Recommender Systems (RS) are fundamental to modern online services. While most existing approaches optimize for short-term engagement, recent work has begun to explore reinforcement learning (RL) to model long-term user value. However, these efforts face significant challenges due to the vast, dynamic action spaces inherent in RS, which hinder stable policy learning. To resolve this bottleneck, we introduce Hierarchical Semantic RL (HSRL), which reframes RL-based recommendation over a fixed Semantic Action Space (SAS). HSRL encodes items as Semantic IDs (SIDs) for policy learning, and maps SIDs back to their original items via a fixed lookup during execution. To align decision-making with SID generation, the Hierarchical Policy Network (HPN) operates in a coarse-to-fine manner, employing hierarchical residual state modeling to refine each level's context from the previous level's residual, thereby reducing representation–decision mismatch. In parallel, a Multi-level Critic (MLC) provides token-level value estimates, enabling fine-grained credit assignment. Across public benchmarks and a large-scale production dataset from a leading short-video advertising platform, HSRL consistently surpasses state-of-the-art baselines. In online deployment over a 7-day A/B testing, it delivers an 18.421% ADVV lift and a 1.251% increase in Revenue, supporting HSRL as a scalable paradigm for RL-based recommendation.
论文检索
输入标题、作者或关键词，从 3,225 篇学术成果中精准定位
Industrial-scale recommender systems rely on a cascade pipeline in which the retrieval stage must return a high-recall candidate set from billions of items under tight latency. Existing solutions either (i) suffer from limited expressiveness in capturing fine-grained user-item interactions, as seen in decoupled dual-tower architectures that rely on separate encoders, or generative models that lack precise target-aware matching capabilities, or (ii) build structured indices (tree, graph, quantization) whose item-centric topologies struggle to incorporate dynamic user preferences and incur prohibitive construction and maintenance costs. We present GRank, a novel structured-index-free retrieval paradigm that seamlessly unifies target-aware learning with user-centric retrieval. Our key innovations include: (1) A target-aware Generator trained to perform personalized candidate generation via GPU-accelerated MIPS, eliminating semantic drift and maintenance costs of structured indexing; (2) A lightweight but powerful Ranker that performs fine-grained, candidate-specific inference on small subsets; (3) An end-to-end multi-task learning framework that ensures semantic consistency between generation and ranking objectives. Extensive experiments on two public benchmarks and a billion-item production corpus demonstrate that GRank improves Recall@500 by over 30% and 1.7× the P99 QPS of state-of-the-art tree- and graph-based retrievers. GRank has been fully deployed in production in our recommendation platform since Q2 2025, serving 400 million active users with 99.95% service availability. Online A/B tests confirm significant improvements in core engagement metrics, with Total App Usage Time increasing by 0.160% in the main app and 0.165% in the Lite version.
Accurately modeling long-term value (LTV) at the ranking stage of short-video recommendation systems remains a practical challenge. Though production systems and recent research have begun exploring delayed feedback and extended user engagement, modeling LTV with fine-grained attribution and robust positional normalization for billion-scale platforms is underdeveloped. In this work, we present a practical ranking-stage LTV framework that systematically addresses three core challenges: position bias, attribution ambiguity, and temporal limitations. First, to address position bias in sequential video feeds, we introduce a Position-aware Debias Quantile (PDQ) module that normalizes engagement signals using quantile-based distributions, enabling position-robust LTV estimation without requiring architectural changes. Second, we propose a multi-dimensional attribution module that learns continuous strengths across contextual, behavioral, and content-related signals, moving beyond static rule sets to capture nuanced influences among videos. Explicit noise filtering is incorporated via a customized hybrid loss, improving causal clarity in LTV attribution. Third, our cross-temporal author modeling module constructs censoring-aware, day-level long-term value targets, capturing creator-driven re-engagement over extended time windows. While our framework currently focuses on the author dimension, it is readily extensible to further aspects such as topics or styles. Extensive offline experiments and online A/B tests demonstrate statistically significant gains in LTV-related metrics and stable trade-offs with short?term objectives. The framework is realized as task augmentation within an existing ranking model, facilitates billion-scale deployment on Taobao's production system with efficient training and serving, achieving sustained user engagement improvements while remaining compatible with industrial constraints.
