The monograph on “Recommendation with Generative Models” extensively discusses the role of generative artificial intelligence in modern recommender systems Generative Recommender Systems (Gen-RecSys), illustrating how it integrates with deep learning models to offer richer, more personalized outputs than traditional item ranking systems. It distinguishes the various types of generative outputs such as structured outputs (e.g. bundles or sequences of items), text generation (e.g. conversational recommendations and explanations via large language models [LLMs]), and image or multimedia generation (e.g. virtual try-ons and personalized advertisements). The monograph’s structure offers a deep dive into three primary types of Gen-RecSys models: ID-based models, LLM-driven models, and multimodal systems. Chapters systematically address each model’s design, functionality, and application, including the challenges and potential of each paradigm in recommendation tasks.
Concretely, the book categorizes generative models into three primary types: ID-based models, LLM-driven models, and multimodal systems. It goes into depth about ID-based models, such as variational autoencoders (VAEs), autoregressive models, generative adversarial networks (GANs), and diffusion models. The text also expands on newer paradigms like flow-based models and sequential recommendation models that extend the standard user-item interaction matrix. For LLM-driven models, the use of textual inputs to generate personalized responses through natural language is highlighted, as is the use of multimodal generative models that integrate text, images, and video to deliver richer, context-aware recommendations. The book also explores how generative models enhance recommendation tasks, offering solutions to common issues such as the cold-start problem and data sparsity, with models like VAEs improving top-k accuracy and LLMs enhancing personalization through natural language processing.
In the second section, the evaluation and risks associated with Gen-RecSys are given separate focus. The evaluation chapter emphasizes the need for new benchmarks and measures tailored to generative models, beyond traditional rank-based metrics. It underscores that offline evaluations must account for the added complexities of generative outputs such as diversity and quality, and the need for human judgment in assessing generated recommendations. The chapter stresses that new evaluation paradigms must also tackle the inefficiencies introduced by generative systems, particularly in large-scale deployments where computational resources are stretched thin. The risks chapter takes a critical look at the ethical concerns arising from these systems. Bias amplification, privacy violations, and the potential for misinformation are key topics, with the authors highlighting how generative models can unintentionally perpetuate societal harms. The book stresses that the safe deployment of Gen-RecSys requires continuous ethical evaluations and the integration of regulatory frameworks to mitigate risks like filter bubbles, manipulation, and reward misspecification.
