Having this information about what order the eye travels in typically can help #CRO guys prioritize what we want to communicate to the visitor in order to persuade them. This is great content keep this kind of stuff coming!

* What clustering algorithm are you using and how you are picking the number of clusters? Since your visualisation is based on the clusters, getting this right will be critical for success. Looking at the cluster shapes I’d guess either k-means or hierarchical with ward’s distance. But how are you picking the number of clusters?

* I think you need to be very careful when you draw cluster convex hulls – this has the tendency to make clusters look much more distinct than they really are (probably because of the gestalt principle of connectedness). Take a sample from a bivariate normal, cluster it and then display the clusters with convex hulls – they look really distinct but you know they’re not. I’d want to explore rectangular bounding boxes since they have a natural correspondence to the generally rectangular underlying page elements.

* I’d be interested to see alternatives without the clustering – e.g. what if you use a rectangular grid? If you made the grid cells smaller and smaller (maybe with some smoothing) you could end up with something like a vector field showing average movement around the page.

* I think you’re wasting colour by mapping it to group membership – that’s obvious from the structure of the clusters anyway. Why not map colour to number/density of views?

Anyway, just a few ideas off the top of my head. Send me an email and I can hook you up with some experts I know who are starting to get interested in eye tracking data.

Great to hear from you! We’re looking into building a video based off of this where the arrows grow over time. This will allow us to aggregate visual patterns over time, adding in the temporal element that is so painfully missing from heatmaps and many other static visualizations.