Case Study: Comcast achieves scalable machine learning to reduce avoidable truck rolls and personalize TV with H2O.ai

Operationalizing Machine Learning at Comcast

Comcast needed to turn massive, multi‑source production data (from tens of millions of customers and hundreds of millions of devices) into reliable, scalable predictive models rather than relying on sampling. The company’s challenges included operationalizing and scaling data science across high‑volume operations and improving predictive accuracy for use cases like avoidable truck rolls, personalized TV recommendations, customer‑experience metrics and self‑healing systems. To address this, Comcast partnered with H2O.ai and adopted the H2O machine learning platform.

Using Datameer for feature engineering and H2O.ai for model training and deployment, Comcast built end‑to‑end pipelines (feature extraction/selection, training/validation, classifier export) to prevent avoidable truck rolls, predict show popularity 24 hours ahead with gradient‑boosted trees, cluster viewers for real‑time recommendations, and power a rules‑based self‑healing system. H2O.ai enabled exporting models as Java or serving them as web services, helping Comcast operationalize ML, reduce costly truck rolls, deliver more accurate personalized TV suggestions, improve customer‑experience metrics, and scale predictive analytics across its footprint.

Comcast

Drew Leamon

Director of Engineering Analysis