Probability matching is a decision strategy where predictions align with the observed frequencies of class memberships. For example, if positive events occur 60% of the time and negative events 40%, a person following this strategy would predict 'positive' outcomes 60% of the time and 'negative' outcomes 40% of the time. In contrast, the optimal Bayesian strategy would always predict the majority, yielding a higher accuracy.
In a study about predicting whether a coin toss will land heads or tails, if a participant is shown that heads comes up 60% of the time in previous tosses, they may predict heads 60% of the time when actually the best strategy would be to consistently predict heads to maximize accuracy.
To overcome probability matching, one can consistently apply the Bayesian decision strategy by always choosing the majority class based on prior probabilities, regardless of frequency observation.