It is also sometimes possible to have some elements of a class marked and many unmarked instances which you know to contain only a small proportion of the class of interest.
In such a case, you can do two-class learning with your marked instances and all others. This is how most fraud models are developed, for instance. On Thu, Nov 25, 2010 at 9:11 AM, Isabel Drost <[email protected]> wrote: > On Thu, 25 Nov 2010 JAGANADH G <[email protected]> wrote: > > > Or is it enough to train with either of good or bad.? > > > > It will be something like train a person to identify 'sweet' by giving > > 'salt' as sample > > There are some domains where it may make sense to formulate a task as > one-class classification problem. E.g. looking at time series data one > might want to train a model to identify "normal" behaviour from > positive data only. > > Though it is possible to come up with algorithms for this so-called > one-class classification problem*, I am not aware of any implementation > in Mahout. > > > Isabel > > * For instance see "One-Class SVMs for Document Classification" by > Larry m. Manevits and Malik Yousef for some references and comparison. > >
