Welcome to my data science portfolio
Explore, model, and predict smartphone app usage behavior in this data project. Leveraging a rich real-world dataset from 12k+ users, I developed a novel multivariate time-series deep-learning model based on Hierarchical Attention Networks (HAN) to accurately forecast users' weekly app usage. The model outperforms baseline models, including persistence and simple GRU/LSTM cells. Uncovering insights into usage patterns and the potential for identifying problematic smartphone usage.
Explore the world of time series analysis and its practical applications in mobile app usage behavior. Leveraging data from 10k+ smartphone users, this project unveils insights into daily screen time, check frequency, and app usage trends. Discover key findings, including the superiority of the Facebook Prophet model in forecasting daily check numbers and adapting to change points compared to seasonal ARIMA.
Developed a state-of-the-art IMDb movie review sentiment classifier. Leveraging the power of the Hugging Face Transformer framework, this machine learning model, based on BERT, automatically categorizes reviews into positive or negative sentiments. Emphasizing MLOps practices, from data prep to deployment.
Explore the world of Deep Learning Image Colorization with this project. Using Convolutional Neural Networks (CNNs) and conditional Generator-Discriminator pipelines, including Attention U-Net and GAN architectures, we tackle the complex task of predicting color channels from grayscale inputs. Our project evaluates model performance using computer vision metrics like PSNR, SSIM, and LPIPS. Dive into the challenging realm of colorization with our exploration of L·a·b Color Space and discover the optimal model for grayscale image colorization.
Focusing on predicting car collisions in New York City using traditional machine learning techniques. This data-driven project investigates contributing factors to offer insights for accident prevention. Experience interactive data storytelling and visualization through thoughtfully designed dashboards.