LogFAQs > #923777056

LurkerFAQs, Active DB, DB1, DB2, DB3, DB4, Database 5 ( 01.01.2019-12.31.2019 ), DB6, DB7, DB8, DB9, DB10, DB11, DB12, Clear

Topic List	Page List: 1
Topic	Bill proposed to make tech companies disclose worth of user data
s0nicfan 06/24/19 1:10:44 PM #26:	Balrog0 posted... wouldn't it probably just be like... idk... we have x profiles that we get y data from that gets us z in revenue so each profile is worth epsilon dollars or something? I imagine it would be significantly more complicated than that. You have the raw data, which is hypothetically weighted based on importance or relevance to their algorithms. on a per-feature basis (so one piece of data in a profile gets you 50% accuracy in ML predictions, but another piece only adds 1% accuracy). Then you have to determine for your services if the source of data is equally valuable: if you sell bread is the value of a child the same as the value of a mom, for example. Then you have to determine to what extent your algorithms have impacted your profits to determine what percentage of those profits can be attributed to the data so you can work backwards to value the data itself. Then you have to decide how staleness of data depreciates the "market value" of the data itself... there's no clear indicator that month old data is worth as much as fresh data, and that value likely depends on the data, the service it supports, and the industry that service is associated with. There's also the fact that the value of data changes with quantity: a single point of data is worthless whereas ten million points of data is worth a lot, but each point individually is still basically useless... it's the aggregate only that has use. Which also depends on the algorithm: a support vector machine requires far less data to train than a convolutional neural net, but also generally produces lower accuracy predictions... the effectiveness of the algorithm is again a factor of quantity of data which ties back to the value of the data as a factor of profits. A bad algorithm producing bad results has no value so... does the data? And I expect the value to vary wildly depending on the industry. And honestly this is just a top-of-my-head estimate of the complexity of the problem. --- "History Is Much Like An Endless Waltz. The Three Beats Of War, Peace And Revolution Continue On Forever." - Gundam Wing: Endless Waltz <cite>s0nicfan posted [p:923777056] in Bill propo... [t:77815366]...</cite> <quote><cite>Balrog0 posted...</cite> <quote>wouldn't it probably just be like... idk... we have x profiles that we get y data from that gets us z in revenue so each profile is worth epsilon dollars or something?</quote> I imagine it would be significantly more complicated than that. You have the raw data, which is hypothetically weighted based on importance or relevance to their algorithms. on a per-feature basis (so one piece of data in a profile gets you 50% accuracy in ML predictions, but another piece only adds 1% accuracy). Then you have to determine for your services if the source of data is equally valuable: if you sell bread is the value of a child the same as the value of a mom, for example. Then you have to determine to what extent your algorithms have impacted your profits to determine what percentage of those profits can be attributed to the data so you can work backwards to value the data itself. Then you have to decide how staleness of data depreciates the "market value" of the data itself... there's no clear indicator that month old data is worth as much as fresh data, and that value likely depends on the data, the service it supports, and the industry that service is associated with. There's also the fact that the value of data changes with quantity: a single point of data is worthless whereas ten million points of data is worth a lot, but each point individually is still basically useless... it's the aggregate only that has use. Which also depends on the algorithm: a support vector machine requires far less data to train than a convolutional neural net, but also generally produces lower accuracy predictions... the effectiveness of the algorithm is again a factor of quantity of data which ties back to the value of the data as a factor of profits. A bad algorithm producing bad results has no value so... does the data? And I expect the value to vary wildly depending on the industry. And honestly this is just a top-of-my-head estimate of the complexity of the problem. --- "History Is Much Like An Endless Waltz. The Three Beats Of War, Peace And Revolution Continue On Forever." - Gundam Wing: Endless Waltz</quote> ... Copied to Clipboard!
Topic List	Page List: 1