一种基于自训练的众包标记噪声纠正算法

杨艺; 蒋良孝; 李超群

doi:10.16383/j.aas.c210051

一种基于自训练的众包标记噪声纠正算法

doi: 10.16383/j.aas.c210051

杨艺^1,,
蒋良孝^{1, 2,},
李超群^3,

1.
中国地质大学(武汉)计算机学院武汉 430074
2.
智能地学信息处理湖北省重点实验室(中国地质大学(武汉))武汉 430074
3.
中国地质大学(武汉)数学与物理学院武汉 430074

基金项目: 国家自然科学基金联合基金(U1711267), 中央高校基本科研业务费专项资金(CUGGC03)资助

详细信息

作者简介:
杨艺：中国地质大学(武汉)计算机学院硕士研究生. 2018年获得中国地质大学(武汉)计算机学院学士学位. 主要研究方向为机器学习与数据挖掘. E-mail: yangyi@cug.edu.cn

蒋良孝：中国地质大学(武汉)计算机学院教授. 2009年获得中国地质大学(武汉)地球探测与信息技术博士学位. 主要研究方向为机器学习与数据挖掘. 本文通信作者. E-mail: ljiang@cug.edu.cn

李超群：中国地质大学(武汉)数学与物理学院副教授. 2012年获得中国地质大学(武汉)地球探测与信息技术博士学位. 主要研究方向为机器学习与数据挖掘. E-mail: chqli@cug.edu.cn

计量
- 文章访问数: 1533
- HTML全文浏览量: 409
- PDF下载量: 207
- 被引次数: 0
出版历程
- 收稿日期: 2021-01-18
- 网络出版日期: 2021-06-20
- 刊出日期: 2023-04-20

A Self-training-based Label Noise Correction Algorithm for Crowdsourcing

YANG Yi^1
,,
JIANG Liang-Xiao^{1, 2
,},
LI Chao-Qun^3
,

1.
School of Computer Science, China University of Geosciences (Wuhan), Wuhan 430074
2.
Hubei Key Laboratory of Intelligent Geo-Information Processing, China University of Geosciences (Wuhan), Wuhan 430074
3.
School of Mathematics and Physics, China University of Geosciences (Wuhan), Wuhan 430074

Funds: Supported by National Natural Science Foundation of China (U1711267) and Fundamental Research Funds for the Central Universities (CUGGC03)

More Information

Author Bio:
YANG Yi　Master student at the School of Computer Science, China University of Geosciences (Wuhan). He received his bachelor degree from China University of Geosciences (Wuhan) in 2018. His research interest covers machine learning and data mining

JIANG Liang-Xiao　Professor at the School of Computer Science, China University of Geosciences (Wuhan). He received his Ph.D. degree in earth prospecting and information technology from China University of Geosciences (Wuhan) in 2009 . His research interest covers machine learning and data mining. Corresponding author of this paper

LI Chao-Qun　Associate professor at the School of Mathematics and Physics, China University of Geosciences (Wuhan). She received her Ph.D. degree in earth prospecting and information technology from China University of Geosciences (Wuhan) in 2012. Her research interest covers machine learning and data mining

摘要

摘要: 针对众包标记经过标记集成后仍然存在噪声的问题, 提出了一种基于自训练的众包标记噪声纠正算法(Self-training-based label noise correction, STLNC). STLNC整体分为3个阶段: 第1阶段利用过滤器将带集成标记的众包数据集分为噪声集和干净集. 第2阶段利用加权密度峰值聚类算法构建数据集中低密度实例指向高密度实例的空间结构关系. 第3阶段首先根据发现的空间结构关系设计噪声实例选择策略; 然后利用在干净集上训练的集成分类器对选择的噪声实例按照设计的实例纠正策略进行纠正, 并将纠正后的实例加入到干净集, 再重新训练集成分类器; 重复实例选择与纠正过程直到噪声集中所有的实例被纠正; 最后用最后一轮训练得到的集成分类器对所有实例进行纠正. 在仿真标准数据集和真实众包数据集上的实验结果表明STLNC比其他5种最先进的噪声纠正算法在噪声比和模型质量两个度量指标上表现更优.
- 众包学习 /
- 自训练 /
- 集成标记 /
- 标记噪声 /
- 噪声纠正
Abstract: In order to solve the problem that a certain level of label noise exists in integrated labels obtained by label integration algorithms, this paper proposes a self-training-based label noise correction (STLNC) algorithm for crowdsourcing. There are three stages in STLNC. At the first stage, STLNC employs a filter to get a clean set and a noisy set. At the second stage, the weighted density peak clustering algorithm is used to construct the spatial structure relationship between low-density instances and high-density instances in the dataset. At the third stage, a noise instance selection strategy is at first designed according to the found spatial structure relationship. Then, these selected noise instances are corrected by the ensemble classifier trained on the clean set according to the designed instance correction strategy, and the corrected instances are added into the clean set and the ensemble classifier is retrained. The process of instance selection and correction is repeated until all noise instances are corrected. Finally, the ensemble classifier trained from the last round is used to correct all the instances. Experimental results on both simulated benchmark datasets and real-world crowdsourced datasets show that STLNC significantly outperforms other five state-of-the-art noise correction algorithms in team of the noise ratio and the model quality.
- Crowdsourcing learning /
- self-training /
- integrated labels /
- label noise /
- noise correction
注释:

