In practice, the process of nding the geodesic neighbors is composed of
two phases. The rst phase consists in constructing a weighted graph G over the
data set where neighboring data points are connected. In principle, any similarity
measure dE can be adopted to determine neighboring relations, and probably
the Euclidean distance is the most common choice. Two points are neighbors if
are closer than a xed distance (-graph), or one is the K nearest point of the