Sensing and monitoring information diffusion in online social networks is a complex problem of prominent importance, typically requiring significant sensing resources to address it properly. In this paper, we propose an inference approach for an information diffusion process where information is considered to belong to different classes, characterized by different spreading dynamics and possibly different topical content. Our framework utilizes social network analysis metrics in order to reduce the sensing resources that would be required in an otherwise exhaustive approach, while employing statistical learning and probabilistic inference for maintaining the accuracy of information tracking, whenever needed. The proposed framework defines an edge coloring scheme, based on which it is possible to keep track of information diffusion. We assume that the latter spreads according to various biased random walks that represent the dynamics of the considered classes of information. We have employed learning for the inference of those cases where backtracking leads to multiple potential choices for information paths. We demonstrate the operation and efficacy of our approach in characteristic online social networks, such as distributed wireless (spatial) and scale-free (relational) topologies, and draw conclusions on the impact of topology on information spreading. Finally, we discuss the emerging trends applicable for each topology and provide broader guidelines on the suitability of the proposed information diffusion inference scheme for each network.