Multi-instance multi-label learning (MIML) is a weakly supervised approach that models relationships between complex objects and multiple labels, where each object is represented as a bag of instances. A key advantage...
详细信息
Multi-instance multi-label learning (MIML) is a weakly supervised approach that models relationships between complex objects and multiple labels, where each object is represented as a bag of instances. A key advantage of MIML is its ability to perform both bag-level and instance-level multi-label predictions, relying solely on bag-level labels. However, a significant performance gap persists between instance-level MIML algorithms and fully supervised learning approaches due to the lack of instance-level labels. Existing MIML algorithms address this challenge by treating bag labels as ambiguous and attempting to reduce supervision imprecision. Moreover, they often assume that instances are independent and identically distributed (i.i.d.) and rely on prior knowledge to learn label correlations, which is impractical in real-world scenarios. To address these challenges, we propose MIMLVAE, a novel dual-granularity MIML algorithm based on a variational autoencoder. MIMLVAE employs a graph attention network to dynamically capture label correlations and instance dependencies, eliminating the i.i.d. assumption and prior knowledge. By treating all instances within a bag equally, it infers effective bag-level and instance-level representations for dual-granularity prediction. At the same time, the label encoder captures label-specific prototype representations, facilitating prototype-based classification at both the bag and instance levels without requiring label disambiguation. Furthermore, MIMLVAE integrates Gaussian mixture model into the shared latent space of features and labels, mitigating posterior collapse and over-regularization. Experiments on six standard MIML datasets demonstrate that MIMLVAE significantly outperforms state-of-the-art methods in both bag-level and instance-level multi-label classification tasks.
In recent years, the integration of hyperbolic geometry with Graph Neural Networks (GNNs) has garnered significant attention due to its effectiveness in capturing complex hierarchical structures, particularly within r...
详细信息
In recent years, the integration of hyperbolic geometry with Graph Neural Networks (GNNs) has garnered significant attention due to its effectiveness in capturing complex hierarchical structures, particularly within real-world graphs and scale-free networks. Although hyperbolic neural networks have shown strong performance across various domains, most existing models rely on static graph structures, limiting their adaptability to dynamic data. Previous studies have primarily focused on improving the modeling capacity of hyperbolic space for latent space data during training, often neglecting the preservation of high-order intrinsic features before training. To address this, we propose a novel Dynamic Hypergraph Hyperbolic Neural Network (DHHNN) based on a variational autoencoder for multimodal data integration. This model combines the advantages of hyperbolic geometry, dynamic hypergraphs, and the self-attention mechanism to enhance multimodal data representation learning. DHHNN introduces a dynamic hypergraph framework that continuously adjusts the relationships between hypernodes and hyperedges during training, effectively capturing higher-order dependencies within complex networks. Furthermore, the self-attention mechanism dynamically regulates the dependency levels between hypernodes and hyperedges, enhancing the model's ability to capture long-range dependencies and complex feature interactions. Leveraging the negative curvature of hyperbolic space, DHHNN compactly and accurately represents complex scale-free networks. Experimental results on seven benchmark datasets and latent space visualizations demonstrate that DHHNN provides compact and effective data representation, outperforming existing models and achieving state-of-the-art performance in node classification tasks.
Among vibration-based structural health monitoring (SHM), methods based on modal analysis are the most popular and efficient ones for assessing the inherent states of structures. The primary techniques involve trackin...
详细信息
Among vibration-based structural health monitoring (SHM), methods based on modal analysis are the most popular and efficient ones for assessing the inherent states of structures. The primary techniques involve tracking changes in modal parameters regularly during the monitoring period. Numerous studies have utilized both modal frequencies and mode shapes for damage detection, as mode shapes help identify damage locations, and the combination is more robust against environmental effects. Obviously, recent advances in machine learning (ML) and artificial intelligence (AI) have revolutionized SHM based on modal analysis. This study aims to develop a clustering technique using variational autoencoder (VAE) for classifying structural damages. To fully expand the entangled spaces that may arise from lower-dimensional representation, the loss function incorporates triplet loss (known as triplet VAE or tVAE) while preserving all dimensions in the encoded space. The proposed approach is thoroughly developed and evaluated using both numerical simulations and experimental investigations of multistory building structures. The effectiveness of detecting changes via VAE-based and tVAE-based modal clustering has been demonstrated under the consideration of environmental variations and uncertainties. Additionally, a procedure is suggested for the field applications and the proposed approach's capability for structural inspection during long-term monitoring is showcased using a practical scenario.
Geophysical methods, such as electrical resistivity tomography (ERT), can be used to image near-surface electrical resistivity, as field measurements depend on subsurface porosity, water saturation, and fluid salinity...
