One problem we've had with data sets is that they don't have enough history. Even a very dense data set with billions of records will allow NuPIC to pick up yearly patterns unless there are years of background data. Sales transactions, for example, will have a yearly pattern, and subsampling or aggregating large data sets should allow the CLA to pick up those patterns. --------- Matt Taylor OS Community Flag-Bearer Numenta
On Fri, Dec 20, 2013 at 4:58 PM, Subutai Ahmad <[email protected]> wrote: > Hi Matt, > > Hmm, that’s a thought provoking question! > > Here’s one suggestion: predicting user clicks on websites. The task would be > to predict the pages a given user might click on next based on their past > sequence of clicks. We did some initial work on this before with one of the > largest news websites and got really good results. However it was never > commercialized. > > The dataset would consist of weblogs from a large e-commerce or news > website. Ideally we’d also have sanitized demographic information about each > user. > > Possible applications would be: 1) much better page recommendations for > users (which results in a user staying on a site longer), and 2) more > relevant ads and hence improved ad-clicks. > > It would be great to hear some other ideas as well. > > —Subutai > > > > On Wed, Dec 18, 2013 at 1:36 PM, Matt Barton <[email protected]> wrote: >> >> Hi all, >> >> >> >> I’ve been loosely following nupic for a while now. I have a question for >> the group: Are there any cool things someone has been wishing they could >> try, if only the right (large) data set was available? >> >> >> >> My company consults to most of the largest businesses around the world, >> including retailers, banks, airlines, and many others, and we can access all >> the data we ask for. Is there a killer application for nupic you’ve always >> wanted to try, if only you had access to a retailer’s full transaction log >> history or online search history, or an insurer’s full claims history, or an >> airline’s bookings and trip history? These datasets can be in the billions >> of rows... >> >> >> >> Best, >> >> Matt Barton @ Oliver Wyman >> >> >> _______________________________________________ >> nupic mailing list >> [email protected] >> http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org >> > > > _______________________________________________ > nupic mailing list > [email protected] > http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org > _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
