IJCAI 2022 Papers

Graph representation learning has attracted tremendous attention due to its remarkable performance in many real-world applications. However, prevailing supervised graph representation learning models for specific tasks often suffer from label sparsity issue as data labeling is always time and resource consuming. In light of this, few-shot learning on graphs (FSLG), which combines the strengths of graph representation learning and few-shot learning together, has been proposed to tackle the performance degradation in face of limited annotated data challenge. There have been many studies working on FSLG recently. In this paper, we comprehensively survey these work in the form of a series of methods and applications. Specifically, we first introduce FSLG challenges and bases, then categorize and summarize existing work of FSLG in terms of three major graph mining tasks at different granularity levels, i.e., node, edge, and graph. Finally, we share our thoughts on some future research directions of FSLG. The authors of this survey have contributed significantly to the AI literature on FSLG over the last few years.

Deep networks are prone to performance degradation when there is a domain shift between the source (training) data and target (test) data. Recent test-time adaptation methods update batch normalization layers of pre-trained source models deployed in new target environments with streaming data. Although these methods can adapt on-the-fly without first collecting a large target domain dataset, their performance is dependent on streaming conditions such as mini-batch size and class-distribution which can be unpredictable in practice. In this work, we propose a framework for few-shot domain adaptation to address the practical challenges of data-efficient adaptation. Specifically, we propose a constrained optimization of feature normalization statistics in pre-trained source models supervised by a small target domain support set. Our method is easy to implement and improves source model performance with as little as one sample per class for classification tasks. Extensive experiments on 5 cross-domain classification and 4 semantic segmentation datasets show that our proposed method achieves more accurate and reliable performance than test-time adaptation, while not being constrained by streaming conditions.

Federated Learning (FL) is an emerging privacy-aware machine learning technique that applies successfully to the collaborative learning of global models for Human Activity Recognition (HAR). As of now, the applications of FL for HAR assume that the data associated with diverse individuals follow the same distribution. However, this assumption is impractical in real-world scenarios where the same activity is frequently performed differently by different individuals. To tackle this issue, we propose FedMAT, a Federated Multi-task ATtention framework for HAR, which extracts and fuses shared as well as individual-specific multi-modal sensor data features. Specifically, we treat the HAR problem associated with each individual as a different task and train a federated multi-task model, composed of a shared feature representation network in a central server plus multiple individual-specific networks with attention modules stored in decentralized nodes. In this architecture, the attention module operates as a mask that allows to learn individual-specific features from the global model, whilst simultaneously allowing for features to be shared among different individuals. We conduct extensive experiments based on publicly available HAR datasets, which are collected in both controlled environments and real-world scenarios. Numeric results verify that our proposed FedMAT significantly outperforms baselines not only in generalizing to existing individuals but also in adapting to new individuals.

Federated learning (FL) provides a privacy-preserving solution for distributed machine learning tasks. One challenging problem that severely damages the performance of FL models is the co-occurrence of data heterogeneity and long-tail distribution, which frequently appears in real FL applications. In this paper, we reveal an intriguing fact that the biased classifier is the primary factor leading to the poor performance of the global model. Motivated by the above finding, we propose a novel and privacy-preserving FL method for heterogeneous and long-tailed data via Classifier Re-training with Federated Features (CReFF). The classifier re-trained on federated features can produce comparable performance as the one re-trained on real data in a privacy-preserving manner without information leakage of local data or class distribution. Experiments on several benchmark datasets show that the proposed CReFF is an effective solution to obtain a promising FL model under heterogeneous and long-tailed data. Comparative results with the state-of-the-art FL methods also validate the superiority of CReFF. Our code is available at https://github.com/shangxinyi/CReFF-FL.

Despite achieving remarkable performance, Federated Learning (FL) suffers from two critical challenges, i.e., limited computational resources and low training efficiency. In this paper, we propose a novel FL framework, i.e., FedDUAP, with two original contributions, to exploit the insensitive data on the server and the decentralized data in edge devices to further improve the training efficiency. First, a dynamic server update algorithm is designed to exploit the insensitive data on the server, in order to dynamically determine the optimal steps of the server update for improving the convergence and accuracy of the global model. Second, a layer-adaptive model pruning method is developed to perform unique pruning operations adapted to the different dimensions and importance of multiple layers, to achieve a good balance between efficiency and effectiveness. By integrating the two original techniques together, our proposed FL model, FedDUAP, significantly outperforms baseline approaches in terms of accuracy (up to 4.8% higher), efficiency (up to 2.8 times faster), and computational cost (up to 61.9% smaller).

