recommendation systems

Addressing the challenge of simultaneously achieving low latency, high accuracy, and scalability in real-time personalized recommendation for large-scale fashion e-commerce platforms, this paper proposes a novel recommendation framework integrating Graph Neural Networks (GNNs) with parsimonious learning. The method jointly models users’ heterogeneous interaction behaviors and dynamic purchase sequences, augmented by real-time streaming feature engineering to enable millisecond-level inference. Its key innovation lies in the first synergistic adoption of GNNs and parsimonious learning in recommendation systems—significantly reducing model complexity without compromising predictive accuracy. Evaluated on a large-scale industrial e-commerce dataset, the framework achieves a 12.7% improvement in Recall@20 over state-of-the-art baselines, maintains a P99 latency under 80 ms, and sustains online serving throughput of up to ten million users per second—effectively balancing accuracy, efficiency, and scalability.

This work addresses the limitations of traditional item-based collaborative filtering and two-tower models, which suffer from rigid truncation strategies and weak interaction modeling, hindering fine-grained user interest capture. The authors propose PI2I, a two-stage retrieval framework: in the first stage, a relaxed truncation threshold expands the candidate set to improve recall; in the second stage, an interactive scoring model replaces inner product computation, and negative samples are constructed from trigger–target item pairs to align training with online inference. By integrating flexible index construction with personalized interaction modeling, PI2I significantly enhances recommendation accuracy. Offline experiments demonstrate superior performance over classical collaborative filtering and parity with two-tower models. Deployed on Taobao’s “Guess You Like” feed, it achieves a 1.05% increase in transaction conversion rate and releases a public dataset containing 130 million interactions.

This work addresses the high computational complexity and inefficiency of Transformer-based sequential recommendation models when handling long user interaction histories. To this end, the authors propose a general-purpose personalized compression mechanism that leverages learnable personalized tokens to efficiently compress extensive historical interactions, which are then fused with recent user behaviors to generate recommendations. The approach is compatible with mainstream Transformer architectures such as HSTU and HLLM. Extensive experiments across multiple benchmark models demonstrate that the proposed method significantly reduces computational overhead while maintaining or even improving recommendation accuracy, effectively balancing efficiency and performance.

This work addresses a critical limitation in existing generative recommender systems, which rely on static and decoupled item tokenization schemes that ignore collaborative signals, thereby hindering end-to-end co-evolution between item indexing and the recommendation model. To overcome this, we propose the PIT framework, which introduces a novel co-generative architecture that jointly trains a dynamic, personalized item tokenizer and a generative recommender, enabling their co-evolution under aligned collaborative signals. Furthermore, we incorporate a one-to-many beam indexing mechanism to enhance scalability and robustness for industrial deployment. Extensive experiments demonstrate that PIT significantly outperforms baseline methods across multiple real-world datasets, and large-scale A/B testing on Kuaishou shows a 0.402% increase in user app dwell time.

This work addresses key challenges in industrial-scale generative recommendation—namely, the difficulty of adapting large language models (LLMs), low inference efficiency, and weak semantic representation. To this end, we propose PLUM, a novel framework featuring: (1) semantic ID tokenization for interpretable and generalizable item encoding; (2) a two-stage adaptation paradigm combining continual pretraining with generative retrieval fine-tuning to integrate world knowledge and domain-specific semantics; and (3) scalable deployment supporting billion-scale users with low latency. PLUM marks the first successful deployment of LLMs for generative retrieval recommendation at YouTube-scale. Empirical evaluation on large-scale video recommendation demonstrates that PLUM significantly outperforms highly optimized production models, validating the effectiveness, scalability, and industrial practicality of the generative paradigm in real-world recommender systems.

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 work addresses a critical limitation in existing large language model (LLM)-based recommendation approaches, which predominantly focus on token-level interactions while overlooking the item as the fundamental unit of recommendation and thus fail to effectively capture collaborative relationships among items. To bridge this gap, the paper proposes an Item-Aware Attention Mechanism (IAM) tailored for recommendation tasks, which explicitly distinguishes and separately models intra-item and inter-item token relationships. Through a two-layer attention architecture, IAM captures semantic content within individual items and models collaborative signals across items, thereby elevating the item to the core unit of recommendation modeling. Extensive experiments on multiple public recommendation datasets demonstrate that IAM significantly outperforms current LLM-based methods, confirming its effectiveness and novelty.

This work addresses a key limitation in existing multimodal foundation model–based recommender systems: despite parameter-efficient fine-tuning (PEFT), they produce uniform item embeddings that fail to capture heterogeneous user interests. To overcome this, we propose PerPEFT, the first approach that integrates user interest clustering with PEFT. Specifically, users are grouped by interest, and each group is assigned a dedicated lightweight fine-tuning module, enabling item embeddings to reflect group-specific, fine-grained characteristics. PerPEFT is compatible with any PEFT technique, introduces only 1.3% additional parameters, and achieves consistent performance gains across multiple PEFT variants—yielding up to a 15.3% improvement in NDCG@20—demonstrating strong generality, scalability, and efficient personalization capability.

This work challenges the prevailing assumption that complex generative sequential recommendation models are necessary for strong performance on mainstream benchmarks. The authors propose an untrained, minimalist graph-based heuristic that ranks items solely using the user’s most recent one or two interactions, leveraging multi-hop item transition graphs and feature similarity—without any sequence encoder or generative objective. Remarkably, this simple method achieves competitive results on 10 out of 14 standard benchmarks and yields relative NDCG@10 improvements of 38.10% and 44.18% on the Amazon Sports and CDs datasets, respectively. These findings reveal that many benchmarks exhibit “shortcut-solvable” structures—such as low-branching transitions, feature smoothness, and short-term historical dependencies—suggesting that sophisticated modeling may often be unnecessary.