ABSTRACT Link prediction has long been a fundamental task in graph data analysis, aimed at identifying potential or missing connections between nodes. This task is particularly important for understanding social dynamics and improving the robustness of networks. However, existing graph learning methods often struggle with modeling complex structures, especially in social networks where data is noisy, uncertain, as well as multi‐faceted. Most traditional graph embedding architectures are limited in their ability to handle noise and effectively represent multi‐view information. Likewise, there are several graph neural networks (GNNs) face challenges in integrating diverse structural perspectives and managing uncertainty, which leads to subpar performance in link prediction tasks. To overcome these challenges, we propose a novel model called AFGRL which is an attention‐driven fuzzy graph representation learning. Our proposed AFGRL combines multiple types of GNNs for multi‐view embedding with a multi‐headed attention‐enhanced fuzzy neural network. This design enables our AFGRL model to better learn richer, as well as more expressive graph representations while effectively managing uncertainty and noise. The attention mechanism integrated into our AFGRL model allows it to focus on various structural aspects of the graph—while the fuzzy logic component captures ambiguity inherent in social network data. Our model is particularly tailored for online social networks—where user relationships are dynamic and characterized by varying degrees of trust and influence. We evaluate AFGRL on several real‐world and benchmark datasets, demonstrating its superior performance in link prediction compared to state‐of‐the‐art baselines. The results confirm that our AFGRL model not only enhances predictive accuracy but also provides robust and meaningful structural representations; as a result, highlighting the value of integrating attention and fuzzy logic into graph learning frameworks.

ABSTRACT Recent advances in deep learning (DL) have significantly improved the performance of time‐series forecasting tasks. While recurrent neural networks (RNNs) have traditionally served as the foundation for such models, DL/RNN‐based forecasting models have frequently struggled with capturing long‐range temporal dependencies as well as handling noise, particularly in multivariate and high‐dimensional settings. Transformer‐based architectures have emerged as promising alternatives due to their ability to model complex temporal patterns across extended time‐dependent sequences. However, most existing transformer‐based forecasting techniques often face limitations in mitigating feature noise as well as ambiguity during the long‐sequence representation learning process. To overcome these challenges, in this paper, we propose a novel DT4TS model—which is a denoising transformer‐based architecture that integrates a cross‐time/dimension embedding mechanism with a radial basis function neural network (RBFNN) layer to effectively enhance noise suppression during the temporal feature extraction process. To evaluate the effectiveness of our proposed DT4TS model, we evaluate DT4TS on a real‐world air quality dataset, focusing on PM2.5 prediction across two monitoring stations in Ho Chi Minh City, Vietnam. On average across datasets, DT4TS reduces RMSE by 25.45% and MAE by 16.52% compared with Crossformer, with even larger gains over Autoformer and FEDformer which are known as state‐of‐the‐art transformer‐based architectures for time‐series learning. These results demonstrate the superior accuracy and robustness of our proposed DT4TS model within the long‐range air quality forecasting problem; as a result, confirming the effectiveness of its noise‐resilient embedding design in capturing fine‐grained temporal dependencies over extended horizons.

In recent years, there are multiple temperature predictive models have demonstrated the effectiveness and outperformances of applying different deep neural architectures, such as convolutional neural network (CNN) and recurrent neural network (RNN) for the temperature forecasting task in forms of time-series analysis problem in comparing with previous traditional machine learning based techniques. However, up to this time, there are still several limitations of existing deep learning-based temperature predictive methods related to the capability of efficiently integrating extra information resources into the temperature data learning process. Moreover, the high-noised/chaotic fluctuations within daily temperature data also lead to downgrades in the accuracy performance of recent deep learning-based techniques. To overcome these challenges, in this paper, we proposed a novel integrated dual attention mechanism with the Convolutional Long Short-Term Memory Network (LSTM), called as: DAttConvLSTM. Our proposed DAttCovLSTM supports to effectively capture the chaotic and dynamic temporal information from daily temperature data, thus significantly improve the prediction results. Extensive experiments and comparative studies in real-world datasets demonstrated the effectiveness of our proposed model in comparing with recent state-of-the-art baselines.

In recent years, with tremendous progresses of deep learning in multiple disciplines, there are several advanced sequential neural-network (NN) based architectures (e.g., recurrent neural network—RNN, Auto-Encoding—AE, transformer, etc.) have been proposed. Recently, there are several well-known GNN-based architectures like as graph convolutional network (GNN) have been proposed to deal with challenges related to the global representation preservation of text. However, most of recent proposed GNN-based text-embedding models still be unable to integrate the global structure with the semantic sequential representations of words/sentences into the unified textual embedding space. Moreover, they are also considered as unable to learn the rich context-varied representations of words. In order to tackle aforementioned challenges, in this paper we proposed a novel integrated text graph representation learning approach, named as: GOWSeqGCN. Our proposed GOWSeqGCN is an integrated semantic graph-of-words sequential textual representation under the graph convolutional network framework. In order to demonstrate for the effectiveness of our proposed GOWSeqGCN model in comparing with recent state-of-the-art text representation learning baselines, we conducted extensive experiments in benchmark textual datasets. The experimental outputs showed the outperformances and necessary of our proposed ideas in this paper.

