Skip to main content
melvinisaacjoshig2675
February 1, 2024
Solved

AI Skill Classification

  • February 1, 2024
  • 8 replies
  • 0 views

Hi Community,

I have a multiple question on the Document classification AI Skill. I have trained the AI Skill classification with 2 types of document. Both type does not have fixed structure (few text change based on the conditions). 

When we upload the document of these 2 type the confident score is always fall between 55 to 60 . 

Another behaviour i observed  when i upload different document type which is totally different , the confident score is 50 which i do not expect.

Questions:

Does anyone know how the confident score is determine? 

The document structure is not same every time  is that because the confident score not going beyond 60? I expect to be more.

Why confident score is still 50 when non trained document uploaded?

Best answer by gabby.rothschild

Hi Melvin,

Because machine learning is not deterministic, a high variance in your confidence score or accuracy scores may mean that your dataset is not representative enough. Please refer to our documentation on how to create a representative dataset: https://docs.appian.com/suite/help/23.4/collect-data.html as well as how to evaluate model performance: https://docs.appian.com/suite/help/23.4/evaluate-ai.html

8 replies

gabby.rothschild
February 1, 2024

Hi Melvin,

Because machine learning is not deterministic, a high variance in your confidence score or accuracy scores may mean that your dataset is not representative enough. Please refer to our documentation on how to create a representative dataset: https://docs.appian.com/suite/help/23.4/collect-data.html as well as how to evaluate model performance: https://docs.appian.com/suite/help/23.4/evaluate-ai.html

melvinisaacjoshig2675
February 5, 2024

Thank you gabby. I will check and revert if i have questions

abhayd3663
February 5, 2024

The confidence score is internal to Appian and it represents how close the document classification and extraction is.

melvinisaacjoshig2675
February 6, 2024

Thanks Abhay but my curiosity is Why confident score is still 50 when non trained document uploaded? confident score is crucial to decided on the flow of the process. 

stefanhelzle0001
February 6, 2024

I think this happens when the identified patterns in the documents are not strong enough and you upload a document that does not deviate enough from the training material.

srmm0001
February 8, 2024

To address the issue of low confidence scores for untrained document types, it's essential to:

  • Clearly separate known and unknown document types during the training phase.
  • Consider training the model to recognize an "other" category for documents that do not fit into the predefined types.