Zhaoyang Wang, Yiming Liang, Xuchao Zhang, Qianhui Wu, Siwei Han, Anson Bastos, Rujia Wang, Chetan Bansal, Baolin Peng, Jianfeng Gao, Saravan Rajmohan, Huaxiu Yao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Current Large Language Models excel in general reasoning yet struggle with specialized tasks requiring proprietary or domain-specific knowledge. Fine-tuning large models for every niche application is often infeasible due to black-box constraints and high computational overhead. To address this, we propose a collaborative framework that pairs a specialized weak model with a general strong model. The weak model, tailored to specific domains, produces initial drafts and background information, while the strong model leverages its advanced reasoning to refine these drafts, extending LLMs’ capabilities to critical yet specialized tasks. To optimize this collaboration, we introduce a collaborative feedback to fine-tunes the weak model, which quantifies the influence of the weak model’s contributions in the collaboration procedure and establishes preference pairs to guide preference tuning of the weak model. We validate our framework through experiments on three domains. We find that the collaboration significantly outperforms each model alone by leveraging complementary strengths. Moreover, aligning the weak model with the collaborative preference further enhances overall performance.

Recent self-rewarding large language models (LLM) have successfully applied LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technologies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee of accuracy in the rewarding and ranking, which is critical for ensuring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning framework for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that leverages the consistency of rewards across different iterations to regularize the self-rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superiority of CREAM in improving both reward consistency and alignment performance. The code is publicly available at https://github.com/Raibows/CREAM.

High-quality preference data is essential for aligning foundation models with human values through preference learning. However, manual annotation of such data is often time-consuming and costly. Recent methods often adopt a self-rewarding approach, where the target model generates and annotates its own preference data, but this can lead to inaccuracies since the reward model shares weights with the target model, thereby amplifying inherent biases. To address these issues, we propose Anyprefer, a framework designed to synthesize high-quality preference data for aligning the target model. Anyprefer frames the data synthesis process as a cooperative two-player Markov Game, where the target model and the judge model collaborate together. Here, a series of external tools are introduced to assist the judge model in accurately rewarding the target model’s responses, mitigating biases in the rewarding process. In addition, a feedback mechanism is introduced to optimize prompts for both models, enhancing collaboration and improving data quality. The synthesized data is compiled into a new preference dataset, Anyprefer-V1, consisting of 58K high-quality preference pairs. Extensive experiments show that Anyprefer significantly improves model alignment performance across four main applications, covering 21 datasets, achieving average improvements of 18.55% in five natural language generation datasets, 3.66% in nine vision-language understanding datasets, 30.05% in three medical image analysis datasets, and 16.00% in four visuo-motor control tasks.

Multi-preference optimization enriches language-model alignment beyond pairwise preferences by contrasting entire sets of helpful and undesired responses, enabling richer training signals for large language models. During self-play alignment, these models often produce numerous candidate answers per query, making it computationally infeasible to include all of them in the training objective. We propose Active Multi-Preference Optimization (AMPO), which combines on-policy generation, a multi-preference group-contrastive loss, and active subset selection. Specifically, we score and embed large candidate pools of responses, then pick a small but informative subset—covering reward extremes and distinct semantic clusters—for preference optimization.The resulting contrastive-training scheme identifies not only the best and worst answers but also subtle, underexplored modes crucial for robust alignment. Theoretically, we provide guarantees of expected reward maximization using our active selection method. Empirically, AMPO achieves state-of-the-art results on AlpacaEval with Llama 8B and Mistral 7B. We release our datasets [here](https://huggingface.co/Multi-preference-Optimization).

As cloud service systems grow in scale and complexity, incidents that indicate unplanned interruptions and outages become unavoidable. Rapid and accurate triage of these incidents to the appropriate responsible teams is crucial to maintain service reliability and prevent significant financial losses. However, existing incident triage methods relying on manual operations and predefined rules often struggle with efficiency and accuracy due to the heterogeneity of incident data and the dynamic nature of domain knowledge across multiple teams.To solve these issues, we propose Triangle, an end-to-end incident triage system based on a Multi-Agent framework. Triangle leverages a semantic distillation mechanism to tackle the issue of semantic heterogeneity in incident data, enhancing the accuracy of incident triage. Additionally, we introduce multi-role agents and a negotiation mechanism to emulate human engineers’ workflows, effectively handling decentralized and dynamic domain knowledge from multiple teams. Furthermore, our system incorporates an automated troubleshooting information collection and mitigation mechanism, reducing the reliance on human labor and enabling fully automated end-to-end incident triage. Extensive experiments conducted on a real-world cloud production environment demonstrate that Triangle significantly improved incident triage accuracy (up to 97%) and reduced Time to Engage (TTE) by as much as 91%, demonstrating substantial operational impact across diverse cloud services.

Large Language Models (LLMs) are increasingly being used to plan, reason, and execute tasks across diverse scenarios. In use cases like repeatable workflows and agentic settings, prompts are often reused with minor variations while having a similar structure for recurring tasks. This opens up opportunities for caching. However, exact prompt matching fails on such structurally similar prompts, while semantic caching may produce incorrect responses by ignoring critical differences. To address this, we introduce GenCache, a generative cache that produces variation-aware responses for structurally similar prompts. GenCache identifies reusable response patterns across similar prompt structures and synthesizes customized outputs for new requests. We show that GenCache achieves 83\% cache hit rate, while having minimal incorrect hits on datasets without prompt repetition. In agentic workflows, it improves cache hit rate by $\sim$20\% and reduces end-to-end execution latency by $\sim$34\% compared to standard prompt matching.

