Help on Optimizing Attribute weights for k-NN modeling
I'm having trouble doing something that seems pretty basic:
I have a set of around 50000 examples with about 20 features.
I'd like to find good weights for those features that optimize k-NN predictions.
Intuitively (and this may be wrong), the most natural way to use attribute weights to improve a k-NN prediction model is to scale the (normalized) data by the weight of the attribute. This way two points are considered further apart in feature space if they differ on a more important attribute than if they differ on a less important one.
Problem: I don't see how to do this in Rapid Miner because "Optimize Weights" processes do not provide direct access to the weight vector being tested, and without direct access to the weight vector being tested, I don't see how to use those weights to affect the clustering performance in the sub-process.
Unfortunately, since my data uses numeric labels, there are not many weight-generation schemes available for the purpose of making some dimensions more important than others in my predictive clustering efforts.
Any help is appreciated!
I have a set of around 50000 examples with about 20 features.
I'd like to find good weights for those features that optimize k-NN predictions.
Intuitively (and this may be wrong), the most natural way to use attribute weights to improve a k-NN prediction model is to scale the (normalized) data by the weight of the attribute. This way two points are considered further apart in feature space if they differ on a more important attribute than if they differ on a less important one.
Problem: I don't see how to do this in Rapid Miner because "Optimize Weights" processes do not provide direct access to the weight vector being tested, and without direct access to the weight vector being tested, I don't see how to use those weights to affect the clustering performance in the sub-process.
Unfortunately, since my data uses numeric labels, there are not many weight-generation schemes available for the purpose of making some dimensions more important than others in my predictive clustering efforts.
Any help is appreciated!