Large Language Models (LLMs), despite their remarkable capabilities, are prone to generating hallucinated or outdated content due to their static internal knowledge. While Retrieval-Augmented Generation (RAG) integrated with Reinforcement Learning (RL) offers a solution, these methods are fundamentally constrained by a single-query mode, leading to prohibitive latency and inherent brittleness. To overcome these limitations, we introduce RAG-R1, a novel two-stage training framework centered around multi-query parallelism. Our framework enables LLMs to adaptively leverage internal and external knowledge during the reasoning process while transitioning from the single-query mode to multi-query parallelism. This architectural shift bolsters reasoning robustness while significantly reducing inference latency. Extensive experiments on seven question-answering benchmarks confirm the superiority of our method, which outperforms the strongest baseline by up to 13.7% and decreases inference time by 11.1%.

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating external knowledge, where the LLM's ability to generate responses based on the combination of a given query and retrieved documents is crucial. However, most benchmarks focus on overall RAG system performance, rarely assessing LLM-specific capabilities. Current benchmarks emphasize broad aspects such as noise robustness, but lack a systematic and granular evaluation framework on document utilization. To this end, we introduce Placeholder-RAG-Benchmark, a multi-level fine-grained benchmark, emphasizing the following progressive dimensions: (1) multi-level filtering abilities, (2) combination abilities, and (3) reference reasoning. To provide a more nuanced understanding of LLMs' roles in RAG systems, we formulate an innovative placeholder-based approach to decouple the contributions of the LLM's parametric knowledge and the external knowledge. Experiments demonstrate the limitations of representative LLMs in the RAG system's generation capabilities, particularly in error resilience and context faithfulness. Our benchmark provides a reproducible framework for developing more reliable and efficient RAG systems.

Lei Liu, Hao Zhu, Xiaoyan Yang, Yue Shen, Zhixuan Chu, Jian Wang, Jinjie Gu, Kui Ren. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Junjie Wang, Zequn Xie, Dan Yang, Jie Feng, Yue Shen, Duolin Sun, Meixiu Long, Yihan Jiao, Zhehao Tan, Jian Wang, Peng Wei, Jinjie Gu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Knowledge transfer aims to apply existing knowledge to different tasks or new data, and it has extensive applications in multi-domain and Multi-Task Learning. The key to this task is quickly identifying a fine-grained object for knowledge sharing and efficiently transferring knowledge. Current methods, such as fine-tuning, layer-wise parameter sharing, and task-specific adapters, only offer coarse-grained sharing solutions and struggle to effectively search for shared parameters, thus hindering the performance and efficiency of knowledge transfer. To address these issues, we propose Channel-Wise Parameter Sharing (CWPS), a novel fine-grained parameter-sharing method for knowledge transfer, which is efficient for parameter sharing, comprehensive, and plug-and-play. For the coarse-grained problem, we first achieve fine-grained parameter sharing by refining the granularity of shared parameters from the level of layers to the level of neurons. The knowledge learned from previous tasks can be utilized through the explicit composition of the model neurons. Besides, we promote an effective search strategy to minimize computational costs, simplifying the selection of shared weights. In addition, our CWPS has strong composability and generalization ability, which theoretically can be applied to any network consisting of linear and convolution layers. We introduce several datasets in both Incremental Learning and Multi-Task Learning scenarios. Our method has achieved state-of-the-art precision-to-parameter ratio performance with various backbones, demonstrating its efficiency and versatility.

The emergence of long-context text applications utilizing large language models (LLMs) has presented significant scalability challenges, particularly in memory footprint. The linear growth of the Key-Value (KV) cache, which stores attention keys and values to reduce redundant computations, can significantly increase memory usage and may prevent models from functioning properly in memory-constrained environments. To address this issue, we propose a novel approach called Cache Sparse Representation (CSR), which converts the KV cache by transforming the dense Key-Value cache tensor into sparse indexes and weights, offering a more memory-efficient representation during LLM inference. Furthermore, we introduce NeuralDict, a novel neural network-based method to automatically generate the dictionary used in our sparse representation. Our extensive experiments demonstrate that CSR matches the performance of state-of-the-art KV cache quantization algorithms while ensuring robust functionality in memory-constrained environments.

This paper presents the outcomes of the PETS2025 challenge, held in conjunction with AVSS 2025 and sponsored by the EU-funded EURMARS project. The challenge introduces a novel maritime surveillance dataset comprising image sequences captured by diverse multi-altitude, multimodal sensors, reflecting the real-world multi-authority environment. The key tasks include: (1) object detection using various sensors across different platforms (ground-based and low-altitude aerial) and spectral ranges (visible, thermal, ultraviolet (UV), and short-wave infrared (SWIR)); (2) long-term tracking of targets in maritime environments spanning both sea and land; and (3) approximating target geolocations by using sensor imagery and telemetry data. Performance evaluations of results submitted by 12 international participants are discussed. The results show the effectiveness of these submissions and highlight ongoing challenges posed by heterogeneous sensors and complex environments. These challenges emphasise the need to further improve detection, tracking, and geolocation approximation for maritime and coastal surveillance.

In this paper, we seek to carefully prompt a Large Language Model (LLM) with domain-level knowledge as a better marketing-oriented knowledge miner for marketing-oriented knowledge graph construction, which is non-trivial, suffering from several inevitable issues in real-world marketing scenarios, i.e., uncontrollable relation generation of LLMs, insufficient prompting ability of a single prompt, unaffordable deployment cost of LLMs. To this end, we propose PAIR, a novel Progressive prompting Augmented mIning fRamework for harvesting marketing-oriented knowledge graph with LLMs. In particular, we reduce the pure relation generation to an LLM-based adaptive relation filtering process through knowledge-empowered prompting technique. Next, we steer LLMs for entity expansion with progressive prompting augmentation, followed by a reliable aggregation with comprehensive consideration of both self-consistency and semantic relatedness. In terms of online serving, we specialize in a small and white-box PAIR (i.e., LightPAIR), which is fine-tuned with a high-quality corpus provided by a strong teacher-LLM. Extensive experiments and practical applications in audience targeting verify the effectiveness of the proposed (Light)PAIR.

