CN111310799A - Active learning algorithm based on historical evaluation result - Google Patents
Active learning algorithm based on historical evaluation result Download PDFInfo
- Publication number
- CN111310799A CN111310799A CN202010063306.7A CN202010063306A CN111310799A CN 111310799 A CN111310799 A CN 111310799A CN 202010063306 A CN202010063306 A CN 202010063306A CN 111310799 A CN111310799 A CN 111310799A
- Authority
- CN
- China
- Prior art keywords
- samples
- unlabeled
- sample
- historical evaluation
- sample set
- 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.)
- Granted
Links
- 238000011156 evaluation Methods 0.000 title claims abstract description 152
- 238000000034 method Methods 0.000 claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 36
- 238000012360 testing method Methods 0.000 claims abstract description 8
- 238000010801 machine learning Methods 0.000 claims abstract description 6
- 230000008569 process Effects 0.000 claims description 26
- 238000012163 sequencing technique Methods 0.000 claims description 13
- 238000002372 labelling Methods 0.000 claims description 8
- 239000000126 substance Substances 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 5
- 239000004576 sand Substances 0.000 claims description 3
- 230000006870 function Effects 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000005070 sampling Methods 0.000 description 3
- 230000003247 decreasing effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012854 evaluation process Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000013604 expression vector Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
Abstract
The invention relates to an active learning algorithm based on historical evaluation results, which is characterized by comprising the following contents: 1) initializing a task model by adopting the labeled sample set; 2) selecting part of unlabeled samples in the unlabeled sample set according to the weighting of the historical evaluation results of the unlabeled samples in the unlabeled sample set and/or the fluctuation of the historical evaluation results and/or the sorting results of the sorting model; 3) marking the selected unmarked samples, adding the marked samples into the marked sample set, and training and updating the task model; 4) and repeating the steps 2) to 3) until the performance of the trained and updated task model on the test set meets the preset requirement, and the method can be widely applied to the field of machine learning.
Description
Technical Field
The invention relates to an active learning algorithm, in particular to an active learning algorithm based on historical evaluation results.
Background
Active learning is a sub-problem in the field of machine learning, and a small part of training data is selected to be labeled in a strategic way, so that a result equivalent to that of training by adopting all data is achieved, the labeling cost of the data is greatly reduced, and an active learning algorithm is acted on the evaluation and selection of unlabeled samples. Existing active learning algorithms are mainly classified into three categories: (1) based on an uncertainty method, the accuracy of the model for sample judgment is used as an evaluation index in the method, and samples which cannot be accurately judged by the current model are considered to be more valuable for model training, so that the samples are preferentially selected for marking; (2) based on a representative method, the method mainly starts from the representativeness of the samples, if the sample A can represent all the samples in a small set S, the model can naturally and correctly judge the samples in the set S after learning the characteristics of the sample A, therefore, the model can preferentially select the sample A to label for training, and most of the existing methods reflect the representativeness of the sample through the similarity between the samples; (3) the method is mainly applied to the batch sampling condition based on the diversity method, so that a batch of selected samples are different from each other and comprise sample characteristics as much as possible, the model training efficiency is greatly improved, and the method is usually matched with the former two methods for use.
The method for training the model by adopting the active learning mode is a cyclic process, firstly a small amount of marked samples are needed for initializing the model, then a batch of samples are sampled by adopting the current model and the active learning algorithm for marking, the samples are used for training and updating the model, and then the next round of sample selection and model updating is carried out. Therefore, when each wheel pair samples the unlabeled sample, a large amount of evaluation data is generated, and each sample is accumulated to have a historical evaluation sequence. However, the above three methods are considered only based on the evaluation result of the current round when selecting samples in each round, and the knowledge and information that can be obtained from the historical evaluation result are ignored, and the data that can be obtained is not fully utilized. And even if researchers pay attention to the historical evaluation result sequence at present, the maximum value in the sequence is simply selected as the current evaluation result, and information in the sequence is not sufficiently mined.
Disclosure of Invention
In view of the above problems, an object of the present invention is to provide an active learning algorithm based on historical evaluation results, which can fully utilize the historical evaluation results to complete sample selection.
