IJCAI 2022 Papers

We tackle an extreme form of imbalanced classification, with up to 105 features but as few as 5 samples from the minority class. This problem occurs in predicting predicting tumor types and fraud detection, among others. Standard imbalanced classification methods are not designed for such severe data scarcity. Sampling-based methods need too many samples due to the high-dimensionality, while cost-based methods must place too high a weight on the limited minority samples. Our proposed method, called DIRECT, bypasses sample generation by training the classifier over a robust smoothed distribution of the minority class. DIRECT is fast, simple, robust, parameter-free, and easy to interpret. We validate DIRECT on several real-world datasets spanning document, image, and medical classification. DIRECT is up to 5x − 7x better than SMOTE-like methods, 30−200% better than ensemble methods, 3x − 7x better than cost-sensitive methods. The greatest gains are for settings with the fewest samples in the minority class, where DIRECT’s robustness is most helpful.

Fine-tuning pretrained language models (PLMs) on downstream tasks has become common practice in natural language processing. However, most of the PLMs are vulnerable, e.g., they are brittle under adversarial attacks or imbalanced data, which hinders the application of the PLMs on some downstream tasks, especially in safe-critical scenarios. In this paper, we propose a simple yet effective fine-tuning method called Match-Tuning to force the PLMs to be more robust. For each instance in a batch, we involve other instances in the same batch to interact with it. To be specific, regarding the instances with other labels as a perturbation, Match-Tuning makes the model more robust to noise at the beginning of training. While nearing the end, Match-Tuning focuses more on performing an interpolation among the instances with the same label for better generalization. Extensive experiments on various tasks in GLUE benchmark show that Match-Tuning consistently outperforms the vanilla fine-tuning by 1.64 scores. Moreover, Match-Tuning exhibits remarkable robustness to adversarial attacks and data imbalance.

Graph neural networks (GNNs) have emerged as powerful approaches for graph representation learning and node classification. Nevertheless, they can be vulnerable (sensitive) to link perturbations due to structural noise or adversarial attacks. This paper introduces RoboGNN, a novel framework that simultaneously robustifies an input classifier to a counterpart with certifiable robustness, and suggests desired graph representation with auxiliary links to ensure the robustness guarantee. (1) We introduce (p,θ)-robustness, which characterizes the robustness guarantee of a GNN-based classifier if its performance is insensitive for at least θ fraction of a targeted set of nodes under any perturbation of a set of vulnerable links up to a bounded size p. (2) We present a co-learning framework that interacts model learning with graph structural learning to robustify an input model M to a (p,θ)-robustness counterpart. The framework also outputs the desired graph structures that ensure the robustness. Using real-world benchmark graphs, we experimentally verify that roboGNN can effectively robustify representative GNNs with guaranteed robustness, and desirable gains on accuracy.

Graph contrastive learning has gained significant progress recently. However, existing works have rarely explored non-aligned node-node contrasting. In this paper, we propose a novel graph contrastive learning method named RoSA that focuses on utilizing non-aligned augmented views for node-level representation learning. First, we leverage the earth mover's distance to model the minimum effort to transform the distribution of one view to the other as our contrastive objective, which does not require alignment between views. Then we introduce adversarial training as an auxiliary method to increase sampling diversity and enhance the robustness of our model. Experimental results show that RoSA outperforms a series of graph contrastive learning frameworks on homophilous, non-homophilous and dynamic graphs, which validates the effectiveness of our work. To the best of our awareness, RoSA is the first work focuses on the non-aligned node-node graph contrastive learning problem. Our codes are available at: https://github.com/ZhuYun97/RoSA

Recent theoretical studies illustrated that kernel ridgeless regression can guarantee good generalization ability without an explicit regularization. In this paper, we investigate the statistical properties of ridgeless regression with random features and stochastic gradient descent. We explore the effect of factors in the stochastic gradient and random features, respectively. Specifically, random features error exhibits the double-descent curve. Motivated by the theoretical findings, we propose a tunable kernel algorithm that optimizes the spectral density of kernel during training. Our work bridges the interpolation theory and practical algorithm.

