Planning of current and future mobile networks is becoming increasingly complex due to the heterogeneity of deployments, which feature not only macrocells, but also an underlying layer of small cells whose deployment is not fully under the control of the operator. In this paper, we focus on selecting the most appropriate Quality of Service (QoS) prediction techniques for assisting network operators in planning future dense deployments. We propose to use machine learning as a tool to extract the relevant information from the huge amount of data generated in current 4G and future 5G networks during normal operation, which is then used to appropriately plan networks. In particular, we focus on radio measurements to develop correlative statistical models with the purpose of improving QoS-based network planning. In this direction, we combine multiple learners by building ensemble methods and use them to do regression in a reduced space rather than in the original one. We then compare the QoS prediction accuracy of various approaches that take as input the 3GPP Minimization of Drive Tests (MDT) measurements collected throughout a heterogeneous network and analyse their trade-offs. We also explain how the collected data is processed and used to predict QoS expressed in terms of Physical Resource Block (PRB)/ Megabit (MB) transmitted. This metric was selected because of the interest it may have for operators in planning, since it relates lower layer resources with their impact in terms of QoS up in the protocol stack, hence closer to the end-user.