CN106156856A - The method and apparatus selected for mixed model - Google Patents
The method and apparatus selected for mixed model Download PDFInfo
- Publication number
- CN106156856A CN106156856A CN201510147815.7A CN201510147815A CN106156856A CN 106156856 A CN106156856 A CN 106156856A CN 201510147815 A CN201510147815 A CN 201510147815A CN 106156856 A CN106156856 A CN 106156856A
- Authority
- CN
- China
- Prior art keywords
- hidden variable
- initialization
- variable
- hidden
- candidate family
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Embodiments of the invention relate to the method and apparatus that mixed model selects.The method includes: determine candidate family based on training sample set;Based on candidate family first initializes at least one in hidden variable and renewal hidden variable, generate the set of the second initialization hidden variable, wherein candidate family initializes hidden variable according to first and generates, and updates the sample packet result that hidden variable represents that candidate family exports;And incompatible determine object module based on the second collection initializing hidden variable.
Description
Technical field
Embodiments of the invention relate generally to machine learning field, and more particularly, to
The method and apparatus selected for mixed model.
Background technology
Mixed model (mixture model) is a kind of to use mixed distribution for density Estimation
Probabilistic model, it can represent in a Ge great colony and there is sub-group.Mixed model can include
Multiple model, such as gauss hybrid models, piecewise linearity mixed model etc., these models by
It is widely used in multiple fields, such as document classification, handwriting recognition, broad image segmentation etc..
In practice, the Model Selection of mixed model be one extremely important and challenging
Problem.Industry has been proposed for certain methods to carry out the Model Selection of mixed model, wherein
Variation reasoning (variational inference) is a kind of relatively effective method, its attempt to
Go out the analytic approximation of limit log-likelihood.But, variation reasoning is more sensitive for initializing,
If initializing inappropriate, then the effect of variation reasoning may become very poor, thus cannot be accurate
True ground preference pattern.Therefore, initialize to have become as and affect the precision of variation reasoning and efficiency
Key factor.
At present, the initial method of variation reasoning such as can include random initializtion, based on poly-
Initialization of class etc..But, random initializtion typically requires the substantial amounts of initialization sample of trial
Realize, process time-consumingly the longest, thus the speed of Model Selection can be caused slower.It addition, base
Initializing for the mixed model (such as, gauss hybrid models) with cluster as target in cluster
Be probably relatively effective, but for for returning/be categorized as the mixed model of target not
It is suitable for using, so initializing universal based on cluster is relatively low.
Accordingly, it would be desirable to a kind of more generally applicable and efficient initialization scheme realizes mixed model choosing
Select.
Summary of the invention
Generally, embodiments of the invention propose a kind of technical scheme selected for mixed model.
In one aspect of the invention, it is provided that a kind of method selected for mixed model.Described
Method includes: determine candidate family based on training sample set;At the beginning of based on candidate family first
At least one in beginningization hidden variable and renewal hidden variable, generates the collection of the second initialization hidden variable
Closing, wherein candidate family initializes hidden variable according to first and generates, and updates hidden variable table
Show the sample packet result that candidate family exports;And based on second initialize hidden variable set
Determine object module.
In another aspect of this invention, it is provided that a kind of device selected for mixed model.Device
Including: candidate family determines unit, is configured to determine candidate family based on training sample set;
Signal generating unit, is configured to based on candidate family first and initializes hidden variable and update hidden variable
In at least one, generate the set of the second initialization hidden variable, wherein candidate family is according to the
One initializes hidden variable and generates, and updates the sample that hidden variable represents that candidate family exports and divide
Group result;And object module determines unit, it is configured to initialize hidden variable based on second
Collect incompatible and determine object module.
According to embodiments of the invention, hidden by initialization based on performance preferably candidate family
Variable and renewal hidden variable generate new initialization hidden variable, then hidden based on new initialization
Variable obtains new model, and this is not limited to certain specific blend types of models.Therefore,
According to the solution of the present invention, there is good versatility.Other features and advantages of the present invention will be logical
Cross and be described below and be easy to understand.
Accompanying drawing explanation
By combining accompanying drawing, exemplary embodiment of the invention is described in more detail, this
Bright above-mentioned and other purpose, feature and advantage will be apparent from. wherein:
Fig. 1 shows the signal of the method selected for mixed model according to embodiments of the present invention
Property flow chart;
Fig. 2 shows the signal of the method selected for mixed model according to embodiments of the present invention
Property flow chart;
Fig. 3 A show according to embodiments of the present invention for determining object initialization hidden variable
The indicative flowchart of the method for set;
Fig. 3 B show according to embodiments of the present invention for determining object initialization hidden variable
The indicative flowchart of the method for set;
Fig. 4 shows of the method selected for mixed model according to embodiments of the present invention
The schematic diagram of graph-based;
Fig. 5 shows the signal of the device selected for mixed model according to embodiments of the present invention
Property block diagram;And
Fig. 6 shows showing of the exemplary computer system be suitable to for realizing the embodiment of the present invention
Meaning property block diagram.
In the accompanying drawings, same or analogous label is used to represent same or analogous element.
Detailed description of the invention
It is more fully described the preferred implementation of the disclosure below with reference to accompanying drawings.Although accompanying drawing
In show the preferred implementation of the disclosure, however, it is to be appreciated that can be real in a variety of manners
Show the disclosure and should not limited by embodiments set forth herein.On the contrary, it is provided that these are implemented
Mode is to make the disclosure more thorough and complete, and can be complete by the scope of the present disclosure
Convey to those skilled in the art.
The mechanism of the embodiment of the present invention described in detail below and principle.Unless specifically stated otherwise,
Hereafter represent " being based at least partially on " with the term "based" of use in claim.Term " bag
Include " represent that opening includes, i.e. " include but not limited to ".Term " multiple " expression " two or more
Many ".Term " embodiment " expression " at least one embodiment ".Term " another embodiment "
Represent " at least one further embodiment ".The definition of other terms is given in will be described below.
Fig. 1 shows the method 100 for mixed model selection according to embodiments of the present invention
Flow chart.Traditionally, the initial method of variation reasoning uses random initializtion, based on cluster
The method such as initialization.But, it is relatively slow, logical that these initial methods are generally of processing speed
By problems such as property are the highest, these are all the problem demanding prompt solutions during mixed model selects.Pass through
It is described below it will be appreciated that utilize method 100 according to an embodiment of the invention, these problems
Can be obtained and efficiently solve.
Method 100 starts from step S110, determines candidate family at this based on training sample set.
In embodiments of the present invention, " hidden variable " can represent and can not be directly observed and need
The variable that sample data to be passed through is derived from.The variation distribution of hidden variable may be used for describing sample
Notebook data is clustered the probability of corresponding classification.It should be noted that, in embodiments of the present invention,
" hidden variable " is not limited to a kind of variable, but can comprise " the variation distribution of hidden variable " and/
Or other suitable information.In the disclosure, hidden variable can include initializing hidden variable, more
New hidden variable, etc., wherein initialize hidden variable and represent the hidden variable for being trained,
And the hidden variable updated obtains hidden variable after representing training.In the context of the disclosure, " more
New hidden variable " it is otherwise referred to as " renewal hidden variable ", the two can replace use.
In an embodiment of the present invention, " model " generally refers to mixed model, such as candidate family,
Mid-module and object module, etc..Model can be by instructing initialization hidden variable
Practice and generate.The model so generated can include the variation distribution of hidden variable, the model updated
Parameter, model structure, etc..Model parameter can according to the type difference of mixed model not
With, owing to mixed model is typically the general name of model I, and a mixed model can be by many
Individual sub-model combines, and therefore model parameter is associated with the type of concrete mixed model.
For example, for gauss hybrid models, model parameter can include each submodel institute
The average of the Gauss distribution obeyed and variance.For up model, model parameter is then
Can include that the condition of a node controls parameter and the regression coefficient of leaf node and deviation.Model
Structure is also associated with the type of mixed model.For example, for gauss hybrid models,
Model structure can include that the number of submodel and submodel merge coefficient, etc..For burst
For linear model, model structure then can include the tree construction learning.Should be appreciated that
State the example purpose merely for the sake of discussion, it is not intended to limit the scope of the present invention by any way.
It addition, " statistical model " mentioned in an embodiment of the present invention is different from mixed model.System
Meter model such as can include Gaussian process model, student-t process model etc..
