New tuning.sh
Boris 'pi' Piwinger
3.14 at logic.univie.ac.at
Mon Jun 16 12:56:14 CEST 2003
Boris 'pi' Piwinger wrote:
> Top 10 results:
> robs min_dev spam_cutoff run0 run1 run2 total
> 0.0100 0.450 0.500000 35 32 32 99
> 0.1000 0.425 0.500000 43 44 46 133
> 0.1000 0.450 0.500000 41 48 51 140
> 0.0320 0.450 0.501000 48 55 56 159
> 0.0320 0.400 0.500000 55 56 53 164
> 0.0320 0.375 0.500000 62 64 67 193
> 0.1000 0.400 0.500000 59 68 67 194
> 0.0320 0.425 0.503000 64 70 68 202
> 0.0100 0.425 0.501000 64 70 70 204
> 0.3200 0.450 0.550000 82 83 93 258
>
> I used target=6 so results have 1 or 2 fp. Actually, those
> in the top10 only have one.
With target=3:
robs min_dev spam_cutoff run0 run1 run2 total
0.1000 0.450 0.501000 51 59 66 176
0.1000 0.425 0.525000 70 82 81 233
0.0320 0.425 0.528000 73 85 88 246
0.0320 0.450 0.681000 94 96 92 282
0.3200 0.450 0.607000 91 94 101 286
0.0320 0.400 0.511000 94 97 104 295
0.0100 0.425 0.537000 99 97 106 302
0.0100 0.450 0.685000 104 110 104 318
0.1000 0.400 0.545000 106 109 111 326
0.0100 0.400 0.517000 106 104 119 329
These are all with 0 fp, but the result list has runs with
up to two. Now this is quite a difference in results.
Using 0.5 seems to risky to me according to earlier tests.
So I currently use .501 as a spam_cuttoff. With min_dev=.45
robs should be .1 or .01 according to the tuning results. So
I test with my current in-production-database and my
complete training archive (which is bigger):
robs=0.01
spam_cutoff=0.501
Spam:
13702 test.spam
False negatives:
43
Ham:
22387 test.ham
False positives:
0
robs=0.1
spam_cutoff=0.501
Spam:
13702 test.spam
False negatives:
100
Ham:
22387 test.ham
False positives:
0
robs=0.01
spam_cutoff=0.5
Spam:
13702 test.spam
False negatives:
16
Ham:
22387 test.ham
False positives:
2
robs=0.1
spam_cutoff=0.5
Spam:
13702 test.spam
False negatives:
82
Ham:
22387 test.ham
False positives:
2
So I will stay with robs=0.01 and spam_cutoff=0.501.
pi
More information about the Bogofilter
mailing list