The adPredictor paper has a good example of how to do this. See http://research.microsoft.com/apps/pubs/default.aspx?id=122779
On Sat, Feb 15, 2014 at 6:38 PM, Pat Ferrel <[email protected]> wrote: > I’m still unclear about how to apply TS to dithering. TS is usually talked > about in conjunction with a feedback loop, which make the examples seem > more like Multi-armed Bandit examples. > > Are you suggesting some feedback in recommender ranking or just using the > same distribution assumptions used in TS? > > On Feb 8, 2014, at 12:13 PM, Ted Dunning <[email protected]> wrote: > > Thompson sampling doesn't require time other than a sense of what do we now > know. It really is just a correct form for dithering that uses our current > knowledge. > > For a worked out version of Thompson sampling with ranking, see this blog: > > http://tdunning.blogspot.com/2013/04/learning-to-rank-in-very-bayesian-way.html > > The reason that we aren't adding this like cross-rec and other things is > that "we" have full-time jobs, mostly. Suneel is full-time on Mahout, but > the rest are not. You seem more active than most. > > > > > > On Sat, Feb 8, 2014 at 8:50 AM, Pat Ferrel <[email protected]> wrote: > > > Didn’t mean to imply I had historical view data—yet. > > > > The Thompson sampling ‘trick’ looks useful for auto converging to the > best > > of A/B versions and a replacement for dithering. Below you are proposing > > another case to replace dithering—this time on a list of popular items? > > Dithering works on anything you can rank but Thompson Sampling usually > > implies a time dimension. The initial guess, first Thompson sample, could > > be thought of as a form of dithering I suppose? Haven’t looked at the > math > > but it wouldn’t surprise me to find they are very similar things. > > > > While we are talking about it, why aren’t we adding things like > > cross-reccomendations, dithering, popularity, and other generally useful > > techniques into the Mahout recommenders? All the data is there to do > these > > things, and they could be packaged in the same Mahout Jobs. They seem to > be > > languishing a bit while technology and the art of recommendations moves > on. > > > > If we add temporal data to preference data a bunch of new features come > to > > mind, like hot lists or asymmetric train/query preference history. > > > > On Feb 6, 2014, at 9:43 PM, Ted Dunning <[email protected]> wrote: > > > > One way to deal with that is to build a model that predicts the ultimate > > number of views/plays/purchases for the item based on history so far. > > > > If this model can be made Bayesian enough to sample from the posterior > > distribution of total popularity, then you can use the Thomson sampling > > trick and sort by sampled total views rather than estimated total views. > > That will give uncertain items (typically new ones) a chance to be shown > > in the ratings without flooding the list with newcomers. > > > > Sent from my iPhone > > > >> On Feb 7, 2014, at 3:38, Pat Ferrel <[email protected]> wrote: > >> > >> The particular thing I’m looking at now is how to rank a list of items > > by some measure of popularity when you don’t have a velocity. There is an > > introduction date though so another way to look at popularity might be to > > decay it with something like e^-t where t is it’s age. You can see the > > decay in the views histogram > > > > > >
