Hacker News Reader

Comment by arkitaip | Hacker News Reader

As a HCI nerd I find this tech very interesting but I am skeptical about the extraordinary claim that EyeQuant - the technology developed and employed by Whitematter labs in this analysis - can, "predicts within seconds where users will look and what users will see" [1]. Where is the equally extraordinary evidence? Any kind of prediction of human behavior is very hard and subject to all kinds of exceptions. For instance, are the predictions applicable to all people wrt age, gender, culture, etc? What about something as basic as people whose language are Right-to-left (arabic, hebrew)? I haven't found any information in English that offers any scientific explanation how EyeQuant performs its predictions.

[1] http://eyequant.com/

View on HN

Hi Arkitaip,

glad you asked - (I'm one of co-founders of EyeQuant.)

In a nutshell, EyeQuant's predictions use a model of human attention that is based on hundreds of empirical eye-tracking studies, which were conducted at the Neurobiopsychology Lab at the University of Osnabrueck, Germany - also, we're cooperating with Caltech and USC.

The models only represent the cultural background of tracked subjects - it's strictly data-driven. Our current model is representative for western subjects; it can't predict, e. g., people who usually read Farsi. It also only predicts the first 3-5 seconds.

Here's the method: In the eye-tracking studies, we present subjects with websites from different categories and track their fixations - more importantly, we also analyze the statistical features of the fixated spots (around 50 of them to date): color contrasts, luminance, shapes and location, to name a few of the very basic ones.

In the next step we build models based on the weight of these features and evaluate them with other eye-tracking studies as a gold standard. One eye-tracking study typically predicts another one with 90-95% accuracy for the first few seconds (ROC). With EyeQuant, and after 5 years of foundational research, we achieve around 84% predictive accuracy - that's pretty close to what you'll get with an empirical study, but the results are available within seconds.

Let me know if you have more questions!

Thanks for providing more information about EyeQuant. One thing that I wonder about is how sensitive your system is wrt the customer's users. If I am running a gaming site where most users are men between ages 12-39, will you only use data from test subjects that share similar characteristics?

Reply on news.ycombinator.com