This thesis proposes approaches for multiscale event detection from social media data streams. The system architecture involves three phases: (1) data retrieval and preprocessing to filter noisy tweets, (2) data representation using text vectorization, entity extraction and time series analysis, and (3) document-pivot and feature-pivot clustering to detect event candidates. Clustering results are classified using features like topic distribution and social engagement to identify meaningful events. Wavelet analysis is also used to measure term similarity across different temporal and spatial scales.