Recommender systems are the backbone of personalized services that
provide tailored experiences to individual users. Still, data sparsity
remains a common challenge, especially for new applications where
training data is limited or unavailable. We'll present a combinatorial
optimization problem that formalizes the selection of item universe for
experimentation in recommender systems. On one hand, a large set of
items is desirable to increase diversity. On the other hand, a smaller
set enables rapid experimentation and minimizes the time and the amount
of data required to train machine learning models. We show how to
optimize for such conflicting criteria using a multi-level optimization
framework. Our approach integrates techniques from discrete
optimization, unsupervised clustering, and latent text embeddings.
Experimental results on well-known recommendation benchmarks demonstrate
the benefits of optimized item selection.