DescriptionUnderstanding the economic nature of consumer decisions in e-Commerce is important to personalized recommendation systems. Established economic theories claim that informed consumers always attempt to maximize their utility by choosing the items of the largest marginal utility per dollar (MUD) within their budget. For example, gaining 5 dollars of extra benefit by spending 10 dollars makes a consumer much more satisfied than having the same amount of extra benefit by spending 20 dollars, although the second product may have a higher absolute utility value. Meanwhile, making purchases online may be risky decisions that could cause dissatisfaction. For example, people may give low ratings towards purchased items that they thought they would like when placing the order. Therefore, the design of recommender systems should also take users' risk attitudes into consideration to better learn consumer behaviors.
Motivated by the first consideration, in this paper, we propose a learning algorithm to maximize marginal utility per dollar for recommendation. With the second, economic theory shows that rational people can be arbitrarily close to risk neutral when stakes are arbitrarily small, and this is generally applicable to consumer online purchase behaviors because most people spend a small portion of their total wealth for a single purchase. To integrate this theory with machine learning, we propose to augment MUD optimization with approximate risk-neural constraint to generate personalized recommendations. Experiments on real-world e-Commerce datasets show that our approach is able to achieve better performance than many classical recommendation methods, in terms of both traditional recommendation measures such as precision and recall, as well as economic measures such as MUD.