Existing multi-modal object tracking approaches primarily focus on dual-modal paradigms, such as RGB-Depth or RGB-Thermal, yet remain challenged in complex scenarios due to limited input modalities. To address this gap, this work introduces a novel multi-modal tracking task that leverages three complementary modalities, including visible RGB, Depth (D), and Thermal Infrared (TIR), aiming to enhance robustness in complex scenarios. To support this task, we construct a new multi-modal tracking dataset, coined RGBDT500, which consists of 500 videos with synchronised frames across the three modalities. Each frame provides spatially aligned RGB, depth, and thermal infrared images with precise object bounding box annotations.Furthermore, we propose a novel multi-modal tracker, dubbed RDTTrack.RDTTrack integrates tri-modal information for robust tracking by leveraging a pretrained RGB-only tracking model and prompt learning techniques.In specific, RDTTrack fuses thermal infrared and depth modalities under a proposed orthogonal projection constraint, then integrates them with RGB signals as prompts for the pre-trained foundation tracking model, effectively harmonising tri-modal complementary cues.The experimental results demonstrate the effectiveness and advantages of the proposed method, showing significant improvements over existing dual-modal approaches in terms of tracking accuracy and robustness in complex scenarios. The dataset and source code are publicly available at https://xuefeng-zhu5.github.io/RGBDT500.

Most start-ups fail, and early-stage ventures face even lower survival rates. Identifying high-potential start-ups remains a critical challenge for venture capital (VC) investors and policymakers. While predictive models exist, the evolving relationships between VC investors, start-ups, and management teams in dynamic networks are underexplored. We propose a method to predict whether a start-up will succeed within 5 years of its first funding round. Using a 40-year global VC dataset, we model the VC ecosystem as a dynamic bipartite network linking start-ups to individuals (investors/managers). Our approach incrementally updates graph embeddings through unsupervised self-attention to incorporate new nodes, edges, and their neighbors. Node embeddings are further fine-tuned via link prediction and classification tasks, while temporal dependencies are captured to form sequential representations. The model identifies early-stage start-ups with twice the success likelihood of those chosen by professional investors. Key factors including networking and education align with VC literature. Additionally, we provide model complexity analysis and open source our implementation to support practical applications and future research.

Accurate prediction of coupon usage is crucial for promoting user consumption through targeted coupon recommendations. However, in real-world coupon recommendations, the coupon allocation process is not solely determined by the model trained with the history interaction data but is also interfered with by marketing tactics desired to fulfill specific commercial goals.This interference creates an imbalance in the interactions, which causes the data to deviate from the user's natural preferences. We refer to this deviation as the matching bias. Such biased interaction data affects the efficacy of the model, and thus it is necessary to employ debiasing techniques to prevent any negative impact. We investigate the mitigation of matching bias in coupon recommendations from a causal-effect perspective. By treating the attributes of users and coupons associated with marketing tactics as confounders, we find the confounders open the backdoor path between user-coupon matching and the conversion, which introduces spurious correlation. To remove the bad effect, we propose a novel training paradigm named Backdoor Adjustment via Group Adaptation (BAGA) for debiased coupon recommendations, which performs intervened training and inference, i.e., separately modeling each user-coupon group pair. However, modeling all possible group pairs greatly increases the computational complexity and cost. To address the efficiency challenge, we further present a simple but effective dual-tower multi-task framework and leverage the Customized Gate Control (CGC) model architecture, which separately models each user and coupon group with a separate expert module. We instantiate BAGA on five representative models: FM, DNN, NCF, MASKNET, and DEEPFM, and conduct comprehensive offline and online experiments to demonstrate the efficacy of our proposed paradigm.

Recommendation systems aim to provide users with relevant suggestions, but often lack interpretability and fail to capture higher-level semantic relationships between user behaviors and profiles. In this paper, we propose a novel approach that leverages large language models (LLMs) to construct personalized reasoning graphs. These graphs link a user's profile and behavioral sequences through causal and logical inferences, representing the user's interests in an interpretable way. Our approach, LLM reasoning graphs (LLMRG), has four components: chained graph reasoning, divergent extension, self-verification and scoring, and knowledge base self-improvement. The resulting reasoning graph is encoded using graph neural networks, which serves as additional input to improve conventional recommender systems, without requiring extra user or item information. Our approach demonstrates how LLMs can enable more logical and interpretable recommender systems through personalized reasoning graphs. LLMRG allows recommendations to benefit from both engineered recommendation systems and LLM-derived reasoning graphs. We demonstrate the effectiveness of LLMRG on benchmarks and real-world scenarios in enhancing base recommendation models.

The application of mixture-of-experts (MoE) is gaining popularity due to its ability to improve model's performance. In an MoE structure, the gate layer plays a significant role in distinguishing and routing input features to different experts. This enables each expert to specialize in processing their corresponding sub-tasks. However, the gate's routing mechanism also gives rise to "narrow vision": the individual MoE's expert fails to use more samples in learning the allocated subtask, which in turn limits the MoE to further improve its generalization ability. To effectively address this, we propose a method called Mixture-of-Distilled-Expert (MoDE), which applies moderate mutual distillation among experts to enable each expert to pick up more features learned by other experts and gain more accurate perceptions on their allocated sub-tasks. We conduct plenty experiments including tabular, NLP and CV datasets, which shows MoDE's effectiveness, universality and robustness. Furthermore, we develop a parallel study through innovatively constructing "expert probing", to experimentally prove why MoDE works: moderate distilling knowledge from other experts can improve each individual expert's test performances on their assigned tasks, leading to MoE's overall performance improvement.

Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Lei Liang, Jinjie Gu, Huajun Chen. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Relevance modeling plays a crucial role in e-commerce search engines, striving to identify the utmost pertinent items corresponding to a given search query. With the rapid advancement of pre-trained large language models (LLMs), recent endeavors have leveraged the capabilities of LLMs in relevance modeling, resulting in enhanced performance. This is usually done through the process of fine-tuning LLMs on specifically annotated datasets to determine the relevance between queries and items. However, there are two limitations when LLMs are naively employed for relevance modeling through fine-tuning and inference. First, it is not inherently efficient for performing nuanced tasks beyond simple yes or no answers, such as assessing search relevance. It may therefore tend to be overconfident and struggle to distinguish fine-grained degrees of relevance (e.g., strong relevance, weak relevance, irrelevance) used in search engines. Second, it exhibits significant performance degradation when confronted with data distribution shift in real-world scenarios. In this paper, we propose a novel Distribution-Aware Robust Learning framework (DaRL) for relevance modeling in Alipay Search. Specifically, we design an effective loss function to enhance the discriminability of LLM-based relevance modeling across various fine-grained degrees of query-item relevance. To improve the generalizability of LLM-based relevance modeling, we first propose the Distribution-Aware Sample Augmentation (DASA) module. This module utilizes out-of-distribution (OOD) detection techniques to actively select appropriate samples that are not well covered by the original training set for model fine-tuning. Furthermore, we adopt a multi-stage fine-tuning strategy to simultaneously improve in-distribution (ID) and OOD performance, bridging the performance gap between them. DaRL has been deployed online to serve the Alipay's insurance product search. Both offline experiments on real-world industry data and online A/B testing show that DaRL effectively improves the performance of relevance modeling.

In the realm of information retrieval, users often engage in multi-turn interactions with search engines to acquire information, leading to the formation of sequences of user feedback behaviors. Leveraging the session context has proven to be beneficial for inferring user search intent and document ranking. A multitude of approaches have been proposed to exploit in-session context for improved document ranking. Despite these advances, the limitation of historical session data for capturing evolving user intent remains a challenge. In this work, we explore the integration of future contextual information into the session context to enhance document ranking. We present the siamese model optimization framework, comprising a history-conditioned model and a future-aware model. The former processes only the historical behavior sequence, while the latter integrates both historical and anticipated future behaviors. Both models are trained collaboratively using the supervised labels and pseudo labels predicted by the other. The history-conditioned model, referred to as ForeRanker, progressively learns future-relevant information to enhance ranking, while it singly uses historical session at inference time. To mitigate inconsistencies during training, we introduce the peer knowledge distillation method with a dynamic gating mechanism, allowing models to selectively incorporate contextual information. Experimental results on benchmark datasets demonstrate the effectiveness of our ForeRanker, showcasing its superior performance compared to existing methods.

Modern recommendation platforms frequently encompass multiple domains to cater to the varied preferences of users. Recently, cross-domain learning has gained traction as a significant paradigm within the context of recommendation systems, enabling the leveraging of rich information from a well-endowed source domain to enhance a target domain, often limited by inadequate data resources. A primary concern in cross-domain recommendation is the mitigation of negative transfer-ensuring the selective transference of pertinent knowledge from the source (domain-shared knowledge) while maintaining the integrity of domain-unique insights within the target domain (domain-specific knowledge).

Retrieval Augmented Generation (RAG) has become prevalent in question-answering (QA) tasks due to its ability of utilizing search engine to enhance the quality of long-form question-answering (LFQA). Despite the emergence of various open source methods and web-enhanced commercial systems such as Bing Chat, two critical problems remain unsolved, i.e., the lack of factuality and clear logic in the generated long-form answers. In this paper, we remedy these issues via a systematic study on answer generation in web-enhanced LFQA. Specifically, we first propose a novel outline-enhanced generator to achieve clear logic in the generation of multifaceted answers and construct two datasets accordingly. Then we propose a factuality optimization method based on a carefully designed doubly fine-grained RLHF framework, which contains automatic evaluation and reward modeling in different levels of granularity. Our generic framework comprises conventional fine-grained RLHF methods as special cases. Extensive experiments verify the superiority of our proposed Factuality-optimized RAG (FoRAG) method on both English and Chinese benchmarks. In particular, when applying our method to Llama2-7B-chat, the derived model FoRAG-L-7B outperforms WebGPT-175B in terms of three commonly used metrics (i.e., coherence, helpfulness, and factuality), while the number of parameters is much smaller (only 1/24 of that of WebGPT-175B). Our datasets and models are made publicly available for better reproducibility.https://huggingface.co/forag łabelfootnote_dataset_url

Recent advancements in Large Language Models (LLMs) and Large Multi-modal Models (LMMs) have shown potential in various medical applications, such as Intelligent Medical Diagnosis. Although impressive results have been achieved, we find that existing benchmarks do not reflect the complexity of real medical reports and specialized in-depth reasoning capabilities. In this work, we establish a comprehensive benchmark in the field of medical specialization and introduced RJUA-MedDQA, which contains 2000 real-world Chinese medical report images poses several challenges: comprehensively interpreting imgage content across a wide variety of challenging layouts, possessing the numerical reasoning ability to identify abnormal indicators and demonstrating robust clinical reasoning ability to provide the statement of disease diagnosis, status and advice based on a collection of medical contexts. We carefully design the data generation pipeline and proposed the Efficient Structural Restoration Annotation (ESRA) Method, aimed at restoring textual and tabular content in medical report images. This method substantially enhances annotation efficiency, doubling the productivity of each annotator, and yields a 26.8% improvement in accuracy. We conduct extensive evaluations, including few-shot assessments of 5 LMMs which are capable of solving Chinese medical QA tasks. To further investigate the limitations and potential of current LMMs, we conduct comparative experiments on a set of strong LLMs by using image-text generated by ESRA method. We report the performance of baselines and offer several observations: (1) The overall performance of existing LMMs is still limited; however LMMs more robust to low-quality and diverse-structured images compared to LLMs. (3) Reasoning across context and image content present significant challenges. We hope this benchmark helps the community make progress on these challenging tasks in multi-modal medical document understanding and facilitate its application in healthcare. Our dataset will be publicly available for noncommercial use at https://github.com/Alipay-Med/medDQA_benchmark.git

Large language models (LLMs) are gaining increasing interests to improve clinical efficiency, owing to their unprecedented performance in modelling natural language.Ensuring the reliable clinical applications, the evaluation of LLMs indeed becomes critical for better mitigating the potential risks, e.g., hallucinations.However, current evaluation methods heavily rely on labor-intensive human participation to achieve human-preferred judgements.To overcome this challenge, we propose an automatic evaluation paradigm tailored to assess the LLMs' capabilities in delivering clinical services, e.g., disease diagnosis and treatment.The evaluation paradigm contains three basic elements: metric, data, and algorithm.Specifically, inspired by professional clinical practice pathways, we formulate a LLM-specific clinical pathway (LCP) to define the clinical capabilities that a doctor agent should possess.Then, Standardized Patients (SPs) from the medical education are introduced as the guideline for collecting medical data for evaluation, which can well ensure the completeness of the evaluation procedure.Leveraging these steps, * Equal Contribution.

