生成式人工智能时代的多目标推荐：最新进展与未来展望综述

2RELATEDWORKWe review existing literature on conventional and genera-tive recommendation systems, as well as MORS, to contex-tualize and motivate the present study.2.1Traditional Recommendation SystemRecommendation systems emerged in response to the infor-mation overload problem, where users are confronted withan overwhelming amount of information and struggle todiscover items or content, such as products, movies, music,or articles, that align with their interests. With the adventof big data, recommendation systems have become indis-pensable tools for filtering and surfacing relevant content tousers.Currently, algorithms in the field of recommendationsystems are generally divided into three main categories:content-based filtering, collaborative filtering, and hybridfiltering methods [24]–[26].Content-based filtering algorithmsrecommend items by analyzing item features and matchingthem with user profiles.Collaborative filtering algorithms, onthe other hand, leverage similarity measurements betweenusers (or items) to identify those with similar preferences tothe target user (or item).Hybrid filtering algorithmsintegratetwo or more recommendation techniques to overcome thelimitations of individual approaches.Regardless of the algorithm used, a crucial step is thedeep exploration of the input data. With advances in AI suchas deep learning [12], [13] and graph neural networks [27],[28], increasingly sophisticated models have been developedto better extract data features, capture complex user–item in-teractions, and integrate heterogeneous data sources. Theseinnovations have significantly improved the modeling ofuser preferences and item characteristics, helping to addresscommon challenges in recommendation systems, such as thecold-start problem and data sparsity.We could also note that numerous surveys have alreadybeen conducted on the application of AI technologies inrecommendation systems. For instance, the article [29] sys-tematically surveys eight key areas of AI technologies andtheir applications in recommendation systems. It provides acomprehensive overview of state-of-the-art AI algorithms,covering models, methods, and applications. Specifically,itdelves into deep neural networks,transfer learning,active learning, reinforcement learning, fuzzy techniques,evolutionary algorithms, natural language processing, andcomputer vision. Similarly, Masciariet al.[30] discussescommonly used AI techniques in recommendation systems,including convolutional neural networks (CNNs), collabo-rative filtering, long- and short-term memory (LSTM), de-cision trees, and Na¨ıve Bayes. These traditional AI-basedrecommendation systems primarily utilize AI technologiesfor mining existing data, including feature extraction andenriching data types, thereby significantly enhancing theperformance of recommendation systems.Now, with the rapid evolution of AI technologies, gener-ative techniques emerge as a transformative force in recom-mendation systems, demonstrating superior performancein data mining and representation learning compared totraditional AI methods. Additionally, they also excel in gen-erating creative content. Consequently, exploring generative isting studies that employ generative technologies to enablemulti-objective recommendation systems. It is crucial to em-phasize that, we consider that any discussion of additionalobjectives in recommendation systems, such as diversity,serendipity, or fairness, implicitly assumes accuracy as afoundational requirement. Thus, we definemulti-objectiverecommendation systems (MORS)as those achieving opti-mization of beyond-accuracy metrics.Our survey identifiesdiversity,serendipity,fairness, andsafetyas the primary beyond-accuracy objectives in genera-tive recommendation systems, along with additional objec-tives such asnovelty,controllability,efficiency, androbustness.Meanwhile, we focus on four generative techniques widelyadopted in recommendation systems, which areGANs,diffu-sion models,VAEs, andLLMs. For each objective, we criticallyreview the underlying model architectures and evaluationmetrics prevalent in generative recommendation systems,and also summarize relevant development challenges, aim-ing to provide foundational insights for future research inmulti-objective generative recommendation.The main contributions are summarized as follows:•This paper is the first comprehensive survey thatintegratesgenerative AI(GANs,VAEs,diffusionmodels and LLMs) with MORS. It proposes a object-oriented classification framework, systematically re-viewingadvancements and limitations in modelarchitecture,optimization strategies,and evalua-tion metrics across four key objectives: diversity,serendipity, fairness, and security.•We systematically summarize specialized evaluationmetrics and corresponding benchmark datasets fordifferentobjective domains such as fairness andserendipity, providing standardized references forexperimental design.•We also discuss key challenges in generative MORSand outline