Jointly modeling products and resource pages for task-oriented recommendation

Modeling high-level user intent in recommender systems can improve performance, although it is often difcult to obtain a ground truth measure of this intent. In this paper, we investigate a novel way to obtain such an intent signal by leveraging resource pages associated with a particular task. We jointly model product interactions and resource page interactions to create a system which can recommend both products and resource pages to users. Our experiments consider the domain of home improvement product recommendation, where resource pages are DIY (do-it-yourself) project pages from Lowes.com. Each DIY page provides a list of tools, materials, and step-by-step instructions to complete a DIY project, such as building a deck, installing cabinets, and fxing a leaking pipe. We use this data as an indicator of the intended project, which is a natural high-level intent signal for home improvement shoppers. We then extend a state-of-the-art system to incorporate this new intent data, and show a signifcant improvement in the ability of the system to recommend products. We further demonstrate that our system can be used to successfully recommend DIY project pages to users. We have taken initial steps towards deploying our method for project recommendation in production on the Lowe's website and for recommendations through marketing emails.