Hi,

Please refer to the picture for the output which I have got. The best output which I have got is under the following conditions:

1. Validations: 10
2. Subject weight: 0.9
3. Body weight: 0.1
4. Operators used under 'Process Documents from Data' - 'Tokenize, Filter Stopwords, Filter Tokens by Length (lower limit: 3 upper limit: 999), Filter Tokens by Content: string filtered - www with 'inverse' condition.

Please put in your outputs for discussions and deciding on the best combination.
Under this condition, 2 texts have been predicted as Davison correctly.

The results are changing in two different conditions. They are:
1. While 'Filter stopwords' and 'Stem(porter)' are used, the result is misclassified for Davison and other categories.
2. While 'Filter tokens by length' and 'Filter tokens by content' are used, the result is classifying emails more correctly.

stem porter is clearly affecting the result , funnily in negative way

Davison: 4
Phone & Network: 4
Web: 3
Others: 9
Weight to Subject: 0.9
Weight to Body: 0.1
Model Accuracy: 71.71% +/- 8.44% (mikro: 71.63%)

Hi

It seems that the model is working, and is automated. Though in some cases, the emails are misclassified. But, this is the best, which I am getting. See the XML file.

trainingdata_validationscheduled_unscheduled_xml.txt


test_output_xml.txt