Creating a One AI Time to Separation Model

Time to Separation adds time to traditional attrition risk analysis. Instead of only estimating whether an employee may separate within one fixed period, it shows how separation risk changes across multiple future time horizons and helps identify the factors associated with employees staying longer or separating sooner.

This guide walks through how to configure, run, review, and deploy a Time to Separation model using the One AI Recipe.

Before you Begin

It is helpful to know the following before you start:

• The date column that records employee separations.
• The headcount population you want to make predictions for.
• The unique identifier that identifies each employee.
• The population date, or the point in time from which you want One AI to make predictions.
• The three future time periods you are interested in predicting.
• The employee attributes you want the model to be able to consider.

It is normal for active employees to have no separation date. One AI treats employees who have not separated during the period it can observe as censored. In simple terms, One AI knows that the employee remained active for at least that long, even though it does not yet know when or whether that employee will eventually separate.

Step 1. Configure the Time to Separation Model

Create the model and select the recipe

1. Select One AI from the top navigation bar.

2. Select + Add Machine Learning Model.

   

3. Enter a Display Name for your model.
4. Under Using Data From, select 'One AI Recipe'.

5. Select 'Configure One AI Recipe'.

   

6. Under "What are you interested in predicting?", select 'Time to Separation'.

   

Review the Model Statement

As you complete the recipe, the statement at the top of the page updates to summarize what One AI is being configured to predict. Use this as a quick check that the selected population, population date, prediction horizons, training history, and attributes match what you intended.

Select the Separation Date

Under "Which column should One AI use to define the separation date?", select the date column that records when an employee separated.

This date is the event the model is learning to predict. Employees who have a separation date contribute information about when the separation occurred. Employees without a separation date by the end of the available observation period are still included and are treated as censored.

The target for a Time to Separation model must be a date column that records when an employee separated. A termination count, yes/no termination field, or other non-date field cannot be used as the target for this model.

Select the Model Population

Under "Which headcount should One AI use to make predictions?", configure the group of employees you want the model to score. This step works the same as any other model. 

• Population Metric: Select the headcount metric that represents the population you want to model - usually a headcount metric. 
• Unique Identifier: Select the field that uniquely identifies each person - usually the employee_id column from the prd_employee table. 
• Population Date: Select the point in time from which the model should make future predictions. Think of this as your model's time anchor. 
• Filters: Add filters if you only want to model a subset of the selected population - for example, if you chose a headcount (EOP) metric, you may want to filter it to only include managers or employees in California. 

One AI also uses this population at prior points in time when it builds the model’s training data.

Select the Time Periods

Under "Which time periods should One AI use for this model?", configure predicted horizons and the training history period.

Prediction Horizons

Prediction horizons are the three future time windows at which One AI will calculate separation risk from the population date. For example, horizons of 12, 24, and 36 months would produce a separation-risk prediction for each employee at each of those three points in time.

Choose horizons that are useful for the workforce decisions you are trying to support. They do not have to be 12, 24, and 36 months.

Training History Period

The Training History Period determines how far backward One AI looks when creating the historical data used to train the model. This is different from the prediction horizons: prediction horizons look forward from the population date, while training history looks backward.

Enter the number of years of history you want One AI to use. The recipe requires at least two years of history. Longer prediction horizons generally require more historical data to support training and evaluation, so use an appropriate amount of available history for the horizons you selected.

Select the Attributes the Model Can Use

Under "Which core attributes do you want to use in your prediction?", select the employee attributes One AI is allowed to consider when building the model. This step works the same as any other model. 

These might include information related to an employee’s role, organization, compensation, tenure, performance, or other available data. You want to provide the model with useful information without including every available field simply because it exists. Avoid including duplicate or highly redundant fields, such as both age and date of birth. Redundant information can add unnecessary noise and make the model harder to interpret.

The recipe shows which columns are Included, Excluded, or Unavailable based on the selected population and data relationships.

Add Generative Attributes, if needed

Generative Attributes are optional. They can add additional information created from your data, such as an employee-level measure or a team-level measure. This step works the same as any other model. 

If you are creating your first Time to Separation model and are unsure whether you need a Generative Attribute, it is reasonable to begin with your core attributes and refine the model later.

Check your Data before Running the Model

Under "Would you like to verify that all of the selections you have made are valid?", select Generate Data Statistics. We recommend doing this before saving the recipe.

The generated statistics let you review information such as row counts, unique counts, minimums, and maximums. They can help you catch unexpected data before spending time running the model. If you included Generative Attributes, review their statistics here as well.

You can also use Download Train/Test Data and Download Predict Data if you need to inspect the underlying datasets in more detail. Downloading the files is optional and is generally not necessary to create a model.

Save the Model

Once you are satisfied with the recipe selections:

1. Save the recipe.

   

2. You will be automatically returned to the Machine Learning Model configuration.

3. Save or create the overall Machine Learning Model on the augmentation screen.

   

Step 2. Run the Model

Find your new Time to Separation model on the Augmentations page and select the 'Run' button. Confirm that you want to start the run when prompted.

One AI will prepare the historical data, train the Time to Separation model, validate it against historical snapshots, and generate predictions for the selected population. This can take some time depending on the amount of data being processed, and you do not need to remain on the page while the run completes.

When processing is complete, the run will show a Pending status. Pending means the results are ready for you to review and decide whether to deploy, deploy and persist, or ignore the run.

Step 3. Review the Model Results

Select 'Runs' for your model and open the run you want to review.

Time to Separation does not have a separate Exploratory Data Analysis (EDA) report. The Time to Separation Results Summary is the primary place to understand what happened during the run and decide whether you are comfortable using its predictions.

Start with any messages or warnings shown for the run. Some messages are informational, while others may indicate that One AI did not have enough historical follow-up to fully evaluate one of your longer prediction horizons.

Then review the Results Summary. At a minimum, you should be able to answer three questions before deploying:

• Did the model run on the population, dates, horizons, and attributes you intended?
• Does the model performance give you enough confidence to use the predictions?
• Do the model features and prediction patterns make sense in the context of your workforce?

For a step-by-step explanation of the report, see Understanding Your Time to Separation Results.

4. Decide What to do with the Model

After reviewing the Results Summary, choose the action that matches what you want to do with this run at the bottom of the report.

Deploy

Choose Deploy when you are comfortable with the model and want its results available in One Model for reporting and analysis. Deploying loads the results of that run into the data model so they can be used in areas such as Explore, metrics, and Storyboards.

Deploy and Persist

Deploy and Persist also makes the results available in One Model, but it additionally keeps the trained model for use in future runs.

After a model is persisted, future runs reuse that trained model rather than training a new model from scratch. The population being scored can change as your data changes, but the trained model itself remains fixed.

Persisting can be useful once you have reviewed a model carefully and intentionally want to continue using the same trained model over time. If you are comfortable deploying the results but are unsure whether the trained model should be reused for future runs, choose Deploy.

Ignore

Choose Ignore if you decide that a particular run should not be used. Ignoring a run prevents that run from later being deployed. If the problem can be fixed by changing the recipe configuration, edit the model and run it again.

Using the Model Results

After deployment, Time to Separation results can be analyzed alongside the rest of your people data. Ask your CS team about the templated storyboards that go with the time to separation model. Depending on your site’s configuration, you may be able to explore separation risk at each configured horizon, employee risk ranking, risk groups, expected active days, and attributes associated with higher or lower separation risk.

These predictions should be used as additional information for workforce analysis and decision-making, not as certainty about what an individual employee will do.