ENSEMBLE LEARNING PREDICTING METHOD AND SYSTEM

Abstract

An ensemble learning prediction method includes: establishing a plurality of base predictors based on a plurality of training data; initializing a plurality of sample weights of a plurality of sample data and initializing a processing set; in each iteration round, based on the sample data and the sample weights, establishing a plurality of predictor weighting functions of the predictors in the processing set and predicting each of the sample data by each of the predictors in the processing set for identifying a prediction result; evaluating the predictor weighting functions, and selecting a respective target predictor weighting function from the predictor weighting functions established in each iteration round and selecting a target predictor from the predictors in the processing set to update the processing set and to update the sample weights of the sample data.

Description

CROSS-REFERENCE TO RELATED ART

This application claims the benefit of Taiwan application Serial No. 107139955, filed Nov. 9, 2018, the disclosure of which is incorporated by reference herein in its entirety.

TECHNICAL FIELD

The disclosure relates in general to an ensemble learning predicting method and system thereof.

BACKGROUND

In manufacturing industry or other industries, predicting future events by analyzing history data is important. In intelligent manufacturing factory, the machine will record a lot of manufacturing parameters, such as temperature, press, gas flow etc. The manufacturing parameters may be used as sample data to predict manufacturing quality or whether the machine is failed or not.

Ensemble learning is one of supervised learning in machine learning. In ensemble learning, the prediction result is obtained by dynamically selecting the highest-confidence predictor (or hypothesis) among a plurality of predictors, wherein each predictor is assigned by respective weighting.

FIG. 1 (prior art) shows an example of an ensemble learning system. As shown in FIG. 1, the sample data x is input into each of the base predictors h₁(x), h₂(x) and h₃(x). Weighting of the base predictors h₁(x), h₂(x) and h₃(x) are w₁, w₂ and w₃, respectively. If the weighting is fixed (i.e. the weighting does not vary according to the sample data x), then the prediction result y is expressed as: y=w₁h₁(x)+w₂h₂(x)+w₃h₃(x). If the weighting is dynamic (i.e. the weighting is expressed as: w_i=g_i(x), and w_i varies according to the sample data x), then the prediction result y is expressed as: y=g₁(x)h₁(x)+g₂(x)h₂(x)+g₃(x)h₃(x). In ensemble learning, the prediction result is better if the weighting is dynamic other than fixed. However, if the sample has complicated distribution, the confidence prediction is not easily accuracy and thus it is difficult to effectively train the dynamic weighting.

Thus, even if the sample space is complicated, effective training of the dynamic weighting is one important consideration in designing the ensemble learning test system.

SUMMARY

According to one embodiment, provided is an ensemble learning prediction method for an electronic device, the method including: (a) establishing a plurality of base predictors based on a plurality of training data from a signal source; (b) initializing a plurality of sample weights of a plurality of sample data and initializing a processing set; (c) in a first iteration round, based on the plurality of sample data and the plurality of sample weights, establishing a plurality of predictor weighting functions of the predictors in the processing set and predicting each of the plurality of sample data by each of the predictors in the processing set for identifying a prediction result; (d) evaluating the plurality of predictor weighting functions, and selecting a target predictor weighting function from the plurality of predictor weighting functions established in the first iteration round and selecting a target predictor from the plurality of predictors in the processing set based on an evaluation result to update the processing set and to update the plurality of sample weights of the plurality of sample data; and (e) performing a next iteration round to repeat the above steps (c) and (d) until all iteration rounds are performed to select a plurality of target predictor weighting functions and a plurality of target predictors to combine into an ensemble predictor, wherein the ensemble predictor includes the plurality of target predictor weighting functions and the plurality of target predictors; and a prediction result of the ensemble predictor is shown in a display.

