Machine learning attribution: Inferring item-level impact from slate recommendation in e-commerce
Slate-level recommendation is widely adopted in online services including e-commerce, video streaming and news services. Customers may observe a set of recommended items and interact with the content accordingly. Due to the combinatorial characteristics of slate level recommendation, various current ranking models are still aiming to optimize item level scores, instead of the slate level score. One key challenge of training such models is to attribute slate level reward correctly to each item. In this paper, we propose two learning based approaches named StepNet and SlateNet to estimate item level attributed reward. In StepNet we train models from different sets of data and use the difference of the estimated scores from two models as the attributed rewards. In SlateNet, we train a model on seen portions of slate-level recommendations and use this model to infer the impact of removing seen content to determine attribution. Simulation results show that both StepNet and SlateNet achieve normalized mean absolute error lower than 0.1 for attribution results with up to 8 items per slate. Validation on real dataset also confirms that SlateNet models are able to learn the intrinsic features of content and their relationship to different types of rewards.