Management Information Systems Quarterly
Abstract
E-commerce platforms often use collaborative filtering (CF) algorithms to recommend products to consumers. What recommendations consumers receive and how they respond to the recommendations largely depend on the design of CF algorithms. However, the extant empirical research on recommender systems has primarily focused on how the presence of recommendations affects product demand, without considering the underlying algorithm design. Leveraging a field experiment on a major e-commerce platform, we examine the differential impact of two widely used CF designs: view-also-view (VAV) and purchase-also-purchase (PAP). We found several striking differences between the impact of these two designs on individual products. First, VAV is about seven times more effective in generating additional product views than PAP but only about twice as effective in generating sales due to a lower conversion rate. Second, VAV is more effective in increasing views for more expensive products, whereas PAP is more effective in increasing the sales of cheaper products. Third, VAV is less effective in increasing the views but more effective in increasing the sales of products with higher purchase incidence rates (PIRs). Finally, when aggregated over all products with the same levels of price or PIRs, VAV dominates PAP in generating views and the difference is more striking for products with higher prices or lower PIRs. Interestingly, PAP is more effective than VAV in increasing the sales of products with low prices or moderate PIRs, though VAV generates more sales than PAP overall. Our findings suggest that platforms may benefit from employing different CF designs for different types of products.