Cancer remains a significant global health burden, underscoring the need for innovative diagnostic tools to enable early detection and improve patient outcomes. While circulating cell-free DNA (cfDNA) methylation has emerged as a promising biomarker for noninvasive cancer diagnostics, existing methods often face limitations in handling the high-dimensionality of methylation data, small sample sizes, and a lack of biological interpretability. To address these challenges, we propose cfMethylPre, a novel deep transfer learning framework tailored for cancer detection using cfDNA methylation data. cfMethylPre leverages large language model pretrained embeddings from DNA sequence information and integrates them with methylation profiles to enhance feature representation. The deep transfer learning process involves pretraining on bulk DNA methylation data encompassing 2801 samples across 82 cancer types and normal controls, followed by fine-tuning with cfDNA methylation data. This approach ensures robust adaptation to cfDNA's unique characteristics while improving predictive accuracy. Our model achieved superior predictive accuracy compared with state-of-the-art methods, with a weighted Matthews Correlation Coefficient of 0.926 and a weighted F1-score of 0.942. Through model interpretation and biological experimental validation, we identified three novel breast cancer genes-PCDHA10, PRICKLE2, and PRTG-demonstrating their inhibitory effects on cell proliferation and migration in breast cancer cell lines. These findings establish cfMethylPre as a powerful and interpretable tool for cancer diagnostics and biological discovery, paving the way for its application in precision oncology.

The rapid growth in the use of Large Language Models (LLMs) and AI Agents as part of software development and deployment is revolutionizing the information technology landscape. While code generation receives significant attention, a higher-impact application lies in using agents for the operational resilience of cloud services, which currently require significant human effort and domain knowledge. There is a growing interest in AI for IT Operations (AIOps), which aims to automate complex operational tasks, like fault localization and root cause analysis, reducing human intervention and customer impact. However, achieving the vision of autonomous and self-healing clouds through AIOps is hampered by the lack of standardized frameworks for building, evaluating, and improving AIOps agents. This vision paper lays the groundwork for such a framework by framing the requirements and then discussing design decisions that satisfy them. We also propose AIOpsLab, a prototype implementation leveraging agent-cloud-interface that orchestrates an application, injects real-time faults using chaos engineering, and interfaces with an agent to localize and resolve the faults. We report promising results and lay the groundwork to build a modular and robust framework for building, evaluating, and improving agents for autonomous clouds.

Ensuring the reliability and availability of cloud services necessitates efficient root cause analysis (RCA) for cloud incidents. Traditional RCA methods, which rely on manual investigations of data sources such as logs and traces, are often laborious, error-prone, and challenging for on-call engineers. In this paper, we introduce RCACopilot, an innovative on-call system empowered by the large language model for automating RCA of cloud incidents. RCACopilot matches incoming incidents to corresponding incident handlers based on their alert types, aggregates the critical runtime diagnostic information, predicts the incident's root cause category, and provides an explanatory narrative. We evaluate RCACopilot using a real-world dataset consisting of a year's worth of incidents from Microsoft. Our evaluation demonstrates that RCACopilot achieves RCA accuracy up to 0.766. Furthermore, the diagnostic information collection component of RCACopilot has been successfully in use at Microsoft for over four years.

Chen Ling, Xujiang Zhao, Xuchao Zhang, Wei Cheng, Yanchi Liu, Yiyou Sun, Mika Oishi, Takao Osaki, Katsushi Matsuda, Jie Ji, Guangji Bai, Liang Zhao, Haifeng Chen. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Artificial intelligence has significantly impacted medical applications, particularly with the advent of Medical Large Vision Language Models (Med-LVLMs), sparking optimism for the future of automated and personalized healthcare. However, the trustworthiness of Med-LVLMs remains unverified, posing significant risks for future model deployment. In this paper, we introduce CARES and aim to comprehensively evaluate the Trustworthiness of Med-LVLMs across the medical domain. We assess the trustworthiness of Med-LVLMs across five dimensions, including trustfulness, fairness, safety, privacy, and robustness. CARES comprises about 41K question-answer pairs in both closed and open-ended formats, covering 16 medical image modalities and 27 anatomical regions. Our analysis reveals that the models consistently exhibit concerns regarding trustworthiness, often displaying factual inaccuracies and failing to maintain fairness across different demographic groups. Furthermore, they are vulnerable to attacks and demonstrate a lack of privacy awareness. We publicly release our benchmark and code in https://github.com/richard-peng-xia/CARES.

Various contrastive learning approaches have been proposed in recent years and achieve significant empirical success. While effective and prevalent, contrastive learning has been less explored for time series data. A key component of contrastive learning is to select appropriate augmentations imposing some priors to construct feasible positive samples, such that an encoder can be trained to learn robust and discriminative representations. Unlike image and language domains where "desired'' augmented samples can be generated with the rule of thumb guided by prefabricated human priors, the ad-hoc manual selection of time series augmentations is hindered by their diverse and human-unrecognizable temporal structures. How to find the desired augmentations of time series data that are meaningful for given contrastive learning tasks and datasets remains an open question. In this work, we address the problem by encouraging both high fidelity and variety based on information theory. A theoretical analysis leads to the criteria for selecting feasible data augmentations. On top of that, we propose a new contrastive learning approach with information-aware augmentations, InfoTS, that adaptively selects optimal augmentations for time series representation learning. Experiments on various datasets show highly competitive performance with up to a 12.0% reduction in MSE on forecasting tasks and up to 3.7% relative improvement in accuracy on classification tasks over the leading baselines.

Meta-learning has emerged as a trending technique to tackle few-shot text classification and achieve state-of-the-art performance.However, the performance of existing approaches heavily depends on the inter-class variance of the support set.As a result, it can perform well on tasks when the semantics of sampled classes are distinct while failing to differentiate classes with similar semantics.In this paper, we propose a novel Task-Adaptive Reference Transformation (TART) network, aiming to enhance the generalization by transforming the class prototypes to per-class fixed reference points in task-adaptive metric spaces.To further maximize divergence between transformed prototypes in task-adaptive metric spaces, TART introduces a discriminative reference regularization among transformed prototypes.Extensive experiments are conducted on four benchmark datasets and our method demonstrates clear superiority over the stateof-the-art models in all the datasets.In particular, our model surpasses the state-of-the-art method by 7.4% and 5.4% in 1-shot and 5-shot classification on the 20 Newsgroups dataset, respectively.

