Learning Distributed Representation of Recipe Flow Graphs via Frequent Subgraphs

Recent rapid increase in health awareness is producing a large amount of user generated cooking recipes in online community sites. For the effective use of such cooking recipes, it is necessary not only to understand their meaning but also to extract certain structures among them, by paying attention to cooking steps in detail. One of the most precise representations of cooking procedure is the recipe flow graph that is a directed acyclic graph having recipe terms in vertices and their relations in edges. In this paper, as a preliminary attempt for acquiring a new vector representation reflecting various aspects of cooking procedures, we propose a simple method to learn a distributed representation of recipe flow graphs using frequent fragments of cooking procedures. Experiments using real world dataset are conducted to compare the distributed representation of recipe flow graphs and that of recipe texts. As a result, we confirm that the proposed representation can capture the difference among recipes well, and it is suitable for the classification tasks.