Web recommendations provide personalized items from massive catalogs for users, which rely heavily on retrieval stages to trade off the effectiveness and efficiency of selecting a small relevant set from billion-scale candidates in online digital platforms. As one of the largest Chinese search engine and news feed providers, Baidu resorts to Deep Neural Network (DNN) and graph-based Approximate Nearest Neighbor Search (ANNS) algorithms for accurate relevance estimation and efficient search for relevant items. However, current retrieval at Baidu fails in comprehensive user-item relational understanding due to dissected interaction modeling, and performs inefficiently in large-scale graph-based ANNS because of suboptimal traversal navigation and the GPU computational bottleneck under high concurrency. To this end, we propose a GPU-accelerated Multi-relational Parallel Graph Retrieval (GMP-GR) framework to achieve effective yet efficient retrieval in web-scale recommendations. First, we propose a multi-relational user-item relevance metric learning method that unifies diverse user behaviors through multi-objective optimization and employs a self-covariant loss to enhance pathfinding performance. Second, we develop a hierarchical parallel graph-based ANNS to boost graph retrieval throughput, which conducts breadth-depth-balanced searches on a large-scale item graph and cost-effectively handles irregular neural computation via adaptive aggregation on GPUs. In addition, we integrate system optimization strategies in the deployment of GMP-GR in Baidu. Extensive experiments demonstrate the superiority of GMP-GR in retrieval accuracy and efficiency. Deployed across more than twenty applications at Baidu, GMP-GR serves hundreds of millions of users with a throughput exceeding one hundred million requests per second.
Federated learning offers a promising paradigm for privacy-preserving traffic prediction, yet its performance is often challenged by the non-identically and independently distributed (non-IID) nature of decentralized traffic data. Existing federated methods frequently struggle with this data heterogeneity, typically entangling globally shared patterns with client-specific local dynamics within a single representation. In this work, we postulate that this heterogeneity stems from the entanglement of two distinct generative sources: client-specific localized dynamics and cross-client global spatial-temporal patterns. Motivated by this perspective, we introduce FedDis, a novel framework that, to the best of our knowledge, is the first to leverage causal disentanglement for federated spatial-temporal prediction. Architecturally, FedDis comprises a dual-branch design wherein a Personalized Bank learns to capture client-specific factors, while a Global Pattern Bank distills common knowledge. This separation enables robust cross-client knowledge transfer while preserving high adaptability to unique local environments. Crucially, a mutual information minimization objective is employed to enforce informational orthogonality between the two branches, thereby ensuring effective disentanglement. Comprehensive experiments conducted on four real-world benchmark datasets demonstrate that FedDis consistently achieves state-of-the-art performance, promising efficiency, and superior expandability. To ease the reproducibility, we release our implementation code online. https://github.com/Jlu-zcy/www2026_FedDis.
The rapid growth of online short texts has made specialized analysis essential, as these texts are sparse and information-limited. Short text clustering (STC) is critical for automatically grouping unlabeled texts into meaningful clusters, supporting applications such as sentiment analysis, spam filtering, and social media personalization. In the context of massive online content, deep clustering seeks to uncover semantic categories by measuring distances in the representation space. Consequently, aligning clustering pseudo-labels with the true category distribution is crucial for effective self-supervised training, particularly under class imbalance and distribution skew commonly observed in web data. To address this challenge, we propose the Cluster Identification-Guided Dual Correction (CIDC) framework, which generates reliable pseudo-labels to guide deep clustering. Specifically, given cluster partitions and model-estimated class distributions, we perform Cluster Category Identification (CCI) at each training epoch to determine the most probable category for each cluster. This identification provides the foundation for the Pseudo-Label Correction (PLC) and Prototype-Based Correction (PBC) modules, which jointly enhance pseudo-label reliability and representation learning. In the PLC module, samples whose model-estimated class distributions conflict with the assigned cluster category are corrected, thereby improving semantic alignment within clusters. In the PBC module, representative and reliable prototypes are selected according to cluster categories and model predictions to guide training, further strengthening representation discriminability. Extensive experiments demonstrate that CIDC consistently outperforms existing methods in terms of clustering accuracy and mutual information, particularly in unsupervised settings characterized by class imbalance and noisy data.
