Search

Factuality evaluation aims to detect factual errors produced by language models (LMs) and hence guide the development of more factual models. Towards this goal, we train a factuality evaluator, FenCE, that provides LM generators with claim-level factuality feedback. We conduct data augmentation on a combination of public judgment datasets to train FenCE to (1) generate textual critiques along with scores and (2) make claim-level judgment based on diverse source documents obtained by various tools. We then present a framework that leverages FenCE to improve the factuality of LM generators by constructing training data. Specifically, we generate a set of candidate responses, leverage FenCE to revise and score each response without introducing lesser-known facts, and train the generator by preferring highly scored revised responses. Experiments show that our data augmentation methods improve the evaluator's accuracy by 2.9% on LLM-AggreFact. With FenCE, we improve Llama3-8B-chat's factuality rate by 14.45% on FActScore, outperforming state-of-the-art factuality finetuning methods by 6.96%.

Large Language Models (LLMs) have demonstrated impressive capabilities in various tasks, including instruction following, which is crucial for aligning model outputs with user expectations. However, evaluating LLMs' ability to follow instructions remains challenging due to the complexity and subjectivity of human language. Current benchmarks primarily focus on single-turn, monolingual instructions, which do not adequately reflect the complexities of real-world applications that require handling multi-turn and multilingual interactions. To address this gap, we introduce Multi-IF, a new benchmark designed to assess LLMs' proficiency in following multi-turn and multilingual instructions. Multi-IF, which utilizes a hybrid framework combining LLM and human annotators, expands upon the IFEval by incorporating multi-turn sequences and translating the English prompts into another 7 languages, resulting in a dataset of 4,501 multilingual conversations, where each has three turns. Our evaluation of 14 state-of-the-art LLMs on Multi-IF reveals that it presents a significantly more challenging task than existing benchmarks. All the models tested showed a higher rate of failure in executing instructions correctly with each additional turn. For example, o1-preview drops from 0.877 at the first turn to 0.707 at the third turn in terms of average accuracy over all languages. Moreover, languages with non-Latin scripts (Hindi, Russian, and Chinese) generally exhibit higher error rates, suggesting potential limitations in the models' multilingual capabilities. We release Multi-IF prompts and the evaluation code base to encourage further research in this critical area.

Self-anthropomorphism in robots manifests itself through their display of human-like characteristics in dialogue, such as expressing preferences and emotions. Our study systematically analyzes self-anthropomorphic expression within various dialogue datasets, outlining the contrasts between self-anthropomorphic and non-self-anthropomorphic responses in dialogue systems. We show significant differences in these two types of responses and propose transitioning from one type to the other. We also introduce Pix2Persona, a novel dataset aimed at developing ethical and engaging AI systems in various embodiments. This dataset preserves the original dialogues from existing corpora and enhances them with paired responses: self-anthropomorphic and non-self-anthropomorphic for each original bot response. Our work not only uncovers a new category of bot responses that were previously under-explored but also lays the groundwork for future studies about dynamically adjusting self-anthropomorphism levels in AI systems to align with ethical standards and user expectations., Comment: Findings of EMNLP 2024, 19 pages

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications., Comment: submitted to conference

The outstanding lubrication of articular cartilage in the major synovial joints such as hips and knees, essential for the joint well-being, has been attributed to boundary layers of lipids at the outer cartilage surfaces, which have very low friction mediated by the hydration lubrication mechanism at their highly hydrated exposed headgroups. However, the role of spontaneously present lipid splays, lipids with an acyl tail in each of the opposing bilayers, in modulating the frictional force between lipid bilayers has not, to date, been considered. In this study, we perform all-atom molecular dynamics simulations to quantitatively assess the significance of splayed molecules within the framework of lubricating lipid bilayers. We demonstrate that, although transient, splayed molecules significantly increase the inter-membrane friction until their retraction back into the lamellar phase, with this effect more steadily occurring at lower sliding velocities that are comparable to the physiological velocities of sliding articular cartilage.

