While significant progress has been made in multi-modal learning driven by large-scale image-text datasets, there is still a noticeable gap in the availability of such datasets within the facial domain. To facilitate and advance the field of facial representation learning, we present FLIP-80M, a large-scale visual-linguistic dataset comprising over 80 million face images paired with text descriptions. FLIP-80M is constructed by leveraging the large openly available image-text-pair dataset LAION-5B and a mixed-method approach to filter face-related pairs from both visual and linguistic perspectives. Our curation process involves face detection, face caption classification, text de-noising, and synthesis-based image augmentation. As a result, FLIP-80M stands as the largest face-text dataset to date. To evaluate the potential of our dataset, we fine-tune the CLIP model using the proposed FLIP-80M, to create FLIP (Facial Language-Image Pretraining) and assess its representation capabilities across various downstream tasks. Our experiments demonstrate that our FLIP model achieves state-of-the-art results in a range of face analysis tasks, including face parsing, face alignment, and face attribute classification. The dataset and models are available at https://github.com/ydli-ai/FLIP.
论文检索
输入标题、作者或关键词，从 7,876 篇学术成果中精准定位
Multimodal tasks require learning a joint representation of the constituent modalities of data. Contrastive learning learns a joint representation by using a contrastive loss. For example, CLIP takes as input image-caption pairs and is trained to maximize the similarity between an image and its corresponding caption in actual image-caption pairs, while minimizing the similarity for arbitrary image-caption pairs. This approach operates on the premise that the caption depicts the image's content. However, this assumption does not always hold true for tweets that contain both text and images. Previous studies have indicated that the connection between the image and the text in a tweet is more intricate and complex. We study the effectiveness of pre-trained multimodal contrastive learning models, specifically, CLIP, and ALIGN, on the task of classifying multimodal crisis related tweets. Our experiments using two publicly available datasets, CrisisMMD and DMD, show that despite the intricate relationships in tweets, pre-trained contrastive learning models fine-tuned with task-specific data produce better results than prior approaches used for the multimodal classification of crisis related tweets. Additionally, the experiments show that the contrastive learning models are effective in low-data few-shot and cross-domain settings.
Human vs ChatGPT: Effect of Data Annotation in Interpretable Crisis-Related Microblog Classification
PDF ↗Recent studies have exploited the vital role of microblogging platforms, such as Twitter, in crisis situations. Various machine-learning approaches have been proposed to identify and prioritize crucial information from different humanitarian categories for preparation and rescue purposes. In crisis domain, the explanation of models' output decisions is gaining significant research momentum. Some previous works focused on human annotations of rationales to train and extract supporting evidence for model interpretability. However, such annotations are usually expensive, require much effort, and are not always available in real-time situations of a new crisis event. In this paper, we investigate the recent advances in large language models (LLMs) as data annotators on informal tweet text. We perform a detailed qualitative and quantitative evaluation of ChatGPT rationale annotations over a few-shot setup. ChatGPT annotations are quite close to humans but less precise in nature. Further, we propose an active learning-based interpretable classification model from a small set of annotated data. Our experiments show that (a). ChatGPT has the potential to extract rationales for the crisis tweet classification tasks, but the performance is slightly less than the model trained on human-annotated rationale data (\sim3-6%), (b). active learning setup can help reduce the burden of manual annotations and maintain a trade-off between performance and data size.
In the global food industry, where the line between legitimate and illicit manufacturing is increasingly blurred by the scale and complexity of the supply chain, safeguarding consumer health and trust necessitates innovative detection methods. Addressing this, this paper presents Graph-aware Self-supervised Contrastive Anomaly Ranking (GraphCAR), a novel unsupervised learning model, devised to identify illicit food factories through the scrutiny of chemical declaration data. GraphCAR tackles the scarcity of labeled data and the intricacies inherent in the vast array of declared chemicals, leveraging a Graph Autoencoder fused with a self-supervised contrastive learning mechanism. This fusion not only simplifies the feature space by embedding chemical declarations within a bipartite graph but also adeptly flags subtle, potentially illicit patterns through contrastively inspecting the learned factory representations. Through rigorous evaluations conducted on real-world factory's chemical declaration data, GraphCAR has demonstrated superior performance over conventional methods on unsupervised outlier detection and one-class classification tasks, showcasing its accuracy, robustness and reliability in flagging potential malpractice. With its successful application in food safety, GraphCAR stands as a testament to the potential of AI-driven solutions to address multifaceted challenges for the greater good.
