The increasing polarization of online political discourse calls for computational tools that are able to automatically detect and monitor ideological divides in social media . Here, we introduce a minimally supervised method that directly leverages the network structure of online discussion forums, specifically Reddit, to detect polarized concepts . We model polarization along the dimensions of agenda setting and framing, drawing upon insights from moral psychology . The architecture we propose combines graph neural networks with structured sparsity learning and results in representations for concepts and subreddits that capture phenomena such as ideological radicalization and subreddit hijacking . We also create a new dataset of political discourse spanning 12 years and covering more than 600 online groups with different ideologies.