iankamar/amazon-food-sentiment-models

Amazon Food Sentiment Models

Bag-of-words and TF-IDF logistic regression models, each trained with and without balanced class weights.

Try the demo

Results

Model	Test accuracy	Test macro F1	Test ROC AUC
Bag of words	93.12%	0.8682	0.9581
Bag of words balanced	93.05%	0.8750	0.9574
TF-IDF	92.06%	0.8290	0.9663
TF-IDF balanced	91.76%	0.8644	0.9683

The default, Bag of words balanced, is selected by validation macro F1, not test results. The constant training-majority baseline has test accuracy 83.69% and macro F1 0.4556.

Data and evaluation

The source has 568,454 reviews. Three-star reviews are excluded. After normalization, 30 conflicting-label rows and 162,446 duplicate rows are removed. A uniform 60,000-review sample is selected from the 363,338 remaining non-neutral reviews using seed 42. The chronological split, with whole-day boundaries, contains 41,973 training, 9,020 validation, and 9,007 test reviews.

Both vectorizers are fitted only on training text. Features include unigrams and bigrams, with a 20,000-feature cap and minimum document frequency 3. Logistic regression uses C=1, liblinear, and a fixed 0.5 decision threshold. The comparison uses class weighting, not synthetic reviews or SMOTE. Reviews from the same product may appear in different time partitions; this is not a new-product generalization test.

Run

Use Python 3.12 and obtain Reviews.csv from Kaggle.

python -m pip install -r requirements.txt
python train.py --csv path/to/Reviews.csv

To use a saved pipeline:

import joblib
from preprocessing import clean_text

models = joblib.load("models.joblib")
model = models["Bag of words balanced"]
print(model.predict_proba([clean_text("The flavor was excellent and delivery was quick.")]))

The output columns are negative and positive. Load joblib files only from trusted sources. metrics.json records the source hash, split IDs, and evaluation; test-predictions.npz supports metric checks. Raw reviews are not redistributed.

Data credit: Stanford SNAP Fine Foods, Julian McAuley and Jure Leskovec; preserve the original data terms. Results describe sentiment prediction, not causal drivers of satisfaction. This update fixes the original full-data vocabulary fit and adds the actual TF-IDF comparison.