As an integral part of people's daily lives, social media is becoming a rich source for automatic mental health analysis. As traditional discriminative methods bear poor generalization ability and low interpretability, the recent large language models (LLMs) have been explored for interpretable mental health analysis on social media, which aims to provide detailed explanations along with predictions in zero-shot or few-shot settings. The results show that LLMs still achieve unsatisfactory classification performance in a zero-shot/few-shot manner, which further significantly affects the quality of the generated explanations. Domain-specific finetuning is an effective solution, but faces two critical challenges: 1) lack of high-quality training data. 2) no open-source foundation LLMs. To alleviate these problems, we formally model interpretable mental health analysis as a text generation task, and build the first multi-task and multi-source interpretable mental health instruction (IMHI) dataset with 105K data samples to support LLM instruction tuning and evaluation. The raw social media data are collected from 10 existing sources covering 8 mental health analysis tasks. We prompt ChatGPT with expert-designed few-shot prompts to obtain explanations. To ensure the reliability of the explanations, we perform strict automatic and human evaluations on the correctness, consistency, and quality of generated data. Based on the IMHI dataset and LLaMA2 foundation models, we train MentaLLaMA, the first open-source instruction-following LLM series for interpretable mental health analysis on social media. We evaluate MentaLLaMA and other advanced methods on the IMHI benchmark, the first holistic evaluation benchmark for interpretable mental health analysis. The results show that MentaLLaMA approaches state-of-the-art discriminative methods in correctness and generates human-level explanations. MentaLLaMA models also show strong generalizability to unseen tasks. The project is available at https://github.com/SteveKGYang/MentaLLaMA.
Anonymous networks employ a triple proxy to transmit packets to enhance user privacy, causing traffic packets from all applications and web services to form a unified flow. The traditional approach of applying flow-level encrypted traffic classification methods to anonymous traffic (i.e., treating consecutive packets as a single flow) is hindered by irrelevant packet noise. Moreover, fluctuations in the network environment can introduce per-packet attribute noise and discrepancies between training and test data. How to extract robust patterns from consecutive packets replete with noise remains a key challenge. In this paper, we propose the Anti-Noise Network (AN-Net) to construct robust short-term representations for a single modality, effectively countering irrelevant packet noise. We also incorporate an enhanced multi-modal fusion approach to combat per-packet attribute noise. AN-Net achieves state-of-the-art performance across two anonymous traffic classification tasks and one VPN traffic classification task, notably elevating the F1 score of SJTU-AN21 to 94.39% (6.24%↑). Our code and dataset are available on https://github.com/SJTU-dxw/AN-Net.
With the increasing popularity of live streaming, the interactions from viewers during a live streaming can provide more specific and constructive feedback for both the streamer and platform. In such scenario, the primary and most direct feedback method from the audience is through comments. Thus, mining these live streaming comments to unearth the intentions behind them and, in turn, aiding streamers to enhance their live streaming quality is significant for the well development of live streaming ecosystem. To this end, we introduce the MMLSCU dataset, containing 50,129 intention-annotated comments across multiple modalities (text, images, vi-deos, audio) from eight streaming domains. Using multimodal pretrained large model and drawing inspiration from the Chain of Thoughts (CoT) concept, we implement an end-to-end model to sequentially perform the following tasks: viewer comment intent detection ➛ intent cause mining ➛ viewer comment explanation ➛ streamer policy suggestion. We employ distinct branches for video and audio to process their respective modalities. After obtaining the video and audio representations, we conduct a multimodal fusion with the comment. This integrated data is then fed into the large language model to perform inference across the four tasks following the CoT framework. Experimental results indicate that our model outperforms three multimodal classification baselines on comment intent detection and streamer policy suggestion, and one multimodal generation baselines on intent cause mining and viewer comment explanation. Compared to the models using only text, our multimodal setting yields superior outcomes. Moreover, incorporating CoT allows our model to enhance comment interpretation and more precise suggestions for the streamers. Our proposed dataset and model will bring new research attention on multimodal live streaming comment understanding.
Graph anomaly detection (GAD) has various applications in finance, healthcare, and security. Graph Neural Networks (GNNs) are now the primary method for GAD, treating it as a task of semi-supervised node classification (normal vs. anomalous). However, most traditional GNNs aggregate and average embeddings from all neighbors, without considering their labels, which can hinder detecting actual anomalies. To address this issue, previous methods try to selectively aggregate neighbors. However, the same selection strategy is applied regardless of normal and anomalous classes, which does not fully solve this issue. This study discovers that nodes with different classes yet similar neighbor label distributions (NLD) tend to have opposing loss curves, which we term it as "loss rivalry". By introducing Contextual Stochastic Block Model (CSBM) and defining NLD distance, we explain this phenomenon theoretically and propose a Bi-level optimization Graph Neural Network (BioGNN), based on these observations. In a nutshell, the lower level of BioGNN segregates nodes based on their classes and NLD, while the upper level trains the anomaly detector using separation outcomes. Our experiments demonstrate that BioGNN outperforms state-of-the-art methods on four benchmarks and effectively mitigates "loss rivalry".
