Combining CBR and Deep Learning to Generate Surprising Recipe Designs

This paper presents a dual-cycle CBR model in the domain of recipe generation. The model combines the strengths of deep learning and similarity-based retrieval to generate recipes that are novel and valuable (i.e. they are creative). The first cycle generates abstract descriptions which we call “design concepts” by synthesizing expectations from the entire case base, while the second cycle uses those concepts to retrieve and adapt objects. We define these conceptual object representations as an abstraction over complete cases on which expectations can be formed, allowing objects to be evaluated for surprisingness (the peak level of unexpectedness in the object, given the case base) and plausibility (the overall similarity of the object to those in the case base). The paper presents a prototype implementation of the model, and demonstrates its ability to generate objects that are simultaneously plausible and surprising, in addition to fitting a user query. This prototype is then compared to a traditional single-cycle CBR system.