arXiv subjects

Liyin Hong

Publications and source records attributed to Liyin Hong.

6 recordsLinked to original sources

arXiv2026

Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation

Industrial recommender systems build candidate pools by assigning explicit quotas to objective-specific retrieval routes. This design offers quota control but increasingly fragments modeling, training, and serving as the route set grows. Semantic-ID-based generative retrieval provides a unified alternative, yet a single decoder entangles objective policies and limits candidate complementarity. We propose Multi-Decoder OneRec, a controllable framework that combines shared representations, isolated objective adaptation, and coordinated decoding. All objectives share a user-context module and the General Decoder, while each objective adds an isolated, parameter-efficient LoRA expert. During training, exposure-sample next-token prediction (NTP) updates the shared base, target-filtered NTP updates the event-based experts, and Kullback-Leibler (KL)-regularized policy optimization updates the Watch-time expert; gradient routing isolates these updates, and the General Decoder supplies a stop-gradient reference. At inference, explicit route quotas allocate the fixed budget and Multi-Decoder Constrained Beam Search reduces cross-route overlap. We publicly release Kwai26, a large-scale multi-objective benchmark with 1.31 billion raw item-level records, 31.85 million Item-ID entries, and 25.03 million items with valid Semantic IDs, together with predefined splits and an evaluation protocol. Under the same 512-item retrieval budget, Multi-Decoder OneRec improves over the single-decoder OneRec baseline by 1.69%-5.62% across four Recall@512 metrics. In a production A/B test, it yields relative gains of 0.37% in app usage time per device, 0.19% in Day-7 retained users, 0.19% in devices with at least one share, and 2.09% in new-content Cold-Start. These results show that generative retrieval can combine shared modeling with objective-specific control and complementary candidate generation.

cs.IR↗

arXiv2026

ConAlign: Conditional Alignment Framework for Balancing Biased and Unbiased Recommendation

Industry recommender systems trained on observational data suffer from various biases that create filter bubbles, causing user interests to collapse into narrow categories and severely degrading long-term engagement. While utilizing unbiased uniform data for debiasing has shown promise, existing methods remain impractical for industrial deployment due to limitations such as neglect of factual (biased) recommendation performance and the substantial computational overhead. To overcome these limitations, we propose ConAlign (Conditional Alignment Framework), a conditional debiasing approach for industrial deployment. The key innovation of ConAlign lies in a discrete gating-based conditional alignment mechanism that selectively transfers knowledge from the biased tower to the unbiased tower. Following a selective intervention paradigm rather than universal correction, it seamlessly balances factual accuracy and unbiased preference estimation while supporting real-time streaming adaptation. To the best of our knowledge, ConAlign is the first streaming debiasing recommendation framework successfully deployed in a large-scale industrial recommendation system that utilizes a small fraction of unbiased random traffic for debiasing. Extensive offline experiments on three real-world datasets rigorously validate the effectiveness of our proposed framework. Furthermore, large-scale online A/B testing on Kuaishou demonstrates significant improvements in long-term user engagement and interest diversity, with negligible latency overhead.

cs.IR↗

arXiv2025

TADT-CSA: Temporal Advantage Decision Transformer with Contrastive State Abstraction for Generative Recommendation

With the rapid advancement of Transformer-based Large Language Models (LLMs), generative recommendation has shown great potential in enhancing both the accuracy and semantic understanding of modern recommender systems. Compared to LLMs, the Decision Transformer (DT) is a lightweight generative model applied to sequential recommendation tasks. However, DT faces challenges in trajectory stitching, often producing suboptimal trajectories. Moreover, due to the high dimensionality of user states and the vast state space inherent in recommendation scenarios, DT can incur significant computational costs and struggle to learn effective state representations. To overcome these issues, we propose a novel Temporal Advantage Decision Transformer with Contrastive State Abstraction (TADT-CSA) model. Specifically, we combine the conventional Return-To-Go (RTG) signal with a novel temporal advantage (TA) signal that encourages the model to capture both long-term returns and their sequential trend. Furthermore, we integrate a contrastive state abstraction module into the DT framework to learn more effective and expressive state representations. Within this module, we introduce a TA-conditioned State Vector Quantization (TAC-SVQ) strategy, where the TA score guides the state codebooks to incorporate contextual token information. Additionally, a reward prediction network and a contrastive transition prediction (CTP) network are employed to ensure the state codebook preserves both the reward information of the current state and the transition information between adjacent states. Empirical results on both public datasets and an online recommendation system demonstrate the effectiveness of the TADT-CSA model and its superiority over baseline methods.

cs.IR↗

arXiv2025

Reinforce Lifelong Interaction Value of User-Author Pairs for Large-Scale Recommendation Systems

Recommendation systems (RS) help users find interested content and connect authors with their target audience. Most research in RS tends to focus either on predicting users' immediate feedback (like click-through rate) accurately or improving users' long-term engagement. However, they ignore the influence for authors and the lifelong interaction value (LIV) of user-author pairs, which is particularly crucial for improving the prosperity of social community in short-video platforms. Currently, reinforcement learning (RL) can optimize long-term benefits and has been widely applied in RS. In this paper, we introduce RL to Reinforce Lifelong Interaction Value of User-Author pairs (RLIV-UA) based on each interaction of UA pairs. To address the long intervals between UA interactions and the large scale of the UA space, we propose a novel Sparse Cross-Request Interaction Markov Decision Process (SCRI-MDP) and introduce an Adjacent State Approximation (ASA) method to construct RL training samples. Additionally, we introduce Multi-Task Critic Learning (MTCL) to capture the progressive nature of UA interactions (click -> follow -> gift), where denser interaction signals are leveraged to compensate for the learning of sparse labels. Finally, an auxiliary supervised learning task is designed to enhance the convergence of the RLIV-UA model. In offline experiments and online A/B tests, the RLIV-UA model achieves both higher user satisfaction and higher platform profits than compared methods.

cs.IR↗

arXiv2025

ChorusCVR: Chorus Supervision for Entire Space Post-Click Conversion Rate Modeling

Post-click conversion rate (CVR) estimation is a vital task in many recommender systems of revenue businesses, e.g., e-commerce and advertising. In a perspective of sample, a typical CVR positive sample usually goes through a funnel of exposure to click to conversion. For lack of post-event labels for un-clicked samples, CVR learning task commonly only utilizes clicked samples, rather than all exposed samples as for click-through rate (CTR) learning task. However, during online inference, CVR and CTR are estimated on the same assumed exposure space, which leads to a inconsistency of sample space between training and inference, i.e., sample selection bias (SSB). To alleviate SSB, previous wisdom proposes to design novel auxiliary tasks to enable the CVR learning on un-click training samples, such as CTCVR and counterfactual CVR, etc. Although alleviating SSB to some extent, none of them pay attention to the discrimination between ambiguous negative samples (un-clicked) and factual negative samples (clicked but un-converted) during modelling, which makes CVR model lacks robustness. To full this gap, we propose a novel ChorusCVR model to realize debiased CVR learning in entire-space.

cs.IR↗

arXiv2025

LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou

Live-streaming, as a new-generation media to connect users and authors, has attracted a lot of attention and experienced rapid growth in recent years. Compared with the content-static short-video recommendation, the live-streaming recommendation faces more challenges in giving our users a satisfactory experience: (1) Live-streaming content is dynamically ever-changing along time. (2) valuable behaviors (e.g., send digital-gift, buy products) always require users to watch for a long-time (>10 min). Combining the two attributes, here raising a challenging question for live-streaming recommendation: How to discover the live-streamings that the content user is interested in at the current moment, and further a period in the future?

cs.IR↗