Recipe card · live prediction
Slow cooker pulled pork
Slow
pork shoulder, brown sugar, smoked paprika, garlic powder, onion…
- Quick3.6% Quick
- Medium6.5% Medium
- Slow89.8% Slow
Words that tipped it
- hours
- slow cooker
- slow
- pork
- cooker
COMP30027 Machine Learning · University of Melbourne · 2021
How long will it take to cook this?
In 2021 our team trained classifiers to guess whether a Food.com recipe is quick, medium or slow from its name, ingredients and method. The code was lost, so this site rebuilds the whole study from the report, reruns it on public data and lets you test it on your own recipes.
- Training recipes
- 40,000
- Classes
- 3
- Models
- 5
The brief
What the coursework asked
Project 2 of COMP30027 gave pairs a sample of Food.com recipes and asked them to build, compare and explain supervised classifiers for preparation time.
Ingredients
The data
Each recipe came with its name, ingredient list, method steps and their counts, plus staff-made text encodings: Doc2Vec vectors and bag-of-words matrices.
Method
The task
Predict one of three labels, then write a conference-style report on why each model behaves the way it does. Pairs had to build at least three models, one of them a stacking ensemble.
Chef's note
The catch
Only 5% of recipes are slow. A model can score well while barely ever saying “slow”, so accuracy alone hides the hardest class.
- Quickon the table in under 35 minutes
- Medium35 minutes to 4 hours
- Slowmore than 4 hours, often overnight
Results
2021 report against the 2026 rebuild
The rebuild reruns the report's five models with the same features and settings on a public copy of Food.com, sampled to the report's exact class counts. 10 of the 11 findings the report draws carry over, including the model ranking. The exception: the C curve peaking at C = 1. The scores come out lower, because the recipes, labels and Doc2Vec vectors are not the staff's.
Five-fold CV
Cross-validated accuracy
- 2021 report
- 2026 rebuild
SettingAccuracy2021 → 2026
Gaussian Naive Bayes350 Doc2Vec features0.6255 → 0.5422
Logistic Regression350 Doc2Vec features0.7201 → 0.6780
SVM350 Doc2Vec features0.7302 → 0.7128
Random ForestSteps bag of words0.7885 → 0.7661
Stacked Naive BayesGNB + LR + SVM labels0.7297 → 0.7103
Best model, rebuilt (Random Forest)
76.6%
Report: 78.9% · 15 repeated folds: 76.5% (76.1% to 76.9%) · 10,000 unseen recipes: 76.4% (75.5% to 77.2%). Intervals are 95%.
Stacking against its best ingredient
71.0%
SVM alone: 71.3%. Paired over 15 folds the stack scores −0.4 pts against the SVM (−0.6 to −0.2 pts), so like 2021, stacking does not help.
Explore
Six ways in
PredictPaste a recipe and see the class, the odds and the words behind them.Open ModelsKernels, C and gamma, tree counts and confusion matrices, side by side.Open EvaluationEvery score with its interval, paired tests between models and a calibration check.Open LLM evaluationBring your own key and test an LLM against the 2021 models on the same recipes.Open StackingCast the three level-0 votes yourself and watch the Naive Bayes judge decide.Open DataWhere the recipes come from and how minutes become three classes.Open
About this project
Then and now
2021
The coursework
- Subject
- COMP30027 Machine Learning
- University
- The University of Melbourne
- When
- 2021 Semester 1, Project 2
- Team
- Sunchuangyu (Rin) Huang and a project partner
- Original stack
- Python, scikit-learn, staff Doc2Vec and CountVectorizer features, Kaggle in-class competition
2026
The revival
- Pipeline
- Python 3.12 with uv, scikit-learn and gensim, rebuilt from the report
- Web
- Next.js, TypeScript, Tailwind CSS, hand-drawn SVG charts
- In the browser
- TF-IDF, Logistic Regression, Naive Bayes and the stacking judge, ported to TypeScript and tested
- Rigour
- Repeated cross-validation, intervals, McNemar tests and calibration, checked against SciPy and R
- AI, optional
- Bring-your-own-key LLM evaluation with an audit log that stays in your browser
- Code
- Private GitHub repository (available on request)
More on provenance, credits and academic integrity on the about page.