详细信息
Geophysical methods, such as electrical resistivity tomography (ERT), can be used to image near-surface electrical resistivity, as field measurements depend on subsurface porosity, water saturation, and fluid salinity. ERT has been widely applied to investigate mineral and groundwater resources and in archaeological, environmental, and engineering studies. The prediction of subsurface electrical conductivity from ERT data requires solving a geophysical inverse problem. For near-surface characterization studies, this is often accomplished with deterministic inverse methods. These methods linearize the problem around an initial solution, and their smoothness depends on an imposed a priori spatial regularization term. Depending on this parameterization, these methodologies might struggle to capture the natural variability of the subsurface. Moreover, deterministic solutions have limited capabilities for uncertainty assessment. In contrast, stochastic inverse methods can assess uncertainties by predicting multiple model realizations that fit the recorded ERT data similarly. However, they are often more computationally expensive than deterministic solutions. Deep-learning algorithms based on deep generative models have been used to reparameterize model and data spaces into low- dimensional domains and efficiently solve geophysical inverse problems. However, within this context, uncertainty assessment is challenging. We develop a deep convolutional variational autoencoder (VAE) coupled with stochastic adaptive optimization to perform stochastic ERT inversion. Geostatistical simulations of electrical resistivity are used as the training data set of the VAE. After training, the VAE generates electrical resistivity models that reproduce the statistics and spatial continuity patterns of the training data set. Then, the VAE latent space is iteratively perturbed and updated with adaptive stochastic sampling based on the misfit between observed and predicted ERT data. Our method
A ventilated acoustic resonator (VAR), a class of acoustic metamaterial, offers an effective noise reduction solution in urban environments, where ventilation is essential. The significant non-linearity between the VA...
详细信息
A ventilated acoustic resonator (VAR), a class of acoustic metamaterial, offers an effective noise reduction solution in urban environments, where ventilation is essential. The significant non-linearity between the VAR structure and the corresponding acoustic response poses substantial challenges to the inverse design of VAR through analytical approaches. Although deep learning has been employed to address the inverse design of such structures, conventional deep learning-based inverse design methods are constrained to parameter-defined structures, which exhibit limited design flexibility, thereby hindering the achievement of highly accurate inverse design. To address this challenge, we propose a novel inverse design framework for non-parametric VAR, composed of a dual-variational autoencoder (Dual-VAE) and iterative transfer learning method. Dual-VAE consists of structure variational autoencoder and acoustic response variational autoencoder with their latent space aligned to facilitate accurate inverse design of non-parametric VAR structures exhibiting the target acoustic response. The iterative transfer learning method is employed to enhance the inverse design performance of the Dual-VAE by progressively augmenting the initial training dataset, which consists solely of parametric VAR structures, with generated non-parametric VAR structures. The transfer-learned Dual-VAE demonstrated approximately a 32.27 % reduction in mean squared error with the target acoustic response compared to the Dual-VAE trained solely on the initial parametric VAR dataset. We present a novel approach to the inverse design of complex structures with high non-linearity by introducing the inverse design framework, demonstrating exceptional performance in generating non-parametric structures that achieve target performance.
The challenges in multi-objective and multi-dimensional optimization design of airfoils, marked by prolonged optimization cycles and low accuracy, call for an efficient solution to expedite airfoil design. This study ...
详细信息
The challenges in multi-objective and multi-dimensional optimization design of airfoils, marked by prolonged optimization cycles and low accuracy, call for an efficient solution to expedite airfoil design. This study presents an innovative airfoil generative design model based on a conditional variational autoencoder (CVAE). Initially, to overcome the limitation of insufficient training data, the model leverages the variational autoencoder (VAE) to learn the spatial distribution of University of Illinois at Urbana-Champaign (UIUC) airfoils, enabling the generation of a diverse set of airfoils with similar distributions. Subsequently, two CVAE-based airfoil generation models, the airfoil freedom design model and the airfoil precision design model, are proposed, which can realize diverse airfoil design under different conditions, such as shape and aerodynamic conditions. Furthermore, two measurements of roughness and diversity are introduced to evaluate the quality of the generated airfoils. The impact of different conditions and network parameters on the model's generation performance is thoroughly analyzed. Results indicate that our proposed model achieves a 65% lower error compared to physics-guided conditional Wasserstein generative adversarial networks (PG-cWGAN) when generating airfoils that satisfy a specific lift coefficient and a 99.99% lower error compared to airfoil pressure distributions generative adversarial networks (Airfoil-Cp-GAN) when generating airfoils that satisfy specific pressure distributions. This method introduces amore creative and accurate approach for aircraft designers in the realm of airfoil design. The code used for this paper is available at https://***/liujun39/airfoilvae.
Compared to traditional single-behavior models, multibehavior recommendation models incorporate the auxiliary behavior information of users. This integration step addresses the cold-start and data sparsity issues and ...