According to another embodiment, provided is an ensemble learning prediction system including: a base predictor training model for establishing a plurality of base predictors based on a plurality of training data from a signal source; a predictor weighting function training module, for initializing a plurality of sample weights of a plurality of sample data and initializing a processing set, in a first iteration round, based on the plurality of sample data and the plurality of sample weights, establishing a plurality of predictor weighting functions of the predictors in the processing set and predicting each of the plurality of sample data by each of the predictors in the processing set for identifying a prediction result; an evaluation module for evaluating the plurality of predictor weighting functions, and selecting a target predictor weighting function from the plurality of predictor weighting functions established in the first iteration round and selecting a target predictor from the plurality of predictors in the processing set based on an evaluation result; and a sample weight adjustment module for updating the processing set and for updating the plurality of sample weights of the plurality of sample data; wherein a next iteration round is performed to repeat the above operations until all iteration rounds are performed to select a plurality of target predictor weighting functions and a plurality of target predictors to combine into an ensemble predictor, wherein the ensemble predictor includes the plurality of target predictor weighting functions and the plurality of target predictors; and a prediction result of the ensemble predictor is shown in a display.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 (prior art) shows an example of an ensemble learning system.

FIG. 2 shows a flow chart of an ensemble learning predicting method according to one embodiment of the application.

FIG. 3 shows a functional diagram of an ensemble learning predicting system according to one embodiment of the application.

In the following detailed description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the disclosed embodiments. It will be apparent, however, that one or more embodiments may be practiced without these specific details. In other instances, well-known structures and devices are schematically shown in order to simplify the drawing.

DESCRIPTION OF THE EMBODIMENTS

Technical terms of the disclosure are based on general definition in the technical field of the disclosure. If the disclosure describes or explains one or some terms, definition of the terms is based on the description or explanation of the disclosure. Each of the disclosed embodiments has one or more technical features. In possible implementation, one skilled person in the art would selectively implement part or all technical features of any embodiment of the disclosure or selectively combine part or all technical features of the embodiments of the disclosure.

FIG. 2 shows a flow chart of an ensemble learning predicting method according to one embodiment of the application, which is applicable in an electronic device having a processing circuit. As shown in FIG. 2, in step 210 (the training phase), the training data D_train is used in establishing base predictors h₁, h₂, . . . h_N (N being a positive integer). Each of the base predictors h_i (i=1˜N) may come from different algorithms, or different hyperparameter or different samples. The application is not limited by this.

In step 220 (the training phase), each sample weighting of each sample in the validation data D_valid is set (initialized) and a processing set H is set (initialized). Each sample weighting k_j^(t=1) of each sample x_j (j=1˜n, n meaning the sample number of the validation data) in the validation data D_valid is initialized (t=1, t meaning the iteration round number). For example, each sample weighting k_j^(t=1) of each sample x_j (j=1˜n) is initialized as 1 (i.e. k_j^(t=1)=1). The initialized processing set H is expressed as: H={h₁, h₂, . . . h_N}. The predictors included in the processing set H are the unselected predictors. As for the principle of selecting the predictors is described as follows. In an embodiment of the application, the training data D_train and the validation data D_valid are both the sample data. For example but not limited by, there are 1000 sample data, 800 of the 1000 sample data are as the training data D_train while the remaining 200 sample data of the 1000 sample data are as the validation data D_validOf course, the application is not limited by this.

In step 230 (the training phase), in the first iteration round, the respective predictor weighting functions of the predictors in the processing set H is established based on all sample data and each sample weighting of all sample data in the validation data D_valid.

After the respective predictor weighting functions of the predictors are established, the predictors h_i in the processing set H predict each sample x_j in the validation data D_valid, respectively. Whether the prediction results are true or false is determined and the prediction results are recorded.

For example, the predictor h₁ predicts each sample x_j in the validation data D_valid. Whether the prediction results predicted by the predictor h₁ are true or false is determined and the prediction results predicted by the predictor h₁ are recorded. Then, the predictor h₂ predicts each sample x_j in the validation data D_valid. Whether the prediction results predicted by the predictor h₂ are true or false is determined and the prediction results predicted by the predictor h₂ are recorded. The steps are repeated until the predictor h_N of the processing set H predicts each sample x_j in the validation data D_valid. Whether the prediction results predicted by the predictor h_N are true or false is determined and the prediction results predicted by the predictor h_N are recorded.