In order to achieve the purpose, the invention adopts the following technical scheme: an active learning algorithm based on historical assessment results, comprising the following: 1) initializing a task model by adopting the labeled sample set; 2) selecting an unlabeled sample in the unlabeled sample set according to the weighted sum of the historical evaluation results of the unlabeled sample in the unlabeled sample set; 3) according to the current task model, marking the selected unmarked samples and adding the unmarked samples into the marked sample set to obtain a trained and updated task model; 4) and repeating the steps 2) to 3) until the performance of the trained and updated task model on the labeled sample set meets the preset requirement.
Further, the specific process of step 2) is as follows: setting the weight of each historical evaluation result and the current evaluation result in the unlabeled sample set; determining the weighted sum of the historical evaluation results of the unlabeled samples in the unlabeled sample set according to the set weight and the corresponding historical evaluation result; selecting the unmarked samples in the unmarked sample set according to the weighted sum of the historical evaluation results of the unmarked samples in the unmarked sample set
Wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; w is ajThe weight corresponding to the j iteration evaluation result;
the evaluation result of the sample x in the jth iteration process is obtained; and t is the total iteration number.
Further, the weight in the step 2) is automatically obtained by adopting a machine learning algorithm.
Further, the weight in step 2) is set as:
wherein l is a hyper-parameter for controlling the size of the history window.
An active learning algorithm based on historical assessment results, comprising the following: A) initializing a task model by adopting the labeled sample set; B) selecting an unlabeled sample in the unlabeled sample set according to the volatility of the historical evaluation result of the unlabeled sample in the unlabeled sample set; C) according to the current task model, marking the selected unmarked samples and adding the unmarked samples into the marked sample set to obtain a trained and updated task model; D) and repeating the steps B) to C) until the performance of the trained and updated task model on the labeled sample set meets the preset requirement.
Further, the specific process of the step B) is as follows: determining volatility of historical evaluation results of unlabeled samples in unlabeled sample set
Wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; l is a hyperparameter; i and j are both counting variables;
the evaluation result of the sample x in the jth iteration process is obtained;
volatility of evaluation results from determined history
Selecting the unlabeled samples in the unlabeled sample set
Wherein, wsAnd wfAre all hyper-parameters.
An active learning algorithm based on historical assessment results, comprising the following: a) initializing a task model by adopting the labeled sample set; b) according to the historical evaluation results of the unlabeled samples in the unlabeled sample set, constructing a sequencing model, sequencing all the unlabeled samples in the unlabeled sample set, and selecting a plurality of unlabeled samples which are ranked in the front; c) according to the current task model, marking the selected unmarked samples and adding the unmarked samples into the marked sample set to obtain a trained and updated task model; d) and repeating the steps b) to c) until the performance of the trained and updated task model on the labeled sample set meets the preset requirement.
The specific process of the step b) is that ① builds a sequencing model according to the historical evaluation result of the unlabeled samples in the unlabeled sample set, wherein the sequencing model comprises the samples to be sequenced, sample characteristics and labels of training samples, ② sequences all the unlabeled samples in the unlabeled sample set according to the built sequencing model, and selects a plurality of unlabeled samples which are ranked at the top in the samples to be sequenced.
Further, the specific process of the step ① is that I) a selection strategy is adopted to select a sample to be sorted in the unlabeled sample set, II) the characteristics of the historical evaluation sequence of the sample to be sorted are extracted to be used as the basis of sorting of the sorting model, and III) the label of the sample to be sorted is trained to obtain the sorting model.
Further, the specific process of the step II) is as follows: selecting the historical evaluation result of the sample to be sorted for the latest time l, wherein l is a hyper-parameter; determining the fluctuation of the historical evaluation result in the historical evaluation sequence of the sample to be sorted; determining the trend of the historical evaluation result in the historical evaluation sequence of the sample to be sorted; determining a predicted value of a next evaluation result of a sample to be sequenced; determining probability distribution of the prediction result of the sequencing model; a representation vector of samples to be sorted is determined.
Due to the adoption of the technical scheme, the invention has the following advantages:
1. in the prior art, the selection strategy of the active learning algorithm only considers the evaluation result of the current round when selecting a sample, ignores the historical evaluation sequence, and designs a plurality of active learning algorithms based on the historical evaluation result in order to fully utilize the effective information in the historical evaluation sequence.
