QPC+NaiveBayes CV test on 10 first projections for Sonar

Test klasyfikacji 10x10CV dla modelu Naive Bayes (przy założeniu rozkładu normalnego, default Matlab settings) uczonego na całym zbiorze Sonar oraz na pierwszych 10 kierunkach znalezionych przez QPC maksymalizowanego za pomocą spadku gradientu oraz za pomocą zmodyfikowanej wersji QPC bazującej na zbiorze prototypów referencyjnych i inicjowanej za pomocą podziału losowych projekcji na interwały.

10x10 CV, 10 first projections (10.05.15)

 10x10 CV
 normalization
 QPC parameters
    lr = 0.1
    optimization = gradient descent 
dataset= sonar
vectors= 208
features= 60
folds= 10
repetitions=10
ttest paired corrected

Naive Bayes (all data)
acc.	std.
67.95	11.28

Dim.	QPC		T-Test vs. NB	QPC2		T-Test vs. NB	T-Test vs. QPC
	acc.	std.	p-value	acc.	std.	p-value	p-value
1	74.41	9.44	0.0819	74.23	8.45	0.0700	0.9290
2	75.19	8.68	0.0482	74.60	9.52	0.0621	0.8029
3	76.68	9.11	0.0195	74.56	9.08	0.0683	0.3609
4	76.61	8.33	0.0182	76.39	9.16	0.0213	0.9255
5	76.87	8.05	0.0195	76.73	8.61	0.0145	0.9479
6	77.50	7.32	0.0114	78.31	8.41	0.0041	0.7347
7	77.49	7.52	0.0112	78.62	7.80	0.0025	0.6373
8	78.69	7.17	0.0039	79.19	8.38	0.0020	0.8445
9	78.55	7.32	0.0039	78.86	8.78	0.0037	0.9017
10	78.56	7.66	0.0039	79.09	9.17	0.0022	0.8328

Dim. - number of linear ptojections generated by QPC
QPC2 - fast QPC method based on subset of reference prototypes
NB - Naive Bayes with normal distribution assumption

${{:projects:models:qpc_and_naivebayes:sonar_10x10_dim_1_10.png|$

1x10 CV , 10 first projections (10.05.14)

 1x10 CV
 normalization
 QPC parameters
    lr = 0.1
    optimization = gradient descent

NaiveBayes
acc.	std.
70.64	13.80

Dim.	QPC + NB		QPC2 + NB
	acc.	std.	acc.	std.
1	74.02	14.25	73.02	10.55
2	74.52	13.29	74.97	10.29
3	76.02	13.93	76.95	12.21
4	77.90	12.51	78.85	10.32
5	77.42	12.65	79.83	10.18
6	76.47	12.72	77.90	11.68
7	76.95	12.79	77.45	11.81
8	76.95	11.77	78.88	10.75
9	78.38	11.91	78.40	11.21
10	79.33	12.30	78.90	11.40

Dim. - number of linear ptojections generated by QPC
QPC2 - fast QPC method based on subset of reference prototypes
NB - Naive Bayes with normal distribution assumption