replace ols(x,y) with x\y in highly computational methods such like lts, lms, and lta

Author: jbytecode

CHANGELOG.md

@@ -1,7 +1,7 @@
 # v0.11.2 (Upcoming Release)
 
 - Optional eps and maxiter parameters in iterateCSteps() in LTS.
-
+- Replace `ols(X[indices, :], y[indices])` with `X[...] \ y[...]` in highly computational ones such like LTS, LMS, and LTA. 
 
 # v0.11.1
 

src/lms.jl

@@ -6,7 +6,6 @@ export lms
 
 import ..Basis:
     RegressionSetting, @extractRegressionSetting, designMatrix, responseVector, applyColumns
-import ..OrdinaryLeastSquares: ols, predict, residuals, coef
 
 import Distributions: sample
 
@@ -77,8 +76,7 @@ function lms(X::AbstractMatrix{Float64}, y::AbstractVector{Float64}; iters = not
         try
             k = rand(kindices, 1)[1]
             sampledindices = sample(indices, k, replace = false)
-            olsreg = ols(X[sampledindices, :], y[sampledindices])
-            betas = coef(olsreg)
+            betas = X[sampledindices, :] \ y[sampledindices]
             origres = y .- X * betas
             res = sort(origres .^ 2.0)
             m2 = res[h]

src/lta.jl

@@ -6,7 +6,7 @@ export lta
 
 import ..Basis:
     RegressionSetting, @extractRegressionSetting, designMatrix, responseVector, applyColumns
-import ..OrdinaryLeastSquares: ols, predict, residuals, coef
+
 
 import Combinatorics: combinations
 import Distributions: sample
@@ -90,7 +90,7 @@ function lta(X::AbstractMatrix{Float64}, y::AbstractVector{Float64}; exact = fal
 
     function lta_cost(subsetindices::Array{Int,1})::Tuple{Float64,Vector{Float64}}
         try
-            betas = coef(ols(X[subsetindices, :], y[subsetindices]))
+            betas = X[subsetindices, :] \ y[subsetindices]
             res_abs = abs.(y .- X * betas)
             ordered_res = sort(res_abs)
             cost = sum(ordered_res[1:h])
@@ -102,7 +102,7 @@ function lta(X::AbstractMatrix{Float64}, y::AbstractVector{Float64}; exact = fal
 
     L = length(psubsets)
     bestobjective = Inf64 
-    bestbetas = coef(ols(X, y))
+    bestbetas = X \ y
     numberofitersunchanged = 0
     for i in 1:L
         objective, betas = lta_cost(psubsets[i])

src/lts.jl

@@ -4,7 +4,7 @@ export lts
 export iterateCSteps
 
 import ..Basis: RegressionSetting, @extractRegressionSetting, designMatrix, responseVector
-import ..OrdinaryLeastSquares: ols, coef, residuals, predict
+
 import Distributions: sample, mean
 
 """
@@ -49,8 +49,7 @@ function iterateCSteps(
     objective::Float64 = Inf64
     iter::Int = 0
     while iter < maxiter
-        olsreg = ols(X[subsetindices, :], y[subsetindices])
-        betas = coef(olsreg)
+        betas = X[subsetindices, :] \ y[subsetindices]
         res = y - X * betas
         sortedresindices = sortperm(abs.(res))
         subsetindices = sortedresindices[1:h]
@@ -82,8 +81,7 @@ function iterateCSteps(
     initialBetas::AbstractVector{Float64},
     h::Int; eps::Float64 = 0.01, maxiter::Int = 10000
 )
-    n, p = size(X)
-    #res = [y[i] - sum(X[i, :] .* initialBetas) for i = 1:n]
+    _, p = size(X)
     res = y - X * initialBetas
     sortedresindices = sortperm(abs.(res))
     subsetindices = sortedresindices[1:p]
@@ -174,10 +172,7 @@ function lts(X::AbstractMatrix{Float64}, y::AbstractVector{Float64}; iters=nothi
         end
     end
 
-
-    ltsreg = ols(X[besthsubset, :], y[besthsubset])
-    ltsbetas = coef(ltsreg)
-    #ltsres = [y[i] - sum(X[i, :] .* ltsbetas) for i = 1:n]
+    ltsbetas = X[besthsubset, :] \ y[besthsubset]
     ltsres = y - X * ltsbetas
     ltsS = sqrt(sum((ltsres .^ 2.0)[1:h]) / (h - p))
     ltsresmean = mean(ltsres[besthsubset])