As an important task in modern web technologies, multivariate time series forecasting drives many core functionalities. However, in realistic web-scale environments, sensor failures, privacy filtering, sampling, and instrumentation churn frequently yield missing variables, making training sets complete while test sets contain only a small subset of variables. The challenge lies in utilizing incomplete data for forecasting, which is known as Variable Subset Forecasting (VSF). Distribution shift is inherent to time series and remains in VSF, including inter-series shift as changes in cross-series correlations and intra-series shift as substantial distribution differences within the same series across different time windows. Existing VSF approaches typically impute missing variables and then forecast on the completed series, yet they overlook these shifts and thus underperform in realistic web-scale scenarios. To address these challenges, we propose Shift Resilient Diffusive Imputation (SRDI), a framework tailored to VSF and robust to distribution shift. Specifically, SRDI integrates a divide-conquer strategy with the denoising process, which decomposes the input into invariant patterns and variant patterns, representing the temporally stable parts of inter-series correlation and the highly fluctuating parts, respectively. By extracting spatiotemporal features from each part separately and then appropriately combining them, inter-series shift can be effectively mitigated. Then, we innovatively organize SRDI and the forecasting model into a meta-learning paradigm tailored for VSF scenarios. We address the intra-series shift by treating time windows as tasks during training and employing an adaptation process before testing, which naturally supports robust online forecasting in dynamic web environments. Extensive experiments on four datasets have demonstrated our superior performance compared with state-of-the-art methods. Our code is available at https://github.com/xhhmacau/SRDI.
The prevalence of sarcasm in multimodal dialogues on the social platforms presents a crucial yet challenging task for understanding the true intent behind online content. Comprehensive sarcasm analysis requires two key aspects: Multimodal Sarcasm Detection (MSD) and Multimodal Sarcasm Explanation (MuSE). Intuitively, the act of detection is the result of the reasoning process that explains the sarcasm. Current research predominantly focuses on addressing either MSD or MuSE as a single task. Even though some recent work has attempted to integrate these tasks, their inherent causal dependency is often overlooked. To bridge this gap, we propose MuVaC, a variational causal inference framework that mimics human cognitive mechanisms for understanding sarcasm, enabling robust multimodal feature learning to jointly optimize MSD and MuSE. Specifically, we first model MSD and MuSE from the perspective of structural causal models, establishing variational causal pathways to define the objectives for joint optimization. Next, we design an alignment-then-fusion approach to integrate multimodal features, providing robust fusion representations for sarcasm detection and explanation generation. Finally, we enhance the reasoning trustworthiness by ensuring consistency between detection results and explanations. Experimental results demonstrate the superiority of MuVaC in public datasets, offering a new perspective for understanding multimodal sarcasm.
Knowledge tracing (KT) aims to personalize online education on large-scale web-based platforms by modeling students' evolving knowledge states from their interaction sequences. However, most KT models rely on a single encoder architecture (e.g., self-attention or RNN), with fixed inductive biases that fails to capture the diversity of learning behaviors. Specifically, student learning unfolds across multiple timescales, and interaction sequences contain diverse frequency components ranging from short-term variations to long-term trends. Our data-driven analysis reveals that existing encoders exhibit characteristic frequency biases (e.g., self-attention tends to emphasize low-frequency patterns), highlighting the limitations of any single architecture. To address this problem, we propose FA-KT, a frequency-aware mixture of heterogeneous experts framework. FA-KT combines self-attention, Mamba, CNN, and LSTM experts, each with complementary frequency biases. A frequency-aware router analyzes each sequence's frequency characteristics and adaptively combines experts to create dynamic, personalized encoders for individual students. Across five benchmark datasets, FA-KT consistently outperforms 20 strong KT baselines in predicting future performance. Code is available at https://pykt.org/.
The rapid spread of online misinformation poses a serious threat to public trust and social stability, making automatic fake news detection increasingly critical. In this work, we propose a novel retrieval-augmented fake news detection framework, Triple-R (Rewriting, Retrieval, and iterative Refinement), which emphasizes optimizing the retrieval query. Unlike prior retrieval-augmented methods that adapt either the retriever or the verification module, often overlooking the gap between news text and the evidence needed for verification, our approach focuses on adapting the search query to retrieve the most relevant evidence. Specifically, we employ a small language model as a trainable query rewriter, optimized via reinforcement learning with feedback from a frozen LLM-based fake news detector, to transform the original news text into effective retrieval queries. To further enhance evidence relevance, we introduce an iterative query refinement mechanism, which progressively updates rewritten queries based on previously retrieved results. Finally, the original news text and the evidence retrieved through refined queries are integrated for verification. Experiments on two real-world datasets demonstrate consistent improvements, validating the effectiveness of our approach.
