I am trying to interpret some t-SNE results (using the Rtsne package in R). I used t-SNE on class probabilities from a multiclass model with 6 levels to try and visualize the relationships between the model's various class predictions. (If that doesn't make sense methodologically, please let me know, even though that isn't my primary question.)
In the plot below, the points are colored by the predicted classes. In general, I'd like to know how someone would interpret this overall structure, and more specifically, whether it's possible to interpret the middle of the "spokes" as where the model struggled to distinguish between classes. My guess is no, based on the fact that t-SNE doesn't preserve the global structure of the data (from here and here, answer 2). So can we only interpret the local structures within each cluster?
