Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
Contents lists available at ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Discrimination between authentic and adulterated liquors by near-infrared spectroscopy and ensemble classification Hui Chen a, Chao Tan a,b,c,⇑, Tong Wu a, Li Wang a, Wanping Zhu a a
Key Lab of Process Analysis and Control, Yibin University, Yibin, Sichuan 644007, PR China Department of Chemistry and Chemical Engineering, Yibin University, Yibin, Sichuan 644007, PR China c Computational Physics Key Laboratory of Sichuan Province, Yibin University, Yibin, Sichuan 644007, PR China b
h i g h l i g h t s
g r a p h i c a l a b s t r a c t
Near-infrared spectroscopy is used to
discriminate liquors. Ensemble support vector machines
serves as a tool. Such a method is fast and objective.
a r t i c l e
i n f o
Article history: Received 9 January 2014 Received in revised form 5 March 2014 Accepted 18 March 2014 Available online 2 April 2014 Keywords: Liquor Authenticity Near-infrared Support vector machines
a b s t r a c t Chinese liquor is one of the famous distilled spirits and counterfeit liquor is becoming a serious problem in the market. Especially, age liquor is facing the crisis of confidence because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. An ideal method for authenticity confirmation of liquors should be non-invasive, non-destructive and timely. The combination of near-infrared spectroscopy with chemometrics proves to be a good way to reach these premises. A new strategy is proposed for classification and verification of the adulteration of liquors by using NIR spectroscopy and chemometric classification, i.e., ensemble support vector machines (SVM). Three measures, i.e., accuracy, sensitivity and specificity were used for performance evaluation. The results confirmed that the strategy can serve as a screening tool applied to verify adulteration of the liquor, that is, a prior step used to condition the sample to a deeper analysis only when a positive result for adulteration is obtained by the proposed methodology. Ó 2014 Elsevier B.V. All rights reserved.
Introduction Chinese liquor (distilled spirits) is one of the famous distilled spirits and has been consumed for centuries in China [1]. As a ⇑ Corresponding author at: Department of Chemistry and Chemical Engineering, Yibin University, Yibin, Sichuan 644007, PR China. Tel./fax: +86 831 3551080. E-mail address: [email protected] (C. Tan). http://dx.doi.org/10.1016/j.saa.2014.03.091 1386-1425/Ó 2014 Elsevier B.V. All rights reserved.
complex matrix, liquor is composed of water, ethanol, inorganic elements and many kinds of fragrance ingredients such as esters, among which water and ethanol are the two major constituents and account for about 98% of the total mass. In recent years, counterfeit liquor is becoming a big, expensive issue in the market. Most of counterfeit liquors are prepared by simple dilution of the original liquors, by water or ethanol, or by a mixture of alcohol, water and aroma. Also, age liquor is facing the crisis of confidence
246
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. The authenticity of liquor is regulated by strict guidelines laid down by the responsible national authorities, which may include official sensory evaluation, chemical analysis, and examination of the records [2]. The identification or classification of liquors, mainly in terms of variety and geographical region of origin, has received increasing attention during the past 10 years. Most methods for screening liquor authenticity has been attempted based several different types of compositional data such as volatile compounds [3,4]. However, all of these methods include the pre-treatment process such as distillation, and the operation are complicated, time-consuming and laborious. Nowadays, the application of spectroscopic techniques in wine analysis and quality control has developed considerably [5,6]. Especially, Near-infrared (NIR) spectroscopy offers the advantages of simplicity of sample presentation, the speed of collecting the information (spectra) and low cost [7–9]. It is well known that the NIR technique is based on the correlation between chemical properties and absorption of light at different wavelengths in the NIR region, measured by reflectance, transmittance or transflectance. Often, the NIR signal is very weak. So, it is inevitable to use modeling techniques including classification and regression to highlight useful information. Also, the performance of a model is decisive to the availability of NIR-based applications. In recent years, great effort has been made to develop model construction methods for improving the performance. In conventional practice, most methods are based on constructing a single model, which in some cases may result in unsatisfactory performance, especially when the number samples is relatively small, which therefore leads to the emergence of the so-called ensemble technique [10–12]. Ensemble has gained increasing attention in many fields and has made a fundamental shift for model construction, i.e., instead of trying to build a single complex model, one can instead resort to combine a group of models. It enables an increase in generalization performance by combining several individual models trained on the same task. The ensemble advantage has been justified both theoretically and empirically. An ideal