I wrote a meta classifier which instead of aggregating various models on the whole data (like bagging, boosting, random forest), it splits the input space into some disjoint regions and builds a base model for each region.
The reasons why one would want to try this approach would be:
- remove all cross-regional interferences, for example build one model for each US state, thus letting the model to fit only on data from a specific state
- for computational reasons, when too much data is available and there is such an input variable which can partition the whole space in smaller regions
The input space is split according to a disjoint predicate set.
Because the input space regions are disjoint, at prediction time, for each input data point, a proper base model would be selected for fitting the results.
The question is how could be named this procedure. My expectation is that there is some theory and this kind of meta algorithm has already a name. If not, what would be a proper name for it?