personalization

This work addresses the efficient mapping of high-dimensional sparse ID features (e.g., user/item IDs) to low-dimensional continuous embeddings in recommender systems. We systematically survey mainstream embedding paradigms from 2015–2023, categorizing them into three families: collaborative filtering, self-supervised learning (contrastive and generative), and graph neural networks (e.g., node2vec). We propose the first unified taxonomy, revealing fundamental trade-offs among accuracy, generalization, and computational cost. Methodologically, we introduce three lightweight optimization directions—AutoML-based hyperparameter and architecture tuning, hash-based compression, and low-bit quantization—to bridge the gap between theoretical modeling and industrial deployment. Our contributions include: (i) a principled classification framework elucidating design principles and limitations of existing methods; (ii) novel lightweight techniques enabling scalable, memory-efficient, and latency-aware embedding generation; and (iii) comprehensive guidelines for developing efficient, deployable, and production-ready embedded recommendation systems.

This paper addresses scalability, real-time responsiveness, and trustworthiness challenges hindering the industrial deployment of recommender systems in e-commerce, healthcare, and finance. It systematically surveys technical advances from 2017 to 2024, integrating classical paradigms—such as collaborative filtering and content-based filtering—with state-of-the-art approaches, including graph neural networks, reinforcement learning, and large language models. Methodologically, it introduces the first “theory–industrial practice” mapping framework and proposes a unified evaluation paradigm incorporating fairness, explainability, and cross-domain transferability. The contributions include a comprehensive taxonomy covering 12 recommendation paradigms across 8 major application domains, an industrial decision-making guide for algorithm selection, and the open-sourcing of multiple toolkits and benchmark datasets. These resources bridge academic research and industrial implementation, significantly advancing interdisciplinary collaboration and reproducible system development.

This work addresses the challenge of effectively integrating user–item and item–item collaborative filtering to enhance Top-N recommendation performance while maintaining computational efficiency. The authors propose a weighted similarity ensemble method based on shared embeddings, which, for the first time, unifies both recommendation pathways within a single framework. By sharing user and item embeddings across strategies, the approach simplifies model architecture and eliminates the need for separate hyperparameter tuning for each pathway, thereby substantially reducing deployment complexity. Experimental results demonstrate that the proposed method achieves competitive recommendation accuracy across multiple datasets and exhibits robust performance in scenarios favoring different collaborative filtering paradigms.

This study addresses the challenge of inferring user demographic attributes across multiple scenarios using only recommendation lists, without access to explicit user profiles or interaction histories. To this end, the authors propose RAPI, a general-purpose framework that simulates the original recommender to generate augmented recommendation lists, integrates BERT-based content embeddings, and employs a dynamically weighted classifier to enhance inference performance. RAPI is the first approach to enable cross-scenario user attribute inference solely from recommendation lists, introducing both sample augmentation and adaptive weighting mechanisms to mitigate data sparsity. Extensive experiments on four real-world datasets demonstrate its effectiveness, achieving accuracy scores of 0.764 and 0.6477, respectively, and significantly outperforming baseline methods, thereby validating its robustness and generalization capability.

This work addresses the high cost, prolonged duration, and user experience disruption associated with online A/B testing in traditional recommender systems, as well as the limitations of existing large language model–based user agents that lack multimodal perception and authentic interactive capabilities, thereby failing to accurately simulate user behavior. To overcome these challenges, the authors propose A/B Agent—a novel user agent framework capable of processing multimodal inputs and supporting multi-page interactions. Within a constructed recommendation sandbox environment, A/B Agent integrates user profiles, action memory, and a fatigue mechanism, enabling, for the first time, visually aware simulation users with cross-page behavioral modeling. Experimental results demonstrate that this framework effectively substitutes real-world A/B testing across model, data, and feature dimensions, and that the synthesized data it generates significantly enhances recommender system performance.

