Text classification using string kernels
The Journal of Machine Learning Research
| Hi-index | 0.00 |
We consider the problem of biosonar landmark classification as an example of random and non-stationary signal classification in which finding robust and structure independent features for classification is not trivial. Time-frequency domain studies show that despite the seemingly randomness of those signals, there are local temporal similarities, independent of the position of occurrence in echoes of each object that reflect the intrinsic similarities between the echoes and also a self similarity in the objects. In this paper we suggest a time resolved spectrum kernel for extracting the local similarities (subsequence similarity) in time series in general, and as an example in biosonar signals. We implemented this kernel using dynamic programming and could get accurate results using a low number of echoes needed for training compared with the methods in which finding specific features in each echo were followed.