method used to verify the quality and authenticity of liquors should be non-invasive, non-destructive and timely. In addition, it should to accomplish a fast data acquisition and data treatment accurately with relatively low costs. The combination of NIR spectroscopy with chemometric is a good way to reach these premises. In this work, it is proposed a new strategy for classification and verification of adulteration of liquors using NIR spectroscopy and chemometric classification, i.e., ensemble support vector machines (SVM). The results confirmed that the strategy can serve as a screening tool applied to verify adulteration of the liquor, that is, a prior step used to condition the sample to a deeper analysis only when a positive result for adulteration is obtained by the proposed methodology. Three measures, i.e., accuracy, sensitivity and specificity were used for model evaluation. Materials and methods Samples and spectra collection A total of 120 samples of commercial bottled liquors were collected in local stores of China. It contained 42 liquor samples belonging to a brand without marked age from the markets (Class 1) and 78 samples with marked age (Class 2). These liquors belong to a type of distilled spirit made from grains based on a series of process such as fermenting, distilling, storage and blending, etc. The samples of class 2 were actually age liquor, i.e., the high-grade liquor product in China and therefore had a high price and the samples of class 1 were low-grade with low price. It is known that through
ageing or maturation, the liquor body becomes more harmonious after undergoing a large number of physical and chemical reactions. Nowadays, age liquor is facing the crisis of confidence because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. Each sample was confirmed by expert and composition analysis. The two classes had different distribution and abundance of esters such as ethyl acetate, which reflects characteristic aroma of liquor. All analyses were done in duplicate. For each sample, liquor bottles were opened and subsamples were scanned on in transmission mode using a near-infrared spectrometer coupled with an automated transmission module (Antaris II, Thermo fisher, USA). Spectral data collection was made using Vision software-TQ Analyst. Samples were scanned in a rectangular curette with a 1-mm path length and temperature equilibrated at 25 °C for 2 min in the instrument before scanning. For each sample, the spectrum was recorded in the region of 4000–12,000 cm 1 at intervals of 4 cm 1, containing 2074 points and was an average of 32 scans. The absorbance in the region of 8000–12,000 cm 1 was very weak. Ensemble support vector machines In the past few years, there has been increased attention in the literature on the use of so-called ensemble methods in pattern recognition such as classification [13,14]. These methods have been shown to have interesting properties. Ensemble modeling is a word originated from machine learning. The idea behind ensemble methods is to generate a large number of alternative predictors obtained on independently perturbed data and combine the predictors either by averaging or majority vote strategies. The most well known method in the class of techniques is perhaps bootstrap aggregating (bagging) [15]. In fact, a known problem in machine learning is that classifiers with good training performance often exhibit a poor generalization performance on unseen data. A possibility to overcome these limitations is to create an ensemble of classifiers and average the output of all independent/member classifiers. This idea can be compared to the process of consulting several experts for making a final decision. For a medical practice, it means to seek the opinion of several doctors. By the way, it certainly reduces the risk of making a poor classification based on a single model. Another major advantage of ensemble classifiers is that they are able to successfully address small sample problem. The mechanism of ensemble modeling can be analyzed indepth by bias/variance decomposition of the error. More specifically, the generalization error of an ensemble model could be significantly improved if the predictions of different member models disagree and if their residual errors are uncorrelated. The success of the ensemble heavily depends on so-called trade-off between accuracy and diversity of member models in it. Different ways of achieving this give birth to different ensemble methods. Several ways are available. The most common way is based on the re-sampling of the training set to produce diverse subsets. In this work, we use an effective scheme, i.e., a mixture of bagging and cross-validation. As one of the first successfully applied ensemble-based technique, bagging can improve the classification by combining classifiers trained on randomly generated sample subsets. Wichard [16] extended it by applying a cross-validation (CV) scheme for model selection on each subset and further combine the selected models to an ensemble model. Unlike traditional cross-validation, it adopts random subsets for cross-validation folds. The ensemble scheme consists of the following steps: 1. For the K-fold CV, the dataset is divided K-times into a training set and a test set, both sets containing randomly drawn subsets of the dataset without replications. The size of each test set was 20% of the entire dataset.
