CN109086825A - A kind of more disaggregated model fusion methods based on model adaptation selection - Google Patents
A kind of more disaggregated model fusion methods based on model adaptation selection Download PDFInfo
- Publication number
- CN109086825A CN109086825A CN201810876135.2A CN201810876135A CN109086825A CN 109086825 A CN109086825 A CN 109086825A CN 201810876135 A CN201810876135 A CN 201810876135A CN 109086825 A CN109086825 A CN 109086825A
- Authority
- CN
- China
- Prior art keywords
- sample
- disaggregated model
- model
- fusion
- base
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The embodiment of the present invention proposes a kind of more disaggregated model fusion methods based on model adaptation selection, it include: to calculate separately base disaggregated model to the classification accuracy of every a kind of sample, it is the disaggregated model of probability value to result output, takes the Top-N tag along sort collection of its classification results；According to each base disaggregated model to the classification accuracy maximum value of every class sample, the dynamic accuracy rate threshold value of Different categories of samples is set, and label is merged to each data set sample setting classification results；It is marked according to the fusion of sample, the base disaggregated model for participating in fusion adaptively selected to each sample, in conjunction with the Top-N tag along sort collection of base disaggregated model, realizes the fusion of base disaggregated model.Technical solution provided in an embodiment of the present invention, it can be respectively two base disaggregated models progress effective integration of probability value and the affiliated class label of sample by result output form, the adaptively selected of base disaggregated model can be realized for each data sample, improve the accuracy rate of disaggregated model after fusion.
Description
[technical field]
Disaggregated model fusion method when more classification problems is solved the present invention relates to machine learning field, more particularly to a kind of
Two base disaggregated model output forms are respectively more disaggregated model fusion methods of probability value and class label.
[background technique]
When merging multiple disaggregated models to solve more classification problems, the learning outcome of comprehensive base disaggregated model is needed
Determine the output of disaggregated model after merging.More classification problems are solved using data mining technology, comprehensive multiple base disaggregated models
Learning outcome realizes that message complementary sense is one of the hot spot studied now further to promote the performance of disaggregated model.At present right
When base disaggregated model result is merged, common Model Fusion method mainly includes most value method, summation, averaging method, ballot
Method, it is contemplated that the decision performance of different classifications model is different, usually can be to difference when merging to base disaggregated model result
Disaggregated model result be arranged weight, thus preferably guarantee fusion after disaggregated model performance promotion.Most be worth method, summation and
Averaging method is when carrying out Model Fusion, it is required that the result output form of base disaggregated model must be that data sample belongs to per the same
The probability value of this classification, and be the base disaggregated model of sample class label for output result and be not suitable for；Utilizing ballot method
When carrying out Model Fusion, to the result output form of base disaggregated model, there is no limit but its decisions that the minority is subordinate to the majority
Mechanism determines that it may be only available for the situation that base disaggregated model number is no less than three.In view of existing Model Fusion method
Limitation to base disaggregated model result output form and base disaggregated model number, the output for two base disaggregated models are respectively
The case where probability value and sample label, existing Model Fusion method can not be effectively applicable in.When disaggregated model output result is
When sample is belonged to probability of all categories, sample can be divided into all kinds of probability sizes by comparing disaggregated model, obtaining should
Disaggregated model is to the top n tag along sort of the sample, hereinafter referred to as Top-N tag along sort collection.
[summary of the invention]
In view of this, the embodiment of the present invention proposes a kind of more disaggregated model fusion sides based on model adaptation selection
Method can solve limitation of the existing blending algorithm to base disaggregated model, and realize for each data sample to base classification mould
Type it is adaptively selected.
A kind of more disaggregated model fusion methods based on model adaptation selection that the embodiment of the present invention proposes, comprising:
Base disaggregated model is calculated separately for the classification accuracy of every a kind of sample, is the classification of probability value to result output
Model takes the Top-N tag along sort collection of its classification results；
According to each base disaggregated model to the classification accuracy maximum value of every class sample, the dynamic accuracy rate of Different categories of samples is set
Threshold value, and label is merged to each data set sample setting classification results；
It is marked according to the fusion of sample, the base disaggregated model for participating in fusion adaptively selected to each sample, in conjunction with base
The Top-N tag along sort collection of disaggregated model realizes the fusion of base disaggregated model.
