Group Validation in Recommender Systems: Framework for Multi-layer Performance Evaluation

Evaluation of recommendation systems continues evolving, especially in recent years. There have been several attempts to standardize the assessment processes and propose replacement metrics better oriented toward measuring effective personalization. However, standard evaluation tools merely possess the capacity to provide a general overview of a system’s performance; they lack consistency and effectiveness in their use, as evidenced by most recent studies on the topic. Furthermore, traditional evaluation techniques fail to detect potentially harmful data on small subsets. Moreover, they generally lack explainable features to interpret how such minor variations could affect the system’s performance. This proposal focuses on data clustering for recommender evaluation and applies a cluster assessment technique to locate such performance issues. Our new approach, named group validation, aids in spotting critical performance variability in compact subsets of the system’s data and unravels hidden weaknesses in predictions where such unfavorable variations generally go unnoticed with typical assessment methods. Group validation for recommenders is a modular evaluation layer that complements regular evaluation and includes a new unique perspective to the evaluation process. Additionally, it allows several applications to the recommender ecosystem, such as model evolution tests, fraud/attack detection, and the capacity for hosting a hybrid model setup.