Symbolic Computation Group

In this work, we explore a noise-robust framework for multi-stroke handwritten symbol recognition by combining polynomial-based feature representations with graph neural networks. Each stroke is treated as a continuous function and represented using orthogonal polynomial bases such as Legendre and Chebyshev polynomials. We also study their Sobolev variants, which include derivative information to capture additional information about the structure of the strokes.

The polynomial coefficients are used as node features in a graph representation, where each node corresponds to a stroke and edges represent the spatial relationships between strokes. The resulting graph is then processed using a Graph Isomorphism Network (GIN).

We investigate the effect of Sobolev regularization on graph-based handwritten symbol recognition, particularly under noisy stroke conditions. This work explores how incorporating derivative-based information in polynomial features can influence the stability and robustness of graph learning models, providing a connection between classical polynomial approximation methods and modern graph neural networks.