| Summary: | Recommender systems exploit user feedback over items they have experienced for
making recommendations of other items that are most likely to appeal to them.
However, users and items are but two of the three types of entities participating in
this ecosystem of recommender systems. The third type of entities are the manufacturers
of the products, and users are really their customers. Traditional recommender
systems research ignores the role of this third entity type and exclusively
focuses on the other two. What might item producers bring to recommender systems
research? Their objectives are related to their business and are captured by
questions such as “what kind of (new) products should I manufacture that will
maximize their popularity?” These questions are not asked in a vacuum: manufacturers
have constraints, e.g., a budget. The idea is that the user feedback data (e.g.,
ratings) capture users’ preferences. The question is whether we can learn enough
intelligence from it, so as to recommend new products to manufacturers that will
help meet their business objectives.
We propose the novel problem of new product recommendation for manufacturers.
We collect real data by crawling popular e-commerce websites, and model
cost and popularity as a function of product attributes and their values. We incorporate
cost constraints into our problem formulation: the cost of the new products
should fall within the desired range while maximizing the popularity. We show
that the above problem is NP-hard and develop a pseudo-polynomial time algorithm
for the recommendations generation. Finally, we conduct a comprehensive
experimental analysis where we compare our algorithm with several natural heuristics
on three real data sets and perform scalability experiments on a synthetic data
set. === Science, Faculty of === Computer Science, Department of === Graduate
|
|---|
