Robin Hirt

Machine Learning (ML) can be used to enrich existing products and services by introducing novel ML-based features. However, these products or services are often used by several customers. In order to train the required ML models, providers need to get their hands on customer data and centralize it - which is often not possible due to data sensitivity. In such situations where data is scattered across distributed IT systems, new techniques like sequential transfer learning can help to train high performing ML models. In this talk, we show a solution for creating and deploying ML models on federated systems using an innovative sequential transfer technique that outperforms existing federated learning frameworks. We present the case of an ERP provider who utilizes this technique to introduce a sales forecast as an ML-based feature in their software. Users of that system experienced a significant decrease in forecast error of up to 30%.