I85PREP
⌘ K
Dataiku ML Practitioner study materials
Dataiku ML Practitioner study resources

Dataiku ML Practitioner: Visual Machine Learning

Dataiku ML Practitioner · Work through your exam, one topic at a time.

All guide sections

Visual Machine Learning

What this domain tests

Dataiku's Visual ML lab is a guided environment for training, comparing, and deploying models without writing code. The exam checks whether you can configure a modeling session correctly: choose the target and prediction type, define a train/test strategy, handle features, select metrics, train candidate algorithms, optimize hyperparameters, interpret results, and deploy the winning model into the flow. Questions often test judgment rather than menu recall, especially around leakage, imbalance, and metric selection.

Core rules and concepts

  • Target and prediction type: Set the target before training. Binary classification, multiclass classification, and regression imply different metrics and algorithms. A numeric target with many repeated values may still be regression; a low-cardinality string target is classification.
  • Train/test split: Hold out data before comparing models. Dataiku supports random and ordered splits; ordered splits are essential for time-dependent data. Never tune on the test set.
  • Cross-validation: Use k-fold CV on the training portion to estimate variance and support hyperparameter search. Report CV mean and spread, not a single fold.
  • Feature handling: Categorical, text, and vector features need explicit handling. High-cardinality categoricals can overfit; text features require tokenization or hashing settings.
  • Feature generation and selection: Generate ratios, dates, and interactions where domain logic supports them; use selection to reduce noise, but validate that removed features do not encode rare but important segments.
  • Metrics: Classification uses AUC, precision, recall, F1, log loss, accuracy, and lift. Regression uses RMSE, MAE, R2, and MAPE. Choose metrics that reflect the business cost.
  • Optimization: Grid or random search explores hyperparameters. Watch for overfitting when the search space is large relative to data.
  • Deployment: Deploy a model as a saved model, then score new data with a Score recipe or API node. An Evaluation recipe compares models on a shared dataset.

Worked example

A churn dataset has 100,000 customers and 18% positives. A baseline logistic regression reaches AUC 0.78; XGBoost reaches 0.84. Accuracy at threshold 0.5 is 83% for both because predicting "no churn" always is already 82% accurate. The correct reasoning is to reject accuracy, compare precision-recall and lift, and choose a threshold based on retention cost. If a retained customer is worth more than a wasted offer, lower the threshold to raise recall. The test set stays untouched until the final choice.

Traps that cost marks

  • Including post-outcome fields such as cancellation date or refund amount, which leaks the target.
  • Selecting the model and threshold on the test set instead of using CV and a validation split.
  • Reporting accuracy on an imbalanced target.
  • Ignoring feature handling for high-cardinality categoricals.
  • Deploying without an evaluation recipe or a scoring check.