Retinal microaneurysms (MA) are the earliest clinical indicators of diabetic retinopathy (DR), and their detection is crucial for early diagnosis of DR and preventing vision loss. However, accurately detecting MA is extremely challenging. Existing object detection models fail to accurately detect MA and suffer from severe under-detection, because MA are characterized by tiny size and low contrast with the fundus background. To address this challenge, we have developed a two-stage learning model, MADET, which is proposed. MADET consists of two models: an enhanced feature attention fusion model (EFU-Net) for segmentation and a residual multi-scale model (RMSNet) for classification. EFU-Net integrates multi-attention gate fusion (MAGF) modules and residual RFCBAM convolution (ResRF-conv) into the standard U-Net architecture to improve the encoder-decoder connection and enhance the model’s ability to acquire strong context features, resulting in better segmenting of the MA regions. RMSNet considers the difference of feature information between different scales to eliminate non-MA regions, thereby achieving more accurate MA detection. MADET achieved a Free-response Receiver Operating Characteristic (FROC) score of 0.625 on the E-Ophtha-MA dataset (148 images), showing an improvement of 0.034 over recent methods. On the IDRiD dataset (81 images), it achieved a score of 0.497, which is 0.021 higher than the latest comparable approach. These results indicate that MADET achieves competitive or near-best performance compared to existing state-of-the-art techniques.

As graph tasks become pervasive in real-time and safetycritical domains (e.g., financial fraud detection and electrical power systems), it is also essential to guarantee their reliable execution beyond pursuing extraordinary performance. However, due to the neglect of consideration for graph-specific execution paradigm, existing Fault Injection (FI) reliability analysis methods typically incur inaccurate system error resilience characterization, making it challenging to provide helpful guidance for efficient and reliable graph processing paradigm design. This paper proposes GraphFI, an efficient Graph Fault Injection framework on the universal parallel tasks acceleration platform (i.e., GPGPUs). Our key insight is progressively excavating the graph-specific error propagation and effect mechanisms, thereby avoiding blind FI trials. Firstly, observing that iterations with similar active vertex set exhibit similar error behavior, we propose iteration-driven GraphFI (ID-GraphFI) to solely select representative iterations for fast error resilience profile assessment. Secondly, by detecting resilience-similarity communities in graph topology, we propose topology-driven GraphFI (TD-GraphFI) that only selects representative vertices for community overall reliability evaluation. Thirdly, by exploring the graph-specific fault monotonic property, we propose the monotonicity-driven GraphFI (MD-GraphFI) to granularly draw system severe error boundaries for predictable/unnecessary fault injection avoidance. Merging them all, GraphFI can reduce system fault site space by up to two orders of magnitude, which achieves $2.1 \sim 15.2 \times$ speedup compared to SOTA methods while providing better reliability assessment accuracy.

Generalized few-shot node classification (GFS-NC) is a very important challenge for graph-based algorithms, as it requires to identify novel classes and base classes simultaneously. Although there are several methods that try to combine metalearning or metric learning with graph neural networks to solve few-shot problem, most of them assume that test samples only come from the novel classes, which is impractical in reality. Besides, they overlook the relationship among classes, which can provide additional information for the novel classes classification. In this article, we propose a graph-based knowledge distillation network (GraphKD) to extract the class relationship and learn better nodes representations for nodes from novel classes in GFS-NC task. GraphKD consists of two modules: balanced pretraining module and class-relation transferring module. Balanced pretraining can optimize network parameters to a suitable manifold for subsequent initialization. The class-relation transferring module leverages a knowledge distillation model, where a teacher model generates soft labels containing interclass relationships and then transfer them to the student model. The student model is optimized to fit both the soft labels and hard labels concurrently. This relationship information can help the student model better understand the similarities and differences between classes, thereby improving its classification performance. In addition, we employee information entropy to distinguish the samples locate at the boundary of a base class and novel class and then assign them larger weights in the student model to enhance its expressive capacity for novel nodes. Our experiments show that the proposed method outperforms state-of-the-art baselines on various few-shot node classification datasets.

Online social networks often enforce a restricted web interface to allow third part users to access their data objects in one-by-one manner along links. The heavy cost of collecting data from them makes it very difficult to train a graph neural networks (GNNs) algorithm by using valuable data resources from those social network services. In this article, we endeavor to harness the potential of collected data in order to cultivate robust node representations, all while taking into account the prevalent class-imbalanced scenario. To this end, we propose a framework based on the reinforcement learning technique that can implement training while sampling. We employee deep Q-learning network (DQN) as the basic reinforcement learning framework which maintains an experience replay buffer to store experiences in order to improve efficiency and generalization ability of the algorithm. The proposed framework can guide the training process according to the label distribution of the currently collected training data. To maximize the utility of the collected data, we introduce two ways to guide the agent to learn more robust node representations under the natural class-imbalanced scenario. Specifically, effective-number GNNs (ENG) try to employ a reward function with the technique of effective number and resampling GNNs (RSG) attempts to modify the state transition function motivated by resampling technique. Extensive experiments are conducted on several real-world imbalanced datasets and the proposed methods significantly outperform the state-of-the-art methods in node classification task, accomplishing higher performance while utilizing a reduced volume of data.

Graph Neural Networks (GNNs) have achieved remarkable success in graph-based tasks. However, learning unbiased node representations under class-imbalanced training data remains challenging. Existing solutions may face overfitting due to extensive reuse of those limited labeled data in minority classes. Furthermore, many works address the class-imbalanced issue based on the embeddings generated from the biased GNNs, which make models intrinsically biased towards majority classes. In this paper, we propose a novel data augmentation strategy GraphGLS for semi-supervised class-imbalanced node classification, which aims to select informative unlabeled nodes to augment minority classes with consideration of both global and local information. Specifically, we first design a Global Selection module to learn global information (pseudo-labels) for unlabeled nodes and then select potential ones from them for minority classes. The Local Selection module further conducts filtering over those potential nodes by comparing their neighbor distributions with minority classes. To achieve this, we further design a neighbor distribution auto-encoder to learn a robust node-level neighbor distribution for each node. Then, we define class-level neighbor distribution to capture the overall neighbor characteristics of nodes within the same class. We conduct extensive experiments on multiple datasets, and the results demonstrate the superiority of GraphGLS over state-of-the-art baselines.

Graph neural networks (GNNs) have achieved great success in node classification tasks. However, existing GNNs naturally bias towards the majority classes with more labelled data and ignore those minority classes with relatively few labelled ones. The traditional techniques often resort over-sampling methods, but they may cause overfitting problem. More recently, some works propose to synthesize additional nodes for minority classes from the labelled nodes, however, there is no any guarantee if those generated nodes really stand for the the corresponding minority classes. In fact, improperly synthesized nodes may result in insufficient generalization of the algorithm. To resolve the problem, in this paper we seek to automatically augment the minority classes from the massive unlabelled nodes of the graph. Specifically, we propose \textit{GraphSR}, a novel self-training strategy to augment the minority classes with significant diversity of unlabelled nodes, which is based on a Similarity-based selection module and a Reinforcement Learning(RL) selection module. The first module finds a subset of unlabelled nodes which are most similar to those labelled minority nodes, and the second one further determines the representative and reliable nodes from the subset via RL technique. Furthermore, the RL-based module can adaptively determine the sampling scale according to current training data. This strategy is general and can be easily combined with different GNNs models. Our experiments demonstrate the proposed approach outperforms the state-of-the-art baselines on various class-imbalanced datasets.