Özet:
Recommendation is one of the the most critical functionalities of an online retailer. Almost every platform wants to boost their sales by providing more relevant and personalized products for their consumer. The customer may not be able to find the product that is most suitable for her, as the product with the desired features is displayed in the lower rank positions on category page. It is a critical improvement to show the products at higher ranks that customers can purchase. It is important here at which rank the product is shown and how it meets the expectations for instance in terms of price, brand, color, size etc. With this motivation; we define a formal framework of strategies for design ing learning to rank algorithms for products that are sold online. First, we propose popularity-based ranking algorithm as the simplest and naive approach. In order to rank alternatives on category pages linear logistic and multinomial logit regression are proposed as model based linear approaches. However, according to computational problems beacuse of high product diversity and various kind of consumer profiles multi nomial logit regression may not perform well on online retail data. So, to handle these obstacles linear pairwise method is applied as model based linear ranking method. Real online retail data sets usually involve non linearity, Boosting is considered as model based nonlinear approach that uses a tree learner. It is determined that models con structed with boosting and popularity-based methods have better ranking accuracy than other methods in synthetic data set. Boosting method gives the highest ranking performance for the real online retail data set.