Early event detection aims to detect events even before the event is complete. However, most of the existing methods focus on an event with a single label but fail to be applied to cases with multiple labels. Another non-negligible issue for early event detection is a prediction with overconfidence due to the high vacuity uncertainty that exists in the early time series. It results in an over-confidence estimation and hence unreliable predictions. To this end, technically, we propose a novel framework, Multi-Label Temporal Evidential Neural Network (MTENN), for multi-label uncertainty estimation in temporal data. MTENN is able to quality predictive uncertainty due to the lack of evidence for multi-label classifications at each time stamp based on belief/evidence theory. In addition, we introduce a novel uncertainty estimation head (weighted binomial comultiplication (WBC)) to quantify the fused uncertainty of a sub-sequence for early event detection. We validate the performance of our approach with state-of-the-art techniques on real-world audio datasets.

Incident management for cloud services is a complex process involving several steps and has a huge impact on both service health and developer productivity. On-call engineers require significant amount of domain knowledge and manual effort for root causing and mitigation of production incidents. Recent advances in artificial intelligence has resulted in state-of-the-art large language models like GPT-3.x (both GPT-3.0 and GPT-3.5), which have been used to solve a variety of problems ranging from question answering to text summarization. In this work, we do the first large-scale study to evaluate the effectiveness of these models for helping engineers root cause and mitigate production incidents. We do a rigorous study at Microsoft, on more than 40,000 incidents and compare several large language models in zero-shot, fine-tuned and multi-task setting using semantic and lexical metrics. Lastly, our human evaluation with actual incident owners show the efficacy and future potential of using artificial intelligence for resolving cloud incidents.

Few-shot text classification has extensive application where the sample collection is expensive or complicated. When the penalty for classification errors is high, such as early threat event detection with scarce data, we expect to know "whether we should trust the classification results or reexamine them.'' This paper investigates the Uncertainty Estimation for Few-shot Text Classification (UEFTC), an unexplored research area. Given limited samples, a UEFTC model predicts an uncertainty score for a classification result, which is the likelihood that the classification result is false. However, many traditional uncertainty estimation models in text classification are unsuitable for implementing a UEFTC model. These models require numerous training samples, whereas the few-shot setting in UEFTC only provides a few or just one support sample for each class in an episode. We propose Contrastive Learning from Uncertainty Relations (CLUR) to address UEFTC. CLUR can be trained with only one support sample for each class with the help of pseudo uncertainty scores. Unlike previous works that manually set the pseudo uncertainty scores, CLUR self-adaptively learns them using our proposed uncertainty relations. Specifically, we explore four model structures in CLUR to investigate the performance of three common-used contrastive learning components in UEFTC and find that two of the components are effective. Experiment results prove that CLUR outperforms six baselines on four datasets, including an improvement of 4.52% AUPR on an RCV1 dataset in a 5-way 1-shot setting. Our code and data split for UEFTC are in https://github.com/he159ok/CLUR_UncertaintyEst_FewShot_TextCls.

We target the task of cross-lingual Machine Reading Comprehension (MRC) in the direct zero-shot setting, by incorporating syntactic features from Universal Dependencies (UD), and the key features we use are the syntactic relations within each sentence. While previous work has demonstrated effective syntax-guided MRC models, we propose to adopt the inter-sentence syntactic relations, in addition to the rudimentary intra-sentence relations, to further utilize the syntactic dependencies in the multi-sentence input of the MRC task. In our approach, we build the Inter-Sentence Dependency Graph (ISDG) connecting dependency trees to form global syntactic relations across sentences. We then propose the ISDG encoder that encodes the global dependency graph, addressing the inter-sentence relations via both one-hop and multi-hop dependency paths explicitly. Experiments on three multilingual MRC datasets (XQuAD, MLQA, TyDiQA-GoldP) show that our encoder that is only trained on English is able to improve the zero-shot performance on all 14 test sets covering 8 languages, with up to 3.8 F1 / 5.2 EM improvement on-average, and 5.2 F1 / 11.2 EM on certain languages. Further analysis shows the improvement can be attributed to the attention on the cross-linguistically consistent syntactic path. Our code is available at https://github.com/lxucs/multilingual-mrc-isdg.

Sound Event Early Detection (SEED) is an essential task in recognizing the acoustic environments and soundscapes. However, most of the existing methods focus on the offline sound event detection, which suffers from the over-confidence issue of early-stage event detection and usually yield unreliable results. To solve the problem, we propose a novel Polyphonic Evidential Neural Network (PENet) to model the evidential uncertainty of the class probability with Beta distribution. Specifically, we use a Beta distribution to model the distribution of class probabilities, and the evidential uncertainty enriches uncertainty representation with evidence information, which plays a central role in reliable prediction. To further improve the event detection performance, we design the backtrack inference method that utilizes both the forward and backward audio features of an ongoing event. Experiments on the DESED database show that the proposed method can simultaneously improve 13.0% and 3.8% in time delay and detection F1 score compared to the state-of-the-art methods.

Analogical reasoning is the process of discovering and mapping correspondences from a target subject to a base subject. As the most well-known computational method of analogical reasoning, Structure-Mapping Theory (SMT) abstracts both target and base subjects into relational graphs and forms the cognitive process of analogical reasoning by finding a corresponding subgraph (i.e., correspondence) in the target graph that is aligned with the base graph. However, incorporating deep learning for SMT is still under-explored due to several obstacles: 1) the combinatorial complexity of searching for the correspondence in the target graph; 2) the correspondence mining is restricted by various cognitive theory-driven constraints. To address both challenges, we propose a novel framework for Analogical Reasoning (DeepGAR) that identifies the correspondence between source and target domains by assuring cognitive theory-driven constraints. Specifically, we design a geometric constraint embedding space to induce subgraph relation from node embeddings for efficient subgraph search. Furthermore, we develop novel learning and optimization strategies that could end-to-end identify correspondences that are strictly consistent with constraints driven by the cognitive theory. Extensive experiments are conducted on synthetic and real-world datasets to demonstrate the effectiveness of the proposed DeepGAR over existing methods. The code and data are available at: https://github.com/triplej0079/DeepGAR.