1) ¹ http://www.mturk.com² http://www.crowdflower.com³ http://www.clickworker.com

2) http://www.crowdflower.com

3) http://www.clickworker.com

HTML全文

图 1 STLNC算法的框架

Fig. 1 Framework of STLNC

下载: 全尺寸图片幻灯片

图 2 不同T值的STLNC在ionosphere数据集上的噪声比结果

Fig. 2 Noise ratio of STLNC with different T values on ionosphere dataset

下载: 全尺寸图片幻灯片

图 3 Leaves数据集上的噪声比对比结果

Fig. 3 Noise ratio comparisons on Leaves datasets

下载: 全尺寸图片幻灯片

图 4 Leaves数据集上的模型质量结果对比结果

Fig. 4 Model quality comparisons on Leaves datasets

下载: 全尺寸图片幻灯片

图 5 LabelMe数据集上的噪声比对比结果

Fig. 5 Noise ratio comparisons on LabelMe datasets

下载: 全尺寸图片幻灯片

图 6 LabelMe数据集上的模型质量对比结果

Fig. 6 Model quality comparisons on LabelMe datasets

下载: 全尺寸图片幻灯片

图 7 STLNC基于不同过滤器在LabelMe4数据集上的实验结果

Fig. 7 Experimental results of STLNC with different filters on LabelMe4 dataset

下载: 全尺寸图片幻灯片

图 8 STLNC在LabelMe4数据集上的消融实验结果

Fig. 8 Results of STLNC ablation experiment on LabelMe4 dataset

下载: 全尺寸图片幻灯片

表 1 22个仿真标准数据集详细描述

Table 1 Description of 22 simulated benchmark datasets

数据集	#Ins	#Att	#Pos	#Neg
biodeg	1055	41	356	699
breast-cancer	268	9	85	201
breast-w	699	10	241	458
credit-a	690	16	383	307
credit-g	1000	21	300	700
diabetes	768	8	268	500
heart-statlog	270	14	120	32
hepatitis	155	20	123	32
horse-colic	368	22	232	136
ionosphere	351	35	225	126
kr-vs-kp	3196	37	1527	1669
labor	57	16	37	20
mushroom	8124	23	3916	4208
sick	3772	30	231	3541
sonar	208	61	111	97
spambase	4601	57	813	2788
tic-tac-toe	958	10	332	626
vote	435	17	168	267
climate	540	20	494	46
colic	368	22	136	232
monks	432	6	228	204
steel-plates-faults	1941	33	673	1268

下载: 导出CSV

表 2 工人质量0.6时的噪声比对比结果 (%)

Table 2 Noise ratio comparisons with p_j = 0.6 (%)

