Fast Data Mining with Sparse Chemical Graph Fingerprints by Estimating the Probability of Unique Patterns

The aim of this work is to introduce a modification of chemical graphs fingerprints for data mining. The algorithm reduces the number of features by taking the probability of producing an unique feature at a specific search depth into account. We observed the probability of generating a non-unique feature depending on a search parameter (which leads to a power-law growths of features) and modeled it by a sigmoid function. This function was integrated into a fingerprinting routine to reduce the features according to their probability. The predictive performance was convincing with a considerable speedup for the training of a linear support vector machine for sparse instances.