Young-In Shin, Donald Fussell
A key challenge in applying kernel-based methods for discriminative learning is to identify a suitable kernel given a problem domain. Many methods instead transform the input data into a set of vectors in a feature space and classify the transformed data using a generic kernel. However, finding an effective transformation scheme for sequence (e.g. time series) data is a difficult task. In this paper, we introduce a scheme for directly designing kernels for the classification of sequence data such as that in handwritten character recognition and object recognition from sensor readings. Ordering information is represented by values of a parameter associated with each input data element. A similarity metric based on the parametric distance between corresponding elements is combined with their problem-specific similarity metric to produce a Mercer kernel suitable for use in methods such as support vector machine (SVM). This scheme directly embeds extraction of features from sequences of varying cardinalities into the kernel without needing to transform all input data into a common feature space before classification. We apply our method to object and handwritten character recognition tasks and compare against current approaches. The results show that we can obtain at least comparable accuracy to state of the art problem-specific methods using a systematic approach to kernel design. Our contribution is the introduction of a general technique for designing SVM kernels tailored for the classification of sequence data.
Subjects: 12. Machine Learning and Discovery; 17. Robotics
Submitted: Oct 12, 2006