Generative Models for Intelligent Recommendation and Creativity Systems

Abstract

The integration of large generative AI models into recommendation systems and creative AI applications represents a fundamental architectural shift: from matching users to existing items in a fixed catalogue to generating novel personalised content, creative artefacts, and recommendations that transcend the boundaries of what exists in the training data. Generative recommendation systems -- which generate personalised content, product descriptions, or creative suggestions rather than ranking existing items -- offer the potential to overcome key limitations of collaborative filtering: cold-start problems, catalogue sparsity, and the inability to suggest truly novel content. This paper proposes the Generative Recommendation and Creativity (GRC) framework, a unified architecture that integrates large generative AI capabilities into three distinct application contexts: personalised content recommendation (PCR) that generates customised explanations, summaries, and previews of candidate items tailored to user preferences; creative AI assistance (CAA) that generates novel creative artefacts (music, visual art, writing prompts, design concepts) conditioned on user creative style and context; and serendipitous discovery facilitation (SDF) that deliberately generates recommendations outside user preference boundaries to facilitate novel interest discovery. GRC is evaluated through a 60-day user study with 144 participants across the three application contexts against collaborative filtering and RAG-based recommendation baselines. GRC achieves 24.8% higher user satisfaction than collaborative filtering (4.4/5.0 vs. 3.5/5.0; Cohen d = 2.48), 38.4% higher novel item acceptance rate, and 84.6% creative output acceptance rate in the CAA context. The serendipity dimension -- user acceptance of genuinely unexpected recommendations -- improves from 28.4% to 48.6% under GRC SDF versus collaborative filtering. The study contributes the GRC specification, a Generative Recommendation Quality (GRQ) metric, and empirical evidence for the substantial user experience advantages of generative over retrieval-based recommendation.