Fusion of lipid bilayers in membranes is important in processes from vesicle-cell interactions (as in drug delivery) to exosome-cell signaling, while transient transmembrane electric fields are known to occur spontaneously. Two contacting phosphatidylcholine (PC) lipid membranes are known to fuse into one under external electric fields, suggesting that the interaction between them is modified by the field as they approach, prior to the fusion event. Here we measure directly the adhesion energy between dimyristoylphosphatidylcholine (DMPC) and between distearoylphosphatidylcholine (DSPC) surface layers attached to solid substrates both without and with a transmembrane electric field. We find a marked pressure-dependent adhesion behavior in the electric field, which we attribute to fusion intermediates that are formed, shedding new light on membrane electro-fusion.

Transformer has become one of the most popular architectures for multivariate time series (MTS) forecasting. Recent Transformer-based MTS models generally prefer channel-independent structures with the observation that channel independence can alleviate noise and distribution drift issues, leading to more robustness. Nevertheless, it is essential to note that channel dependency remains an inherent characteristic of MTS, carrying valuable information. Designing a model that incorporates merits of both channel-independent and channel-mixing structures is a key to further improvement of MTS forecasting, which poses a challenging conundrum. To address the problem, an injection method for global information into channel-independent Transformer, InjectTST, is proposed in this paper. Instead of designing a channel-mixing model directly, we retain the channel-independent backbone and gradually inject global information into individual channels in a selective way. A channel identifier, a global mixing module and a self-contextual attention module are devised in InjectTST. The channel identifier can help Transformer distinguish channels for better representation. The global mixing module produces cross-channel global information. Through the self-contextual attention module, the independent channels can selectively concentrate on useful global information without robustness degradation, and channel mixing is achieved implicitly. Experiments indicate that InjectTST can achieve stable improvement compared with state-of-the-art models.

Despite vision-language models' (VLMs) remarkable capabilities as versatile visual assistants, two substantial challenges persist within the existing VLM frameworks: (1) lacking task diversity in pretraining and visual instruction tuning, and (2) annotation error and bias in GPT-4 synthesized instruction tuning data. Both challenges lead to issues such as poor generalizability, hallucination, and catastrophic forgetting. To address these challenges, we construct Vision-Flan, the most diverse publicly available visual instruction tuning dataset to date, comprising 187 diverse tasks and 1,664,261 instances sourced from academic datasets, and each task is accompanied by an expert-written instruction. In addition, we propose a two-stage instruction tuning framework, in which VLMs are firstly finetuned on Vision-Flan and further tuned on GPT-4 synthesized data. We find this two-stage tuning framework significantly outperforms the traditional single-stage visual instruction tuning framework and achieves the state-of-the-art performance across a wide range of multi-modal evaluation benchmarks. Finally, we conduct in-depth analyses to understand visual instruction tuning and our findings reveal that: (1) GPT-4 synthesized data does not substantially enhance VLMs' capabilities but rather modulates the model's responses to human-preferred formats; (2) A minimal quantity (e.g., 1,000) of GPT-4 synthesized data can effectively align VLM responses with human-preference; (3) Visual instruction tuning mainly helps large-language models (LLMs) to understand visual features., Comment: 8 Pages, visual instruction tuning

Graph neural networks (GNNs) have found widespread application in modeling graph data across diverse domains. While GNNs excel in scenarios where the testing data shares the distribution of their training counterparts (in distribution, ID), they often exhibit incorrect predictions when confronted with samples from an unfamiliar distribution (out-of-distribution, OOD). To identify and reject OOD samples with GNNs, recent studies have explored graph OOD detection, often focusing on training a specific model or modifying the data on top of a well-trained GNN. Despite their effectiveness, these methods come with heavy training resources and costs, as they need to optimize the GNN-based models on training data. Moreover, their reliance on modifying the original GNNs and accessing training data further restricts their universality. To this end, this paper introduces a method to detect Graph Out-of-Distribution At Test-time (namely GOODAT), a data-centric, unsupervised, and plug-and-play solution that operates independently of training data and modifications of GNN architecture. With a lightweight graph masker, GOODAT can learn informative subgraphs from test samples, enabling the capture of distinct graph patterns between OOD and ID samples. To optimize the graph masker, we meticulously design three unsupervised objective functions based on the graph information bottleneck principle, motivating the masker to capture compact yet informative subgraphs for OOD detection. Comprehensive evaluations confirm that our GOODAT method outperforms state-of-the-art benchmarks across a variety of real-world datasets. The code is available at Github: https://github.com/Ee1s/GOODAT, Comment: 9 pages, 5 figures