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
2. Training some candidate member models/classifiers for every CV-fold. Here, support vector machine (SVM) is selected as the member classifier. 3. By evaluating the prediction errors on the unseen data of the test set, only those best models are taken out and become ensemble members. 4. The average of K member models is used for a final prediction. In fact, the models with the lowest misclassified error in each CV-fold are taken out and added to the final ensemble for future prediction. The SVM can be seen as a method of training polynomial, radial basis function, or multilayer perception classifiers, in which the weights of the network are found by solving a quadratic programming (QP) problem with linear inequality and equality constraints. The SVM uses structural risk minimization, rather than empirical risk minimization in standard neural network training technique. Even the SVM is considered to have good generalization ability, the prediction result of SVM model is often far from the theoretically expected level. This promotes the emergence of the ensemble SVM such as boosting SVM, in which each individual SVM is trained by the training samples chosen according to the sample’s probability distribution that is updated in proportional to the errorness of the sample [17]. Often, the training process is slow because the training set depends on the performance of a previous model. In the present work, all member models of SVM can be trained in parallel and the process is therefore faster. More details on SVM can be found in other papers [18,19]. The performance of all the classifiers was assessed in terms of three measures, i.e., total accuracy (ACC), sensitivity (SENS, the percentage of authentic liquors classified as authentic) and the specificity (SPEC, the percentage of adulterated liquors considered as adulterated). All of the calculations and modeling were performed with Matlab version 7.1 under Windows Xp.
247
with CAH stretch first overtone, and at 5128 cm 1 with OAH stretch first overtone of water and ethanol. The absorption band at 4413 cm 1 is likely related to CAH combinations and OAH stretch overtones and that at 4338 cm 1 to CAH overtones. Absorptions at 4413 and 4438 cm 1 are most likely CAH combination bands of methanol. Spectral data were analyzed by principal component analysis (PCA) which is a common tool used to obtain useful information about the latent structure of spectral data, sample distribution, and spectral outliers. Fig. 2 shows the percentage of explained variance versus the number of principal component (PCs). It is clear in Fig. 2 that the first three PCs account for more than 99% the variation in the spectra. To investigate the basis for the observed spectral discrimination between the two liquor varieties, Fig. 3 shows the PCA loadings (eigenvectors) of the first three PCs. The PC1 explains most of the total variance in the samples, and the highest loadings are found around 6993 cm 1 associated with OAH absorption bands (second overtone), at 5285 cm 1 with OAH stretch and C@O second overtone combinations, at 5165 cm 1 with OAH first overtone, respectively. These spectral regions are characteristic of either water absorption of OAH overtones, ethanol, or sugars in the liquors. Therefore, it is suggested that particular chemical constituents, such as ethanol, water, sugars, phenolic compounds, lactic acid, and oxidation products, either in combination or alone, contribute the strongest influences that explain the basis for the observed discrimination between the two commercial varieties. Fig. 4 gives the scores scatter plot based on the first three
Results and discussion Spectra analysis Fig. 1 shows the mean spectra for the authentic and adulterated liquors analyzed. No obvious differences were detected from a visual observation of the spectra between the two varieties in the NIR region. Both varieties have absorption bands at 6896 cm 1 related to the OAH second overtone of water and ethanol, at 5917 cm 1 related with either CAH3 stretch first overtone or compounds containing CAH aromatic groups, at 5586 cm 1 related
Fig. 2. The percentage of explained variance versus the number of PCs.
Fig. 1. Mean spectra for the authentic and adulterated liquors.
Fig. 3. The PCA loadings (eigenvectors) of the first three PCs.
248
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
PCs, among which (a), (b) and (c) subplots correspond to PC1–PC2, PC1–PC3 and PC2–PC4, respectively. The asterisk represents class 1 while the circles represent class 2. As can be seen in Fig 4 a and b, the points corresponding to class 1 gather in a very compact region while those points related to class 2 scatter in a broad region. An obvious overlap will make the discrimination very difficult. Also, there exist some clusters with class 2. It appears that age liquors of a brand maybe exhibit greater dispersion, e.g., there exist two samples marked as asterisk located some distance away from the remainder of this class. It can be partially due to different aging conditions and processes. Classification model Based on the ensemble method described above, a series of SVM classifiers were constructed. Considering the fact that the sample set was relatively small, the cross-validation and two-dimensional grid search were used to optimize each classifier. Furthermore, to verify the robustness of ensemble classifiers, randomly partitioning the dataset into training and validation subsets runs 20 times. The performance of the classifiers was estimated in terms of sensitivity (SENS), i.e. the percentage of objects of class 2 identified correctly by the classifier, and specificity (SPEC), i.e. the percentage of
Fig. 5. Shows the performance measures for 20 runs of ensemble SVM and the optimal single SVM model.