2. Since the historical evaluation results are obtained in each iteration, the present invention is easy to implement and does not add additional time complexity.
3. The method is applied to two common natural language processing tasks of named entity identification and text classification, and experimental results prove that the effect and efficiency of selecting samples in active learning can be greatly improved by integrating historical evaluation results into an active learning algorithm, and the method can be widely applied to the field of machine learning.
Drawings
FIG. 1 is an overall architecture diagram of the process of the present invention;
fig. 2 is a variation trend graph of four different historical evaluation sequences in the present invention, in which fig. 2(a) is a relatively stable variation trend graph, fig. 2(b) is a gradually increasing variation trend graph, fig. 2(c) is a gradually decreasing variation trend graph, fig. 2(d) is a fluctuating variation trend graph, the abscissa is iteration number, and the ordinate is evaluation result.
Detailed Description
The present invention is described in detail below with reference to the attached drawings. It is to be understood, however, that the drawings are provided solely for the purposes of promoting an understanding of the invention and that they are not to be construed as limiting the invention.
Definition of active learning problem:
training the model under an active learning framework, and assuming that a smaller labeled sample set L and a larger unlabeled sample set U are available, the goal is to train a model M that meets the requirements. Firstly, a model M1 is preliminarily trained by using a labeled sample set L, then, samples are continuously selected from a non-labeled sample set U for labeling according to a selection strategy through a loop iteration process, and the selected samples are used for labelingAnd updating the model subsequently until the model achieves a satisfactory effect. When selecting samples in each iteration, scoring the samples in each unlabeled sample set U by adopting a scoring function in a current model and a selection strategy, sequencing all the samples according to the scores, and selecting the sample with the highest score for labeling. Then, in the t-th iteration, the scoring function in the selection strategy calculates the score of a sample x as phi t (x), and the scores in the previous t iterations are accumulated to obtain a historical evaluation sequence
The existing active learning algorithm generally evaluates unlabeled samples according to the score calculated at the current time directly through a scoring function, and the invention provides a brand-new selection strategy as follows:
wherein x is*Samples that are preferred for the algorithm;
a history evaluation sequence at the t iteration;
is a scoring function of the present invention; s is a selection strategy for active learning;
the evaluation of the sample x during the j-th iteration is performed for the scoring function of the selection strategy S.
The overall framework of the active learning algorithm of the present invention is shown in fig. 1, and fig. 1 vividly shows the active learning algorithm of the present invention. By observing and analyzing the historical evaluation sequences of the sample population, the historical evaluation sequences can be found to approximately show the change trends shown in fig. 2(a) to 2(d), and the four change trends reflect different performances and properties of the sample in the active learning process. As can be seen from the figure, the evaluation results of the four samples in the current round are very close, and the selection strategy in the existing active learning algorithm can regard the samples as equivalent, but the samples do not behave similarly for many times in the historical evaluation process, and it is unreasonable to judge the equivalence of the samples only according to the evaluation results of the last round of iteration. Therefore, the invention provides a brand-new active learning algorithm based on historical evaluation results by combining the information contained in the historical evaluation sequence with a general or selection strategy of the invention.
Example one
As shown in fig. 2, it can be seen that the evaluation results of the several samples in the last iteration are the same, but have quite different performances in the historical evaluation process, the variation curve of the samples similar to fig. 2(a) and fig. 2(c) is evaluated to contain larger information quantity many times in the process of model iterative training, the variation curve of the samples similar to fig. 2(b) only obtains higher evaluation scores in the current evaluation, and the former has larger information quantity and value than the latter obviously for the training of the model.
Based on the above description, the present invention provides an active learning algorithm based on historical evaluation results, comprising the following steps:
1) and initializing the task model by adopting the labeled sample set.
2) The method comprises the following steps of evaluating samples by adopting a current task model, and selecting part of unlabeled samples in an unlabeled sample set according to the weighted sum of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set, wherein the specific steps are as follows:
2.1) setting the weight of each historical evaluation result and the current evaluation result in the unlabeled sample set.