In upper the method, base disaggregated model is calculated separately to the classification accuracy of every a kind of sample, result is exported
For the disaggregated model of probability value, the method for taking the Top-N tag along sort collection of its classification results are as follows: base disaggregated model number is 2, sample
This classification number is M, and training set sample size is n, and x is sampling feature vectors collection, x=[x1 x2 … xn], xiIndicate i-th of sample
This feature vector, i=1,2 ..., n, the model that k-th of base disaggregated model learns are fk(x)；If the output of disaggregated model k
As a result belong to posterior probability of all categories for sample, then fk(xi) indicate model to the maximum of probability in i-th of sample classification result
It is worth corresponding sample class label, i.e. its Top-1 tag along sort；If the output of disaggregated model k is the prediction class label of sample,
fk(xi) presentation class model k is to the tag along sort of i-th of sample.aK, jIn the output result for indicating k-th of base disaggregated model, point
Class label is the sample predictions accuracy rate of jth class.For calculating base disaggregated model for the classification accuracy of every a kind of sample
Expression formula is
Wherein, I is indicator function, I (j=fk(xi), yi=fk(xi)) indicate to work as j=fk(xi) and yi=fk(xi) set up
Whether duration is 1, is otherwise 0, correctly classified for judgement sample.
In upper the method, according to each base disaggregated model to the classification accuracy maximum value of every class sample, all kinds of samples are set
This dynamic accuracy rate threshold value, and to the method that each data set sample setting classification results merge label are as follows: according to base classification mould
Type classification results calculate Different categories of samples accuracy rate threshold value σ to accuracy rate of all categoriesj:
σj=λ max (a1, j, a2, j), j=1,2 ..., M
Wherein, σjIndicate the accuracy rate threshold value of jth class sample, λ is threshold coefficient, and meets λ ∈ [0,1].
Fusion label δ is respectively set to the sample of all categories in disaggregated model prediction resultj:
Wherein, δjIndicate the fusion label of jth class sample；δj=1 presentation class label is the sample of jth class, disaggregated model
1 accuracy rate is apparently higher than disaggregated model 2, and the classification results after Model Fusion take the classification results of disaggregated model 1；δj=2
It indicates to export the classification results that result takes disaggregated model 2 after the sample pattern for being predicted as jth class merges；δj=0 indicates for pre-
The sample that result label is jth class is surveyed, the classification accuracy between base disaggregated model does not have significant difference, needs mould that base is classified
The result of type further judges to determine the fused output of final mask.
In conjunction with base disaggregated model classification results and its result in Different categories of samples fusion label, and then it is each to obtain data set
The fusion of a sample marks εi:
εi=δj
s.t.f1(xi)=j
Wherein, εiIndicate the fusion label of i-th of sample, δjIndicate the fusion label of jth class sample, f1(xi) divide for base
Class model 1 has i=1 to the classification results of i-th of sample, 2 ..., n, j=1,2 ..., M.
Θ is marked up to the Model Fusion of each sample of data set:
Θ=[ε1 ε2 … εi … εn]
It in upper the method, is marked according to the fusion of sample, the base for participating in fusion point adaptively selected to each sample
Class model, in conjunction with the Top-N tag along sort collection of base disaggregated model, the method for realizing the fusion of base disaggregated model are as follows: Hypothetical classification mould
The output result of type 1 is the probability that sample is divided into each sample class, is belonging respectively to by the sample that disaggregated model 1 exports all kinds of general
Rate size obtains the top n tag along sort set Top-N biggish to posterior probability in the classification results of the sample of disaggregated model 1.It is false
Determine fk j(xi) indicate in the Top-N set of i-th of sample of k-th of disaggregated model output, by the jth that posterior probability is descending
A tag along sort, then have:
Wherein, n is data set sample size.
