A multi-criteria recommender system incorporating intensity of preferences

Many websites provide visitors with the possibility to evaluate each item on more than one criteria. A commonly used rating scale is the one to five-star rating system or similar linguistic scales. Such scales are ordinal but the symbolic or lexical semantics convey information about the strength of user references in addition to the order of rated items. We refer to such scales as discrete ordered scales. We present AHP-Rec a method that treats user ratings as interval scale data and uses a multi-criteria approach for deriving predictions for user ratings. We use the data provided by Yahoo! Movies to demonstrate and evaluate the AHP-Rec recommender method. AHP-Rec takes as input the ratings each user gives to movies, calculates weights for each scale item that are personal for each user and provides its recommendation by aggregating preferences of similar users. Our method provides improved results over the state of the art single criterion method SVD++ and the multi-criteria method UTARec.