According to one embodiment of present invention, in step S110, can be based on training sample
Collection determines that one or more first initializes hidden variable, it is possible to initialize hidden change according to first
Amount generates candidate family.First initializes hidden variable can be by the sample concentrating training sample
Carry out random packet to obtain, or can be clustered by the sample that training sample is concentrated
Obtain.First initializes hidden variable such as can be implemented as the form of matrix, data acquisition system
Form or other any suitable forms.In one embodiment, first hidden variable is initialized
Can be the matrix of k × n dimension, wherein k be the line number of this matrix, represents that first is initial
Change the sample group number of hidden variable;N is this matrix column number, represents the sample that training sample is concentrated
This number.A line in this matrix corresponds to a sample group, the often value of each element in row
Can be 0 or 1.Such as, if the i-th element in a line in this matrix is 1, then
Represent and the sample group corresponding with this row comprises the i-th sample that training sample is concentrated;If
Jth element in this row is 0, then it represents that do not comprise instruction in the sample group corresponding with this row
Practice the jth sample in sample set.Should be appreciated that the value of element in above-mentioned matrix not necessarily
It is necessary for 0 or 1, it is also possible to be any other suitable numerical value.Above example is only
The purpose that discussion is convenient, it is not intended to limit the scope of the present invention.In another embodiment,
One initialization hidden variable can be a data acquisition system, such as, can include k sample group, often
Individual sample group there is one or more sample.Should be appreciated that above-mentioned example is only used to discuss
Purpose, be not intended to limit the scope of the present invention.Those skilled in the art can be by any suitable
When mode realize the first initialization hidden variable.
When determining multiple first initialization hidden variable based on training sample set, can generate many
Individual model is used as in candidate family determined by step S110.Alternatively or additionally, may be used
To select one to be used as in candidate's mould determined by step S110 from the multiple models generated
Type.The mode selected can have multiple, such as, can randomly choose from the multiple models generated
At least one candidate family, it is also possible to select according to certain order or requirement.Should be appreciated that
The example above is only used to the purpose of discussion, it is not intended to limit the scope of the present invention.This area
Technical staff can utilize any suitable system of selection to select at least one from multiple models
Candidate family.
Such as, according to one embodiment of present invention, in step S110, can be based on instruction
Practice sample set and determine that multiple first initializes hidden variable, initialize hidden variable based on multiple first
Generate multiple initial model, and according to the performance of multiple initial models select the plurality of initially
At least one in model is as candidate family.
The performance of model can include many factors, such as precision, ageing, etc..Model
Performance can determine in several ways, such as calculate root-mean-square error, average absolute value
Error, likelihood ratio etc..As a example by precision, for prediction/classification mixed model, can be by
Candidate family is for testing the prediction/classification of data, thus obtains the prediction/nicety of grading of correspondence.
For Clustering Model, the clustering precision of candidate family, such as standard mutual information can be calculated
Deng.
It follows that method 100 proceeds to step S120, at this at the beginning of based on candidate family first
At least one in beginningization hidden variable and renewal hidden variable, generates the collection of the second initialization hidden variable
Close.
According to embodiments of the invention, can based on candidate family first initialize hidden variable,
Update hidden variable or the two, generate the set of the second initialization hidden variable.Wherein, this time
Modeling type first initializes hidden variable according to this and generates, and updates hidden variable and represent this candidate
The sample packet result of model output.
According to one embodiment of present invention, in step S120, can be based on candidate family
First initialization hidden variable, generate the 3rd initialization hidden variable set;Based on candidate family
Renewal hidden variable, generate the 4th initialization hidden variable set;And can be based at the beginning of the 3rd
The set of beginningization hidden variable and the 4th initializes the set of hidden variable, determines the second hidden change of initialization
The set of amount.
3rd set initializing hidden variable can comprise and to initialize hidden variable with first and be associated
One or more hidden variables, it is possible to generate in several ways.In one embodiment,
By initializing hidden variable to the first of candidate family and can be weighted with reference to hidden variable
Summation, generates the 3rd initialization hidden variable in the set of the 3rd initialization hidden variable.
4th set initializing hidden variable can comprise with renewal be associated one of hidden variable or
Multiple hidden variables, it is possible to generate in several ways.In one embodiment, Ke Yicong
Multiple sample groups in the renewal hidden variable of candidate family select a sample group;By selected
Sample components be multiple subgroup;And determine the 4th initialization hidden variable based on multiple subgroups
The 4th initialization hidden variable in set.
Referring still to Fig. 1, method 100 proceeds to step S130, initializes based on second at this
The collection of hidden variable is incompatible determines object module.
According to embodiments of the invention, initialize hidden variable in the second of step S120 generation
Set can including, one or more second initializes hidden variable.In one embodiment, as
Really the number of the second initialization hidden variable obtained by step S120 is 1, then in step S130
Second can initialize hidden variable by training this and obtain a model, and using this model as
Object module.
In another embodiment, if generating multiple second in step S120 to initialize hidden change
Amount, the most in step s 130, can initialize hidden variable next life by training the plurality of second
Become multiple mid-module, may then based on the performance of mid-module to determine object module.Example
As a mid-module selected from which as mesh according to the performance of each mid-module
Mark model.
According to embodiments of the invention, optionally, can be at the beginning of second in step S130
The set of beginningization hidden variable is further processed, therefrom selects the second hidden change of initialization
One subset of amount, and determine object module based on this subset.In one embodiment, exist
Step S130, it is possible to use predetermined statistical model, based on the second collection initializing hidden variable
The incompatible set determining object initialization hidden variable, initializes the collection of hidden variable by training objective
Each object initialization hidden variable in conjunction generates mid-module, and property based on mid-module
Object module can be determined.About object initialization hidden variable set determination process will with
Lower discussion combines the flow chart shown in Fig. 3 be described.
It is appreciated by the above description that embodiments of the invention are not related to certain types of mixing
The concrete property of model or requirement, therefore have good versatility, it is not limited to certain kinds
The mixed model of type.Meanwhile, embodiments of the invention are also a kind of efficient solutions.With
Lower the further advantage of embodiments of the invention will be discussed by the embodiment shown in Fig. 2.
Fig. 2 shows the method 200 for mixed model selection according to embodiments of the present invention
Indicative flowchart.Method 200 is considered the specific embodiment of method 100.Should
Understanding, shown embodiment is only in order at the purpose of discussion, it is not intended to limit this by any way
Bright scope.
Method 200 starts from step S210, determines multiple first at this based on training sample set
Initialize hidden variable.
In the embodiment shown in Figure 2, training sample set is one and includes multiple sample (also referred to as
For sample data) set.Each sample can be characterized by multiple parameters.One sample
Can be expressed as that (x, y), wherein x represents the feature of this sample, and y represents estimating for this sample
Evaluation.Such as, the x of the sample of power consumption data can include power consumption today, temperature,
Humidity etc., y then can represent estimated power consumption tomorrow, i.e.
X=[ele-consumption_today, temperature, humidity ...],
Y=ele-consumption_tomorrow.In the following discussion, mentioned " sample group "
Represent one group of sample, this sample group potentially includes one or more sample.In discussion below
In, mentioned " to sample packet " or " being grouped sample " refers to divide sample
For some groups, each group is i.e. a sample group.
According to embodiments of the invention, in step S210, can give birth to randomly from training sample set
Multiple first is become to initialize hidden variable.Optionally, it is also possible to training sample is concentrated
Sample clusters, thus generates multiple first and initialize hidden variable.In addition, it is also possible to
Use those skilled in the art can additive method determine at the beginning of first based on training sample set
Beginningization hidden variable, above-mentioned exemplary embodiment is not limitation of the present invention.
In step S220, initialize hidden variable based on the plurality of first and generate multiple initial models.
According to embodiments of the invention, can for multiple first initialize in hidden variables each
First initializes hidden variable, learns on training sample set, thus generates and each first
Initialize the initial model that hidden variable is corresponding.In one embodiment, at the beginning of each first
Beginningization hidden variable, can exercise supervision on training sample set or unsupervised learning, automatically learn
Practise the model parameter of mixed model, update hidden variable (or the hidden variable variation distribution updated)
And model structure, thus obtain the initial model of correspondence.In this way, can obtain
Multiple initial models that hidden variable is corresponding are initialized with multiple first.
It follows that method 200 proceeds to step S230, property based on multiple initial models at this
At least one can be selected to be used as candidate family from multiple initial models.
The performance of model such as can be obtained by the precision of computation model, ageing etc..With
As a example by precision, for prediction/classification mixed model for, can test data prediction/point
Apoplexy due to endogenous wind uses the distribution of hidden variable variation, model parameter and the model structure of the renewal of initial model,
Thus obtain the prediction/nicety of grading of correspondence.For Clustering Model, introductory die can be calculated
The clustering precision of type, such as standard mutual information etc..Should be appreciated that in addition to precision, also may be used
To be determined by the ageing grade of initial model about the performance because usually obtaining initial model.On
State example and be intended merely to facilitate discussion in this, be not limitation of the present invention.
The process selecting candidate family from multiple initial models can be real in several ways
Existing.In one embodiment, this bigger set comprising multiple initial model can be contracted
Little is a less set, thus the most one or more for performance initial models are selected with
As candidate family.In one embodiment, can every by based in multiple initial models
The distribution of hidden variable variation, model parameter and the model structure of the renewal of individual initial model, determines
The performance of each initial model.It is then possible to performance determined by based on is to multiple initial models
It is ranked up, for example, it is possible to arrange the plurality of initial model according to performance order from high to low.