数据集	MV	PL	STC	CC	AVNC	CENC	STLNC
biodeg	28.25	29.95	28.34	19.53	18.48	21.90	15.83
breast-cancer	27.62	26.92	25.87	31.12	26.57	29.37	24.84
breast-w	28.76	9.01	19.31	10.30	9.16	8.44	7.30
credit-a	26.67	20.00	15.94	18.84	13.04	13.33	12.90
credit-g	26.60	27.40	28.40	26.60	25.30	27.50	26.40
diabetes	26.69	32.29	26.56	26.95	23.70	23.96	22.79
heart-statlog	25.19	19.26	23.70	22.96	24.07	25.93	18.52
hepatitis	30.32	19.35	26.45	20.65	27.74	25.16	30.97
horse-colic	27.72	32.34	17.39	21.20	17.66	14.13	14.13
ionosphere	27.92	16.24	21.65	9.12	10.83	13.39	11.68
kr-vs-kp	27.38	21.96	10.45	19.34	2.19	2.85	2.28
labor	31.58	24.56	24.56	15.79	12.28	31.58	7.02
mushroom	26.71	12.65	6.43	4.30	0.04	0.10	0
sick	27.60	2.60	8.83	10.31	1.78	2.28	3.37
sonar	26.92	31.73	29.33	24.04	25.00	24.52	18.75
spambase	27.02	27.47	19.50	14.78	9.11	10.56	8.06
tic-tac-toe	26.20	34.13	23.07	24.43	22.44	22.23	14.61
vote	25.98	4.60	10.34	11.26	3.91	4.14	4.14
climate	27.41	8.52	27.41	14.07	8.52	8.52	8.52
colic	27.45	22.28	20.92	23.10	14.13	14.95	13.59
monks	26.39	25.00	11.34	21.76	5.32	6.71	2.78
steel-plates-faults	27.51	34.98	9.22	18.70	0	0.10	0.15
平均值	27.45	21.97	19.77	18.60	13.69	15.08	12.21

下载: 导出CSV

表 3 工人质量0.6时的模型质量对比结果(%)

Table 3 Model quality comparisons with p_j = 0.6 (%)

数据集	MV	PL	STC	CC	AVNC	CENC	STLNC
biodeg	71.37	75.91	72.13	78.77	74.34	74.21	78.29
breast-cancer	67.00	71.28	69.98	69.08	70.92	68.13	69.73
breast-w	92.85	92.47	90.68	93.38	94.00	92.85	95.54
credit-a	82.03	84.78	84.49	84.78	83.91	83.04	83.48
credit-g	62.20	68.30	67.40	69.70	67.30	63.90	70.40
diabetes	71.74	70.56	71.50	70.80	71.72	72.12	74.00
heart-statlog	65.56	74.81	67.78	70.00	74.07	69.26	78.15
hepatitis	69.00	77.67	71.33	77.50	72.00	70.17	74.17
horse-colic	77.42	78.11	80.00	80.15	81.62	81.36	82.35
ionosphere	83.48	85.45	86.90	85.19	85.77	84.04	85.76
kr-vs-kp	95.18	95.49	96.62	90.29	96.81	96.59	97.03
labor	70.67	61.83	73.50	68.33	68.33	72.33	76.33
mushroom	99.85	98.56	99.88	99.90	99.90	99.86	99.83
sick	96.74	94.62	96.98	94.48	97.77	97.75	97.11
sonar	58.93	55.57	50.07	58.29	55.00	59.14	58.29
spambase	85.92	88.87	87.44	84.20	89.68	88.98	90.39
tic-tac-toe	77.17	69.74	74.54	71.63	74.63	74.63	78.36
vote	89.89	95.37	94.21	90.59	94.21	93.98	94.21
climate	91.48	91.48	91.48	91.48	91.48	91.48	91.48
colic	79.97	81.09	81.09	82.47	81.09	82.47	81.09
monks	90.75	90.73	93.51	83.35	93.51	93.51	93.28
steel-plates-faults	100.00	89.64	100.00	92.01	100.00	100.00	100.00
平均值	80.87	81.47	81.89	81.20	82.64	82.26	84.06

下载: 导出CSV

表 4 工人质量0.6时的噪声比的威尔科克森测试结果

Table 4 Noise ratio summary of Wilcoxon tests with p_j = 0.6

	MV	PL	STC	CC	AVNC	CENC	STLNC
MV	—	◦	◦	◦	◦	◦	◦
PL		—		◦	◦	◦	◦
STC	•		—		◦	◦	◦
CC	•			—	◦	◦	◦
AVNC	•	•	•	•	—	•	◦
CENC	•	•	•	•	◦	—	◦
STLNC	•	•	•	•	•	•	—

