基于RF_AdaBoost模型的血液种属鉴别算法

doi:10.7503/cjcu20190400

摘要/Abstract

摘要：

针对人类和非人类血液种属鉴别对无损、高效分析方法的需求, 结合随机森林(Random Forest)和AdaBoost(Adaptive Boosting Algorithm)算法, 提出了一种血液种属鉴别方法(RF_AdaBoost). 该方法将RF作为AdaBoost的弱分类器, 以达到提高模型鉴别准确度, 增强模型鲁棒性的目的. 采用RF、支持向量机(SVM)、极限学习机(ELM)、核极限学习机(KELM)、堆栈自编码网络(SAE)、反向传播网络(BP)、主成分分析-线性判别法(PCA-LDA)及偏最小二乘判别分析(PLS-DA)与RF_AdaBoost模型进行对比, 以不同规模血液拉曼光谱数据训练集进行鉴别实验评估其性能. 结果表明, 随着训练样本的增加, RF_AdaBoost鉴别准确度最高达100%, 预测标准偏差趋于0. 与其它模型相比, RF_AdaBoost具有较高的分类准确度及较强的稳定性, 为血液种属的鉴别工作提供了新方法.

关键词: 拉曼光谱, 随机森林, AdaBoost算法, 集成学习, 血液种属鉴别

Abstract: Aim ing at the requirements of non-destructive and high-efficiency analysis methods for ide.pngying human and non-human blood species, a method of blood species ide.pngication based on Random Forest combined Adaptive Boosting Algorithm(RF_AdaBoost) was proposed. This method uses RF as the weak classifier of AdaBoost to improve the ide.pngication accuracy and enhance the robustness of the model. RF, Extreme Learning Machine(ELM), Kernel Extreme Learning Machine(KELM), Stacked Auto-Encoder(SAE), Back Propagation(BP), Principal Component Analysis and Linear Discriminant Analysis(PCA-LDA), Partial Least Squares Discriminant Analysis(PLS-DA) are used to compare with the RF_AdaBoost model, and the training sets of different scales of blood Raman spectroscopy data were used for ide.pngication experiments to evaluate its performance. With the increase of training samples in the experiment, the ide.pngication accuracy of RF_AdaBoost is up to 100%, and the prediction standard deviation tends to zero. The results show that RF_AdaBoost has higher classification accuracy and stronger stability than other models, which provides an effective new method for the ide.pngication of blood species.

Key words: Raman spectroscopy, Random forest, AdaBoost algorithm, Ensemble learning, Boold species ide.pngication

中图分类号:

O652

TrendMD:

魏曼曼,路皓翔,杨辉华. 基于RF_AdaBoost模型的血液种属鉴别算法. 高等学校化学学报, 2020, 41(1): 94.

WEI Manman,LU Haoxiang,YANG Huihua. Research on Boold Species Ide.pngication Algorithm Based on RF_AdaBoost Model ^†. Chem. J. Chinese Universities, 2020, 41(1): 94.