Thus, it is possible to the initial model that selected and sorted is the most forward from multiple initial models is used as
Candidate family.In this case, the number of the candidate family obtained in step S230 is 1.
Optionally, can based on sequence from multiple initial models, selected and sorted is forward two
Or two or more initial model is used as candidate family.In this case, in step S230
Having of the candidate family obtained is multiple, and is the most multiple model of performance in initial model.
Referring still to Fig. 2, method 200 proceeds to step S240, based on candidate family at this
First initializes hidden variable, generates the set of the 3rd initialization hidden variable.
In one embodiment, can by the first of candidate family the initialization hidden variable and
It is weighted summation with reference to hidden variable, generates the 3rd in the set of the 3rd initialization hidden variable
Initialize hidden variable.Can be the hidden variable of a uniform sampling with reference to hidden variable, it can be with
The form of the first initialization hidden variable is identical, and the matrix of e.g. k × n dimension, wherein k is this
The line number of matrix, represents the sample group number of the first initialization hidden variable;N is this matrix column
Number, represents the number of samples that training sample is concentrated.In one embodiment, it is assumed that first is initial
Change hidden variable and be represented as qz_initial, be represented as qz_new with reference to hidden variable, then the 3rd
Initialize hidden variable to be calculated by following:
Qz=a*qz_initial+ (1+a) * qz_new (1)
Wherein, a represents the first weighter factor initializing hidden variable, and qz represents that the 3rd initialization is hidden
Variable, and qz is the matrix of k × n dimension the most accordingly, each element in this matrix
Value can be 0 or 1, or other suitable numerical value that those skilled in the art commonly use.
Correspondingly, the 3rd set initializing hidden variable can be determined that:
S1={qz | qz=a*qz_initial+ (1-a) qz_new, a ∈ (0,1), qz ∈ [0,1]k×n}。
In step S250, renewal hidden variable based on candidate family, generate the 4th hidden change of initialization
The set of amount.
According to embodiments of the invention, update hidden variable and can represent the sample that candidate family exports
Group result.Multiple sample groups from the renewal hidden variable of candidate family can select one
Sample group.In an embodiment of the present invention, it is possible to use various ways is from the renewal of candidate family
Multiple sample groups (the such as number of sample group is N) of hidden variable select a sample group,
Such as can randomly choose or can be according to the scale of sample group to carry out above-mentioned selection.One
In individual embodiment, it may be determined that the number of samples of each sample group in the plurality of sample group,
And the sample group that selection number of samples is maximum from multiple sample groups, i.e. select largest
Sample group.Optionally, in one embodiment, can select from multiple sample groups
Select the number of samples sample group more than predetermined threshold, i.e. select scale to exceed the sample of predetermined threshold
Group.Should be appreciated that above example is only schematically, be not the protection model to the present invention
The restriction enclosed.
Select from multiple sample groups of the renewal hidden variable of candidate family sample group it
After, can be multiple subgroup by selected sample components.According to one embodiment of present invention,
Selected sample group can be randomly divided into multiple subgroup.As an alternative solution, can be by institute
The sample group selected is equally divided into multiple subgroup, and the most each subgroup has equal number of sample.
It is to be understood that, it is also possible to by other modes multiple, selected sample components is become multiple subgroup,
More than being merely exemplary, being not construed as is limiting the scope of the present invention.Additionally,
The number of the subgroup obtained from selected sample group can be predefined, and this number is such as
Can be 2,3 ..., any one in k ', wherein k ' is less than or equal to selected
The integer of the number of samples in sample group, simultaneously k ' and described renewal dependent variable except selected sample
Residue sample group number sum outside group need to be less than or equal to the sample of the first initialization hidden variable
This group number.
After selected sample components is multiple subgroup, can be true based on the plurality of subgroup
The 4th initialization hidden variable in the set of fixed 4th initialization hidden variable.In one embodiment,
The number of subgroup can be calculated and remaining updated in hidden variable in addition to selected sample group
Sample group number sum, and the result calculated is initialized the sample group in hidden variable with first
Number compares.
If the result calculated is less than the sample group number in the first initialization hidden variable, then may be used
So that selected sample group is removed from renewal hidden variable.Then, can continue executing with above
Operation, i.e. from one sample group of middle reselection of this renewal hidden variable, and by selected sample
This component is multiple subgroup.It is then possible to calculate the sum of all subgroups (if from this renewal
Hidden variable selected N number of sample group, then corresponding to not obtaining from this N number of sample components
The sum of subgroup) and this renewal hidden variable in remaining in addition to selected all sample groups
Sample group was (if selecting N number of sample group, remaining sample the most now from this renewal hidden variable
This group is the residue sample group in renewal hidden variable in addition to this N number of sample group) sum, and
Based on the comparison to the result calculated with the sample group number in the first initialization hidden variable, come
Judge whether that continuing iteration performs aforesaid operations.
If it is determined that the result calculated is equal to the sample group number in the first initialization hidden variable,
Then by all subgroups of obtaining and this renewal hidden variable in addition to selected all sample groups
Remaining sample group initialize the sample group in hidden variable as the 4th.Thus, it is possible to determine
One the 4th initializes hidden variable.
If it is determined that the result calculated is more than the sample group number in the first initialization hidden variable,
From remaining sample group of all subgroups and this renewal hidden variable, then select minimum both of scale,
And both are merged to generate a new sample group, then replace above-mentioned two by new sample group
Person's (that is, two sample group/subgroups selected by replacement).Repeat aforesaid operations until all sons
Remaining sample group number sum of group and this renewal hidden variable is equal in the first initialization hidden variable
Sample group number, then can be by remaining the sample group in all subgroups and this renewal hidden variable
It is used as the sample group in the 4th initialization hidden variable, thereby determines that out that one the 4th initialization is hidden
Variable.
Should be appreciated that above-mentioned example is only used to the purpose of discussion, be not intended to the present invention's
Scope limits.Those skilled in the art can use other in the scope of the invention any
Suitable means are carried out renewal hidden variable based on candidate family and are generated the 4th initialization hidden variable.
Step S260, initializes hidden variable based on the 3rd set initializing hidden variable and the 4th
Set, determines the set of the second initialization hidden variable.
In one embodiment, the set of hidden variable, the 4th initialization can be initialized by the 3rd
The set of hidden variable or the union of the two are defined as the set of the second initialization hidden variable.Make
For substituting, in one embodiment, can be from the third and fourth set initializing hidden variable
Determine a subset respectively, then the union of two subsets can be defined as the second initialization hidden
The set of variable.The determination method of above-mentioned subset can have multiple, such as can be randomly from
Three and the 4th set initializing hidden variable determines a subset respectively.Should be appreciated that above-mentioned
Example is only used to the purpose discussed, and is not intended to limit the scope of the present invention.Ability
Field technique person personnel can utilize other any suitable means to come based on the 3rd and initialize hidden change
The set of amount and the 4th initializes the set of hidden variable, determines the set of the second initialization hidden variable.
Referring still to Fig. 2, method 200 proceeds to step S270, at this by training at the beginning of second
In the set of beginningization hidden variable each second initializes hidden variable and generates mid-module.
In one embodiment, can be for each the in the second set initializing hidden variable
Two initialize hidden variable, learn on training sample set, such as, exercise supervision or without prison
Educational inspector practises, automatically the study model parameter of mixed model, the hidden variable variation distribution of renewal and
Model structure, initializes the model that hidden variable is corresponding, this model quilt to generate with each second
It is referred to as mid-module.
In step S280, performance based on mid-module determines object module.
The performance of model can include many factors, such as precision, ageing etc., it is possible to logical
Cross various ways to determine.As a example by precision, if mid-module is prediction/classification mixed model,
Mid-module can be used for testing the prediction/classification of data, thus obtain the prediction of correspondence/point
Class precision.If mid-module is Clustering Model, the clustering precision of mid-module can be calculated,
Such as standard mutual information etc..
In one embodiment, can be become by the hidden variable of renewal based on each mid-module
Point distribution, model parameter and model structure, determine each mid-module precision, ageing or
Other characteristics of person, can obtain the performance of each mid-module then.It is then possible to really
Multiple mid-modules are ranked up by fixed performance, such as, according to performance order from high to low
Arrange the plurality of mid-module.Subsequently, best performance can be selected from multiple mid-modules
One mid-module is as object module.
As optional step, after step S280, can be by according to embodiments of the invention
The object module obtained is supplied to user, in order to user uses.
