactly the value of input.

II. hide layer: multilayer can be arranged, be made up of n neuron for every layer, these neurons are exactly the basic neuron of introducing previously.

III. output layer: having only one deck, be made up of p neuron, also is basic neuron.

It is the connection of multi-to-multi that the neuron that (comprises between a plurality of hiding layers) between above-mentioned each layer connects, and the input and output layer is man-to-man the connection with the external world, shown in Fig. 5 artificial neural network basic block diagram.

Substantially neuronic activation function adopts the Logistic function, and expression formula is:

$O_j=\frac{1}{1+e^{-I_j}}$

Algorithm divides following step:

Calculate the input value of each the neuronic j that hides layer and output layer, thus the output valve of calculating:

A) input:

I is that preceding one deck all and neuron j have the neuron that connects in the formula.

B) output: adopt Logistic function calculation output valve.

Calculate the error of output layer neuron j:

E _rrj＝O _i(1-O _j)(T _j-O _j)

T in the formula _jBe sample class label.

Calculate the error of hiding each neuron j of layer:

$E_{\mathrm{rrj}}=O_j(1-O_j)\underset{k}{\mathrm{\Σ}}{E}_{\mathrm{rrk}}{\mathrm{\ω}}_{\mathrm{jk}}$

In the formula k be the back one deck all with neuron j the neuron that is connected, E are arranged _RrkThen be these neuronic errors.

Each connecting line weights ω in the computational grid _IjModified value:

Δω _ij＝(l)E _rriO _j

(l) is the learning rate of algorithm in the formula, and this value is formulated voluntarily by the trainer.The selection of learning rate helps to seek the minimum weights of the overall situation, selects too for a short time, and learning process can be carried out very slowly, and is too big, may appear at swing between unsuitable the separating.Generally can select a constant between (0,1), empirical value commonly used is 1/t, and t is a number of iterations.

Calculate the new weights of this connecting line then, and revise it:

ω _ij＝ω _ij+Δω _ij

Hide the modified value of each neuron deviate in layer and the output layer in the computational grid:

Δθ _j＝(l)E _rrj

Calculate the new deviate of this neuron then:

θ _j＝θ _j+Δθ _j

2) check end condition, several are generally arranged, as:

A) Δ ω _iWith Δ θ _jAll enough little, less than a certain designated value;

B) iterations has reached specified quantity.

This process has been set up the mathematical model of artificial neural network, by the neural network model that trains, can calculate the input sample, thereby obtain predicted value.

5. algorithm of support vector machine modeling

The initial thought of support vector machine is how to seek the optimal classification face for the linear separability problem, for feature space neutral line separable problem, the optimal classification face is exactly the interphase of interval γ maximum, according to the analysis of above-mentioned nuclear theory as can be known, it really under guaranteeing that sample is by the prerequisite of correctly classifying, have the interphase of best generalization ability.For the inseparable problem of feature space neutral line, can take all factors into consideration at interval and the influence of relaxation factor by a penalty factor.

The lineoid of a usefulness feature space of consideration is done the problem of two-value classification to given training dataset.For given sample point: (x ₁, y ₁) ..., (x _l, y _l), x _i∈ R ⁿ, y _i{ 1 ,+1} is vector x wherein for ∈ _iMay be to extract the directly vector of structure of some feature from the object samples collection, also may be original vector be mapped to mapping vector in the nuclear space by certain kernel function.In feature space, construct segmentation plane:

(wx)+b=0 is feasible:

$\left\{\begin{array}{cc}(w\·x_i)+b\≥1& y_i=1\\ (w\·x_i)+b\≤-1& \mathrm{yi}=-1\end{array}\right.\⇔y_i[(w\·x_i)+b]\≥(i=\mathrm{1,2},...,l)$

Can calculate, the minor increment of the segmentation plane that training dataset to is given is:

$p(w,b)=\underset{\left\{x_i\right|y_i=1\}}{\min }\frac{w\·x_i+b}{\left|w\right|}-\underset{\left\{x_i\right|y_i=-1\}}{\max }\frac{w\·x_i+b}{\left|w\right|}=\frac{2}{\left|w\right|}$

To optimizing the definition of segmentation plane, the problem of finding the solution to this plane can be reduced to as can be seen: under the situation of the formula of satisfying condition (3), calculating can maximize p (w, the normal vector w of segmentation plane b) and side-play amount b according to SVM.People such as Vapnik prove:

The normal vector w of cutting apart lineoid ₀Be the linear combination of all training set vectors.Be w ₀Can be described as:

Definition discriminant function f (x)=w ₀X+b ₀Then the classification function of test set can be described as: label (x)=sgn (f (x))=sign (w ₀X+b ₀)

Under the situation of linear separability, all should satisfy all training samples | f (x) | 〉=1, hereinafter, we are satisfying | f (x) |＜1 zone calls cuts apart the pairing borderline region of lineoid.

