I'm Matthew Wilchek

This project attempted to develop a proof of concept to classify one of seven emotions from pictures of people that include Anger, Contempt, Disgust, Fear, Happy, Sadness and Surprise. After training a Convolutional Neural Network, we then applied our model to images of ourselves like in the example above. To view our approach check out our presentation of the project here.

This project attempted to develop a proof of concept to predict any stock performance for current day and following day, based on history and the assumption news sentiment in a given day has a significant relationship. The above shows a Web Application developed in R to visualize the results of our project. To view our approach check out our presentation of the project here.

Stock and News Web Scraping

In order to create a sufficient supervised machine learning algorithm our project required sufficient data collection. We extracted stock data using an API from Yahoo Finance and news headline data from a Reddit API called PRAW. Once the data was collected on a daily basis, we executed a separate sentiment algorithm on the news data using R (that can be found here), and joined all to the master dataset for the algorithm. To view our data collection script click here.

Modeling

In order to account for changing sentiment on any given day from our previous sentiment algorithm, we needed to identify the most significant sentiment for the current and predicted following day for our primary models. To do so, we leveraged Python Sklearn's library for a number of machine learning libraries such as Random Forest Regressor, Neural Net Regressor, and Support Vector Machine Regressor. Click here to view a plot for one of our results in a single execution here.

Finally after identifying the most significant sentiment in the given day, we were able to refine our models for a fit regression. As new data is collected and added to our master dataset, our train/test prediction model will continually be refined and accuracy improved. To check out our final algorithm, click here.

The team gathered climate data from Berkeley Earth, and focused on global land temperatures for the analysis. Data is available from 1750-2013. Our analysis and climate predictions are presented through ArcGIS and a R Shiny web application. All source code is availble in our Github repository here.

U.S. Temperatures

By using ArcGIS, I developed two interactive web map application to help visualize how CONUS temperatures changed over the years in major cities. The shapefiles used for the application can also be found in the Github repository here.

Below are where the 2 applications can be viewed:

• http://www.arcgis.com/apps/View/index.html?appid=a29ddc77e43a4013a982d08d8344cb6b
• http://www.arcgis.com/apps/StorytellingSwipe/index.html?appid=bd128209373d4b8a850f05310cbe0a6e#

Analyzing Models and Interpreting Results

The 1975-2013 model is Y = -49.148211+0.048966X, with both the coefficient and intercept statistically significant at the alpha <0.05 level. The overall model has a statistically F-statistic as well and is a good fit. The adjusted R^2 = 0.7813. While this R^2 is not as good as the first, it is still an acceptable R^2.

From an R^2 perspective, model 1 has the best result. However, from an interpretation standpoint, the coefficient for time in Model 1 is close to 0 and thus shows just a small rise in temperature. This model does not provide a good fit to recent temperatures while Model 4 does. To check out our predictions for future temperatures using Model 4, check out our web application here.

This was one of my early assignments as I started my graduate degree in Data Science. The The class focused on learning the epicycle of data analysis and using R Statistical Computing to strenghten quantitative analysis skills. To see the results of my analysis click here.

In this assignment I performed an exploratory data analysis on a population of women who were at least 21 years old, of Pima Indian heritage and living near Phoenix, Arizona, and were tested for diabetes according to World Health Organization criteria. The data was collected by the US National Institute of Diabetes and Digestive and Kidney Diseases. It comprises of 532 complete records after dropping the (mainly missing) data on serum insulin. More information on the dataset can be found here.