It should be noted that, what the method 200 shown in Fig. 2 was merely exemplary, it is not intended to appoint
Where formula limits the scope of the present invention.According to embodiments of the invention, it is not necessary to determining second
Directly generate mid-module and then determine object module after initializing the set of hidden variable, but
Second set initializing hidden variable can be further processed, therefrom select at the beginning of second
One subset of beginningization hidden variable, and determine object module based on the subset selected.At one
In embodiment, alternative steps S270 and S280, it is possible to use predetermined statistical model, based on
The second incompatible set determining object initialization hidden variable of collection initializing hidden variable, is then based on
The collection of object initialization hidden variable is incompatible determines object module.Can have there are ways to based on
The second incompatible set determining object initialization hidden variable of collection initializing hidden variable.Fig. 3 A shows
Go out showing of the method 300 for determining object initialization hidden variable according to embodiments of the present invention
Meaning property flow chart.
Method 300 starts from step S310, and wherein the set according to the second initialization hidden variable is right
The set of object initialization hidden variable initializes.
The set of object initialization hidden variable is carried out initialization can enter in several ways
OK.In one embodiment, at least one can be selected from the second set initializing hidden variable
Individual second initializes hidden variable, as the object initialization in the set of object initialization hidden variable
Hidden variable.In an additional embodiment, it is also possible to from the second set initializing hidden variable
This at least one second initialization hidden variable selected by removal.Above-mentioned selection course can utilize
Multiple method realizes, such as, can randomly choose, namely initializes hidden variable from second
Set in randomly choose one or more second initialize hidden variable be used as object initialization
Hidden variable.Again for example, it is possible to from second initialize hidden variable set chosen distance farthest
Multiple second initializes multiple mesh that hidden variable is used as in the set of object initialization hidden variable
Mark initializes hidden variable.This distance such as can by calculate Euclidean distance, Hamming distance,
Mahalanobis distance, included angle cosine or art technology person personnel can other suitable computational methods
Obtain.In this way, it is possible to based on the second incompatible initialized target of collection initializing hidden variable
Initialize the set of hidden variable.
In step S320, utilize the hidden change of the object initialization in the set of object initialization hidden variable
The performance of amount and the temporary pattern corresponding with object initialization hidden variable is trained predetermined
Statistical model.
In an embodiment according to the present invention, predetermined statistical model can be such as user according to
Need or preference specify or system statistical model set in advance.Predetermined statistics mould
Type can be Gaussian process model, student-t process model or other suitable statistical models.
In one embodiment, it is possible to use all mesh in the set of object initialization hidden variable
Mark initializes the performance of hidden variable and corresponding temporary pattern thereof and constructs a new training
Sample set.Each sample that this new training sample is concentrated can be by (x, y) represents, wherein
X represents object initialization hidden variable, and y represents and this corresponding facing of object initialization hidden variable
Time model performance.It is then possible to carry out supervised learning on the training sample set that this is new with instruction
Practice predetermined statistical model.
In step S330, initialize hidden change based on trained predetermined statistical model from second
The set of amount selects the second initialization hidden variable, to update the set of object initialization hidden variable.
According to embodiments of the invention, it is possible to use trained predetermined statistical model comes pre-
Survey and initialize corresponding interim of hidden variable with second in the second set initializing hidden variable
The performance of model.In one embodiment, it is possible to use trained predetermined statistical model comes
Calculate average and the variance of the performance initializing the corresponding temporary pattern of hidden variable with second, so
Determine with variance according to this average afterwards and initialize, with second, the temporary pattern that hidden variable is corresponding
Performance prediction value.
After completing performance prediction, the second initialization can be selected according to the performance predicted
It is hidden that at least one second initialization hidden variable in the set of hidden variable updates object initialization
The set of variable.In one embodiment, can be by the second set initializing hidden variable
The second initialization hidden variable corresponding with the optimal performance predicted is defined as object initialization
An object initialization hidden variable in the set of hidden variable.In an additional embodiment, also
Can remove corresponding with the optimal performance predicted from the second set initializing hidden variable
Second initialization hidden variable.So, while updating the set of object initialization hidden variable,
The renewal to the second set initializing hidden variable can also be realized.
Should be appreciated that what method 300 was merely exemplary, and be not to the scope of the present invention
Restriction.According to embodiments of the invention, update object initialization hidden variable in step S330
Set after, can continue return step S320, utilize the object initialization after renewal
Object initialization hidden variable in the set of hidden variable and temporary pattern corresponding thereto
Performance continues to train predetermined statistical model.It is then possible to based on trained predetermined
Statistical model selects the second initialization hidden variable from the second set initializing hidden variable, again
Update the set of object initialization hidden variable.Above-mentioned steps S320-S330 can be performed a plurality of times,
Until reaching default execution number of times.
Fig. 3 B show according to embodiments of the present invention for determining object initialization hidden variable
The indicative flowchart of method 340.Method 340 may be considered being embodied as of method 300
Example.Should be appreciated that shown embodiment is only in order at the purpose of discussion, it is not intended to by any way
Limit the scope of the present invention.
Method 340 starts from step S341, selects from the second set initializing hidden variable at this
Select at least one second initialization hidden variable, as the mesh in the set of object initialization hidden variable
Mark initializes hidden variable.
In one embodiment, one can be randomly choosed from the second set initializing hidden variable
Individual or multiple second initialization hidden variables.In an alternative embodiment, can be initial from second
The multiple second initialization hidden variables changing chosen distance in the set of hidden variable farthest are used as mesh
Mark initializes the multiple object initialization hidden variables in the set of hidden variable.This distance such as may be used
With by calculating Euclidean distance, Hamming distance, mahalanobis distance, included angle cosine or this area skill
Art person personnel can other suitable computational methods obtain.Should be appreciated that above example is only
It is only exemplary, is not limiting the scope of the present invention.In addition, it is also possible to logical
Cross other suitable modes select from the second set initializing hidden variable at least one the
Two initialize hidden variable.
In step S342, from the second set initializing hidden variable, remove at least one of selection
Second initializes hidden variable.
In one embodiment, it is assumed that second initialize hidden variable set in original N number of second
Initialize hidden variable, be expressed as Z1、Z2……、ZN, and in step S341 from the beginning of second
The set of beginningization hidden variable have selected one second and initialize hidden variable Z1, then in step S342
By Z1Remove from the second set initializing hidden variable.So, second hidden variable is initialized
Set in be updated to have N-1 second and initialize hidden variable, i.e. Z2……、ZN。
In another embodiment, it is assumed that in step S341 from the second collection initializing hidden variable
Conjunction have selected multiple (such as j) second and initialize hidden variable Z1、……Zj, then in step
Rapid S342 is by Z1、……ZjRemove from the second set initializing hidden variable.So,
Two set initializing hidden variable are updated to have N-j second and initialize hidden variable, i.e.
Zj+1……、ZN。
In step S343, generate hidden with the object initialization in the set of object initialization hidden variable
The temporary pattern that variable is corresponding.
According to embodiments of the invention, can every in the set of object initialization hidden variable
Individual object initialization hidden variable, learns on training sample set, thus generates and each mesh
Mark initializes the model that hidden variable is corresponding, hereon referred to as " temporary pattern ".An enforcement
In example, for each object initialization hidden variable, can exercise supervision on training sample set or
Person's unsupervised learning, automatically learn mixed model model parameter, update hidden variable (or
The hidden variable variation distribution updated) and model structure, thus obtain the temporary pattern of correspondence.
In this way, can generate with each target in the set of object initialization hidden variable at the beginning of
The temporary pattern that beginningization hidden variable is corresponding.
In one embodiment, above-mentioned training process can be light-duty training (lightly train),
That is, perform variation reasoning algorithm, to preset step number relative with object initialization hidden variable to learn
The mixed model answered, it is not necessary to perform variation reasoning algorithm until it is restrained.In this way,
Can improve the speed of service, minimizing processes the time.Should be appreciated that above-mentioned example is not to this
The restriction of the scope of invention, in alternative embodiments of the present invention, can during above-mentioned training
To learn completely (fully train), namely perform variation reasoning algorithm until its convergence with
Learn out the mixed model corresponding with object initialization hidden variable.
In step S344, utilize the hidden change of the object initialization in the set of object initialization hidden variable
The performance of amount and corresponding temporary pattern thereof trains predetermined statistical model.
In one embodiment, in step S344 firstly the need of determining change hidden with object initialization
Measure the performance of corresponding temporary pattern.In one embodiment, can be for object initialization
Hidden variable, exercises supervision or unsupervised learning on training sample set, automatically learns hybrid guided mode
The model parameter of type, renewal hidden variable (or the hidden variable variation distribution updated) and model
Structure, thus obtain the temporary pattern corresponding with object initialization hidden variable.It is then possible to
The performance of temporary pattern is obtained by the precision of computation model, ageing etc..As a example by precision,
For prediction/classification mixed model, can use in the prediction/classification of test data and face
Time the hidden variable variation distribution of renewal of model, model parameter and model structure, thus obtain right
Prediction/the nicety of grading answered.For Clustering Model, the cluster essence of temporary pattern can be calculated
Degree, such as standard mutual information etc..
