Unified and Scalable Incremental Recommenders with Consumed Item Packs

Recommenders personalize the web content using collaborative filtering to relate users (or items). This work proposes to unify user-based, item-based and neural word embeddings types of recommenders under a single abstraction for their input, we name Consumed Item Packs (CIPs). In addition to genericity, we show this abstraction to be compatible with incremental processing, which is at the core of low latency recommendation to users. We propose three such algorithms using CIPs, analyze them, and describe their implementation and scalability for the Spark platform. We demonstrate that all three provide a recommendation quality that is competitive with three algorithms from the state-of-the-art.