Evaluate taste, nutrition, and sustainability separately
The three targets have different label sources and failure meanings. Separate models preserve those differences before weighting them in the final ranking.
Machine learning · Recommendation
PantryWise
A Georgia Tech team recommendation project that explores taste, nutrition, sustainability, and visual similarity from the ingredients a user already has.
- Role
- Georgia Tech CS 4641 team project · Modeling and evaluation
- Technologies
- LightGBM · K-Means · TF-IDF · ResNet50 · Pandas
01
Context
Recommending from available ingredients while accounting for taste, nutrition, sustainability, and visual similarity requires different datasets and evaluation criteria. The project compares TF-IDF/K-Means ranking, LightGBM models, and ResNet50 image embeddings, then examines how far offline evaluation can explain actual preference.
02
Engineering
Ingredient text is normalized and pantry coverage narrows the candidate set before TF-IDF similarity and model scores produce a ranking. Separate LightGBM tasks learn Food.com ratings and Open Food Facts Nutri-Score and Eco-Score signals. ResNet50 embeddings retrieve visually similar recipe images.
My contribution
- Contributed to modeling that compared features and evaluation flows across several recommendation approaches.
- Found and removed features that exposed the diet label, then reran the complete evaluation.
- Documented strong and weak results together in the team report and public demo.
Key technical decisions
Remove the leakage behind 99.9% accuracy
The initial score was implausibly high. plant_ratio and meat_ratio effectively exposed the vegan label, so both were removed before retraining on TF-IDF ingredient semantics.
03
Validation
After removing leakage features, vegan accuracy measured 92.5%. At a 0.8 threshold, precision was 91.2% but recall fell to 44.1%, which is why the label-generation process was investigated instead of accepting the original score.
The final LightGBM pipeline measured R² 0.923 on the nutrition hold-out and Spearman 0.836 on sustainability ranking. Taste classification reached 0.625 accuracy, nearly identical to the 0.622 majority baseline. The report preserves the difference between strong and weak targets.
Next steps
- Current results come from a course dataset and offline split. A user study should test whether pantry coverage and recommendation explanations improve actual selection and repeat use.
- Nutrition and sustainability labels approximate recipe properties from external product data. Future work should obtain recipe-level ground truth or represent label uncertainty, while the taste model should wait for real interaction data before retraining.