It is critical and important to detect anomalies in event sequences, which becomes widely available in many application domains. Indeed, various efforts have been made to capture abnormal patterns from event sequences through sequential pattern analysis or event representation learning. However, existing approaches usually ignore the semantic information of event content. To this end, in this paper, we propose a self-attentive encoder-decoder transformer framework, Content-Aware Transformer CAT, for anomaly detection in event sequences. In CAT, the encoder learns preamble event sequence representations with content awareness, and the decoder embeds sequences under detection into a latent space, where anomalies are distinguishable. Specifically, the event content is first fed to a content-awareness layer, generating representations of each event. The encoder accepts preamble event representation sequence, generating feature maps. In the decoder, an additional token is added at the beginning of the sequence under detection, denoting the sequence status. A one-class objective together with sequence reconstruction loss is collectively applied to train our framework under the label efficiency scheme. Furthermore, CAT is optimized under a scalable and efficient setting. Finally, extensive experiments on three real-world datasets demonstrate the superiority of CAT.

Document authoring involves a lengthy revision process, marked by individual edits that are frequently linked to comments. Modeling the relationship between edits and comments leads to a better understanding of document evolution, potentially benefiting applications such as content summarization, and task triaging. Prior work on understanding revisions has primarily focused on classifying edit intents, but falling short of a deeper understanding of the nature of these edits. In this paper, we present explore the challenge of describing an edit at two levels: identifying the edit intent, and describing the edit using free-form text. We begin by defining a taxonomy of general edit intents and introduce a new dataset of full revision histories of Wikipedia pages, annotated with each revision’s edit intent. Using this dataset, we train a classifier that achieves a 90% accuracy in identifying edit intent. We use this classifier to train a distantly-supervised model that generates a high-level description of a revision in free-form text. Our experimental results show that incorporating edit intent information aids in generating better edit descriptions. We establish a set of baselines for the edit description task, achieving a best score of 28 ROUGE, thus demonstrating the effectiveness of our layered approach to edit understanding.

In today’s era of big data, robust least-squares regression becomes a more challenging problem when considering the extremely corrupted labels along with explosive growth of datasets. Traditional robust methods can handle the noise but suffer from several challenges when applied in huge dataset including (1) computational infeasibility of handling an entire dataset at once, (2) existence of heterogeneously distributed corruption, and (3) difficulty in corruption estimation when data cannot be entirely loaded. This article proposes online and distributed robust regression approaches, both of which can concurrently address all the above challenges. Specifically, the distributed algorithm optimizes the regression coefficients of each data block via heuristic hard thresholding and combines all the estimates in a distributed robust consolidation. In addition, an online version of the distributed algorithm is proposed to incrementally update the existing estimates with new incoming data. Furthermore, a novel online robust regression method is proposed to estimate under a biased-batch corruption. We also prove that our algorithms benefit from strong robustness guarantees in terms of regression coefficient recovery with a constant upper bound on the error of state-of-the-art batch methods. Extensive experiments on synthetic and real datasets demonstrate that our approaches are superior to those of existing methods in effectiveness, with competitive efficiency.

Using attention weights to identify information that is important for models’ decision making is a popular approach to interpret attention-based neural networks. This is commonly realized in practice through the generation of a heat-map for every single document based on attention weights. However, this interpretation method is fragile and it is easy to find contradictory examples. In this article, we propose a corpus-level explanation approach, which aims at capturing causal relationships between keywords and model predictions via learning the importance of keywords for predicted labels across a training corpus based on attention weights. Based on this idea, we further propose a concept-based explanation method that can automatically learn higher level concepts and their importance to model prediction tasks. Our concept-based explanation method is built upon a novel Abstraction-Aggregation Network (AAN), which can automatically cluster important keywords during an end-to-end training process. We apply these methods to the document classification task and show that they are powerful in extracting semantically meaningful keywords and concepts. Our consistency analysis results based on an attention-based Naïve Bayes classifier (NBC) also demonstrate that these keywords and concepts are important for model predictions.

Forecasting on sparse multivariate time series (MTS) aims to model the predictors of future values of time series given their incomplete past, which is important for many emerging applications. However, most existing methods process MTS’s individually, and do not leverage the dynamic distributions underlying the MTS’s, leading to sub-optimal results when the sparsity is high. To address this challenge, we propose a novel generative model, which tracks the transition of latent clusters, instead of isolated feature representations, to achieve robust modeling. It is characterized by a newly designed dynamic Gaussian mixture distribution, which captures the dynamics of clustering structures, and is used for emitting time series. The generative model is parameterized by neural networks. A structured inference network is also designed for enabling inductive analysis. A gating mechanism is further introduced to dynamically tune the Gaussian mixture distributions. Extensive experimental results on a variety of real-life datasets demonstrate the effectiveness of our method.

We consider the models of deep multi-task learning with recurrent architectures that exploit regularities across tasks to improve the performance of multiple sequence processing tasks jointly. Most existing architectures are painstakingly customized to learn task relationships for different problems, which is not flexible enough to model the dynamic task relationships and lacks generalization abilities to novel test-time scenarios. We propose multi-task recurrent modular networks (MT-RMN) that can be incorporated in any multi-task recurrent models to address the above drawbacks. MT-RMN consists of a shared encoder and multiple task-specific decoders, and recurrently operates over time. For better flexibility, it modularizes the encoder into multiple layers of sub-networks and dynamically controls the connection between these sub-networks and the decoders at different time steps, which provides the recurrent networks with varying degrees of parameter sharing for tasks with dynamic relatedness. For the generalization ability, MT-RMN aims to discover a set of generalizable sub-networks in the encoder that are assembled in different ways for different tasks. The policy networks augmented with the differentiable routers are utilized to make the binary connection decisions between the sub-networks. The experimental results on three multi-task sequence processing datasets consistently demonstrate the effectiveness of MT-RMN.