By the way, the obtained temporary pattern corresponding with object initialization hidden variable
Performance likely there is one or more feature (such as precision, ageing etc.).It addition,
For precision, it is also possible to have various ways, every kind of form can also constitute the one of performance
Individual feature.In one embodiment, when with the i-th in the set of object initialization hidden variable
The performance of the temporary pattern that object initialization hidden variable is corresponding has a feature, such as, be only
During the precision of a certain form of this temporary pattern, it is possible to use the set of object initialization hidden variable
In object initialization hidden variable and the performance of corresponding temporary pattern construct one new
Training sample concentrate a sample.This sample such as can pass through (xi,yi) represent, its
Middle xiRepresent the i-th object initialization hidden variable in the set of object initialization hidden variable, yi
Represent that the performance of the temporary pattern corresponding with i-th object initialization hidden variable is (the most a certain
The precision of form).
In another embodiment, when with the jth mesh in the set of object initialization hidden variable
Mark initializes the performance of the corresponding temporary pattern of hidden variable and has multiple (such as m) feature,
Such as during the precision of various ways and/or various ways ageing, new training sample can be constructed
Multiple samples of this concentration.The plurality of sample such as can pass through (xj,yjp) represent, wherein xj
Represent jth object initialization hidden variable, yipRepresent and i-th object initialization hidden variable phase
Pth the feature of the performance of corresponding temporary pattern, wherein p=1,2 ..., m.
After building new training sample set, can carry out on the training sample set that this is new
Supervised learning is to train predetermined statistical model.Predetermined statistical model can be such as that user is pre-
That first arrange or system requirements statistical model.Predetermined statistical model can be such as high
This process model, student-t process model or other suitable statistical models.
It is to be understood that what above-described embodiment was merely exemplary, it is not intended to the scope of the present invention is carried out
Limit.In other embodiments of the invention, when constructing new training sample set, can be with base
All object initialization hidden variables in the set of object initialization hidden variable are carried out, it is possible to
Be based only upon partial target initialize hidden variable carry out.
In step S345, trained predetermined statistical model is utilized to predict initial with second
Change the performance of the corresponding temporary pattern of the second initialization hidden variable in the set of hidden variable.
In this step, it is possible to use trained predetermined statistical model calculate with second at the beginning of
The average of the performance of the temporary pattern that beginningization hidden variable is corresponding and variance.It is then possible to according to
The average calculated determines the temporary pattern corresponding with the second initialization hidden variable with variance
Performance prediction value.
In one embodiment, if the change hidden with object initialization obtained in step S344
The performance of the corresponding temporary pattern of i-th object initialization hidden variable in the set of amount has
One feature, corresponding sample is (xi,yi), then utilize trained pre-in step S345
The average determining the performance that statistical model can calculate this temporary pattern (is such as expressed as
meani) and a variance (be such as expressed as variancei).It is then possible to based on this average
A performance predictive value is obtained with variance.For example, it is possible to calculated as below:
Pi=meani+α*variancei (2)
Wherein PiRepresent the performance prediction value corresponding with i-th object initialization hidden variable;α represents
Weighter factor, it such as can be set by user or be predefined for certain fixing as required
Value.
In another embodiment, if the performance of the temporary pattern obtained in step S344
Having multiple feature, corresponding sample is (xj,yjp), then utilize through instruction in step S345
The predetermined statistical model practiced can calculate multiple averages of the performance of this temporary pattern (such as
meanjp) and multiple variance (such as variancejp).It is then possible to based on these averages and
Variance obtains a performance predictive value.For example, it is possible to calculated as below:
Wherein PjRepresent the performance prediction value corresponding with jth object initialization hidden variable;α represents
Weighter factor, it such as can be set by user or be predefined for certain fixing as required
Value.
Should be appreciated that in an embodiment of the present invention, performance prediction value can by multiple other
Suitable mode obtains.Above example is intended merely to facilitate the purpose of discussion, it is not intended to right
The scope of the present invention limits.
In step S346, according to the performance predicted, select the set of the second initialization hidden variable
In at least one second initialize hidden variable and update the set of object initialization hidden variable.
In one embodiment, can by second initialize hidden variable set in predicted
Optimal performance corresponding second initialization hidden variable be defined as object initialization hidden variable
An object initialization hidden variable in set.It is then possible to initialize hidden variable from second
Set is removed the second initialization hidden variable corresponding with the optimal performance predicted, thus can
To update the set of the second initialization hidden variable.In one implementation, can based on really
The second initialization hidden variable in second set initializing hidden variable is arranged by fixed performance
Sequence.For example, it is possible to initialize hidden variable according to performance prediction value order from high to low to second
It is ranked up.Thus, it is possible to select and optimal performance (that is, the performance that performance prediction value is maximum)
Corresponding one second initializes hidden variable.In this case, can be by step S346
One the second initialization hidden variable selected joins in the set of object initialization hidden variable, from
And complete the renewal of the set to object initialization hidden variable.
Optionally, in one embodiment, can select and sort forward two or
The second initialization hidden variable that two or more performance is associated.In this case, in step
Two or more the second initialization hidden variables selected can be joined at the beginning of target by S346
In the set of beginningization hidden variable, thus complete the renewal of the set to object initialization hidden variable.
In step S347, determine whether the number of object initialization hidden variable reaches predetermined threshold.
In one embodiment, can be for the first prime number in the set of object initialization hidden variable
Mesh set a predetermined threshold, this predetermined threshold can be those skilled in the art rule of thumb or
Person's preference is arranged, also being that system is fixedly installed.If it is determined that object initialization hidden variable
Set in the number of object initialization hidden variable be not up to this predetermined threshold, then can be with iteration
Ground performs step S343 to S346, with the new object initialization hidden variable determined.If really
It is pre-that the number of the object initialization hidden variable in the set initializing hidden variable that sets the goal reaches this
Determine threshold value, then can stop iteration;Come it is then possible to initialize hidden variable by training objective
Generate mid-module, and performance based on mid-module determines object module.
Fig. 4 shows of the method selected for mixed model according to embodiments of the present invention
The schematic diagram of graph-based 400.Mixed model selection course shown in Fig. 4 and the step of Fig. 1
Rapid S110-S130 is corresponding.As shown in Figure 4, it is possible, firstly, to determine based on training sample set
Candidate family.It is then possible to based on this candidate family first initializes hidden variable and updates hidden
At least one in variable, generates the set of the second initialization hidden variable.This second initialization is hidden
The set of variable can include that one or more second initializes hidden variable, and it is denoted respectively as
Z1,Z2,…,Zk1, wherein Z represents the second initialization hidden variable, and k1 represents obtained second
Initialize the number of hidden variable.It follows that can be based on the set of this second initialization hidden variable
Determine object module.As shown in Figure 4, hidden variable can be initialized by training multiple second
Z1,Z2,…,Zk1Generate multiple mid-module, i.e. mid-module 1, mid-module 2 ...
Mid-module k1, may then based on the performance of this k1 mid-module to select one of them
It is used as object module.
Should be appreciated that what the figure shown in Fig. 4 was merely exemplary, it is not intended to limit by any way
The scope of the present invention.According to embodiments of the invention, it is not necessarily to from hidden with the second initialization
Variable Z1,Z2,…,Zk1Corresponding mid-module 1, mid-module 2 ... mid-module
K1 obtains object module.Such as, in an alternative embodiment, it is possible to use predetermined system
Meter model, initializes hidden variable Z based on second1,Z2,…,Zk1Determine the hidden change of object initialization
The set of amount, may then based on that the collection of this object initialization hidden variable is incompatible determines object module.
Should be appreciated that what above-described embodiment was merely exemplary, be not the limit to the scope of the present invention
System.Within the scope of the invention, embodiments of the invention can be carried out by those skilled in the art
Various deformation.
Fig. 5 shows the device 500 for mixed model selection according to embodiments of the present invention
Schematic block diagram.As it can be seen, device 500 includes: candidate family determines unit 510, quilt
It is configured to training sample set to determine candidate family;Signal generating unit 520, is configured to base
In the first initialization hidden variable of candidate family and update in hidden variable at least one, generate the
Two set initializing hidden variable, wherein candidate family initializes hidden variable according to first and generates,
And update the sample packet result that hidden variable represents that candidate family exports;And object module is true
Cell 530, is configured to incompatible determine object module based on the second collection initializing hidden variable.
In one embodiment, candidate family determines that unit 510 may include that the first initialization
Hidden variable determines unit, is configured to determine the first initialization hidden variable based on training sample set;
And the first candidate family signal generating unit, it is configured to initialize hidden variable according to first and generates and wait
Modeling type.
In one embodiment, candidate family determines that unit 510 may include that the first initialization
Hidden variable determines unit, is configured to determine that multiple first initialization is hidden based on training sample set
Variable;Initial model determines unit, is configured to initialize hidden variable based on multiple first and generates
Multiple initial models;And the second candidate family signal generating unit, be configured to according to multiple initially
The performance of model, selects at least one in multiple initial model as candidate family.