Temporal Point Processes (TPPs) have been widely used for modeling event sequences on the Web, such as user reviews, social media posts, and online transactions. However, traditional TPP models often struggle to effectively incorporate the rich textual descriptions that accompany these events, while Large Language Models (LLMs), despite their remarkable text processing capabilities, lack mechanisms for handling the temporal dynamics inherent in Web-based event sequences. To bridge this gap, we introduce Language-TPP, a unified framework that seamlessly integrates TPPs with LLMs for enhanced Web event sequence modeling. Our key innovation is a novel temporal encoding mechanism that converts continuous time intervals into specialized byte-tokens, enabling direct integration with standard language model architectures for TPP modeling without requiring TPP-specific modifications. This approach allows Language-TPP to achieve state-of-the-art performance across multiple TPP benchmarks, including event time prediction and type prediction, on real-world Web datasets spanning e-commerce reviews, social media and online Q&A platforms. More importantly, we demonstrate that our unified framework unlocks new capabilities for TPP research: incorporating temporal information improves the quality of generated event descriptions, as evidenced by enhanced ROUGE-L scores, and better aligned sentiment distributions. Through comprehensive experiments, including qualitative analysis of learned distributions and scalability evaluations on long sequences, we show that Language-TPP effectively captures both temporal dynamics and textual patterns in Web user behavior, with important implications for content generation, user behavior understanding, and Web platform applications. Code is available at https://github.com/qykong/Language-TPP.
Logs serve as a primary source of information for engineers to diagnose failures in large-scale online service systems. Log parsing, which extracts structured events from massive unstructured log data, is a critical first step for downstream tasks like anomaly detection and failure diagnosis. With advances in large language models (LLMs), leveraging their strong text understanding capabilities has proven effective for accurate log parsing. However, existing LLM-based log parsers all focus on the constant part of logs, ignoring the potential contribution of the variable part to log parsing. This constant-centric strategy brings four key problems. First, inefficient log grouping and sampling with only constant information. Second, a relatively large number of LLM invocations due to constant-based cache, leading to low log parsing accuracy and efficiency. Third, a relatively large number of consumed constant tokens in prompts leads to high LLM invocation costs. At last, these methods only retain placeholders in the results, losing the system visibility brought by variable information in logs. Facing these problems, we propose a variable-centric log parsing strategy named VarParser. Through variable contribution sampling, variable-centric parsing cache, and adaptive variable-aware in-context learning, our approach can efficiently capture the variable parts of logs and leverage their contributions to parsing. By introducing variable units, we preserve rich variable information, enhancing the integrity of log parsing results. Extensive evaluations on large-scale datasets demonstrate that VarParser achieves higher accuracy compared to existing methods, significantly improving parsing efficiency while reducing the LLM invocation costs.
Click-through rate (CTR) prediction plays a pivotal role in online advertising and recommender systems. Despite notable progress in modeling user preferences from historical behaviors, two key challenges persist. First, exsiting discriminative paradigms focus on matching candidates to user history, often overfitting to historically dominant features and failing to adapt to rapid interest shifts. Second, a critical information chasm emerges from the point-wise ranking paradigm. By scoring each candidate in isolation, CTR models discard the rich contextual signal implied by the recalled set as a whole, leading to a misalignment where long-term preferences often override the user's immediate, evolving intent. To address these issues, we propose GenCI, a generative user intent framework that leverages semantic interest cohorts to model dynamic user preferences for CTR prediction. The framework first employs a generative model, trained with a next-item prediction (NTP) objective, to proactively produce candidate interest cohorts. These cohorts serve as explicit, candidate-agnostic representations of a user's immediate intent. A hierarchical candidate-aware network then injects this rich contextual signal into the ranking stage, refining them with cross-attention to align with both user history and the target item. The entire model is trained end-to-end, creating a more aligned and effective CTR prediction pipeline. Extensive experiments on three widely used datasets demonstrate the effectiveness of our approach.
