I loaded the dataset and applied NLP preprocessing steps such as tokenization, cleaning, and normalization.
CodeML Hackathon: Developing AI Models to Classify Product Feedback
- Featured
- AI
Introduction
What was this project about?
This project was developed during CodeML Hackathon 2020, a 48-hour hackathon organized by PolyAI. The challenge was to predict whether product feedback was positive or negative.
My Role
I led the development of the machine learning pipeline, including data preprocessing, text feature extraction, baseline model training, and experimentation with pre-trained language models.
Technical Details
Machine Learning: Python, Scikit-learn
Dataset: Product feedback comments (text) labeled as positive or negative.
Timeline
CodeML Hackathon, 2020
Links
Problem
Companies receive large amounts of product feedback that are difficult to classify manually.
Could an AI model predict if a new product comment is positive or negative?
Solution
I built machine learning models to classify product feedback as positive or negative.
- I created a baseline model with TF-IDF features and Logistic Regression.
- I implemented and tuned a Google ELECTRA model using SimpleTransformers.
- The ELECTRA-based model reached 3rd place on the Kaggle leaderboard for the challenge.
Process
I transformed the comments into TF-IDF features and trained a Logistic Regression model as a first baseline for positive vs. negative feedback classification.
I then implemented a second version using Google ELECTRA through the SimpleTransformers library and tuned the model to improve performance.
I submitted my predictions and compared my performance on the Kaggle leaderboard, where my ELECTRA-based approach ranked 3rd for the sentiment classification challenge.
Impact
3rd place
Our ELECTRA-based model reached 3rd place on the Kaggle leaderboard for the sentiment classification challenge.
ELECTRA model
The project gave me hands-on experience with a pre-trained language model and showed how it could outperform a simpler baseline.