ML Review Classifier

Version: v0.1.0-baseline

Status: Experimental baseline, not production-ready.

A reproducible sentiment classification model developed with Python, scikit-learn, TF-IDF and Logistic Regression.

Model Description

This model classifies English-language reviews into two categories:

  • 0: Negative
  • 1: Positive

It uses a scikit-learn pipeline combining TfidfVectorizer and LogisticRegression.

Dataset

Source: UCI Sentiment Labelled Sentences

The original dataset contains 3,000 reviews from Amazon, IMDb and Yelp.

Following preprocessing and duplicate removal, 2,979 examples remained.

Split Samples
Training 1,784
Validation 593
Test 602

Dataset source: https://archive.ics.uci.edu/dataset/331/sentiment+labelled+sentences

Dataset citation: Kotzias, D. (2015). Sentiment Labelled Sentences. UCI Machine Learning Repository. DOI: 10.24432/C57604.

The dataset is distributed under CC BY 4.0.

Model Architecture

  • Feature extraction: TF-IDF
  • N-grams: 1โ€“2
  • Minimum document frequency: 2
  • Classifier: Logistic Regression
  • Maximum iterations: 1,000
  • Random state: 42

Evaluation Results

Validation

Metric Result
Accuracy 79.76%
Macro F1 79.76%

Final Test

Metric Result
Accuracy 81.56%
Macro Precision 81.56%
Macro Recall 81.56%
Macro F1 81.56%

The final test was performed on 602 held-out reviews.

Confusion Matrix

Actual / Predicted Negative Positive
Negative 247 54
Positive 57 244

Total correct predictions: 491 / 602.

Usage

Install the dependencies listed in the environment section.

Download tfidf_logistic_regression.joblib from this repository.

import joblib

model = joblib.load("tfidf_logistic_regression.joblib")

reviews = [
    "The food was amazing!",
    "The service was terrible."
]

predictions = model.predict(reviews)

print(predictions)

Security: Joblib files use pickle-based serialization and can execute arbitrary code when loaded. Only load this artifact from a trusted source after verifying its integrity.

Reproducibility

Environment used:

  • Python 3.14.7
  • scikit-learn 1.9.1
  • NumPy 2.5.5
  • SciPy 1.18.1
  • Joblib 1.6.0

Model artifact SHA-256:

8395cce17d942589785e2afaec14d1fa267244ea5b9a248f13b2619ad8047d5

Reproducing the training experiment requires the source code and dataset preparation procedure in the GitHub repository.

Limitations

  • Supports English-language reviews only.
  • Binary classification; no neutral sentiment class.
  • Trained on a small, relatively old dataset.
  • Performance on real-world Brazilian customer reviews has not been established.
  • Predictions may be incorrect on sarcasm, ambiguous statements or unfamiliar vocabulary.
  • Evaluation results do not demonstrate production readiness.

Source Code

GitHub โ€” ml-review-classifier

Author

Lucas Miranda

Software Engineering | Machine Learning | Generative AI

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support