Manja Gersholm Grønberg:
How do we provide the user with the best recommendations?
A new approach to analysis of data from recommender systems.
How many stars did you give your clothing shop? How did you score Shawshank Redemption? and how many chef’s hats did you give to the last restaurant you visited? These are all examples of questions asked by online user rating systems, which are often used to provide the user with relevant recommendations for which movie to watch or which product to buy. The data generated by these systems are generally identified as ordered categorical data, also known as ordinal data and comes in massive amounts. With the increasing use of rating systems, the need for methods to analyze big ordinal data has never been more topical.
Current methods to analyze data from rating systems do not model the scoring scale, nor the use of it, appropriately. As an example, these methods assume that it is equally hard to move a user from one star to two stars as it is to move a user from two stars to three stars (equal distance between groups). Furthermore, the methods do not consider user’s different use of the scale. In classical statistics ordinal data can be modelled by ordinal mixed models. These models make proper inference of the rating scale and allow for complex dependency structures, hopefully making it possible to model user’s different use of the scale. The problem is, however, that these models are not yet available for big data.
The aims of this PhD project are to investigate how ordinal mixed models can be combined with big data in order to make the models applicable to online user ratings. By making ordinal mixed models applicable to online user ratings we might be able to take the scoring scale, and the use of it, into account and thereby hopefully improve the recommendations for the user.