Xiusheng Huang, Zhe Li, Xuanwu Yin, Lu Wang, Yequan Wang, Dong Li, Emad Barsoum, Kang Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhiyuan Peng, Xin Yin, Pu Zhao, Fangkai Yang, Lu Wang, Ran Jia, Xu Chen, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Huawen Feng, Pu Zhao, Qingfeng Sun, Can Xu, Fangkai Yang, Lu Wang, Qianli Ma, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Qi Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Junting Lu, Zhiyang Zhang, Fangkai Yang, Jue Zhang, Lu Wang, Chao Du, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Qi Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Kaikai An, Fangkai Yang, Liqun Li, Junting Lu, Sitao Cheng, Shuzheng Si, Lu Wang, Pu Zhao, Lele Cao, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Baobao Chang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Yichen Ouyang, Lu Wang, Fangkai Yang, Pu Zhao, Chenghua Huang, Jianfeng Liu, Bochen Pang, Yaming Yang, Yuefeng Zhan, Hao Sun, Qingwei Lin, Saravan Rajmohan, Weiwei Deng, Dongmei Zhang, Feng Sun. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In-context learning (ICL) and Retrieval-Augmented Generation (RAG) have gained attention for their ability to enhance LLMs' reasoning by incorporating external knowledge but suffer from limited contextual window size, leading to insufficient information injection. To this end, we propose a novel framework to automatically distill large volumes of offline data into interpretable first-order logic rules, which are injected into LLMs to boost their reasoning capabilities. Our method begins by formulating the search process relying on LLMs' commonsense, where LLMs automatically define head and body predicates. Then, we apply Monte Carlo Tree Search (MCTS) to address the combinational searching space and efficiently discover logic rules from data. The resulting logic rules are translated into natural language, allowing targeted knowledge injection and seamless integration into LLM prompts for LLM's downstream task reasoning. We evaluate our framework on public and private industrial tasks, including Natural Language Processing (NLP), time-series, decision-making, and industrial tasks, demonstrating its effectiveness in enhancing LLM's capability over diverse tasks.

Reinforcement Learning from Human Feedback (RLHF) is a crucial technique for aligning language models with human preferences and is a key factor in the success of modern conversational models like GPT-4, ChatGPT, and Llama 2. A significant challenge in employing RLHF lies in training a reliable RM, which relies on high-quality labels. Typically, these labels are provided by human experts or a stronger AI, both of which can be costly and introduce bias that may affect the language model's responses. As models improve, human input may become less effective in enhancing their performance. This paper explores the potential of using the RM itself to generate additional training data for a more robust RM. Our experiments demonstrate that reinforcement learning from self-feedback outperforms baseline approaches. We conducted extensive experiments with our approach on multiple datasets, such as HH-RLHF and UltraFeedback, and models including Mistral and Llama 3, comparing it against various baselines. Our results indicate that, even with a limited amount of human-labeled data, learning from self-feedback can robustly enhance the performance of the RM, thereby improving the capabilities of large language models.

User profiling is pivotal for recommendation systems, as it transforms raw user interaction data into concise and structured representations that drive personalized recommendations. While traditional embedding-based profiles lack interpretability and adaptability, recent advances with large language models (LLMs) enable text-based profiles that are semantically richer and more transparent. However, existing methods often adhere to fixed formats that limit their ability to capture the full diversity of user behaviors. In this paper, we introduce LettinGo, a novel framework for generating diverse and adaptive user profiles. By leveraging the expressive power of LLMs and incorporating direct feedback from downstream recommendation tasks, our approach avoids the rigid constraints imposed by supervised fine-tuning (SFT). Instead, we employ Direct Preference Optimization (DPO) to align the profile generator with task-specific performance, ensuring that the profiles remain adaptive and effective. LettinGo operates in three stages: (1) exploring diverse user profiles via multiple LLMs(2) evaluating profile quality based on their impact in recommendation systems, and (3) aligning the profile generation through pairwise preference data derived from task performance. Experimental results demonstrate that our framework significantly enhances recommendation accuracy, flexibility, and contextual awareness. This work enhances profile generation as a key innovation for next-generation recommendation systems.