2.2) determining the weighted sum of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set according to the set weight and the corresponding historical evaluation result.
2.3) weighted sum of several historical evaluation results based on unlabeled samples in the unlabeled sample set
Selecting part of unlabeled samples in the unlabeled sample set:
wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; w is ajThe weight corresponding to the j iteration evaluation result;
the evaluation result of the sample x in the jth iteration process is obtained; and t is the current iteration number.
3) And marking the selected unmarked sample, adding the marked sample into the marked sample set, training and updating to obtain a task model M'.
4) And repeating the steps 2) to 3) until the expression Eval (M ') of the trained and updated task model M' on the test set (a labeled sample set) meets the preset requirement, wherein the expression Eval (M ') of the trained and updated task model M' on the test set can be tested in a mode of accuracy and the like, and the specific process is not repeated herein.
In the step 2.1), a machine learning algorithm may be adopted to automatically obtain the weight corresponding to each round of evaluation result.
In the above step 2.1), the weight may be set according to the following rule, that the closer to the historical evaluation result of the current round of iteration, the more reference information is provided for the sample selection of the round, and it is assumed that the importance is exponentially decreased from back to front, so the weight corresponding to each round of evaluation result is set as:
wherein l is a hyper-parameter for controlling the size of the history window. Because the model is continuously updated in an iteration mode, if the evaluation result which is too far away from the iteration is taken into account, the selection of the sample can be interfered, and therefore, a history window is set, and only the evaluation result of the last round in the selection of the sample can participate in the calculation.
Example two
Uncertainty is one of the most common evaluation indexes in active learning, and a sampling method based on uncertainty tends to select samples of which the model cannot accurately predict or of which the confidence coefficient of the predicted result is low. At the t-th iteration, each unlabeled sample has a corresponding historical evaluation sequence, and each historical evaluation result is calculated based on a specific uncertainty sampling method S. From the viewpoint of sequence fluctuation, it can be found that there are two typical sequences as shown in fig. 2(a) and fig. 2(d), wherein the sample with the variation curve similar to fig. 2(a) shows relatively stable performance, and the sample with the variation curve similar to fig. 2(d) has larger fluctuation. With the iterative update of the model, the sample with low uncertainty and relatively stable change of the historical evaluation sequence is easy to learn and judge by the model. However, samples with high uncertainty and obvious fluctuation of historical evaluation sequences in the updating process are likely to be located near the decision hyperplane of the model, so that the prediction result is unstable, and such samples are more critical for the model training and the decision hyperplane determination. Therefore, the volatility of the sample history evaluation sequence can also be used as an important index of the evaluation uncertainty.
Therefore, the active learning algorithm based on the history evaluation result of the present embodiment is basically the same as that of the first embodiment: step 1) initializing a task model by adopting a labeled sample set; step 2) selecting unmarked samples in the unmarked sample set; step 3) marking the selected unmarked samples and adding the marked samples into the marked sample set, training and updating to obtain a task model M'; and 4) repeating the steps 2) to 3) until the expression Eval (M ') of the trained and updated task model M' on the test set meets the preset requirement. The difference is that in step 2) of this embodiment, the current task model is used to evaluate the samples, and according to the weighted sum of several historical evaluation results of the unlabeled samples in the unlabeled sample set, a part of the unlabeled samples in the unlabeled sample set is selected, instead of: adopting a current task model to evaluate samples, and selecting part of unlabeled samples in the unlabeled sample set according to the volatility of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set, wherein the specific process comprises the following steps:
2.1) determining the volatility of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set:
the embodiment adopts the volatility of the historical evaluation sequence of the unlabeled sample set to measure the uncertainty of the samples in the unlabeled sample set, and the volatility of the historical evaluation sequence
Measured by sequence variance:
wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; l is a hyper-parameter used for controlling the size of the historical sequence window; j and j are both count variables;
is the evaluation result of the sample x in the j iteration process.
2.2) volatility according to several historical evaluation results determined
Selecting part of unlabeled samples in the unlabeled sample set
Wherein, wsAnd wfThe evaluation parameters are hyper-parameters and are used for controlling the influence of the current evaluation result and the historical evaluation sequence fluctuation on sample selection.
