This project explores several strategies to develop the most accurate models to solve two problems: identifying a cuisine when a recipe is given and inferring an ingredient missing from a recipe. Therefore, we focus on two tasks: classification and completion. As an exploratory stage, we implement two categories of models: baseline and target. Baseline models consist of Logistic Regression, Random Forest Classifier, Multinomial Naive Bayes, Linear Support Vector Classifier, and Collaborative Filtering, while our target model is based on Transformer architecture. We propose a model named RecFormer - Recipe Transformer - that can perform both classification and completion tasks in a single model with similar or better accuracy than our baseline models. Finally, we employ model blending to achieve maximum performance. An ensemble of 2 models, LSVC and RecFormer, achieves 79.20% and 14.74% validation accuracy for classification and completion tasks.
Further details are available in the report and presentation.
You must have the pytorch, tensorflow, numpy, pandas, sklearn, matplotlib, jupyterlab libraries installed.
- Train and evaluate RecFormer: recipe_transformers.ipynb
- Train and evaluate baselines: baselines/BaseLineModelExperiment_final.ipynb