Each dot = one labeled training row plotted on two of the 22 features. Patterns to look for:

• Clean separation: green clusters on one side, red on the other → these two features alone predict outcome well
• No separation: dots mixed → these features aren't useful together (model has to rely on other dimensions)
• Edge cluster: corner of the plot is all one color → strong rule (e.g., "high RVOL + low IV = always wins")