Hierarchical Temporal-Contextual Recommenders

Recommendation systems have developed beyond simple matrix factorization to focus on two important sources of information: the temporal order of events (Hidasi et al., 2015) and side (e.g., spatial) information encoded in user and item features (Rendle, 2012). However, state-of-art temporal modeling is often limited by model capacity for long user histories. In addition, meta data are rarely used in generic sequence models, perhaps due to a lack of improvement guarantees in end-to-end training. Important kinds of meta-data, like interaction feedback (e.g. click vs. add to cart, view duration) are not modeled. In this paper we propose a hierarchical recurrent network with meta data (HRNN-meta) model to solve both problems. To compactly store long histories, and propagate gradients through them, we use HRNN to group user interactions into hierarchical sessions of activity intervals, within which a user tends to maintain related interests. Different from previous hierarchical models (Quadrana et al., 2017), which manipulate model hidden states, HRNN encodes session information in the embedded inputs. We show that this change not only yields up to 10x better computational efficiency due to better ability to align batches, but also allows us to extend from GRUs (Cho et al., 2014) to the entire family of RNN models, and further increases model capacities when combined with temporal convolutional networks. To use meta data in sequential models, we extend the HRNN decoder with a factorization machine inspired network, between the HRNN output embedding and item meta data, which improves over the vanilla HRNN which is a special case of the model. We also extend HRNN-meta model to handle user features and interaction feedback to learn different objectives such as click or rating predictions. We report significant improvements both in simulation studies and in real-world datasets.