A Guideline-Based Approach for Assisting with the Reproducibility of Experiments in Recommender Systems Evaluation

摘要

Recommender systems' evaluation is usually based on predictive accuracy and information retrieval metrics, with better scores meaning recommendations are of higher quality. However, new algorithms are constantly developed and the comparison of results of algorithms within an evaluation framework is difficult since different settings are used in the design and implementation of experiments. In this paper, we propose a guidelines-based approach that can be followed to reproduce experiments and results within an evaluation framework. We have evaluated our approach using a real dataset, and well-known recommendation algorithms and metrics; to show that it can be difficult to reproduce results if certain settings are missing, thus resulting in more evaluation cycles required to identify the optimal settings.