The use of millimeter wave technologies offer a promising solution for dense small cell networks, despite having to contend with challenging propagation characteristics. In particular, user-induced effects can lead to significant channel variations depending on the user equipment (UE) usage mode which in turn, can impact the quality of service. Estimation of UE operating conditions is therefore critical for optimal radio resource management. We propose a new approach to user activity recognition which makes use of both supervised and unsupervised machine learning. In particular, using information extracted from the received signal strength (RSS), a common metric readily available from many receiver chipsets, we perform a classification of user state ( static or mobile relative to an access point) and UE mode of operation ( voice call, using an app or in pocket ). To develop and then train our classification system, measured RSS data was obtained using a custom 60 GHz measurement system for a range of indoor office scenarios which considered various UE to ceiling mounted access point configurations. In our approach, differentiation between static and mobile states is performed in preprocessing using a k-means algorithm. Small-scale fading features are then estimated from the RSS data and, using different feature scaling mechanisms, various supervised learning approaches are applied to investigate the optimal classification accuracy for the considered use cases in this work. We compare the classification performance of various window sizes and types, and show that a sliding window length of 1s without overlap performs best for time series segmentation at 60 GHz for the activities considered in this study. Among the different supervised learning approaches, the Decision Tree (DT) classifier performs best for both the user static and mobile cases with an accuracy of 100% and 98.0%, respectively. For static cases, user orientation, i.e., line-of-sight (LOS), quasi-LOS, and non-LOS, can also be classified and here the DT classifier also performs best with an accuracy of 98.2%, 97.6% and 100% for the voice call , using an app or in pocket use cases. Additionally, a feature ranking algorithm, called ReliefF , is adopted to determine the small-scale fading features that have the most significant influences on the classification accuracy and three different feature sets, namely Full , Reduced and Constrained sets, are then proposed based on feature ranking results. This allows the proposed techniques to be deployed on wireless platforms with different levels of processing capability.