CŐDE

Train our PhDs with us through our Data Science Boot Camps and Certificate Programs

Designed to train PhDs with the skill sets necessary for successful employment in your organization.

FALL 2020 DATA SCIENCE CERTIFICATE PROGRAM

ONLINE: FALL 2020

We're happy to be adapting our May Data Science Boot Camp curriculum into a semester-long certificate program!

​

This Fall 50+ OSU PhDs will be diving deep into the theory and practice of essential Data Science methodologies including:

  • Web Scraping

  • Regression Algorithms

  • Classification Techniques

  • Time Series Analysis & Forecasting

  • Basic Unsupervised Learning

  • And More!

MAY 2020 DATA SCIENCE BOOT CAMPS

ONLINE: May 4-31, 2020

Overview

This past May, the Erdős Institute moved its Data Science Boot Camps online. We had over 140 PhD students and postdocs from 7 universities participate in the program.  We thank our corporate members and sponsors for making this program such a huge success! 

​

​

​

​

​

Educational Aim

The goal of our Python Boot Camp is to provide you with the skills needed to produce a portfolio worthy data science/machine learning project.

​

Topics
In fitting with our educational aim, this course will walk you through the steps involved in a typical data science/machine learning project. While covering a variety of important techniques and algorithms we will illuminate themes and motivations that can be adapted to many data science settings.


Below we outline the specific topics we will be covering in this boot camp.

​

 

  • Data Gathering Techniques

    • Searching common online sources for data

    • Basic web scraping with BeautifulSoup

    • Interacting with APIs
       

  • Data Cleaning

    • Data Types

    • Basic data exploration with pandas, and numpy

    • Basic plotting with matplotlib

    • Handling Missing Data

    • Common Data Transformations
       

  • Supervised Learning

    • Regression

      • Simple Linear Regression

      • Multiple Linear Regression

      • Polynomial Regression

      • Ridge Regression

      • LASSO

      • Kernel Regression (if time permits)

      • Local Regression (if time permits)

    • Classification

      • Nearest Neighbor Methods

      • Naive Bayes

      • Logistic Regression

      • Decision Trees

      • Random Forests

      • Support Vector Machines
         

  • Unsupervised Learning

    • Dimensionality Reduction

      • Principal Components Analysis

      • t-SNE

    •  Clustering

      • k-Means

      • Hierarchical

      • DBScan
         

  • Forecasting for Time Series Data

    • Handling and cleaning time series data

    • Simple forecasting methods

    • Time series regression models

    • Smoothing

    • Exponential Smoothing
       

  • Neural Networks

    • Perceptrons

    • Shallow Networks
       

  • Presenting Results

    • Pandas for presentation

    • Advanced matplotlib

    • Plotting in seaborn

    • Introduction to Interactive Plotting With Python
       

  • Machine Learning Concepts

    • Training Test Split

    • Loss Functions

    • Gradient Descent

    • Model Validation

    • Bias Variance Trade-Off

    • Cross Validation

SAMPLE PROJECTS

Some winning projects from previous years

bookend: an authorship attribution classifier

Kyle Dettman, Elaad Applebaum, Diptanil Roy, Nikhil Tilak

Goal: Given some small snippet from a book, can we devise a classifier that can predict the author with some accuracy?

Data Source:

  • E-books from open source Project Gutenburg by authors such as Jane Austen, Mark Twain, Jack London

  • A purpose-built class in Python to handle the reading in, cleaning, and general analysis of these e-books.

​

Methods:

  • A series of Natural Language Processing techniques to build features: n-grams, syntactic tagging, bag-of-words, and lexical features.

  • The predictions from each of these models was fed to a soft- voting classifier to obtain the results.
     

Packages: Scikit-learn, Nltk, Ngram-graphs

Results:

  • Classifier achieves >90% accuracy on texts it had never seen before.
     

  • Algorithm correctly classifies books written by J.K. Rowling under her pseudonym Robert Galbraith, demonstrating cross-genre success in identifying author-level characteristics.
     

  • Classifier correctly attributes the disputed Federalist papers to Madison.

Birds in Random Forests

Dananjaya Liyanage, Caleb Dilsavor, Hiran Wijesinghe

Results:​

​

The following is a brief summary of classifiers generated along with their accuracy. Cross validation was performed using random 1:5 testing to training splits of the data set.​

Summary:

​

Extracted ~23k feature vectors from ~8k recordings and tested various classifiers using labels provided in metadata.

​

  • Scraping and preprocessing data from xeno-canto.org using R

  • Fourier spectral analysis using TuneR (R library)

  • WarbleR (R library) for feature extraction

  • Python with Jupyter Notebook

  • Matplotlib (python library) for visualization

  • Scikit-Learn (python library) for machine learning

Two Tweet Too Furious

Matthew Osborne, Austin Antoniou, Dan McGregor, Luke Andrejek

Brief Description

In a set of tweets containing #Charlottesville from the week of the Charlottesville Unite the Right rally, can we gain any insight on who is actually tweeting these tweets?

 

Thanks to Center for the Study of Networks and Society for the Excellent Data!

 

Data and Analysis

We had tweets from over 1000 unique Twitter users to analyze. Each user in the data set tweeted #Charlottesville at least once in the week following Unite the Right rally. Other than that we knew nothing about the accounts.

 

We arranged these accounts into what we called a mutual retweet network (see Figure 1). Each node was one of the unique accounts from the data set and two accounts were connected with an edge if they retweeted the same accounts.

 

Our thought process was that people that are similar will retweet the same content.

 

Analysis was done using the Python packages: pandas, numpy, matplotlib, and networkx.

Figure 1

Outcome

Our approach resulted in clear communities, that describe the accounts quite well. We’ve highlighted those communities along with pictures of the top accounts that the encircled users retweeted (see Figure 2). A quick breakdown of the communities is provided below.

 

We expected to get two communities corresponding to republicans and democrats, however, we were surprised to see two outlier communities that are seemingly unrelated to the politics of the Charlottesville incident.

 

Perhaps this technique of Twitter data analysis could be a fruitful way to identify who is talking about or leveraging an event/issue.

​

Community Breakdown

Green – Possibly followers of a religious group from India

Yellow – Media accounts with a similar political skew

Blue – Democrat leaning accounts

Red – Republican leaning accounts

Figure 2

Identifying Leaves with Python

Alex Beckwith, Jason Bello, Sheng Guo, Kiwon Lee, Francisco Martínez Figueroa

PHOTO GALLERY

Photos courtesy of Stephen Takacs Photography.

Goal: Classify species of leaves from a database of leaf images using techniques from image processing and computer vision.

Successive approximations of leaf outlines using elliptic Fourier analysis.

Principal component analysis to identify leaf species.

Some eigen-“leaves”

Primary packages used:

  • cv2 (for computer vision)

  • sklearn (for statistical analysis)

​​

Examples of techniques used:

  • Principal component analysis

  • Elliptic Fourier analysis

  • ~FAST corner detection algorithm

  • Scale-invariant feature transform (SIFT)

  • Template matching via Hausdorff distance

  • Canny edge detection

  • Misc. statistics on leaves (aspect ratio, solidity, isoperimetric factor)

  • Distance-to-mean comparison

~FAST corner detection

Template matching via Hausdorff distance

©2020 by The Erdős Institute.