I created a round based natural selection simulator.

In each round, n herbivores are spawned in the center of the screen, and m carnivores are spawned a configurable distance away from the center. Vegetables will be spawned outside of the center. Carnivores will try to eat the herbivores and herbivores will try to eat vegetables. A round ends once all vegetables are eaten. If a creature eats enough food, they will produce offspring for the next round and pass on their traits. The traits of their offspring will be randomly mutated based on the parent's values. Traits include speed, size, and sight distance.

After the end of each simulation, a log is created. The github repository contains examples and functions to use to analyze a run.

A graph depicting how average velocity changes throughout the rounds.

• Date: November 2019
• Github: Link
• Tools: Python, Photoshop

Close

The goal for this project was to use a Twitter user’s tweets to recommend news articles that they would be interested in. Data was collected from the Twitter API, I separated different topics using hashtags. After processing the data I applied bag of words, SVM, multinomialNB, and logistic regression to classify the topic of a tweet. I then incorporated these models into apps that queried tweets from a chosen Twitter user and generated appropriate news topics.

I created two versions of a news recommendation web app. The RShiny app can be found here. I have also made an alternate app using Python and Dash that can be accessed by running this file.

GUI for the Dash based recommendation system.

• Date: April 2018
• Github: Link
• Tools: Python, Dash, R, RShiny

Close

This was an independent project based on data provided by Santander bank. The dataset consisted of demographic and transactional behavior of customers over a 17-month period. I used this data to predict which products customers would buy in a certain timeframe and also to create a measure for customer satisfaction.

The models used were association rule learning, logistic regression, random forests, gradient tree boosting, and neural networks. The main challenge of this project was coming up with ways to deal with heavily imbalanced classes (<1%). Class rebalancing via over-sampling or under-sampling proved to be an efficent way to improve results.

Key Takeaways

• Demographic data is important when gauging customer satisfaction but it is not useful when predicting which products a customer will buy.

• It is much easier to predict if a customer will drop an item than it is to predict when they will buy a new item.

• When predicting customer satisfaction and identifying which customers would leave the bank, I was able to create 4 new features that gave strong indications for at risk customers.

• Date: January 2018
• Github: Link
• Tools: R

Close

This analysis was done for an AI project competition hosted by the Columbia Statistics Club. My group of 3 received 2^nd place. We used the Speed Dating Data Set compiled by the Columbia Business School in 2002-2004. Each observation in the dataset represents a speed date that a single person went on. Demographic information, interests, and subjective qualities about each person are recorded. After data cleaning, we transformed the entire dataset so that each observation represented a date between two participants.

The models used were Random Forests, Linear SVMs, and Neural Networks. The main challenge in the modeling was that it was difficult to get a high true positive rate.

Key Takeaways

• For most participants, their perception of their date's attractiveness was the most important variable in determining if that person wanted to go on another date.

• Participants' perceptions about their own qualities like attractiveness and intelligence were only loosely correlated with what their date thought about those qualities.

• Date: October 2017
• Tools: R

Close