Design of kernel methods for classification of structured data: methodologies and applications

Many problems in machine learning and pattern recognition involve variable size structured data, such as sets, sequences, trees, and graphs. This structural representation of data overcomes the intrinsic limitations of the traditional, fixed-length vectorial (feature-based) representation. Among machine learning techniques, kernel methods (e.g., Support Vector Machines, etc.) can naturally deal with structured data. This thesis focuses on the design of novel methodologies for kernel-based classification of structured data. In particular, we propose new contributions in the field of generative kernels and kernels based on the generalized dissimilarity representation of data. The effectiveness of the proposed approaches are assessed on real-world classification tasks.