Recommending ephemeral items at web scale

We describe an innovative and scalable recommendation system successfully deployed at eBay. To build recommenders for long-tail marketplaces requires projection of volatile items into a persistent space of latent products. We first present a generative clustering model for collections of unstructured, heterogeneous, and ephemeral item data, under the assumption that items are generated from latent products. An item is represented as a vector of independently and distinctly distributed variables, while a latent product is characterized as a vector of probability distributions, respectively. The probability distributions are chosen as natural stochastic models for different types of data. The learning objective is to maximize the total intra-cluster coherence measured by the sum of log likelihoods of items under such a generative process. In the space of latent products, robust recommendations can then be derived using naive Bayes for ranking, from historical transactional data. Item-based recommendations are achieved by inferring latent products from unseen items. In particular, we develop a probabilistic scoring function of recommended items, which takes into account item-product membership, product purchase probability, and the important auction-end-time factor. With the holistic probabilistic measure of a prospective item purchase, one can further maximize the expected revenue and the more subjective user satisfaction as well. We evaluated the latent product clustering and recommendation ranking models using real-world e-commerce data from eBay, in both forms of offline simulation and online A/B testing. In the recent production launch, our system yielded 3-5 folds improvement over the existing production system in click-through, purchase-through and gross merchandising value; thus now driving 100% related recommendation traffic with billions of items at eBay. We believe that this work provides a practical yet principled framework for recommendation in the domains with affluent user self-input data.