Deep Belief Networks Used to Generate Meaningful Patterns

Trappl, Juraj

This thesis investigates quality-guided Deep Belief Networks for recipe generation. We collect 18,255 recipes from major cooking websites and de- velop three representation schemes: one-hot encoding, FastText embeddings, and BERT embeddings. To address survivorship bias in published recipes, we generate synthetic negative examples using culinary anti-patterns. Quality prediction models are trained to distinguish between good and poor ingredi- ent combinations. One-hot encoding emerges as the optimal choice, balancing architectural compatibility with direct interpretability. We integrate quality predictors into Deep Belief Network training through supervised ne-tuning, adding a regression head to predict recipe quality scores. Systematic evalua- tion of nine architectures reveals that uniform con gurations with moderate capacity achieve the best balance between representation learning and gen- eralization. Manual evaluation of 100 generated recipes con rms the system produces plausible combinations, while also revealing gaps between statistical optimization and culinary coherence.

guest :: login Digital Repository
		Search		Submit		Help		About