Consumers face a large number of choices while shopping online. Studies have shown, that they are already expecting to be targeted with content addressing their personal needs. In a web shop, products are presented as lists based on a selected category or as results of a product search. To support the users in their decision making, they can be provided with a personalized product ranking fitted to their current interests. In this piece of work, three levels of personalized product rankings are proposed: explicit personalization, cluster-based personalization and individualization. To estimate the potential effect of the personalization and its required effort, two prototypes for the second and third level are developed and evaluated. The prototypes are based on a previously existing non-personalized ranking, which ranks the products in descending order according to a sales prediction. The cluster-based prototype enhances this product ranking by determining customer clusters beforehand using both situative and behavioural data. The individualized product rankings rely on the combination of the ranking with a recommendation system realized as a matrix factorization. In doing so, the concept of learning to rank is considered. By evaluating the cluster-based and individualized prototype on a sampled data set in comparison to the non-personalized ranking, it is shown that the created personalized rankings are in fact closer to the users’ needs. Furthermore, a subjective evaluation confirms that the cluster-based rankings can reflect the users’ interests in a better way.
|Qualification||Master of Science|
|Award date||22 Apr 2019|
|Publication status||Published - Apr 2019|
- Machine Learning
- Recommendation system
- Matrix Factorization