Revision by Comparison (RbC) is a non-prioritized belief revision mechanism on epistemic states that specifies constraints on the plausibility of an input sentence via a designated reference sentence, allowing for kind of relative belief revision. In this paper, we make the strategy underlying RbC more explicit and transfer the mechanism together with its intuitive strengths to a semi-quantitative framework based on ordinal conditional functions where a more elegant implementation of RbC is possible. We furthermore show that RbC can be realized as an iterated revision by so-called weak conditionals. Finally, we point out relations of RbC to credibility-limited belief revision, illustrating the versatility of RbC for advanced belief revision operations.

The causal mechanism between climate and political violence is fraught with complex mechanisms. Current quantitative causal models rely on one or more assumptions: (1) the climate drivers persistently generate conflict, (2) the causal mechanisms have a linear relationship with the conflict generation parameter, and/or (3) there is sufficient data to inform the prior distribution. Yet, we know conflict drivers often excite a social transformation process which leads to violence (e.g., drought forces agricultural producers to join urban militia), but further climate effects do not necessarily contribute to further violence. Therefore, not only is this bifurcation relationship highly non-linear, there is also often a lack of data to support prior assumptions for high resolution modeling. Here, we aim to overcome the aforementioned causal modeling challenges by proposing a neural forward-intensity Poisson process (NFIPP) model. The NFIPP is designed to capture the potential non-linear causal mechanism in climate induced political violence, whilst being robust to sparse and timing-uncertain data. Our results span 20 recent years and reveal an excitation-based causal link between extreme climate events and political violence across diverse countries. Our climate-induced conflict model results are cross-validated against qualitative climate vulnerability indices. Furthermore, we label historical events that either improve or reduce our predictability gain, demonstrating the importance of domain expertise in informing interpretation.

We argue that the present setting of semisupervised learning on graphs may result in unfair comparisons, due to its potential risk of over-tuning hyper-parameters for models. In this paper, we highlight the significant influence of tuning hyper-parameters, which leverages the label information in the validation set to improve the performance. To explore the limit of over-tuning hyperparameters, we propose ValidUtil, an approach to fully utilize the label information in the validation set through an extra group of hyper-parameters. With ValidUtil, even GCN can easily get high accuracy of 85.8% on Cora. To avoid over-tuning, we merge the training set and the validation set and construct an i.i.d. graph benchmark (IGB) consisting of 4 datasets. Each dataset contains 100 i.i.d. graphs sampled from a large graph to reduce the evaluation variance. Our experiments suggest that IGB is a more stable benchmark than previous datasets for semisupervised learning on graphs. Our code and data are released at https://github.com/THUDM/IGB/.

Graph Contrastive Learning (GCL) has proven highly effective in promoting the performance of Semi-Supervised Node Classification (SSNC). However, existing GCL methods are generally transferred from other fields like CV or NLP, whose underlying working mechanism remains underexplored. In this work, we first deeply probe the working mechanism of GCL in SSNC, and find that the promotion brought by GCL is severely unevenly distributed: the improvement mainly comes from subgraphs with less annotated information, which is fundamentally different from contrastive learning in other fields. However, existing GCL methods generally ignore this uneven distribution of annotated information and apply GCL evenly to the whole graph. To remedy this issue and further improve GCL in SSNC, we propose the Topology InFormation gain-Aware Graph Contrastive Learning (TIFA-GCL) framework that considers the annotated information distribution across graph in GCL. Extensive experiments on six benchmark graph datasets, including the enormous OGB-Products graph, show that TIFA-GCL can bring a larger improvement than existing GCL methods in both transductive and inductive settings. Further experiments demonstrate the generalizability and interpretability of TIFA-GCL.

