Machine Learning For Dummies

Chapter 18

Resorting to Ensembles of Learners

IN THIS CHAPTER

Discovering why many guesses are better than one

Making uncorrelated trees work well together in Random Forests

Learning to map complex target functions piece by piece using boosting

Getting better predictions by averaging models

After discovering so many complex and powerful algorithms, you might be surprised to discover that a summation of simpler machine learning algorithms can often outperform the most sophisticated solutions. Such is the power of ensembles, groups of models made to work together to produce better predictions. The amazing thing about ensembles is that they are made up of groups of singularly nonperforming algorithms.

Ensembles don’t work much differently from the collective intelligence of crowds, through which a set of wrong answers, if averaged, provides the right answer. Sir Francis Galton, the English Victorian age statistician known for having formulated the idea of correlation, narrated the anecdote of a crowd in a county fair that could guess correctly the weight of an ox after all the people’s previous answers were averaged. You can find similar examples everywhere and easily recreate the experiment by asking friends to guess the number of sweets in a jar and averaging their answers. The more friends who participate in the game, the more precise the averaged answer.

Luck isn’t what’s behind the result — it’s simply the law of large numbers in action (see more at https://en.wikipedia.org/wiki/Law_of_large_numbers ...