Networks are everywhere: social networks, transportation systems, biological interactions. To make sense of them, researchers often look for communities --- groups of elements that are more strongly connected to one another than to the rest of the network.
But these networks constantly evolve over time, and the interactions they record come in different forms. Some happen in an instant, some extend over a period of time, others still carry a delay between their two endpoints.
Beside, detecting communities that form, transform, and dissolve through such diverse data remains a challenge. It is one that existing methods typically address by transforming the data, slicing continuous time into snapshots and flattening interaction types, at the cost of discarding the very information the analysis aimed to capture.

This thesis proposes a new approach that identifies dynamic communities by pinpointing the exact moments when each member joins or leaves them, without any artificial time slicing. It relies on a new mathematical measure, Longitudinal Modularity, and an efficient algorithm to optimize it. Designed to preserve the data's original features --- exact timing, weights, directions, multiple types of nodes, and the diverse temporal nature of interactions themselves --- the framework avoids the destructive transformations that conventional methods require.
Without claiming to cover every scenario, it is validated on data from social interactions, voting systems, urban mobility, and online social platforms. This opens new perspectives for analyzing the dynamics of complex systems.