From: dblock <dblockdotorg@...>
Date: 2011-01-20T01:25:27+09:00
Subject: Re: Improving performance of hash math

> Use Hash's default value mechanism or use "|| 0" and kick out
> #has_key? calls.  I'd also rather use a proper Matrix or Vector class
> instead of a Hash.  NArray also seems like a good and fast option.

- has_key has almost no cost, made no difference
- my data is sparse, there're a lot of zero values - using an array
forces you to do N calculations (N is the max width of the array), an
Array was 3x slower, especially with a fast growing data set (haven't
tried NArray)

> Btw, why are you using an instance variable here?  That does not seem
> to make any sense.

- fixed

> Also, it seems that your calculation is flawed because you calculate
> the distance for keys which are present in both Hashes twice.  I'd
> rather

- re-read the core, you're wrong, the second half only adds values
that are not present in the first array

> require 'set'
>
> def euclidian_distance(lhs, rhs)
>   s = 0
>
>   (lhs.keys.to_set.merge(rhs.keys)).each do |k|
>     s += ( (rhs[k] || 0) - (lhs[k] || 0) ) ** 2
>   end
>
>   s
> end

- this is 3x slower because of the cost of to_set.merge (the code is
nicer though :)