Federated learning (FL) aims to protect data privacy by enabling clients to build machine learning models collaboratively without sharing their private data. Recent works demonstrate that information exchanged during FL is subject to gradient-based privacy attacks and, consequently, a variety of privacy-preserving methods have been adopted to thwart such attacks. However, these defensive methods either introduce orders of magnitudes more computational and communication overheads (e.g., with homomorphic encryption) or incur substantial model performance losses in terms of prediction accuracy (e.g., with differential privacy). In this work, we propose FEDCG, a novel federated learning method that leverages conditional generative adversarial networks to achieve high-level privacy protection while still maintaining competitive model performance. FEDCG decomposes each client's local network into a private extractor and a public classifier and keeps the extractor local to protect privacy. Instead of exposing extractors, FEDCG shares clients' generators with the server for aggregating clients' shared knowledge aiming to enhance the performance of each client's local networks. Extensive experiments demonstrate that FEDCG can achieve competitive model performance compared with FL baselines, and privacy analysis shows that FEDCG has a high-level privacy-preserving capability.

Feature selection and instance selection are two important techniques of data processing. However, such selections have mostly been studied separately, while existing work towards the joint selection conducts feature/instance selection coarsely; thus neglecting the latent fine-grained interaction between feature space and instance space. To address this challenge, we propose a reinforcement learning solution to accomplish the joint selection task and simultaneously capture the interaction between the selection of each feature and each instance. In particular, a sequential-scanning mechanism is designed as action strategy of agents and a collaborative-changing environment is used to enhance agent collaboration. In addition, an interactive paradigm introduces prior selection knowledge to help agents for more efficient exploration. Finally, extensive experiments on real-world datasets have demonstrated improved performances.

Existing learning-based dehazing methods do not fully use non-local information, which makes the restoration of seriously degraded region very tough. We propose a novel dehazing network by defining the Feature Dense Relevance module (FDR) and the Shallow Feature Mapping module (SFM). The FDR is defined based on multi-head attention to construct the dense relationship between different local features in the whole image. It enables the network to restore the degraded local regions by non-local information in complex scenes. In addition, the raw distant skip-connection easily leads to artifacts while it cannot deal with the shallow features effectively. Therefore, we define the SFM by combining the atmospheric scattering model and the distant skip-connection to effectively deal with the shallow features in different scales. It not only maps the degraded textures into clear textures by distant dependence, but also reduces artifacts and color distortions effectively. We introduce contrastive loss and focal frequency loss in the network to obtain a realitic and clear image. The extensive experiments on several synthetic and real-world datasets demonstrate that our network surpasses most of the state-of-the-art methods.

Recent work for extracting relations from texts has achieved excellent performance. However, most existing methods pay less attention to the efficiency, making it still challenging to quickly extract relations from massive or streaming text data in realistic scenarios. The main efficiency bottleneck is that these methods use a Transformer-based pre-trained language model for encoding, which heavily affects the training speed and inference speed. To address this issue, we propose a fast relation extraction model (FastRE) based on convolutional encoder and improved cascade binary tagging framework. Compared to previous work, FastRE employs several innovations to improve efficiency while also keeping promising performance. Concretely, FastRE adopts a novel convolutional encoder architecture combined with dilated convolution, gated unit and residual connection, which significantly reduces the computation cost of training and inference, while maintaining the satisfactory performance. Moreover, to improve the cascade binary tagging framework, FastRE first introduces a type-relation mapping mechanism to accelerate tagging efficiency and alleviate relation redundancy, and then utilizes a position-dependent adaptive thresholding strategy to obtain higher tagging accuracy and better model generalization. Experimental results demonstrate that FastRE is well balanced between efficiency and performance, and achieves 3-10$\times$ training speed, 7-15$\times$ inference speed faster, and 1/100 parameters compared to the state-of-the-art models, while the performance is still competitive. Our code is available at \url{https://github.com/seukgcode/FastRE}.

