Figure 2From: Activism via attention: interpretable spatiotemporal learning to forecast protest activitiesOverview of our proposed ActAttn architecture. It incorporates hierarchical attentional networks where the top level (a) differentiates the intra-region and inter-region importance, and the second level (b) identifies the hub regions. The temporal dependency of time-varying features in both intra- and inter-regions are modeled using LSTM (c), with sparse feature learning using Group Lasso (d)Back to article page