Hepatoma is the most common type of primary liver cancer with a high mortality rate in the world. The genetic causes of the disease pathology remain largely unknown. Effective discovery of the genes associated with hepatoma has become important in disease prevention, early diagnosis, and therapeutic treatments. With the developments of molecular networks, graph-based methods have been tremendously successful in predicting disease genes based on the hypothesis of guilt-by-association. Network representation learning (NRL) techniques have accelerated disease gene discovery in recent years because of their powerful network feature extraction ability. However, the current network representation learning-based methods for disease gene discovery did not consider the gene features derived from gene ontology annotations, which apriori group genes with similar functions. To fill this gap, here we propose a novel framework to predict hepatoma-related genes based on representation learning from both protein-protein interactions (PPI) network and gene ontology annotations. Our framework has three steps: learning features from PPI network and gene ontologies using NRL techniques, integrating different features based on autoencoder, predicting hepatoma-related genes using machine learning classifiers. Experiments have demonstrated that our framework could accurately predict hepatoma-related genes with AUROC and AUPRC reaching 0.93 and 0.94, respectively. Compared with other methods using only single representation features, our framework also shows superior performance on hepatoma gene prediction.

In many high-stakes applications of machine learning models, outputting only predictions or providing statistical confidence is usually insufficient to gain trust from end users, who often prefer a transparent reasoning paradigm. Despite the recent encouraging developments on deep networks for sequential data modeling, due to the highly recursive functions, the underlying rationales of their predictions are difficult to explain. Thus, in this paper, we aim to develop a sequence modeling approach that explains its own predictions by breaking input sequences down into evidencing segments (i.e., sub-sequences) in its reasoning. To this end, we build our model upon convolutional neural networks, which, in their vanilla forms, associates local receptive fields with outputs in an obscure manner. To unveil it, we resort to case-based reasoning, and design prototype modules whose units (i.e., prototypes) resemble exemplar segments in the problem domain. Each prediction is obtained by combining the comparisons between the prototypes and the segments of an input. To enhance interpretability, we propose a training objective that delicately adapts the distribution of prototypes to the data distribution in latent spaces, and design an algorithm to map prototypes to human-understandable segments. Through extensive experiments in a variety of domains, we demonstrate that our model can achieve high interpretability generally, together with a competitive accuracy to the state-of-the-art approaches.

Recent multilingual pre-trained language models have achieved remarkable zero-shot performance, where the model is only finetuned on one source language and directly evaluated on target languages.In this work, we propose a self-learning framework that further utilizes unlabeled data of target languages, combined with uncertainty estimation in the process to select high-quality silver labels.Three different uncertainties are adapted and analyzed specifically for the cross lingual transfer: Language Heteroscedastic/Homoscedastic Uncertainty (LEU/LOU), Evidential Uncertainty (EVI).We evaluate our framework with uncertainties on two cross-lingual tasks including Named Entity Recognition (NER) and Natural Language Inference (NLI) covering 40 languages in total, which outperforms the baselines significantly by 10 F1 on average for NER and 2.5 accuracy score for NLI.

Aspect-based sentiment classification aims to predict sentimental polarities of one or multiple aspects in texts. As texts always contain a large proportion of task-irrelevant words, accurate alignment between aspects and their sentimental descriptions is the most crucial and challenging step. State-of-the-art approaches are mainly based on word-level attention learned from recurrent neural network variants (e.g., LSTM) or graph neural networks. From another view, these methods essentially weight and aggregate all possible alignments. However, this mechanism heavily relies on large-scale supervision training: without enough labels, it could easily overfit with difficulty in generalization. To address this challenge, we propose SentRL, a reinforcement learning-based framework for aspect-based sentiment classification. In this framework, input texts are transformed into their dependency graphs. Then, an agent is deployed to walk on the graphs, explores paths from target aspect nodes to their potential sentimental regions, and differentiates the effectiveness of different paths. By limiting the agent’s exploration budget, our method encourages the agent to skip task-irrelevant information and focus on the most effective paths for alignment purpose. Our method considerably reduces the impact of task-irrelevant words and improves generalization performance. Compared with competitive baseline methods, our approach achieves the highest performance on public benchmark datasets with up to 3.7% improvement.

Despite the great progress made by deep neural networks in the semantic segmentation task, traditional neural-network-based methods typically suffer from a shortage of large amounts of pixel-level annotations. Recent progress in few-shot semantic segmentation tackles the issue by only a few pixel-level annotated examples. However, these few-shot approaches cannot easily be applied to multi-way or weak an-notation settings. In this paper, we advance the few-shot segmentation paradigm towards a scenario where image-level an-notations are available to help the training process of a few pixel-level annotations. Our key idea is to learn a better prototype representation of the class by fusing the knowledge from the image-level labeled data. Specifically, we propose a new framework, called PAIA, to learn the class prototype representation in a metric space by integrating image-level annotations. Furthermore, by considering the uncertainty of pseudo-masks, a distilled soft masked average pooling strategy is designed to handle distractions in image-level annotations. Extensive empirical results on two datasets show superior performance of PAIA.

Xuchao Zhang, Bo Zong, Wei Cheng, Jingchao Ni, Yanchi Liu, Haifeng Chen. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

With the advance of sensor technologies, the Multivariate Time Series classification (MTSC) problem, perhaps one of the most essential problems in the time series data mining domain, has continuously received a significant amount of attention in recent decades. Traditional time series classification approaches based on Bag-of-Patterns or Time Series Shapelet have difficulty dealing with the huge amounts of feature candidates generated in high-dimensional multivariate data but have promising performance even when the training set is small. In contrast, deep learning based methods can learn low-dimensional features efficiently but suffer from a shortage of labelled data. In this paper, we propose a novel MTSC model with an attentional prototype network to take the strengths of both traditional and deep learning based approaches. Specifically, we design a random group permutation method combined with multi-layer convolutional networks to learn the low-dimensional features from multivariate time series data. To handle the issue of limited training labels, we propose a novel attentional prototype network to train the feature representation based on their distance to class prototypes with inadequate data labels. In addition, we extend our model into its semi-supervised setting by utilizing the unlabeled data. Extensive experiments on 18 datasets in a public UEA Multivariate time series archive with eight state-of-the-art baseline methods exhibit the effectiveness of the proposed model.