In this paper, we explore a new way for user targeting, where non-expert marketers could select their target users solely given demands in natural language form. The key to this issue is how to transform natural languages into practical structured logical languages, i.e., the structured understanding of marketer demands. In practical scenarios, the demands of non-expert marketers are often abstract and diverse. Considering the impressive natural language processing ability of large language models (LLMs), we try to leverage LLMs to solve this issue. To stimulate the LLMs' reasoning ability, the chain-of-thought (CoT) prompting method is widely used, but existing methods still have some limitations in our scenario: (1) Previous methods either use simple "Let's think step by step" spells or provide fixed examples in demonstrations without considering compatibility between prompts and concrete questions, making LLMs ineffective when the marketers' demands are abstract and diverse. (2) Previous methods are often implemented in closed-source models or excessively large models, which is not suitable in industrial practical scenarios. Based on these, we propose ARALLM (i.e., Analogical Reasoning Augmented Large Language Models) consisting of two modules: Analogical Reasoning based Prompting and Reasoning-Augmented Multi-Task Model Distillation. Then, we adopt a retrieval-based method to conduct analogical reasoning with the help of the reasoning library. The experimental results show that this prompting strategy achieves better performance than the ordinary prompting method. Beyond that, we distill knowledge from super LLMs (GPT-3.5) to fine-tune smaller student LLMs in a multi-task training paradigm, enabling the models to be easily deployed in practical environments. Part of our data and code can be found at https://github.com/alipay/Analogic-Reasoning-Augmented-Large-Language-Model.

As Large Language Models (LLMs) have made significant advancements across various tasks, such as question answering, translation, text summarization, and dialogue systems, the need for accuracy in information becomes crucial, especially for serious financial products serving billions of users like Alipay. However, for a real-world product serving millions of users, the inference speed of LLMs becomes a critical factor compared to a mere experimental model.

In the era of data proliferation, efficiently sifting through vast information to extract meaningful insights has become increasingly crucial. This paper addresses the computational overhead and resource inefficiency prevalent in existing Sequential Recommender Systems (SRSs). We introduce an innovative approach combining pruning methods with advanced model designs. Furthermore, we delve into resource-constrained Neural Architecture Search (NAS), an emerging technique in recommender systems, to optimize models in terms of FLOPs, latency, and energy consumption while maintaining or enhancing accuracy. Our principal contribution is the development of a Data-aware Neural Architecture Search for Recommender System (DNS-Rec). DNS-Rec is specifically designed to tailor compact network architectures for attention-based SRS models, thereby ensuring accuracy retention. It incorporates data-aware gates to enhance the performance of the recommendation network by learning information from historical user-item interactions. Moreover, DNS-Rec employs a dynamic resource constraint strategy, stabilizing the search process and yielding more suitable architectural solutions. We demonstrate the effectiveness of our approach through rigorous experiments conducted on three benchmark datasets, which highlight the superiority of DNS-Rec in SRSs. Our findings set a new standard for future research in efficient and accurate recommendation systems, marking a significant step forward in this rapidly evolving field.

Click-through rate (CTR) prediction plays a crucial role in recommendation systems, with significant impact on user experience and platform revenue generation. Despite the various public CTR datasets available due to increasing interest from both academia and industry, these datasets have limitations. They cover a limited range of scenarios and predominantly focus on ID-based features, neglecting the vital role of multi-modal features for effective multi-scenario CTR prediction. Moreover, their scale is modest compared to real-world industrial datasets, hindering robust and comprehensive evaluation of complex models. To address these challenges, we introduce a large-scale Multi-Scenario Multi-Modal CTR dataset named AntM2 C, built from real industrial data from Alipay. This dataset offers an impressive breadth and depth of information, covering CTR data from four diverse business scenarios, including advertisements, consumer coupons, mini-programs, and videos. Unlike existing datasets, AntM2 C provides not only ID-based features but also five textual features and one image feature for both users and items, supporting more delicate multi-modal CTR prediction. AntM2 C is also substantially larger than existing datasets, comprising 100 million CTR data. This scale allows for robust and comprehensive evaluation and comparison of CTR prediction models. We employ AntM2 C to construct several typical CTR tasks, including multi-scenario modeling, item and user cold-start modeling, and multi-modal modeling. Initial experiments and comparisons with baseline methods have shown that AntM2 C presents both new challenges and opportunities for CTR models, with the potential to significantly advance CTR research. The AntM2 C dataset is available at https://www.atecup.cn/OfficalDataSet.

With the growing popularity of various mobile devices, user targeting has received a growing amount of attention, which aims at effectively and efficiently locating target users that are interested in specific services. Most pioneering works for user targeting tasks commonly perform similarity-based expansion with a few active users as seeds, suffering from the following major issues: the unavailability of seed users for new-coming services and the unfriendliness of black-box procedures towards marketers. In this paper, we design an Entity Graph Learning (EGL) system to provide explainable user targeting ability meanwhile applicable to addressing the cold-start issue. EGL System follows the hybrid online-offline architecture to satisfy the requirements of scalability and timeliness. Specifically, in the offline stage, the system focuses on the heavyweight entity graph construction and user entity preference learning, in which we propose a Three-stage Relation Mining Procedure (TRMP), breaking loose from the expensive seed users. At the online stage, the system offers the ability of user targeting in real-time based on the entity graph from the offline stage. Since the user targeting process is based on graph reasoning, the whole process is transparent and operation-friendly to marketers. Finally, extensive offline experiments and online A/B testing demonstrate the superior performance of the proposed EGL System.

