perClass | Shouldn’t sdprox return 0’s on the diagonal...

Shouldn’t sdprox return 0’s on the diagonal…

Jean-Michel

Posted: 01 October 2010 03:43 PM

[ Ignore ]

Member

Total Posts: 68

Joined 2010-04-27

Hello,

Am I right to expect zeros on the diagonal of a proximity mapping that uses Euclidian distance or at least a bit of symmetry?

>> sda = +(cdatadiff2*sdprox(cdatadiff2,'EUCL'));
>> sda(1:5,1:5)

ans =

   43.2359   52.3670   44.0920   42.4990   47.1889
   96.7733   98.6114   98.6162   98.9042  102.2701
  111.9697  108.5412  108.2605  108.8375  110.1390
   69.8523   67.3323   68.1756   66.5191   71.3832
   97.2699   97.9262   94.2977   95.2910   93.6762

a = +(dataset(cdatadiff2)*proxm(dataset(cdatadiff2),'d',1));
>> +a(1:5,1:5)

ans =

         0   99.6569  112.8007   77.6335  101.5290
   99.6569         0  137.9108  111.0354  119.8310
  112.8007  137.9108         0  113.5863  140.3564
   77.6335  111.0354  113.5863    0.0000  120.1441
  101.5290  119.8310  140.3564  120.1441    0.0000

Do I use sdprox correctly?

regards,

Profile

pavel

Posted: 01 October 2010 05:39 PM

[ Ignore ] [ # 1 ]

Administrator

Total Posts: 240

Joined 2008-04-26

Hi Jean-Michel,

yes, you use sdprox in the correct way. I cannot repeat your problem with non-zero diagonal:

>> load fruit
>> a
'Fruit set' 260 by 2 sddata, 3 classes: 'apple'(100) 'banana'(100) 'stone'(60) 
>> d=a*sdprox(a,'EUCL')
'Fruit set' 260 by 260 sddata, 3 classes: 'apple'(100) 'banana'(100) 'stone'(60) 
>> +d(1:5,1:5)

ans =

         0    5.2580    3.1564    3.5293    5.4365
    5.2580         0    7.6838    2.8937    1.6194
    3.1564    7.6838         0    6.5988    8.2748
    3.5293    2.8937    6.5988         0    2.1833
    5.4365    1.6194    8.2748    2.1833         0

Can you post a small data set where you get this problem?

With Kind Regards,

Pavel

Profile

pavel

Posted: 06 October 2010 08:48 AM

[ Ignore ] [ # 2 ]

Administrator

Total Posts: 240

Joined 2008-04-26

Hi Jean-Michel,

thanks for providing the data for repeating the problem. The issue was in sdprox which was in certain situations re-organizing prototypes for execution.

It is fixed in the latest release 2.2.4 5-oct-2010 available for download.

With Kind Regards,

Pavel

Profile

‹‹ sdsvc: problem during grid search. modifying priors ››