During the last decade, a variety of social networks and applications has been developed, providing to the users a set of potential functionalities. Thanks to these functionalities, they have become vital part of the daily life of many people. As a result, a great volume of data has been created. Due to the different nature of the functionalities, datasets of different nature and schema are created. This paper introduces GeoTe-Gra, a system that targets to reveal non-obvious knowledge by connecting datasets that derive from multiple heterogeneous sources. GeoTeGra is a scalable framework to compare different machine learning algorithms in terms of scalability and effectiveness, finding semantic similarities between entities. Our system is based on a distributed storage and parallel map-reduce manipulation for the fast retrieval of information from multi-class feature representations.