This page showcases some of my work related to Data Science
This project focuses on applying Machine Learning techniques to categorize a piece of text into three distinct categories, which are "hate speech", "offensive language" and "neither". The classifier and a prediction ... schema with visualization are deployed on a web app hosted on HEROKU server on a free DYNO plan, which means that the initial boot of the website will be delayed by about 30 seconds.
During the analysis of the subject #quarantine, I collaborated with two fellow MSc students to bring the project into life. Specifically, I collected data and extracted ... features from Instagram. I created an advanced natural language pre-processing framework that proved really benefitial in the task of emerging topics. On the latter, I utilized the LDA algorithm and detected various topics based on their day-to-day popularity. Additionally, I compared the performance of two methods in the task of like prediction, which are a natural language approach and a numeric/statistical feature approach. Finally, I created interactive graphs for each task that made sense and compiled the overal results of this analysis into a web page that showcases some of the results. There is an online version of the web app hosted by HEROKU on a free DYNO plan, which means that the initial boot of the website will take about 30 seconds to load.
In this project I utilized the database PostgreSQL in order to create an image search web application. Images in the database are saved as vectors produced by ... algorithms that describe images, such as SIFT, histogram, SIFT with BOVW. The web app is hosted online by HEROKU on a free DYNO plan, which means that the initial boot of the website will need about 30 seconds to load.
In this project I reviewed the vast majority of cutting-edge techniques that handle class imbalance, an essential real world problem. I drew some conclusions about ... each method and got some guidlines that indicate what method to deploy depending on the problem at hand. I experimented with under-sampling, over-sampling, cost-sensitive-learning and ensembles with special mechanisms that handles class imbalance. Finally, I joined forces with two fellow MSc students to combine our subtask findings to get a complete model. That model utilizes an optimal subset of the dataset for training produced by active learning, handles the class imbalance efficiently, and offers proper interpretability of the results.
Extensive study of machine learning algorithms and data mining methods in view of detecting ironic text ... excerpts. The code is in Python and it features Data Cleaning, Natural language processing, Feature Engineering (extraction-selection), Machine Learning, Data Visualization, Deep learning (Tensorflow), and Dimensionality reduction.