Text classification is a fundamental task in web content mining. Although the existing supervised contrastive learning (SCL) approach combined with pre-trained language models (PLMs) has achieved leading performance in text classification, it lacks fundamental principles. Theoretically motivated by a derived lower bound of mutual information maximization, we propose a dual contrastive learning framework DualCL that satisfies three properties, i.e., parameter-free, augmentation-easy and label-aware. DualCL generates classifier parameters from the PLM and simultaneously uses them for classification and as augmented views of the input text for supervised contrastive learning. Extensive experiments conclusively demonstrate that DualCL excels in learning superior text representations and consistently outperforms baseline models.
Text classification is one vital tool assisting web content mining. Semi-supervised text classification (SSTC) offers an approach to alleviate the burden of annotation costs by training on a few labeled texts alongside many unlabeled texts. Unsolved challenges in SSTC are the overfitting problem caused by the limited labeled data and the mislabeling problem of unlabeled texts. To address these issues, this paper proposes a Self-Paced PairWise representation learning (SPPW) model. Concretely, SPPW alleviates the overfitting problem by replacing the overfitting-prone learning of a parameterized classifier with representation learning in a pair-wise manner. Besides, we propose a novel self-paced text filtering method that effectively integrates both label confidence and text hardness to reduce mislabeled texts synergistically. Extensive experiments on 3 benchmark SSTC datasets show that SPPW outperforms baselines and is effective in mitigating overfitting and mislabeling problems.
Inductive Graph Alignment Prompt: Bridging the Gap between Graph Pre-training and Inductive Fine-tuning From Spectral Perspective
PDF ↗The "Graph pre-training and fine-tuning" paradigm has significantly improved Graph Neural Networks(GNNs) by capturing general knowledge without manual annotations for downstream tasks. However, due to the immense gap of data and tasks between the pre-training and fine-tuning stages, the model performance is still limited. Inspired by prompt fine-tuning in Natural Language Processing(NLP), many endeavors have been made to bridge the gap in graph domain. But existing methods simply reformulate the form of fine-tuning tasks to the pre-training ones. With the premise that the pre-training graphs are compatible with the fine-tuning ones, these methods typically operate in transductive setting. In order to generalize graph pre-training to inductive scenario where the fine-tuning graphs might significantly differ from pre-training ones, we propose a novel graph prompt based method called Inductive Graph Alignment Prompt(IGAP). Firstly, we unify the mainstream graph pre-training frameworks and analyze the essence of graph pre-training from graph spectral theory. Then we identify the two sources of the data gap in inductive setting: (i) graph signal gap and (ii) graph structure gap. Based on the insight of graph pre-training, we propose to bridge the graph signal gap and the graph structure gap with learnable prompts in the spectral space. A theoretical analysis ensures the effectiveness of our method. At last, we conduct extensive experiments among nodes classification and graph classification tasks under the transductive, semi-inductive and inductive settings. The results demonstrate that our proposed method can successfully bridge the data gap under different settings.
Explaining stock predictions is generally a difficult task for traditional non-generative deep learning models, where explanations are limited to visualizing the attention weights on important texts. Today, Large Language Models (LLMs) present a solution to this problem, given their known capabilities to generate human-readable explanations for their decision-making process. However, the task of stock prediction remains challenging for LLMs, as it requires the ability to weigh the varying impacts of chaotic social texts on stock prices. The problem gets progressively harder with the introduction of the explanation component, which requires LLMs to explain verbally why certain factors are more important than the others. On the other hand, to fine-tune LLMs for such a task, one would need expert-annotated samples of explanation for every stock movement in the training set, which is expensive and impractical to scale. To tackle these issues, we propose our Summarize-Explain-Predict (SEP) framework, which utilizes a verbal self-reflective agent and Proximal Policy Optimization (PPO) that allow a LLM teach itself how to generate explainable stock predictions, in a fully autonomous manner. The reflective agent learns how to explain past stock movements through a self-reasoning process, while the PPO trainer trains the model to generate the most likely explanations given the input texts at test-time. The training samples for the PPO trainer are also the responses generated during the reflective process, which eliminates the need for human annotators. Using our SEP framework, we fine-tune a specialized LLM that can outperform both traditional deep-learning and LLM methods in prediction accuracy and Matthews correlation coefficient, for the stock classification task. To justify the generalization capability of our framework, we further test it on the portfolio construction task, and demonstrate its effectiveness through various portfolio metrics. Our code can be accessed through https://github.com/koa-fin/sep.
