Paper #2: Preference-based User Rate Correction Process for Interactive Recommendation Systems

The objective of the paper is to improve the performance of rating systems to account for user error. The importance of this analysis is due to recommeder systems being based on user ratings; by improving the user ratings the system would give more accurate recommendations.

User ratings are subject to two phenomenon: the missing value problem (user doesn't rate an item), and the noisy rating problem (a user makes a mistake giving a specific rating). Rating noise accounts for: user changing their opinion over time and the user can fail to express their personal preference. The paper focuses on the second problem, users making mistakes when rating.


In their approach they "have focused on a set of items on which the user has taken the action (i.e., rating), and designed an attribute selection scheme to represent user preferences." The example given in the document features a user who likes a movie director (attribute), however the user gave a high rating to most of that director's movies but gave a much lower rating to a particular movie made by the same director. The system assumes that the user has a preference over that attribute and compensates for the lower score in the aforementioned movie.

Section 4 features some of the different techniques which can be used together with the User-Item-Attribute-rating model the paper describes on section 3. The first method calculates a dominant attribute and item, and from there find discrepancies in the ratings. The second method determines a set of 'expert' users, to which the average users can look up to in terms of movie preferences. The third method is self based correction, and the last is a hybrid between the previous two.