In one embodiment, signal generating unit 520 may include that the 3rd hidden variable determines unit,
It is configured to based on candidate family first and initializes hidden variable, generate the 3rd initialization hidden variable
Set;4th hidden variable determines unit, is configured to renewal hidden variable based on candidate family,
Generate the set of the 4th initialization hidden variable;And second hidden variable determine unit, be configured to
Initialize the set of hidden variable based on the 3rd set initializing hidden variable and the 4th, determine second
Initialize the set of hidden variable.
In one embodiment, the 3rd hidden variable of signal generating unit 520 determines that unit may include that
3rd hidden variable signal generating unit, is configured to the to candidate family first initialization hidden variable
And it is weighted summation with reference to hidden variable, generate the in the set of the 3rd initialization hidden variable
Three initialize hidden variable.
In one embodiment, the 4th hidden variable of signal generating unit 520 determines that unit may include that
First selects unit, in the multiple sample groups being configured to from the renewal hidden variable of candidate family
Select a sample group;Grouped element, being configured to selected sample components is many height
Group;And the 4th hidden variable determines unit can be further configured to based on multiple subgroups and determines
The 4th initialization hidden variable in 4th set initializing hidden variable.
In one embodiment, the 4th hidden variable of signal generating unit 520 determines the first choosing of unit
Select unit and may include that number of samples determines unit, be configured to determine that in multiple sample group
The number of samples of each sample group;And sample group selection unit, it is configured to from multiple samples
Group selects the sample group that number of samples is maximum.
In one embodiment, object module determines that unit 530 may include that mid-module is raw
Become unit, be configured to train at the beginning of each second in the set of the second initialization hidden variable
Beginningization hidden variable generates mid-module;And object module signal generating unit, be configured to based on
The performance of mid-module determines object module.
In one embodiment, object module determines that unit 530 may include that object initialization
Hidden variable determines unit, is configured to, with predetermined statistical model, initializes hidden based on second
The incompatible set determining object initialization hidden variable of collection of variable;Mid-module signal generating unit, quilt
It is configured to pass the hidden change of each object initialization in the set of training objective initialization hidden variable
Amount generates mid-module;And object module signal generating unit, it is configured to based on mid-module
Performance determine object module.
In one embodiment, object module determines that the object initialization hidden variable of unit 530 is true
Cell may include that initialization unit, is configured to according to the second collection initializing hidden variable
Close the set to object initialization hidden variable to initialize;Statistical model training unit, is joined
It is set to utilize the object initialization hidden variable in the set of object initialization hidden variable and and mesh
Mark initializes the performance of the corresponding temporary pattern of hidden variable and trains predetermined statistical model;With
And first updating block, it is configured to based on trained predetermined statistical model from the beginning of second
The set of beginningization hidden variable selects the second initialization hidden variable, to update the hidden change of object initialization
The set of amount.
In one embodiment, object module determines that the object initialization hidden variable of unit 530 is true
The initialization unit of cell may include that the second selection unit, is configured to from second initial
Change and the set of hidden variable selects at least one second initialization hidden variable, as object initialization
Object initialization hidden variable in the set of hidden variable.
In one embodiment, object module determines that the object initialization hidden variable of unit 530 is true
First updating block of cell may include that performance prediction unit, be configured to, with through
The predetermined statistical model of training is predicted and second in the second set initializing hidden variable
Initialize the performance of the corresponding temporary pattern of hidden variable;And second updating block, it is configured
By according to the performance predicted, select at least one in the set of the second initialization hidden variable the
Two initialize hidden variable updates the set of object initialization hidden variable.
In one embodiment, performance prediction unit may include that the first computing unit, is joined
It is set to utilize trained predetermined statistical model to calculate relative with the second initialization hidden variable
The average of the performance of the temporary pattern answered and variance;And second computing unit, it is configured to root
Determine and the second performance initializing the corresponding temporary pattern of hidden variable with variance according to average
Predictive value.
In one embodiment, the second updating block may include that object initialization hidden variable obtains
Take unit, be configured to by second initialize hidden variable set in the optimality predicted
The second initialization hidden variable that can be corresponding is defined as in the set of object initialization hidden variable
One object initialization hidden variable;And second initialize hidden variable updating block, be configured to
Corresponding with the optimal performance predicted the is removed from the second set initializing hidden variable
Two initialize hidden variable.
For clarity, selectable unit that device 500 comprises and each it is shown without in Figure 5
The subelement that individual unit is comprised.Should be appreciated that device 500 can utilize to be embodied in various ways.
Such as, in certain embodiments, device 500 can utilize software and/or firmware to realize.Example
As, device 500 may be implemented as the computer program product comprised on a computer-readable medium
Product, each unit therein is the program module being realized its function by computer instruction.Standby
Selection of land or additionally, device 500 can partially or fully realize based on hardware.Example
As, device 500 can be implemented as integrated circuit (IC) chip, special IC (ASIC)
Or SOC(system on a chip) (SOC).Other modes currently known or exploitation in the future are also feasible
, the scope of the present invention is not limited in this respect.
Fig. 6 shows the exemplary computer system 600 be suitable to for realizing the embodiment of the present invention
Schematic block diagram.As it can be seen, computer system 600 includes CPU (CPU)
601, it can be according to the program being stored in read only memory (ROM) 602 or from depositing
Storage unit 608 is loaded into the program in random access storage device (RAM) 603 and performs various
Suitable action and process.In RAM 603, also storage has CPU 601 to perform various place
The data that reason etc. are required.CPU 601, ROM 602 and RAM 603 are by bus 604
It is connected with each other.Input/output (I/O) unit 605 is also connected to bus 604.
It is connected to I/O interface 605: include the input block 606 of keyboard, mouse etc. with lower component;
Including such as cathode ray tube (CRT), liquid crystal display (LCD) etc. and speaker etc.
Output unit 607;Memory element 608 including hard disk etc.;And include such as LAN card,
The communication unit 609 of the NIC of modem etc..Communication unit 609 is via such as
The network of the Internet performs communication process.Driver 610 is connected to I/O interface also according to needs
605.Removable medium 611, such as disk, CD, magneto-optic disk, semiconductor memory etc.,
It is arranged on as required in driving 610, in order to the computer program read from it is according to need
Memory element 608 to be mounted into.
Especially, according to embodiments of the invention, each process above-described can be implemented
For computer software programs.Such as, embodiments of the invention include a kind of computer program,
It includes the computer program being tangibly embodied on machine readable media, described computer program
Comprise the program code for performing each method.In such embodiments, this computer program
Can be downloaded and installed from network by communication unit 609, and/or from removable medium
611 are mounted.
It is said that in general, the various example embodiment of the present invention can be at hardware or special circuit, soft
Part, logic, or its any combination are implemented.Some aspect can be implemented within hardware, and its
His aspect can by controller, microprocessor or other calculate firmware that equipment performs or
Software is implemented.When each side of embodiments of the invention is illustrated or described as block diagram, flow process
Figure or when using some other figure to represent, it will be appreciated that square frame described herein, device, system,
Techniques or methods can be as nonrestrictive example at hardware, software, firmware, special circuit
Logic, common hardware or controller or other calculate equipment, or its some combinations is implemented.
And, each frame in flow chart can be counted as method step, and/or computer program
The operation that the operation of code generates, and/or it is interpreted as performing the logic of multiple couplings of correlation function
Component.Such as, embodiments of the invention include computer program, this computer journey
Sequence product includes the computer program visibly realized on a machine-readable medium, this computer journey
Sequence comprises the program code being configured to realize method described above.
In the context of the disclosure, machine readable media can be to comprise or store for or have
Any tangible medium about the program of instruction execution system, device or equipment.Machine readable is situated between
Matter can be machine-readable signal medium or machinable medium.Machine readable media is permissible
Include but not limited to electronics, magnetic, optics, electromagnetism, infrared or semiconductor system,
Device or equipment, or the combination of its any appropriate.The more detailed example of machinable medium
Including with the electrical connection of one or more wire, portable computer diskette, hard disk, with
Machine memory access device (RAM), read only memory (ROM), erasable programmable is read-only deposits
Reservoir (EPROM or flash memory), light storage device, magnetic storage apparatus, or its any appropriate
Combination.
Can compile with one or more for realizing the computer program code of the method for the present invention
Cheng Yuyan writes.These computer program codes can be supplied to general purpose computer, dedicated computing
Machine or the processor of other programmable data processing meanss so that program code is by computer
Or the when of the execution of other programmable data processing meanss, cause in flow chart and/or block diagram
Function/the operation of regulation is carried out.Program code can the most on computers, part calculate
On machine, as independent software kit, part the most on computers and part the most on the remote computer or
Perform on remote computer or server completely.
Although it addition, operation is depicted with particular order, but this and should not be construed and require this
Generic operation with the particular order illustrated or completes with sequential order, or performs the behaviour of all diagrams
Make to obtain expected result.In some cases, multitask or parallel processing can be useful.
