I am trying to implement a Laplacian SVM classifier (trained in primal) using algorithm from this paper.
I would like to know what is the most common way of constructing adjacency matrix and the most common weights to put in the matrix. By most common I mean something that give acceptable results on average.
In the paper the author talks about k-nearest neighbors or graph kernel as examples, I have also seen distance with threshold (what threshold value leads to acceptable average results?). For the weights he talks about binary weights and heat kernel weights (I don't know others).