Bridging Recommendation and Marketing via Recurrent Intensity Modeling

Personalized item-recommendation (ItemRec) and targeted item-marketing are two sides of the same coin. The key in the latter problem, which we call UserRec, is to recommend users in their temporal contexts on behalf of the item providers. As ItemRec progresses to using end-to-end sequence models which directly ingest the item-consumption history of each user to build real-time hidden states, we study whether we can turn these models around to recommend users in their real-time contexts. Our main challenge is to compare these item-consumption sequences of variable lengths and duration. To overcome it, we discover a novel connection to Temporal Point Process (TPP), where we use its intensity parameter to predict the user's likelihood of future engagements, based on but not biased by the past observations. We develop a Recurrent Intensity Modeling (RIM) trick to convert sequential ItemRec models to UserRec based on Marked-TPP decomposition. RIM allows us to rethink recommendation in a Matching (Mtch) scenario, where the benefits of the users (e.g., ItemRec relevance) and item providers (e.g., item-exposure guarantees) are considered at the same time. Notably, as RIM is based on sequence models, we extend to Online Matching (OnlnMtch) to serve ItemRec in real-time while satisfying the item guarantees by approximate ratios in the long-term, borrowing techniques from online dual decomposition in convex optimization. Finally, to complement our discussions on RIM and Mtch, we provide empirical simulations on real-world datasets with open-source codes.