Improving daily deals recommendation using explore-then-exploit strategies

摘要

Daily-Deals Sites (DDSs) enable local businesses, such as restaurants and stores, to promote their products and services and to increase their sales by offering customers significantly reduced prices. If a customer finds a relevant deal in the catalog of electronic coupons, she can purchase it and the DDS receives a commission. Thus, offering relevant deals to customers maximizes the profitability of the DDS. An immediate strategy, therefore, would be to apply existing recommendation algorithms to suggest deals that are potentially relevant to specific customers, enabling more appealing, effective and personalized catalogs. However, this strategy may be innocuous because (1) most of the customers are sporadic bargain hunters, and thus past preference data is extremely sparse, (2) deals have a short living period, and thus data is extremely volatile, and (3) customers' taste and interest may undergo temporal drifts. In order to address such a particularly challenging scenario, we propose a new algorithm for daily deals recommendation based on an explore-then-exploit strategy. Basically, we choose a fraction of the customers to gather feedback on the current catalog in the exploration phase, and the remaining customers to receive improved recommendations based on the previously gathered feedback in a posterior exploitation phase. During exploration, a co-purchase network structure is updated with customer feedback (i.e., the purchases of the day), and during exploitation the updated network is used to enrich the recommendation algorithm. An advantage of our approach is that it is agnostic to the underlying recommender algorithm. Using real data obtained from a large DDS in Brazil, we show that the way in which we split customers into exploration and exploitation impacts by large the effectiveness of the recommendations. We evaluate different splitting strategies based on network centrality metrics and show that our approach offers gains in mean average precision and mean reciprocal rank ranging from 14 to 34 % when applied on top of state-of-the-art recommendation algorithms.