If you want relative error, you should model the log of the target variable. This is very commonly done with prices.
My beefs with SVD methods in general are a) they are often implemented without regularization b) they are typically used to model ratings instead of the desired target behavior c) LLR sparsified binary interaction matrices are easier to implement, typically perform as well or better with sufficient data and can be deployed using a search engine. On Wed, Feb 6, 2013 at 2:04 AM, Sean Owen <[email protected]> wrote: > Yes that would be valid in the sense that the neighborhood based approaches > are outputting a weighted average of prices here which is also a price. You > would have to think about which similarity metrics are meaningful though. > > The SVD has a perhaps undesirable behavior here. Because it treats the > squared error of each datum equally it will tend to pay too much attention > to large values. A 5% error produces 100 times more squared error when the > value is large. > > Ted I think this is why you say the SVD is bad for count-like data? > On Feb 6, 2013 8:28 AM, "万代豊" <[email protected]> wrote: > > > Hi > > I also have a similar question regarding result interpretation based on > how > > we provide data to recommender. > > Typcally , we provide rating data say in scale from 1-5 and get the > result > > in the same scale range.(and need to be consistent as Sean points out) > > > > If we assume the provided data with other dimension such as sales > > revenue,price or some other physical dimension and can we still use the > > recommender output as precdicted sales revenue or simulated possible > price > > in the same dimension scale? > > I'm assuming user-based recommendation here. > > > > Looks like SVD recommender is more suitable for this purpose. > > Apprecaite your advise. > > Yutaka > > > > > > > > 2013/2/5 Sean Owen <[email protected]> > > > > > You don't have to fix a scale. But your data needs to be consistent. > > > It wouldn't work to have users rate on a 1-5 scale one day, and 1-100 > > > tomorrow (unless you go back and normalize the old data to 1-100). > > > > > > On Mon, Feb 4, 2013 at 3:56 PM, Zia mel <[email protected]> > wrote: > > > > Hi , is there a necessity to have a fix rating scale while running > > > > recommendations or it can be dynamic based on the users' data ? > > > > > > > > Many Thanks > > > > > >