Denoising diffusion probabilistic models (DDPMs) have recently achieved leading performances in many generative tasks. However, the inherited iterative sampling process costs hindered their applications to speech synthesis. This paper proposes FastDiff, a fast conditional diffusion model for high-quality speech synthesis. FastDiff employs a stack of time-aware location-variable convolutions of diverse receptive field patterns to efficiently model long-term time dependencies with adaptive conditions. A noise schedule predictor is also adopted to reduce the sampling steps without sacrificing the generation quality. Based on FastDiff, we design an end-to-end text-to-speech synthesizer, FastDiff-TTS, which generates high-fidelity speech waveforms without any intermediate feature (e.g., Mel-spectrogram). Our evaluation of FastDiff demonstrates the state-of-the-art results with higher-quality (MOS 4.28) speech samples. Also, FastDiff enables a sampling speed of 58x faster than real-time on a V100 GPU, making diffusion models practically applicable to speech synthesis deployment for the first time. We further show that FastDiff generalized well to the mel-spectrogram inversion of unseen speakers, and FastDiff-TTS outperformed other competing methods in end-to-end text-to-speech synthesis. Audio samples are available at https://FastDiff.github.io/.

On computing clusters, the autoscaler is responsible for allocating resources for jobs or fine-grained tasks to ensure their Quality of Service. Due to a more precise resource management, fine-grained autoscaling can generally achieve better performance. However, the fine-grained autoscaling for streaming jobs needs intensive computation to model the complicated running states of tasks, and has not been adequately studied previously. In this paper, we propose a novel fine-grained autoscaler for streaming jobs based on reinforcement learning. We first organize the running states of streaming jobs as spatio-temporal graphs. To efficiently make autoscaling decisions, we propose a Neural Variational Subgraph Sampler to sample spatio-temporal subgraphs. Furthermore, we propose a mutual-information-based objective function to explicitly guide the sampler to extract more representative subgraphs. After that, the autoscaler makes decisions based on the learned subgraph representations. Experiments conducted on real-world datasets demonstrate the superiority of our method over six competitive baselines.

Fallacies play a prominent role in argumentation since antiquity due to their contribution to argumentation in critical thinking education. Their role is even more crucial nowadays as contemporary argumentation technologies face challenging tasks as misleading and manipulative information detection in news articles and political discourse, and counter-narrative generation. Despite some work in this direction, the issue of classifying arguments as being fallacious largely remains a challenging and an unsolved task. Our contribution is twofold: first, we present a novel annotated resource of 31 political debates from the U.S. Presidential Campaigns, where we annotated six main categories of fallacious arguments (i.e., ad hominem, appeal to authority, appeal to emotion, false cause, slogan, slippery slope) leading to 1628 annotated fallacious arguments; second, we tackle this novel task of fallacious argument classification and we define a neural architecture based on transformers outperforming state-of-the-art results and standard baselines. Our results show the important role played by argument components and relations in this task.

In recent years, most fairness strategies in machine learning have focused on mitigating unwanted biases by assuming that the sensitive information is available. However, in practice this is not always the case: due to privacy purposes and regulations such as RGPD in EU, many personal sensitive attributes are frequently not collected. Yet, only a few prior works address the issue of mitigating bias in such a difficult setting, in particular to meet classical fairness objectives such as Demographic Parity and Equalized Odds. By leveraging recent developments for approximate inference, we propose in this paper an approach to fill this gap. To infer a sensitive information proxy, we introduce a new variational auto-encoding-based framework named SRCVAE that relies on knowledge of the underlying causal graph. The bias mitigation is then done in an adversarial fairness approach. Our proposed method empirically achieves significant improvements over existing works in the field. We observe that the generated proxy’s latent space correctly recovers sensitive information and that our approach achieves a higher accuracy while obtaining the same level of fairness on two real datasets.

Consider the practical goal of making a desired action profile played, when the planner can only change the payoffs, bound by stringent constraints. Applications include motivating people to choose the closest school, the closest subway station, or to coordinate on a communication protocol or an investment strategy. Employing subsidies and tolls, we adjust the game so that choosing this predefined action profile becomes strictly dominant. Inspired mainly by the work of Monderer and Tennenholtz, where the promised subsidies do not materialise in the not played profiles, we provide a fair and individually rational game adjustment, such that the total outside investments sum up to zero at any profile, thereby facilitating easy and frequent usage of our adjustment without bearing costs, even if some players behave unexpectedly. The resultant action profile itself needs no adjustment. Importantly, we also prove that our adjustment minimises the general transfer among all such adjustments, counting the total subsidising and taxation.

