Weighted Recommender System

This recommender system will compute a score based on the weighted accuracy, run time, and interpretability of each method.

The recommender will identify which method has a score that is significantly better than that of another method. For example, methodA and methodB receive weighted scores that are not significantly different from each other. methodA and methodB are both significantly better than methodC. The ranking for this instance would look like this:
1) methodA and methodB
3) methodC

The weights will be parameterized and will add to 1. For example, scores will be updated like so:
new_scores = weightA * _accuracy_ + weightB * _runtime_ + weightC * _interpret_
and
weightA + weightB + weightC = 1

The _accuracy_ of variable will be the traditional accuracy of each method. The _runtime_ variable is also straightforward and is the time it takes each method. The _interpret_ variable represents the interpretability of each method. For example, there will be an objective ranking of each of the methods by their complexity for the user to understand. There could be a ranking system that tiers methods and gives them a score of 1, 2, or 3. Methods like logistic regression and decision trees are easier to understand and could be assigned weights of 3 whereas methods that are more complex, such as random forest, may be assigned a 1.

As the framework outlines this would all be computed in the update method. A separate method, _assign_score would be used find the value of interpret variable for each run.

This recommender system will identify which scores are significantly better than other scores. This would be calculated in the recommend method which ranks the methods based on scores. Two or more methods with scores that are not significantly different would be ranked equally.

Finally, a grid search will determine the optimal weights (values of weightA, weightB, and weightC) for the recommender.