KNN-Regression/Classification should allow "self" points

[bwerth]

In the prediction of the kNN-Model it is currently disallowed to utilize data points with zero distance to the query point, which is not only counter intuitive but might lead to worse prediction results (especially in cases where the features are de facto ordinal/integers and zero distances are common). The inclusion of zero-distance points should at least be optional. This however requires an alteration in the way weights are assigned to neighboring points (currently 1/distance which would cause division-by-zero exceptions).

[gkronber]

bwerth has already prepared a document describing the necessary steps for implementation and will gladly help.

[msemenki]

r16408: Add for KNN-Regression/Classification ability to utilize data points with zero distance to the query point. Alteration in the way weights are assigned to neighboring points (to except division-by-zero).

[gkronber]

r16488: added SelfMatch parameters in the AfterDeserialization hook (for loading files stored with the old version)

[gkronber]

r16489: merged changes from trunk to support testing with trunk

[gkronber]

r16490: added a comment on the handling of weights in the self-matching case

[gkronber]

r16491: merged r16408, r16488, r16490 from branch to trunk (manually)

[gkronber]

Thank you @msemenki. Good work. I only added the parameters in the AfterDeserialization hooks (see r16488). This is necessary so that we can open and re-run stored kNN experiments.

[gkronber]

r16493: deleted branch after changes have been merged to trunk

[gkronber]

This has been closed unintended. Changes still need to be merged to stable.

[gkronber]

r17164: merged r16491 from trunk to stable