The model that Hypothetical classification model 2 learns is f2It (x), is f to the tag along sort of each sample in learning outcome2
(xi).Disaggregated model 2 is obtained to the classification results F of data set2(x) are as follows:
F2(x)=[f2(x1)f2(x2)…f2(xi)…f2(xn)]
After Hypothetical classification Model Fusion, the classification results output of sample is p (xi), it can be obtained by the fusion label of sample:
Wherein, f1(xi)=f1 1(xi), feature samples x in 1 result of presentation class modeliGeneric posterior probability is maximum
The corresponding class label of value；pf(xi) indicate fusion label εiOutput label after=0 samples fusion:
s.t.εi=0, i=1,2 ..., n
Wherein, Top-NiIndicate feature samples xiTag along sort collection in corresponding Top-N set:
Top-Ni=[f1 1(xi)f1 2(xi)…f1 j(xi)…f1 N(xi)]
Data set sample can be obtained after the study of base disaggregated model, utilize the Model Fusion side based on Top-N tag along sort collection
After method fusion, the tag along sort P (x) of final output can be determined:
P (x)=[p (x1)p(x2)…p(xi)…p(xn)]
As can be seen from the above technical solutions, the embodiment of the present invention has the advantages that
In the technical solution that the present invention is implemented, base disaggregated model is calculated separately for the classification accuracy of every a kind of sample,
It is the disaggregated model of probability value to result output, takes the Top-N tag along sort collection of its classification results；According to each base disaggregated model pair
The classification accuracy maximum value of every class sample is arranged the dynamic accuracy rate threshold value of Different categories of samples, and each data set sample is arranged
Classification results fusion label；It is marked according to the fusion of sample, the base classification mould for participating in fusion adaptively selected to each sample
Type realizes the fusion of base disaggregated model in conjunction with the Top-N tag along sort collection of base disaggregated model.It provides according to embodiments of the present invention
Result output form can be respectively that two base disaggregated models of probability value and the affiliated class label of sample have by technical solution
Effect fusion, and can realize for each specific data sample to the adaptively selected of base disaggregated model.
[Detailed description of the invention]
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this field
For those of ordinary skill, without any creative labor, it can also be obtained according to these attached drawings other attached
Figure.
Fig. 1 is the process for more disaggregated model fusion methods based on model adaptation selection that the embodiment of the present invention is proposed
Schematic diagram；
Fig. 2 is more disaggregated model fusion method models instruction based on model adaptation selection that the embodiment of the present invention is proposed
Practice stage and test phase algorithm frame flow chart；
Fig. 3 is when solving public data according to the collection more classification tasks of vehicle, and output form is the base disaggregated model of probability value
The curve synoptic diagram of classification accuracy and Top-N tag along sort collection；
When Fig. 4 is that the embodiment of the present invention is applied to public data collection vehicle solution more classification tasks, model point after fusion
The curve synoptic diagram of class accuracy rate and Top-N tag along sort collection；
When Fig. 5 is that the embodiment of the present invention is applied to 10 public data collection solution more classification tasks, disaggregated model after fusion
The contrast schematic diagram of accuracy rate and base disaggregated model accuracy rate.
[specific embodiment]
For a better understanding of the technical solution of the present invention, being retouched in detail to the embodiment of the present invention with reference to the accompanying drawing
It states.
It will be appreciated that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Base
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts it is all its
Its embodiment, shall fall within the protection scope of the present invention.
The embodiment of the present invention provides more disaggregated model fusion methods based on model adaptation selection, referring to FIG. 1, it is
The flow diagram for more disaggregated model fusion methods based on model adaptation selection that the embodiment of the present invention is proposed, such as Fig. 1
It is shown, method includes the following steps:
Step 101, base disaggregated model is calculated separately for the classification accuracy of every a kind of sample, is general for result output
The disaggregated model of rate value takes the Top-N tag along sort collection of its classification results.