For example, in the first iteration round (t=1), the prediction results from the predictor h₁ are as follows:

	the prediction	the prediction
	results from the	results from the
	predictor h1	predictor h1 are
	are true	false
f1(1) having high	R1	R3
output value
f1(1) having low	R2	R4
output value

Whether f₁⁽¹⁾ is high or low is determined based on whether f₁⁽¹⁾ is higher than a predetermined threshold or not (in an embodiment, f₁⁽¹⁾ varies according to the sample data x). The set R1 means the set of the sample data which are correctly predicted by the predictor h₁ when f₁⁽¹⁾ has a high output value. The set R2 means the set of the sample data which are correctly predicted by the predictor h₁ when f₁⁽¹⁾ has a low output value. The set R3 means the set of the sample data which are not correctly predicted by the predictor h₁ when f₁⁽¹⁾ has a high output value. The set R4 means the set of the sample data which are not correctly predicted by the predictor h₁ when f₁⁽¹⁾ has a low output value. In an embodiment of the application, f_i^(t) means the predicting weighting function of the predictor h_i established in the t-th iteration round.

In the step 240 (the training phase), the established predictor weighting functions are evaluated. Based on the evaluation result, a target predictor weighting function is selected among the predictor weighting functions established in the current iteration round, a target predictor is selected among the processing set H and the processing set H is updated (for example, the selected target predictor is removed from the processing set H).

Details of the step 240 are as follows. In each iteration round, when each predictor weighting function of each predictor is evaluated, the confidence score s_i is evaluated as follows. Of course, the application is not limited by this and other possible embodiment of the application may use other different confidence formula which is still within the scope of the application.

$s_i=\frac{\sum _{x_j\in R_1\bigcup R_4}k_j^{\left(t\right)}}{\sum _jk_j^{\left(t\right)}}$

k_j^(t) means the sample weighting of the sample data x_i which is used in training the predictor weighting function in t-th iteration round. Among all predictor weighting functions, the predictor weighting function having highest confidence score is selected. If the weighting function f_i has the highest confidence score, in the t-th iteration round, it is set h_(t)=h_i and g_(t)=f_i^(t) and the predictor h_i is removed from the processing set H. For example, in the current iteration round, the weighting function f₂ has the highest confidence score, then the predictor h₂ and the weighting function f₂ are selected and the predictor h₂ is removed from the processing set H. h_(t) and g_(t) mean the target predictor and the target predictor weighting function selected in the t-th iteration round, respectively. In an embodiment of the application, the confidence score of the predictor weighting function means whether the prediction result of the predictor is consistent with the output value of the predictor weighting function or not. Higher the confidence score, more consistence between the prediction result of the predictor and the output value of the predictor weighting function. On the contrary, lower the confidence score, less consistence between the prediction result of the predictor and the output value of the predictor weighting function.

In step 250 (the training phase), each sample weight of each sample of the validation data D_valid is updated and the details will be described later. If all iteration rounds are not completed yet (step 260), then the flow returns to the step 230 for the next iteration round. On the contrary, if all iteration rounds are completed, then all target predictor weighting functions are established and trained and thus the training phase is over. The ensemble predictor is obtained (step 270).

Based on the established base predictors and the respective predictor weighting functions, the ensemble predictor is obtained. In the test phase, the test data x is input into the ensemble predictor for outputting the prediction result y (step 280). In an embodiment of the application, the prediction result may be shown on a display (not shown). Further, in an embodiment of the application, the test data (used in the test phase), the training data D_train (used in the training phase) and the validation data D_valid (used in the training phase) are generated by the same signal source, the signal source being for example but not limited by, a sensor in an intelligent manufacturing factory for sensing manufacturing parameters (the temperature, the humidity and the pressure etc.).