Learning from human feedback is a prominent technique to align the output of large language models (LLMs) with human expectations. Reinforcement learning from human feedback (RLHF) leverages human preference signals that are in the form of ranking of response pairs to perform this alignment. However, human preference on LLM outputs can come in much richer forms including natural language, which may provide detailed feedback on strengths and weaknesses of a given response. In this work we investigate data efficiency of modeling human feedback that is in natural language. Specifically, we fine-tune an open-source LLM, e.g., Falcon-40B-Instruct, on a relatively small amount (1000 records or even less) of human feedback in natural language in the form of critiques and revisions of responses. We show that this model is able to improve the quality of responses from even some of the strongest LLMs such as ChatGPT, BARD, and Vicuna, through critique and revision of those responses. For instance, through one iteration of revision of ChatGPT responses, the revised responses have 56.6% win rate over the original ones, and this win rate can be further improved to 65.9% after applying the revision for five iterations., Comment: Accepted by Workshop on Instruction Tuning and Instruction Following at NeurIPS 2023, Submitted to AAAI 2024

As information filtering services, recommender systems have extremely enriched our daily life by providing personalized suggestions and facilitating people in decision-making, which makes them vital and indispensable to human society in the information era. However, as people become more dependent on them, recent studies show that recommender systems potentially own unintentional impacts on society and individuals because of their unfairness (e.g., gender discrimination in job recommendations). To develop trustworthy services, it is crucial to devise fairness-aware recommender systems that can mitigate these bias issues. In this survey, we summarise existing methodologies and practices of fairness in recommender systems. Firstly, we present concepts of fairness in different recommendation scenarios, comprehensively categorize current advances, and introduce typical methods to promote fairness in different stages of recommender systems. Next, after introducing datasets and evaluation metrics applied to assess the fairness of recommender systems, we will delve into the significant influence that fairness-aware recommender systems exert on real-world industrial applications. Subsequently, we highlight the connection between fairness and other principles of trustworthy recommender systems, aiming to consider trustworthiness principles holistically while advocating for fairness. Finally, we summarize this review, spotlighting promising opportunities in comprehending concepts, frameworks, the balance between accuracy and fairness, and the ties with trustworthiness, with the ultimate goal of fostering the development of fairness-aware recommender systems., Comment: 27 pages, 9 figures

Transient electric fields across cell bilayer membranes can lead to electroporation, as well as to cell fusion, and have been extensively studied. We find that transmembrane electric fields similar to those in cells can lead to a massive, reversible modulation--by up to 200-fold--of the interfacial energy dissipation between surfaces sliding across the lipid bilayer membranes. Atomistic simulations reveal that this arises from (fully reversible) electroporation of the interfacially-confined bilayers, and formation of bilayer bridges analogous to stalks preceding intermembrane fusion. These cell-membrane-mimicking effects topologically-force the slip to partially-revert from the low-dissipation, hydrated lipid-headgroups plane to the intra-bilayer, high-dissipation acyl tail interface. Our results demonstrate that lipid bilayers under transmembrane electric fields can have striking materials-modification properties, and shed new light on membrane hemifusion., Comment: Yu Zhang and Di Jin contributes equally to this work

