Driven by the remarkable task-level generalization ability of large language models, an emerging trend of graph learning is to enable fast adaptation to new tasks with limited annotations, and has found applications across a spectrum of domains. Graph meta-learning and graph prompting techniques have demonstrated potential in task generalization by transferring knowledge acquired from prior experiences to new tasks. However, these methods often overlook distribution shifts between training and testing data in real-world scenarios. To fill this gap, we delve into a novel and practical challenge, namely joint task and distribution generalization. Motivated by recent studies that explicitly identifying key substructures related to task prediction can help generalization, we introduce a refiner module to highlight key substructures robust to distribution shifts. To efficiently adapt the refiner to new tasks, we introduce a few extra parameters as prompt vectors to instruct its behavior. Specifically, we employ a global prompt to acquire universal knowledge and task-specific prompts to capture task-relevant information. We pretrain model parameters on known tasks, and efficiently adapt to a target task by merely learning a corresponding classifier and task-specific prompt. Extensive experiments in task generalization show that, the proposed Graph Substructure Prompting (GSP) significantly outperforms recent state-of-the-art (SOTA) methods on both in-distribution (ID) and out-of-distribution (OOD) data, instead of a trade-off between them. GSP also enjoys comparable or even less computational cost as compared to baselines.

Master data is commonly utilized in business scenarios, serving as the bedrock for describing core entities. Studies have been presented to manage master data in curating relational data. However, the research of master data remains untouched in the context of time series. This raises concerns regarding time series data management and application, as we have found that master data is also prevalent in time series scenarios. For example, to reduce the fuel consumption of vehicles, it relies on the complete map of engine speed and torque to fuel consumption rate, known as the engine universal characteristic map, i.e., master data. Additionally, to conduct anomaly detection and weather prediction in weather scenario, we need to understand the fine-grained relationship between temperature, wind speed, and atmospheric pressure, which also represents master data. Unfortunately, such master data is often incomplete and inconsistent. In this paper, we propose to predict the fine-grained master data by learning a model from the incomplete and inconsistent observations. A novel MasterNet is designed based on Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs). Deconvolution is employed to predict incomplete data, while the discriminator can successfully tolerate inconsistent observations. We also consider two attention features to capture more time-related information. Experiments over real-world time series datasets show that our MasterNet outperforms the existing approaches in both imputing the incomplete and repairing the inconsistent master data.

This paper studies optical flow estimation, a critical task in motion analysis with applications in autonomous navigation, action recognition, and film production. Traditional optical flow methods require consecutive frames, which are often unavailable due to limitations in data acquisition or real-world scene disruptions. Thus, single-frame optical flow estimation is emerging in the literature. However, existing single-frame approaches suffer from two major limitations: (1) they rely on labeled training data, making them task-specific, and (2) they produce deterministic predictions, failing to capture motion uncertainty. To overcome these challenges, we propose ProbDiffFlow, a training-free framework that estimates optical flow distributions from a single image. Instead of directly predicting motion, ProbDiffFlow follows an estimation-by-synthesis paradigm: it first generates diverse plausible future frames using a diffusion-based model, then estimates motion from these synthesized samples using a pre-trained optical flow model, and finally aggregates the results into a probabilistic flow distribution. This design eliminates the need for task-specific training while capturing multiple plausible motions. Experiments on both synthetic and real-world datasets demonstrate that ProbDiffFlow achieves superior accuracy, diversity, and efficiency, outperforming existing single-image and two-frame baselines.

Model-based diagnosis (MBD) with multiple pseudo-normal observations enables the effective detection of latent faults even when some actual observations are consistent with the system’s expected observations. Current state-of-the-art algorithms overlook this diagnostically rich scenario and incorporate too many components, which frequently leads to stack overflow errors in large-scale experiments. To address these challenges, we introduce our algorithm DKIPNO (Diagnosis with Key node and IterativeDFS for Pseudo-Normal Observations) in this paper, which mainly focuses on MBD with multiple pseudo-normal observations and incorporates two original proposed ideas. Firstly, we introduce the novel concept of the ‘key node’. By virtually flipping the outputs of key nodes and comparing flipped observations with the system’s expected observations, we effectively identify more functionally normal components and significantly reduce the number of potentially faulty components in the diagnostic process. Secondly, we present the IterativeDFS method, an original iterative traversal technique that can prevent stack overflow errors commonly encountered with the recursive method used in previous algorithms. We evaluate the performance of DKIPNO using the ISCAS’85 benchmark in our experiment. Experimental results demonstrate that DKIPNO not only significantly reduces the time required for diagnosis but also prevents stack overflow errors, thereby outperforming other state-of-the-art algorithms.