EXAMPLE III
The first embodiment and the second embodiment are two heuristic active learning algorithms that use historical evaluation results of unlabeled samples in an unlabeled sample set, and use information in a historical evaluation sequence to select the unlabeled samples in the unlabeled sample set, but there may be much uncovered information. The embodiment further provides an active learning algorithm based on learning from a history evaluation sequence, a process of selecting samples by active learning is regarded as a ranking problem, a ranking model is trained by using the history evaluation sequence as training data to rank all unlabeled samples in each iteration, and the unlabeled samples ranked in the top are selected, so that the active learning algorithm based on learning from the history evaluation sequence of the embodiment is basically the same as the contents of the first embodiment and the second embodiment except for the training of the ranking model, and the method comprises the following steps:
1) training a sequencing model on a plurality of unlabeled data according to the historical evaluation result of the samples, wherein the training of the sequencing model requires the samples to be sequenced, the sample characteristics and the labels of the training samples, and the unlabeled data adopted by the training sequencing model and the unlabeled samples in the unlabeled sample set to be selected are not the same batch of samples, and specifically comprises the following steps:
1.1) adopting a simple selection strategy (such as a selection strategy based on cross entropy, a selection strategy based on minimum confidence coefficient and the like), selecting partial samples in the unmarked sample set as samples to be sorted:
generally, in each iteration, the active learning algorithm needs to evaluate all unlabeled samples and select the sample with the highest score for labeling. However, the unlabeled sample set is typically large, which results in a large ordering space and is prone to large training errors. In order to reduce the sorting space, a smaller unlabeled sample candidate set needs to be selected according to some simple selection strategies, and then the samples in the unlabeled sample candidate set are sorted by adopting a sorting model and the sample with the best performance is selected for labeling.
1.2) extracting the characteristics of the historical evaluation sequence of the samples to be sorted as the basis of sorting of the sorting model, wherein the characteristics comprise the historical evaluation result of the last time l of the samples to be sorted, the volatility of the historical evaluation sequence, the trend of the historical evaluation sequence, the predicted value of the next evaluation result, the probability distribution of the model prediction result and the expression vector of the samples:
a) and selecting the historical evaluation result of the sample to be sorted for the latest time l as a characteristic, wherein l is a hyper-parameter.
b) Determining volatility of historical evaluation results in a historical evaluation sequence of samples to be ranked
c) Determining the trend of the historical evaluation result in the historical evaluation sequence of the sample to be sorted by adopting a Mann-Kendall method (Mann-Kendall test method)
The trend of the historical evaluation result in the historical evaluation sequence reflects the change of the performance of the sample along with the update of the task model in the active learning process, for example, the historical evaluation sequence is calculated by a selection strategy based on cross entropy, and the trend is increased to make the sample become more uncertain along with the update of the task model.
d) Determining the predicted value of the next evaluation result of the samples to be sorted by adopting an LSTM (Long Short-Term Memory) algorithm
The historical evaluation sequence is accumulated by the evaluation result in each iteration, and can be regarded as a time sequence to some extent. Based on the characteristics of the time series, the next evaluation result can be predicted and used for guiding the selection of the sample. The invention adopts the LSTM algorithm as a prediction algorithm of the time series.
e) Extracting probability distribution of task model sample prediction results
Many conventional evaluation methods for selecting strategies calculate a sample score based on a prediction result of a task model on a sample, and therefore, the probability distribution of the prediction result is directly used as a part of sample characteristics in the invention.
f) Determining a representation vector of samples to be ordered
In addition to the above features, the features of the samples to be sorted themselves are also important, and therefore, the present invention also takes the representation vector of the samples to be sorted themselves as a part of the features. For example: and if the sample is a sentence, adopting a Doc2Vec model or obtaining a representation vector of the sample from a full connection layer of the task model.
1.3) according to the effectiveness degree of the samples to the training of the task model, labeling the samples to be sorted, and training to obtain a sorting model:
firstly testing the expression of the current task model M on the test set as Eval (M), then labeling the sample x and adding the sample x into the labeled sample set, training and updating to obtain a task model M ', and measuring the effectiveness of the sample x by adopting score (x) -Eval (M') -Eval (M). In addition, considering that only one sample is added for updating the task model, the absolute scores of all unlabeled samples are divided into a plurality of levels at certain intervals (which can be set according to actual conditions), and the levels are used as labels of training samples, for example: there are five samples with scores of 0.01,0.015, 0.02, 0.008, and 0.025, which are divided into 3 levels 1[0.008], 2[0.01,0.015], and 3[0.02,0.025], and these three levels are labeled with the five samples.