Weakly supervised text classification (WSTC), also called zero-shot or dataless text classification, has attracted increasing attention due to its applicability in classifying a mass of texts within the dynamic and open Web environment, since it requires only a limited set of seed words (label names) for each category instead of labeled data. With the help of recently popular prompting Pre-trained Language Models (PLMs), many studies leveraged manually crafted and/or automatically identified verbalizers to estimate the likelihood of categories, but they failed to differentiate the effects of these category-indicative words, let alone capture their correlations and realize adaptive adjustments according to the unlabeled corpus. In this paper, in order to let the PLM effectively understand each category, we at first propose a novel form of rule-based knowledge using logical expressions to characterize the meanings of categories. Then, we develop a prompting PLM-based approach named RulePrompt for the WSTC task, consisting of a rule mining module and a rule-enhanced pseudo label generation module, plus a self-supervised fine-tuning module to make the PLM align with this task. Within this framework, the inaccurate pseudo labels assigned to texts and the imprecise logical rules associated with categories mutually enhance each other in an alternative manner. That establishes a self-iterative closed loop of knowledge (rule) acquisition and utilization, with seed words serving as the starting point. Extensive experiments validate the effectiveness and robustness of our approach, which markedly outperforms state-of-the-art weakly supervised methods. What is more, our approach yields interpretable category rules, proving its advantage in disambiguating easily-confused categories.
Graphs, as a fundamental data structure, have proven efficacy in modeling complex relationships between objects and are therefore found in wide web applications. Graph classification is an essential task in graph data analysis, which can effectively assist in extracting information and mining content from the web. Recently, few-shot graph classification, a more realistic and challenging task, has garnered great research interest. Existing few-shot graph classification models are all supervised, assuming abundant labeled data in base classes for meta-training. However, sufficient annotation is often challenging to obtain in practice due to high costs or demand for expertise. Moreover, they commonly adopt complicated meta-learning algorithms via episodic training to transfer prior knowledge from base classes. To break free from these constraints, in this paper, we propose a simple yet effective approach named SMART for unsupervised few-shot graph classification without using any labeled data. SMART employs transfer learning philosophy instead of the previously prevailing meta-learning paradigm, avoiding the need for sophisticated meta-learning algorithms. Additionally, we adopt a novel mixup strategy to augment the original graph data and leverage unsupervised pretraining on these data to obtain the expressive graph encoder. We also utilize the prompt tuning technique to alleviate the overfitting and low fine-tuning efficiency caused by the limited support samples of novel classes. Extensive experimental results demonstrate the superiority of our proposed approach, significantly surpassing even leading supervised few-shot graph classification models. Our code is available here.
Web-based applications such as chatbots, search engines and news recommendations continue to grow in scale and complexity with the recent surge in the adoption of large language models (LLMs). Online model selection has thus garnered increasing attention due to the need to choose the best model among a diverse set while balancing task reward and exploration cost. Organizations faces decisions like whether to employ a costly API-based LLM or a locally finetuned small LLM, weighing cost against performance. Traditional selection methods often evaluate every candidate model before choosing one, which are becoming impractical given the rising costs of training and finetuning LLMs. Moreover, it is undesirable to allocate excessive resources towards exploring poor-performing models. While some recent works leverage online bandit algorithm to manage such exploration-exploitation trade-off in model selection, they tend to overlook the increasing-then-converging trend in model performances as the model is iteratively finetuned, leading to less accurate predictions and suboptimal model selections. In this paper, we propose a time-increasing bandit algorithm TI-UCB, which effectively predicts the increase of model performances due to training or finetuning and efficiently balances exploration and exploitation in model selection. To further capture the converging points of models, we develop a change detection mechanism by comparing consecutive increase predictions. We theoretically prove that our algorithm achieves a lower regret upper bound, improving from prior works' polynomial regret to logarithmic in a similar setting. The advantage of our method is also empirically validated through extensive experiments on classification model selection and online selection of LLMs. Our results highlight the importance of utilizing increasing-then-converging pattern for more efficient and economic model selection in the deployment of LLMs.