In the t-th iteration round, the sample weighting k_j^(t+1) of each sample x_j is updated as follows, wherein c>1 and α₃>α₄>α₂>α₁. The updated sample weighting may be used in training the predictor weighting functions in the (t+1)-th iteration round.

	xj is correctly predicted	xj is not correctly
	by h(t)	predicted by h(t)
g(t)(xj) is low	kj(t+1) = kj(t)cα1	kj(t+1) = kj(t)cα3
g(t)(xj) is high	kj(t+1) = kj(t)cα2	kj(t+1) = kj(t)cα4

That is, in updating the sample weighting of each sample x_j, if in the t-th iteration round, the sample x_j is correctly predicted by the (selected) predictor h_(t) and the predictor weighting function g_(t)(x_j) of the (selected) predictor h_(t) is high, the sample weighting of each sample x_j is updated as: k_j^(t+1)=k_j^(t)c^α1 (i.e. the sample weighting is lowered or decreased). If in the t-th iteration round, the sample x_j is correctly predicted by the (selected) predictor h_(t) and the predictor weighting function g_(t)(x_j) of the (selected) predictor h_(t) is low, the sample weighting of each sample x_j is updated as: k_j^(t+1)=k_j^(t)c^α2 (i.e. the sample weighting is lowered or decreased). If in the t-th iteration round, the sample x_j is not correctly predicted by the (selected) predictor h_(t) and the predictor weighting function g_(t)(x_j) of the (selected) predictor h_(t) is high, the sample weighting of each sample x_j is updated as: k_j^(t+1)=k_j^(t)c^α3 (i.e. the sample weighting is increased). If in the t-th iteration round, the sample x_j is not correctly predicted by the (selected) predictor h_(t) and the predictor weighting function g_(t)(x_j) of the (selected) predictor h_(t) is low, the sample weighting of each sample x_j is updated as: k_j^(t+1)=k_j^(t)c^α4 (i.e. the sample weighting is increased).

In an embodiment of the application, the sample which is not correctly predicted by the (selected) predictor h_(t) will have a higher sample weighting in the next iteration round; and on the contrary, the sample which is correctly predicted by the (selected) predictor h_(t) will have a lower sample weighting in the next iteration round.

Thus, in an embodiment of the application, updating the sample weighting of the sample is based on evaluating the consistence between the predictor weighting functions and the prediction results of the selected predictor. The sample weighting of the sample is decreased when a high consistence between the predictor weighting functions and the prediction results of the selected predictor. The sample weighting of the sample is increased when a low consistence between the predictor weighting functions and the prediction results of the selected predictor. Further, in an embodiment of the application, updating (adjusting) the sample weighting is based on “whether the prediction results are correct or not” and “whether the prediction results of the predictor and the output value of the predictor weighting function are consistent or not”.

Further, when the predicting results of the selected predictor are correct and the selected predictor is assigned by a high predictor weighting function, it is defined as the consistence is high. When the predicting results of the selected predictor are wrong and the selected predictor is assigned by a low predictor weighting function, it is defined as the consistence is high. When the predicting results of the selected predictor is correct and the selected predictor is assigned by a low predictor weighting function, it is defined as the consistence is low. When the predicting results of the selected predictor is wrong and the selected predictor is assigned by a high predictor weighting function, it is defined as the consistence is low.

An example is used in describing an embodiment of the application. For simplicity, the situations in the example are: three iteration rounds (t=3), the validation data D_valid having 5 sample data x1-x5, c=1.5 and (α₁, α₂, α₃, α₄)=(−1, 0, 2, 1). The processing set H is initialized as H={h₁, h₂, h₃}. After initialization, the sample weights of the sample x1-x5 are as: (k₁⁽¹⁾, k₂⁽¹⁾, k₃⁽¹⁾, k₄⁽¹⁾, k₅⁽¹⁾)=(1, 1, 1, 1, 1).

In the first iteration round, the predictor weighting functions f₁⁽¹⁾, f₂⁽¹⁾, f₃⁽¹⁾ of the predictors h₁, h₂, h₃ are established based on the sample data x1-x5 and the sample weighting k₁⁽¹⁾, k₂⁽¹⁾, k₃⁽¹⁾, k₄⁽¹⁾, k₅⁽¹⁾ of the sample data x1-x5. Details of establishing the weighting functions are not specified here.