2) And initializing the task model by adopting the labeled sample set.
3) And sequencing all unlabeled samples in the unlabeled sample set by adopting the trained sequencing model, and selecting partial unlabeled samples in the unlabeled sample set, namely selecting a plurality of unlabeled samples with the top rank in the samples to be sequenced, wherein the number of the selected unlabeled samples can be set according to the case condition.
4) And marking the selected unmarked sample, adding the marked sample into the marked sample set, training and updating to obtain a task model M'.
5) And repeating the steps 3) to 4) until the expression Eval (M ') of the trained and updated task model M' on the test set meets the preset requirement.
The above embodiments are only used for illustrating the present invention, and the structure, connection mode, manufacturing process, etc. of the components may be changed, and all equivalent changes and modifications performed on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.
Claims (10)
1. An active learning algorithm based on historical evaluation results, comprising:
1) initializing a task model by adopting the labeled sample set;
2) evaluating the samples by adopting a current task model, and selecting part of unlabeled samples in the unlabeled sample set according to the weighted sum of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set;
3) marking the selected unmarked samples, adding the marked samples into the marked sample set, and training and updating the task model;
4) and repeating the steps 2) to 3) until the performance of the trained and updated task model on the training set meets the preset requirement.
2. The active learning algorithm based on historical evaluation results as claimed in claim 1, wherein the specific process of sample selection in step 2) is:
setting the weight of each historical evaluation result and the current evaluation result in the unlabeled sample set;
determining a weighted sum of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set according to the set weight and the corresponding historical evaluation result;
weighted sum of several historical evaluation results based on unlabeled samples in the unlabeled sample set
Selecting part of unlabeled samples in the unlabeled sample set:
wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; w is ajThe weight corresponding to the j iteration evaluation result;
the evaluation result of the sample x in the jth iteration process is obtained; and t is the current iteration number.
3. The active learning algorithm based on historical evaluation results as claimed in claim 2, wherein the weights in step 2) are set as:
wherein l is a hyper-parameter used for controlling the size of the history window;
or, the weight in the step 2) is automatically obtained by adopting a machine learning algorithm.
4. An active learning algorithm based on historical evaluation results, comprising:
A) initializing a task model by adopting the labeled sample set;
B) evaluating the samples by adopting a current task model, and selecting part of unlabeled samples in the unlabeled sample set according to the volatility of a plurality of historical evaluation results of the unlabeled samples in the unlabeled sample set;
C) marking the selected unmarked samples, adding the marked samples into the marked sample set, training and updating to obtain a task model;
D) and repeating the steps B) to C) until the performance of the trained and updated task model on the test set meets the preset requirement.
5. The active learning algorithm based on historical evaluation results as claimed in claim 4, wherein the specific process of sample selection in step B) is:
determining volatility of a plurality of historical evaluation results of unlabeled samples in a set of unlabeled samples
Wherein the content of the first and second substances,
a historical evaluation sequence of the sample x at the t-th iteration; l is a hyperparameter; i and j are both counting variables;
the evaluation result of the sample x in the jth iteration process is obtained;
volatility of evaluation results from a number of determined histories
Selecting part of unlabeled samples in the unlabeled sample set
Wherein, wsAnd wfAre all hyper-parameters.
6. An active learning algorithm based on learning from a historical evaluation sequence, comprising:
a) training a sequencing model on a plurality of unlabeled data according to the historical evaluation result of the sample;
b) initializing a task model by adopting the labeled sample set;
c) sorting all unlabeled samples in the unlabeled sample set by adopting the trained sorting model, and selecting part of the unlabeled samples in the unlabeled sample set;
d) marking the selected unmarked samples, adding the marked samples into the marked sample set, training and updating to obtain a task model;
e) and repeating the steps c) to d) until the performance of the trained and updated task model on the training set meets the preset requirement.