In this work we introduce a new class of mechanisms composed of a traditional Generalized Second Price (GSP) auction, and a fair division scheme in order to achieve some desired level of fairness between groups of Bayesian strategic advertisers. We propose two mechanisms, beta-Fair GSP and GSP-EFX, that compose GSP with, respectively, an envy-free up to one item, and an envy-free up to any item fair division scheme. The payments of GSP are adjusted in order to compensate advertisers that suffer a loss of efficiency due the fair division stage. We investigate the strategic learning implications of the deployment of sponsored search auction mechanisms that obey to such fairness criteria. We prove that, for both mechanisms, if bidders play so as to minimize their external regret they are guaranteed to reach an equilibrium with good social welfare. We also prove that the mechanisms are budget balanced, so that the payments charged by the traditional GSP mechanism are a good proxy of the total compensation offered to the advertisers. Finally, we evaluate the quality of the allocations through experiments on real-world data.

Allocating resources to individuals in a fair manner has been a topic of interest since the ancient times, with most of the early rigorous mathematical work on the problem focusing on infinitely divisible resources. Recently, there has been a surge of papers studying computational questions regarding various different notions of fairness for the indivisible case, like maximin share fairness (MMS) and envy-freeness up to any good (EFX). We survey the most important results in the discrete fair division literature, focusing on the case of additive valuation functions and paying particular attention to the progress made in the last 10 years.

Network quantization significantly reduces model inference complexity and has been widely used in real-world deployments. However, most existing quantization methods have been developed mainly on Convolutional Neural Networks (CNNs), and suffer severe degradation when applied to fully quantized vision transformers. In this work, we demonstrate that many of these difficulties arise because of serious inter-channel variation in LayerNorm inputs, and present, Power-of-Two Factor (PTF), a systematic method to reduce the performance degradation and inference complexity of fully quantized vision transformers. In addition, observing an extreme non-uniform distribution in attention maps, we propose Log-Int-Softmax (LIS) to sustain that and simplify inference by using 4-bit quantization and the BitShift operator. Comprehensive experiments on various transformer-based architectures and benchmarks show that our Fully Quantized Vision Transformer (FQ-ViT) outperforms previous works while even using lower bit-width on attention maps. For instance, we reach 84.89% top-1 accuracy with ViT-L on ImageNet and 50.8 mAP with Cascade Mask R-CNN (Swin-S) on COCO. To our knowledge, we are the first to achieve lossless accuracy degradation (~1%) on fully quantized vision transformers. The code is available at https://github.com/megvii-research/FQ-ViT.

Traffic flow forecasting plays a vital role in the transportation domain. Existing studies usually manually construct correlation graphs and design sophisticated models for learning spatial and temporal features to predict future traffic states. However, manually constructed correlation graphs cannot accurately extract the complex patterns hidden in the traffic data. In addition, it is challenging for the prediction model to fit traffic data due to its irregularly-shaped distribution. To solve the above-mentioned problems, in this paper, we propose a novel learning-based method to learn a spatial-temporal correlation graph, which could make good use of the traffic flow data. Moreover, we propose First-Order Gradient Supervision (FOGS), a novel method for traffic flow forecasting. FOGS utilizes first-order gradients, rather than specific flows, to train prediction model, which effectively avoids the problem of fitting irregularly-shaped distributions. Comprehensive numerical evaluations on four real-world datasets reveal that the proposed methods achieve state-of-the-art performance and significantly outperform the benchmarks.

The k-means problem is an extensively studied unsupervised learning problem with various applications in decision making and data mining. In this paper, we propose a fast and practical local search algorithm for the k-means problem. Our method reduces the search space of swap pairs from O(nk) to O(k^2), and applies random mutations to find potentially better solutions when local search falls into poor local optimum. With the assumption of data distribution that each optimal cluster has "average" size of \Omega(n/k), which is common in many datasets and k-means benchmarks, we prove that our proposed algorithm gives a (100+\epsilon)-approximate solution in expectation. Empirical experiments show that our algorithm achieves better performance compared to existing state-of-the-art local search methods on k-means benchmarks and large datasets.