The respective confidence scores of each predictor weighting function f₁⁽¹⁾, f₂⁽¹⁾, f₃⁽¹⁾ are calculated in the first iteration round. The prediction result of the predictor h₁ is for example as follows:

	Samples which are	Samples which are not
	correctly predicted	correctly predicted
	by h1	by h1
f1(1) is high	x1, x2	x5
f1(1) is low	x4	x3

The confidence score s1 of the predictor weighting function f₁⁽¹⁾ is as follows:

$s_1=\frac{\sum _{x_j\in R_1\bigcup R_4}k_j^{\left(1\right)}}{\sum _jk_j^{\left(1\right)}}=\frac{k_2^{\left(1\right)}+k_2^{\left(1\right)}+k_3^{\left(1\right)}}{k_1^{\left(1\right)}+k_2^{\left(1\right)}+k_3^{\left(1\right)}+k_4^{\left(1\right)}+k_5^{\left(1\right)}}=\frac{1+1+1}{1+1+1+1}$

The confidence scores of each of the predictor weighting functions f₁⁽¹⁾, f₂⁽¹⁾, f₃⁽¹⁾ are calculated, If the predictor weighting function f₂⁽¹⁾ has the highest confidence score, it means that in the first iteration round, the predictor weighting function f₂⁽¹⁾ is best. In the first iteration round, the predictor h₂ is selected as the target predictor and the predictor weighting function f₂⁽¹⁾ is selected as the target predictor weighting function, that is, h₍₁₎=h₂ and g₍₁₎=f₂⁽¹⁾ (h₍₁₎ and g₍₁₎ are the target predictor and the target predictor weighting function selected in the first iteration round) and the processing set H is updated as: H={h₁, h₃} (i.e. h₂ is removed from H). Based on the results in the first iteration round, the sample weighting of the sample data x1-x5 are updated as: (k₁⁽²⁾, k₂⁽²⁾, k₃⁽²⁾, k₄⁽²⁾, k₅⁽²⁾)=(k₁⁽¹⁾*1.5⁻¹, k₂⁽¹⁾*1.5⁻¹, k₃⁽¹⁾*1.5¹, k₄⁽¹⁾*1.5⁰, k₅⁽¹⁾*1.5²), wherein k₁⁽²⁾, k₂⁽²⁾, k₃⁽²⁾, k₄⁽²⁾, k₅⁽²⁾ are used in the second iteration round.

In the second iteration round, the predictor weighting functions f₁⁽²⁾, f₃⁽²⁾ of the predictors h₁, h₃ are established based on the sample data x1-x5 and the sample weighting k₁⁽²⁾, k₂⁽²⁾, k₃⁽²⁾, k₄⁽²⁾, k₅⁽²⁾ of the sample data x1-x5.

Similarly, the confidence scores s₁ and s₃ of the predictor weighting functions f₁⁽²⁾, f₃⁽²⁾ of the predictors h₁, h₃ in the processing set H are calculated in the second iteration round. The prediction result of the predictor h₃ is for example as follows:

	Samples which are	Samples which are
	correctly predicted	not correctly
	by h3	predicted by h3
f3(2) is high	x4, x5	x3
f3(2) is low	x1	x2

Thus, the confidence score s₃ of the weighting function f₃⁽²⁾ is as follow:

$s_3=\frac{\sum _{x_j\in R_1\bigcup R_4}k_j^{\left(2\right)}}{\sum _jk_j^{\left(2\right)}}=\frac{k_2^{\left(2\right)}+k_4^{\left(2\right)}+k_5^{\left(2\right)}}{k_1^{\left(2\right)}+k_2^{\left(2\right)}+k_3^{\left(2\right)}+k_4^{\left(2\right)}+k_5^{\left(2\right)}}$

