Infinite Attribute Space

In Nieme, you never have to specify the full feature set. The features are data-driven: a feature does not exists until it has been seen in one example. At the user level, a feature can be any alpha numeric identifier. All computations in Nieme are done with sparse representations. When a new feature appears, the parameters of the learning machines are automatically extended to include this new feature. Classes are handled in a similar way. In a multiclass classification, if an example is presented with a new class, the learning machines just adapt their architecture and weights transparently.

This infinite attribute space model makes Nieme a very good candidate for tasks such as Automatic Features Generation and Structured Problems in general.

next