A machine learning approach that combines predictions from multiple models to produce a single output that is more accurate and more stable than any individual model’s prediction. For agencies, ensemble learning is the technique behind many of the highest-performing predictive tools in advertising technology, including the scoring models that drive bidding, lead qualification, and audience targeting.
Also known as model ensembling, ensemble methods, ensemble modeling
Ensemble learning works on a simple premise: different models make different mistakes. A model that is wrong about 20% of examples will be wrong about a different 20% than another model with the same aggregate accuracy. By combining their predictions and requiring agreement or taking a weighted average, the ensemble reduces errors in cases where individual models disagree and makes more confident predictions in cases where they agree. The result is lower overall error than any single model achieves independently.
The two dominant ensemble approaches are bagging and boosting. Bagging, exemplified by random forests, trains many models independently on different random subsets of the training data and averages their predictions. This approach reduces variance: the ensemble is more stable than any individual tree and less prone to overfitting. Boosting, exemplified by gradient boosting methods like XGBoost and LightGBM, trains models sequentially where each new model focuses on correcting the errors of the previous ones. This approach reduces bias: the ensemble can learn complex patterns that simpler models miss.
Ensemble methods consistently outperform single models on tabular data, which is why they dominate machine learning competitions and production deployments in advertising technology. The tradeoff is computational cost: an ensemble of 500 trees requires 500 times the inference computation of a single tree, and an ensemble of neural networks requires 500 times the training compute. For many advertising applications, this cost is paid by the vendor and invisible to the agency, but it matters when evaluating inference latency and training budget.
Ensemble methods are inside most of the AI tools a working ad agency uses: bidding algorithms, lead scoring systems, audience qualification tools, and churn prediction models all use ensemble-based approaches. Understanding why helps agencies evaluate these tools more precisely, configure them more deliberately, and communicate their behavior to clients more accurately.
Stability across data variations is a practical benefit for client deployments. A single model trained on client data can behave erratically when input distributions shift slightly, such as when a new product line or audience segment is added to the campaign. An ensemble trained on diverse subsets of the same data is inherently more stable under these shifts. For production systems that clients depend on, stability is often more valuable than peak accuracy.
It provides a natural confidence measure. When all models in an ensemble agree on a prediction, the ensemble is confident. When models disagree significantly, the ensemble’s aggregate prediction is less reliable and the individual prediction warrants human review. This disagreement signal is a built-in uncertainty estimate that single models do not provide, and it is useful for routing borderline cases in lead qualification or content moderation workflows.
Gradient boosting on tabular data outperforms neural networks more often than the AI narrative suggests. For structured client data like CRM records, email engagement histories, and campaign attribution tables, gradient boosted ensemble models frequently outperform neural networks. Agencies selecting modeling approaches for client data should not default to neural network architectures simply because they are prominent in AI coverage; the choice should be driven by what works best on the actual data type.
An agency builds a customer churn prediction model for a subscription retail client using three approaches: logistic regression, a single decision tree, and a gradient boosted ensemble of 200 trees. Tested on a held-out validation set, logistic regression achieves 71% accuracy, the single decision tree achieves 74%, and the gradient boosted ensemble achieves 83%. When deployed on live monthly subscriber data and evaluated against actual churn outcomes 90 days later, the accuracy gap narrows slightly but holds. The agency presents all three results to the client with a plain-language explanation of why the ensemble outperforms: it corrects its own errors iteratively rather than finding a single decision boundary and stopping. The client approves the ensemble model for production and the explanation of why it works becomes part of the model documentation delivered at handoff.
The generative AI foundations module of the workshop covers how today’s models work, what the performance differences between approaches actually mean, and how to choose the right tool for the data and objective at hand.