On time series features and kernels for machine olfaction

Kernel methods such as support vector machines are a powerful technique to solve pattern recognition problems. One of the important properties of kernel methods is that they can be applied to any kind of input domain, for which it is possible to construct an appropriate kernel. Over the past years, there has been a tremendous interest and progress in the machine learning community to design kernels for “non-standard” data sets, i.e., for data without a vectorial feature representation; examples include graphs, strings, trees, and other such discrete objects. In this paper, we investigate the benefit of using time series kernels to solve machine olfaction applications. In particular, we apply these time series kernels for two pattern recognition problems in machine olfaction, namely, odor classification and odor localization in an open sampling system. We also study the use of time series feature extraction methods, in which features are extracted by making assumptions on the underlying mechanism that generate the time series. Experimental results clearly indicate the advantage of using these kernels when compared to naïve techniques that discard the temporal information in the data, and, even more interestingly, these kernels also perform better that techniques that rely on an explicit feature extraction step prior to solving the pattern recognition problem.