Specifically, base disaggregated model number is 2, sample class number is M, and training set sample size is n, and x is sample characteristics
Vector set, x=[x1 x2 … xn], xiIndicate the feature vector of i-th of sample, i=1,2 ..., n, k-th of base disaggregated model
The model learnt is fk(x)；If the output result of disaggregated model k is that sample belongs to posterior probability of all categories, fk(xi) table
Representation model is to the corresponding sample class label of maximum value of probability in i-th of sample classification result, i.e. its Top-1 tag along sort；
If the output of disaggregated model k is the prediction class label of sample, fk(xi) presentation class model k is to the contingency table of i-th of sample
Label；aK, jIn the output result for indicating k-th of base disaggregated model, tag along sort is the sample predictions accuracy rate of jth class；Based on
Calculate base disaggregated model is for the expression formula of the classification accuracy of every a kind of sample
Wherein, I is indicator function, I (j=fk(xi), yi=fk(xi)) indicate to work as j=fk(xi) and yi=fk(xi) set up
Whether duration is 1, is otherwise 0, correctly classified for judgement sample.
Step 102, the dynamic of Different categories of samples is arranged to the classification accuracy maximum value of every class sample according to each base disaggregated model
State accuracy rate threshold value, and label is merged to each data set sample setting classification results.
Specifically, calculating Different categories of samples accuracy rate threshold value to accuracy rate of all categories according to base disaggregated model classification results
σj:
σj=λ max (a1, j, a2, j), j=1,2 ..., M
Wherein, σjIndicate the accuracy rate threshold value of jth class sample, λ is threshold coefficient, and meets λ ∈ [0,1].
Fusion label δ is respectively set to the sample of all categories in disaggregated model prediction resultj:
Wherein, δjIndicate the fusion label of jth class sample；δj=1 presentation class label is the sample of jth class, disaggregated model
1 accuracy rate is apparently higher than disaggregated model 2, and the classification results after Model Fusion take the classification results of disaggregated model 1；δj=2
It indicates to export the classification results that result takes disaggregated model 2 after the sample pattern for being predicted as jth class merges；δj=0 indicates for pre-
The sample that result label is jth class is surveyed, the classification accuracy between base disaggregated model does not have significant difference, needs mould that base is classified
The result of type further judges to determine the fused output of final mask.
In conjunction with base disaggregated model classification results and its result in Different categories of samples fusion label, and then it is each to obtain data set
The fusion of a sample marks εi:
εi=δj
s.t.f1(xi)=j
Wherein, εiIndicate the fusion label of i-th of sample, δjIndicate the fusion label of jth class sample, f1(xi) divide for base
Class model 1 has i=1 to the classification results of i-th of sample, 2 ..., n, j=1,2 ..., M.
Θ is marked up to the Model Fusion of each sample of data set:
Θ=[ε1 ε2 … εi … εn]
Step 103, it is marked according to the fusion of sample, the base classification mould for participating in fusion adaptively selected to each sample
Type realizes the fusion of base disaggregated model in conjunction with the Top-N tag along sort collection of base disaggregated model.
Specifically, the output result of Hypothetical classification model 1 is the probability that sample is divided into each sample class, by disaggregated model 1
The sample of output is belonging respectively to all kinds of probability sizes, and it is larger to posterior probability in the classification results of the sample to obtain disaggregated model 1
Top n tag along sort set Top-N.It is assumed that fk j(xi) indicate k-th of disaggregated model output i-th of sample Top-N collection
In conjunction, by j-th of tag along sort that posterior probability is descending, then have:
Wherein, n is data set sample size.