Similarly, contain some specific implementation detail although discussed above, but this should not explain
For limiting any invention or the scope of claim, and should be interpreted that can be for specific invention
The description of specific embodiment.In this specification described in the context of separate embodiment
Some feature can also combined implementation in single embodiment.On the contrary, in single embodiment
Various features described in context can also be discretely in multiple embodiments or in any appropriate
Sub-portfolio in implement.
The various amendments of example embodiment, change for the aforementioned present invention will looked into together with accompanying drawing
When seeing described above, those skilled in the technology concerned are become obvious.Any and all amendment
Unrestriced and the present invention example embodiment scope will be still fallen within.Additionally, aforementioned specification and
There is the benefit inspired in accompanying drawing, relates to the technology people of the technical field of embodiments of the invention
Member will appreciate that other embodiments of the present invention herein illustrated.
It will be appreciated that embodiments of the invention are not limited to disclosed specific embodiment, and revise
All should be contained in scope of the appended claims with other embodiments.Although being used here spy
Fixed term, but they only use in the sense that describing general, and be not limited to
Purpose.
Claims (28)
1. the method selected for mixed model, described method includes:
Candidate family is determined based on training sample set;
Based on described candidate family first initializes hidden variable and update in hidden variable at least
Individual, generate the set of the second initialization hidden variable, wherein said candidate family is according to described first
Initialize hidden variable and generate, and described renewal hidden variable represents what described candidate family exported
Sample packet result;And
Collection based on described second initialization hidden variable is incompatible determines object module.
Method the most according to claim 1, wherein determines candidate based on training sample set
Model includes:
The first initialization hidden variable is determined based on described training sample set;And
Initialize hidden variable according to described first and generate candidate family.
Method the most according to claim 1, wherein determines candidate based on training sample set
Model includes:
Determine that multiple first initializes hidden variable based on described training sample set;
Initialize hidden variable based on the plurality of first and generate multiple initial models;And
According to the performance of the plurality of initial model, select in the plurality of initial model at least
One as candidate family.
Method the most according to claim 1, wherein at the beginning of based on described candidate family first
At least one in beginningization hidden variable and renewal hidden variable, generates the collection of the second initialization hidden variable
Conjunction includes:
Based on described candidate family first initializes hidden variable, generates the 3rd initialization hidden variable
Set;
Renewal hidden variable based on described candidate family, generates the set of the 4th initialization hidden variable;
And
The collection of hidden variable is initialized based on the described 3rd set initializing hidden variable and the described 4th
Close, determine the described second set initializing hidden variable.
Method the most according to claim 4, wherein at the beginning of based on described candidate family first
Beginningization hidden variable, the set generating the 3rd initialization hidden variable includes:
By initializing hidden variable to the first of described candidate family and adding with reference to hidden variable
Power summation, generates the 3rd initialization hidden variable in the described 3rd set initializing hidden variable.
Method the most according to claim 4, wherein renewal based on described candidate family is hidden
Variable, the set generating the 4th initialization hidden variable includes:
Multiple sample groups from the renewal hidden variable of described candidate family select a sample group;
It is multiple subgroup by selected sample components;And
The 4th initialization in the set of the 4th initialization hidden variable is determined based on the plurality of subgroup
Hidden variable.
Method the most according to claim 6, wherein from the hidden change of renewal of described candidate family
Multiple sample groups in amount select a sample group include:
Determine the number of samples of each sample group in the plurality of sample group;And
The sample group that number of samples is maximum is selected from the plurality of sample group.
Method the most according to claim 1, wherein initializes hidden variable based on described second
Collection incompatible determine that object module includes:
Each second initialized by training described second in the set of hidden variable initializes hidden change
Amount generates mid-module;And
Performance based on described mid-module determines described object module.
Method the most according to claim 1, wherein initializes hidden variable based on described second
Collection incompatible determine that object module includes:
Utilizing predetermined statistical model, collection based on described second initialization hidden variable is incompatible to be determined
The set of object initialization hidden variable;
By the hidden change of each object initialization in the set of the described object initialization hidden variable of training
Amount generates mid-module;And
Performance based on described mid-module determines described object module.
Method the most according to claim 9, wherein utilizes predetermined statistical model, based on
Described second set initializing hidden variable determines that the set of object initialization hidden variable includes:
The set of the hidden variable collection to described object initialization hidden variable is initialized according to described second
Conjunction initializes;And
The following one or many that operates of execution:
Utilize object initialization hidden variable in the set of described object initialization hidden variable with
And the performance of the temporary pattern corresponding with described object initialization hidden variable trains described
Predetermined statistical model;And
Hidden change is initialized from described second based on trained described predetermined statistical model
The set of amount selects the second initialization hidden variable, to update described object initialization hidden variable
Set.
11. methods according to claim 10, wherein initialize hidden change according to described second
The set of amount carries out initialization to the set of described object initialization hidden variable and includes:
At least one second hidden change of initialization is selected from the described second set initializing hidden variable
Amount, as the object initialization hidden variable in the set of described object initialization hidden variable.
12. methods according to claim 10, wherein based on trained described predetermined
Statistical model from described second initialize hidden variable set select second initialization hidden variable,
Include with the set of the described object initialization hidden variable of renewal:
Trained described predetermined statistical model is utilized to predict hidden with described second initialization
In the set of variable second initializes the performance of the corresponding temporary pattern of hidden variable;And
According to the performance predicted, select in the described second set initializing hidden variable at least
One second initializes hidden variable and updates the set of described object initialization hidden variable.
13. methods according to claim 12, wherein utilize trained described predetermined
Statistical model predict with described second initialize hidden variable set in second initialization hidden
The performance of the temporary pattern that variable is corresponding includes:
Utilize trained described predetermined statistical model to calculate and initialize hidden change with described second
Measure average and the variance of the performance of corresponding temporary pattern;And
Determine corresponding with described second initialization hidden variable with described variance according to described average
The performance prediction value of temporary pattern.
14. methods according to claim 12, wherein according to the performance predicted, select
At least one second initialization hidden variable in described second set initializing hidden variable updates
The set of described object initialization hidden variable includes:
By described second initialize hidden variable set in corresponding with the optimal performance predicted
The second initialization mesh being defined as in the set of described object initialization hidden variable of hidden variable
Mark initializes hidden variable;And
Remove relative with the optimal performance predicted from the described second set initializing hidden variable
Described second answered initializes hidden variable.
15. 1 kinds of devices selected for mixed model, described device includes:
Candidate family determines unit, is configured to determine candidate family based on training sample set;
Signal generating unit, is configured to based on described candidate family first and initializes hidden variable and more
At least one in new hidden variable, generates the set of the second initialization hidden variable, wherein said time
Modeling type initializes hidden variable according to described first and generates, and described renewal hidden variable represents
The sample packet result of described candidate family output;And
Object module determines unit, is configured to based on the described second set initializing hidden variable
Determine object module.
16. devices according to claim 15, wherein said candidate family determines unit bag
Include:
First initializes hidden variable determines unit, is configured to come really based on described training sample set
Fixed first initialization hidden variable;And
First candidate family signal generating unit, is configured to initialize hidden variable according to described first raw
Become candidate family.
17. devices according to claim 15, wherein said candidate family determines unit bag
Include:
First initializes hidden variable determines unit, is configured to come really based on described training sample set
Fixed multiple first initializes hidden variable;
Initial model determines unit, is configured to initialize hidden variable based on the plurality of first raw
Become multiple initial model;And
Second candidate family signal generating unit, is configured to the performance according to the plurality of initial model,
Select at least one in the plurality of initial model as candidate family.
18. devices according to claim 15, wherein said signal generating unit includes:
3rd hidden variable determines unit, is configured to based on described candidate family first and initializes
Hidden variable, generates the set of the 3rd initialization hidden variable;
4th hidden variable determines unit, is configured to renewal hidden variable based on described candidate family,
Generate the set of the 4th initialization hidden variable;And
Second hidden variable determines unit, is configured to based on the described 3rd collection initializing hidden variable
Close and the set of described 4th initialization hidden variable, determine the described second collection initializing hidden variable
Close.
19. devices according to claim 18, wherein said 3rd hidden variable determines unit
Including:
3rd hidden variable signal generating unit, is configured to described candidate family first initial
Change hidden variable and be weighted summation with reference to hidden variable, generating the described 3rd and initialize hidden variable
Set in the 3rd initialization hidden variable.
20. devices according to claim 18, wherein said 4th hidden variable determines unit
Including:
First selects unit, and be configured to from the renewal hidden variable of described candidate family is multiple
Sample group selects a sample group;
Grouped element, being configured to selected sample components is multiple subgroup;And
Described 4th hidden variable determines unit to be further configured to based on the plurality of subgroup and determines
The 4th initialization hidden variable in 4th set initializing hidden variable.
21. devices according to claim 20, wherein said first selects unit to include:
Number of samples determines unit, is configured to determine that each sample in the plurality of sample group
The number of samples of group;And
Sample group selection unit, is configured to from the plurality of sample group select number of samples
A big sample group.