7. The active learning algorithm based on learning from a historical evaluation sequence as claimed in claim 6, wherein the specific process of step a) is:
I) selecting a part of samples in the unlabeled sample set as samples to be sorted by adopting a selection strategy;
II) extracting the characteristics of the historical evaluation sequence of the sample to be sorted, and using the characteristics as the basis of sorting of a sorting model;
and III) labeling the samples to be sorted according to the effectiveness degree of the samples to the task model training, and training to obtain a sorting model.
8. The active learning algorithm based on learning from a historical evaluation sequence as claimed in claim 7, wherein the specific process of step II) is:
selecting the historical evaluation result of the sample to be sorted for the latest time l, wherein l is a hyper-parameter;
determining the fluctuation of the historical evaluation result in the historical evaluation sequence of the sample to be sorted;
determining the trend of the historical evaluation result in the historical evaluation sequence of the sample to be sorted;
determining a predicted value of a next evaluation result of a sample to be sequenced;
extracting probability distribution of a task model sample prediction result;
a representation vector of samples to be sorted is determined.
9. The active learning algorithm based on learning from a historical evaluation sequence of claim 8, wherein the determining the trend of the historical evaluation results in the historical evaluation sequence of the samples to be ranked adopts a mankennedel method;
and determining the predicted value of the next evaluation result of the sample to be sorted by adopting an LSTM algorithm.
10. The active learning algorithm based on learning from a historical evaluation sequence as claimed in claim 7, wherein the specific process of step III) is:
and dividing the absolute scores of all unlabeled samples into a plurality of layers according to intervals, and taking the layers as labels of the training samples.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010063306.7A CN111310799B (en) | 2020-01-20 | Active learning method based on historical evaluation result |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010063306.7A CN111310799B (en) | 2020-01-20 | Active learning method based on historical evaluation result |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111310799A true CN111310799A (en) | 2020-06-19 |
| CN111310799B CN111310799B (en) | 2024-04-26 |
Family
ID=
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112766390A (en) * | 2021-01-26 | 2021-05-07 | 上海明略人工智能(集团)有限公司 | Method, device and equipment for determining training sample |
| CN112785585A (en) * | 2021-02-03 | 2021-05-11 | 腾讯科技（深圳）有限公司 | Active learning-based training method and device for image video quality evaluation model |
| CN112818755A (en) * | 2021-01-13 | 2021-05-18 | 华中科技大学 | Gait recognition method based on active learning |
| CN113361278A (en) * | 2021-06-21 | 2021-09-07 | 中国人民解放军国防科技大学 | Small sample named entity identification method based on data enhancement and active learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103530321A (en) * | 2013-09-18 | 2014-01-22 | 上海交通大学 | Sequencing system based on machine learning |
| CN107808661A (en) * | 2017-10-23 | 2018-03-16 | 中央民族大学 | A kind of Tibetan voice corpus labeling method and system based on collaborative batch Active Learning |
| US20180285771A1 (en) * | 2017-03-31 | 2018-10-04 | Drvision Technologies Llc | Efficient machine learning method |
| CN109697289A (en) * | 2018-12-28 | 2019-04-30 | 北京工业大学 | It is a kind of improved for naming the Active Learning Method of Entity recognition |
| CN110298415A (en) * | 2019-08-20 | 2019-10-01 | 视睿（杭州）信息科技有限公司 | A kind of training method of semi-supervised learning, system and computer readable storage medium |
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103530321A (en) * | 2013-09-18 | 2014-01-22 | 上海交通大学 | Sequencing system based on machine learning |
| US20180285771A1 (en) * | 2017-03-31 | 2018-10-04 | Drvision Technologies Llc | Efficient machine learning method |
| CN107808661A (en) * | 2017-10-23 | 2018-03-16 | 中央民族大学 | A kind of Tibetan voice corpus labeling method and system based on collaborative batch Active Learning |
