We will meet four days per week for three weeks. In class, in lieu of lecture, we will collectively work through carefully designed Jupyter notebooks to develop both your skill in "coding" but especially to deepen your perspective on "data science." After our time together each day, you will work through some homework (ideally in small groups) to check your understanding.
At the end of the course, you have the opportunity for a final project to be built off the various topics introduced in class. Keep your eyes open for interesting data sets that could form the basis for your final project.
This course relies on Python. I assume that you don't know Python, but I know that you can learn very quickly, so you'll meet Python on Day 1 by diving head first into some Jupyter notebooks. Rest assured that a great many people learn Python this way and don't drown.
Understanding data motivates our learning Python. And when presented
with data, the very first thing to do is to look at your data. So
we also meet matplotlib on Day 1 to put some tools in our
visualization toolkit. Those visualizations encourage us to dig into
various techniques to summarize (like linear regression in one
variable) the distributions we are seeing and to "bootstrap" our way
to some confidence intervals. Along the way, we'll do some "data
wrangling" by merging data frames, grouping rows, etc.
- Day 1: Python, matplotlib
- Day 2: models, simulation
- Day 3: linear regression in one variable
- Day 4: estimates, bootstrapping
By the end of Week 1 you will know how to get into a Jupyter notebook, wrangle and look at your data, estimate some feature of your data, and to quantify how good your estimate is. In short, Week 1 prepares you to clean and prepare your data.
Week 2 forms the meaty middle of our course; more than estimates, in
Week 2 we aim to make predictions. One week isn't much time for this!
But our tools, like scikit-learn, make it possible to play in a
playground full of powerful algorithms before fully understanding how
those algorithms work. To address the pedagogical pitfall of too many
black boxes, we spend some time in Week 2 "boosting" our vocabulary
with words like training and testing sets, cross-validation,
underfitting and overfitting, the bias/variance tradeoff.
- Day 5: supervised learning, training and testing
- Day 6: cost functions, gradient descent
- Day 7: logistic regression
- Day 8: overfitting and regularization
By the end of Week 2 you will do some machine learning, e.g., take some labeled training data and predict the labels on the testing set. In other words, Week 2 prepares you to train and test a model.
In this final week, we encourage you to seek out data sets that speak to you personally. As always, the first step is to look at your data but your chosen data might be high-dimensional or otherwise hard to explore. So this week we consider clustering algorithms (k-means, and hierarchical clustering) and we use some linear algebra (e.g., singular value decomposition) to reduce the dimension of our data.
- Day 9: high-dimensional data, correlation
- Day 10: k-means
- Day 11: hierarchical clustering
- Day 12: linear algebra
By the end of Week 3 you will be prepared to find a (potentially high-dimensional!) data set you are interested in, find some insights in your data, and present evidence supporting your hypothesis.