I just finished reading "Bounding Rationality to the World" by Todd and Gigerenzer, published in the Journal of Economic Psychology in 2003. It’s an interesting article with the core theme that the reason we see cognitive limitations in human decision making is that people may be optimized to make decisions in a particular environment. In particular, this argument suggests that the apparent deficits in human decision making discovered in various economic experiments may be due to those experiments creating environments that do not correspond well to the sorts of environments humans regularly have to make decisions about.
Several of the issues seem to have interesting implications for recommenders, including:
- people are often making decisions in an environment in which information is not free. collecting more information to make a better decision might lead to an overall decision that is worse, ironically. (This reminds me of Colin Powell’s talk in which he explained that he was trained to make decisions with 2/3 of the information; waiting for more information would mean that the decision would not be available in time to help during a battle.) Can a recommender be set to help a person satisfice appropriately?
- how do known stable cognitive illusions like the overconfidence bias and the hard-easy effect relate to recommenders? Can and should recommenders be tuned to avoid these illusions?
- knowing more sometimes decreases decision performance. Can this sort of problem be seen in recommenders?
- one common decision approach is sequential choice, in which a decider must make a decision without having the opportunity to choose a prior choice once it has been passed on. What would a recommender be like for a sequential choice problem? Could it improve performance?
- the paper mentions research that suggests that humans benefit during language acquisition from having limited cognitive ability, so higher abilities don’t get in the way while learning the basics. This feels related to concepts in machine learning, like those that limit the complexity of the function search space. However, humans over time increase the complexity of their cognitive function. What does this correspond to in machine learning?
Overall, recommender research that explores the ways recommenders interact with human decision processes would be fascinating.