Given the enormous number of users and items, industrial cascade recommendation systems (RS) are continuously expanded in size and complexity to deliver relevant items, such as news, services, and commodities, to the appropriate users. In a real-world scenario with hundreds of thousands requests per second, significant computation is required to infer personalized results for each request, resulting in a massive energy consumption and carbon emission that raises concern. This paper proposes GreenFlow, a practical computation allocation framework for RS, that considers both accuracy and carbon emission during inference. For each stage (e.g., recall, pre-ranking, ranking, etc.) of a cascade RS, when a user triggers a request, we define two actions that determine the computation: (1) the trained instances of models with different computational complexity; and (2) the number of items to be inferred in the stage. We refer to the combinations of actions in all stages as action chains. A reward score is estimated for each action chain, followed by dynamic primal-dual optimization considering both the reward and computation budget. Extensive experiments verify the effectiveness of the framework, reducing computation consumption by 41% in an industrial mobile application while maintaining commercial revenue. Moreover, the proposed framework saves approximately 5000kWh of electricity and reduces 3 tons of carbon emissions per day.

Model-free RL-based recommender systems have recently received increasing research attention due to their capability to handle partial feedback and long-term rewards. However, most existing research has ignored a critical feature in recommender systems: one user's feedback on the same item at different times is random. The stochastic rewards property essentially differs from that in classic RL scenarios with deterministic rewards, which makes RL-based recommender systems much more challenging. In this paper, we first demonstrate in a simulator environment where using direct stochastic feedback results in a significant drop in performance. Then to handle the stochastic feedback more efficiently, we design two stochastic reward stabilization frameworks that replace the direct stochastic feedback with that learned by a supervised model. Both frameworks are model-agnostic, i.e., they can effectively utilize various supervised models. We demonstrate the superiority of the proposed frameworks over different RL-based recommendation baselines with extensive experiments on a recommendation simulator as well as an industrial-level recommender system.

In marketing recommendations, the campaign organizers will distribute coupons to users to encourage consumption. In general, a series of strategies are employed to interfere with the coupon distribution process, leading to a growing imbalance between user-coupon interactions, resulting in a bias in the estimation of conversion probabilities. We refer to the estimation bias as the matching bias. In this paper, we explore how to alleviate the matching bias from the causal-effect perspective. We regard the historical distributions of users and coupons over each other as confounders and characterize the matching bias as a confounding effect to reveal and eliminate the spurious correlations between user-coupon representations and conversion probabilities. Then we propose a new training paradigm named De-Matching Bias Recommendation (DMBR) to remove the confounding effects during model training via the backdoor adjustment. We instantiate DMBR on two representative models: DNN and MMOE, and conduct extensive offline and online experiments to demonstrate the effectiveness of our proposed paradigm.

Transformer-based models have achieved tremendous success in sequential recommendation (SR), but they suffer from consuming excessive computational resources, particularly in the inference stage. Thus, developing lightweight yet effective SR models has become a frequent demand in industrial applications, which is also in line with the ideals of Green AI and Green IR. In this applied paper, we introduce GreenSeq deployed in Alipay to automatically design Green networks that can provide appropriate recommendations with lower computational consumption in SR. Specifically, GreenSeq uses a novel multi-layer search space that allows for flexible network design and a Greenness-aware loss term for balancing efficiency and effectiveness. Experiments on benchmark datasets and A/B testing show that GreenSeq performs well while using fewer resources. GreenSeq also reduces electricity and carbon emissions in Alipay.

As web applications continue to expand and diversify their services, user interactions exist in different scenarios. To leverage this wealth of information, cross-domain recommendation (CDR) has gained significant attention in recent years. However, existing CDR approaches mostly focus on information transfer between observed domains, with little attention paid to generalizing to unseen domains. Although recent research on invariant learning can help for the purpose of generalization, relying only on invariant preference may be overly conservative and result in mediocre performance when the unseen domain shifts slightly. In this paper, we present a novel framework that considers both CDR and domain generalization through a united causal invariant view. We assume that user interactions are determined by domain-invariant preference and domain-specific preference. The proposed approach differentiates the invariant preference and the specific preference from observational behaviors in a way of adversarial learning. Additionally, a novel domain routing module is designed to connect unseen domains to observed domains. Extensive experiments on public and industry datasets have proved the effectiveness of the proposed approach under both CDR and domain generalization settings.

We study the budget allocation problem in online marketing campaigns that utilize previously collected offline data. We first discuss the long-term effect of optimizing marketing budget allocation decisions in the offline setting. To overcome the challenge, we propose a novel game-theoretic offline value-based reinforcement learning method using mixed policies. The proposed method reduces the need to store infinitely many policies in previous methods to only constantly many policies, which achieves nearly optimal policy efficiency, making it practical and favorable for industrial usage. We further show that this method is guaranteed to converge to the optimal policy, which cannot be achieved by previous value-based reinforcement learning methods for marketing budget allocation. Our experiments on a large-scale marketing campaign with tens-of-millions users and more than one billion budget verify the theoretical results and show that the proposed method outperforms various baseline methods. The proposed method has been successfully deployed to serve all the traffic of this marketing campaign.

Uplift modeling aims to model the incremental impact of a treatment on an individual outcome, which has attracted great interests of researchers and practitioners from different communities. Existing uplift modeling methods rely on either the data collected from randomized controlled trials (RCTs) or the observational data which is more realistic. However, we notice that on the observational data, it is often the case that only a small number of subjects receive treatment, but finally infer the uplift on a much large group of subjects. Such highly imbalanced data is common in various fields such as marketing and medical treatment but it is rarely handled by existing works. In this paper, we theoretically and quantitatively prove that the existing representative methods, transformed outcome (TOM) and doubly robust (DR), suffer from large bias and deviation on highly imbalanced datasets with skewed propensity scores, mainly because they are proportional to the reciprocal of the propensity score. To reduce the bias and deviation of uplift modeling with an imbalanced dataset, we propose an imbalance-aware uplift modeling (IAUM) method via constructing a robust proxy outcome, which adaptively combines the doubly robust estimator and the imputed treatment effects based on the propensity score. We theoretically prove that IAUM can obtain a better bias-variance trade-off than existing methods on a highly imbalanced dataset. We conduct extensive experiments on a synthetic dataset and two real-world datasets, and the experimental results well demonstrate the superiority of our method over state-of-the-art.

