Relationship semantics refer to the types of social relationships between users in a social network, e.g., friend, family, enemy, etc. Inferring the semantics of social relationships using digital social footprints plays an important role in understanding the social network and utilizing them for further application. In this paper, we propose a semi-supervised machine-learning model based on a dual-view features co-training framework by employing both interaction behaviors between social dyads and structure features of the social network. Specifically, the intensity of social interaction and geographical co-occurrence are used to characterize interaction behaviors between social dyads, while network representation learning is used to extracting structure features of the dyads in their ego-networks. We evaluated our approach on a real mobile terminal usage dataset. Results show that our method can significantly improve the performance of social relationship semantics inference in the case of limited labeled data compared to the state-of-art methods.