下载: 导出CSV

表 5 工人质量0.6时的模型质量的威尔科克森测试结果

Table 5 Model quality summary of Wilcoxon tests with p_j = 0.6

	MV	PL	STC	CC	AVNC	CENC	TTLNC
MV	—		◦		◦	◦	◦
PL		—					◦
STC			—		◦		◦
CC				—			◦
AVNC	•		•		—	•	◦
CENC	•					—	◦
STLNC	•	•	•	•	•	•	—

下载: 导出CSV

表 6 工人质量[0.55, 0.75]时的噪声比对比结果 (%)

Table 6 Noise ratio comparisons with p_j ∈ [0.55, 0.75] (%)

数据集	MV	PL	STC	CC	AVNC	CENC	STLNC
biodeg	14.22	21.14	16.02	13.84	13.46	13.08	12.89
breast-cancer	16.43	26.22	20.98	19.93	23.43	24.83	24.48
breast-w	20.46	3.72	10.01	4.15	4.15	4.43	3.72
credit-a	18.41	20.58	14.93	13.62	13.77	13.04	12.17
credit-g	17.70	29.60	22.70	22.90	21.60	22.30	24.60
diabetes	20.18	22.66	24.09	22.27	23.44	22.66	23.44
heart-statlog	16.30	20.37	15.19	20.00	16.67	16.67	18.52
hepatitis	12.26	20.65	14.19	14.84	16.77	12.90	12.26
horse-colic	17.66	15.49	13.86	18.75	14.67	14.13	15.22
ionosphere	17.38	18.80	13.68	9.69	11.11	10.83	13.96
kr-vs-kp	17.43	25.19	5.60	11.55	1.31	1.88	2.44
labor	17.54	29.82	17.54	12.28	21.05	21.05	14.04
mushroom	18.07	4.94	4.84	1.67	0.10	0.11	0
sick	13.94	1.78	4.98	3.76	1.46	1.54	2.04
sonar	15.38	37.50	21.63	25.96	19.23	22.60	20.67
spambase	19.32	37.54	15.11	9.04	7.00	7.04	6.67
tic-tac-toe	20.67	27.45	19.31	17.54	15.76	14.41	6.47
vote	22.07	6.90	10.57	8.97	4.37	4.83	4.60
climate	22.96	8.52	22.96	10.74	8.52	8.52	8.52
colic	16.58	19.57	15.49	17.93	15.22	14.40	15.49
monks	17.13	12.73	7.18	23.38	2.78	4.86	2.78
steel-plates-faults	22.46	34.83	7.32	15.92	0.26	0.26	0.21
平均值	17.93	20.27	14.46	14.49	11.64	11.65	11.15

下载: 导出CSV

表 7 工人质量[0.55, 0.75]时的模型质量对比结果 (%)

Table 7 Model quality comparisons with p_j ∈ [0.55, 0.75] (%)

数据集	MV	PL	STC	CC	AVNC	CENC	STLNC
biodeg	74.58	81.87	76.36	79.99	81.59	80.25	81.78
breast-cancer	69.43	71.81	69.50	70.27	71.28	71.64	71.64
breast-w	90.76	94.54	91.10	94.34	93.85	92.40	94.69
credit-a	82.17	85.36	84.78	85.65	85.65	84.64	84.93
credit-g	69.50	69.80	70.50	69.60	71.10	69.00	72.00
diabetes	71.67	74.44	71.15	74.21	73.13	72.87	74.56
heart-statlog	70.00	78.89	76.30	75.93	79.63	78.52	80.37
hepatitis	76.17	79.17	77.33	80.50	76.83	77.83	79.00
horse-colic	82.50	80.35	81.57	82.63	83.51	83.01	82.68
ionosphere	80.62	83.48	82.33	88.03	87.73	86.90	84.07
kr-vs-kp	97.94	95.34	97.66	92.86	98.28	98.00	98.06
labor	78.33	68.17	78.33	64.33	77.17	77.17	84.33
mushroom	99.99	98.52	99.95	99.96	100.00	100.00	99.95
sick	97.72	96.95	97.48	95.47	97.64	97.83	97.14
sonar	63.79	68.93	67.14	70.36	69.29	69.50	70.86
spambase	86.65	88.87	88.07	84.83	90.37	88.96	90.76
tic-tac-toe	77.83	73.00	77.58	76.30	77.89	76.85	80.08
vote	93.35	93.79	94.98	92.68	95.24	95.00	94.77
climate	91.48	91.48	91.48	91.48	91.48	91.48	91.48
colic	83.74	77.90	84.19	79.17	82.53	82.63	81.84
monks	98.37	85.21	98.60	88.16	100.00	100.00	100.00
steel-plates-faults	99.90	100.00	100.00	99.69	100.00	100.00	100.00
平均值	83.48	83.54	84.38	83.47	85.65	85.20	86.14