Among common tasks in natural language processing (NLP) domain, text classification is considered as an important primitive task which is widely applied in multiple disciplines. Recent advanced deep learning-based architectures such as sequence-to-sequence (seq2seq) with attention mechanism have demonstrated remarkable improvements in multiple NLP’s tasks, including classification. However, recent seq2seq-based models still encounter challenges related to the limitation in effectively capturing long-range dependent relationships between words in a text corpus. Recent integrated graph neural network and textual graph transformer (TGT)-based models have demonstrated significant improvements in preserving the structural n-hop co-occurring relationships between words in a given text corpus. However, these models still suffer problems related to the thorough considerations on the sequential and contextual relations of words within a single document’s graph. To meet these challenges, in this article we proposed a novel semantic-enhanced graph transformer-based textual representation learning approach, called as: SemTGT. Our proposed SemTGT can support to effectively learn both local rich-contextual and global long-range structural latent representations of texts for leveraging the performance of classification task. Extensive experiments in standard datasets demonstrate the effectiveness of our proposed SemTGT model in comparing with recent seq2seq-based and textual graph embedding-based baselines.

In recent years, along with the dramatic developments of deep learning in the natural language processing (NLP) domain, notable multilingual pre-trained language techniques have been proposed. These recent multilingual text analysis and mining models have demonstrated state-of-the-art performance in several primitive NLP tasks, including cross-lingual text classification (CLC). However, these recent multilingual pre-trained language models still suffer limitations regarding their adaptation for specific task-driven fine-tuning in the context of low-resource languages. Moreover, they also encounter problems related to the capability of preserving the global semantic (e.g., topic, etc.) and long-range relationships between words to better fine-tune and effectively handle the cross-lingual text classification task. To meet these challenges, in this article, we propose a novel topic-driven multi-typed text graph attention–based representation learning method for dealing with the cross-lingual text classification problem called TG-CTC. In the proposed TG-CTC model, we utilize a novel fused topic-driven multi-typed text graph representation to jointly learn the rich-schematic structural and global semantic information of texts to effectively handle the CLC task. More specifically, we integrate the heterogeneous text graph attention network with the neural topic modelling approach to enrich the semantic information of learned textual representations in the context of multiple languages. Extensive experiments in benchmark multilingual datasets showed the effectiveness of the proposed TG-CTC model compared with the contemporary state-of-the-art baselines.

Recently, rapid growth of social networks and online news resources from Internet have made text stream clustering become an insufficient application in multiple domains (e.g.: text retrieval diversification, social event detection, text summarization, etc.) Different from traditional static text clustering approach, text stream clustering task has specific key challenges related to the rapid change of topics/clusters and high-velocity of coming streaming document batches. Recent well-known model-based text stream clustering models, such as: DTM, DCT, MStream, etc. are considered as word-independent evaluation approach which means largely ignoring the relations between words while sampling clusters/topics. It definitely leads to the decrease of overall model accuracy performance, especially for short-length text documents such as comments, microblogs, etc. in social networks. To tackle these existing problems, in this paper we propose a novel approach of graph-of-words (GOWs) based text stream clustering, called GOW-Stream. The application of common GOWs which are generated from each document batch while sampling clusters/topics can support to overcome the word-independent evaluation challenge. Our proposed GOW-Stream is promising to significantly achieve better text stream clustering performance than recent state-of-the-art baselines. Extensive experiments on multiple benchmark real-world datasets demonstrate the effectiveness of our proposed model in both accuracy and time-consuming performances.

Recently, advanced techniques in deep learning such as recurrent neural network (GRU, LSTM and Bi-LSTM) and auto-encoding (attention-based transformer and BERT) have achieved great successes in multiple application domains including text summarization. Recent state-of-the-art encoding-based text summarization models such as BertSum, PreSum and DiscoBert have demonstrated significant improvements on extractive text summarization tasks. However, recent models still encounter common problems related to the language-specific dependency which requires the supports of the external NLP tools. Besides that, recent advanced text representation methods, such as BERT as the sentence-level textual encoder, also fail to fully capture the representation of a full-length document. To address these challenges, in this paper we proposed a novel s emantic-ware e mbedding approach for ex tractive text sum marization , called as: SE4ExSum. Our proposed SE4ExSum is an integration between the use of feature graph-of-words (FGOW) with BERT-based encoder for effectively learning the word/sentence-level representations of a given document. Then, the g raph c onvolutional n etwork (GCN) based encoder is applied to learn the global document's representation which is then used to facilitate the text summarization task. Extensive experiments on benchmark datasets show the effectiveness of our proposed model in comparing with recent state-of-the-art text summarization models.