objects of class 1 identified correctly by the classifier and the accuracy (ACC), i.e., the total percentage of objects classified correctly. The means and standard deviations (S.D.) of the ACC, SENS and SPEC are 0.9752, 0.9636, 0.9959 and 0.0189, 0.0296, 0.0126, respectively. Fig. 5 shows the performance measures of 20 runs. On the same dataset, a single SVM classifier was trained and optimized. The accuracy (ACC0), sensitivity (SENS0) and specificity (SPEC0) are 94.9%, 93.1% and 97.9%, respectively, and are represented by horizontal lines in Fig. 5. Clearly, among 20 runs, 18 ensemble SVM classifiers outperform the optimal single SVM classifier, meaning that on average, such an ensemble model is superior to a single model with a probability of 90%. The ability of the NIR-based model to discriminate or identify liquor varieties is related to the vibrational responses of chemical bonds to NIR radiation. It is probable that the higher the variability between sample types in those chemical entities, which respond in these regions of the spectrum, the better the accuracy of the model. This suggests that it is not simply a specific constituent, such as ethanol and esters but also the compositional characteristics of the liquor as a whole that provides the necessary information for discrimination by NIR techniques. It seems that NIR techniques might offer the possibility to analyze liquor for authenticity of variety without the need for costly and laborious chemical and sensory analysis. Conclusions The results obtained in this study showed the potential of nearinfrared spectroscopy to discriminate by variety between commercially available bottles of liquors. NIR spectroscopy together with chemometric techniques could serve as a tool for the identification of Chinese liquor varieties or their blends. The work reported here constitutes a feasibility study and requires further development with considerably more commercial samples of different varieties before its potential may be realized and adopted by the liquor industry. Acknowledgements
Fig. 4. The scores scatter plot based on the first three PCs. (a), (b) and (c) subplots correspond to PC1–PC2, PC1–PC3 and PC2–PC3, respectively.
This work was supported by the National Natural Science Foundation of China (21375118), the Applied Basic Research Programs of Science and Technology Department of Sichuan Province of China (2013JY0101), Scientific Research Foundation of Sichuan Provincial Education Department of China (12ZA201, 13ZB0300), Yibin Municipal Key Research Foundation (2013GY018), and Innovative Research and Teaching Team Program of Yibin University (Cx201104).
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
References [1] C.W. Li, J.P. Wei, Q. Zhou, S.Q. Sun, J. Mol. Struct. 883 (2008) 99–102. [2] D. Cozzolino, H.E. Smyth, M. Gishen, J. Agric. Food Chem. 51 (2003) 7703–7708. [3] M.J.C. Pontes, S.R.B. Santos, M.C.U. Araújo, L.F. Almeida, R.A.C. Lima, E.N. Gaiáo, U.T.C.P. Souto, Food Res. Int. 39 (2006) 182–189. [4] D. Cozzolino, M. Holdstock, R.G. Dambergs, W.U. Cynkar, P.A. Smith, Food. Chem. 116 (2009) 761–765. [5] H.Y. Yu, X.Y. Niu, H.J. Liu, Y.B. Ying, B.B. Li, X.X. Pan, Food Chem. 113 (2009) 291–296. [6] Z. Ge, Z.H. Song, Chemometr. Intell. Lab. Syst. 125 (2013) 51–57. [7] X.Y. Niu, F. Shen, Y.F. Yu, Z. Yan, K. Xu, H.Y. Yu, J. Agric. Food. Chem. 56 (2008) 7271–7278. [8] C. Tan, T. Wu, Z.H. Xu, W.Y. Li, K.S. Zhang, Vib. Spectrosc. 58 (2012) 44–49.
249
[9] J. Duan, Y. Huang, Z.H. Li, B. Zheng, Q.Q. Li, Y.M. Xiong, L.J. Wu, S.G. Min, Ind. Crop. Prod. 40 (2012) 21–26. [10] B.H. Mevik, V.H. Segtnan, T. Næs, J. Chemometr. 18 (2004) 498–507. [11] Q.D. Su, W.D. Tong, L.M. Shi, X.G. Shao, W.S. Cai, Anal. Lett. 39 (2006) 2073. [12] H. Chen, C. Tan, Z. Lin, T. Wu, Y.B. Diao, Comput. Biol. Med. 43 (2013) 865–869. [13] .R.A. Viscarra Rossel, J. Near Infrared Spectrosc. 15 (2007) 39–47. [14] C. Tan, X. Qin, M.L. Li, Anal. Lett. 42 (2009) 1693–1710. [15] L. Breiman, Mach. Learn. 24 (1996) 123–140. [16] J.D. Wichard, H. Cammann, C. Stephan, T. Tolxdorff, J. Biomed. Biotechnol. 21 (2008) 1–5. [17] H.C. Kim, S.N. Pang, H.M. Je, Pattern Recogn. 36 (2003) 2757–2767. [18] C. Cortes, V. Vapnik, Support vector network, Mach. Learn. 20 (1995) 273–297. [19] J.H. Hong, S.B. Cho, Neurocomputing 71 (2008) 3275–3281.
