SMS Spam Detection using Machine Learning with TF-IDF and Naive Bayes

Spam SMS still drains attention and enables fraud: prize scams, fake bank alerts and phishing links arrive as ordinary-looking texts, and rule-based filters miss the wording variations spammers constantly invent. Machine learning treats this as a text-classification problem — learn the statistical signature of spam from thousands of labeled examples, then score new messages by how closely they match. Naive Bayes pairs naturally with TF-IDF here: TF-IDF turns each message into a weighted word vector, and Naive Bayes converts those weights into a clean probability for spam vs ham, while remaining fast enough to run in a browser and transparent enough that every decision can be traced back to individual words. Students often train a black-box model they cannot explain; this project does the opposite — a classical, inspectable pipeline with token-level log-odds explanations, so the "why" behind each verdict is visible and viva-defensible.

1. The 5,574 labeled SMS messages are loaded from the SMS Spam Collection and split into a stratified training and test set so both classes are represented.
2. Text is preprocessed: lowercasing, tokenization on word boundaries, and stopword removal, keeping digits since phone-number fragments are genuine spam signals.
3. A TF-IDF vectorizer converts each message into a weighted vector — common words are down-weighted, distinctive words like "prize" or "claim" are up-weighted.
4. A Multinomial Naive Bayes classifier learns per-class word likelihoods with Laplace smoothing (alpha = 1.0) and class priors from the training vectors.
5. The held-out test set is scored once: precision, recall, F1 and the confusion matrix are computed and written into the report.
6. In the demo, a typed message passes through the same vectorizer and model; the app displays the verdict, class probabilities and the top contributing tokens with their log-odds.

Tech stack:

• Python 3, scikit-learn
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• TF-IDF vectorizer · Multinomial Naive Bayes
• Single-file HTML/CSS/JS demo app
• SMS Spam Collection (UCI) dataset

Dataset & model details

• Dataset: SMS Spam Collection — UCI Machine Learning Repository (Almeida & Hidalgo, 2011). 5,574 English SMS messages: 4,827 ham, 747 spam (approximately 13.4% spam). Tab-separated file with label and message-text columns.
• Task: Binary text classification; input = raw SMS text, output = spam/ham label with class probabilities.
• Model: TF-IDF vectorization (word tokens, stopword removal) feeding a Multinomial Naive Bayes classifier with Laplace smoothing (alpha = 1.0); prediction in log space.
• Metrics: Precision, recall, F1-score per class and confusion matrix on a held-out stratified test split. Literature baselines for TF-IDF + Naive Bayes on this dataset are commonly reported in the high-90s for accuracy and high-80s to mid-90s F1, varying with split and preprocessing — reported here as literature context only, not a claim for the delivered build. Final performance depends on the split, preprocessing and configuration used in the buyer's training run.

• [TF-IDF + Multinomial Naive Bayes classifier] The core model: TF-IDF vectorization over word tokens with Laplace-smoothed Multinomial Naive Bayes, implemented with scikit-learn and fully documented from preprocessing to prediction.
• [Live classify-any-SMS web demo] Paste any message and get an instant spam/ham verdict with class probabilities, rendered as a single-file app that runs offline in the browser.
• [Token-level explanations] Every verdict is broken down into per-token log-odds contributions, so the student can show exactly which words pushed the decision toward spam or ham.
• [Full training notebook] Dataset loading, stratified train/test split, vectorizer fitting, model training and evaluation in one reproducible Jupyter notebook.
• [Honest evaluation report] Precision, recall and F1-score per class plus a confusion matrix, with a written interpretation of where the model errs — typically aggressive prize-scam spam caught easily, short ambiguous messages harder.
• [Imbalance-aware methodology] The dataset is ~13% spam, so the build treats spam recall and the precision/recall trade-off explicitly instead of leaning on misleading raw accuracy.
• [Viva-ready theory notes] Concise write-ups of TF-IDF, Bayes' theorem, smoothing, log-space computation and the independence assumption, mapped to viva questions.

• Complete training & evaluation Jupyter notebook
• Trained vectorizer and model with preprocessing code
• Live classify-any-SMS demo app (single-file, offline-capable)
• Per-class precision/recall/F1 report with confusion matrix and interpretation
• Project report PDF (background, TF-IDF/NB theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (Bayes theorem, smoothing, TF-IDF, evaluation metrics)