RealML

Before you do anything
- What question are you trying to answer?
- How can you measure the answer?
- What data can you get?
- How well do you think you can do?
- Is the number of observations and features reasonable?
Overfitting is your worst enemy
- Train, Validate (Quiz), Test
- Out-of-sample in the real world is subtle
  - new people, products, words, time periods, countries, ...
Loss functions
- What is the end-use loss function?
- what is a good ML loss function?
- L2 vs. L1 vs. L0 vs. Hinge vs. Exponential vs. cost
- Classification problems often have asymmetric costs
- precision/recall, sensitivity/specificity, ROC
Feature generation is critical
- Think about the problem!!
- How might you transform the features?
  - Do you want a scale-invariant method or not?
- What else could you measure?
- Is semi-supervised learning possible?
- Are there surrogate labels you might use?
  - 'distant supervision'
Feature Blocks
- Different feature sets need different regularization
- One solution: block-stagewise regression
Ensemble methods
- Combinations of multiple methods are almost always the most accurate
- Averaging methods (or experts)
  - equal weighting
    - {$ \hat{y} = (1/k) \sum_k \hat{y}_k$}
  - inverse variance-based weighting
    - {$ \hat{y} = (\sum_k \hat{y}_k / \sigma_k^2 ) /( \sum_k 1/ \sigma_k^2)$}
  - regression-based weighting
    - {$ \hat{y} = \sum_k w_k \hat{y}_k$}
- Boosting
- Random Forests
Missing data
- Missing at random (MAR) or not requires different handling
  - Imputation works well for MAR, but most data are not MAR, so it is best to add a new variable that indicates whether or not the feature is missing.
Explanation/Insight is often important
- Look at the data!
  - posts, images scoring highest in some feature or outcome
  - error analysis
- variable importance
  - How "important" is each feature for the prediction?
- visualization: word clouds, PCA, MDS
  - MDS: given an {$n x n$} matrix of distances between points, find a new (usually 2-D) representation of each of the points that as closely as possible preserves that distance matrix
Feature representation
- reduce dimension using PCA, Autoencoding, or clustering like LDA
- Think about how to best handle categorical variables
Correlation is not causality ?

Back to Lectures