Click-through Rate (CTR) module is the foundation block of recommendation system and used for search, content selection, advertising, video streaming etc. CTR is modelled as a classification problem and extensive research is done to improve the CTR models. However, uncertainty method for these models are still an unexplored area. In this work we analyse popular uncertainty methods in the context of recommendation system. We found that popular uncertainty models fails to capture the predictive uncertainty of the CTR model that exist unique to the recommendation models and is not prevalent in the traditional classification models. We empirical show why a different uncertainty measure is required for the recommendation system CTR prediction models. We propose PRU (Predictive Relevance Uncertainty), a single forward pass uncertainty approach for a sample as a distance from the predictive relevance samples of the training data. We show the efficacy of the proposed predictive relevance uncertainty (PRU) on selective prediction. Further, we demonstrate the utility of the proposed framework on the downstream task of OOD detection and active learning while maintaining the latency of a single pass deterministic model.
Trajectory representation learning plays a pivotal role in supporting various downstream tasks, such as travel time estimation, trajectory classification and Top-k similar trajectory search. Traditional methods in order to filter the noise in GPS trajectories tend to focus on routing-based methods to simplify the trajectories. However, these approaches ignore the motion details contained in the GPS data, limiting the representation capability of trajectory representation learning. To fill this gap, we propose a novel representation learning framework that is Jointly G PS and Route Modeling based on self-supervised technology, namely JGRM. We consider GPS trajectory and route trajectory as the two modals of a single movement observation and fuse information through inter-modal information interaction. Specifically, we develop two encoders, each tailored to capture representations of GPS trajectories and route trajectories respectively. The representations from these two modalities are fed into a shared transformer for inter-modal information interaction. Eventually, we design three self-supervised tasks to train the model. We validate the effectiveness of the proposed method on two real-world datasets through extensive experiments. The experimental results show that JGRM significantly outperforms existing methods in both road segment representation and trajectory representation tasks. Our source code is available at Github https://github.com/mamazi0131/JGRM.
Sub-model extraction based federated learning has emerged as a popular strategy for training models on resource-constrained devices. However, existing methods treat all clients equally and extract sub-models using predetermined rules, which disregard the statistical heterogeneity across clients and may lead to fierce competition among them. Specifically, this paper identifies that when making predictions, different clients tend to activate different neurons of the entire model related to their respective distributions. If highly activated neurons from some clients with one distribution are incorporated into the sub-model allocated to other clients with different distributions, they will be forced to fit the new distributions, which can hinder their activation over the previous clients and result in a performance reduction. Motivated by this finding, we propose a novel method called FedDSE, which can reduce the conflicts among clients by extracting sub-models based on the data distribution of each client. The core idea of FedDSE is to empower each client to adaptively extract neurons from the entire model based on their activation over the local dataset. We theoretically show that FedDSE can achieve an improved classification score and convergence over general neural networks with the ReLU activation function. Experimental results on various datasets and models show that FedDSE outperforms all state-of-the-art baselines.
Local news outlets play a vital role in providing trusted and relevant information to communities and addressing their specific needs and concerns. The emergence of news outlets posing as local sources and their spread on social media present a significant challenge in the digital information landscape. This paper presents a comprehensive study investigating posts featuring "pink slime'' news, which is a term that has been used to refer to these news outlets due to its deceptive nature. By analyzing a large dataset of posts, we gain valuable insights into the patterns of these posts and the origin of these posts. We show in this work that extracting syntactical features proves valuable in developing a classification approach for detecting such posts and that the approach achieves 92.5% accuracy. We also show that our approach achieves near-perfect detection when grouping the posts by URL.
This paper develops a symbiotic human-AI collective learning framework that explores the complementary strengths of both AI and crowdsourced human intelligence to address a novel Web-based healthcare-policy-adherence assessment (WebHA) problem. In particular, the objective of the WebHA problem is to automatically assess people's public health policy adherence during emergent global health crisis events (e.g., COVID-19, MonkeyPox) by exploring massive social media imagery data. Recent advances in human-AI systems exhibit a significant potential in addressing the intricate imagery-based classification problems like WebHA by leveraging the collective intelligence of both humans and AI. This paper aims to address the limitation of existing human-AI systems that often rely heavily on human intelligence to improve AI model performance while overlooking the fact that humans themselves can be fallible and prone to errors. To address the above limitation, this paper develops SymLearn, a symbiotic human-AI co-learning framework that leverages human intelligence to troubleshoot and fine-tune the AI model while using AI models to guide human crowd workers to reduce the inherent human errors in their labels. Extensive experiments on two real-world WebHA applications show that SymLearn clearly outperforms the state-of-the-art baselines by improving WebHA performance and reducing crowd response delay.