In numerous reinforcement learning (RL) problems involving safety-critical systems, a key challenge lies in balancing multiple objectives while simultaneously meeting all stringent safety constraints. To tackle this issue, we propose a primal-based framework that orchestrates policy optimization between multi-objective learning and constraint adherence. Our method employs a novel natural policy gradient manipulation method to optimize multiple RL objectives and overcome conflicting gradients between different objectives, since the simple weighted average gradient direction may not be beneficial for specific objectives due to misaligned gradients of different objectives. When there is a violation of a hard constraint, our algorithm steps in to rectify the policy to minimize this violation. Particularly, We establish theoretical convergence and constraint violation guarantees, and our proposed method also outperforms prior state-of-the-art methods on challenging safe multi-objective RL tasks.

Knowledge Tracing (KT) is a crucial task in computer-aided education and intelligent tutoring systems, predicting students’ performance on new questions from their responses to prior ones. An accurate KT model can capture a student’s mastery level of different knowledge topics, as reflected in their predicted performance on different questions. This helps improve the learning efficiency by suggesting appropriate new questions that complement students’ knowledge states. However, current KT models have significant drawbacks that they neglect the imbalanced discrimination of historical responses. A significant proportion of question responses provide limited information for discerning students’ knowledge mastery, such as those that demonstrate uniform performance across different students. Optimizing the prediction of these cases may increase overall KT accuracy, but also negatively impact the model’s ability to trace personalized knowledge states, especially causing a deceptive surge of performance. Towards this end, we propose a framework to reweight the contribution of different responses based on their discrimination in training. Additionally, we introduce an adaptive predictive score fusion technique to maintain accuracy on less discriminative responses, achieving proper balance between student knowledge mastery and question difficulty. Experimental results demonstrate that our framework enhances the performance of three mainstream KT methods on three widely used datasets.

Ensuring the safety of Reinforcement Learning (RL) is crucial for its deployment in real-world applications. Nevertheless, managing the trade-off between reward and safety during exploration presents a significant challenge. Improving reward performance through policy adjustments may adversely affect safety performance. In this study, we aim to address this conflicting relation by leveraging the theory of gradient manipulation. Initially, we analyze the conflict between reward and safety gradients. Subsequently, we tackle the balance between reward and safety optimization by proposing a soft switching policy optimization method, for which we provide convergence analysis. Based on our theoretical examination, we provide a safe RL framework to overcome the aforementioned challenge, and we develop a Safety-MuJoCo Benchmark to assess the performance of safe RL algorithms. Finally, we evaluate the effectiveness of our method on the Safety-MuJoCo Benchmark and a popular safe benchmark, Omnisafe. Experimental results demonstrate that our algorithms outperform several state-of-the-art baselines in terms of balancing reward and safety optimization.

We address the real problem of safe, robust, adaptive resource oversubscription in uncertain environments with our proposed novel technique of chance-constrained imitation learning. Our objective is to enhance resource efficiency while ensuring safety against congestion risk. Traditional supervised or forecasting models are ineffective in learning adaptive oversubscription policies, and conventional online optimization or reinforcement learning is difficult to deploy on real systems. Offline policy learning methods, such as Imitation Learning (IL) can leverage historical resource utilization telemetry data to learn effective policies if we can ensure robustness and safety from the underlying uncertainty in the domain, and thus the data. Our work investigates the nature of this uncertainty, how it can be quantified and proposes a novel chance-constrained IL that implicitly models such uncertainty in a principled manner via additional knowledge in the form of stochastic constraints on the associated risk, to learn provably safe and robust policies. We show empirically a substantial improvement (~ 3-4×) in capacity efficiency and congestion safety in test as well as real deployments.

Effective incident management is pivotal for the smooth operation of Microsoft cloud services. In order to expedite incident mitigation, service teams gather troubleshooting knowledge into Troubleshooting Guides (TSGs) accessible to On-Call Engineers (OCEs). While automated pipelines are enabled to resolve the most frequent and easy incidents, there still exist complex incidents that require OCEs’ intervention. In addition, TSGs are often unstructured and incomplete, which requires manual interpretation by OCEs, leading to on-call fatigue and decreased productivity, especially among new-hire OCEs. In this work, we propose Nissist which leverages unstructured TSGs and incident mitigation history to provide proactive incident mitigation suggestions, reducing human intervention. Leveraging Large Language Models (LLM), Nissist extracts knowledge from unstructured TSGs and incident mitigation history, forming a comprehensive knowledge base. Its multi-agent system design enhances proficiency in precisely discerning OCE intents, retrieving relevant information, and delivering systematic plans consecutively. Through our user experiments, we demonstrate that Nissist significantly reduce Time to Mitigate (TTM) in incident mitigation, alleviating operational burdens on OCEs and improving service reliability. Our webpage is available at https://aka.ms/nissist.