图/表 10

参考文献 19

[1]	Xu Y. J., Fan X. G.,Lin Z. L.,Wang X.,Zuo Y., J. Instrum. Anal., 2017, 36(8), 1047— 1050
	( 许英杰, 范贤光, 林智乐, 王昕, 左勇. 分析测试学报, 2017,36(8), 1047— 1050)
[2]	Zou T. T., Xu Z. L.,Yang J. Y.,Wang H.,Sun Y. M.,Shen Y. D., J. Instrum. Anal., 2018,37(10), 85— 92
	( 邹婷婷, 徐振林, 杨金易, 王弘, 孙远明, 沈玉栋. 分析测试学报, 2018,37(10), 85— 92)
[3]	Ge H., Ye Z., He R ., J. Environ. Sci., 2019,79(5), 74— 80 doi: 10.1016/j.jes.2018.11.001 URL
[4]	Zhang M., Peng W., Lai Z. Q., Wang H. P.,Yuan R. J.,He Q.,Wan X., Spectroscopy and Spectral Analysis, 2019,39(3), 797— 801
	( 张铭, 彭文, 赖珍荃, 王泓鹏, 袁汝俊, 何强, 万雄. 光谱学与光谱分析, 2019,39(3), 797— 801)
[5]	Tomasz B., Zuzanna B ., Analyst, 2019,144(4), 1409— 1419 doi: 10.1039/c8an01494a URL pmid: 30601472
[6]	Kelly V., Lednev I. K ., Anal. Chem., 2009,81(18), 7773— 7777 doi: 10.1021/ac901350a URL pmid: 19670872
[7]	Mclaughlin G., Doty K. C ., Lednev I. K.,. Forensic Science International, 2014,238(5), 91— 95 doi: 10.1016/j.forsciint.2014.02.027 URL
[8]	Mistek E., Lednev I. K ., Anal. Bioanal. Chem., 2015,407(24), 7435— 7442 doi: 10.1007/s00216-015-8909-6 URL pmid: 26195028
[9]	Fujihara J., Fujita Y., Yamamoto T., Nishimoto N., Kimura-Kataoka K., Kurata S., Takinami Y., Yasuda T., Takeshita H ., International Journal of Legal Medicine, 2017,131(2), 319— 322 doi: 10.1007/s00414-016-1396-2 URL pmid: 27262482
[10]	Chen X. L., Wang G. W.,Tao Z. H.,Liu J. X.,Yao H. L.,Huang S. S.,Li Y. Q., Chinese Journal of Lasers, 2009,36(9), 2448— 2454 doi: 10.3788/JCL URL
	( 陈秀丽, 王桂文, 陶站华, 刘军贤, 姚辉璐, 黄庶识, 黎永青. 中国激光, 2009,36(9), 2448— 2454) doi: 10.3788/JCL URL
[11]	Pan J. J., Chen Q. S.,Huang W.,Lin S. J.,Su Y.,Li C.,Feng S. Y.,Chen R., China Oncology, 2011,21(9), 708— 712
	( 潘建基, 陈奇松, 黄伟, 林少俊, 苏颖, 力超, 冯尚源, 陈荣. 中国癌症杂志, 2011,21(9), 708— 712)
[12]	Bai P. L., Wang J.,Yin H. C.,Tian Y. B.,Yao W. M.,Gao J., Journal of Light Scattering, 2016,28(2), 163— 167
	( 白鹏利, 王钧, 尹焕才, 田玉冰, 姚文明, 高静. 光散射学报, 2016,28(2), 163— 167)
[13]	Zheng X. Q., Liao X.,Xu Y.,Hong M. J., Chem. J. Chinese Universities, 2017,38(4), 575— 582
	( 郑祥权, 廖鑫, 徐溢, 洪明坚. 高等学校化学学报, 2017,38(4), 575— 582)
[14]	Helma C., Cramer T., Kramer S., De R. L ., J. Chem. Inf. Comput. Sci., 2018,35(39), 1402— 1411
[15]	Zhang F. D., Liu J.,Wang Z. H., Chem. J. Chinese Universities, 2016,37(10), 1792— 1798
	( 张福东, 刘杰, 王智宏. 高等学校化学学报, 2016,37(10), 1792— 1798)
[16]	Zhang X. C., Zhao Z. X.,Zheng Y.,Li J. Y., IEEE Transactions on Intelligent Transportation Systems , doi: 10.1109/TITS.2018.2888587 doi: 10.1109/TITS.2018.2795381 URL pmid: 29881331
[17]	Kong Q. Q., Ding X. Q.,Gong H. L.,Li Z. R.,Tang X. H.,Yu C. X., J. Instrum. Anal., 2017,36(10), 1203— 1207
	( 孔清清, 丁香乾, 宫会丽, 李忠任, 唐兴宏, 于春霞. 分析测试学报, 2017,36(10), 1203— 1207)
[18]	Breiman L ., Machine Learning, 2001,45(1), 5— 32 doi: 10.1023/A:1010933404324 URL
[19]	Freund Y., Schapire R. E ., . Lecture Notes in Computer Science, 1970,55(1), 23— 37

