Currently the distributed RF algo uses the rank-local cdf to select rows for bootstrap sampling. (The rank-local cdf is computed by computing cumulative sums over sample weights.)
To obtain consistent results that's independent of data partitioning, we should compute the global cdf and have the workers use the cdf to draw bootstrap samples.
Currently the distributed RF algo uses the rank-local cdf to select rows for bootstrap sampling. (The rank-local cdf is computed by computing cumulative sums over sample weights.)
To obtain consistent results that's independent of data partitioning, we should compute the global cdf and have the workers use the cdf to draw bootstrap samples.