Project Status: Mastery Milestone Completed
A fundamental challenge for modern enterprises is that 80% of data exists in unstructured, qualitative formats—such as product reviews, emails, and support tickets. This data is invisible to traditional analytics until it crosses the engineering threshold into quantitative spatial representations.
I built this project to solve that exact problem. My mandate was to engineer a strict, automated, multi-stage machine learning assembly line capable of ingesting raw, chaotic human language and generating high-accuracy probability distributions of emotional polarity.
This repository is built for production robustness, reproducibility, and clarity. It strictly follows a 5-step engineering lifecycle:
My pipeline dynamically manages local zip file extraction to structured data directories, handling potential character encoding issues automatically.
Unprocessed text is full of syntactic noise, creates exponential redundancy, and leads to memory-based system crashes. To resolve this, I implemented three critical constraints:
A. Character Normalization & Tokenization: Standardization through regex-based URL removal, special character stripping, and strict lowercasing reduces the feature dimensions the model must learn.
B. Resolving the Stop-Word Trap: Default NLP dictionaries often contain negations. I implemented explicit, set-based union operations to ensure negations (like "not", "nor", "never") are preserved, preventing prediction failures on phrases like "not good".
C. Accurate, Morphological Reduction (Lemmatization): My pipeline uses NLTK's WordNetLemmatizer. To correct the pitfall where the lemmatizer defaults to noun reductions (e.g., failing to reduce "went" to "go"), I implemented POS tagging. My mandatory engineering rule is to map Penn Treebank tags to WordNet tags, ensuring accurate morphological reduction for verbs, adjectives, and nouns.
This step converts English text into mathematics.
- Weight Calculation: I implemented a combined TF (how often does the word appear in this document?) and IDF (how rare is the word across all documents?) logic to assign weights, ensuring non-sentiment fluff words receive low weights while rare, high-sentiment words (e.g., 'terrible') receive high weights.
- N-Gram Capturing: I included Bigrams (2-word relationships) to explicitly capture negated sentiment (e.g., 'not good'), while setting boundaries on
max_featuresto manage vocabulary explosion and dimensionality issues.
Vectorizing 10,000 reviews against a vocabulary of 20,000 terms results in a matrix that is 99% empty zeros. Attempting to store this as a dense array exhausts RAM, causes fatalities on system crashes, and operates at an abysmal O(N³). Following best practices for training linear classifiers, my implementation compresses this void using SciPy's Compressed Sparse Row (CSR) format, ensuring industry-standard efficiency and memory usage.
With my text mathematically vectorized, I benchmarked two robust models:
- Multinomial Naive Bayes: A model that is best suited for perfectly balanced datasets and applies Laplace Smoothing to handle zero-frequency problems in testing.
- Linear Support Vector Machine (SVC): This classifier also proved effective in identifying polarity boundaries.
The pipeline dynamically generates and saves comparative visualizations, providing immediate clarity on results.
My script confirms a highly balanced distribution of sentiment in the training data, with slightly over 8,000 instances of each class (Positive and Negative). This balance prevents extreme bias during training.
| Model Performance Comparison Matrix (Confusion Matrices) | ROC Curve Comparison (Benchmarking Metrics) |
|---|---|
 |
 |
Side-by-side matrices demonstrate exceptional accuracy for both models. SVM correctly predicted 1,869 examples. NB slightly behind, with 1,862 examples.
ROC benchmarking shows both models are exceptionally strong, achieving a matching AUC (Area Under Curve) of 0.96. NB and SVM are virtually overlayed, demonstrating excellent separability.
============================================================
NAIVE BAYES METRIC REPORT (Accuracy: 0.88)
============================================================
precision recall f1-score support
Negative 0.89 0.87 0.88 1001
Positive 0.88 0.90 0.89 1103
weighted avg 0.88 0.88 0.88 2104
============================================================
SVM CLASS METRIC REPORT (Accuracy: 0.89)
============================================================
precision recall f1-score support
Negative 0.88 0.89 0.88 1001
Positive 0.90 0.89 0.89 1103
weighted avg 0.89 0.89 0.89 2104