Many recommender systems optimize click through rates (CTRs) as one of their core goals, and it further breaks down to predicting each item's click probability for a user (user-item click probability) and recommending the top ones to this particular user. User-item click probability is then estimated as a single term, and the basic assumption is that the user has different preferences over items. This is presumably true, but from real-world data, we observe that some people are naturally more active in clicking on items while some are not. This intrinsic tendency contributes to their user-item click probabilities. Besides this, when a user sees a particular item she likes, the click probability for this item increases due to this user-item preference.

Modeling users' historical behaviors is an essential task in many industrial recommender systems. The user interest representation, in previous works, is obtained through the following paradigm: concrete behaviors are firstly embedded as low-dimensional behavior representations, which are then aggregated conditioning on the target item for final user interest representation. Most existing researches focus on the aggregation process that explores the intrinsic structure of the behavior sequences. However, the quality of behavior representation is largely ignored. In this paper, we present a pluggable module, FwSeqBlock, to enhance the expressiveness of behavior representations. Specifically, FwSeqBlock introduces the multiplicative operation among users' historical behaviors and the target item, where a field memory unit is designed to dynamically identify the dominant features from the behavior sequence and filter out the noise. Extensive experiments validate that FwSeqBlock consistently generates higher-quality user representations compared with competitive methods. Besides, online A/B testing reports a 4.46% improvement in Click-Through Rate (CTR), confirming the effectiveness of the proposed method.

Uplift modeling is widely used in performance marketing to estimate effects of promotion campaigns (e.g., increase of customer retention rate). Since it is impossible to observe outcomes of a recipient in treatment (e.g., receiving a certain promotion) and control (e.g., without promotion) groups simultaneously (i.e., counter-factual), uplift models are mainly trained on instances of treatment and control groups separately to form two models respectively, and uplifts are predicted by the difference of predictions from these two models (i.e., two-model method). When responses are noisy and the treatment effect is fractional, induced individual uplift predictions will be inaccurate, resulting in targeting undesirable customers. Though it is impossible to obtain the ideal ground-truth individual uplifts, known as Individual Treatment Effects (ITEs), alternatively, an average uplift of a group of users, called Average Treatment Effect (ATE), can be observed from experimental deliveries. Upon this, similar to Multiple Instance Learning (MIL) in which each training sample is a bag of instances, our framework sums up individual user uplift predictions for each bag of users as its bag-wise ATE prediction, and regularizes it to its ATE label, thus learning more accurate individual uplifts. Additionally, to amplify the fractional treatment effect, bags are composed of instances with adjacent individual uplift predictions, instead of random instances. Experiments conducted on two datasets show the effectiveness and universality of the proposed framework.

In this paper, we study the digital marketing where marketing officers (MOs) have to commit to creating brand new promotion ads/contents based on understandings of users' needs or preferences. Users' behaviors are typically high dimensional and hard to understand. Therefore, dimension reduction of users' behaviors from high dimensions and explainability are important to help MOs launch operation-friendly marketings. As such, it is natural to exploit topic models to help MOs understand users' intents from users' behaviors (e.g., user-item visits) in case we treat each user as a document and users' behaviors of visiting an item as a word. However, users of low activities and items followed by power law distributions are common in user-item visit data, which pose significant challenges to traditional topic models. We present a social and semantic enhanced topic model (S2TM) for users' intent mining. We optimize the user-intent estimates based on a graph neural network atop of a social network, and optimize the intent-item estimates based on a skip-gram word embedding approach by linking the semantics of items to pre-trained word embeddings. We propose an efficient stochastic vari-ational inference algorithm for the inference of latent variables and learning of parameters. Extensive experiments on real-world data show the effectivenesses of S2TM in terms of perplexities, topic coherence and semantic coherence compared with state-of-the-art topic models. We further show how MOs interact with our operation-friendly intent mining system, and results on real-world marketing campaigns in terms of click-through rate at Alipay.

Modeling sequential data is essential to many applications such as natural language processing, recommendation systems, time series predictions, anomaly detection, etc. When processing sequential data, one of the critical issues is how to capture the temporal-correlation among events. Though prevalent and effective in many applications, conventional approaches such as RNNs and Transformers, struggle with handling the non-stationary characteristics (i.e., such temporal-correlation among events would change over time), which is indeed encountered in many real-world scenarios. In this paper, we present a non-stationary time-aware kernelized attention approach for input sequences of neural networks. By constructing the Generalized Spectral Mixture Kernel (GSMK), and integrating it to the attention mechanism, we mathematically reveal its representation capability in terms of the time-dependent temporal-correlation. Following that, a novel neural network structure is proposed, which would enable us to encode both stationary and non-stationary time event series. Finally, we demonstrate the performance of the proposed method on both synthetic data which presents the theoretical insights, and a variety of real-world datasets which shows its competitive performance against related work.

\emph{Classification with rejection} (CwR) refrains from making a prediction to avoid critical misclassification when encountering test samples that are difficult to classify. Though previous methods for CwR have been provided with theoretical guarantees, they are only compatible with certain loss functions, making them not flexible enough when the loss needs to be changed with the dataset in practice. In this paper, we derive a novel formulation for CwR that can be equipped with arbitrary loss functions while maintaining the theoretical guarantees. First, we show that $K$-class CwR is equivalent to a $(K\!+\!1)$-class classification problem on the original data distribution with an augmented class, and propose an empirical risk minimization formulation to solve this problem with an estimation error bound. Then, we find necessary and sufficient conditions for the learning \emph{consistency} of the surrogates constructed on our proposed formulation equipped with any classification-calibrated multi-class losses, where consistency means the surrogate risk minimization implies the target risk minimization for CwR. Finally, experiments on benchmark datasets validate the effectiveness of our proposed method.