22. devices according to claim 15, wherein said object module determines unit bag
Include:
Mid-module signal generating unit, is configured to train described second to initialize hidden variable
In set each second initializes hidden variable and generates mid-module;And
Object module signal generating unit, is configured to performance based on described mid-module and determines institute
State object module.
23. devices according to claim 15, wherein said object module determines unit bag
Include:
Object initialization hidden variable determines unit, is configured to, with predetermined statistical model, base
In the described second incompatible set determining object initialization hidden variable of collection initializing hidden variable;
Mid-module signal generating unit, is configured to train described object initialization hidden variable
Each object initialization hidden variable in set generates mid-module;And
Object module signal generating unit, is configured to performance based on described mid-module and determines institute
State object module.
24. devices according to claim 23, wherein said object initialization hidden variable is true
Cell includes:
Initialization unit, is configured to initialize the set of hidden variable to described according to described second
The set of object initialization hidden variable initializes;
Statistical model training unit, is configured to, with the set of described object initialization hidden variable
In object initialization hidden variable and the interim mould corresponding with described object initialization hidden variable
The performance of type trains described predetermined statistical model;And
First updating block, be configured to based on trained described predetermined statistical model from
Described second set initializing hidden variable selects the second initialization hidden variable, described to update
The set of object initialization hidden variable.
25. devices according to claim 24, wherein said initialization unit includes:
Second selects unit, is configured to select from the described second set initializing hidden variable
At least one second initialization hidden variable, in the set as described object initialization hidden variable
Object initialization hidden variable.
26. devices according to claim 24, wherein said first updating block includes:
Performance prediction unit, is configured to, with trained described predetermined statistical model
Predict corresponding with the second initialization hidden variable in the set that described second initializes hidden variable
The performance of temporary pattern;And
Second updating block, is configured to according to the performance predicted, selects described second initial
It is initial that at least one second initialization hidden variable in the set of change hidden variable updates described target
Change the set of hidden variable.
27. devices according to claim 26, wherein said performance prediction unit includes:
First computing unit, is configured to, with trained described predetermined statistical model and counts
Calculate average and the variance of the performance initializing the corresponding temporary pattern of hidden variable with described second;
And
Second computing unit, is configured to determine with described according to described average and described variance
The second performance prediction value initializing the corresponding temporary pattern of hidden variable.
28. devices according to claim 26, wherein said second updating block includes:
Object initialization hidden variable acquiring unit, is configured to initialize hidden variable by described second
Set in corresponding with the optimal performance predicted the second initialization hidden variable be defined as institute
State an object initialization hidden variable in the set of object initialization hidden variable;And
Second initializes hidden variable updating block, is configured to initialize hidden variable from described second
Set in remove corresponding with the optimal performance predicted described second initialization hidden variable.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510147815.7A CN106156856A (en) | 2015-03-31 | 2015-03-31 | The method and apparatus selected for mixed model |
JP2016040848A JP6233432B2 (en) | 2015-03-31 | 2016-03-03 | Method and apparatus for selecting mixed model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510147815.7A CN106156856A (en) | 2015-03-31 | 2015-03-31 | The method and apparatus selected for mixed model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106156856A true CN106156856A (en) | 2016-11-23 |
Family
ID=57322798
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510147815.7A Pending CN106156856A (en) | 2015-03-31 | 2015-03-31 | The method and apparatus selected for mixed model |
Country Status (2)
Country | Link |
---|---|
JP (1) | JP6233432B2 (en) |
CN (1) | CN106156856A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021051505A1 (en) * | 2019-09-18 | 2021-03-25 | 平安科技(深圳)有限公司 | Method, device, and apparatus for performing voiceprint clustering on the basis of sample size, and storage medium |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113191824A (en) * | 2021-05-24 | 2021-07-30 | 北京大米科技有限公司 | Data processing method and device, electronic equipment and readable storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101923662A (en) * | 2009-06-11 | 2010-12-22 | 索尼公司 | Messaging device, information processing method and program |
WO2010035690A9 (en) * | 2008-09-24 | 2011-02-03 | 日本電気株式会社 | Structure training device, structure training system, structure training method, program, and recording medium |
CN101994001A (en) * | 2010-09-14 | 2011-03-30 | 上海海事大学 | Support vector machine algorithm based method for predicting vibration aging effect |
WO2011108632A1 (en) * | 2010-03-03 | 2011-09-09 | 日本電気株式会社 | Model selection device, model selection method and model selection program |
CN103221945A (en) * | 2011-03-18 | 2013-07-24 | 日本电气株式会社 | Multivariate data mixture model estimation device, mixture model estimation method, and mixture model estimation program |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2000163395A (en) * | 1998-11-30 | 2000-06-16 | Nippon Telegr & Teleph Corp <Ntt> | Maximum likelihood estimation for mixed model and recording medium where maximum likelihood estimation program for mixed model is recorded |
JP4254623B2 (en) * | 2004-06-09 | 2009-04-15 | 日本電気株式会社 | Topic analysis method, apparatus thereof, and program |
JP4902378B2 (en) * | 2007-02-06 | 2012-03-21 | 日本放送協会 | Mixed model initial value calculation device and mixed model initial value calculation program |
-
2015
- 2015-03-31 CN CN201510147815.7A patent/CN106156856A/en active Pending
-
2016
- 2016-03-03 JP JP2016040848A patent/JP6233432B2/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2010035690A9 (en) * | 2008-09-24 | 2011-02-03 | 日本電気株式会社 | Structure training device, structure training system, structure training method, program, and recording medium |
CN101923662A (en) * | 2009-06-11 | 2010-12-22 | 索尼公司 | Messaging device, information processing method and program |
WO2011108632A1 (en) * | 2010-03-03 | 2011-09-09 | 日本電気株式会社 | Model selection device, model selection method and model selection program |
CN101994001A (en) * | 2010-09-14 | 2011-03-30 | 上海海事大学 | Support vector machine algorithm based method for predicting vibration aging effect |
CN103221945A (en) * | 2011-03-18 | 2013-07-24 | 日本电气株式会社 | Multivariate data mixture model estimation device, mixture model estimation method, and mixture model estimation program |
Non-Patent Citations (1)
Title |
---|
徐勤丰 等: ""基于隐变量的有限混合模型对有序数据的Bayes聚类分析"", 《应用概率统计》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021051505A1 (en) * | 2019-09-18 | 2021-03-25 | 平安科技(深圳)有限公司 | Method, device, and apparatus for performing voiceprint clustering on the basis of sample size, and storage medium |
Also Published As
Publication number | Publication date |
---|---|
JP6233432B2 (en) | 2017-11-22 |
JP2016194912A (en) | 2016-11-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Vitelli et al. | Probabilistic preference learning with the Mallows rank model | |
Eggensperger et al. | Efficient benchmarking of hyperparameter optimizers via surrogates | |
Raghu et al. | Evaluation of causal structure learning methods on mixed data types | |
Wang et al. | A grey prediction-based evolutionary algorithm for dynamic multiobjective optimization | |
CN103226595B (en) | The clustering method of the high dimensional data of common factor analyzer is mixed based on Bayes | |
CN106845530A (en) | character detection method and device | |
Strelioff et al. | Inferring Markov chains: Bayesian estimation, model comparison, entropy rate, and out-of-class modeling | |
Gadhavi et al. | Student final grade prediction based on linear regression | |
Van Mechelen et al. | Benchmarking in cluster analysis: A white paper | |
CN104102917A (en) | Construction method of domain self-adaptive classifier, construction device for domain self-adaptive classifier, data classification method and data classification device | |
CN106156857A (en) | The method and apparatus selected for mixed model | |
Hamim et al. | Student profile modeling using boosting algorithms | |
CN112420125A (en) | Molecular attribute prediction method and device, intelligent equipment and terminal | |
Fong et al. | Applying a hybrid model of neural network and decision tree classifier for predicting university admission | |
Bu et al. | Cognitive diagnostic model made more practical by genetic algorithm | |
CN106156856A (en) | The method and apparatus selected for mixed model | |
Xiang et al. | Understanding propagation error and its effect on collective classification | |
CN106156077A (en) | The method and apparatus selected for mixed model | |
Shou et al. | Learning path planning algorithm based on kl divergence and d-value matrix similarity | |
CN115081515A (en) | Energy efficiency evaluation model construction method and device, terminal and storage medium | |
Tribelhorn et al. | Parametric Analysis and Validation of Machine Learning in Chaotic Transitions of the Lorenz System | |
Desmarais et al. | –Methods to Refine the Mapping of Items to Skills | |
Tang et al. | Semi-supervised Contrastive Memory Network for Industrial Process Working Condition Monitoring | |
Nikolikj et al. | Sensitivity Analysis of RF+ clust for Leave-one-problem-out Performance Prediction | |
Zhang et al. | Student Achievement Analysis and Prediction Based on Educational Data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161123 |