Zezhong Wang, Fangkai Yang, Lu Wang, Pu Zhao, Hongru Wang, Liang Chen, Qingwei Lin, Kam-Fai Wong. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

The advent of abundant image data has catalyzed the advancement of visual control in reinforcement learning (RL) systems, leveraging multiple view- points to capture the same physical states, which could enhance control performance theoretically. However, integrating multi-view data into representation learning remains challenging. In this paper, we introduce SMuCo, an innovative multi-view reinforcement learning algorithm that constructs robust latent representations by optimizing multi- view sequential total correlation. This technique effectively captures task-relevant information and temporal dynamics while filtering out irrelevant data. Our method supports an unlimited number of views and demonstrates superior performance over leading model-free and model-based RL algorithms. Empirical results from the DeepMind Control Suite and the Sapien Basic Manipulation Task confirm SMuCo’s enhanced efficacy, significantly improving task performance across diverse scenarios and views.

Imitation learning, which learns agent policy by mimicking expert demonstration, has shown promising results in many applications such as medical treatment regimes and self-driving vehicles. However, it remains a difficult task to interpret control policies learned by the agent. Difficulties mainly come from two aspects: 1) agents in imitation learning are usually implemented as deep neural networks, which are black-box models and lack interpretability; 2) the latent causal mechanism behind agents' decisions may vary along the trajectory, rather than staying static throughout time steps. To increase transparency and offer better interpretability of the neural agent, we propose to expose its captured knowledge in the form of a directed acyclic causal graph, with nodes being action and state variables and edges denoting the causal relations behind predictions. Furthermore, we design this causal discovery process to be state-dependent, enabling it to model the dynamics in latent causal graphs. Concretely, we conduct causal discovery from the perspective of Granger causality and propose a self-explainable imitation learning framework, CAIL. The proposed framework is composed of three parts: a dynamic causal discovery module, a causality encoding module, and a prediction module, and is trained in an end-to-end manner. After the model is learned, we can obtain causal relations among states and action variables behind its decisions, exposing policies learned by it. Experimental results on both synthetic and real-world datasets demonstrate the effectiveness of the proposed CAIL in learning the dynamic causal graphs for understanding the decision-making of imitation learning meanwhile maintaining high prediction accuracy.

Cloud providers often have resources that are not being fully utilized, and they may offer them at a lower cost to make up for the reduced availability of these resources. However, customers may be hesitant to use such offerings (such as spot VMs) as making trade-offs between cost and resource availability is not always straightforward. In this work, we propose Snape (Spot On-demand Perfect Mixture), an intelligent framework to optimize the cost and resource availability by dynamically mixing on-demand VMs with spot VMs. Through a detailed characterization based on real production traces, we verify that the eviction of spot VMs is predictable to some extent. Snape also leverages constrained reinforcement learning to adjust the mixture policy online. Experiments across different configurations show that Snape achieves 44% savings compared to using only on-demand VMs while maintaining 99.96% availability, which is 2.77% higher than using only spot VMs.

Imitation learning that replicates experts' skills via their demonstrations has shown significant success in various decision-making tasks. However, two critical challenges still hinder the deployment of imitation learning techniques in real-world application scenarios. First, existing methods lack the intrinsic interpretability to explicitly explain the underlying rationale of the learned skill and thus making learned policy untrustworthy. Second, due to the scarcity of expert demonstrations from each end user (client), learning a policy based on different data silos is necessary but challenging in privacy-sensitive applications such as finance and healthcare. To this end, we present a privacy-preserved interpretable skill learning framework (FedSkill) that enables global policy learning to incorporate data from different sources and provides explainable interpretations to each local user without violating privacy and data sovereignty. Specifically, our proposed interpretable skill learning model can capture the varying patterns in the trajectories of expert demonstrations, and extract prototypical information as skills that provide implicit guidance for policy learning and explicit explanations in the reasoning process. Moreover, we design a novel aggregation mechanism coupled with the based skill learning model to preserve global information utilization and maintain local interpretability under the federated framework. Thoroughly experiments on three datasets and empirical studies demonstrate that our proposed FedSkill framework not only outperforms state-of-the-art imitation learning methods but also exhibits good interpretability under a federated setting. Our proposed FedSkill framework is the first attempt to bridge the gaps among federated learning, interpretable machine learning, and imitation learning.

