In my prior blog post, I wrote of a clever elf that could predict the outcome of a mathematically fair process roughly ninety percent of the time. Actually, it is ninety-three percent of the time and why it is ninety-three percent instead of ninety percent is also important. The purpose of the prior blog post was to illustrate the weakness of using the minimum variance unbiased estimator (MVUE) in applied finance. Nonetheless, that begs a more general question of when and why it should be used, or a Bayesian or Likelihood-based method should be applied. Fortunately, the prior blog post provides a way of looking at the problem. Fisher's Likelihood-based, Pearson and Neyman's Frequency-based and Laplace's method of inverse probability really are at odds with one another. Indeed, much of the literature of the mid-twentieth century had a polemical ring to it.

There are three quite distinct ways to train a machine learning model on recommender system logs. The first method is to model the reward prediction for each possible recommendation to the user, at the scoring time the best recommendation is found by computing an argmax over the personalized recommendations. This method obeys principles such as the conditionality principle and the likelihood principle. A second method is useful when the model does not fit reality and underfits. In this case, we can use the fact that we know the distribution of historical recommendations (concentrated on previously identified good actions with some exploration) to adjust the errors in the fit to be evenly distributed over all actions. Finally, the inverse propensity score can be used to produce an estimate of the decision rules expected performance. The latter two methods violate the conditionality and likelihood principle but are shown to have good performance in certain settings. In this paper we review the literature around this fundamental, yet often overlooked choice and do some experiments using the RecoGym simulation environment.