Chain-of-Thought (CoT) prompting enables large language models to solve complex reasoning problems by generating intermediate steps. However, confined by its inherent single-pass and sequential generation process, CoT heavily relies on the initial decisions, causing errors in early steps to accumulate and impact the final answers. In contrast, humans adopt recursive thinking when tackling complex reasoning problems, i.e., iteratively breaking the original problem into approachable sub-problems and aggregating their answers to resolve the original one. Inspired by the human cognitive process, we propose SOCRATIC QUESTIONING, a divide-and-conquer style algorithm that mimics the recursive thinking process. Specifically, SOCRATIC QUESTIONING leverages large language models to raise and answer sub-questions until collecting enough information to tackle the original question. Unlike CoT, SOCRATIC QUESTIONING explicitly navigates the thinking space, stimulates effective recursive thinking, and is more robust towards errors in the thinking process. Extensive experiments on several complex reasoning tasks, including MMLU, MATH, LogiQA, and visual question-answering demonstrate significant performance improvements over the state-of-the-art prompting methods, such as CoT, and Tree-of-Thought. The qualitative analysis clearly shows that the intermediate reasoning steps elicited by SOCRATIC QUESTIONING are similar to humans' recursively thinking process of complex reasoning problems., Comment: ACL 2023 main

The very low sliding friction of articular cartilage in the major synovial joints such as hips and knees is crucial for their well-being, and has been attributed to lubrication by phospholipid boundary layers. While single-component lipid layers have demonstrated efficient lubricity in model studies, in living joints there is a large number of different lipids, raising the question of whether this is natural redundancy, or whether this multiplicity confers any benefits. Here we examine lubrication by progressively more complex mixtures of lipids representative of those in joints, using a surface forces balance at physiologically-relevant salt concentrations and pressures. We find that different lipid combinations differ very significantly in their lubricating ability, as manifested by their robustness to hemifusion under physiological loads, pointing to a clear lubrication synergy arising from multiple lipid types in the lubricating layers. Insight into the origins of this synergy is provided by molecular dynamics (MD) simulations of the different lipid mixtures used in the experiments, which directly reveal how hemifusion - associated with greatly increased friction - depends on the detailed lipid composition. Our results provide insight into the role of lipid type proliferation in healthy synovial joints, and point to new treatment modalities for osteoarthritis.

Task-oriented Dialogue (TOD) Systems aim to build dialogue systems that assist users in accomplishing specific goals, such as booking a hotel or a restaurant. Traditional TODs rely on domain-specific APIs/DBs or external factual knowledge to generate responses, which cannot accommodate subjective user requests (e.g., "Is the WIFI reliable?" or "Does the restaurant have a good atmosphere?"). To address this issue, we propose a novel task of subjective-knowledge-based TOD (SK-TOD). We also propose the first corresponding dataset, which contains subjective knowledge-seeking dialogue contexts and manually annotated responses grounded in subjective knowledge sources. When evaluated with existing TOD approaches, we find that this task poses new challenges such as aggregating diverse opinions from multiple knowledge snippets. We hope this task and dataset can promote further research on TOD and subjective content understanding. The code and the dataset are available at https://github.com/alexa/dstc11-track5., Comment: SIGDIAL 2023

Recommender systems are essential to various fields, e.g., e-commerce, e-learning, and streaming media. At present, graph neural networks (GNNs) for session-based recommendations normally can only recommend items existing in users' historical sessions. As a result, these GNNs have difficulty recommending items that users have never interacted with (new items), which leads to a phenomenon of information cocoon. Therefore, it is necessary to recommend new items to users. As there is no interaction between new items and users, we cannot include new items when building session graphs for GNN session-based recommender systems. Thus, it is challenging to recommend new items for users when using GNN-based methods. We regard this challenge as '\textbf{G}NN \textbf{S}ession-based \textbf{N}ew \textbf{I}tem \textbf{R}ecommendation (GSNIR)'. To solve this problem, we propose a dual-intent enhanced graph neural network for it. Due to the fact that new items are not tied to historical sessions, the users' intent is difficult to predict. We design a dual-intent network to learn user intent from an attention mechanism and the distribution of historical data respectively, which can simulate users' decision-making process in interacting with a new item. To solve the challenge that new items cannot be learned by GNNs, inspired by zero-shot learning (ZSL), we infer the new item representation in GNN space by using their attributes. By outputting new item probabilities, which contain recommendation scores of the corresponding items, the new items with higher scores are recommended to users. Experiments on two representative real-world datasets show the superiority of our proposed method. The case study from the real-world verifies interpretability benefits brought by the dual-intent module and the new item reasoning module. The code is available at Github: https://github.com/Ee1s/NirGNN, Comment: 10 Pages, 6 figures, WWW'2023