In microservice systems, the identification of root causes of anomalies is imperative for service reliability and business impact. This process is typically divided into two phases: (i)constructing a service dependency graph that outlines the sequence and structure of system components that are invoked, and (ii) localizing the root cause components using the graph, traces, logs, and Key Performance Indicators (KPIs) such as latency. However, both phases are not straightforward due to the highly dynamic and complex nature of the system, particularly in large-scale commercial architectures like Microsoft Exchange.

As cloud computing continues to expand globally, the need for effective management of decommissioned cloud assets in data centers becomes increasingly important. This work focuses on predicting the physical decommissioning date of cloud assets as a crucial component in reverse cloud supply chain management and data center warehouse operation. The decommissioning process is modeled as a contextual self-attentive temporal point process, which incorporates contextual information to model sequences with parallel events and provides more accurate predictions with more seen historical data. We conducted extensive offline and online experiments in 20 sampled data centers. The results show that the proposed methodology achieves the best performance compared with baselines and improves remarkable 94% prediction accuracy in online experiments. This modeling methodology can be extended to other domains with similar workflow-like processes.

Imitation learning has achieved great success in many sequential decision-making tasks, in which a neural agent is learned by imitating collected human demonstrations. However, existing algorithms typically require a large number of high-quality demonstrations that are difficult and expensive to collect. Usually, a trade-off needs to be made between demonstration quality and quantity in practice. Targeting this problem, in this work we consider the imitation of sub-optimal demonstrations, with both a small clean demonstration set and a large noisy set. Some pioneering works have been proposed, but they suffer from many limitations, e.g., assuming a demonstration to be of the same optimality throughout time steps and failing to provide any interpretation w.r.t knowledge learned from the noisy set. Addressing these problems, we propose \method by evaluating and imitating at the sub-demonstration level, encoding action primitives of varying quality into different skills. Concretely, SDIL consists of a high-level controller to discover skills and a skill-conditioned module to capture action-taking policies, and is trained following a two-phase pipeline by first discovering skills with all demonstrations and then adapting the controller to only the clean set. A mutual-information-based regularization and a dynamic sub-demonstration optimality estimator are designed to promote disentanglement in the skill space. Extensive experiments are conducted over two gym environments and a real-world healthcare dataset to demonstrate the superiority of SDIL in learning from sub-optimal demonstrations and its improved interpretability by examining learned skills.

Learning from positive and unlabeled data is known as positive-unlabeled (PU) learning in literature and has attracted much attention in recent years. One common approach in PU learning is to sample a set of pseudo-negatives from the unlabeled data using ad-hoc thresholds so that conventional supervised methods can be applied with both positive and negative samples. Owing to the label uncertainty among the unlabeled data, errors of misclassifying unlabeled positive samples as negative samples inevitably appear and may even accumulate during the training processes. Those errors often lead to performance degradation and model instability. To mitigate the impact of label uncertainty and improve the robustness of learning with positive and unlabeled data, we propose a new robust PU learning method with a training strategy motivated by the nature of human learning: easy cases should be learned first. Similar intuition has been utilized in curriculum learning to only use easier cases in the early stage of training before introducing more complex cases. Specifically, we utilize a novel ''hardness'' measure to distinguish unlabeled samples with a high chance of being negative from unlabeled samples with large label noise. An iterative training strategy is then implemented to fine-tune the selection of negative samples during the training process in an iterative manner to include more ''easy'' samples in the early stage of training. Extensive experimental validations over a wide range of learning tasks show that this approach can effectively improve the accuracy and stability of learning with positive and unlabeled data. Our code is available at https://github.com/woriazzc/Robust-PU.

Offline reinforcement learning faces a significant challenge of value over-estimation due to the distributional drift between the dataset and the current learned policy, leading to learning failure in practice. The common approach is to incorporate a penalty term to reward or value estimation in the Bellman iterations. Meanwhile, to avoid extrapolation on out-of-distribution (OOD) states and actions, existing methods focus on conservative Q-function estimation. In this paper, we propose Conservative State Value Estimation (CSVE), a new approach that learns conservative V-function via directly imposing penalty on OOD states. Compared to prior work, CSVE allows more effective state value estimation with conservative guarantees and further better policy optimization. Further, we apply CSVE and develop a practical actor-critic algorithm in which the critic does the conservative value estimation by additionally sampling and penalizing the states around the dataset, and the actor applies advantage weighted updates extended with state exploration to improve the policy. We evaluate in classic continual control tasks of D4RL, showing that our method performs better than the conservative Q-function learning methods and is strongly competitive among recent SOTA methods.