The finding the solution of optimum segmentation plane is equivalent under the former constraint below the maximization

Introduce Lagrange multiplier α _i, i=1,2 ..., l, and definition

Use the Wolfe antithesis fixed

$\mathrm{MaxW}\left(\mathrm{\α}\right)=\underset{i}{\mathrm{\Σ}}{\mathrm{\α}}_i-\frac{1}{2}w\left(\mathrm{\α}\right)\·w\left(\mathrm{\α}\right)$

$\mathrm{subject\; to}{\mathrm{\α}}_i\≥0,\underset{i}{\mathrm{\Σ}}{a}_i{y}_i=0$

Reason is converted into its dual problem to the problems referred to above:

For the inseparable training set of linearity, can introduce slack variable ξ _i, be rewritten as following:

$\mathrm{Min}(\frac{1}{2}{\left|\right|w\left|\right|}^2+C\underset{i}{\mathrm{\Σ}}{\mathrm{\ξ}}_i)$

Subject?to?y _i(w·x _i+b)≥1-ξ _i，ξ _i≥0

Similarly can obtain corresponding dual problem:

$\mathrm{MaxW}\left(\mathrm{\α}\right)=\underset{i}{\mathrm{\Σ}}{\mathrm{\α}}_i-\frac{1}{2}w\left(\mathrm{\α}\right)\·w\left(\mathrm{\α}\right)$

$\mathrm{subject\; to}0\≤{\mathrm{\α}}_i\≤C,\underset{i}{\mathrm{\Σ}}{a}_i{y}_i=0,$

Finding the solution of this form is a typical constrained quadratic form optimization problem, the derivation algorithm that a lot of maturations have been arranged, in recent years, V.Vapnik, C.Burges, E.Osuna, T.Joachims, people's such as J.Platt a series of activities makes the algorithm of support vector machine of extensive training set is realized becoming possibility.

By the mathematical model that above description is set up, can seek out those automatically has the support vector of better separating capacity to classification, and the sorter that constructs thus can maximize the interval of class and class, thereby adaptive faculty and higher differentiation rate are preferably arranged.This method only need decide last classification results by the classification of the boundary sample of each class field, has finally set up the support vector machine mathematical model.

6. association rule algorithm modeling

Association rule mining is in a class problem of excavating the relevance between the collection on the big transaction data set (TDS).The correlation rule analysis is a big generic task in the machine learning.It originates from the analysis to dichotomic variable, expresses two relations between the dichotomic variable with the mode of rule, and the relation between a plurality of dichotomic variable.Certainly, later development also makes correlation rule not only be confined to dichotomic variable, also can analyze many classified variables and continuous variable.So correlation rule can be regarded as between the situational variables and concerns, and this relation table is reached the method for the rule of be very easy to explaining.

The correlation rule analytical approach distributes to data and does not do any requirement, and the result of gained is complete in data, without any the subjectivity supposition, has reflected the essence of data objectively, and very strong cogency is arranged.The result that correlation rule obtains data analysis can be regarded as in the data between variable regular summary.Therefore correlation rule has obtained a large amount of application in all trades and professions after proposing.

The algorithm of correlation rule is exactly by the solution procedure of input to output.If I={i ₁, i ₂..., i _mBe m different item destination aggregation (mda), element wherein is called (Item).Note D is the set of item for the set of transaction T (Transaction), the T that concludes the business here, and

Corresponding each transaction has unique sign, and as Transaction Identification Number, note is made TID.Correlation rule be shape as

Implications, here,

And X ∩ Y=θ.X is called the prerequisite of rule, and Y is the result.Rule Support in transaction set D (Support) is meant the ratio of the number of deals that comprises X and Y and All Activity number, is designated as

Promptly

Rule

Confidence level in transaction set D (confidence) is meant the ratio of number of deals that comprises X and Y and the number of deals that comprises X, promptly

A given transaction set D excavates the correlation rule problem and is exactly and produces support and confidence level respectively greater than the correlation rule of given minimum support of user (Minsupp) and minimum confidence level (Minconf), is called strong rule.

The task of association rule mining is exactly to excavate strong rules all among the data set D.Strong regular X] Item Sets (X ∪ Y) of Y correspondence must be collection frequently, collection (∪ Y) correlation rule of deriving frequently

Degree of confidence can calculate with the support that frequently collect X and (X ∪ Y).

The mathematical model that contains description rule by above process obtains is the correlation rule modeling.

