Scalable and Personalized Item Recommendations Framework

Balancing scalability and relevance is important for industry-level recommender systems. A large scale e-commerce website may have millions of customers and millions of items. Typically, researchers and data scientists generate models which are anchored on customer level, or item level. Customer anchored recommendation models are typically surfaced on front pages of e-commerce websites. Similarly, item pages typically host various item anchored models. In each of the two usecases, developers frequently store models offline, i.e., models are stored for each customer, or are stored for each item. Scalability challenges arise when one wishes to personalize item anchored models. Offline based approaches, where both customer and item ids are stored as anchors become rapidly unscalable as number of customers and items increase. Another approach is to utilize an online approach on a pre-computed recall set (for example: item recommendations for a given anchor item), and then to apply users’ preferences. In this paper, we describe a scalable personalized item recommender system which follows the latter approach. We take historical user preferences (customer understanding) and existing item recommendation models to personalize item anchored model. We showcase several usecases to show how we apply online inferencing algorithms and scale it up to millions of customers.