下载: 导出CSV

表 8 工人质量[0.55, 0.75]时的噪声比的威尔科克森测试结果

Table 8 Noise ratio summary of Wilcoxon tests with p_j ∈ [0.55, 0.75]

	MV	PL	STC	CC	AVNC	CENC	STLNC
MV	—		◦	◦	◦	◦	◦
PL		—	◦	◦	◦	◦	◦
STC		•	—		◦	◦	◦
CC		•		—	◦	◦	◦
AVNC	•	•	•	•	—
CENC	•	•	•	•		—
STLNC	•	•	•	•			—

下载: 导出CSV

表 9 工人质量[0.55, 0.75]时的模型质量的威尔科克森测试结果

Table 9 Model quality summary of Wilcoxon tests with p_j ∈ [0.55, 0.75]

	MV	PL	STC	CC	AVNC	CENC	STLNC
MV	—		◦		◦	◦	◦
PL		—			◦	◦	◦
STC	•		—		◦	◦	◦
CC				—	◦	◦	◦
AVNC	•	•	•	•	—	•
CENC	•		•		◦	—	◦
STLNC	•	•	•	•		•	—

下载: 导出CSV

表 10 8个真实众包数据的详细描述

Table 10 Description of eight real-world crowdsourced datasets

数据集	分类任务	#Instances	#Positives	#Negatives	#Labelers	#Labels
Leaves1	maple/alder	142	96	46	70	1093
Leaves2	maple/tilia	140	96	44	74	1044
Leaves3	alder/eucalyptus	93	46	47	58	407
Leaves4	alder/poplar	89	46	43	54	400
LabelMe1	highway/street	199	89	110	50	395
LabelMe2	highway/forest	227	89	138	54	476
LabelMe3	highway/opencountry	240	89	151	54	375
LabelMe4	highway/insidecity	205	89	116	49	339

下载: 导出CSV

参考文献(27)