InfoNCE is a widely used contrastive training loss. It aims to estimate the mutual information between a pair of variables by discriminating between each positive pair and its associated K negative pairs. It is proved that when the sample labels are clean, the lower bound of mutual information estimation is tighter when more negative samples are incorporated, which usually yields better model performance. However, in practice the labels often contain noise, and incorporating too many noisy negative samples into model training may be suboptimal. In this paper, we study how many negative samples are optimal for InfoNCE in different scenarios via a semi-quantitative theoretical framework. More specifically, we first propose a probabilistic model to analyze the influence of the negative sampling ratio K on training sample informativeness. Then, we design a training effectiveness function to measure the overall influence of training samples based on their informativeness. We estimate the optimal negative sampling ratio using the K value that maximizes the training effectiveness function. Based on our framework, we further propose an adaptive negative sampling method that can dynamically adjust the negative sampling ratio to improve InfoNCE-based model training. Extensive experiments in three different tasks show our framework can accurately predict the optimal negative sampling ratio, and various models can benefit from our adaptive negative sampling method.

Challenges in image aesthetics assessment (IAA) arise from that images of different themes correspond to different evaluation criteria, and learning aesthetics directly from images while ignoring the impact of theme variations on human visual perception inhibits the further development of IAA; however, existing IAA datasets and models overlook this problem. To address this issue, we show that a theme-oriented dataset and model design are effective for IAA. Specifically, 1) we elaborately build a novel dataset, called TAD66K, that contains 66K images covering 47 popular themes, and each image is densely annotated by more than 1200 people with dedicated theme evaluation criteria. 2) We develop a baseline model, TANet, which can effectively extract theme information and adaptively establish perception rules to evaluate images with different themes. 3) We develop a large-scale benchmark (the most comprehensive thus far) by comparing 17 methods with TANet on three representative datasets: AVA, FLICKR-AES and the proposed TAD66K, TANet achieves state-of-the-art performance on all three datasets. Our work offers the community an opportunity to explore more challenging directions; the code, dataset and supplementary material are available at https://github.com/woshidandan/TANet.

This paper is concerned with contrastive learning (CL) for low-level image restoration and enhancement tasks. We propose a new label-efficient learning paradigm based on residuals, residual contrastive learning (RCL), and derive an unsupervised visual representation learning framework, suitable for low-level vision tasks with noisy inputs. While supervised image reconstruction aims to minimize residual terms directly, RCL alternatively builds a connection between residuals and CL by defining a novel instance discrimination pretext task, using residuals as the discriminative feature. Our formulation mitigates the severe task misalignment between instance discrimination pretext tasks and downstream image reconstruction tasks, present in existing CL frameworks. Experimentally, we find that RCL can learn robust and transferable representations that improve the performance of various downstream tasks, such as denoising and super resolution, in comparison with recent self-supervised methods designed specifically for noisy inputs. Additionally, our unsupervised pre-training can significantly reduce annotation costs whilst maintaining performance competitive with fully-supervised image reconstruction.

In the context of aggregating intervals reflecting the views of several agents into a single interval, we investigate the impact of the form of representation chosen for the intervals involved. Specifically, we ask whether there are natural rules we can define both as rules that aggregate separately the left and right endpoints of intervals and as rules that aggregate separately the left endpoints and the interval widths. We show that on discrete scales it is essentially impossible to do so, while on continuous scales we can characterise the rules meeting these requirements as those that compute a weighted average of the endpoints of the individual intervals.

Compressed video action recognition has recently drawn growing attention, since it remarkably reduces the storage and computational cost via replacing raw videos by sparsely sampled RGB frames and compressed motion cues (e.g., motion vectors and residuals). However, this task severely suffers from the coarse and noisy dynamics and the insufficient fusion of the heterogeneous RGB and motion modalities. To address the two issues above, this paper proposes a novel framework, namely Attentive Cross-modal Interaction Network with Motion Enhancement (MEACI-Net). It follows the two-stream architecture, i.e. one for the RGB modality and the other for the motion modality. Particularly, the motion stream employs a multi-scale block embedded with a denoising module to enhance representation learning. The interaction between the two streams is then strengthened by introducing the Selective Motion Complement (SMC) and Cross-Modality Augment (CMA) modules, where SMC complements the RGB modality with spatio-temporally attentive local motion features and CMA further combines the two modalities with selective feature augmentation. Extensive experiments on the UCF-101, HMDB-51 and Kinetics-400 benchmarks demonstrate the effectiveness and efficiency of MEACI-Net.