Assessing the alignment between textual descriptions and corresponding images is fundamental to multi-modal research. In recent years, there has been a surge in the adoption of reference-free methods that utilize visual-language pre-trained models (VLMs). Empirical evidence supports that these innovative methods correlate more closely with human judgment, representing a notable progression in the field. However, due to the unknown underlying judgment mechanisms within VLMs, the metrics designed based on VLMs may exhibit some unidentified flaws. To uncover potential issues with the reference-free metrics, we employ a reinforcement learning approach to hack these metrics, guiding the model to generate sentences that better align with the metric criteria. If the metrics contain some flaws, these deficiencies will manifest in the generated sentences. On the hacking experiment, we observe that the generated sentences achieve higher metric scores, yet they also become unreadable. These inconsistencies reflect the inherent flaws within the metrics themselves. To address these issues, we propose a simple but effective approach by introducing sentences with flaws as negative samples in contrastive learning called Negative Text Contrastive Learning (NTCL). We utilize GPT-4V as an evaluation tool to analyze the generated sentences, and our results demonstrate that the NTCL method is more robust and achieves state-of-the-art performance. We hope our findings can raise awareness in the community about the importance of reference-free image captioning metrics hacking and pave the way for the design of more robust metrics.

Open source communities have a wealth of digital talents, who are urgently needed by various industries under the digitalization process of the entire society. However, barriers exist between digital talents in open source communities and employers. On one hand, open source contributors wonder whether their expertise matches the requirements of specific jobs; On the other hand, developers working on small open source projects are less likely to get recognition from employers, compared with those contributing to well-known projects. To bridge this gap, we propose a new task, matching digital talents and job titles in open source communities, which measures the matching degrees between digital talents with open source experience and job titles requiring digital skills. To solve the task, we construct a heterogeneous information network connecting open source communities and job markets, and propose a semi-supervised network alignment model to augment the connectivity of the network. Then we employ a graph neural network to learn the representations of the digital talents and the job titles from the augmented network, based on which we measure the matching degrees between them. Experimental results demonstrate that our method achieves improvements of at least 5.34, 3.52, 2.37, 2.93, and 8.21 in accuracy, precision, recall, F1, and AUC compared to other possible solutions.

The capacity to interpret the brain’s processing of visual art via brain imaging techniques provides significant understanding of the cognitive mechanisms behind aesthetic appreciation. This study investigates these mechanisms through analyzing electroencephalography (EEG) data from participants performing two different tasks: gazing at a blank wall and viewing artworks. We created the ArtEEGAttention model, a novel deep learning architecture that employs sliding window convolution and multi-head self-attention mechanisms to accurately identify these varied viewing scenarios. Evaluated on a selected dataset of 16 individuals, with EEG signals separated into 3-second epochs and classified according to viewing environment, our model exhibited outstanding performance, with a remarkable cross-subject accuracy of 77.96%. The model’s remarkable accuracy, especially evident in specific subjects, highlights its robustness and superior generalization skills across various brain responses to art.

Natural language generation (NLG) models have emerged as a focal point of research within natural language processing (NLP), exhibiting remarkable performance in tasks such as text composition and dialogue generation. However, their intricate architectures and massive model parameters pose significant challenges to interpretability, limiting their applicability in high-stakes decision-making scenarios. To address this issue, human-computer interaction (HCI) and visualization techniques offer promising avenues to enhance the transparency and usability of NLG models by making their decision-making processes more interpretable. In this paper, we provide a comprehensive investigation into the roles, limitations, and impact of HCI and visualization in facilitating human understanding and control over NLG systems. We introduce a taxonomy of interaction methods and visualization techniques, categorizing three major research domains and their corresponding six key tasks in the application of NLG models. Finally, we summarize the shortcomings in the existing work and investigate the key challenges and emerging opportunities in the era of large language models (LLMs).

The alldifferent constraint is one of the most essential global constraints that has been used to model many classic Constraint Satisfaction Problems (CSPs). To tackle this constraint, previous work generally focuses on proposing filtering algorithms to reduce the value domain of variables, thereby pruning the backtracking search. However, due to the inherent complexity of alldifferent constraints, it remains challenging to solve CSPs containing large-scale alldifferent constraints for the existing search-based solvers. In this paper, we propose an efficient local search algorithm, called AllDiff-LS. We first represent the set of alldifferent constraints as graphs and simplify the graphs by two polynomial-time reduction rules. Afterward, we propose a new two-step strategy for selecting moves, as well as the carefully crafted tabu and restart strategies, to enable the algorithm to explore the search space more efficiently. Experiments on a representative set of benchmarks show that AllDiff-LS successfully solves more instances compared with state-of-the-art complete and heuristic methods, especially on large-scale instances, with much less running time.