[1]	Pollicelli D, Coscarella M, Delrieux C. RoI detection and segmentation algorithms for marine mammals photo-identification. Ecological Informatics, 2020, 56: Article No. 101038
[2]	Wang H, Zhao D, Ma H D. Informative image selection for crowdsourcing-based mobile location recognition. Multimedia Systems, 2019, 25(5): 513−523 doi: 10.1007/s00530-017-0562-9
[3]	Lotfian R, Busso C. Building naturalistic emotionally balanced speech corpus by retrieving emotional speech from existing podcast recordings. IEEE Transactions on Affective Computing, 2019, 10(4): 471−483 doi: 10.1109/TAFFC.2017.2736999
[4]	Sheng V S, Provost F, Ipeirotis P G. Get another label? Improving data quality and data mining using multiple, noisy labelers. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Las Vegas, USA: ACM, 2008. 614−622
[5]	Demartini G, Difallah D E, Cudré-Mauroux P. ZenCrowd: Leveraging probabilistic reasoning and crowdsourcing techniques for large-scale entity linking. In: Proceedings of the 21st International Conference on World Wide Web. Lyon, France: ACM, 2012. 469−478
[6]	Zhang H, Jiang L Z, Xu W Q. Multiple noisy label distribution propagation for crowdsourcing. In: Proceedings of the 28th International Joint Conference on Artificial Intelligence. Macao, China: AAAI Press, 2019. 1473−1479
[7]	Tian T, Zhu J, You Q B. Max-margin majority voting for learning from crowds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(10): 2480−2494 doi: 10.1109/TPAMI.2018.2860987
[8]	Zhong J H, Yang P, Tang K. A quality-sensitive method for learning from crowds. IEEE Transactions on Knowledge and Data Engineering, 2017, 29(12): 2643−2654 doi: 10.1109/TKDE.2017.2738643
[9]	Nicholson B, Sheng V S, Zhang J. Label noise correction and application in crowdsourcing. Expert Systems With Applications, 2016, 66: 149−162 doi: 10.1016/j.eswa.2016.09.003
[10]	Xu W Q, Jiang L X, Li C Q. Resampling-based noise correction for crowdsourcing. Journal of Experimental & Theoretical Artificial Intelligence, 2021, 33(6): 985−999
[11]	Zhang J, Sheng V S, Li T, Wu X D. Improving crowdsourced label quality using noise correction. IEEE Transactions on Neural Networks and Learning Systems, 2018, 29(5): 1675−1688 doi: 10.1109/TNNLS.2017.2677468
[12]	Li C Q, Jiang L X, Xu W Q. Noise correction to improve data and model quality for crowdsourcing. Engineering Applications of Artificial Intelligence, 2019, 82: 184−191 doi: 10.1016/j.engappai.2019.04.004
[13]	Xu W Q, Jiang L X, Li C Q. Improving data and model quality in crowdsourcing using cross-entropy-based noise correction. Information Sciences, 2021, 546: 803−814 doi: 10.1016/j.ins.2020.08.117
[14]	Wu D, Shang M S, Luo X, Xu J, Yan H Y, Deng W H, et al. Self-training semi-supervised classification based on density peaks of data. Neurocomputing, 2018, 275: 180−191 doi: 10.1016/j.neucom.2017.05.072
[15]	Khuri S A, Sayfy A. A laplace variational iteration strategy for the solution of differential equations. Applied Mathematics Letters, 2012, 25(12): 2298−2305 doi: 10.1016/j.aml.2012.06.020
[16]	Hershey J R, Olsen P A. Approximating the kullback leibler divergence between Gaussian mixture models. In: Proceedings of the 32nd IEEE International Conference on Acoustics, Speech and Signal Processing. Honolulu, USA: IEEE, 2007. IV-317−IV-320
[17]	Zhang J, Sheng V S, Nicholson B, Wu X D. CEKA: A tool for mining the wisdom of crowds. The Journal of Machine Learning Research, 2015, 16(1): 2853−2858
[18]	Witten I H, Frank E. Data Mining: Practical Machine Learning Tools and Techniques. (3rd edition). Beijing: China Machine Press, 2005.
[19]	Gamberger D, Lavrac N, Groselj C. Experiments with noise filtering in a medical domain. In: Proceedings of the 16th International Conference on Machine Learning. Bled, Slovenia: ACM, 1999. 143−151
[20]	García S, Herrera F. An extension on “statistical comparisons of classifiers over multiple data sets” for all pairwise comparisons. Journal of Machine Learning Research, 2008, 9(12): 2677−2694
[21]	Jiang L X, Zhang L G, Li C Q, Wu J. A correlation-based feature weighting filter for naive Bayes. IEEE Transactions on Knowledge and Data Engineering, 2019, 31(2): 201−213 doi: 10.1109/TKDE.2018.2836440
[22]	Rodrigues F, Lourenčo M, Ribeiro B, Pereira F C. Learning supervised topic models for classification and regression from crowds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(12): 2409−2422 doi: 10.1109/TPAMI.2017.2648786
[23]	Li F F, Perona P. A Bayesian hierarchical model for learning natural scene categories. In: Proceedings of the 14th IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Diego, USA: IEEE, 2005. 524−531
[24]	McCallum A, Nigam K. A comparison of event models for naive Bayes text classification. In: Proceedings of the AAAI-98 Workshop on Learning for Text Categorization. Palo Alto, USA: AAAI Press, 1998. 41−48
[25]	Rennie J D M, Shih L, Teevan J, Karger D R. Tackling the poor assumptions of naive Bayes text classifiers. In: Proceedings of the 20th International Conference on Machine Learning. Washington, USA: AAAI Press, 2003. 616−623
[26]	Khoshgoftaar T M, Rebours P. Improving software quality prediction by noise filtering techniques. Journal of Computer Science and Technology, 2007, 22(3): 387−396 doi: 10.1007/s11390-007-9054-2
[27]	Brodley C E, Friedl M A. Identifying mislabeled training data. Journal of Artificial Intelligence Research, 1999, 11(1): 131−167