Welcome to Matthew Lee's Data Science Blog
Enthusiastic about applying data science and machine learning on
real world problems for practical and data driven results.
Currently seeking data scientist & machine learning engineer positions.
About Me
I've been always curious about everything and all about problem-solving.
I can't count how many broken objects I disassembled to understand how
they worked and try to fix them myself as a kid. After getting a Bachelor of
Science in Mechanical Engineering from University of Texas at Austin in 2015, I
worked in manufacturing industry.
While working in the field, I encountered
a problem for which conventional solutions were inadequate, but maching learning
approach was very promising. Since then, I pursued self-teaching data science and
machine learning while doing personal projects and competing in Kaggle where I
placed in top 2% and 9% in two competitions. Over time, I fell in love with data
science and also realized limitations in resources to apply data science skills with
the position I held at the time. Eventually, I decided to fully pivot into data science
and I would like to share some of my end-to-end projects I completed in Portfolio section below.
Portfolio
In this project, I trained LSTM models to generate melody and percussion, which are then combined to generate EDM music with sprinkles of Classical music. Then, neural network classifier was built to gauge output from LSTM models. Finally, interactive Flask app was built to utilize pretrained models to easily generate and play songs.
I built a tree based classifier LightGBM to classify a person's emotion based
on his/her actual voice. I used LightGBM instead of neural network architecture
such as CNN with RNN or LSTM since non-NN models are generally quicker than NN
at training and predicting, and are more fit for implementation in real world
applications. However, a lot of manual feature engineering was necessary to achieve
meaningful results since non-NN models don't have automatic feature engineering
from filters in CNN architecture.
Link to Project
I applied natural language processing (NLP) on news articles to perform topic
modeling using bag-of-words approach and sentiment analysis using open source
modules. Topic modeling gives a very concise visual for the user to understand
topics and trends revolving around Bitcoin and cryptocurrency over time. Also,
features created from sentiment analysis were combined with other features
(Bitcoin's open price, close price, volume, etc.) to build a LSTM model to predict
Bitcoin price.
Link to Project
