com.planetj.taste.eval
Interface RecommenderEvaluator

All Known Implementing Classes:

AverageAbsoluteDifferenceRecommenderEvaluator, RMSRecommenderEvaluator

public interface RecommenderEvaluator

Implementations of this interface evaluate the quality of a Recommender's recommendations.

Author:

Sean Owen

Method Summary

double

evaluate(RecommenderBuilder recommenderBuilder, DataModel dataModel, double trainingPercentage, double evaluationPercentage) Evaluates the quality of a Recommender's recommendations.

Method Detail

evaluate

double evaluate(@Nullable
                RecommenderBuilder recommenderBuilder,
                @Nullable
                DataModel dataModel,
                double trainingPercentage,
                double evaluationPercentage)
                throws TasteException

Evaluates the quality of a Recommender's recommendations. The range of values that may be returned depends on the implementation, but lower values must mean better recommendations, with 0 being the lowest / best possible evaluation, meaning a perfect match. This method does not accept a Recommender directly, but rather a RecommenderBuilder which can build the Recommender to test on top of a given DataModel.

Implementations will take a certain percentage of the preferences supplied by the given DataModel as "training data". This is typically most of the data, like 90%. This data is used to produce recommendations, and the rest of the data is compared against estimated preference values to see how much the Recommender's predicted preferences match the user's real preferences. Specifically, for each user, this percentage of the user's ratings are used to produce recommendatinos, and for each user, the remaining preferences are compared against the user's real preferences.

For large datasets, it may be desirable to only evaluate based on a small percentage of the data. evaluationPercentage controls how many of the DataModel's users are used in evaluation.

To be clear, trainingPercentage and evaluationPercentage are not related. They do not need to add up to 1.0, for example.

Parameters:

recommenderBuilder - object that can build a Recommender to test

dataModel - dataset to test on

trainingPercentage - percentage of each user's preferences to use to produce recommendations; the rest are compared to estimated preference values to evaluate Recommender performance

evaluationPercentage - percentage of users to use in evaluation

Returns:

a "score" representing how well the Recommender's estimated preferences match real values; lower scores mean a better match and 0 is a perfect match

Throws:

TasteException - if an error occurs while accessing the DataModel