Classification accuracy results

The Classification Accuracy Details window provides a comprehensive view of your solution's classification performance. This interactive window displays key metrics and a visual result matrix to help you monitor and improve classification quality in your DocAI solution.

Metrics

The window displays the following classification accuracy metrics at the top:

Metric	Description
OVERALL SCORE	This score combines precision and recall and only scores well when DocAI Studio is both correct in the document type it assigns and thorough in catching every document of that type. Strong performance on one side is not enough. The score ranges from 0 to 100 percent. A higher percentage indicates better performance. Formula: F1 Score = 2 × (Precision × Recall) / (Precision + Recall)
PRECISION	Precision measures how accurately DocAI Studio classifies documents. High precision means the system rarely assigns documents to the wrong document type. Formula: Precision = Correctly Classified / (Correctly Classified + Misclassified)
RECALL	Recall measures how many documents DocAI Studio successfully classifies. This metric indicates the system's ability to classify documents rather than leave them unclassified. High recall indicates that the system confidently classifies most documents. Formula: Recall = Correctly Classified / Document Count
DOCUMENTS TESTED	The total count of test documents used in the classification accuracy calculation.
LAST RUN	The date and time of the most recent classification accuracy calculation.

A summary below the metrics displays the classification results:

CORRECTLY CLASSIFIED:: The number of documents that were classified to the correct document type. This value is displayed with a green indicator.
MISCLASSIFIED:: The number of documents that were classified to an incorrect document type. This value is displayed with a red indicator.
UNCLASSIFIED:: The number of documents that could not be classified. This value is displayed with a yellow indicator.

Result matrix

The result matrix is an interactive grid that displays how documents were classified compared to their actual document types. This matrix helps you identify specific misclassification patterns.

Rows represent the actual document types - what the documents really are and their F1 value.
Columns represent the predicted document types - what classification determined.
Cells display the count of documents for each actual/predicted combination.
An Unclassified column displays the number of documents that could not be classified.

The result matrix helps you identify which document types are being confused with each other. For example:

A high number in a red cell indicates that documents of one type are frequently misclassified as another type.
A high number in the Unclassified column indicates that documents of that type are not being classified at all.
Ideally, all numbers should be in green cells, with zeros elsewhere.

Select a cell with a number to navigate to the document type that cell represents.

If the matrix contains many document types, you can scroll to view all rows and columns.

Improving classification accuracy

Use the result matrix results to improve your solution:

For document types with high misclassification rates (red cells), add more training documents to help the system better distinguish between similar document types.
For document types with high unclassified counts, ensure you have sufficient training documents and verify that your classification settings are appropriate.
Adjust classification confidence thresholds in Settings > Processing > Classification to balance precision and recall based on your requirements.
Try different classifiers:

For example, if you are using Generative AI classification, try the following:
- Edit the Generative AI classification descriptions for each class to see if this improves results.
- If the Run before other classifiers setting is selected, clear this and enable one or both layout and content classification.
Similarly, if you are using a single classifier only, try including others to see if classification results improve.