| CN109697289A (en) * | 2018-12-28 | 2019-04-30 | 北京工业大学 | It is a kind of improved for naming the Active Learning Method of Entity recognition |
| CN110298415A (en) * | 2019-08-20 | 2019-10-01 | 视睿（杭州）信息科技有限公司 | A kind of training method of semi-supervised learning, system and computer readable storage medium |
Non-Patent Citations (3)
| Title |
|---|
| 赵悦;穆志纯;李霞丽;潘秀琴;: "一种基于EM和分类损失的半监督主动DBN学习算法", 小型微型计算机系统, no. 04, 30 April 2007 (2007-04-30) * |
| 赵悦;穆志纯;潘秀琴;李霞丽;: "一种基于半监督主动学习的动态贝叶斯网络算法", 信息与控制, no. 02, 30 April 2007 (2007-04-30) * |
| 赵悦;穆志纯;董洁;付冬梅;何伟;: "基于QBC主动学习方法建立电信客户信用风险等级评估模型", 北京科技大学学报, no. 04, 30 April 2007 (2007-04-30) * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112818755A (en) * | 2021-01-13 | 2021-05-18 | 华中科技大学 | Gait recognition method based on active learning |
| CN112818755B (en) * | 2021-01-13 | 2022-05-06 | 华中科技大学 | Gait recognition method based on active learning |
| CN112766390A (en) * | 2021-01-26 | 2021-05-07 | 上海明略人工智能(集团)有限公司 | Method, device and equipment for determining training sample |
| CN112785585A (en) * | 2021-02-03 | 2021-05-11 | 腾讯科技（深圳）有限公司 | Active learning-based training method and device for image video quality evaluation model |
| CN112785585B (en) * | 2021-02-03 | 2023-07-28 | 腾讯科技（深圳）有限公司 | Training method and device for image video quality evaluation model based on active learning |
| CN113361278A (en) * | 2021-06-21 | 2021-09-07 | 中国人民解放军国防科技大学 | Small sample named entity identification method based on data enhancement and active learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109164343B (en) | Transformer fault diagnosis method based on characteristic information quantization and weighted KNN | |
| CN110928764B (en) | Automated evaluation method for crowdsourcing test report of mobile application and computer storage medium | |
| CN108921604B (en) | Advertisement click rate prediction method based on cost-sensitive classifier integration | |
| CN111343147B (en) | Network attack detection device and method based on deep learning | |
| CN115115389A (en) | Express customer loss prediction method based on value subdivision and integrated prediction | |
| CN116720620A (en) | Grain storage ventilation temperature prediction method based on IPSO algorithm optimization CNN-BiGRU-Attention network model | |
| CN108268979A (en) | A kind of cut deal qualitative forecasting method based on evolution fuzzy association rules | |
| JP5684084B2 (en) | Misclassification detection apparatus, method, and program | |
| CN109145685A (en) | Fruits and vegetables EO-1 hyperion quality detecting method based on integrated study | |
| CN108681742B (en) | Analysis method for analyzing sensitivity of driver driving behavior to vehicle energy consumption | |
| CN111948561B (en) | Battery life prediction method based on actual measurement big data and artificial intelligence learning algorithm | |
| CN113282747A (en) | Text classification method based on automatic machine learning algorithm selection | |
| CN117371608A (en) | Pig house multi-point temperature and humidity prediction method and system based on deep learning | |
| CN116629716A (en) | Intelligent interaction system work efficiency analysis method | |
| CN115201394B (en) | Multi-component transformer oil chromatography online monitoring method and related device | |
| CN111310799B (en) | Active learning method based on historical evaluation result | |
| CN111310799A (en) | Active learning algorithm based on historical evaluation result | |
| CN114637620B (en) | Database system abnormal classification prediction method based on SVM algorithm | |
| CN116304941A (en) | Ocean data quality control method and device based on multi-model combination | |
| CN113496255B (en) | Power distribution network mixed observation point distribution method based on deep learning and decision tree driving | |
| CN116010884A (en) | Fault diagnosis method of SSA-LightGBM oil-immersed transformer based on principal component analysis | |
| CN113657726B (en) | Personnel risk analysis method based on random forest | |
| CN111026661B (en) | Comprehensive testing method and system for software usability | |
| CN110956528B (en) | Recommendation method and system for e-commerce platform | |
| CN112347162A (en) | Multivariate time sequence data rule mining method based on online learning |
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 | ||
| GR01 | Patent grant |