Contents lists available at ScienceDirect
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy journal homepage: www.elsevier.com/locate/saa
Discrimination between authentic and adulterated liquors by near-infrared spectroscopy and ensemble classification Hui Chen a, Chao Tan a,b,c,⇑, Tong Wu a, Li Wang a, Wanping Zhu a a
Key Lab of Process Analysis and Control, Yibin University, Yibin, Sichuan 644007, PR China Department of Chemistry and Chemical Engineering, Yibin University, Yibin, Sichuan 644007, PR China c Computational Physics Key Laboratory of Sichuan Province, Yibin University, Yibin, Sichuan 644007, PR China b
h i g h l i g h t s
g r a p h i c a l a b s t r a c t
Near-infrared spectroscopy is used to
discriminate liquors. Ensemble support vector machines
serves as a tool. Such a method is fast and objective.
a r t i c l e
i n f o
Article history: Received 9 January 2014 Received in revised form 5 March 2014 Accepted 18 March 2014 Available online 2 April 2014 Keywords: Liquor Authenticity Near-infrared Support vector machines
a b s t r a c t Chinese liquor is one of the famous distilled spirits and counterfeit liquor is becoming a serious problem in the market. Especially, age liquor is facing the crisis of confidence because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. An ideal method for authenticity confirmation of liquors should be non-invasive, non-destructive and timely. The combination of near-infrared spectroscopy with chemometrics proves to be a good way to reach these premises. A new strategy is proposed for classification and verification of the adulteration of liquors by using NIR spectroscopy and chemometric classification, i.e., ensemble support vector machines (SVM). Three measures, i.e., accuracy, sensitivity and specificity were used for performance evaluation. The results confirmed that the strategy can serve as a screening tool applied to verify adulteration of the liquor, that is, a prior step used to condition the sample to a deeper analysis only when a positive result for adulteration is obtained by the proposed methodology. Ó 2014 Elsevier B.V. All rights reserved.
Introduction Chinese liquor (distilled spirits) is one of the famous distilled spirits and has been consumed for centuries in China [1]. As a ⇑ Corresponding author at: Department of Chemistry and Chemical Engineering, Yibin University, Yibin, Sichuan 644007, PR China. Tel./fax: +86 831 3551080. E-mail address: [email protected] (C. Tan). http://dx.doi.org/10.1016/j.saa.2014.03.091 1386-1425/Ó 2014 Elsevier B.V. All rights reserved.
complex matrix, liquor is composed of water, ethanol, inorganic elements and many kinds of fragrance ingredients such as esters, among which water and ethanol are the two major constituents and account for about 98% of the total mass. In recent years, counterfeit liquor is becoming a big, expensive issue in the market. Most of counterfeit liquors are prepared by simple dilution of the original liquors, by water or ethanol, or by a mixture of alcohol, water and aroma. Also, age liquor is facing the crisis of confidence
246
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. The authenticity of liquor is regulated by strict guidelines laid down by the responsible national authorities, which may include official sensory evaluation, chemical analysis, and examination of the records [2]. The identification or classification of liquors, mainly in terms of variety and geographical region of origin, has received increasing attention during the past 10 years. Most methods for screening liquor authenticity has been attempted based several different types of compositional data such as volatile compounds [3,4]. However, all of these methods include the pre-treatment process such as distillation, and the operation are complicated, time-consuming and laborious. Nowadays, the application of spectroscopic techniques in wine analysis and quality control has developed considerably [5,6]. Especially, Near-infrared (NIR) spectroscopy offers the advantages of simplicity of sample presentation, the speed of collecting the information (spectra) and low cost [7–9]. It is well known that the NIR technique is based on the correlation between chemical properties and absorption of light at different wavelengths in the NIR region, measured by reflectance, transmittance or transflectance. Often, the NIR signal is very weak. So, it is inevitable to use modeling techniques including classification and regression to highlight useful information. Also, the performance of a model is decisive to the availability of NIR-based applications. In recent years, great effort has been made to develop model construction methods for improving the performance. In conventional practice, most methods are based on constructing a single model, which in some cases may result in unsatisfactory performance, especially when the number samples is relatively small, which therefore leads to the emergence of the so-called ensemble technique [10–12]. Ensemble has gained increasing attention in many fields and has made a fundamental shift for model construction, i.e., instead of trying to build a single complex model, one can instead resort to combine a group of models. It enables an increase in generalization performance by combining several individual models trained on the same task. The ensemble advantage has been justified both theoretically and empirically. An ideal method used to verify the quality and