The model that Hypothetical classification model 2 learns is f2It (x), is f to the tag along sort of each sample in learning outcome2
(xi).Disaggregated model 2 is obtained to the classification results F of data set2(x) are as follows:
F2(x)=[f2(x1)f2(x2)…f2(xi)…f2(xn)]
After Hypothetical classification Model Fusion, it is p (x that the classification results of sample, which have output,i), it can be obtained by the fusion label of sample:
Wherein, f1(xi)=f1 1(xi), feature samples x in 1 result of presentation class modeliGeneric posterior probability is maximum
The corresponding class label of value；pf(xi) indicate fusion label εiOutput label after=0 samples fusion:
s.t.εi=0, i=1,2 ..., n
Wherein, Top-NiIndicate feature samples xiTag along sort collection in corresponding Top-N set:
Top-Ni=[f1 1(xi)f1 2(xi)…f1 j(xi)…f1 N(xi)]
Data set sample can be obtained after the study of base disaggregated model, utilize the Model Fusion side based on Top-N tag along sort collection
After method fusion, the tag along sort P (x) of final output can be determined:
P (x)=[p (x1)p(x2)…p(xi)…p(xn)]
The puppet for more disaggregated model fusion methods based on model adaptation selection that algorithm 1 is proposed by the embodiment of the present invention
Code:
Fig. 2 is the training for more disaggregated model fusion methods based on model adaptation selection that the embodiment of the present invention is proposed
Stage and test phase algorithm frame flow chart, the method for proposition mainly includes training stage and test phase.In the training stage,
Selected base disaggregated model is respectively trained first with training dataset, obtains corresponding base disaggregated model.Calculate separately base
Disaggregated model is the disaggregated model of probability value for result output, takes its classification results to the classification accuracy of every a kind of sample
Top-N tag along sort collection.According to each base disaggregated model to the classification accuracy maximum value of every class sample, Different categories of samples is set
Dynamic accuracy rate threshold value, and label is merged to each data set sample setting classification results.It is marked according to the fusion of sample, to each
The adaptively selected base disaggregated model for participating in fusion of a sample realizes base point in conjunction with the Top-N tag along sort collection of base disaggregated model
Class model fusion.In test phase, the situation that model accuracy rate promotes amplitude maximum after selecting the training stage to merge corresponds to Top-N
The N value of tag along sort base, as the N value of Top-N tag along sort collection selected by test phase model, and base disaggregated model
Each learning parameter is consistent with the training stage.
Fig. 3 is when solving public data according to the collection more classification tasks of vehicle, and output form is the base disaggregated model of probability value
The curve synoptic diagram of classification accuracy and Top-N tag along sort collection, when by concrete application of the embodiment of the present invention, for public data
It is calculated according to the multi-classification algorithm that the base disaggregated model for integrating vehicle the present embodiment is respectively adopted as LightGBM algorithm and random forest
Method, wherein the output of LightGBM algorithm is that sample is belonging respectively to a probability for sample class, and the output of random forest is sample
Specific prediction class label.As seen from Figure 3, the Top-1 tag along sort accuracy rate compared to LightGBM is based on Top-2
It has with the classification accuracy of Top-3 tag along sort collection and is substantially obviously improved, further illustrate this paper embodiment of the present invention institute
Propose the feasibility in theory of more disaggregated model fusion methods based on model adaptation selection.
When Fig. 4 is that the embodiment of the present invention is applied to public data collection vehicle solution more classification tasks, model is pre- after fusion
The curve synoptic diagram for surveying accuracy rate and Top-N tag along sort collection, by schematic diagram this it appears that fused more disaggregated models
Classification accuracy compared to base disaggregated model has and is obviously improved, and when available accuracy rate promotes maximum is corresponding
The N value of Top-N data set may thereby determine that the selected Top-N contingency table of base disaggregated model in the test module of algorithm
Sign the N value of collection.In embodiments of the present invention, when N value is 2, the classification accuracy of model is compared to base disaggregated model after fusion
Amplification it is maximum.
Table is first is that more disaggregated model fusion method applications based on model adaptation selection that the embodiment of the present invention is proposed
In public data collection vehicle solve more classification tasks when, after fusion model with and disaggregated model model comparing result.
Table one
As can be seen from table 1, melted using more disaggregated models based on model adaptation selection that the embodiment of the present invention proposes
After conjunction method is merged base disaggregated model model, compared to the classification results of base disaggregated model, recall rate is improved
3.6%, and its accurate rate is also obviously improved.
When Fig. 5 is that the embodiment of the present invention is applied to 10 public data collection solution more classification tasks, category of model after fusion
Accuracy rate with and base disaggregated model classification accuracy contrast schematic diagram.
