Increasing Generalizability: Naïve Bayes Vs K-Nearest Neighbors

Abstract Marketing research is often criticized for lacking generalizability and inability to reproduce results. The problem lies in using models to fit data, rather than determining the predictive power of models in conditions of uncertainty. For instance, how does the predictive power of a model change when customer dynamics change? The current study suggests that marketing researchers can supplement existing research methods with non-probabilistic prediction methods, such as the kNN algorithm-based model. Unlike probabilistic models that rely on past outcomes to predict future events – and lose predictive power when newer events are observed - non-probabilistic models better capture uncertainty. In the current study, the predictive power of the kNN algorithm-based model and the Naïve Bayes model is compared using data from two real markets. The kNN algorithm-based model provides more accurate predictions, showing the utility of combining the kNN algorithm-based model with existing marketing research to improve the predictability and generalizability of models. Implications for research and future research are discussed.