The confidence scores of the predictor weighting functions f₁⁽²⁾, f₃⁽²⁾ are calculated, If the predictor weighting function f₃⁽²⁾ has the higher confidence score, it means that in the second iteration round, the predictor weighting function f₃⁽²⁾ is better. In the second iteration round, the predictor h₃ is selected as the target predictor and the predictor weighting function f₃⁽²⁾ is selected as the target predictor weighting function, that is, h₍₂₎=h₃ and g₍₂₎=f₃⁽²⁾ (h₍₂₎ and g₍₂₎ are the target predictor and the target predictor weighting function selected in the second iteration round) and the processing set H is updated as: H={h₁} (i.e. h₃ is removed from H). Based on the results in the second iteration round, the sample weighting of the sample data x1-x5 are updated as: (k₁⁽³⁾, k₂⁽³⁾, k₃⁽³⁾, k₄⁽³⁾, k₅⁽³⁾)=(k₁⁽²⁾*1.5⁰, k₂⁽²⁾*1.5¹, k₃⁽²⁾*1.5², k₄⁽²⁾*1.5⁻¹, k₅⁽²⁾*1.5⁻¹)=(k₁⁽¹⁾*1.5⁻¹*1.5⁰, k₂⁽¹⁾*1.5⁻¹1.5¹, k₃⁽¹⁾*1.5¹*1.5², k₄⁽¹⁾*1.5⁰*1.5⁻¹, k₅⁽¹⁾*1.5²*1.5⁻¹), wherein k₁⁽³⁾, k₂⁽³⁾, k₃⁽³⁾, k₄⁽³⁾, k₅⁽³⁾ are used in the third iteration round.

In the third iteration round, the predictor weighting function f₁⁽³⁾ of the predictor h₁ is established based on the sample data x1-x5 and the sample weighting k₁⁽³⁾, k₂⁽³⁾, k₃⁽³⁾, k₄⁽³⁾, k₅⁽³⁾ of the sample data x1-x5. Thus, in the third iteration round, the predictor h₁ is selected as the target predictor and the predictor weighting function f₁⁽³⁾ is selected as the target predictor weighting function, that is, h₍₃₎=h₁ and g₍₃₎=f₁⁽³⁾ (h₍₃₎ and g₍₃₎ are the target predictor and the target predictor weighting function selected in the third iteration round) and the processing set H is updated as: H={ϕ} (i.e. h₁ is removed from H). By so, the ensemble predictor is established.

In another example, the prediction result y of a test data x_test have two possible results: 1 or −1. After x_test is introduced into h_(i)(x) and g_(i)(x):

(h₍₁₎(x_test),h₍₂₎(x_test),h₍₃₎(x_test))=(1,−1,−1)

(g₍₁₎(x_test),g₍₂₎(x_test),g₍₃₎(x_test)=(0.3,0.7,0.6)

The prediction result is obtained by the following two ensemble implementations, which is not to limit the application:

Implementation 1:

The threshold is set as 0.5. Based on a sequence of g₍₁₎, g₍₂₎, g₍₃₎, the first weighting function whose output value is higher than the threshold is selected. The weighing function g₍₁₎(x_test) has an output value of 0.3, lower than the threshold. Then, the weighing function g₍₂₎(x_test) has an output value of 0.7, higher than the threshold. The weighing function g₍₂₎(x_test) is selected. Thus, the predictor h₍₂₎(x_test) which is corresponding to the selected weighing function g₍₂₎(x_test) has a value of −1, and the ensemble prediction result of the test data x_test is −1 (i.e. y=h₍₂₎(x_test)=−1).

Thus, in the implementation 1, the function output value of each of the predictor weighting functions are calculated. The weighting function which is the first one to have an output value higher than the threshold is assigned by a highest weighting and the other weighting functions are assigned by zero weighting.

Implementation 2:

The g_(i)(x) are normalized and averaged:

$\frac{\sum \left(g_{\left(i\right)}*h_{\left(i\right)}\right)}{\sum g_{\left(i\right)}}=\frac{0.3*1}{0.3+0.7+0.6}+\frac{0.7*\left(-1\right)}{0.3+0.7+0.6}+\frac{0.6*\left(-1\right)}{0.3+0.7+0.6}=-0.625.$

If the normalized-average result is smaller than 0, then −1 is as the ensemble prediction result of the test data x_test (y=−1). That is, in the implementation 2, the function output values of the predictor weighting functions are normalized as the weighting of each predictor.

