Causal Combinatorial Factorization Machines for Set-Wise Recommendation

With set-wise (exact-k, slate, combinatorial) recommendation, we aim to optimize the whole set of items to recommend while taking the dependency among items into consideration. This enables us to model, for example, the substitution relationship of items, i.e., a customer tends to purchase only one item in the same category, in contrast to the top-k recommendation in which the independency of items is assumed. Recent efforts in this context have focused on the computational aspects of optimizing the set of items to recommend. However, they have not taken into account sample selection bias in datasets. Real-world datasets for recommendation have missing entries not completely at random due to biased exposure or user preferences. Addressing the selection bias is important for the set-wise recommendation since methods with larger hypothesis spaces are more likely to overfit biased training data. In light of recent top-k recommendation research that has addressed this issue by using causal inference techniques, we therefore propose a set-wise recommendation model with debiased training methods based on recent causal inference techniques. We demonstrate the advantage of our method using real-world recommendation datasets consisting of biased training sets and randomized test sets.