Root Cause Analysis (RCA) is becoming increasingly crucial for ensuring the reliability of microservice systems. However, performing RCA on modern microservice systems can be challenging due to their large scale, as they usually comprise hundreds of components, leading significant human effort. This paper proposes TraceDiag, an end-to-end RCA framework that addresses the challenges for large-scale microservice systems. It leverages reinforcement learning to learn a pruning policy for the service dependency graph to automatically eliminates redundant components, thereby significantly improving the RCA efficiency. The learned pruning policy is interpretable and fully adaptive to new RCA instances. With the pruned graph, a causal-based method can be executed with high accuracy and efficiency. The proposed TraceDiag framework is evaluated on real data traces collected from the Microsoft Exchange system, and demonstrates superior performance compared to state-of-the-art RCA approaches. Notably, TraceDiag has been integrated as a critical component in the Microsoft M365 Exchange, resulting in a significant improvement in the system's reliability and a considerable reduction in the human effort required for RCA.

Ensuring reliability in large-scale cloud systems like Microsoft 365 is crucial. Cloud failures, such as disk and node failure, threaten service reliability, causing service interruptions and financial loss. Existing works focus on failure prediction and proactively taking action before failures happen. However, they suffer from poor data quality, like data missing in model training and prediction, which limits performance. In this paper, we focus on enhancing data quality through data imputation by the proposed Diffusion+, a sample-efficient diffusion model, to impute the missing data efficiently conditioned on the observed data. Experiments with industrial datasets and application practice show that our model contributes to improving the performance of downstream failure prediction.

Oversubscription is a common practice for improving cloud resource utilization. It allows the cloud service provider to sell more resources than the physical limit, assuming not all users would fully utilize the resources simultaneously. However, how to design an oversubscription policy that improves utilization while satisfying some safety constraints remains an open problem. Existing methods and industrial practices are over-conservative, ignoring the coordination of diverse resource usage patterns and probabilistic constraints. To address these two limitations, this paper formulates the oversubscription for cloud as a chance-constrained optimization problem and proposes an effective Chance-Constrained Multi-Agent Reinforcement Learning (C2MARL) method to solve this problem. Specifically, C2MARL reduces the number of constraints by considering their upper bounds and leverages a multi-agent reinforcement learning paradigm to learn a safe and optimal coordination policy. We evaluate our C2MARL on an internal cloud platform and public cloud datasets. Experiments show that our C2MARL outperforms existing methods in improving utilization () under different levels of safety constraints.

Time series classification is a popular and important topic in machine learning, and it suffers from the class imbalance problem in many real-world applications. In this paper, to address the class imbalance problem, we propose a novel and practical oversampling method named T-SMOTE, which can make full use of the temporal information of time-series data. In particular, for each sample of minority class, T-SMOTE generates multiple samples that are close to class border. Then, based on those samples near class border, T-SMOTE synthesizes more samples. Finally, a weighted sampling method is called on both generated samples near class border and synthetic samples. Extensive experiments on a diverse set of both univariate and multivariate time-series datasets demonstrate that T-SMOTE consistently outperforms the current state-of-the-art methods on imbalanced time series classification. More encouragingly, our empirical evaluations show that T-SMOTE performs better in the scenario of early prediction, an important application scenario in industry, which indicates that T-SMOTE could bring benefits in practice.

Large-scale distributed systems, such as Microsoft 365's database system, require timely mitigation solutions to address failures and improve service availability and reliability. Still, mitigation actions can be costly as they may cause temporal performance degradation and even incur monetary expenses. Mitigation actions can be either administrated in a reactive fashion to contain detected failures or a proactive fashion to reduce potential failures. The proactive mitigation approach typically relies on a two-stage strategy: the prediction model will firstly identify instances (such as databases or disks) with high failure risk, then appropriate mitigation actions chosen by engineers or an automatic bandit learning model can be applied. As information is not fully shared across those two stages, important factors such as mitigation costs and states of instances are often ignored in one of those two stages. To address these issues, we propose NENYA, an end-to-end mitigation solution for a large-scale database system powered by a novel cascade reinforcement learning model. By taking the states of databases as input, NENYA directly outputs mitigation actions and is optimized based on jointly cumulative feedback on mitigation costs and failure rates. As the overwhelming majority of databases do not require mitigation actions, NENYA utilizes a novel cascade decision structure to firstly reliably filter out such databases and then focus on choosing appropriate mitigation actions for the rest. Extensive offline and online experiments have shown that our methods can outperform existing practices in reducing both failure rates of databases and mitigation costs. NENYA has been integrated into Microsoft 365, a productive platform, with sounding success.

