Parzen Window Approximation on Riemannian Manifold

In graph motivated learning, label propagation largely depends on data affinity represented as edges between connected points. The assignment implicitly assumes even distribution of the manifold. This assumption may not hold and lead to inaccurate metric due drift towards high-density regions. affected heat kernel based with a globally fixed Parzen window either discards genuine neighbors or forces distant points become member neighborhood. yields biased matrix. this paper, bias uneven sampling Riemannian manifold is catered by variable determined function neighborhood size, ambient dimension, flatness range, etc. Additionally, adjustment used which offsets effect responsible for bias. An takes into consideration irregular yield accurate proposed. Extensive experiments synthetic real-world sets confirm that proposed method increases classification accuracy significantly outperforms existing estimators in Laplacian regularization methods.