Good stuff! I wonder if you could somehow classify the conversation trees using depth or branching factor or whatever to predict and/or discover hot topics. Like you said, AMAs tend to have a pretty obvious structure but maybe that could be extended to discover "controversial" threads -- or even people!
For those that are curious, what happens under the covers is a random sampling of the image followed by a conversion to LAB color space (even better than YUV in terms of human perception) and finally some K-means clustering.