Open brezniczky opened 5 years ago
Additionally, note that the entropy drops and the digit repetitions are not at all independent. Maybe these could be examined as a joint probability distribution simulation? Guess wouldn't be easy computationally - although next to nothing compared to today's workloads :) (and neither perhaps analytically).
Using a dissimilarity statistic it may be easy to verify which features are relevant and to be kept at all, i.e. yield the most disjoint vote fingerprints. Raising the overhearing prob. to some >1 power could also result in an improvement.