Performance Ranking of Recommender Systems on Simulated Data
Elizaveta Stavinova, Andrey Gurov, Anton Lysenko, Petr Chunaev · Procedia Computer Science · 2022
Recent studies have shown that modelling and simulation of interactions between a recommender system (RS) and its users have a great potential for accelerating the research and industrial deployment of RSs. Frameworks providing such simulations are called simulators and are widely used for RSs of different types. Nevertheless, there exist the problem of simulation validation and of the inconsistency of RS performance ranking on real-world and the corresponding synthetic data. In this paper, using and extending the recently developed SynEvaRec simulator we propose a validation procedure for simulators of this type and study the consistency of RS performance ranking on response matrices of different sparsity. It is observed in our experiments that (i) the procedure is an effective tool to see what one may expect from the simulation on real-world data, (ii) the consistency of RS performance ranking depend on the data considered and even the sample size used for RS training.