This work investigates whether pretrained language models (PLMs) implicitly encode user preferences and collaborative signals in their representation spaces. To this end, we propose TextCF—a purely text-driven collaborative filtering framework that eliminates the need for item ID embeddings. TextCF employs lightweight linear projections to map item title representations extracted from PLMs (e.g., LLaMA, BERT) into a recommendation-aware latent space. Theoretically, we establish a homomorphic structure between linguistic and collaborative representation spaces. Empirically, TextCF significantly outperforms state-of-the-art ID-based CF methods across multiple public benchmarks—demonstrating, for the first time, that title text alone suffices to achieve superior recommendation performance. Moreover, TextCF exhibits zero-shot recommendation capability and inherent potential for user intent awareness. Collectively, it introduces a new paradigm for recommender systems that is initialization-friendly, generalizable, and inherently interpretable.

This work addresses the challenge of unified representation learning from multi-source heterogeneous user texts—such as profiles, job histories, and search logs—in large-scale job platforms. We propose a novel unsupervised framework that leverages reinforcement learning to fuse these diverse textual sources into concise, interpretable user representations tailored for large language models (LLMs). The approach utilizes implicit user interaction signals (e.g., clicks and applications) as the primary reward, augmented with rule-based rewards to constrain output format and length, thereby achieving both business relevance and interpretability without manual annotations. Extensive offline experiments across multiple LinkedIn product lines demonstrate significant improvements in key downstream recommendation metrics, validating the method’s effectiveness, scalability, and practical utility.

This work proposes RecPilot, a novel multi-agent paradigm for in-depth recommendation that moves beyond traditional item-list outputs by automatically generating user-centric, comprehensive natural language reports. Conventional recommender systems place the full burden of exploration, comparison, and decision-making on users, limiting user experience. RecPilot addresses this limitation through the collaboration of two specialized agents: a user-trajectory simulation agent and a self-evolving report generation agent. This approach redefines recommender systems from passive filtering tools into proactive intelligent services. Evaluated on public datasets, RecPilot demonstrates superior user behavior modeling, significantly reduces users’ cognitive load in evaluating recommendations, and produces highly persuasive and decision-supportive reports.

This work addresses the limitations of conventional recommendation systems, where fixed-dimensional dense user embeddings exhibit constrained representational capacity, and existing large language model (LLM)-based textual user tokens struggle to align with item attributes and model deep behavioral sequences. To overcome these challenges, the authors propose TokenMinds, a novel system that introduces semantic IDs (SIDs)—discrete representations—into user modeling for the first time. TokenMinds employs a pretrained LLM-based encoder-decoder architecture to jointly generate interpretable SID user tokens and dense embeddings, effectively balancing semantic expressiveness with compatibility for downstream recommendation tasks while enabling unified cross-scenario modeling. The system has been fully deployed across multiple YouTube production environments, demonstrating significant improvements in ranking performance and validating the effectiveness and complementary value of SID tokens in billion-scale industrial recommendation systems.

This work addresses the challenge of balancing interpretability and performance in personalized recommendation by proposing BLUE, a novel framework that unifies textual user profiles with implicit embedding representations. BLUE leverages reinforcement learning to align semantic text generated by large language models with embedding-based recommendation objectives. The approach jointly optimizes interpretability and retrieval effectiveness through an embedding-space reward and a text-space supervision signal, while also enabling cross-domain transfer. Experimental results demonstrate that BLUE significantly outperforms baseline methods on the Amazon Reviews 2023 and Google Local Reviews datasets, maintaining robust performance regardless of whether embeddings are frozen or trainable. Furthermore, it exhibits strong capabilities in cross-domain recommendation and personalized question-answering tasks.

This work addresses the high computational cost of full retraining and the challenge of adapting to user behavior drift in large-scale streaming recommender systems by proposing an efficient continual learning framework. The approach integrates gradient-informed user interaction representations with a distribution-matching sampling strategy to intelligently select a small, informative subset of data for model updates. Experimental results demonstrate that this method substantially reduces training overhead while effectively enhancing model robustness against temporal distribution shifts, thereby validating the practical utility and scalability of intelligent data selection in industrial-scale recommendation systems.