Graph Neural Networks (GNNs), a type of neural network that can learn from graph-structured data through neighborhood information aggregation, have shown superior performance in various downstream tasks. However, as the number of layers increases, node representations become indistinguishable, which is known as over-smoothing. To address this issue, many residual methods have emerged. In this paper, we focus on the over-smoothing issue and related residual methods. Firstly, we revisit over-smoothing from the perspective of overlapping neighborhood subgraphs, and based on this, we explain how residual methods can alleviate over-smoothing by integrating multiple orders neighborhood subgraphs to avoid the indistinguishability of the single high-order neighborhood subgraphs. Additionally, we reveal the drawbacks of previous residual methods, such as the lack of node adaptability and severe loss of high-order neighborhood subgraph information, and propose a \textbf{Posterior-Sampling-based, Node-Adaptive Residual module (PSNR)}. We theoretically demonstrate that PSNR can alleviate the drawbacks of previous residual methods. Furthermore, extensive experiments verify the superiority of the PSNR module in fully observed node classification and missing feature scenarios. Our code is available at https://github.com/jingbo02/PSNR-GNN., Comment: NeurIPS2024

It was recently discovered that friction between surfaces bearing phosphatidylcholine (PC) lipid bilayers can be increased by two orders of magnitude or more via an externally-applied electric field, and that this increase is fully reversible when the field is switched off. While this striking effect holds promising application potential, its molecular origin remains unknown due to difficulty in experimentally probing confined membrane structure at a molecular level. Our earlier molecular dynamics simulations revealed the equilibrium electroporated structure of such confined lipid membranes under an electric field; here we extend this approach to study the associated sliding friction between two solid surfaces across such PC bilayers. We identify the enhanced friction in the field as arising from membrane undulations due to the electroporation; this leads to some dehydration at the lipid-water interfaces, leading to closer contact and thus increased attraction between the zwitterionic headgroups, which results in increased frictional dissipation between the bilayers as they slide past each other. Additionally, the electric field facilitates formation of lipid bridges spanning the intersurface gap; at the sliding velocities of the experiments, these bridges increase the friction by topologically-forcing the slip-plane to pass through the acyl tail-tail interface, associated with higher dissipation during sliding. Our results account quantitatively for the experimentally-observed electro-modulated friction with boundary lipid bilayers, and indicate more generally how they may affect interactions between contacting surfaces, where high local transverse fields may be ubiquitous.

It is well known that lipid membranes respond to a threshold transmembrane electric field through a reversible mechanism called electroporation, where hydrophilic water pores form across the membrane, an effect widely used in biological systems. The effect of such fields on interfacially-confined (stacked or supported) lipid membranes, on the other hand, which may strongly modulate interfacial properties, has not to our knowledge been previously studied. Motivated by recent surface forces experiments showing a striking effect of electric fields on lubrication by confined lipid bilayers, we carried out all-atom molecular dynamics simulations of such membranes under transverse electric fields. We find that in addition to electroporation, a new feature emerges of locally merged bilayers which act to bridge the confining interfaces. These features shed light on the remodelling of confined lipid membrane stacks by electric fields, and provides insight into how such fields may modulate frictional and more generally surface interactions in the presence of lipid-based boundary layers.

Social Internet of Things (SIoT), a promising and emerging paradigm that injects the notion of social networking into smart objects (i.e., things), paving the way for the next generation of Internet of Things. However, due to the risks and uncertainty, a crucial and urgent problem to be settled is establishing reliable relationships within SIoT, that is, trust evaluation. Graph neural networks for trust evaluation typically adopt a straightforward way such as one-hot or node2vec to comprehend node characteristics, which ignores the valuable semantic knowledge attached to nodes. Moreover, the underlying structure of SIoT is usually complex, including both the heterogeneous graph structure and pairwise trust relationships, which renders hard to preserve the properties of SIoT trust during information propagation. To address these aforementioned problems, we propose a novel knowledge-enhanced graph neural network (KGTrust) for better trust evaluation in SIoT. Specifically, we first extract useful knowledge from users' comment behaviors and external structured triples related to object descriptions, in order to gain a deeper insight into the semantics of users and objects. Furthermore, we introduce a discriminative convolutional layer that utilizes heterogeneous graph structure, node semantics, and augmented trust relationships to learn node embeddings from the perspective of a user as a trustor or a trustee, effectively capturing multi-aspect properties of SIoT trust during information propagation. Finally, a trust prediction layer is developed to estimate the trust relationships between pairwise nodes. Extensive experiments on three public datasets illustrate the superior performance of KGTrust over state-of-the-art methods., Comment: Accepted by WWW-23