FIG. 3 shows a functional block of an ensemble learning prediction system 300. The ensemble learning prediction system 300 includes: a base predictor training model 310, a predictor weighting function training module 320, an evaluation module 330 and a sample weighting adjusting module 340. The base predictor training model 310, the predictor weighting function training module 320, the evaluation module 330 and the sample weighting adjusting module 340 may implement the ensemble learning prediction method of the above embodiments of the application. The ensemble learning prediction system 300 may further include a display (not shown) to display the prediction results. The ensemble learning prediction system 300 may be an electronic device (for example but not limited by a server) having information processing functions. Further, The base predictor training model 310, the predictor weighting function training module 320, the evaluation module 330 and the sample weighting adjusting module 340 may be implemented by the hardware circuit (for example CPU) having information processing functions.

In the ensemble learning prediction of embodiments of the application, by considering the versatility and complementary of the predictors (or the prediction model), the weighting functions are established via sample weighting adjustment and sequence-optimized weighting function training mechanism. Thus, an embodiment of the application may effectively find the predictors which are suitable respective sample area to dynamically train the weighting functions and to improve the ensemble prediction effect. In here, “sequence-optimized weighting function training mechanism” means that it is to find respective target weighting function in each iteration round (i.e. to select the target weighting function g_(t) in the t-th iteration round), rather than to select all target weighting functions in a single iteration round.

An embodiment of the application may cause specific effect on computers (via predicting the results by the respective target weighting function selected in each iteration round), i.e. the computers are not just means. Thus, an embodiment of the application is more than just to use computers. An embodiment of the application uses specific pattern rules to improve specific effects of the ensemble learning prediction system.

An embodiment of the application is directed to a specific field, for example but not limited by, a computer prediction system. The conventional computer prediction system is not able to effectively train the predictor weighting functions in complicated sample space. An embodiment of the application selects the target predictor and the target predictor weighting function in each iteration round based on the sequence-optimized weighting function evaluation module and steps, and therefore an embodiment of the application may effectively train the predictor weighting function even in complicated sample space (that is because an embodiment of the application selects the target predictor and the target predictor weighting function in each iteration round, rather than selecting all target predictors and all target predictor weighting functions in a single iteration round).

It will be apparent to those skilled in the art that various modifications and variations can be made to the disclosed embodiments. It is intended that the specification and examples be considered as exemplary only, with a true scope of the disclosure being indicated by the following claims and their equivalents.

Claims

1. An ensemble learning prediction method for an electronic device, the method including: (a) establishing a plurality of base predictors based on a plurality of training data from a signal source;(b) initializing a plurality of sample weights of a plurality of sample data and initializing a processing set;(c) in a first iteration round, based on the plurality of sample data and the plurality of sample weights, establishing a plurality of predictor weighting functions of the predictors in the processing set and predicting each of the plurality of sample data by each of the predictors in the processing set for identifying a prediction result;(d) evaluating the plurality of predictor weighting functions, and selecting a target predictor weighting function from the plurality of predictor weighting functions established in the first iteration round and selecting a target predictor from the plurality of predictors in the processing set based on an evaluation result to update the processing set and to update the plurality of sample weights of the plurality of sample data; and(e) performing a next iteration round to repeat the above steps (c) and (d) until all iteration rounds are performed to select a plurality of target predictor weighting functions and a plurality of target predictors to combine into an ensemble predictor, wherein the ensemble predictor includes the plurality of target predictor weighting functions and the plurality of target predictors; and a prediction result of the ensemble predictor is shown in a display.

2. The method according to claim 1, wherein in initializing: the plurality of sample weights kj(t) of the plurality of sample data xj are initialized, j=1˜n, n meaning a number of the sample data and t meaning an iteration round number;the initialized processing set H is expressed as H={h1, h2, . . . , hN}, h1, h2, . . . , hN mean the plurality of base predictors and the processing set H includes unselected predictors.

3. The method according to claim 2, wherein the step (c) includes: in each of the iteration rounds, evaluating respective scores of each of the predictor weighting functions, the scores representing whether respective prediction result of the predictor is consistent with an output value of the predictor weighting function;the predictor weighting function having the highest score among the plurality of the predictor weighting functions is selected as the target predictor weighting function in a current iteration round, h(t)=hi and g(t)=fi(t), h(t) and g(t) are the target predictor and the target predictor weighting function selected in a t-th iteration round.

4. The method according to claim 1, wherein the step of updating the processing set includes: removing the selected target predictor from the processing set.

