Exploiting Surrogate Submodular and Cost-Effective Lazy Forward Algorithms for Calibrated Recommendations

Recommendation systems are tools to suggest items that might interest users. These systems rely on the user’s preference history to generate a list of suggestions most similar to items in the user’s history, aiming for better accuracy and minimal error. Pursuing higher accuracy can lead to side effects such as overspecialization, reduced diversity, and imbalances in categories, genres, or niches. Thus, this work explores algorithms for selecting the items that form a calibrated recommendation list. The hypothesis is that calibration can positively contribute to fairer recommendations based on user preferences. We introduce a variation of an existing algorithm, which performs better than the baselines in a commonly used dataset in two different divergence measurements.