Watch time prediction (WTP) has emerged as a pivotal task in short video recommendation systems, designed to quantify user engagement through continuous interaction modeling. Predicting users' watch times on videos often encounters fundamental challenges, including wide value ranges and imbalanced data distributions, which can lead to significant estimation bias when directly applying regression techniques. Recent studies have attempted to address these issues by converting the continuous watch time estimation into an ordinal regression task. While these methods demonstrate partial effectiveness, they exhibit notable limitations: (1) The discretization process frequently relies on bucket partitioning, inherently reducing prediction flexibility and accuracy. (2) The interdependencies among different partition intervals remain underutilized, missing opportunities for effective error correction. Inspired by language modeling paradigms, we propose a novel Generative Regression (GR) framework that reformulates WTP as a sequence generation task. Our approach employs structural discretization to enable nearly lossless value reconstruction while maintaining prediction flexibility. Through carefully designed vocabulary construction and label encoding schemes, each watch time is bijectively mapped to a token sequence. To mitigate the training-inference discrepancy caused by teacher-forcing, we introduce a curriculum learning with embedding mixup strategy that gradually transitions from guided to free-generation modes. We test our models extensively on two public datasets, a large-scale offline industrial dataset, and an online A/B test on Kuaishou App with over 400 million daily active users (DAU) and GR consistently outperforms existing state-of-the-art approaches significantly. Our code is available at https://github.com/snailma0229/GR.git.
Large language models (LLMs) have recently demonstrated strong potential for sequential recommendation. However, current LLM-based approaches face critical limitations in modeling users' long-term and diverse interests. First, due to inference latency and feature fetching bandwidth constraints, existing methods typically truncate user behavior sequences to include only the most recent interactions, resulting in the loss of valuable long-range preference signals. Second, most current methods rely on next-item prediction with a single predicted embedding, overlooking the multifaceted nature of user interests and limiting recommendation diversity. To address these challenges, we propose HyMiRec, a hybrid multi-interest sequential recommendation framework, which leverages a lightweight recommender to extracts coarse interest embeddings from long user sequences and an LLM-based recommender to captures refined interest embeddings. To alleviate the overhead of fetching features, we introduce a residual codebook based on cosine similarity, enabling efficient compression and reuse of user history embeddings. To model the diverse preferences of users, we design a disentangled multi-interest learning module, which leverages multiple interest queries to learn disentangles multiple interest signals adaptively, allowing the model to capture different facets of user intent. Extensive experiments are conducted on both benchmark datasets and a collected industrial dataset, demonstrating our effectiveness over existing state-of-the-art methods. Furthermore, online A/B testing shows that HyMiRec brings consistent improvements in real-world recommendation systems.
Multimodal recommendation advocates integrating the multimodal features of items with historical user behaviors to enhance recommendation accuracy across various online media platforms. The majority of existing methods concentrate on leveraging cross-modal learning over multimodal features to augment node representations. However, these approaches are confronted with two key challenges: i) augmented representations offer limited information gain for interactive prediction in the collaborative view, and ii) semantic discrepancy between the collaborative view and modality-augmented features remains inadequately addressed. To overcome these obstacles, we present a new Multi-view Semantic Contrastive Alignment (MSCA) approach for multimodal recommendation, which models and aligns node representations from multiple views. Specifically, we introduce a multi-view semantic pattern encoder that learns basic embeddings from the collaborative view and independently captures augmented semantic patterns from the item-item structural view and intra-modal view. Furthermore, a semantic contrastive alignment task is designed to mitigate the semantic divergence between collaborative embeddings and augmented representations by maximizing the mutual consistency between them, thereby facilitating an effective integration of both. Comprehensive experiments on three benchmark datasets confirm that the proposed MSCA consistently excels over diverse state-of-the-art baselines.