authenticity of liquors should be non-invasive, non-destructive and timely. In addition, it should to accomplish a fast data acquisition and data treatment accurately with relatively low costs. The combination of NIR spectroscopy with chemometric is a good way to reach these premises. In this work, it is proposed a new strategy for classification and verification of adulteration of liquors using NIR spectroscopy and chemometric classification, i.e., ensemble support vector machines (SVM). The results confirmed that the strategy can serve as a screening tool applied to verify adulteration of the liquor, that is, a prior step used to condition the sample to a deeper analysis only when a positive result for adulteration is obtained by the proposed methodology. Three measures, i.e., accuracy, sensitivity and specificity were used for model evaluation. Materials and methods Samples and spectra collection A total of 120 samples of commercial bottled liquors were collected in local stores of China. It contained 42 liquor samples belonging to a brand without marked age from the markets (Class 1) and 78 samples with marked age (Class 2). These liquors belong to a type of distilled spirit made from grains based on a series of process such as fermenting, distilling, storage and blending, etc. The samples of class 2 were actually age liquor, i.e., the high-grade liquor product in China and therefore had a high price and the samples of class 1 were low-grade with low price. It is known that through
ageing or maturation, the liquor body becomes more harmonious after undergoing a large number of physical and chemical reactions. Nowadays, age liquor is facing the crisis of confidence because it is difficult for consumer to identify the marked age, which prompts unscrupulous traders to pose off low-grade liquors as high-grade liquors. Each sample was confirmed by expert and composition analysis. The two classes had different distribution and abundance of esters such as ethyl acetate, which reflects characteristic aroma of liquor. All analyses were done in duplicate. For each sample, liquor bottles were opened and subsamples were scanned on in transmission mode using a near-infrared spectrometer coupled with an automated transmission module (Antaris II, Thermo fisher, USA). Spectral data collection was made using Vision software-TQ Analyst. Samples were scanned in a rectangular curette with a 1-mm path length and temperature equilibrated at 25 °C for 2 min in the instrument before scanning. For each sample, the spectrum was recorded in the region of 4000–12,000 cm 1 at intervals of 4 cm 1, containing 2074 points and was an average of 32 scans. The absorbance in the region of 8000–12,000 cm 1 was very weak. Ensemble support vector machines In the past few years, there has been increased attention in the literature on the use of so-called ensemble methods in pattern recognition such as classification [13,14]. These methods have been shown to have interesting properties. Ensemble modeling is a word originated from machine learning. The idea behind ensemble methods is to generate a large number of alternative predictors obtained on independently perturbed data and combine the predictors either by averaging or majority vote strategies. The most well known method in the class of techniques is perhaps bootstrap aggregating (bagging) [15]. In fact, a known problem in machine learning is that classifiers with good training performance often exhibit a poor generalization performance on unseen data. A possibility to overcome these limitations is to create an ensemble of classifiers and average the output of all independent/member classifiers. This idea can be compared to the process of consulting several experts for making a final decision. For a medical practice, it means to seek the opinion of several doctors. By the way, it certainly reduces the risk of making a poor classification based on a single model. Another major advantage of ensemble classifiers is that they are able to successfully address small sample problem. The mechanism of ensemble modeling can be analyzed indepth by bias/variance decomposition of the error. More specifically, the generalization error of an ensemble model could be significantly improved if the predictions of different member models disagree and if their residual errors are uncorrelated. The success of the ensemble heavily depends on so-called trade-off between accuracy and diversity of member models in it. Different ways of achieving this give birth to different ensemble methods. Several ways are available. The most common way is based on the re-sampling of the training set to produce diverse subsets. In this work, we use an effective scheme, i.e., a mixture of bagging and cross-validation. As one of the first successfully applied ensemble-based technique, bagging can improve the classification by combining classifiers trained on randomly generated sample subsets. Wichard [16] extended it by applying a cross-validation (CV) scheme for model selection on each subset and further combine the selected models to an ensemble model. Unlike traditional cross-validation, it adopts random subsets for cross-validation folds. The ensemble scheme consists of the following steps: 1. For the K-fold CV, the dataset is divided K-times into a training set and a test set, both sets containing randomly drawn subsets of the dataset without replications. The size of each test set was 20% of the entire dataset.
