Using a constructed score table over four features (A, B, C, D), the animation traces how forward selection and backward elimination can land on different two-feature subsets. Forward selection greedily builds up from A to AB (.73), while backward elimination trims down from the full set to BC (.85), the true best pair confirmed by exhaustive search. The video clarifies this is a single example, not a general proof, and previews floating search and branch-and-bound as alternatives. Useful for first-year AI and machine learning students studying feature selection strategies.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Greedy feature searches can follow different paths Four features A, B, C, and D form a lattice of subsets. Use a clearly labeled constructed score table: single-feature scores A=.70, B=.68, C=.60, D=.55; pair scores AB=.73, AC=.72, AD=.70, BC=.85, BD=.69, CD=.62; triple scores ABC=.84, ABD=.72, ACD=.71, BCD=.86; and ABCD=.83. The goal is to choose two features. Forward selection chooses A first and then AB, ending at .73. Backward selection removes A from the full set to reach BCD, then removes D to reach BC, ending at .85. Exhaustive comparison of all six pairs confirms BC as the best pair in this table. Explain that this example does not prove backward selection always wins. Briefly show how floating search can revisit additions and removals, while branch-and-bound requires valid bounds. Use black backgrounds and distinct colored search paths. Audience: 一年级人工智能专业学生