research · 2024
IMDB Sentiment Capstone
Team of five. My part: data cleaning and feature engineering, plus contributions across the project.
- Python
- Bidirectional LSTM
- NLP
Summary
A five-person capstone from the IBM Generative AI and Machine Learning program, Batch 7. The model classifies the sentiment of an IMDB review from its text.
Problem
Star ratings can be biased and do not always reflect what the review text actually says, so a rating alone is a weak label for sentiment.
Approach
Data cleaning, a word tokenizer, word embeddings, a Bidirectional LSTM architecture, and then evaluation. The dataset is TensorFlow Datasets `imdb_reviews`.
Outcome
Accuracy, recall, and F1 all reached 88%, balanced across the positive and negative classes.
Honest limitations
The 88% figure is reported on the evaluation split. It is not confirmed as a held-out test-set result, so it is not described as one.
Gallery
Next project
Rice Production Clustering