Value Stream Repair Using Graph Structure Learning

Value streams are attributed graphs used for modeling and simulating production processes. We suggest a machine learning-based approach to identify and repair modeling errors in value streams, specifically incorrect edges or product annotations. Our approach recasts graph attribution as a link prediction problem and uses graph-based features describing the local constellation in the value stream, such as the classes of successor and predecessor nodes or the product consistency in the material flow. By wrapping our model – which suggests single repair steps – into a beam search process, we can derive entire repair sequences. An expert study shows that for all 16 constellations tested, our model suggests the right changes to repair typical errors. Furthermore, our experiments based on five simultaneous random edge corruptions on a set of 70 value streams achieves an average precision up to 96.4%.