Cloud computing systems have become increasingly popular in recent years. A typical cloud system utilizes millions of computing nodes as the basic infrastructure. Node failure has been identified as one of the most prevalent causes of cloud system downtime. To improve the reliability of cloud systems, many previous studies collected monitoring metrics from nodes and built models to predict node failures before the failures happen. However, based on our experience with large-scale real-world cloud systems in Microsoft, we find that the task of predicting node failure is severely hampered by missing data. There is a large amount of missing data, and the online latest data utilized for prediction is even worse. As a result, the real-time performance of the node prediction model is limited. In this paper, we first characterize the missing data problem for node failure prediction. Then, we evaluate several existing data interpolation approaches, and find that node dimension interpolation approaches outperform time dimension ones and deep learning based interpolation is the best for early prediction. Our findings can help academics and engineers address the missing data problem in cloud node failure prediction and other data-driven software engineering scenarios.

Dynamic Treatment Recommendation (DTR) is a sequence of tailored treatment decision rules which can be grouped as individual sub-tasks. As the reward signals in DTR are hard to design, Imitation Learning (IL) has achieved great success as it is effective in mimicking doctors' behaviors from their demonstrations without explicit reward signals. As a patient may have several different symptoms, the behaviors in doctors' demonstrations can often be grouped to handle individual symptoms. However, a single flat policy learned by IL is difficult to mimic doctors' demonstrations with such hierarchical structure, where low-level behaviors are switching from one symptom to another controlled by high-level decisions. Due to this observation, we consider Hierarchical Imitation Learning methods as good solutions for DTR. In this paper, we propose a novel Subgoal conditioned HIL framework (short for SHIL), where a high-level policy sequentially sets a subgoal for each sub-task without prior knowledge, and the low-level policy for sub-tasks is learned to reach the subgoal. To get rid of prior knowledge, a self-supervised learning method is proposed to learn an effective representation for each subgoal. More specifically, we carefully designed to encourage diverse representations among different subgoals. To demonstrate that SHIL is able to learn meaningful high-level policy and low-level policy that accurately reproduces complex doctors' demonstrations, we conduct experiments on a real-world medical data from health care domain, MIMIC-III. Compared with state-of-the-art baselines, SHIL improves the likelihood of patient survival by a significant margin and provides explainable recommendation with hierarchical structure.

In this article, we study the problem of guaranteed display ads (GDAs) allocation, which requires proactively allocate display ads to different impressions to fulfill their impression demands indicated in the contracts. Existing methods for this problem either assume the impressions that are static or solely consider a specific ad's benefits. Thus, it is hard to generalize to the industrial production scenario where the impressions are dynamical and large-scale, and the overall allocation optimality of all the considered GDAs is required. To bridge this gap, we formulate this problem as a sequential decision-making problem in the scope of multiagent reinforcement learning (MARL), by assigning an allocation agent to each ad and coordinating all the agents for allocating GDAs. The inputs are the states (e.g., the demands of the ad and the remaining time steps for displaying the ads) of each ad and the impressions at different time steps, and the outputs are the display ratios of each ad for each impression. Specifically, we propose a novel hierarchical MARL (HMARL) method that creates hierarchies over the agent policies to handle a large number of ads and the dynamics of impressions. HMARL contains: 1) a manager policy to navigate the agent to choose an appropriate subpolicy and 2) a set of subpolicies that let the agents perform diverse conditioning on their states. Extensive experiments on three real-world data sets from the Tencent advertising platform with tens of millions of records demonstrate significant improvements of HMARL over state-of-the-art approaches.

