We research sensible heuristics to enhance the efficiency of prefix-tree based mostly algorithms for differentially non-public heavy hitter detection. Our mannequin assumes every consumer has a number of knowledge factors and the purpose is to study as most of the most frequent knowledge factors as potential throughout all customers’ knowledge with mixture and native differential privateness. We suggest an adaptive hyperparameter tuning algorithm that improves the efficiency of the algorithm whereas satisfying computational, communication and mixture privateness constraints. We discover the affect of various data-selection schemes in addition to the affect of introducing deny lists throughout a number of runs of the algorithm. We take a look at these enhancements utilizing in depth experimentation on the Reddit dataset on the duty of studying most frequent phrases.