Support Vector Machines Approach To Functional Group Prediction From Infrared Spectra

Along with the bigger of the infrared spectra database, the deepen development of the infrared technology and of the computer, it is a badly in need of solution about how to utilize and enlarge the application of infrared spectra. In the past decades, people are trying to search the way to interpret the infrared spectra. Along with the computerization of the commercialized infrared spectrometry, there are many computer-assisted interpretation of infrared spectra emerged. The automatic structure elucidation of infrared spectra generally falls into three groups: library search, knowledge-based systems, or pattern recognition. Among the last group of method, artificial neural networks (ANNs) and partial least squares (PLS) were most frequently used. However, their prediction accuracy of present functional groups is not satisfactory. Furthermore, ANNs have several major drawbacks: unsteadiness, local minima and very low speed of convergence. In this paper, we investigated the potential of support vector machine for the structure elucidation of infrared spectra. SVM is a kind of learning machine which is easy to obtain good results when there are a limited number of examples and a large number of variables. A set of 823 compounds were used in this paper. The compounds for the training set were selected by simply taking all of the even-numbered samples, while the test set were the odd-numbered samples. The training and test sets consisted of 411 and 412 FTIR spectra, respectively. Parameters (C and Sigma) influencing SVMs'training and prediction of functional groups from infrared spectra were scrutinized for the maximal achievable present prediction rate of the 16 functional groups. When the present prediction rate is very high but the absent prediction rate is very low, the striving for the best possible present prediction is relaxed a little in order to gain a higher accuracy of the absent prediction. The trained SVMs can identify the presence or absence of the functional groups with the average prediction accuracy of 93.3% for the presence (Pc) and 99.0% for the absence (Ac) of functionalities. The quality of found present response (Qpr) and absent response (Qar) was 96.0% and 98.8%, respectively. The average extrastatistical quality (EQr) was 93.4%. As to the ANNs, the overall prediction accuracy of 91.5% for the presence (Pc) and 98.5% for the absence (Ac) of functionalities were achieved. The quality of found present response (Qpr) and absent response (Qar) was 92.0% and 98.7%, respectively. The average extrastatistical quality (EQr) was 90.2%. the average values of SVMs were all higher than those of ANNs. Besides, the results showed that SVMs over-perform ANNs in most of the cases. From the results we can come to the conclusion that SVM approach is a powerful tool for the interpretation of infrared spectra.