Beneficial effect: because networks development, information is explosive increase, how effectively to utilize these letters, and uses these information to boost productivity to become the problem that presses for solution.Present present situation is to have only few part can be by correct use in the information that can effectively obtain in a large number, the information that has consumed ample resources not only can not used effectively, and because Useful Information deeper is buried among the garbage, becoming more is difficult to utilize.Machine learning is to solve one of effective ways of this class problem.Along with going deep into and the specifically expansion of application of machine learning research, a large amount of machine learning modeling mission requirements have been brought.It is same because machine learning is of a great variety, the machine learning algorithm that need adapt at concrete problem description could be set up the mathematical model of compound preferably question essence feature, has often spent the machine learning model that the plenty of time seeks and can not well reflect objective reality.

Waste time and energy at the model that specific tasks are based upon on the machine learning basis,, need select according to the personal experience because the distinctiveness of specific tasks details is difficult to directly use for reference the machine learning model that other have built.Even correctly selected relatively to meet the machine learning algorithm of objective fact essence, how complicated parameter is set, also need rule of thumb or the long computing of subscriber computer obtains, single user's computing power is difficult to deal with problems fast.Simultaneously, the user need learn and use concrete machine learning software, the machine learning algorithm numerous and complicated, and user's autonomous learning need spend the plenty of time, each task that suitable user surely need solve and certain some algorithm of user's autonomous learning also differ.

The solution route that this programme provides makes full use of the strong cloud computing platform of computing power on the one hand, the computational problem of complexity when solving machine learning, utilize simple and easy usability, the transparency of cloud computing on the other hand at the user, solved the machine learning algorithm that domestic consumer is difficult to select to meet objective reality, thereby foundation can solve the machine learning model of practical problems fast, and finds suitable parameters as far as possible automatically.

Description of drawings

Fig. 1 machine learning modeling cloud computing flowchart,

Fig. 2 module relation diagram,

Fig. 3 decision tree basic configuration figure,

Fig. 4 genetic algorithm basic process figure,

Fig. 5 artificial neural network basic block diagram,

Embodiment

The present invention is the autonomous system of selection of a kind of machine learning based on cloud computing environment.By using cloud computing platform, the user need not to build the running environment of machine learning, also need not to select machine learning algorithm, more need not adjust the machine learning function of numerous and complicated and subsidiary parameter thereof, only need to use the Web mode to upload sample data, with regard to the machine learning mathematical model of setting up realistic problem of energy automated intelligent.The present invention makes the use of machine learning break away from the constraint of environment, has brought into play the advantage of cloud computing platform, makes the machine learning modeling at user transparent, has farthest reduced the use threshold of machine learning.Shortcomings such as the artificial experience that is difficult to predictability, parameter adjustment that modeling is selected, domestic consumer's difficulty have been solved when the practical application machine learning.The final platform of setting up can put together all computational resources fully in conjunction with the cloud computing advantage, realizes management automatically by software.In data analysis process, it integrates historical data and available data, makes the information of collecting more accurate, can provide the intelligence service for machine learning.The user no longer needs to be concerned about how to buy server, machine learning software according to the business demand of oneself, as long as the demand of basis oneself just can obtain the machine learning achievement by cloud computing platform, obtains the machine learning mathematical model, is used to solve practical problems.

Concrete steps are:

1. under the uniform dispatching of cloud administration module, at first by the Web interactive interface, obtain the required rough description of dealing with problems of user, comprise the problem kind, big class under promptly selecting is as selecting from expert system, cognitive simulation, planning and problem solving, data mining, the network information service, pattern recognition, fault diagnosis, natural language understanding, robot and game, other classification;

2. enable initial modeling cloud,, enter different subclass interfaces, fill in corresponding more detailed information by the big class that the user in the step 1 provides, comprise carry out that sample is uploaded, selected method for expressing, determines the interpretation of result method, usable range, expectation territory etc.

3. startup method is found cloud, compares according to the historical typical example of information of same that the user provides, and determines because of taking the machine learning algorithm of which kind of or which kind.This cloud module is accompanied by the subsequent step operation, thereby adjusts constantly according to each stage result of calculation.

4. then with the information of step 2 kind of user input, input machine learning input/output module, must unitize, after the datumization, carry out operations such as missing values processing, noise data processing, data scrubbing, data integration, data conversion, data reduction successively, so that obtain the intermediate result that general algorithm can use.

5. start the valuation functions cloud, set up valuation functions in the information of step 2 input, the quality judgement that machine learning is separated is prepared according to the user.This step mainly is to formulate the specific algorithm of estimated performance, whether uses verification methods such as cross validation and leaving-one method, bootstrap method.

