3 ms·

This looks really cool. One question I have is about how much the synthetic data can protect privacy. For example, my company has geospatial event data from our

This looks really cool. One question I have is about how much the synthetic data can protect privacy. For example, my company has geospatial event data from our customers. We're very protective of customer identities, and wouldn't want to expose which cities our customers are in. If a model trained on our database notices that the "longitude" column marginal distribution has a spike around (just as an example) -71 degrees (longitude of Boston, where we're located), then presumably the synthetic data would also include a bunch of longitudes near -71 degrees? But there aren't that many cities at longitude -71 degrees, so even the marginal distribution of the synthetic longitudes would reveal something private about our data. Second question is whether y'all support geospatial data? Both in the sense of "the topology of latitudes and longitudes is not a plane" and "can the model be trained on databases which encode geometries as a single column?"

openquery 6y ago
That's a great question. I had to divert to my co-founder Damien who is spear-heading the research side of the company. The gist of it is that if the the original data has a spike around -71, you will indeed see a spike in the synthetic copy as well. What it boils down to under the hood is a decision on the value of a continuous degree of freedom between two pieces of information: - the information that you have a significant number of users located in Boston, and - the information that any given particular user is located in Boston. At a high level, we are taking the view that for your synthetic data to be realistic, it would need to spike around Boston if and only if most of your users are in Boston. This also means that you are not leaking information about any given individual user and that the behavior of the crowd is OK. Put more simply, if you have a single user located in Boston and all the others in, say, San Francisco, then your synthetic data should not end up having users in Boston at all. Currently we do not have any bespoke support for lat/lon data, beyond it being like any other float of course. It is planned for the next release though! So check back in a couple weeks and it'll be there

openquery 6y ago

That's a great question. I had to divert to my co-founder Damien who is spear-heading the research side of the company. The gist of it is that if the the original data has a spike around -71, you will indeed see a spike in the synthetic copy as well. What it boils down to under the hood is a decision on the value of a continuous degree of freedom between two pieces of information: - the information that you have a significant number of users located in Boston, and - the information that any given particular user is located in Boston. At a high level, we are taking the view that for your synthetic data to be realistic, it would need to spike around Boston if and only if most of your users are in Boston. This also means that you are not leaking information about any given individual user and that the behavior of the crowd is OK. Put more simply, if you have a single user located in Boston and all the others in, say, San Francisco, then your synthetic data should not end up having users in Boston at all. Currently we do not have any bespoke support for lat/lon data, beyond it being like any other float of course. It is planned for the next release though! So check back in a couple weeks and it'll be there