Thresholds are not sorted due to disproportionate scores

[janfrancu]

Computing thresholds on short vectors (by default 300 values or less) may produce unsorted results due to floating point arithmetic errors.

In particular given these lables/scores data.txt, containing values with different magnitudes, the call to quantile at

EvalMetrics.jl/src/thresholds.jl

Lines 50 to 64 in 466ea78

	function thresholds(
	scores::RealVector,
	n::Int = length(scores);
	reduced::Bool = true,
	zerorecall::Bool = true,
	)
	N = reduced ? min(length(scores), n) : n
	thres = quantile(scores, range(0, 1, length = N))
	if zerorecall
	return vcat(thres, nextfloat(thres[end]))
	else
	return thres
	end
	end

without any additional argument produces thresholds

 0.3426349461078644                                                                                                   
 0.3456488847732544                                                                                                   
 0.3648797571659088                                                                                                   
 0.3659798502922058                                                                                                   
 0.3748885691165924                                                                                                   
 0.3797685503959656                                                                                                   
 0.38884931802749634                                                                                                  
 0.39463597536087036                                                                                                  
 0.39494413137435913                                                                                                  
 0.40267691016197205                                                                                                  
 0.4065932333469391                                                                                                   
 0.41210252046585083                                                                                                  
 0.41331562399864197                                                                                                  
 0.4157656729221344                                                                                                   
 0.43393915891647333                                                                                                  
 0.4365665018558502                                                                                                   
 0.43950754404067993                                                                                                  
 0.4522964358329773                                                                                                   
 0.454873651266098                                                                                                    
 0.46290522813796997                                                                                                  
 0.490060031414032                                                                                                    
 0.5430693030357361                                                                                                   
 0.5849942564964294                                                                                                   
 0.6120826601982117                                                                                                   
 0.6970598101615906                                                                                                   
 0.8463529348373413                                                                                                   
 0.9729131460189819                                                                                                   
 1.3229070901870714   
-----------------------------------------                                                                                                
 1.083465853229067e8                                                                                  
 1.08346584e8                                                                                                         
 1.10690304e8                                                                                                         
 1.10690304e8                                                                                                         
 1.16699368e8                                                                                                         
 1.16699368e8                                                                                                         
 1.2348864e8                                                                                                          
 1.2348864e8                                                                                                          
 1.54131584e8                                                                                                         
 1.54131584e8                                                                                                         
 1.67473424e8                                                                                                         
 1.67473424e8                                                                                                         
 5.70319872e9                                                                                                         
 5.70319872e9                                                                                                         
 1.5408278528e10                                                                                                      
 1.5408278528e10                                                                                                      
 3.36217088e10
 3.3621708800000824e10
 1.49447049216e11
 1.49447049216e11
 2.8364769329152e16
 2.8364769329152e16

which are sorted with the exception of one element differing from the next by machine eps. Similar issue has been already reported in StatsBase pkg JuliaStats/StatsBase.jl#164 and solved by reordering of arithmetic operations in case of values with similar magnitude. In the case of different magnitudes the problem seems to be even more edge case and may be better to solve it on this end rather than rely on changes to base Julia.

[nalimilan]

I tried reproducing the problem and I couldn't, on both Julia 1.6 and 1.9. What am I doing differently?

scores = [2.836477e16, 2.836477e16, 1.4944705e11, 1.4944705e11, 3.3621709e10, 3.3621709e10, 1.5408279e10, 1.5408279e10, 5.7031987e9, 5.7031987e9, 1.6747342e8, 1.6747342e8, 1.5413158e8, 1.5413158e8, 1.2348864e8, 1.2348864e8, 1.1669937e8, 1.1669937e8, 1.106903e8, 1.106903e8, 1.08346584e8, 1.08346584e8, 1.3229071, 0.97291315, 0.84635293, 0.6970598, 0.61208266, 0.58499426, 0.5430693, 0.49006003, 0.46290523, 0.45487365, 0.45229644, 0.43950754, 0.4365665, 0.43393916, 0.41576567, 0.41331562, 0.41210252, 0.40659323, 0.4026769, 0.39494413, 0.39463598, 0.38884932, 0.37976855, 0.37488857, 0.36597985, 0.36487976, 0.34564888, 0.34263495]

thresholds(scores)
quantile(scores, range(0, 1, length=length(scores)))