Decipher Classification model quality getting worse

[TushigbayarTsev]

Hello,

I trained my classification model with Create Classification Training Batch action for invoices and its attachments at the beginning. After 250 documents I though I could make it better by sending another Training Batch to recognize invoices and attachments. Documents were not that different in my opinion. Suddenly the Classification model became worse and basically had to start all over again with the training. 

Is that possible???

------------------------------
Tushigbayar Tseveenbayar
------------------------------

[Ben.Lyons1 · Accepted solution]

Hi Tushigbayar,

The classification model can only be trained as a "one-off" process. So when you do your initial training, that sets the model up. It will not be updated with manually verified documents, so it should not get better or worse.

If you upload new training batches, it will create a new model and delete the previous model. So this could potentially reduce your success rate.

I would recommend training the model, then testing with as many key documents as possible. If the success rate is below target, then revisit the training sets and create a new model.

Let me know if that helps or not.

Thanks

------------------------------
Ben Lyons
Product Consultant - Decipher Specialist
Blue Prism
UK
------------------------------