The conflicting gradients problem is one of the major bottlenecks for the effective training of machine learning models that deal with multiple objectives. To resolve this problem, various gradient manipulation techniques, such as PCGrad, MGDA, and CAGrad, have been developed, which directly alter the conflicting gradients to refined ones with alleviated or even no conflicts. However, the existing design and analysis of these techniques are mainly conducted under the full-batch gradient setting, ignoring the fact that they are primarily applied with stochastic mini-batch gradients. In this paper, we illustrate that the stochastic gradient manipulation algorithms may fail to converge to Pareto optimal solutions. Firstly, we show that these different algorithms can be summarized into a unified algorithmic framework, where the descent direction is given by the composition of the gradients of the multiple objectives. Then we provide an explicit two-objective convex optimization instance to explicate the non-convergence issue under the unified framework, which suggests that the non-convergence results from the determination of the composite weights solely by the instantaneous stochastic gradients. To fix the non-convergence issue, we propose a novel composite weights determination scheme that exponentially averages the past calculated weights. Finally, we show the resulting new variant of stochastic gradient manipulation converges to Pareto optimal or critical solutions and yield comparable or improved empirical performance.

There exists the cold-start problem in the recommendation systems when observed user-item interactions are insufficient. To alleviate this problem, most existing works aim to learn globally shared prior knowledge across all items and be fast adapted to a new item with few interactions. However, such learning techniques are data demanding and work poorly on new items with no interactions. In this applied paper, we present an industrial framework recently deployed on Alipay to address the item cold-start problem in zero-shot scenarios. The proposed framework provides both efficient and high-quality recommendations for cold items with no log data. Specifically, we formulate the cold-start problem as a zero-shot learning problem and build a highly efficient infrastructure to accomplish online zero-shot recommendations used on large-scale platforms. Extensive offline experiments and online A/B testing demonstrate that the proposed framework has superior performance and recommends cold items to preferred users more effectively than other state-of-the-art methods.

Modern recommendation systems introduce the re-ranking stage to optimize the entire list directly. This paper focuses on the design of re-ranking framework in feed to optimally model the mutual influence between items and further promote user engagement. On mobile devices, users browse the feed almost in a top-down manner and rarely compare items back and forth. Besides, users often compare item with its adjacency based on their partial observations. Given the distinct user behavior patterns, the modeling of mutual influence between items should be carefully designed. Existing re-ranking models encode the mutual influence between items with sequential encoding methods. However, previous works may be dissatisfactory due to the ignorance of connections between items on different scopes. In this paper, we first discuss Unidirectivity and Locality on the impacts and consequences, then report corresponding solutions in industrial applications. We propose a novel framework based on the empirical evidence from user analysis. To address the above problems, we design a \underlineS cope-aware \underlineR e-ranking with \underlineG ated \underlineA ttention model (SRGA ) to emulate the user behavior patterns from two aspects: 1) we emphasize the influence along the user's common browsing direction; 2) we strength the impacts of pivotal adjacent items within the user visual window. Specifically, we design a global scope attention to encode inter-item patterns unidirectionally from top to bottom. Besides, we devise a local scope attention sliding over the recommendation list to underline interactions among neighboring items. Furthermore, we design a learned gate mechanism to aggregating the information dynamically from local and global scope attention. Extensive offline experiments and online A/B testing demonstrate the benefits of our novel framework. The proposed SRGA model achieves the best performance in offline metrics compared with the state-of-the-art re-ranking methods. Further, empirical results on live traffic validate that our recommender system, equipped with SRGA in the re-ranking stage, improves significantly in user engagement.

In the mobile Internet era, mobile payment service becomes the foundation of inclusive finance, which brings convenience and security to people. Various marketing strategies are designed to encourage mobile payment activities by allocating incentives such as coupons or commissions to customers or merchants. We summary two significant issues. First, there is a phenomenon of mutual influence between merchants and customers, i.e., bipartite influence issue, thus making the independent optimization of customers and merchants non-optimal. Second, the redemptions of coupons are partially observed, as we can only observe that the customer redeems the coupon or not at a specific incentive value, but cannot observe that at other incentive value, i.e., data censorship issue. In this paper, we propose a novel joint incentive optimization framework to address the above two issues. We propose to use a graph neural network to represent customers and merchants jointly by modeling the underlying bipartite influences. We then formulate the response model under the hazard regression setting and model the hazard rate with a piecewise nonlinear function to capture the changes of responses to different incentive values. Finally, we propose a linear programming method to allocate approximated optimal incentive values to customers and merchants in real-time. Extensive offline and online experimental results demonstrate the effectiveness of our proposed approach.

Many payment platforms hold large-scale marketing campaigns, which allocate incentives to encourage users to pay through their applications. To maximize the return on investment, incentive allocations are commonly solved in a two-stage procedure. After training a response estimation model to estimate the users' mobile payment probabilities (MPP), a linear programming process is applied to obtain the optimal incentive allocation. However, the large amount of biased data in the training set, generated by the previous biased allocation policy, causes a biased estimation. This bias deteriorates the performance of the response model and misleads the linear programming process, dramatically degrading the performance of the resulting allocation policy. To overcome this obstacle, we propose a bias correction adversarial network. Our method leverages the small set of unbiased data obtained under a full-randomized allocation policy to train an unbiased model and then uses it to reduce the bias with adversarial learning. Offline and online experimental results demonstrate that our method outperforms state-of-the-art approaches and significantly improves the performance of the resulting allocation policy in a real-world marketing campaign.

Understanding inactive users is the key to user growth and engagement for many Internet companies. However, learning inactive users' representations and their preferences is still challenging because the features available are missing and the positive responses or labels are insufficient. In this paper, we propose a cross domain learning approach to exclusively recommend customized items to inactive users by leveraging the knowledge of active users. Particularly, we represent users, no matter active or inactive users, by their friends' browsing behaviors based on a graph neural network (GNN) layer atop of a heterogeneous graph defined on social networks (user-user friendships) and browsing behaviors (user-page clicks). We jointly optimize the learning tasks of active users in source domain and inactive users in target domain based on the domain invariant features extracted from the embedding of our GNN layer, where the domain invariant features that are learned to benefit both tasks on active/inactive users, and are indiscriminate with respect to the shift between the domains. Extensive experiments show that our approach can well capture the preference of inactive users using both public data and real-world data at Alipay.