In conclusion the embodiment of the present invention has the advantages that
In the technical solution that the present invention is implemented, base disaggregated model is calculated separately to the classification accuracy of every a kind of sample, it is right
In the disaggregated model that result output is probability value, the Top-N tag along sort collection of its classification results is taken；According to each base disaggregated model pair
The classification accuracy maximum value of every class sample is arranged the dynamic accuracy rate threshold value of Different categories of samples, and each data set sample is arranged
Classification results fusion label；It is marked according to the fusion of sample, the base classification mould for participating in fusion adaptively selected to each sample
Type realizes the fusion of base disaggregated model in conjunction with the Top-N tag along sort collection of base disaggregated model.It provides according to embodiments of the present invention
Technical solution, when the base disaggregated model result output form for participating in Model Fusion is respectively probability value and sample class label,
This method can extend to the case where other solve more classification problems using Model Fusion and method, to realize base disaggregated model
Effective integration provides the classification accuracy of model entirety after fusion.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent substitution, improvement and etc. done be should be included within the scope of the present invention.
Claims (4)
1. a kind of more disaggregated model fusion methods based on model adaptation selection, which is characterized in that the method step includes:
(1) base disaggregated model is calculated separately to the classification accuracy of every a kind of sample, is the classification mould of probability value to result output
Type takes the Top-N tag along sort collection of its classification results；
(2) the dynamic accuracy rate of Different categories of samples is arranged to the classification accuracy maximum value of every class sample according to each base disaggregated model
Threshold value, and label is merged to each data set sample setting classification results；
(3) it is marked according to the fusion of sample, the base disaggregated model for participating in fusion adaptively selected to each sample, in conjunction with base point
The Top-N tag along sort collection of class model realizes the fusion of base disaggregated model.
2. the method according to claim 1, wherein calculating separately classification of the base disaggregated model to every a kind of sample
Accuracy rate is the disaggregated model of probability value for result output, takes the Top-N tag along sort collection of its classification results, illustrate
As follows: base disaggregated model number is 2, and sample class number is M, and training set sample size is n, and x is sampling feature vectors collection, x=
[x1 x2 … xn], xiIndicate the feature vector of i-th of sample, i=1,2 ..., n, the mould that k-th of base disaggregated model learns
Type is fk(x)；If the output result of disaggregated model k is that sample belongs to posterior probability of all categories, fk(xi) indicate model to the
The corresponding sample class label of the maximum value of probability in i sample classification result, i.e. its Top-1 tag along sort；If disaggregated model k
Output be sample prediction class label, then fk(xi) presentation class model k is to the tag along sort of i-th of sample；ak,jIndicate kth
In the output result of a base disaggregated model, tag along sort is the sample predictions accuracy rate of jth class；For calculating base disaggregated model pair
It is in the expression formula of the classification accuracy of every a kind of sample
Wherein, I is indicator function, I (j=fk(xi),yi=fk(xi)) indicate to work as j=fk(xi) and yi=fk(xi) set up duration be
1, it is otherwise 0, whether is correctly classified for judgement sample.
3. the method according to claim 1, wherein quasi- according to classification of the base disaggregated model to each sample class
True rate is respectively set the dynamic accuracy rate threshold value of Different categories of samples, and merges label, tool to each data set sample setting classification results
Body is described as follows: according to base disaggregated model classification results to accuracy rate of all categories, calculating Different categories of samples accuracy rate threshold value σj:
σj=λ max (a1,j,a2,j), j=1,2 ..., M
Wherein, σjIndicate the accuracy rate threshold value of jth class sample, λ is threshold coefficient, and meets λ ∈ [0,1]；
