From e1fec2fe24e4caf3e86ddba51e1d69c0beb4af23 Mon Sep 17 00:00:00 2001
From: Fredrik Bagge Carlson <baggepinnen@gmail.com>
Date: Thu, 14 Sep 2023 16:20:35 +0200
Subject: [PATCH] add exclusion zone option

---
 Project.toml         |  2 +-
 src/MatrixProfile.jl |  8 +++++---
 test/runtests.jl     | 11 +++++++++++
 3 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/Project.toml b/Project.toml
index cd87321..9ceb334 100644
--- a/Project.toml
+++ b/Project.toml
@@ -19,7 +19,7 @@ Distances = "0.7, 0.8, 0.9, 0.10"
 LoopVectorization = "0.7.4, 0.8, 0.9, 0.10, 0.11, 0.12"
 ProgressMeter = "1.2, 1.3"
 RecipesBase = "0.8, 1.0"
-SlidingDistancesBase = "0.2, 0.3"
+SlidingDistancesBase = "0.2, 0.3.5"
 julia = "1"
 
 [extras]
diff --git a/src/MatrixProfile.jl b/src/MatrixProfile.jl
index 1d882a4..733c102 100644
--- a/src/MatrixProfile.jl
+++ b/src/MatrixProfile.jl
@@ -23,13 +23,15 @@ struct Profile{TT,TP,QT}
 end
 
 """
-    profile = matrix_profile(T, m, [dist = ZEuclidean()]; showprogress=true)
+    profile = matrix_profile(T, m, [dist = ZEuclidean()]; showprogress=true, exclusion_zone = 0)
 
 Return the matrix profile and the profile indices of time series `T` with window length `m`. See fields `profile.P, profile.I`. You can also plot the profile. If `dist = ZEuclidean()` the STOMP algorithm will be used.
 
+- `exclusion_zone` denotes an integer number of samples around the trivial match to avoid. The paper suggests using `exclusion_zone = m ÷ 4`. This is likely most beneficial for time-series dominated by low frequencies.
+
 Reference: [Matrix profile II](https://www.cs.ucr.edu/~eamonn/STOMP_GPU_final_submission_camera_ready.pdf).
 """
-function matrix_profile(T::AbstractVector{<:Number}, m::Int; showprogress=true)
+function matrix_profile(T::AbstractVector{<:Number}, m::Int; showprogress=true, kwargs...)
     n   = lastlength(T)
     l   = n-m+1
     n > 2m+1 || throw(ArgumentError("Window length too long, maximum length is $((n+1)÷2)"))
@@ -47,7 +49,7 @@ function matrix_profile(T::AbstractVector{<:Number}, m::Int; showprogress=true)
             # The expression with fastmath appears to be both more accurate and faster than both muladd and fma
         end
         QT[1] = QT₀[i]
-        distance_profile!(D, ZEuclidean(), QT, μ, σ, m, i)
+        distance_profile!(D, ZEuclidean(), QT, μ, σ, m, i; kwargs...)
         update_min!(P, I, D, i)
         showprogress && i % 5 == 0 && next!(prog)
     end
diff --git a/test/runtests.jl b/test/runtests.jl
index 51f59e7..d8d100a 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -38,6 +38,17 @@ end
    @test m[1] < 1e-6
    @test m[2] == 51 || m[2] == 112
 
+   profileez = @inferred matrix_profile(T, length(y0), exclusion_zone=10)
+   Pez,Iez = profileez.P, profileez.I
+   @test_nowarn plot(profileez)
+   # plot(T, layout=2)
+   # plot!(P, sp=2)
+
+   m = findmin(Pez)
+   @test m[1] < 1e-6
+   @test m[2] == 51 || m[2] == 112
+   @test all(Pez .>= P .- 1e-12)
+
 
    # Test Euclidean between two series
    profile3 = @inferred matrix_profile(T, T, length(y0))