Recent developments in discovering dynamic treatment regimes (DTRs) have heightened the importance of deep reinforcement learning (DRL) which are used to recover the doctor’s treatment policies. However, existing DRL-based methods expose the following limitations: 1) supervised methods based on behavior cloning suffer from compounding errors; 2) the self-defined reward signals in reinforcement learning models are either too sparse or need clinical guidance; 3) only positive trajectories (e.g. survived patients) are considered in current imitation learning models, with negative trajectories (e.g. deceased patients) been largely ignored, which are examples of what not to do and could help the learned policy avoid repeating mistakes. To address these limitations, in this paper, we propose the adversarial cooperative imitation learning model, ACIL, to deduce the optimal dynamic treatment regimes that mimics the positive trajectories while differs from the negative trajectories. Specifically, two discriminators are used to help achieve this goal: an adversarial discriminator is designed to minimize the discrepancies between the trajectories generated from the policy and the positive trajectories, and a cooperative discriminator is used to distinguish the negative trajectories from the positive and generated trajectories. The reward signals from the discriminators are utilized to refine the policy for dynamic treatment regimes. Experiments on the publicly real-world medical data demonstrate that ACIL improves the likelihood of patient survival and provides better dynamic treatment regimes with the exploitation of information from both positive and negative trajectories.

Existing representation learning methods based on graph neural networks and their variants rely on the aggregation of neighborhood information, which makes it sensitive to noises in the graph, e.g. erroneous links between nodes, incorrect/missing node features. In this paper, we propose Graph Denoising Policy Network (short for GDPNet) to learn robust representations from noisy graph data through reinforcement learning. GDPNet first selects signal neighborhoods for each node, and then aggregates the information from the selected neighborhoods to learn node representations for the down-stream tasks. Specifically, in the signal neighborhood selection phase, GDPNet optimizes the neighborhood for each target node by formulating the process of removing noisy neighborhoods as a Markov decision process and learning a policy with task-specific rewards received from the representation learning phase. In the representation learning phase, GDPNet aggregates features from signal neighbors to generate node representations for down-stream tasks, and provides task-specific rewards to the signal neighbor selection phase. These two phases are jointly trained to select optimal sets of neighbors for target nodes with maximum cumulative task-specific rewards, and to learn robust representations for nodes. Experimental results on node classification task demonstrate the effectiveness of GDNet, outperforming the state-of-the-art graph representation learning methods on several well-studied datasets.

Dynamic treatment recommendation systems based on large-scale electronic health records (EHRs) become a key to successfully improve practical clinical outcomes. Prior relevant studies recommend treatments either use supervised learning (e.g. matching the indicator signal which denotes doctor prescriptions), or reinforcement learning (e.g. maximizing evaluation signal which indicates cumulative reward from survival rates). However, none of these studies have considered to combine the benefits of supervised learning and reinforcement learning. In this paper, we propose Supervised Reinforcement Learning with Recurrent Neural Network (SRL-RNN), which fuses them into a synergistic learning framework. Specifically, SRL-RNN applies an off-policy actor-critic framework to handle complex relations among multiple medications, diseases and individual characteristics. The "actor'' in the framework is adjusted by both the indicator signal and evaluation signal to ensure effective prescription and low mortality. RNN is further utilized to solve the Partially-Observed Markov Decision Process (POMDP) problem due to lack of fully observed states in real world applications. Experiments on the publicly real-world dataset, i.e., MIMIC-3, illustrate that our model can reduce the estimated mortality, while providing promising accuracy in matching doctors' prescriptions.

Detecting anomalous events from spatial data has important applications in real world. The spatial scan statistic methods are popular in this area. With maximizing the spatial statistical discrepancy by comparing observed data with a given baseline data distribution, significant spatial overdensity and underdensity can be detected. In reality, the spatial discrepancy is often irregularly shaped and has a structure of multiple spatial scales. However, a large-scale discrepancy pattern may not be significant when conducting fine granularity analysis. Meanwhile, local irregular boundaries of a maximized discrepancy cannot be well approximated with a coarse granularity analysis. Existing methods mostly work either on a fixed granularity, or with a regularly shaped scanning window. Thus, they have difficulties in characterizing such flexible spatial discrepancies. To solve the problem, in this paper we propose a novel discrepancy maximization algorithm, RefineScan. A grid hierarchy encoding multi-scale information is employed, making the algorithm capable of maximizing spatial discrepancies with multi-scale structures and irregular shapes. Experiments on a wide range of datasets demonstrate the advantages of RefineScan over the state-of-the-art algorithms: It always finds the largest discrepancy scores and remarkably better characterizes multi-scale discrepancy boundaries. Theoretical and empirical analyses also show that RefineScan has a moderate computational complexity and a good scalability.