Large Language Models (LLMs) can assist multimodal fake news detection by predicting pseudo labels.However, LLM-generated pseudo labels alone demonstrate poor performance compared to traditional detection methods, making their effective integration nontrivial.In this paper, we propose Global Label Propagation Network with LLM-based Pseudo Labeling (GLPN-LLM) for multimodal fake news detection, which integrates LLM capabilities via label propagation techniques.The global label propagation can utilize LLMgenerated pseudo labels, enhancing prediction accuracy by propagating label information among all samples.For label propagation, a mask-based mechanism is designed to prevent label leakage during training by ensuring that training nodes do not propagate their own labels back to themselves.Experimental results on benchmark datasets show that by synergizing LLMs with label propagation, our model achieves superior performance over state-ofthe-art baselines.Our code is available online 1 .

Rumor detection on social media has become crucial due to the rapid spread of misinformation.Existing approaches primarily focus on within-domain tasks, resulting in suboptimal performance in cross-domain scenarios due to domain shift.To address this limitation, we draw inspiration from the strong generalization capabilities of Test-Time Adaptation (TTA) and propose a novel framework to enhance rumor detection performance across different domains.Specifically, we introduce Test-Time Adaptation for Rumor Detection (T 2 ARD), which incorporates both single-domain model and target graph adaptation strategies tailored to the unique requirements of cross-domain rumor detection.T 2 ARD utilizes a graph adaptation module that updates the graph structure and node attributes through multi-level self-supervised contrastive learning, aiming to derive invariant graph representations.To mitigate the impact of significant distribution shifts on self-supervised signals, T 2 ARD performs model adaptation by using annotations from Large Language Models (LLMs) on target graph to produce pseudo-labels as supervised signals.Experiments conducted on four widely used cross-domain datasets demonstrate that T 2 ARD achieves state-of-the-art performance, surpassing existing methods in rumor detection.

In recent years, numerous Time Series Anomaly Detection methods have emerged, focusing primarily on detecting anomalies within time series, with limited work on open set recognition. In real-world scenarios, obtaining anomaly data is challenging. We can acquire a limited number of samples representing unknown anomaly classes, but these samples are far from representing the full distribution of all possible unknown classes. A direct approach would be to learn from known samples to achieve the goal of recognizing unknown samples. This paper contributes twofold: First, we introduce a cross-domain approach combining time-domain and frequency-domain information for time series Open Set Recognition, achieving superior detection under limited samples. This is the first application of a crossdomain method in the field of time series Open Set Recognition. Second, we propose a learning scheme using the Cross-domain Weibull distribution for classifier calibration. The proposed method, Fusion-OSR, achieves state-of-the-art accuracy, F1score, and recall across 30 datasets.

Video Copy Localization (VCL) aims to identify all copied segments within untrimmed video pairs. Fully supervised methods, which require annotating the boundaries of copied segments, are labor-intensive and susceptible to distraction from non-copied segments. To address this problem, we propose a more efficient annotation paradigm called "single frame supervision", which only requires two randomly selected timestamps within the copied segments. With single frame supervision, we introduce a method called VCSA. This method combines spatial-temporal perception and fusion modules to extract and fuse features, which are then segmented for contrastive learning. Additionally, we propose grid alignment loss to enhance the ability of model to distinguish copied from non-copied segments. Extensive experiments demonstrate that VCSA outperforms other methods within our paradigm, sometimes even reaching performance levels comparable to fully supervised methods.

Early Time Series Classification (ETSC) aims to predict class labels using only prefixes (partial sequences) of time series, which is crucial for applications requiring prompt decision-making. However, time series prefixes, with their limited data length, pose significant challenges for models to recognize patterns and provide reliable predictions based on incomplete data. Enhancing ETSC by addressing this limitation is essential. Existing ETSC methods primarily focus on improving feature extractors or refining stopping strategies, yet they often fail to overcome the core issue of data insufficiency in early segments. We present EarlyMix, which employs a hierarchical mixing strategy comprising Sample Mixing and Latent Mixing to augment data and enhance feature representation. This method significantly improves model performance by enabling the extraction of more informative and robust features from time series prefixes. Extensive experiments on widely-used benchmark datasets demonstrate the superior performance of EarlyMix over baselines.

Early Time Series Classification (ETSC) is pivotal in time-sensitive real-world applications. However, existing methods face two significant challenges: (1) the lack of explicit representation of uncertainty, which leads to inaccurate decision-making in the early stages, and (2) the difficulty in handling heteroscedastic uncertainty, where uncertainty dynamically changes as more data is observed, especially in early stages of time series classification. Traditional ETSC methods assume that the predicted probability distributions at each time step are completely reliable and accurate. This assumption neglects the inherent uncertainty in the classification process, leading to cumulative errors that ultimately affect the overall accuracy and reliability of the classification decision. To address these challenges, we propose a novel Uncertainty-aware framework for early Time Series Classification (UTSC), designed to model and adapt to uncertainty during early classification dynamically. UTSC incorporates an Uncertainty Probability Decoder (UPD), which captures the inherent randomness and uncertainty in early-stage data by leveraging hidden-layer information to model variability adaptively. Additionally, UTSC employs an enhanced probability reweighting method to adjust the probability distributions dynamically, enabling the model to account for uncertainty and make more informed decisions. We evaluate UTSC on 45 diverse datasets, comparing its performance against eight baseline models. The results demonstrate that UTSC significantly outperforms existing methods. Our code is publicly available1.