An ensemble machine learning technique that trains multiple models on different random samples of the same dataset and combines their predictions. It is one of the core reasons AI-powered audience scoring and propensity models are more reliable than any single model trained on the full dataset.
Also known as bootstrap aggregating, ensemble bagging, resampling ensemble
Bagging, short for bootstrap aggregating, works by generating multiple random samples from the training data (with replacement), training a separate model on each sample, and then averaging or voting across all model outputs. Because each model saw a slightly different slice of the data, they make different errors. Combining them reduces the overall error rate.
Random forests, one of the most widely used and reliable machine learning algorithms, are built directly on bagging. Each tree in the forest is trained on a bootstrapped sample of the data, and the forest’s prediction is the average of all tree predictions.
The benefit of bagging is primarily in reducing variance: it makes predictions more stable and less sensitive to quirks in any one training sample. This matters most in domains where data is noisy or limited, which describes most marketing and audience datasets.
Agencies rely on predictive models for lead scoring, audience qualification, churn prediction, and campaign optimization. Many of those models use bagging under the hood. Understanding it helps agencies evaluate vendor claims about model reliability and ask the right questions when results look inconsistent.
Ensemble stability is a feature worth asking for. A vendor claiming their model produces consistent predictions across varied audience segments is implicitly claiming some form of ensemble behavior. Asking whether their model is a single estimator or an ensemble is a reasonable technical due diligence question for any predictive tool used in client work.
Sample size affects whether bagging helps. Bagging improves reliability when training data is sufficiently large. For niche B2B audiences or short campaign histories, bootstrapped samples may all look similar, reducing the benefit. Small-data models require different evaluation criteria.
It does not fix bad features. Bagging reduces variance in predictions, but it does not fix a model trained on irrelevant or biased inputs. Garbage in, garbage out applies regardless of how many copies of the model vote on the answer.
An agency is evaluating two lead scoring vendors. One uses a single logistic regression model retrained quarterly. The other uses an ensemble of gradient-boosted trees retrained monthly. The agency runs both on the same historical pipeline data and finds the ensemble produces more consistent scores across different industry verticals in the client’s database. When one vertical underperforms a quarter later, the ensemble degrades more gracefully than the single model, which flips its predictions on several borderline accounts.
The generative AI foundations module of the workshop covers how today’s models work, what they can and can’t do, and how to choose between them.