For many Internet companies, it has been an important focus to improve user retention rate. To achieve this goal, we need to recommend proper services in order to meet the demands of users. Unlike conventional click-through rate (CTR) estimation, there are lots of noise in the collected data when modeling retention, caused by two major issues: 1) implicit impression-revisit effect: users could revisit the APP even if they do not explicitly interact with the recommender system; 2) selection bias: recommender system suffers from selection bias caused by user's self-selection. To address the above challenges, we propose a novel method named UR-IPW (User Retention Modeling with Inverse Propensity Weighting), which 1) makes full use of both explicit and implicit interactions in the observed data. 2) models revisit rate estimation from a causal perspective accounting for the selection bias problem. The experiments on both offline and online environments from different scenarios demonstrate the superiority of UR-IPW over previous methods. To the best of our knowledge, this is the first work to model user retention by estimating the revisit rate from a causal perspective.

Most existing algorithms in decentralized online learning are conducted in the synchronous setting. However, synchronization makes these algorithms suffer from the straggler problem, i.e., fast learners have to wait for slow learners, which significantly reduces such algorithms' overall efficiency. To overcome this problem, we study decentralized online learning in the asynchronous setting, which allows different learners to work at their own pace. We first formulate the framework of Asynchronous Decentralized Online Convex Optimization, which specifies the whole process of asynchronous decentralized online learning using a sophisticated event indexing system. Then we propose the Asynchronous Decentralized Online Gradient-Push (AD-OGP) algorithm, which performs asymmetric gossiping communication and instantaneous model averaging. We further derive a regret bound of AD-OGP, which is a function of the network topology, the levels of processing delays, and the levels of communication delays. Extensive experiments show that AD-OGP runs significantly faster than its synchronous counterpart and also verify the theoretical results.

Encoding a large-scale network into a low-dimensional space is a fundamental step for various network analytic problems, such as node classification, link prediction, community detection, etc. Existing methods focus on learning the network representation from either the static graphs or time-aggregated graphs (e.g., time-evolving graphs). However, many real systems are not static or time-aggregated as the nodes and edges are timestamped and dynamically changing over time. For examples, in anti-money laundering analysis, cycles formed with time-ordered transactions might be red flags in online transaction networks; in novelty detection, a star-shaped structure appearing in a short burst might be an underlying hot topic in social networks. Existing embedding models might not be able to well preserve such fine-grained network dynamics due to the incapability of dealing with continuous-time and the negligence of fine-grained interactions. To bridge this gap, in this paper, we propose a fine-grained temporal network embedding framework named FiGTNE, which aims to learn a comprehensive network representation that preserves the rich and complex network context in the temporal network. In particular, we start from the notion of fine-grained temporal networks, where the temporal network can be represented as a series of timestamped nodes and edges. Then, we propose the time-reinforced random walk (TRRW) with a bi-level context sampling strategy to explore the essential structures and temporal contexts in temporal networks. Extensive experimental results on real graphs demonstrate the efficacy of our FiGTNE framework.

With the revolution of mobile internet, online finance has grown explosively. In this new area, one challenge of significant importance is how to effectively deliver the financial products or services to a set of target users by marketing. Given a product or service to be promoted and a set of users as seeds, audience expansion is such a targeting technique, which aims to find potential audience among a large number of users. However, in the context of finance, financial products and services are dynamic in nature as they co-vary with the socio-economic environment. Moreover, marketing campaigns for promoting products or services always consist of different rules of play, even for the same type of products or services. As a result, there is a strong demand for the timeliness of seeds in financial targeting. Conventional one-stage audience expansion methods, which generate expanded users by expanding over seeds, would encounter two problems under this setting: (1) the seeds would inevitably involve a number of users that are not representative for expansion, and direct expansion over these noisy seeds would dramatically deteriorate the performance; (2) one-stage expansion over fixed seeds cannot timely and accurately capture users' preferences over the currently running campaign due to the lack of timeliness of seeds.

Recently, in order to take a preemptive opportunity in the mobile economy, the Internet companies conduct thousands of marketing campaigns every day, to promote their mobile products and services. In the mobile marketing scenario, one of the fundamental issues is the audience expansion task for marketing campaigns. Given a set of seed users, audience expansion aims to seek more users (audiences), who are similar to the seeds and will finish the business goal of the targeted campaign (ie convert). However, the problem is challenging in three aspects. First, a company will run hundreds of campaigns to serve massive users every day. The requirements of scalability and timeliness make training model for each campaign extremely resource-consuming thus impractical. Therefore, we proposed to solve the problem in a two-stage manner, in which the offline stage employs heavyweight user representation learning and the online stage performs embedding-based lightweight audience expansion. Second, conventional two-stage audience expansion systems neglect the high-order user-campaign interactions and usually generate entangled user embeddings, thus fail to achieve high-quality user representation. Third, the seeds, which are usually provided by experts or collected from users' feedbacks, could be noisy and cannot cover the entire actual audiences, thus introduce coverage bias. Unfortunately, to our best knowledge, none of the related literatures tackle this crucial issue of audience expansion.

Mobile payment such as Alipay has been widely used in our daily lives. To further promote the mobile payment activities, it is important to run marketing campaigns under a limited budget by providing incentives such as coupons, commissions to merchants. As a result, incentive optimization is the key to maximizing the commercial objective of the marketing campaign. With the analyses of online experiments, we found that the transaction network can subtly describe the similarity of merchants' responses to different incentives, which is of great use in the incentive optimization problem. In this paper, we present a graph representation learning method atop of transaction networks for merchant incentive optimization in mobile payment marketing. With limited samples collected from online experiments, our end-to-end method first learns merchant representations based on an attributed transaction networks, then effectively models the correlations between the commercial objectives each merchant may achieve and the incentives under varying treatments. Thus we are able to model the sensitivity to incentive for each merchant, and spend the most budgets on those merchants that show strong sensitivities in the marketing campaign. Extensive offline and online experimental results at Alipay demonstrate the effectiveness of our proposed approach.