This paper explores the problem of how to construct lazy decision tree ensembles. We present and empirically evaluate a relevancebased boosting-style algorithm that builds a lazy decision tree ensemble customized for each test instance. From the experimental results, we conclude that our boosting-style algorithm significantly improves the performance of the base learner. An empirical comparison to boosted regular decision trees shows that ensembles of lazy decision trees achieve comparable accuracy and better comprehensibility. We also introduce a novel distance-based pruning strategy for the lazy decision tree algorithm to address the problem of over-fitting. Our experiments show that the pruning strategy improves the accuracy and comprehensibility of both single lazy decision trees and boosted ensembles.