2026-24679-coffee-autogluon-regressor

Coffee Price-per-Ounce AutoGluon Regressor

Purpose

This model was created for CMU 24-679 Homework 2. It predicts the continuous price per ounce of a retail coffee bag from package attributes.

Data

Source dataset: ssg1/coffee-bags-tabular

The dataset was created by a classmate for Homework 1. I used the published splits without resplitting:

• Training: 345 rows
• Validation: 5 rows
• Test: 6 rows

The training split contains training originals and their augmented versions. Validation and test contain held-out original samples.

Features and Target

Target:

• price_per_oz ($/oz)

Features:

• altitude_masl
• weight_g
• altitude_stated
• origin
• processing
• roast_level
• grind

price_usd was excluded because price_per_oz is derived from price and package weight, which would introduce target leakage. brand, identifiers, augmentation metadata, and the separate is_premium classification target were also excluded.

Preprocessing

The existing train, validation, and test splits were retained.

Numeric features were converted to numeric values. Categorical features were stored as categorical variables and handled automatically by AutoGluon.

No information from the test split was used for model selection.

AutoML Training

Framework: AutoGluon Tabular

Task: Regression

Model-selection metric: Root Mean Squared Error (RMSE)

AutoML preset: medium_quality

Compute budget: 300 seconds

Bagging folds: 0

Stack levels: 0

Models evaluated: WeightedEnsemble_L2, NeuralNetTorch, XGBoost, RandomForestMSE, ExtraTreesMSE, LightGBMLarge, LightGBM, LightGBMXT, NeuralNetFastAI

Best model: WeightedEnsemble_L2

Best-model hyperparameters: {'use_orig_features': False, 'valid_stacker': True, 'max_base_models': 0, 'max_base_models_per_type': 'auto', 'save_bag_folds': True, 'stratify': 'auto', 'bin': 'auto', 'n_bins': None, 'vary_seed_across_folds': False, 'custom_splits': None, 'model_random_seed': 0}

The best model was selected using validation RMSE. The test set was used only after model selection.

Test Results

• RMSE: 0.650 $/oz
• MAE: 0.447 $/oz
• Approximate 95% bootstrap interval for RMSE: 0.111 to 0.918 $/oz

The bootstrap uncertainty calculation used random seed 24679 for reproducibility.

Because the test split contains only 6 samples, these results have substantial uncertainty.

Hardware

Training was performed in Google Colab.

The AutoGluon training run used a CPU runtime with no CUDA GPU available.

Compute budget for the AutoML search: 300 seconds.

Limitations

The number of independently collected coffee bags is small, especially the validation and test sets. Performance on these samples may not generalize to coffee products from other stores, geographic regions, brands, or time periods.

The training set also contains augmented versions of original samples. These increase the amount of training data but do not represent new independently collected coffee products.

Coffee prices can change over time, so a model trained on this dataset should not be interpreted as a general or permanent coffee-pricing model.

Ethical Considerations

The dataset describes retail coffee products rather than people and does not contain sensitive personal information.

Predictions should not be treated as authoritative estimates of fair market price because the dataset is small and geographically and temporally limited.

License

MIT, following the source dataset license.

AI Usage Disclosure

ChatGPT was used as a coding copilot to assist with syntax, library APIs, and adapting the course-provided AutoML lecture code to this regression task. I used it to help identify the changes needed for the selected dataset and workflow. ChatGPT also helped draft and polish the Model Card based on my inputs, implementation choices, and final results.