Automated Machine Learning for Interaction Functions in Collaborative Filtering

The work introduces a novel approach that treats the design of interaction functions (IFC) in collaborative filtering as an Automated Machine Learning (AutoML) problem, marking the first effort to apply AutoML for feature engineering of interaction functions.

Key contributions include formalizing the IFC design as an AutoML task, constructing a structured search space that balances expressiveness and computational cost, and proposing a one‑shot search algorithm that enables efficient stochastic gradient descent and point‑to‑point AutoML searches.

Two variants of the proposed Searching Interaction Functions (SIF) algorithm—SIF and SIF(no‑auto)—are evaluated against classic collaborative filtering methods (e.g., CP, Tucker, HOFM, Deep&Wide, NCF) and other AutoML baselines (Random, Reinforce, Bayes). Experimental results on MovieLens‑100K, MovieLens‑1M, and YouTube datasets show that SIF consistently achieves lower RMSE and faster convergence.

The paper also provides background on collaborative filtering, low‑rank matrix factorization, and the evolution of interaction functions from simple inner products to complex neural architectures such as ConvMF and ConvNCF. It reviews the history of AutoML, including rule‑based, statistical, deep learning, and AutoML‑driven approaches, and discusses Neural Architecture Search (NAS) as a core component of the proposed method.

Detailed ablation studies examine the impact of search space size, the number of operations, element‑wise transformations, and the use of linear versus non‑linear predictors (MLP), confirming that richer operation sets and appropriately sized search spaces improve performance.

Overall, the study demonstrates that integrating AutoML and NAS into recommendation system pipelines can automatically discover high‑quality interaction functions that outperform manually designed counterparts, offering a promising direction for future research in AI‑driven recommendation technologies.