6. call the EM algorithm simultaneously and support cloud, solution space is carried out maximal possibility estimation, calculate optimum solution or more excellent approximate location in solution space of separating, increase search efficiency.

7. after arriving this step, illustrate that preliminary work finishes, be about to carry out the training process of machine learning, automatic judgement by above step, call one or several concrete machine learning cloud modules respectively and learn, comprise decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud.As User Defined machine learning algorithm expand cloud, then preferentially call machine learning algorithm and expand cloud.

8. after decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud and machine learning algorithm expand the cloud startup, constantly as achievement in the middle of cloud administration module, calculating cloud, EM algorithm support cloud feedback result and the acquisition, thereby adjust self strategy automatically, approach outstanding separating.Pass through the Web interactive interface simultaneously to field feedback, comprise the step of calculating operation, the intermediate result that obtains, current optimum solution variation etc.

9. support in the process that iterates of cloud at the EM algorithm, constantly turn back to step 6, step 7 is calculated, the valuation functions of using step 5 to formulate obtains Performance Evaluation, thereby prediction algorithm is judged the outstanding degree of separating, this step needs a large amount of computational resources, thereby need utilize the calculating advantage of cloud computing, must calculate outstanding separating as far as possible.

10. when end condition satisfies, as computing time to, some generations iteration do not have more excellent separate or the iteration of algorithm own finishes, result of calculation is converted to information by the machine learning input/output module with readability, return the client by the Web interactive interface again, and provide detailed data to download, preserve the machine learning result simultaneously,, avoid double counting so that reuse.

Claims (1)

1. automatic selection method for machine learning in cloud computing environment is characterized in that the step that this method comprises is:

Step 1) is under the uniform dispatching of cloud administration module, at first by the Web interactive interface, obtain the required rough description of dealing with problems of user, comprise the problem kind, big class under promptly selecting, from expert system, cognitive simulation, planning and problem solving, data mining, the network information service, pattern recognition, fault diagnosis, natural language understanding, robot and game, other classification, select

Step 2) enable initial modeling cloud,, enter different subclass interfaces, fill in corresponding more detailed information by the big class that the user in the step 1) provides, comprise carry out that sample is uploaded, selected method for expressing, determines the interpretation of result method, usable range, expectation territory,

Step 3) startup method is found cloud, compares according to the historical typical example of information of same that the user provides, and determines because of taking the machine learning algorithm of which kind of or which kind; This cloud module is accompanied by the subsequent step operation, thereby adjusts constantly according to each stage result of calculation,

The information that step 4) is imported step 2 kind of user then, input machine learning input/output module, must unitize, after the datumization, carry out operations such as missing values processing, noise data processing, data scrubbing, data integration, data conversion, data reduction successively, so that obtain the intermediate result that general algorithm can use

Step 5) starts the valuation functions cloud, according to the user in step 2) information of input sets up valuation functions, the quality that machine learning is separated judges and prepares, thereby the specific algorithm performance is predicted,

Step 6) is called the EM algorithm simultaneously and is supported cloud, and solution space is carried out maximal possibility estimation, calculates optimum solution or more excellent approximate location in solution space of separating, and increases search efficiency,

After step 7) arrives this step, illustrate that preliminary work finishes, be about to carry out the training process of machine learning, automatic judgement by above step, call one or several concrete machine learning cloud modules respectively and learn, comprise decision Tree algorithms cloud, genetic algorithm cloud, Bayesian statistics algorithm cloud, artificial neural network algorithm cloud, algorithm of support vector machine cloud, association rule algorithm cloud; As User Defined machine learning algorithm expand cloud, then preferentially call machine learning algorithm and expand cloud,

Step 8) is calculated through above step, selects one or several algorithm clouds, with its startup, passes through the Web interactive interface simultaneously to field feedback, comprises the step of calculating operation, the intermediate result that obtains, and current optimum solution changes,

Step 9) supports in the process that iterates of cloud at the EM algorithm, constantly turn back to step 6, step 7 is calculated, judge whether to reach end condition simultaneously, if reach end condition then jump procedure 10, otherwise the performance prediction algorithm that uses step 5 to formulate is judged the outstanding degree of separating, this step needs a large amount of computational resources, thereby need utilize the calculating advantage of cloud computing, must calculate outstanding separating as far as possible

When step 10) satisfies at end condition, as arriving computing time, there be not more excellent separating or the iteration end of algorithm own, result of calculation is converted to information by the machine learning input/output module with readability, return the client by the Web interactive interface again, and provide detailed data to download, preserve the machine learning result simultaneously, so that reuse, avoid double counting.

Images