Fusion label δ is respectively set to the sample of all categories in disaggregated model prediction resultj:
Wherein, δjIndicate the fusion label of jth class sample；δj=1 presentation class label is the sample of jth class, disaggregated model 1
Accuracy rate is apparently higher than disaggregated model 2, and the classification results after Model Fusion take the classification results of disaggregated model 1；δj=2 indicate
The classification results that result takes disaggregated model 2 are exported after being predicted as the sample pattern fusion of jth class；δj=0 indicates to tie prediction
Fruit label is the sample of jth class, and the classification accuracy between base disaggregated model does not have significant difference, needs base disaggregated model
As a result further judge to determine the fused output of final mask；
In conjunction with base disaggregated model classification results and its result in Different categories of samples fusion label, and then each sample of data set can be obtained
This fusion marks εi:
εi=δj
s.t.f1(xi)=j
Wherein, εiIndicate the fusion label of i-th of sample, δjIndicate the fusion label of jth class sample, f1(xi) it is base disaggregated model
The classification results of 1 pair of i-th of sample, and have i=1,2 ..., n, j=1,2 ..., M；
Θ is marked up to the Model Fusion of each sample of data set:
Θ=[ε1 ε2 … εi … εn]。
4. the method according to claim 1, wherein accurate according to classification of each base disaggregated model to every class sample
Rate maximum value is arranged the dynamic accuracy rate threshold value of Different categories of samples, and merges label, tool to each data set sample setting classification results
Body explanation are as follows: the output result of Hypothetical classification model 1 is the probability that sample is divided into each sample class, is exported by disaggregated model 1
Sample is belonging respectively to all kinds of probability sizes, obtains disaggregated model 1 to the biggish top n of posterior probability in the classification results of the sample
Tag along sort set Top-N；It is assumed thatIt indicates to press in the Top-N set of i-th of sample of k-th of disaggregated model output
J-th descending of tag along sort of posterior probability, then have:
Wherein, n is data set sample size；
The model that Hypothetical classification model 2 learns is f2It (x), is f to the tag along sort of each sample in learning outcome2(xi)；?
Classification results F of the disaggregated model 2 to data set2(x) are as follows:
F2(x)=[f2(x1) f2(x2) … f2(xi) … f2(xn)]
After Hypothetical classification Model Fusion, the classification results output of sample is p (xi), it can be obtained by the fusion label of sample:
Wherein, f1(xi)=f1 1(xi), feature samples x in 1 result of presentation class modeliGeneric posterior probability maximum value institute
Corresponding class label；pf(xi) indicate fusion label εiOutput label after=0 samples fusion:
s.t.εi=0, i=1,2 ..., n
Wherein, Top-NiIndicate feature samples xiTag along sort collection in corresponding Top-N set:
Top-Ni=[f1 1(xi) f1 2(xi) … f1 j(xi) … f1 N(xi)]
Data set sample can be obtained after the study of base disaggregated model, melted using the Model Fusion method based on Top-N tag along sort collection
After conjunction, the tag along sort P (x) of final output can be determined:
P (x)=[p (x1) p(x2) … p(xi) … p(xn)]。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810876135.2A CN109086825A (en) | 2018-08-03 | 2018-08-03 | A kind of more disaggregated model fusion methods based on model adaptation selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810876135.2A CN109086825A (en) | 2018-08-03 | 2018-08-03 | A kind of more disaggregated model fusion methods based on model adaptation selection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109086825A true CN109086825A (en) | 2018-12-25 |
Family
ID=64833444
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810876135.2A Pending CN109086825A (en) | 2018-08-03 | 2018-08-03 | A kind of more disaggregated model fusion methods based on model adaptation selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109086825A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109726764A (en) * | 2018-12-29 | 2019-05-07 | 北京航天数据股份有限公司 | A kind of model selection method, device, equipment and medium |
| CN109829490A (en) * | 2019-01-22 | 2019-05-31 | 上海鹰瞳医疗科技有限公司 | Modification vector searching method, objective classification method and equipment |
| CN110688482A (en) * | 2019-09-12 | 2020-01-14 | 新华三大数据技术有限公司 | Multi-label identification method, training method and device |
| CN111143560A (en) * | 2019-12-26 | 2020-05-12 | 厦门市美亚柏科信息股份有限公司 | Short text classification method, terminal equipment and storage medium |
