Building ML Platform & executing ML Project: Revenue impact within 6 months

18.03. | 10:45 - 12:15 | tech.stage

This presentation is in 2 parts.

Data Engineering:
First, we introduce a cloud-based architecture for training and operating ML-models at scale. This framework is based on continuous integration/deployment, loose coupling and generic building blocks. We talk about Machine-Learning-Operations (ML-Ops) as a necessity for all machine learning based platforms. What should be the design considerations to avoid high technical debt.

Data Science:
Secondly, we to introduce a ML based TV attribution model to connect the TV-viewing and online-purchasing behaviour of millions of customers. Even though the concept of attribution model has been widely used for multi-channel estimation, it is now for the first time, employed specifically for the TV advertisements. The obtained data-driven attribution model is further utilized to predict the optimal spot allocation maximizing the revenue and ROI. We talk about how we harvested big data and analyzed it using Spark framework to attribute the revenue to the ad. We deep-dive into the project, the corresponding ETL pipelines, the attribution models and the ML alogrithms used.


Andreas Jäck (ProsiebenSat.1 Digital Data GmbH)
Dr. Rostyslav Shevchenko (ProSiebenSat.1 Digital Data GmbH)