Generating Training Data Sets for Machine Learning Approaches with GIPS

Machine Learning (ML) and its application is a research area that has become increasingly important, especially in the last decade. ML approaches in the field of supervised learning depend on labeled data sets for their training process. But, in some problem domains, the collection and generation of labeled training data can be hard because, for example, the underlying problem is in general hard to solve algorithmically. This paper proposes a conceptual framework for the generation of labeled training data sets for supervised learning approaches based on the GIPS framework. For this purpose, GIPS combines Graph Transformation (GT) with Integer Linear Programming (ILP) techniques to solve graph-based optimization problems to obtain labels for data points. A prototypical implementation is used to demonstrate the functionality of our solution in the context of a small-scale Virtual Network Embedding (VNE) example. The evaluation of our prototype shows promising results when compared to an optimal ILP-based implementation.