Revenue optimization and customer targeting in daily-deals sites

Daily-deals sites (DDSs), such as Groupon and LivingSocial, attract millions of customers looking for products and services at significantly reduced prices. The challenge of DDSs is to find the best match between deals and customers while generating as much revenue as possible. Hence, successful DDSs need to maximize revenue and increase customer satisfaction. These two problems are investigated in this thesis. The revenue of DDSs depend how much they charge merchants for each deal sold and on the number of coupons sold for each deal (aka. deal size). The percentage charged is a business decision, but the deal size can be predicted. This thesis introduces a new model for deal size prediction that considers both intra-market competition and market interplay among deals. The deal size prediction method offers gains in precision ranging from 8.18% to 17.67% in comparison with state-of-the-art methods. Maximizing revenue depends on deal sizes and on customers, which most of the time are reached through email marketing. If customers receive uninteresting, nonpersonalized emails on a daily-basis, they stop paying attention to emails and treat them as spam. When dealing with customer satisfaction, we tackled two problems: (i) how to reduce the volume of emails sent daily without reducing revenue and (ii) how to personalize emails. For reducing the percentage of emails sent, we propose criteria to sort customers according to their purchase probability and to apply a muti-armed bandit algorithm to choose which criterion is more appropriate to select the user that should receive the email. Experiments showed that reducing the number of emails sent by 40% does not affect the number clicks on deals advertised in the emails. For the task of email personalization, we model the problem so as to obtain user feedback on deals of the day as soon as possible. We start emailing deals recommended by current methods and wait for user feedback. Having feedback, the remaining users receive better recommendations by reranking current recommendation lists using feedback. Experiments show that our algorithms offer gains in precision ranging from 7.9% to 34.0% in comparison with state-of-the-art recommendation algorithms.