Training and Results

The final step - Train & deploy - is where ML Studio does the heavy lifting. It searches many models to find the one that predicts your target best, shows you the search as it happens, and presents a clear scorecard of the winner. This is the "AutoML" part of ML Studio: you don't choose or tune an algorithm - the search does it for you.

Training runs on the server

Training runs in the background on the server, so you can close the browser and come back - the run keeps going. The screen updates itself as the search progresses. You can Stop a run at any time and resume it later, and long runs pick up where they left off.

While the search runs, ML Studio tries different model families against different feature-sets (for example Process features alone, versus + Enrichments). It scores each combination on data the model hasn't seen, so the numbers reflect how well the prediction will work on new cases.

The live search grid comparing model families against feature-sets

The search grid shows model families down the side and feature-sets across the top. Each cell is the best out-of-sample score that family reached on that feature-set:

  • Darker cells are stronger.
  • The green cell is the current best.
  • An outlined cell is training right now.
  • Empty cells haven't been tried yet.

Above the grid, a live band tells you how many configurations have been scored out of the total, how much time is left in the search, what's training now, and the best result so far.

The training-loss curve

Alongside the grid, a Training loss - lower is better chart plots two lines as the search deepens:

  • out-of-sample (test) - how the model does on cases it hasn't seen.
  • in-sample (train) - how it does on the cases it learned from.

When the two lines stay close, the model is generalizing to new cases rather than memorizing the old ones - exactly what you want.

The scorecard

When the search finishes, the heading changes to Your prediction is ready and ML Studio presents the winning model.

Binary scorecard showing the winning model and its metrics

At the top, Winner: [model family] shows which family won, which feature-set it used, its key settings, and how many signals and cases it was trained and tested on. The headline metric cards depend on your prediction type:

Prediction type Metric cards
Binary F1 score (the headline), Accuracy, Precision, Recall
Classification Accuracy, Balanced accuracy, Log-loss
Regression R-squared, Typical error (MAE) in the target's own units, RMSE

Classification scorecard with Accuracy, Balanced accuracy, and Log-loss

Each metric has a plain-language tooltip, and a footnote names the single headline number the search optimized. Scores are cross-validated across several folds and shown as a mean with a small plus/minus spread, so you can see how stable the result is.

What drives the prediction, and the full leaderboard

Below the metrics, ML Studio explains why the model predicts what it does and shows every family it tried.

Feature drivers and the model leaderboard

  • What drives this prediction most ranks the signals that most influence the prediction, as a simple bar chart.
  • Leaderboard - every family's best configuration lists each model family's best result: its rank, the winning feature-set, the key settings, the score, and how long it took. The winner is marked (best).

If any signal was removed for looking like it already encoded the answer, ML Studio calls that out too, along with the columns it dropped - so you can trust that the score reflects a genuine prediction, not a leak.

From here, ML Studio automatically puts the winning model to work as a live prediction on your cases. That output - and how to use it - is covered in Using your predictions.