Coverage-constrained multi-objective evolutionary recommendation algorithm for balancing accuracy, diversity, and novelty.

Abstract

Deep learning mitigates data sparsity and cold-start issues in recommender systems through automatic feature extraction and pre-training techniques. However, the homogenization negatively impacts the diversity of recommendation results. To balance accuracy, diversity, and novelty, we propose a coverage-constrained multi-objective evolutionary recommendation algorithm, named cCMOERA. The algorithm employs two cooperating populations to approximate the Pareto Frontier (PF) and adjusts the fitness evaluation strategy during evolution. Additionally cCMOERA uses an improved probabilistic crossover operation, with accuracy, diversity, and novelty as objective functions and coverage as a constraint. To initialize the algorithm, we leverage a candidate recommendation list generated by a Multi-Grained Attention Recommendation(MGAR) model. The final recommendation list is generated by calculating population fitness and performing operations including cloning, crossover, and mutation. Experimental results demonstrate that cCMOERA generates recommendation lists with high accuracy, diversity, and novelty.