| CN111275133A (en) * | 2020-02-24 | 2020-06-12 | 腾讯科技（深圳）有限公司 | Fusion method and device of classification models and storage medium |
| CN112232417A (en) * | 2020-10-16 | 2021-01-15 | 北京紫光展锐通信技术有限公司 | Classification method and device, storage medium and terminal |
| CN114897096A (en) * | 2022-06-02 | 2022-08-12 | ***股份有限公司 | Model fusion method, device, equipment and storage medium |
-
2018
- 2018-08-03 CN CN201810876135.2A patent/CN109086825A/en active Pending
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109726764A (en) * | 2018-12-29 | 2019-05-07 | 北京航天数据股份有限公司 | A kind of model selection method, device, equipment and medium |
| CN109829490A (en) * | 2019-01-22 | 2019-05-31 | 上海鹰瞳医疗科技有限公司 | Modification vector searching method, objective classification method and equipment |
| CN109829490B (en) * | 2019-01-22 | 2022-03-22 | 上海鹰瞳医疗科技有限公司 | Correction vector searching method, target classification method and device |
| CN110688482A (en) * | 2019-09-12 | 2020-01-14 | 新华三大数据技术有限公司 | Multi-label identification method, training method and device |
| CN110688482B (en) * | 2019-09-12 | 2022-07-12 | 新华三大数据技术有限公司 | Multi-label identification method, training method and device |
| CN111143560A (en) * | 2019-12-26 | 2020-05-12 | 厦门市美亚柏科信息股份有限公司 | Short text classification method, terminal equipment and storage medium |
| CN111143560B (en) * | 2019-12-26 | 2022-07-01 | 厦门市美亚柏科信息股份有限公司 | Short text classification method, terminal equipment and storage medium |
| CN111275133A (en) * | 2020-02-24 | 2020-06-12 | 腾讯科技（深圳）有限公司 | Fusion method and device of classification models and storage medium |
| CN111275133B (en) * | 2020-02-24 | 2023-09-29 | 腾讯科技（深圳）有限公司 | Fusion method, device and storage medium of classification model |
| CN112232417A (en) * | 2020-10-16 | 2021-01-15 | 北京紫光展锐通信技术有限公司 | Classification method and device, storage medium and terminal |
| CN114897096A (en) * | 2022-06-02 | 2022-08-12 | ***股份有限公司 | Model fusion method, device, equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109086825A (en) | A kind of more disaggregated model fusion methods based on model adaptation selection | |
| CN103632168B (en) | Classifier integration method for machine learning | |
| CN105426826A (en) | Tag noise correction based crowd-sourced tagging data quality improvement method | |
| CN101329734B (en) | License plate character recognition method based on K-L transform and LS-SVM | |
| CN103106265B (en) | Similar image sorting technique and system | |
| CN101604322B (en) | Decision level text automatic classified fusion method | |
| CN109523520A (en) | A kind of chromosome automatic counting method based on deep learning | |
| CN101256631B (en) | Method and apparatus for character recognition | |
| US11335086B2 (en) | Methods and electronic devices for automated waste management | |
| CN105046673B (en) | High spectrum image and visual image fusion sorting technique based on self study | |
| CN109218223A (en) | A kind of robustness net flow assorted method and system based on Active Learning | |
| CN101251896B (en) | Object detecting system and method based on multiple classifiers | |
| CN105095884A (en) | Pedestrian recognition system and pedestrian recognition processing method based on random forest support vector machine | |
| CN106326916A (en) | Multi-scale feature estimation and high-order BING feature-based target detection method | |
| CN106650823A (en) | Probability extreme learning machine integration-based foam nickel surface defect classification method | |
| CN101930549A (en) | Second generation curvelet transform-based static human detection method | |
| CN106445908A (en) | Text identification method and apparatus | |
| CN106326458A (en) | Method for classifying city management cases based on text classification | |
| CN106250913B (en) | A kind of combining classifiers licence plate recognition method based on local canonical correlation analysis | |
| CN103426004A (en) | Vehicle type recognition method based on error correction output code | |
| CN109523514A (en) | To the batch imaging quality assessment method of Inverse Synthetic Aperture Radar ISAR | |
| CN107633264A (en) | Linear common recognition integrated fusion sorting technique based on empty spectrum multiple features limit study | |
| CN101226521A (en) | Machine learning method for ambiguity data object estimation modeling | |
| CN110120035A (en) | A kind of tire X-ray defect detection method differentiating defect grade | |
| CN111814917B (en) | Character wheel image digital identification method with fuzzy state |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181225 |
|
| RJ01 | Rejection of invention patent application after publication |