About Predictive Coding

Predictive Coding uses Continuous Active Learning (CAL) to generate predictive scores. CAL learns from reviewer coding decisions on a small set of documents and applies those patterns to a larger document population.

After each round of training, the model assigns predictive scores to the to the entire model population, indicating how likely they are to receive a positive or negative coding value. These scores are used to build a review queue that prioritizes documents most likely to be coded as positive.

Here is the basic workflow for generating and retrieving predictive scores using CAL.

  1. Prepare the document population. Create a saved search or move the documents into a folder. Identify the single-choice field and the positive and negative values for coding documents. Then assign these values to some of the documents. Ensure that the document population includes at least one document coded as positive and one coded as negative in order to run the initial training of the model. Adding more positively and negatively coded examples improves the model’s accuracy.

  2. Create a predictive model. When creating the model, specify the document population using the saved search or folder you created, and select the field along with the positive and negative values you identified earlier.

  3. Create an assignment for the predictive model you created. The system generates review batches that include the highest-priority documents identified by the predictive model. You can retrieve batches from the prioritized queue after the model is initially trained. The batches include only the promoted documents.

  4. Review batches. Reviewers review each retrieved batch and code documents with a positive or negative value. For any coded documents in completed or released batches, reviewer coding decisions are used to retrain the model for more accurate predictions. The system sends only the positive and negative coding decisions to the model for retraining. The retrained model generates the predictive scores for the entire document set and updates the review queue according to the predictions.

  5. Retrieve the predictive scores. When the model generates the predictive scores, you can retrieve them and view them in the field labeled <model name> Score. The scores are searchable and exportable for analysis and reporting purposes. At any point in the process, you can retrieve the model’s predictive scores to review the score range for the retrieved batches and to assess the accuracy of the predictions against reviewer coding decisions.

You can add more documents to the existing model and retrieve their predictive scores using Document Actions>AI Model. When you add documents using this option, the system transfers the coding values associated with the documents.

Also, the AI Model field displays ‘Chat’ when used for chat functionality, ‘Predictive Coding’ when used for predictive coding, or both values when used for both features.

To view your model metrics and assess the maturity of the model, access Epiq AI. Refer to Epiq Suite shortcuts for more information.

The following list provides related topics.