02 — PROBLEM
Problem
Turn large-scale short-form text into a deployable sentiment classification service while keeping preprocessing consistent between training and inference.
PROJECT / 02
Sentiment Analysis
A sentiment classification system using the Sentiment140 dataset. Bernoulli Naive Bayes and Logistic Regression were compared, with Logistic Regression selected for production after evaluation.
SYSTEM PATH
01DATA
02MODEL
03EVAL
04API
05CONTAINER
06CI/CD
07CLOUD
1.6MtweetsSentiment140 dataset
40Kbalanced training subsetused for training
76.60%accuracyselected production model
0.840ROC-AUCselected production model
500KTF-IDF max featuresunigrams + bigrams
01 — SYSTEM ARCHITECTURE
How the pieces connect.
The case study keeps architecture data-driven too. Each project can describe a different system path without changing the renderer.
01
Sentiment140 · 1.6M tweets
02
40K balanced training subset
03
NLP preprocessing pipeline
04
TF-IDF · 500K max features
05
Model comparison → Logistic Regression
06
FastAPI → Docker → Docker Hub
03 — APPROACH
Solution
Build a reusable NLP preprocessing pipeline covering stopword, URL, punctuation removal, tokenisation, stemming, and lemmatisation, then vectorize text with TF-IDF using unigrams and bigrams.
04 — RESULTS
Results
Logistic Regression was selected for production at 76.60% accuracy and 0.840 ROC-AUC after comparison with Bernoulli Naive Bayes.
05 — IMPLEMENTATION
Built for the path to deployment.
01
Preprocessing is shared between model training and inference to keep input handling consistent.
02
TF-IDF represents unigrams and bigrams with a 500K feature cap.
03
Confusion matrix and ROC curve analysis were used for evaluation.
04
The selected model is exposed with FastAPI and containerized with Docker for Docker Hub publication and AWS EC2 deployment.
06 — STACK & DEPLOYMENT
FastAPI + Docker + Docker Hub; prepared for AWS EC2 deployment