Processing Recipe Data with Transformers

Summary & Methodology Analysis

Recipe data consists of heterogeneous fields organized into sequential procedural steps with variable schemas and dependencies, which existing tabular learning methods handle poorly due to flattening and padding requirements. To solve this, the authors introduce RecipeNet, a hierarchical transformer architecture. Standard mechanics process sequences using standard Transformer encoder architectures, which use self attention to weigh the relationships between tokens, along with linear projections or embedding tables for features. RecipeNet improves on this by using field level tokenization, which maps heterogeneous numerical and categorical fields into a shared latent space and combines value, step-position, and field-identity embeddings into token representations.

The architecture relies on a multi-stage approach. First, a step fusion encoder applies a transformer encoder over observed field tokens within each step, prepending a special step classification token to capture intra-step field interactions. Second, a recipe-level sequence encoder processes the sequence of step embeddings using a second transformer encoder with a prepended recipe classification token to capture long-range procedural dependencies. Finally, the final recipe representation is passed to a task-specific prediction head for classification or regression tasks. This model is evaluated alongside alternatives like XGBoost, CatBoost, TabNet, NODE, standard Transformer, Set Transformer, FT-Transformer, TabTransformer, Deep Sets, RNNs, and LSTMs across several datasets including solid-state reactions, sol-gel precursor synthesis, solution synthesis, and the Text-Mined Synthesis Project.

Regarding performance and limitations, the paper does not specify parameter counts, memory footprints, or dollar costs. However, it notes that RecipeNet reduces training time by approximately 18% compared with the fastest competing transformer-based baseline. The extracted facts contain no information regarding limitations, so the paper does not specify any specific limitations.