5. The method according to claim 3, wherein in updating the sample weight of the sample data xj: if in the t-th iteration round, the sample data xj is correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a high output value, the sample weight of the sample data xj is decreased;if in the t-th iteration round, the sample data xj is correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a low output value, the sample weight of the sample data xj is decreased;if in the t-th iteration round, the sample data xj is not correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a high output value, the sample weight of the sample data xj is increased; andif in the t-th iteration round, the sample data xj is not correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a low output value, the sample weight of the sample data xj is increased.

6. The method according to claim 1, wherein as for the ensemble predictor, respective function output values of the target predictor weighting functions are calculated, a first one among the target predictor weighting functions having a function output value higher than a threshold is assigned by a highest weighting and the other target predictor weighting functions are assigned by zero weighting.

7. The method according to claim 1, wherein as for the ensemble predictor, respective function output values of the target predictor weighting functions are normalized as respective weights of the target predictors.

8. An ensemble learning prediction system including: a base predictor training model for establishing a plurality of base predictors based on a plurality of training data from a signal source;a predictor weighting function training module, for initializing a plurality of sample weights of a plurality of sample data and initializing a processing set, in a first iteration round, based on the plurality of sample data and the plurality of sample weights, establishing a plurality of predictor weighting functions of the predictors in the processing set and predicting each of the plurality of sample data by each of the predictors in the processing set for identifying a prediction result;an evaluation module for evaluating the plurality of predictor weighting functions, and selecting a target predictor weighting function from the plurality of predictor weighting functions established in the first iteration round and selecting a target predictor from the plurality of predictors in the processing set based on an evaluation result; anda sample weight adjustment module for updating the processing set and for updating the plurality of sample weights of the plurality of sample data;wherein a next iteration round is performed to repeat the above operations until all iteration rounds are performed to select a plurality of target predictor weighting functions and a plurality of target predictors to combine into an ensemble predictor, wherein the ensemble predictor includes the plurality of target predictor weighting functions and the plurality of target predictors; and a prediction result of the ensemble predictor is shown in a display.

9. The system according to claim 8, wherein in initializing: the predictor weighting function training module initializes the plurality of sample weights kj(t) of the plurality of sample data xj, j=1˜n, n meaning a number of the sample data and t meaning an iteration round number;the initialized processing set H is expressed as H={h1, h2, . . . , hN}, h1, h2, . . . , hN mean the plurality of base predictors and the processing set H includes unselected predictors.

10. The system according to claim 9, wherein the evaluation module is configured for: in each of the iteration rounds, evaluating respective scores of each of the predictor weighting functions, the scores representing whether respective prediction result of the predictor is consistent with an output value of the predictor weighting function;selecting the predictor weighting function having the highest score among the plurality of the predictor weighting functions as the target predictor weighting function in a current iteration round, h(t)=hi and g(t)=fi(t), h(t) and g(t) are the target predictor and the target predictor weighting function selected in a t-th iteration round.

11. The system according to claim 8, wherein in updating the processing set, the selected target predictor is removed from the processing set.

12. The system according to claim 11, wherein in updating the sample weight of the sample data xj, the sample weight adjustment module is configured for: if in the t-th iteration round, the sample data xj is correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a high output value, decreasing the sample weight of the sample data xj;if in the t-th iteration round, the sample data xj is correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a low output value, decreasing the sample weight of the sample data xj;if in the t-th iteration round, the sample data xj is not correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a high output value, increasing the sample weight of the sample data xj; andif in the t-th iteration round, the sample data xj is not correctly predicted by the target predictor h(t) and the predictor weighting function g(t) of the target predictor h(t) has a low output value, increasing the sample weight of the sample data xj.

13. The system according to claim 8, wherein as for the ensemble predictor, respective function output values of the target predictor weighting functions are calculated, a first one among the target predictor weighting functions having a function output value higher than a threshold is assigned by a highest weighting and the other target predictor weighting functions are assigned by zero weighting.

14. The system according to claim 8, wherein as for the ensemble predictor, respective function output values of the target predictor weighting functions are normalized as respective weights of the target predictors.