Extracting relational triples from unstructured text is an essential task in natural language processing and knowledge graph construction. Existing approaches usually contain two fundamental steps: (1) finding the boundary positions of head and tail entities; (2) concatenating specific tokens to form triples. However, nearly all previous methods suffer from the problem of error accumulation, i.e., the boundary recognition error of each entity in step (1) will be accumulated into the final combined triples. To solve the problem, in this paper, we introduce a fresh perspective to revisit the triple extraction task and propose a simple but effective model, named DirectRel. Specifically, the proposed model first generates candidate entities through enumerating token sequences in a sentence, and then transforms the triple extraction task into a linking problem on a ``head -> tail" bipartite graph. By doing so, all triples can be directly extracted in only one step. Extensive experimental results on two widely used datasets demonstrate that the proposed model performs better than the state-of-the-art baselines.

Reinforcement learning in problems with symbolic state spaces is challenging due to the need for reasoning over long horizons. This paper presents a new approach that utilizes relational abstractions in conjunction with deep learning to learn a generalizable Q-function for such problems. The learned Q-function can be efficiently transferred to related problems that have different object names and object quantities, and thus, entirely different state spaces. We show that the learned, generalized Q-function can be utilized for zero-shot transfer to related problems without an explicit, hand-coded curriculum. Empirical evaluations on a range of problems show that our method facilitates efficient zero-shot transfer of learned knowledge to much larger problem instances containing many objects.

Reinforcement learning (RL) is a powerful framework for learning complex behaviors, but lacks adoption in many settings due to sample size requirements. We introduce a framework for increasing sample efficiency of RL algorithms. Our approach focuses on optimizing environment rewards with high-level instructions. These are modeled as a high-level controller over temporally extended actions known as options. These options can be looped, interleaved and partially ordered with a rich language for high-level instructions. Crucially, the instructions may be underspecified in the sense that following them does not guarantee high reward in the environment. We present an algorithm for control with these so-called option machines (OMs), discuss option selection for the partially ordered case and describe an algorithm for learning with OMs. We compare our approach in zero-shot, single- and multi-task settings in an environment with fully specified and underspecified instructions. We find that OMs perform significantly better than or comparable to the state-of-art in all environments and learning settings.

In this paper, we propose a new hyper-heuristic approach that uses reinforcement learning to automatically learn the selection of low-level heuristics across a wide range of problem domains. We provide a detailed analysis and evaluation of the algorithm components, including different ways to represent the hyper-heuristic state space and reset strategies to avoid unpromising areas of the solution space. Our methods have been evaluated using HyFlex, a well-known benchmarking framework for cross-domain hyper-heuristics, and compared with state-of-the-art approaches. The experimental evaluation shows that our reinforcement-learning based approach produces results that are competitive with the state-of-the-art, including the top participants of the Cross Domain Hyper-heuristic Search Competition 2011.

Multivariate time-series forecasting is a critical task for many applications, and graph time-series network is widely studied due to its capability to capture the spatial-temporal correlation simultaneously. However, most existing works focus more on learning with the explicit prior graph structure, while ignoring potential information from the implicit graph structure, yielding incomplete structure modeling. Some recent works attempts to learn the intrinsic or implicit graph structure directly, while lacking a way to combine explicit prior structure with implicit structure together. In this paper, we propose Regularized Graph Structure Learning (RGSL) model to incorporate both explicit prior structure and implicit structure together, and learn the forecasting deep networks along with the graph structure. RGSL consists of two innovative modules. First, we derive an implicit dense similarity matrix through node embedding, and learn the sparse graph structure using the Regularized Graph Generation (RGG) based on the Gumbel Softmax trick. Second, we propose a Laplacian Matrix Mixed-up Module (LM3) to fuse the explicit graph and implicit graph together. We conduct experiments on three real-word datasets. Results show that the proposed RGSL model outperforms existing graph forecasting algorithms with a notable margin, while learning meaningful graph structure simultaneously. Our code and models are made publicly available at https://github.com/alipay/RGSL.git.