Skip to main content

Learners : Matchbox recommender : Learner API : Evaluation

In order to evaluate finding related users, the predicted set of related users needs to be restricted to such users who have rated some items in common with the query user. Here we cover how to make such predictions and how to evaluate them.

Related users for evaluation can be found using the FindRelatedUsersWhoRatedSameItems method of the evaluator:

var evaluator = new  RecommenderEvaluator<Dataset, User, Item, int, int, Discrete>(  
var relatedUsersForEvaluation = evaluator.FindRelatedUsersWhoRatedSameItems(  

In this example trainedRecommender is a trained recommender and testDataset is the instance source of the test set. Related users will be found for each unique user in the dataset.

The item in each instance will also be queried by the data mapping. This allows a list of rated items to be constructed for this user which can then be examined to see which other users have also rated these items. This gives a list of “potentially related” users, and then predictions are made from this list. Ratings will not be queried by the data mapping at this stage. They will only be needed during evaluation. The parameter maxRelatedUserCount specifies the number of related users to find for each user. But if the number of possible related users for a given user is less than the value of minRelatedUserPoolSize, then the user is skipped. The last parameter allows for easy removal of users from the predictions for whom there is not sufficient information for later evaluation. Another parameter which is used to control this is minCommonRatingCount - it is guaranteed that all related users have rated at least that many items in common with the query user.

Once the restricted related user predictions are produced, they can be evaluated using the RelatedUsersMetric method of the evaluator:

var l1SimNdcg = evaluator.RelatedUsersMetric( testDataset,  
    minCommonRatingCount, Metrics.Ndcg, Metrics.NormalizedManhattanSimilarity));

The ranking metric used to evaluate related users is one of the following:

The fifth argument to the RelatedUsersMetric method is the rating similarity function, the value of which used as gain in the computation of the metrics above. It takes in two vectors and returns a real number. Pre-defined similarity functions include:

The way evaluation works is the following. First, the sets of rated items for each user are extracted from the input data. These sets are then reduced to the items that both users rated in common. Then, for each item in these sets the rating given by the corresponding user is taken. This forms two user rating vectors, which are used as inputs to the functions listed above.