This work focuses on in-context data augmentation for intent detection. Having found that augmentation via in-context prompting of large pre-trained language models (PLMs) alone does not improve performance, we introduce a novel approach based on PLMs and pointwise V-information (PVI), a metric that can measure the usefulness of a datapoint for training a model. Our method first fine-tunes a PLM on a small seed of training data and then synthesizes new datapoints - utterances that correspond to given intents. It then employs intent-aware filtering, based on PVI, to remove datapoints that are not helpful to the downstream intent classifier. Our method is thus able to leverage the expressive power of large language models to produce diverse training data. Empirical results demonstrate that our method can produce synthetic training data that achieve state-of-the-art performance on three challenging intent detection datasets under few-shot settings (1.28% absolute improvement in 5-shot and 1.18% absolute in 10-shot, on average) and perform on par with the state-of-the-art in full-shot settings (within 0.01% absolute, on average)., Comment: Accepted at EACL 2023

While large neural-based conversational models have become increasingly proficient dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, which often perpetuates social biases or stereotypes. We investigate a retrieval-based method for reducing bias and toxicity in responses from chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe responses to similar dialogue contexts. We find our method performs competitively with strong baselines without requiring training. For instance, using automatic evaluation, we find our best fine-tuned baseline only generates safe responses to unsafe dialogue contexts from DiaSafety 4.04% more than our approach. Finally, we also propose a re-ranking procedure which can further improve response safeness., Comment: Findings of EMNLP 2023

Dialogue state tracking (DST) is an important step in dialogue management to keep track of users' beliefs. Existing works fine-tune all language model (LM) parameters to tackle the DST task, which requires significant data and computing resources for training and hosting. The cost grows exponentially in the real-world deployment where dozens of fine-tuned LM are used for different domains and tasks. To reduce parameter size and better utilize cross-task shared information, we propose to use soft prompt token embeddings to learn task properties. Without tuning LM parameters, our method drastically reduces the number of parameters needed to less than 0.5% of prior works while achieves better low-resource DST performance., Comment: In the INTERSPEECH 2023, and the Second Workshop on Efficient Natural Language and Speech Processing (ENLSP) at NeurIPS 2022

Graph Neural Networks (GNNs) have been a prevailing technique for tackling various analysis tasks on graph data. A key premise for the remarkable performance of GNNs relies on complete and trustworthy initial graph descriptions (i.e., node features and graph structure), which is often not satisfied since real-world graphs are often incomplete due to various unavoidable factors. In particular, GNNs face greater challenges when both node features and graph structure are incomplete at the same time. The existing methods either focus on feature completion or structure completion. They usually rely on the matching relationship between features and structure, or employ joint learning of node representation and feature (or structure) completion in the hope of achieving mutual benefit. However, recent studies confirm that the mutual interference between features and structure leads to the degradation of GNN performance. When both features and structure are incomplete, the mismatch between features and structure caused by the missing randomness exacerbates the interference between the two, which may trigger incorrect completions that negatively affect node representation. To this end, in this paper we propose a general GNN framework based on teacher-student distillation to improve the performance of GNNs on incomplete graphs, namely T2-GNN. To avoid the interference between features and structure, we separately design feature-level and structure-level teacher models to provide targeted guidance for student model (base GNNs, such as GCN) through distillation. Then we design two personalized methods to obtain well-trained feature and structure teachers. To ensure that the knowledge of the teacher model is comprehensively and effectively distilled to the student model, we further propose a dual distillation mode to enable the student to acquire as much expert knowledge as possible., Comment: Accepted by AAAI23