The success of training accurate models strongly depends on the availability of a sufficient collection of precisely labeled data. However, real-world datasets contain erroneously labeled data samples that substantially hinder the performance of machine learning models. Meanwhile, well-labeled data is usually expensive to obtain and only a limited amount is available for training. In this paper, we consider the problem of training a robust model by using large-scale noisy data in conjunction with a small set of clean data. To leverage the information contained via the clean labels, we propose a novel self-paced robust learning algorithm (SPRL) that trains the model in a process from more reliable (clean) data instances to less reliable (noisy) ones under the supervision of well-labeled data. The self-paced learning process hedges the risk of selecting corrupted data into the training set. Moreover, theoretical analyses on the convergence of the proposed algorithm are provided under mild assumptions. Extensive experiments on synthetic and real-world datasets demonstrate that our proposed approach can achieve a considerable improvement in effectiveness and robustness to existing methods.

Jianfeng He, Xuchao Zhang, Shuo Lei, Zhiqian Chen, Fanglan Chen, Abdulaziz Alhamadani, Bei Xiao, ChangTien Lu. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.

Multi-target regression has recently drawn increasing attention in the machine learning community due to its capability of simultaneously predicting multiple continuous target variables based on a given set of input features. Jointly handling the inter-target correlations and input-output relationships is very challenging. That task becomes even more intricate in the presence of correlated data corruption. We observe that traditional robust methods can hardly deal with several emerging challenges, including 1) presence of correlated corruption among targets in the datasets, 2) difficulty in estimating the data corruption ratio, and 3) scalability to massive datasets. This paper proposes a novel approach that addresses all the above challenges by developing a distributed robust regression algorithm. Specifically, the algorithm optimizes regression coefficients of each target in parallel with a heuristically estimated corruption ratio and then consolidates the uncorrupted set in two strategies: global consensus and majority voting. Also, we prove that our algorithm benefits from strong guarantees in terms of convergence rates and coefficient recovery, which can be applied as a generic framework for robust regression problems with correlated corruption property. Extensive experiments on synthetic and real-world datasets demonstrate that our algorithm is superior to existing methods in both effectiveness and efficiency.

Question routing (QR) aims at recommending newly posted questions to the potential answerers who are most likely to answer the questions. The existing approaches that learn users' expertise from their past question-answering activities usually suffer from challenges in two aspects: 1) multi-faceted expertise and 2) temporal dynamics in the answering behavior. This paper proposes a novel temporal context-aware model in multiple granularities of temporal dynamics that concurrently address the above challenges. Specifically, the temporal context-aware attention characterizes the answerer's multi-faceted expertise in terms of the questions' semantic and temporal information simultaneously. Moreover, the design of the multi-shift and multi-resolution module enables our model to handle temporal impact on different time granularities. Extensive experiments on six datasets from different domains demonstrate that the proposed model significantly outperforms competitive baseline models.

Cybersecurity event detection is a crucial problem for mitigating effects on various aspects of society. Social media has become a notable source of indicators for detection of diverse events. Though previous social media based strategies for cyber-security event detection focus on mining certain event-related words, the dynamic and evolving nature of online discourse limits the performance of these approaches. Further, because these are typically unsupervised or weakly supervised learning strategies, they do not perform well in an environment of biased samples, noisy context, and informal language which is routine for online, user-generated content. This paper takes a supervised learning approach by proposing a novel multi-task learning based model. Our model can handle diverse structures in feature space by learning models for different types of potential high-profile targets simultaneously. For parameter optimization, we develop an efficient algorithm based on the alternating direction method of multipliers. Through extensive experiments on a real world Twitter dataset, we demonstrate that our approach consistently outperforms existing methods at encoding and identifying cyber-security incidents.

Network embedding, which aims to learn low-dimensional vector representations for nodes in a network, has shown promising performance for many real-world applications, such as node classification and clustering. While various embedding methods have been developed for network data, they are limited in their assumption that nodes are correlated with their neighboring nodes with the same similarity degree. As such, these methods can be suboptimal for embedding network data. In this paper, we propose a new method named SANE, short for Similarity-Aware Network Embedding, to learn node representations by explicitly considering different similarity degrees between connected nodes in a network. In particular, we develop a new framework based on self-paced learning by accounting for both the explicit relations (i.e., observed links) and implicit relations (i.e., unobserved node similarities) in network representation learning. To justify our proposed model, we perform experiments on two real-world network data. Experiments results show that SNAE outperforms state-of-the-art embedding models on the tasks of node classification and node clustering.

With the increase of temporal data availability, time series classification has drawn a lot of attention in the literature because of its wide spectrum of applications in diverse domains (e.g., healthcare, bioinformatics and finance), ranging from human activity recognition to financial pattern identification. While significant progress has been made to solve time series classification problem, the success of such methods relies on data sufficiency, and may not well capture the quality embeddings when training triple instances are scarce and highly imbalance across classes. To address these challenges, we propose a prototype embedding framework-Deep Prototypical Networks (DPN), which leverages a main embedding space to capture the discrepancies of difference time series classes for alleviating data scarcity. In addition, we further augment DPN framework with a relationship-dependent masking module to automatically fuse relevant information with a distance metric learning process, which addresses the data imbalance issue and performs robust time series classification. Experimental results show significant and consistent improvements compared to state-of-the-art techniques.

Xuchao Zhang, Dheeraj Rajagopal, Michael Gamon, Sujay Kumar Jauhar, ChangTien Lu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Online purchase forecasting is of great importance in e-commerce platforms, which is the basis of how to present personalized interesting product lists to individual customers. However, predicting online purchases is not trivial as it is influenced by many factors including: (i) the complex temporal pattern with hierarchical inter-correlations; (ii) arbitrary category dependencies. To address these factors, we develop a Graph Multi-Scale Pyramid Networks (GMP) framework to fully exploit users' latent behavioral patterns with both multi-scale temporal dynamics and arbitrary inter-dependencies among product categories. In GMP, we first design a multi-scale pyramid modulation network architecture which seamlessly preserves the underlying hierarchical temporal factors--governing users' purchase behaviors. Then, we employ convolution recurrent neural network to encode the categorical temporal pattern at each scale. After that, we develop a resolution-wise recalibration gating mechanism to automatically re-weight the importance of each scale-view representations. Finally, a context-graph neural network module is proposed to adaptively uncover complex dependencies among category-specific purchases. Extensive experiments on real-world e-commerce datasets demonstrate the superior performance of our method over state-of-the-art baselines across various settings.