We study off-policy learning (OPL) in contextual bandits, which plays a key role in a wide range of real-world applications such as recommendation systems and online advertising. Typical OPL in contextual bandits assumes an unconstrained environment where a policy can select the same item infinitely. However, in many practical applications, including coupon allocation and e-commerce, limited supply constrains items through budget limits on distributed coupons or inventory restrictions on products. In these settings, greedily selecting the item with the highest expected reward for the current user may lead to early depletion of that item, making it unavailable for future users who could potentially generate higher expected rewards. As a result, OPL methods that are optimal in unconstrained settings may become suboptimal in limited supply settings. To address the issue, we provide a theoretical analysis showing that conventional greedy OPL approaches may fail to maximize the policy performance, and demonstrate that policies with superior performance must exist in limited supply settings. Based on this insight, we introduce a novel method called Off-Policy learning with Limited Supply (OPLS). Rather than simply selecting the item with the highest expected reward, OPLS focuses on items with relatively higher expected rewards compared to the other users, enabling more efficient allocation of items with limited supply. Our empirical results on both synthetic and real-world datasets show that OPLS outperforms existing OPL methods in contextual bandit problems with limited supply.
Large-scale online marketplaces and recommender systems are crucial technological foundations for the development of e-commerce. In industrial recommender systems, features play a vital role as they carry essential information for downstream models. Accurate estimation of feature importance is critical, as it helps identify the most useful feature subsets from thousands of candidates for online services. Such a selection enables optimization of online performance while reducing computational burden. To address the feature selection challenges in deep learning, trainable gate-based and sensitivity-based methods have been proposed and proven effective in the industry. However, by analyzing real-world examples, we identified three bias issues that cause feature importance estimation to rely on partial model layers, samples, or gradients, ultimately leading to inaccurate feature importance estimates. We refer to these biases as layer bias, baseline bias, and approximation bias. To mitigate these biases, we propose FairFS, a fair and accurate feature selection algorithm. On one hand, FairFS directly regularizes feature importance estimation across all non-linear transformational layers to avoid layer bias. On the other hand, it employs a smooth baseline feature close to the classifier's decision boundary and an aggregated approximation method to mitigate bias issues. Extensive experiments demonstrate how FairFS mitigates these three biases and achieves state-of-the-art feature selection results.
Large-scale vision-language models enable powerful cross-modal understanding and generation, driving rapidly growing demand for online inference services. However, cloud-centric serving often suffers from high latency, rising costs, and network dependency, while purely on-device deployment is constrained by limited memory and reduced accuracy on complex tasks. To address this accuracy–latency–cost trilemma, we propose ShiftVL, a task-aware end–cloud serving framework that shifts suitable execution to the end device with a cloud fallback. ShiftVL serves high-frequency requests on an end-side small VLM enhanced with ViTexLoRA, a modality-disentangled parameter-efficient tuning method that preserves cross-modal alignment, while routing low-frequency or complex requests to a cloud-hosted large VLM for higher accuracy. Under tight device budgets, ShiftVL employs a predictive adapter scheduler that combines LRU-style caching with imitation learning to pre-load task-specific adapters. Experiments with InternVL models show that ShiftVL reduces cloud cost by up to 76.3% and latency by up to 42.9% while maintaining high multi-task accuracy, demonstrating its practicality for real-world vision-language model serving.
Social media engagement prediction is a central challenge in computational social science, particularly for understanding how users interact with misinformation. Existing approaches often treat engagement as a homogeneous time-series signal, overlooking the heterogeneous social mechanisms and platform designs that shape how misinformation spreads. In this work, we ask: ''Can neural architectures discover social exchange principles from behavioral data alone?'' We introduce Dreams (Disentangled Representations and Episodic Adaptive Modeling for Social media misinformation engagements), a social exchange theory-guided framework that models misinformation engagement as a dynamic process of social exchange. Rather than treating engagement as a static outcome, Dreams models it as a sequence-to-sequence adaptation problem, where each action reflects an evolving negotiation between user effort and social reward conditioned by platform context. It integrates adaptive mechanisms to learn how emotional and contextual signals propagate through time and across platforms. On a cross-platform dataset spanning 7 platforms and 2.37M posts collected between 2021 and 2025, Dreams achieves state-of-the-art performance in predicting misinformation engagements, reaching a mean absolute percentage error of 19.25%. This is a 43.6% improvement over the strongest baseline. Beyond predictive gains, the model reveals consistent cross-platform patterns that align with social exchange principles, suggesting that integrating behavioral theory can enhance empirical modeling of online misinformation engagement. The source code is available at: https://github.com/ltian678/DREAMS.