Overview of the Data Science workflow (asking the right question, collecting/cleaning/exploring data, modeling, and deployment/reporting). A minimal example using concrete compressive strength.

Supervised learning models are trained on a dataset that includes both inputs (features) and the correct outputs (labels or target values). We discuss what they are and how they are trained.

Train/test splits, cross-validation, and the frequent need for custom splits to give an "honest assessment" of model performance. Always think about what you want your model to generalize to!

Derivation of cross-entropy loss via negative log likelihood. Confusion matrix based classification metrics (accuracy, precision, recall, etc)

A very short video which gives you pointers on where to find additional content on data collection. We do not cover this content in depth in the synchronous lectures.

Scikit-learn models follow a simple model.fit(X,y), model.predict(X) API. This allows you to use many different models as black boxes so you can start using them right away.

We discuss different loss functions (MSE, MAE, Huber) and evaluation metrics (MAPE, R^2, etc) for regression problems.

Plots related to threshold tuning, namely PR-curve and ROC-curve. The area under the ROC curve (AUC-ROC) measures the probability that class 1s are ranked as more probable than class 0s.

Overview of the basics of data cleaning and exploratory data analysis. Showcases some fundamental tools (pandas, matplotlib, plotly) as well as some more specialized tools (panderas, missingno).

We often want to ingest raw data and apply a sequence of transformations (imputation, scaling, and feature engineering) to that data before modeling. Pipelines let us do that in a clean way.

Diagnostic plots for regression analysis, including residual vs. feature plots, residual vs. predicted value plots, and QQ-plots of residuals.

The expected generalization error of a learning algorithm can be decomposed into two terms: the bias and the variance.