From A to PC – Running a Defensible Predictive Coding Workflow

by David Bao on September 11th, 2012

So far in our ongoing predictive coding blog series, we’ve touched on the “whys” and “whats” of predictive coding, and now I’d like to address the “hows” of using this new technology. Given that predictive coding is groundbreaking technology in the world of eDiscovery, it’s no surprise that a different workflow is required in order to run the review process.

The traditional linear review process utilizes a “brute force” approach of manually reading each document and processing it for responsiveness and privilege. In order to reduce the high cost of this process, many organizations now farm out documents to contract attorneys for review. Often, however, contract attorneys possess less expertise and knowledge of the issues, which means that multiple review passes along with additional checks and balances are often needed in order to ensure review accuracy. This process commonly results in a significant number of documents being reviewed multiple times, which in turn increases the cost of review. When you step away from an “eyes-on review” of every document and use predictive coding to leverage the expertise of more experienced attorneys, you will naturally aim to review as few documents as possible in order to achieve the best possible results.

How do you review the minimum number of documents with predictive coding? For starters, organizations should prepare their case to use predictive coding by performing an early case assessment (ECA) in order to cull down to your review population prior to review. While some may suggest that predictive coding can be run without any ECA up front, you will actually save a significant amount of review time if you put in the effort to cull out the profoundly irrelevant documents in your case. Doing so will prevent a “junk in, junk out” situation where leaving too much junk in the case will result in having to necessarily review a number of junk documents throughout the predictive coding workflow.

Next, segregating documents that are unsuitable for predictive coding is important. Most predictive coding solutions leverage the extracted text content within documents to operate. That means any documents that do not contain extracted text, such as photographs and engineering schematics, should be manually reviewed so they are not overlooked by the predictive coding engine. The same concept applies to any other document that has other reviewable limitations, such as encrypted and password protected files. All of these documents should be reviewed separately as to not miss any relevant documents.

After culling down to your review population, the next step in preparing to use predictive coding is to create a Control Set by drawing a randomly selected statistical sample from the document population. Once the Control Set is manually reviewed, it will serve two main purposes. First, it will allow you to estimate the population yield, otherwise referred to as the percentage of responsive documents contained within the larger population. (The size of the control set may need to be adjusted to insure the yield is properly taken into account). Second, it will serve as your baseline for a true “apples-to-apples” comparison of your prediction accuracy across iterations as you move through the predictive coding workflow. The Control Set will only need to be reviewed once up front to be used for measuring accuracy throughout the workflow.

It is essential that the documents in the Control Set are selected randomly from the entire population. While some believe that taking other sampling approaches give better peace of mind, they actually may result in unnecessary review. For example, other workflows recommend sampling from the documents that are not predicted to be relevant to see if anything was left behind. If you instead create a proper Control Set from the entire population, you can get the necessary precision and recall metrics that are representative of the entire population, which in turn represents the documents that are not predicted to be relevant.

Once the Control Set is created, you can begin training the software to evaluate documents by the review criteria in the case. Selecting the optimal set of documents to train the system (commonly referred to as the training set or seed set) is one of the most important steps in the entire predictive coding workflow as it sets the initial accuracy for the system, and thus it should be chosen carefully. Some suggest creating the initial training set by taking a random sample (much like how the control set is selected) from the population instead of proactively selecting responsive documents. However, the important thing to understand is that any items used for training should accurately represent the responsive items instead. The reason selecting responsive documents for inclusion in the training set is important is related to the fact that most eDiscovery cases generally have low yield – meaning the prevalence of responsive documents contained within the overall document population is low. This means the system will not be able to effectively learn how to identify responsive items if enough responsive documents are not included in the training set.

An effective method for selecting the initial training set is to use a targeted search to locate a small set of documents (typically between 100-1000) that is expected to be about 50% responsive. For example, you may choose to focus on only the key custodians in the case and use a combination of tighter keyword/date range/etc search criteria. You do not have to perform exhaustive searches, but a high quality initial training set will likely minimize the amount of additional training needed to achieve high prediction accuracy.

After the initial training set is selected, it must then be reviewed. It is extremely important that the review decisions made on any training items are as accurate as possible since the systems will be learning from these items, which typically means that the more experienced case attorneys should be used for this review. Once review is finished on all of the training documents, then the system can learn from the tagging decisions in order to be able to predict the responsiveness or non-responsiveness of the remaining documents.

While you can now predict on all of the other documents in the population, it is most important to predict on the Control Set at this time. Not only may this decision be more time effective than applying predictions to all the documents in the case, but you will need predictions on all of the documents in the Control Set in order to assess the accuracy of the predictions. With predictions and tagging decisions on each of the Control Set documents, you will be able to get accurate precision and recall metrics that you can extrapolate to the entire review population.

At this point, the accuracy of the predictions is likely to not be optimal, and thus the iterative process begins. In order to increase the accuracy, you must select additional documents to use for training the system. Much like the initial training set, this additional training set must also be selected carefully. The best documents to use for an additional training set are those that the system would be unable to accurately predict. Rather than choosing these documents manually, the software is often able to mathematically determine this set more effectively than human reviewers. Once these documents are selected, you simply continue the iterative process of training, predicting and testing until your precision and recall are at an acceptable point. Following this workflow will result in a set of documents identified to be responsive by the system along with trustworthy and defensible accuracy metrics.

You cannot simply produce all of these documents at this point, however. The documents must still go through a privileged screen in order to remove any documents that should not be produced, and also go through any other review measures that you usually take on your responsive documents. This does, however, open up the possibility of applying additional rounds of predictive coding on top of this set of responsive documents. For example, after running the privileged screen, you can train on the privileged tag and attempt to identify additional privileged documents in your responsive set that were missed.

The important thing to keep in mind is that predictive coding is meant to strengthen your current review workflows. While we have outlined one possible workflow that utilizes predictive coding, the flexibility of the technology lends itself to be utilized for a multitude of other uses, including prioritizing a linear review. Whatever application you choose, predictive coding is sure to be an effective tool in your future reviews.

Leave a Reply