Building a Prediction

After you choose a prediction type, ML Studio walks you through three short steps that define exactly what to predict, when to predict it, and what to learn from. This page covers steps 2, 3, and 4 of the flow.

Step 2 - Pick the target

The Pick the target screen asks "What should mindzie predict?" You choose the column whose value you want the model to learn.

Pick the target screen showing candidate columns as cards

Each candidate column is shown as a card with its name, a hint of its contents, and a badge:

  • For Classification, the badge shows how many distinct values the column has (for example 5 values), with a few examples such as Home, Rehab, SNF, Home Health, Expired.
  • For Binary, only columns with exactly two values are offered.
  • For Regression, the badge reads number, and only numeric columns are offered.

Only columns that fit the prediction type you chose appear here, so you can't accidentally pick something that doesn't make sense. The header tells you which dataset you're working from and how many candidate columns were found.

Every other column becomes a signal. You only pick the target here. Every other column in your dataset - raw fields and enrichment-derived columns alike - becomes a candidate signal the model can learn from. You review and adjust those in step 4.

If no suitable columns appear, try a different prediction type, or add a derived column to your dataset with an enrichment first.

Step 3 - Prediction setup

The Prediction setup screen asks "When does mindzie predict?" A case is a sequence of steps over time, so you choose the moment a running case receives its prediction. ML Studio then learns from your historical cases at that same moment.

Prediction setup screen with the three prediction-moment choices

Pick one prediction moment:

Choice What it does
Continuously, as each step happens (recommended) Every open case gets a fresh prediction each time a new step is recorded, sharpening as the case unfolds.
At a milestone One prediction, made when the case reaches a step you choose - a real decision point. Select the milestone step from the dropdown.
After the first N steps Wait until a case has a little history before predicting. Set how many steps to wait (1 to 50).

A prediction uses everything known up to the moment of prediction - the steps so far, how long the case has run, the time since the last step, the time of day, and case attributes. It never uses the outcome or anything that happens after that moment, so predictions stay honest.

When you're happy, click Continue to features.

Step 4 - Features

The Features screen shows "What mindzie will learn from." Features are the signals the model considers. ML Studio profiles your dataset and pre-selects a strong set for you; you can add or remove signals, but you rarely need to.

Features screen showing suggested, existing, and engineered feature groups

At the top, a summary shows how many candidate features are selected, how many were suggested for this dataset, how many come from your existing enrichments, and confirms that AutoML selects the final subset automatically.

Features are grouped so you can see where each signal comes from:

  • Suggested for this dataset - signals ML Studio spotted as a good fit for your data, each with a short reason. Examples:
    • Time spent in each step - durations per step, the strongest signal in almost every process.
    • Resource handoffs - how often a case changes hands; frequent handoffs often go with a poor outcome.
    • Rework & loops - repeated steps and loops, which frequently precede a bad outcome.
    • Smart bands for numeric attributes - bucketing a raw number into bands, which often works better than the raw value.
  • Your enrichments - the enrichment columns already on this dataset, included by default. Toggle any off you don't want considered.
  • Engineered from your process - signals ML Studio builds automatically from your event log, such as activity patterns, flow and rework, timing, and resource handoffs. These are on by default. Raw case attributes are available too, but off by default because the engineered signals usually predict better.

You don't have to prune. Add anything that might help and leave the rest to the model. The training step ranks every signal, keeps only the subset that genuinely predicts, and shows you what it dropped and why - including a safeguard that removes any signal that already gives the answer away.

Click Continue to training to start the model search, covered in Training and results.