A highlight from Your Mouse Reveals Your Gender and Age

Data Skeptic


Interview. I'm Louis leiber and I'm currently assistant Professor of computer science in the university of Luxembourg, in Europe. So I basically work at the intersection of machine learning and human computer interaction. Local quickly, I feel which is called computational interaction where I basically build computational models to try to explain or predict user behavior. So typically we are either adapt or create new machine learning models to account for any behavioral traits about the user. Let's say how they pay attention to these plays or how they move the mouse or how they use the eye movements to fixate on something kind of things. So if you build a good model of a user that means it's sort of representative of and predictive of the user's behavior, you're not trying to replace the user. What's the use of the model? So the main application of user modeling is simulation. Imagine that you need to recruit representative user sample for measuring whatever. Then you have to spend time designing this experiment, recruiting people reaching out to them, allocated some budget for paying them and so on. So instead of recruiting, let's say, 1000 people for that and you can regret maybe ten, run some statistical tests, do some machine learning on data that you can collect from them. And then you can create a user model that you can try to then to infer. Some trades or behavioral trades that could be extended to a larger user group. Of course, I mean, it's just an example, typically you would need fairly large representative dataset if you run. If you want to really do some production level machine learning code. But for research purposes, sometimes even less than 100 user is really more like enough to start drawing compressions about your user population. Yeah, it just depends on the effect size you're trying to measure. With that in mind, can you expand on what some of the behaviors are? It depends on the study or the kind of analysis that you want to do. For instance, something that I have been doing for more than a decade is to try to infer how people allocate attention on different screens based on how they move them out. And why this isn't interesting because then you don't need to really install any webcam any eye tracker. So you can use the mouse as a proxy of you should behavior and see depending on how they move how fast they move, how reach out to targets or how they click on something. So it's not only about the click itself, but how the process or what happened before the click what is in interesting and this kind of user models can tell us a lot about, for instance, how the layout or the user interface is designed and how can we change things or what happened if we move, I don't know what button from the top left corner to the bottom right corner, but kind of behavior this will enable or if this will facilitate people find an information quicker, these are things that you can measure for free basically by Lara and running larger scale studies. If you really want to pay attention to how people behave on the wave for instance, then the mouse has been shown to be a good proxy of the eye gaze. I mean, not for every single thing, but for most of the tasks that we do online, the mouse is a very reasonable proxy of user behavior. Well, I'm glad you made the comparison to eye tracking or gaze measurements. Obviously, we're going to talk mostly about the mouse today, but in terms of making that compare and contrast for listeners who don't really know much about the scholarship of eye tracking, is that a big thing in how highly is it regarded? I guess I asked because I think of my own eyes and it's not really a precision instrument, right? Just because I look at an ad doesn't mean I'm interested, sometimes I was distracted or it seems like a very noisy dataset. How reliable is eye tracking? Well, I track it is actually one of the earliest measuring instruments to analyze and investigate you should behavior. Typically I'm in New York computer interaction. We are interested in understanding how people pay attention to things or how things are arranged on the screen. And the eye tracking is an essential device in most ACI labs today. And it's not really noisy. Actually, the mouse is way more noisy than the eye. So yeah, I'm for sure we can go through that later. But I can tell you that by looking at how people experience or how people look at the content and how this is a range of the screen is something that people in marketing and in neuroscience has been using for decades. So I would say that it's pretty thunder. And while the use of this in human building interaction context, so I guess that's not all your audience will be familiar with hci when you can compute interaction literature by just to let you know that the eye tracking is a measuring device is really, really popular in NCI. So based on your research and I guess also just things you've read and looked into. Do you think that the mouse approach can be a full proxy if for some reason eye tracking or was too expensive or just not available for my project, but I do have mouse data. Can I get just the same rich amount of information as data center? Is there like a fidelity loss by looking at only the mouse? Yeah, well, it depends on the task. Of course, I mean, I can not give you a 100% accurate answer about that. But for instance, on web search, how people search engines are displaying information, these typical search engine page, this page with time purple snippets that you can click on. This is a highly structured information.

Coming up next