Say I'm using scikits implementation of Local Outlier Factor with euclidean distance being used by the reachability function. My input features are magnitudes apart, so is it advisable that I normalize my input features before training the function even though there might be several separate clusters.
-
$\begingroup$ This research might be helpful to find an answer. $\endgroup$ – Arunoda Samarasinghe Mar 22 '20 at 17:00
That would be the recommended approach since Local Outlier Factors are based on Nearest Neighbor approach which is a similarity based algorithm. Normalization is recommended for most cases where similarity measures are used, unless you'd want high magnitude features to dominate the distance calculation. If you have a mixture of continuous, ordinal and binary variables in there, you may want to evaluate a few different scales for them.