The presence of data noise and corruptions has recently invoked increasing attention on robust least-squares regression ( RLSR ), which addresses this fundamental problem that learns reliable regression coefficients when response variables can be arbitrarily corrupted. Until now, the following important challenges could not be handled concurrently: (1) rigorous recovery guarantee of regression coefficients, (2) difficulty in estimating the corruption ratio parameter, and (3) scaling to massive datasets. This article proposes a novel Robust regression algorithm via Heuristic Corruption Thresholding ( RHCT ) that concurrently addresses all the above challenges. Specifically, the algorithm alternately optimizes the regression coefficients and estimates the optimal uncorrupted set via heuristic thresholding without a pre-defined corruption ratio parameter until its convergence. Moreover, to improve the efficiency of corruption estimation in large-scale data, a Robust regression algorithm via Adaptive Corruption Thresholding ( RACT ) is proposed to determine the size of the uncorrupted set in a novel adaptive search method without iterating data samples exhaustively. In addition, we prove that our algorithms benefit from strong guarantees analogous to those of state-of-the-art methods in terms of convergence rates and recovery guarantees. Extensive experiments demonstrate that the effectiveness of our new methods is superior to that of existing methods in the recovery of both regression coefficients and uncorrupted sets, with very competitive efficiency.

Predicting individuals personality traits with their social media profile has proved to be feasible, but researchers recently have run into bottlenecks on further improving the prediction accuracy. One major limitation is that existing studies failed to consider context information in predicting social media users' behaviors. In this paper, we adopted the DIAMONDS situation theory in psychology to capture the context information in Facebook posts. To solve this issue, we proposed a novel situation-based feature interaction learning model. In this study, we extracted situation features according to the DIAMONDS lexicon and computed the interaction values between these situation features and the commonly used n-gram features at the post level. Features at the post level were aggregated up to the user level using the averaging strategy. A group lasso penalty was employed to enforce strong heredity in the model, which addressed the overfitting challenge introduced by the interaction features. Empirical tests on a large-scale data set have demonstrated the effectiveness of the proposed method.

For node level graph encoding, a recent important state-of-art method is the graph convolutional networks (GCN), which nicely integrate local vertex features and graph topology in the spectral domain. However, current studies suffer from several drawbacks: (1) graph CNNs rely on Chebyshev polynomial approximation which results in oscillatory approximation at jump discontinuities; (2) Increasing the order of Chebyshev polynomial can reduce the oscillations issue, but also incurs unaffordable computational cost; (3) Chebyshev polynomials require degree Ω(poly(1/ε)) to approximate a jump signal such as |x|, while rational function only needs O(poly log(1/ε)). However, it is non-trivial to apply rational approximation without increasing computational complexity due to the denominator. In this paper, the superiority of rational approximation is exploited for graph signal recovering. RatioanlNet is proposed to integrate rational function and neural networks. We show that the rational function of eigenvalues can be rewritten as a function of graph Laplacian, which can avoid multiplication by the eigenvector matrix. Focusing on the analysis of approximation on graph convolution operation, a graph signal regression task is formulated. Under graph signal regression task, its time complexity can be significantly reduced by graph Fourier transform. To overcome the local minimum problem of neural networks model, a relaxed Remez algorithm is utilized to initialize the weight parameters. Convergence rate of RatioanlNet and polynomial based methods on a jump signal is analyzed for a theoretical guarantee. The extensive experimental results demonstrated that our approach could effectively characterize the jump discontinuities, outperforming competing methods by a substantial margin on both synthetic and real-world graphs.

The presence of data corruption in user-generated streaming data, such as social media, motivates a new fundamental problem that learns reliable regression coefficient when features are not accessible entirely at one time. Until now, several important challenges still cannot be handled concurrently: 1) corrupted data estimation when only partial features are accessible; 2) online feature selection when data contains adversarial corruption; and 3) scaling to a massive dataset. This paper proposes a novel RObust regression algorithm via Online Feature Selection (RoOFS) that concurrently addresses all the above challenges. Specifically, the algorithm iteratively updates the regression coefficients and the uncorrupted set via a robust online feature substitution method. Extensive empirical experiments in both synthetic and real-world data sets demonstrated that the effectiveness of our new method is superior to that of existing methods in the recovery of both feature selection and regression coefficients, with very competitive efficiency.

Self-paced learning (SPL) mimics the cognitive process of humans, who generally learn from easy samples to hard ones. One key issue in SPL is the training process required for each instance weight depends on the other samples and thus cannot easily be run in a distributed manner in a large-scale dataset. In this paper, we reformulate the self-paced learning problem into a distributed setting and propose a novel Distributed Self-Paced Learning method (DSPL) to handle large scale datasets. Specifically, both the model and instance weights can be optimized in parallel for each batch based on a consensus alternating direction method of multipliers. We also prove the convergence of our algorithm under mild conditions. Extensive experiments on both synthetic and real datasets demonstrate that our approach is superior to those of existing methods.

In the era of information overload, people are struggling to make sense of complex story events in massive social media data. Most existing approaches are designed to address event extraction in news reports, documents and abstracts, but such approaches are not suitable for Twitter data streams due to their unstructured language, short-length messages, and heterogeneous features; few existing approach generates a story by considering both the shared topics throughout the story and the smooth connection between successive nodes simultaneously. In this paper, a novel Twitter stoRy generation framework via shAred subspaCe and tEmporal Smoothness called TRACES is proposed. Given a query of an ongoing event, a novel multi-task clustering method integrated with shared subspace and temporal smoothness (STMTC) is proposed to generate the event stories. Extensive experimental evaluations of data sets for different events demonstrate the effectiveness of this new approach.

