Figure 2

Depiction of the Non-Negative Matrix Factorization (NMF) stage of the proposed pipeline. We built a Waypoint Matrix W where rows encode users and columns encode towers, in analogy to document-term matrices. A cell in W denotes how frequently its corresponding user connects to its corresponding tower while commuting. The NMF decomposes W into two matrices that can be interpreted as the association between users and latent components (A), and towers and latent components (B). To interpret these latent components as modes of transportation, we used a Semi-Supervised approach to perform the NMF