Number of sample in the training set	100	200	300	400	500	600	700	800	900
Number of negative sample	50	100	150	200	250	300	350	400	450
Ratio of negative sample	0.100	0.201	0.301	0.402	0.502	0.602	0.703	0.803	0.904
Number of positive sample	50	100	150	200	250	300	350	400	450
Ratio of positive sample	0.093	0.187	0.280	0.374	0.467	0.561	0.654	0.748	0.841

Number of sample in the training set	100	200	300	400	500	600	700	800	900
Number of negative sample	50	100	150	200	250	300	350	400	450
Ratio of negative sample	0.100	0.201	0.301	0.402	0.502	0.602	0.703	0.803	0.904
Number of positive sample	50	100	150	200	250	300	350	400	450
Ratio of positive sample	0.093	0.187	0.280	0.374	0.467	0.561	0.654	0.748	0.841

Train/test data	RF_AdaBoost	RF	SVM	ELM	KELM	SAE	BP	PCA-LDA	PLS-DA
100/933	98.95	89.25	96.46	82.64	72.88	70.31	75.99	74.29	93.05
200/833	99.21	93.77	96.88	91.40	79.66	70.18	75.64	76.70	93.93
300/733	99.58	95.89	98.61	95.21	83.41	71.84	76.86	76.21	93.83
400/633	99.90	96.54	98.34	95.64	88.39	70.32	74.04	76.43	93.94
500/533	99.96	97.39	98.48	96.75	88.87	67.04	72.83	77.65	94.19
600/433	100.00	97.45	98.41	96.65	91.92	71.25	77.14	76.37	94.13
700/333	100.00	97.79	99.55	97.81	93.00	69.85	75.29	78.02	94.24
800/233	100.00	98.08	99.18	97.77	93.26	69.27	74.89	76.91	94.28
900/133	100.00	98.33	99.25	98.20	96.39	70.30	82.26	79.47	94.30

Train/test data	RF_AdaBoost	RF	SVM	ELM	KELM	SAE	BP	PCA-LDA	PLS-DA
100/933	98.95	89.25	96.46	82.64	72.88	70.31	75.99	74.29	93.05
200/833	99.21	93.77	96.88	91.40	79.66	70.18	75.64	76.70	93.93
300/733	99.58	95.89	98.61	95.21	83.41	71.84	76.86	76.21	93.83
400/633	99.90	96.54	98.34	95.64	88.39	70.32	74.04	76.43	93.94
500/533	99.96	97.39	98.48	96.75	88.87	67.04	72.83	77.65	94.19
600/433	100.00	97.45	98.41	96.65	91.92	71.25	77.14	76.37	94.13
700/333	100.00	97.79	99.55	97.81	93.00	69.85	75.29	78.02	94.24
800/233	100.00	98.08	99.18	97.77	93.26	69.27	74.89	76.91	94.28
900/133	100.00	98.33	99.25	98.20	96.39	70.30	82.26	79.47	94.30

Train/test data	Training time/s
Train/test data	RF_AdaBoost	RF	SVM	ELM	KELM	SAE	BP	PCA-LDA	PLS-DA
100/933	21.23	2.07	0.03	0.01	0.04	13.88	28.20	1.03	0.03
200/833	18.88	1.71	0.12	0.01	0.04	25.16	29.85	1.04	0.06
300/733	31.69	3.02	0.14	0.01	0.05	36.63	31.89	1.04	0.10
400/633	92.20	9.11	0.23	0.01	0.05	48.04	34.01	1.04	0.13
500/533	61.65	5.94	0.36	0.02	0.06	59.43	36.18	1.04	0.16
600/433	52.73	7.61	0.51	0.03	0.07	71.75	39.01	1.05	0.19
700/333	64.10	9.35	0.53	0.03	0.07	84.56	41.24	1.05	0.23
800/233	77.07	10.80	0.66	0.04	0.08	95.00	44.32	1.06	0.26
900/133	135.97	16.67	0.75	0.05	0.08	107.02	58.71	1.08	0.29