A feature selection technique that evaluates subsets of input variables by training and testing a predictive model on each candidate subset, selecting the combination of features that produces the best model performance on a validation set.
Also known as wrapper feature selection, recursive feature elimination
A wrapper method is an approach to feature selection that uses a machine learning model as a black box to evaluate the predictive value of different subsets of features. Rather than scoring features independently based on statistical properties (filter methods) or encoding feature importance into the model structure (embedded methods), wrapper methods train a complete model on each candidate feature subset and measure how well that model performs on a held-out validation set. The feature subset that produces the best validation performance is selected.
Common wrapper method algorithms include forward selection (start with no features and add one at a time, each time adding the feature that most improves performance), backward elimination (start with all features and remove them one at a time), and recursive feature elimination (RFE), which iteratively removes the least important features according to the model’s built-in importance scores. Exhaustive search over all possible subsets is computationally infeasible for more than a small number of features—with 20 features there are over one million possible subsets—so wrapper methods use greedy or heuristic search strategies.
The key advantage of wrapper methods is that they account for feature interactions: a feature that is individually weak may become important when combined with another feature, and wrapper methods can capture this. The disadvantage is computational cost: each candidate subset requires training a complete model, which is expensive for large datasets or complex models. Wrapper methods are most practical with fast models (like linear models or decision trees) and modest feature counts.
Wrapper methods are relevant to agencies in the context of building predictive models from campaign data, where the available features include a mix of clearly valuable signals and potentially redundant or noisy variables. When an agency or their analytics vendor builds a model to predict creative performance, conversion likelihood, or audience value, the choice of which input features to include directly affects model quality. Wrapper methods provide a principled way to identify which combination of campaign variables actually improves prediction.
Feature selection affects model interpretability and client reporting. A model trained on 50 features produces outputs that are harder to explain to clients than a model trained on the 8 features that actually drive prediction. Wrapper methods can identify the minimal set of features that achieves near-maximum performance, producing simpler, more interpretable models whose outputs can be translated into actionable insights. For agency use cases where client stakeholders need to understand why a model flags a campaign for optimization, a smaller feature set is a significant advantage.
Recursive feature elimination is a common vendor technique. Many commercial ML platforms used in ad tech—for lookalike modeling, bidding optimization, and creative scoring—use recursive feature elimination as part of their model training pipeline. When a vendor reports that their model ‘selects the most predictive signals from your first-party data,’ they are often describing a wrapper-method-based selection process. Understanding what this means helps agencies ask the right questions: how many features were evaluated, what model was used for evaluation, and how stable is the selected feature set across different data splits.
An agency analytics team is building a conversion prediction model for a client using 35 available campaign attributes: creative format, device type, time of day, day of week, audience segment, publisher category, viewability score, historical CTR, bid price, and many others. Training a model on all 35 features produces a validation AUC of 0.71 but requires 35 data points per prediction request and is difficult to explain to the client. They run recursive feature elimination using a logistic regression base model, iteratively removing the least important feature until validation performance drops meaningfully. After RFE, 9 features account for 0.70 AUC—nearly identical performance with far fewer inputs. The simplified model is faster, cheaper to run, and the 9 selected features map cleanly to campaign decisions the client can actually act on: format, device, time of day, audience segment, and viewability threshold.
The workshop covers how AI tools actually work, how to evaluate them, and how to apply them to real agency workflows.