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
2. Training some candidate member models/classifiers for every CV-fold. Here, support vector machine (SVM) is selected as the member classifier. 3. By evaluating the prediction errors on the unseen data of the test set, only those best models are taken out and become ensemble members. 4. The average of K member models is used for a final prediction. In fact, the models with the lowest misclassified error in each CV-fold are taken out and added to the final ensemble for future prediction. The SVM can be seen as a method of training polynomial, radial basis function, or multilayer perception classifiers, in which the weights of the network are found by solving a quadratic programming (QP) problem with linear inequality and equality constraints. The SVM uses structural risk minimization, rather than empirical risk minimization in standard neural network training technique. Even the SVM is considered to have good generalization ability, the prediction result of SVM model is often far from the theoretically expected level. This promotes the emergence of the ensemble SVM such as boosting SVM, in which each individual SVM is trained by the training samples chosen according to the sample’s probability distribution that is updated in proportional to the errorness of the sample [17]. Often, the training process is slow because the training set depends on the performance of a previous model. In the present work, all member models of SVM can be trained in parallel and the process is therefore faster. More details on SVM can be found in other papers [18,19]. The performance of all the classifiers was assessed in terms of three measures, i.e., total accuracy (ACC), sensitivity (SENS, the percentage of authentic liquors classified as authentic) and the specificity (SPEC, the percentage of adulterated liquors considered as adulterated). All of the calculations and modeling were performed with Matlab version 7.1 under Windows Xp.
247
with CAH stretch first overtone, and at 5128 cm 1 with OAH stretch first overtone of water and ethanol. The absorption band at 4413 cm 1 is likely related to CAH combinations and OAH stretch overtones and that at 4338 cm 1 to CAH overtones. Absorptions at 4413 and 4438 cm 1 are most likely CAH combination bands of methanol. Spectral data were analyzed by principal component analysis (PCA) which is a common tool used to obtain useful information about the latent structure of spectral data, sample distribution, and spectral outliers. Fig. 2 shows the percentage of explained variance versus the number of principal component (PCs). It is clear in Fig. 2 that the first three PCs account for more than 99% the variation in the spectra. To investigate the basis for the observed spectral discrimination between the two liquor varieties, Fig. 3 shows the PCA loadings (eigenvectors) of the first three PCs. The PC1 explains most of the total variance in the samples, and the highest loadings are found around 6993 cm 1 associated with OAH absorption bands (second overtone), at 5285 cm 1 with OAH stretch and C@O second overtone combinations, at 5165 cm 1 with OAH first overtone, respectively. These spectral regions are characteristic of either water absorption of OAH overtones, ethanol, or sugars in the liquors. Therefore, it is suggested that particular chemical constituents, such as ethanol, water, sugars, phenolic compounds, lactic acid, and oxidation products, either in combination or alone, contribute the strongest influences that explain the basis for the observed discrimination between the two commercial varieties. Fig. 4 gives the scores scatter plot based on the first three
Results and discussion Spectra analysis Fig. 1 shows the mean spectra for the authentic and adulterated liquors analyzed. No obvious differences were detected from a visual observation of the spectra between the two varieties in the NIR region. Both varieties have absorption bands at 6896 cm 1 related to the OAH second overtone of water and ethanol, at 5917 cm 1 related with either CAH3 stretch first overtone or compounds containing CAH aromatic groups, at 5586 cm 1 related
Fig. 2. The percentage of explained variance versus the number of PCs.
Fig. 1. Mean spectra for the authentic and adulterated liquors.
Fig. 3. The PCA loadings (eigenvectors) of the first three PCs.
248
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
PCs, among which (a), (b) and (c) subplots correspond to PC1–PC2, PC1–PC3 and PC2–PC4, respectively. The asterisk represents class 1 while the circles represent class 2. As can be seen in Fig 4 a and b, the points corresponding to class 1 gather in a very compact region while those points related to class 2 scatter in a broad region. An obvious overlap will make the discrimination very difficult. Also, there exist some clusters with class 2. It appears that age liquors of a brand maybe exhibit greater dispersion, e.g., there exist two samples marked as asterisk located some distance away from the remainder of this class. It can be partially due to different aging conditions and processes. Classification model Based on the ensemble method described above, a series of SVM classifiers were constructed. Considering the fact that the sample set was relatively small, the cross-validation and two-dimensional grid search were used to optimize each classifier. Furthermore, to verify the robustness of ensemble classifiers, randomly partitioning the dataset into training and validation subsets runs 20 times. The performance of the classifiers was estimated in terms of sensitivity (SENS), i.e. the percentage of objects of class 2 identified correctly by the classifier, and specificity (SPEC), i.e. the percentage of
Fig. 5. Shows the performance measures for 20 runs of ensemble SVM and the optimal single SVM model.
