emgrasmeder / clj-knn Goto Github PK

The current, most basic implementation will return a Hamming distance of 1 between

{:abc {:hello "world"
       :goodnight "moon"
       :farewell "auf wiedersehen, goodbye!"}}

{:abc {:hello "darkness my old friend"
       :goodnight "moon"
       :farewell "farewell"}}

because the :abc key's value is not equal. By not flattening the input, we lose fidelity and miss the one similarity that does exist in these maps.

One proposal is to handle flattened maps without adding a weight, thus comparing the following maps instead:

{:abc:hello "world"
 :abc:goodnight "moon"
 :abc:farewell "auf wiedersehen, goodbye!"}}

{:abc:hello "darkness my old friend"
 :abc:goodnight "moon"
 :abc:farewell "farewell"}}

Right now the Hamming Distance function does a simple (sort (vals vect)) which can't differentiate between {:a 1 :b 2 :c 3} and {:a 1 :b 2 :d 3}.

The two vectors above should have a Hamming distance of either 1 or 2, but definitely not 0, which it currently calculates.

Predictive power could be increased if each feature were assigned weight w ∈ [0,1].

Dropping features could be weighted 0, more important features would be given weight 1. By default, every feature should be weighted as 1