Last decade has witnessed a tremendous expansion of mobile devices, which brought an unprecedented opportunity to reach a large number of mobile users at any point in time. This resulted in a surge of interest of mobile operators and ad publishers to understand usage patterns of mobile apps and allow more relevant content recommendations. Due to a large input space, a critical step in understanding app usage patterns is reducing sparseness by classifying apps into predefined interest taxonomies. However, besides short name and noisy description majority of apps have very limited information available, which makes classification a challenging task. We address this issue and present a novel method to classify apps into interest categories by: 1) embedding apps into low-dimensional space using a neural language model applied on smartphone logs; and 2) applying k-nearest-neighbors classification in the embedding space. To validate the method we run experiments on more than one billion device logs covering hundreds of thousands of apps. To the best of our knowledge this is the first app categorization study at this scale. Empirical results show that the proposed method outperforms the current state-of-the-art.