In today's era of information overload, people are struggling to detect the evolution of hot topics from massive news media and microblogs such as Twitter. Reports from mainstream news agencies and discussions from microblogs could complement each other to form a complete picture of major events. Existing work has generally focused on a single source, seldom attempting to combine multiple sources to track the evolution of topics: emerging, evolving and fading phrases as this would require a considerably more sophisticated model. This paper proposes a novel story discovery model that integrates evolutionary topics in news and Twitter data sources using an incremental algorithm by 1) discovering complementary information from news and microblogs that provides a more complete view of major events; 2) modeling emerging, evolving and fading topics and features throughout ongoing events; and 3) creating a scalable algorithm that is capable of handling massive data from news and social media. The parameters of the new model are optimized using a novel algorithm based on the alternative direction method of multipliers (ADMM). Extensive experimental evaluations on multiple datasets from different domains demonstrate the effectiveness and efficiency of our proposed approach.

Hyper-local pricing data, e.g., about foods and commodities, exhibit subtle spatiotemporal variations that can be useful as crucial precursors of future events. Three major challenges in modeling such pricing data include: i) temporal dependencies underlying features; ii) spatiotemporal missing values; and iii) constraints underlying economic phenomena. These challenges hinder traditional event forecasting models from being applied effectively. This paper proposes a novel spatiotemporal event forecasting model that concurrently addresses the above challenges. Specifically, given continuous price data, a new soft time-lagged model is designed to select temporally dependent features. To handle missing values, we propose a data tensor completion method based on price domain knowledge. The parameters of the new model are optimized using a novel algorithm based on the Alternative Direction Methods of Multipliers (ADMM). Extensive experimental evaluations on multiple datasets demonstrate the effectiveness of our proposed approach.

In today's era of big data, robust least-squares regression becomes a more challenging problem when considering the adversarial corruption along with explosive growth of datasets. Traditional robust methods can handle the noise but suffer from several challenges when applied in huge dataset including 1) computational infeasibility of handling an entire dataset at once, 2) existence of heterogeneously distributed corruption, and 3) difficulty in corruption estimation when data cannot be entirely loaded. This paper proposes online and distributed robust regression approaches, both of which can concurrently address all the above challenges. Specifically, the distributed algorithm optimizes the regression coefficients of each data block via heuristic hard thresholding and combines all the estimates in a distributed robust consolidation. Furthermore, an online version of the distributed algorithm is proposed to incrementally update the existing estimates with new incoming data. We also prove that our algorithms benefit from strong robustness guarantees in terms of regression coefficient recovery with a constant upper bound on the error of state-of-the-art batch methods. Extensive experiments on synthetic and real datasets demonstrate that our approaches are superior to those of existing methods in effectiveness, with competitive efficiency.

Deriving event storylines is an effective summarization method to succinctly organize extensive information, which can significantly alleviate the pain of information overload. The critical challenge is the lack of widely recognized definition of storyline metric. Prior studies have developed various approaches based on different assumptions about users' interests. These works can extract interesting patterns, but their assumptions do not guarantee that the derived patterns will match users' preference. On the other hand, their exclusiveness of single modality source misses cross-modality information. This paper proposes a method, multimodal imitation learning via Generative Adversarial Networks(MIL-GAN), to directly model users' interests as reflected by various data. In particular, the proposed model addresses the critical challenge by imitating users' demonstrated storylines. Our proposed model is designed to learn the reward patterns given user-provided storylines and then applies the learned policy to unseen data. The proposed approach is demonstrated to be capable of acquiring the user's implicit intent and outperforming competing methods by a substantial margin with a user study.

The presence of data noise and corruptions recently invokes increasing attention on Robust Least Squares Regression (RLSR), which addresses the fundamental problem that learns reliable regression coefficients when response variables can be arbitrarily corrupted. Until now, several important challenges still cannot be handled concurrently: 1) exact recovery guarantee of regression coefficients 2) difficulty in estimating the corruption ratio parameter; and 3) scalability to massive dataset. This paper proposes a novel Robust Least squares regression algorithm via Heuristic Hard thresholding (RLHH), that concurrently addresses all the above challenges. Specifically, the algorithm alternately optimizes the regression coefficients and estimates the optimal uncorrupted set via heuristic hard thresholding without corruption ratio parameter until it converges. We also prove that our algorithm benefits from strong guarantees analogous to those of state-of-the-art methods in terms of convergence rates and recovery guarantees. We provide empirical evidence to demonstrate that the effectiveness of our new method is superior to that of existing methods in the recovery of both regression coefficients and uncorrupted sets, with very competitive efficiency.

Connecting the dots between diverse entities such as people and organizations is a vital task for forming hypotheses and uncovering latent relationships among complex and large datasets. Most existing approaches are designed to address the relationship of entities in news reports, documents and abstracts, but such approaches are not suitable for Twitter data streams due to their unstructured languages, short-length messages, heterogeneous features and massive size. The sheer size of Twitter data requires more efficient algorithms to connect the dots within a short period of time. We present a system that automatically constructs stories by connecting entities in Twitter datasets. An entity similarity model is designed that combines both traditional entity-related features and social network attributes and a novel story generation algorithm applied on the similarity model is proposed to cope with the massive Twitter datasets. Extensive experimental evaluations were conducted to demonstrate the effectiveness of this new approach.

Storyline detection aims to connect seemly irrelevant single documents into meaningful chains, which provides opportunities for understanding how events evolve over time and what triggers such evolutions. Most previous work generated the storylines through unsupervised methods that can hardly reveal underlying factors driving the evolution process. This paper introduces a Bayesian model to generate storylines from massive documents and infer the corresponding hidden relations and topics. In addition, our model is the first attempt that utilizes Twitter data as human input to ``supervise'' the generation of storylines. Through extensive experiments, we demonstrate our proposed model can achieve significant improvement over baseline methods and can be used to discover interesting patterns for real world cases.