From 279e64e81bd6203b01d940373774004530e4a300 Mon Sep 17 00:00:00 2001
From: Joakim <joakim.j.linja@jyu.fi>
Date: Fri, 5 May 2023 20:27:42 +0300
Subject: [PATCH] Uploaded Feature Importance Detector (FID) implementations
(Python + and Matlab) related to arXiv:2303.09453
---
fid.m | 74 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
fid.py | 69 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 143 insertions(+)
create mode 100644 fid.m
create mode 100644 fid.py
diff --git a/fid.m b/fid.m
new file mode 100644
index 0000000..cb4600d
--- /dev/null
+++ b/fid.m
@@ -0,0 +1,74 @@
+function [rX, rY, f2] = fid(X, Y, cutoff, plotToggle)
+ % Feature Importance Detector, published in the paper:
+ % "Knowledge Discovery from Atomic Structures using Feature Importances" (arXiv:2303.09453).
+ % Description:
+ % Performs FID feature selection.
+ %
+ % Inputs:
+ % X - Index vector.
+ % Y - Feature score vector. Assumed to be in descending
+ % order.
+ % cutoff - The hyperparameter of FID. Range [0,1]. Determines the number
+ % of features to be removed.
+ % plotToggle - Optional boolean input, default is false.
+ %
+ % Ouputs:
+ % rX - Feature selected X
+ % rY - Feature selected Y
+ % f2 - Boolean vector, used to from rX and rY out of X and Y.
+ %
+
+ if exist('plotToggle','var')
+ plotT = plotToggle;
+ else
+ plotT = false;
+ end
+
+ fun = @(a,b,c,d,x) -a./(b*x+c)+d;
+
+ x1 = X(1); y1 = Y(1);
+ Y1 = zeros(size(Y));
+
+ if plotT
+ figure, plot(X,Y,'LineWidth',2);
+ title('FID: Input')
+ end
+
+ Y1(2:end) = (Y(2:end,1)-y1)./(X(2:end,1)-x1);
+ Y1(1) = Y1(2);
+
+ if plotT
+ figure, plot(X,Y1,'LineWidth',2);
+ title('FID: Rolling slope')
+ end
+
+ fopt = fitoptions('Method','NonLinearLeastSquares','Lower',[0 -Inf -Inf -Inf],'Upper',[Inf Inf Inf Inf],'StartPoint',[1 10 10 0]);
+ if plotT
+ disp(fopt)
+ end
+
+ fitobj = fit(X,Y1,fun,fopt);
+
+ Y2 = fitobj(X);
+
+ if plotT
+ figure, plot(X,Y2,'LineWidth',2);
+ title('FID: Curve fit')
+ end
+
+ dY2 = Y2(end)-Y2(1);
+ f2 = Y2 < Y2(end) - dY2*cutoff;
+
+ rX = X(f2);
+ rY = Y(f2);
+
+ if plotT
+ nf2 = ~f2;
+ figure, plot(X,Y,'LineWidth',1);
+ hold on;
+ nrX = X(nf2);
+ nrY = Y(nf2);
+ plot(nrX,nrY,'LineWidth',2);
+ title('FID: Result')
+ end
+end
\ No newline at end of file
diff --git a/fid.py b/fid.py
new file mode 100644
index 0000000..03ca8ae
--- /dev/null
+++ b/fid.py
@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Feature Importance Detector presented in
+Knowledge Discovery from Atomic Structures using Feature Importances (arXiv:2303.09453)
+
+@author: Joakim Linja
+"""
+import numpy as np
+import scipy as sp
+
+def fid(X,Y,cutoff = 0.01):
+ """
+ 1. Bound rolling slope
+ 2. Curve fit a/(b*x+c)+d
+ 3. Remove features based on cutoff-defined percentage of total variance
+
+ Parameters
+ ----------
+
+ X: numpy.ndarray
+ Feature score index array in ascending numerical order.
+
+ Y: numpy.ndarray
+ Feature score array. Assumed to be sorted to descending order.
+
+ cutoff: float, optional, [0,1]
+ Cutoff value to be used in the feature selection.
+
+ Default: 0.01
+
+
+ Returns
+ -------
+
+ rX: numpy.matrix
+ Feature selected X.
+
+ rY: numpy.matrix
+ Feature selected Y.
+
+ f2: numpy.ndarray
+ Boolean indexing array, the feature selection result.
+
+ """
+ try:
+ def f(x,a,b,c,d):
+ return -a/(b*x+c)+d
+
+ x0 = X[0]; y0 = Y[0]
+ Y1 = np.zeros(len(Y))
+ for i in range(1,len(X)):
+ Y1[i] = (Y[i]-y0)/(X[i]-x0)
+ Y1[0] = Y1[1]
+
+
+ popt, pcov = sp.optimize.curve_fit(f,X,Y1,p0=[1,10,10,0],
+ bounds=([0,-np.inf,-np.inf,-np.inf],[np.inf,np.inf,np.inf,np.inf]))
+ Y2 = f(X,*popt)
+
+ dY2 = Y2[-1]-Y2[0]
+ f2 = Y2 < Y2[-1] - dY2*cutoff
+
+ rX = X[f2]
+ rY = Y[f2]
+
+ return rX, rY, f2
+ except RuntimeError as e:
+ raise e
--
GitLab