objects of class 1 identified correctly by the classifier and the accuracy (ACC), i.e., the total percentage of objects classified correctly. The means and standard deviations (S.D.) of the ACC, SENS and SPEC are 0.9752, 0.9636, 0.9959 and 0.0189, 0.0296, 0.0126, respectively. Fig. 5 shows the performance measures of 20 runs. On the same dataset, a single SVM classifier was trained and optimized. The accuracy (ACC0), sensitivity (SENS0) and specificity (SPEC0) are 94.9%, 93.1% and 97.9%, respectively, and are represented by horizontal lines in Fig. 5. Clearly, among 20 runs, 18 ensemble SVM classifiers outperform the optimal single SVM classifier, meaning that on average, such an ensemble model is superior to a single model with a probability of 90%. The ability of the NIR-based model to discriminate or identify liquor varieties is related to the vibrational responses of chemical bonds to NIR radiation. It is probable that the higher the variability between sample types in those chemical entities, which respond in these regions of the spectrum, the better the accuracy of the model. This suggests that it is not simply a specific constituent, such as ethanol and esters but also the compositional characteristics of the liquor as a whole that provides the necessary information for discrimination by NIR techniques. It seems that NIR techniques might offer the possibility to analyze liquor for authenticity of variety without the need for costly and laborious chemical and sensory analysis. Conclusions The results obtained in this study showed the potential of nearinfrared spectroscopy to discriminate by variety between commercially available bottles of liquors. NIR spectroscopy together with chemometric techniques could serve as a tool for the identification of Chinese liquor varieties or their blends. The work reported here constitutes a feasibility study and requires further development with considerably more commercial samples of different varieties before its potential may be realized and adopted by the liquor industry. Acknowledgements
Fig. 4. The scores scatter plot based on the first three PCs. (a), (b) and (c) subplots correspond to PC1–PC2, PC1–PC3 and PC2–PC3, respectively.
This work was supported by the National Natural Science Foundation of China (21375118), the Applied Basic Research Programs of Science and Technology Department of Sichuan Province of China (2013JY0101), Scientific Research Foundation of Sichuan Provincial Education Department of China (12ZA201, 13ZB0300), Yibin Municipal Key Research Foundation (2013GY018), and Innovative Research and Teaching Team Program of Yibin University (Cx201104).
H. Chen et al. / Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 130 (2014) 245–249
References [1] C.W. Li, J.P. Wei, Q. Zhou, S.Q. Sun, J. Mol. Struct. 883 (2008) 99–102. [2] D. Cozzolino, H.E. Smyth, M. Gishen, J. Agric. Food Chem. 51 (2003) 7703–7708. [3] M.J.C. Pontes, S.R.B. Santos, M.C.U. Araújo, L.F. Almeida, R.A.C. Lima, E.N. Gaiáo, U.T.C.P. Souto, Food Res. Int. 39 (2006) 182–189. [4] D. Cozzolino, M. Holdstock, R.G. Dambergs, W.U. Cynkar, P.A. Smith, Food. Chem. 116 (2009) 761–765. [5] H.Y. Yu, X.Y. Niu, H.J. Liu, Y.B. Ying, B.B. Li, X.X. Pan, Food Chem. 113 (2009) 291–296. [6] Z. Ge, Z.H. Song, Chemometr. Intell. Lab. Syst. 125 (2013) 51–57. [7] X.Y. Niu, F. Shen, Y.F. Yu, Z. Yan, K. Xu, H.Y. Yu, J. Agric. Food. Chem. 56 (2008) 7271–7278. [8] C. Tan, T. Wu, Z.H. Xu, W.Y. Li, K.S. Zhang, Vib. Spectrosc. 58 (2012) 44–49.
249
[9] J. Duan, Y. Huang, Z.H. Li, B. Zheng, Q.Q. Li, Y.M. Xiong, L.J. Wu, S.G. Min, Ind. Crop. Prod. 40 (2012) 21–26. [10] B.H. Mevik, V.H. Segtnan, T. Næs, J. Chemometr. 18 (2004) 498–507. [11] Q.D. Su, W.D. Tong, L.M. Shi, X.G. Shao, W.S. Cai, Anal. Lett. 39 (2006) 2073. [12] H. Chen, C. Tan, Z. Lin, T. Wu, Y.B. Diao, Comput. Biol. Med. 43 (2013) 865–869. [13] .R.A. Viscarra Rossel, J. Near Infrared Spectrosc. 15 (2007) 39–47. [14] C. Tan, X. Qin, M.L. Li, Anal. Lett. 42 (2009) 1693–1710. [15] L. Breiman, Mach. Learn. 24 (1996) 123–140. [16] J.D. Wichard, H. Cammann, C. Stephan, T. Tolxdorff, J. Biomed. Biotechnol. 21 (2008) 1–5. [17] H.C. Kim, S.N. Pang, H.M. Je, Pattern Recogn. 36 (2003) 2757–2767. [18] C. Cortes, V. Vapnik, Support vector network, Mach. Learn. 20 (1995) 273–297. [19] J.H. Hong, S.B. Cho, Neurocomputing 71 (2008) 3275–3281.
