-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfRunBoLassoClassifiers.m~
137 lines (122 loc) · 5.35 KB
/
fRunBoLassoClassifiers.m~
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
function out = fRunBoLassoClassifiers(X,Y)
%% options
ntrees = 75;
nboots = 129;
nxvalLasso = 3;
options = statset('UseParallel','always');
Loptions = glmnetSet;
Loptions.maxit = 500;
%% preprocessing
X = fCenterSphereData(X')';
N = length(Y);
D = size(X,2);
[OutlierNDX] = fkNNprototyping(X,X,round(sqrt(N)),'linear');
Info.Outliers.N = sum(OutlierNDX);
Info.Outliers.OutlierNDX = OutlierNDX;
Info.Outliers.KeepedData = ~OutlierNDX;
X = X(~OutlierNDX,:);
Y = Y(~OutlierNDX,:);
N = length(Y);
NBlocks = min(21,max(round(sqrt(N)),4));
[RBI,NBlocks] = fGenXvalBlockIndex(N,NBlocks);
BoLasso.coeffs = [];
%% Correct logs
Lasso.correctlog = [zeros(N,1) (1:N)'];
BoLasso.GLM.correctlog = [zeros(N,1) (1:N)'];
BoLasso.LDA.correctlog = [zeros(N,1) (1:N)'];
BoLasso.QDA.correctlog = [zeros(N,1) (1:N)'];
BoLasso.SVM.correctlog = [zeros(N,1) (1:N)'];
SVM.correctlog = [zeros(N,1) (1:N)'];
RF.correctlog = [zeros(N,1) (1:N)'];
Info.Xval.NBlocks = NBlocks;
Info.Xval.Index = RBI;
for xval = 1:NBlocks
%% parse
trIDX = RBI~=xval;
teIDX = RBI==xval;
Xtr = X(trIDX,:);Xte = X(teIDX,:);
Ytr = Y(trIDX);Yte = Y(teIDX);
Info.Xval.Tr(xval).TrIDX = trIDX;
Info.Xval.Te(xval).TeIDX = teIDX;
%% Lasso
Lassofit = cvglmnet(Xtr,Ytr,nxvalLasso,[],'response','binomial',Loptions,0);
Ypred = glmnetPredict(Lassofit.glmnet_object,'response',Xte,Lassofit.lambda_min);
Lasso.mse(xval) = mean(((Yte-1)-Ypred).^2,1);
Lasso.err(xval) = mean((Yte-1)~=round(Ypred),1);
Lasso.correctlog(teIDX,1) = (Yte-1)==round(Ypred);
Lasso.YteYp(xval).YteYp = [Yte Ypred];
%% BoLasso
for ii = 1:5
[BoLasso.Vars(xval).SelectedVars,...
BoLasso.Vars(xval).Boots,...
BoLasso.Vars(xval).coeffsBestL,...
BoLasso.Vars(xval).coeffarrays,...
BoLasso.Vars(xval).BestL,...
BoLasso.Vars(xval).MSE] = fBootLogisticLasso(Xtr,Ytr,nboots,nxvalLasso);
xvsum = sum(BoLasso.Vars(xval).SelectedVars);
if xvsum > 0
break;
else
disp('BoLasso failed... Retrying');
end
end %
SelectedVars = logical(BoLasso.Vars(xval).SelectedVars);
Ztr = Xtr(:,SelectedVars);
Zte = Xte(:,SelectedVars);
% Refit and eval
% eval
fit = glmfit(Ztr,Ytr-1,'binomial');
BoLasso.GLM.glmfit(xval).fit = fit;
Ypred = glmval(fit,Zte,'logit');
BoLasso.GLM.mse(xval) = mean(((Yte-1)-Ypred).^2,1); % mse
BoLasso.GLM.err(xval) = mean((Yte-1)~=round(Ypred),1); % perc err 0/1 loss
BoLasso.GLM.correctlog(teIDX,1) = (Yte-1)==round(Ypred);
BoLasso.GLM.Posterior(xval).Posterior = Ypred;
BoLasso.GLM.YteYp(xval).YteYp = [Yte Ypred];
% lDA
[Ypred,~,Ypost] = classify(Zte,Ztr,Ytr,'linear');
BoLasso.LDA.Posterior(xval).Posterior = Ypost;
BoLasso.LDA.mse(xval) = mean(((Yte-1)-Ypost(:,2)).^2,1); % mse
BoLasso.LDA.err(xval) = mean(Yte~=Ypred);
BoLasso.LDA.correctlog(teIDX,1) = Yte==Ypred;
BoLasso.LDA.YteYp(xval).YteYp = [Yte Ypost(:,2)];
% QDA
[Ypred,~,Ypost] = classify(Zte,Ztr,Ytr,'quadratic');
BoLasso.QDA.Posterior(xval).Posterior = Ypost;
BoLasso.QDA.mse(xval) = mean(((Yte-1)-Ypost(:,2)).^2,1); % mse
BoLasso.QDA.err(xval) = mean(Yte~=Ypred);
BoLasso.QDA.correctlog(teIDX,1) = Yte==Ypred;
BoLasso.QDA.YteYp(xval).YteYp = [Yte Ypost(:,2)];
% SVM
Model = train(Ytr,sparse(Ztr),'-s 1 -q');
[Ypred,~,Ypost] = predict(Yte,sparse(Zte),Model,'-b -q');
BoLasso.SVM.Posterior(xval).Posterior = normcdf(Ypost);
BoLasso.SVM.err(xval) = mean(Ypred~=Yte);
BoLasso.SVM.mse(xval) = mean((normcdf(Ypost)-(Yte-1)).^2);
BoLasso.SVM.YteYp(xval).YteYp = [Yte normcdf(Ypost)];
BoLasso.SVM.correctlog(teIDX,1) = Yte==Ypred;
%% non-BoLasso
% SVM-Linear
Model = train(Ytr,sparse(Xtr),'-s 1 -q');
[Ypred,~,Ypost] = predict(Yte,sparse(Xte),Model,'-b -q');
SVM.Posterior(xval).Posterior = normcdf(Ypost);
SVM.err(xval) = mean(Ypred~=Yte);
SVM.mse(xval) = mean((normcdf(Ypost)-(Yte-1)).^2);
SVM.YteYp(xval).YteYp = [Yte normcdf(Ypost)];
SVM.correctlog(teIDX,1) = Yte==Ypred;
% RF
Model = TreeBagger(ntrees,Xtr,Ytr,'method','classification','OOBVarImp','on','Options',options);
[~,Ypost] = predict(Model,Xte);
[~,Ypred] = max(Ypost,[],2); % annoyingly outputs winning class labels as cells, but we can get it from the probability outputs
Rf.ProbOuts(xval).Posterior = Ypost;
RF.err(xval) = mean(Yte~=Ypred);
RF.mse(xval) = mean(((Yte-1)-Ypost(:,2)).^2,1);
RF.YteYp(xval).YteYp = [Yte Ypost(:,2)];
%% store
end % xvalidation
out.Lasso = Lasso;
out.BoLasso = BoLasso;
out.SVM = SVM;
out.RF = RF;
out.Info = Info;
end % function