详细信息
Compared to traditional single-behavior models, multibehavior recommendation models incorporate the auxiliary behavior information of users. This integration step addresses the cold-start and data sparsity issues and provides more comprehensive and detailed interaction information for the model. Despite the efforts made by multibehavior recommendation models to analyze user behavior semantics and capture user preferences, challenges remain in terms of effectively modeling the relationships between different types of user feedback. This issue is exacerbated by the heavy reliance on hyperparameters, which leads to overparameterization. In this paper, we propose a variational autoencoder (VAE) and graph-based heterogeneous multibehavior recommendation model (V-GMR), which aims to capture user behavior preferences and mitigate the aforementioned issues. First, we employ VAEs to encode user behaviors and learn feature representations that effectively capture multibehavior information. Second, we develop a preference fusion enhancer based on a VAE to integrate auxiliary user behaviors with the target behavior, effectively addressing the problem concerning sparse interaction data. Third, we design a special behavior decoding layer to handle the latent variables acquired from the preference fusion enhancer. In this layer, we reconstruct the loss function and resolve the issue of optimizing the neural network parameters through backpropagation in the presence of deterministic input values. The effectiveness of V-GMR is validated through experiments conducted on three real-world datasets, and the contributions of the V-GMR model components are verified through ablation experiments.
Implementing efficient automatic fault diagnosis is critical for saving energy and minimizing financial losses in the heating ventilation air-conditioning (HVAC) systems of commercial buildings. However, the limited q...
详细信息
Implementing efficient automatic fault diagnosis is critical for saving energy and minimizing financial losses in the heating ventilation air-conditioning (HVAC) systems of commercial buildings. However, the limited quantity and weak features of fault samples acquired during HVAC operations hinder the effectiveness of conventional machine learning -based fault diagnosis methodologies. This paper proposes a method based on an improved conditional variational autoencoder (MCVAE) and co -training (CT) ideology to address the issue of insufficient training samples. Initially, we employ MCVAE to synthesize an extensive dataset of chiller fault samples from the original training dataset. Subsequently, the beneficial samples for training our fault diagnostic classifier, namely high -quality samples, are selected from the generated dataset using the CT -based framework. Finally, the selected high -quality samples are merged into the original training dataset to train the ultimate fault classifiers. Experimental results demonstrate that our proposed method outperforms in effectiveness and efficiency compared to recently published methods. For instance, in the case of fault level 1 compared to the suboptimal model, our approach exhibits improvements of 2.41% when each type has 5 fault samples.
Recent advancements in deep neural networks have shown great potential in generating realistic data and performing clustering tasks. This is due to their ability to capture intricate patterns. However, current generat...
详细信息
Recent advancements in deep neural networks have shown great potential in generating realistic data and performing clustering tasks. This is due to their ability to capture intricate patterns. However, current generative models face challenges such as poor performance and computational complexity caused by the issue of dimension disaster. The variational autoencoder (VAE), a commonly used method, also encounters problems such as posterior collapse and poor performance in multiclass classification when using the latent variables of VAE. Our goal in this study is to tackle the issue of effective disentanglement in image generation, classification and clustering tasks. We develop a generative network based on VAE incorporating a Gaussian mixture distribution as the prior. This enhancement improves the representation of latent variables and helps to overcome the challenges of matching the ground truth posterior. To further improve clustering performance, we introduce the total correlation as a kernel for computing latent features between embedding points and cluster centers. This technique is particularly useful in cases with complex latent variables and can also be applied for hierarchical disentanglement. Moreover, we employ the Fisher discriminant as a regularization term to minimize the within-class distance and maximize the between-class distance for samples, which has an important effect on the performance of our model viewed from the experimental results. We evaluate our proposed network on four datasets, and the experimental results demonstrate its effectiveness across multiple metrics.
In Emotion Recognition in Conversations (ERC), the emotions of target utterances are closely dependent on their context. Therefore, existing works train the model to generate the response of the target utterance, whic...
详细信息
In Emotion Recognition in Conversations (ERC), the emotions of target utterances are closely dependent on their context. Therefore, existing works train the model to generate the response of the target utterance, which aims to recognise emotions leveraging contextual information. However, adjacent response generation ignores long-range dependencies and provides limited affective information in many cases. In addition, most ERC models learn a unified distributed representation for each utterance, which lacks interpretability and robustness. To address these issues, we propose a VAD-disentangled variational autoencoder (VAD-VAE), which first introduces a target utterance reconstruction task based on variational autoencoder, then disentangles three affect representations Valence-Arousal-Dominance (VAD) from the latent space. We also enhance the disentangled representations by introducing VAD supervision signals from a sentiment lexicon and minimising the mutual information between VAD distributions. Experiments show that VAD-VAE outperforms the state-of-the-art model on two datasets. Further analysis proves the effectiveness of each proposed module and the quality of disentangled VAD representations.
暂无评论