In action recognition, while combining spatio-temporal videos with skeleton features can enhance recognition performance, it necessitates distinct models and balanced feature representations for cross-modal data. Addressing these challenges, we introduce the Deep Residual Vision Transformer (Deep-RVT), a novel architecture that combines the training stability and resilience provided by residual connections with the representational power of Vision Transformers in a seamless manner. The spatio-temporal dynamics required for identifying human actions in video sequences are precisely captured by DeepRVT. Our model efficiently propagates low-level features across layers by embedding residual pathways into the Transformer blocks, which resolves the vanishing gradient issue and speeds up the training process. Our design leverages the integration of local and global features to encode subtle motion cues and spatial arrangements, which are essential for identifying a broad range of human activities. The proposed network is validated by lots of mainstream benchmarks. Many experimental results, conducted on the Penn-Action, ImageNet-1K and ImageNet22K, show that the proposed network outperforms most state-of-the-art methods. Our code is available at https://anonymous.4open.science/r/ RVT- 6D6F/README.md.
Questioning is an effective method of extracting information from documents. While Large Language Models excel at answering user queries in document question-answering systems, users often struggle to formulate effective questions when encountering unfamiliar documents. Fortunately, expert-formulated questions embodying professional knowledge can be transferred to new documents to assist ordinary users. Therefore, we propose a question recommendation approach that transfers expert questions on historical documents to new ones, enabling users to "stand on the shoulders of giants" for enhanced document comprehension. Our approach comprises two modules: 1) A "question reusability classification" module identifies domain-general questions applicable across similar documents; 2) A "document-bridged question ranking" module selects semantically appropriate questions for new documents. Experiments on our self-constructed expert question dataset demonstrate that both components significantly impact recommendation accuracy, and performance improves as historical data volume increases.
In the dynamic field of data analytics, handling diverse, large-scale datasets poses substantial obstacles, with traditional methods, like data warehousing and manual programming, proving insufficient due to limitations in flexibility and efficiency. Dataframes, facilitated by libraries like Pandas, have gained significant popularity. However, their utility is hampered by limitations in quick construction and modification of queries by non-programming experts, as well as handling large datasets and lacking support for big data without additional libraries. To address these gaps, we introduce a novel model-based approach for dataframe analysis based on the Data Virtual Machine (DVM) framework. Building on this framework, we facilitate the construction and analysis of dataframes queries without requiring programming expertise, lowering the entry barrier for beginners and streamlining the analysis process for experts. The proposed tool facilitates easy formulation and modification of dataframe queries, efficient query evaluation, and demonstrates superior performance compared to traditional Pandas implementation as evidenced by our experimental results.
Understanding user navigation patterns from clickstream data is crucial for improving business software, yet remains challenging due to the complexity and variability of real-world environments. Unlike controlled settings, real-world clickstreams are noisy, fragmented, and often incomplete, due to session timeouts, network issues, caching, or third-party interactions-making it difficult to reconstruct coherent user journeys. Additionally, the absence of labeled data hinders the use of supervised learning, pushing researchers toward unsupervised or heuristic-based approaches that struggle to fully capture user behavior. In this paper, we present a benchmark of embedding techniques for modeling user navigation behavior on task-oriented software. We identify distinct user behaviors across three real-world case studies. Results show that Pattern2Vec outperforms Word2Vec in capturing meaningful task-based navigation patterns, confirming its suitability for clickstream analysis.
The demand for environmentally friendly cloud computing is on the rise, leading cloud service providers to focus on reducing carbon emissions by using renewable energy sources and energy-efficient computing models. This study assesses the performance and energy consumption of serverless and serverful architectures, specifically looking at join operations using Apache Spark for big data processing in a private cloud combined with Kubernetes. By using the TPC-DS benchmark, we examine the impact of cold-start and warm-start phases in the serverless environment, as well as the auto-scaling capabilities of Spark in serverless environments within the private cloud. The results show that the efficient and flexible resource management in serverless environments in private clouds leads to more optimal processing times and energy consumption compared to serverful architectures, especially in warm-start scenarios. These findings offer valuable insights for organizations seeking to streamline their big data infrastructure while also making a positive environmental impact within the IT industry.
Efficient autoscaling in Kubernetes (K8s)-managed in-memory systems like Redis remains a critical challenge, especially under highly dynamic workloads. Traditional threshold-based mechanisms (e.g., HPA) often fail to anticipate sudden demand surges, leading to poor performance and inefficient resource use. We introduce DInos, a Deep Reinforcement Learning (Deep RL) agent enhanced with LSTM layers and transfer learning, designed for proactive and adaptive autoscaling in Kubernetes. As an evolution of our earlier agent DERP, DInos leverages temporal workload modeling and pre-trained policies to generalize across deployments with minimal retraining. DInos utilizes a customizable reward function balancing throughput, latency, resource usage, and pod efficiency. DInos achieves up to 17.3 × higher rewards in simulation and a 5.5 × improvement in real-world K8s-Redis deployments by forecasting spikes, optimizing pod counts and maintaining low latency, providing a robust autoscaling solution for volatile, cloud-native environments.
Autism Spectrum Disorder (ASD) is a neurodevelopmental disability that significantly increases the difficulties and risks associated with driving. Individuals with ASD often face a variety of challenges, such as increased sensory sensitivities, difficulty adapting to changing environments, and struggles with unexpected situations on the road. These difficulties can lead to sensory overload, panic attacks, and impaired decision-making, all of which increase the risk of accidents and make driving an especially overwhelming task. In this paper, we propose a novel IoT-based smart mobility framework for predictive stress detection in drivers with ASD, enabling the early identification of potential stressors before they encounter them on the road. This approach leverages AI-based models, including LSTM and CNN-based architectures. Unlike existing methods that focus on reactive stress detection, which may be too late, our approach predicts stress triggers in advance, enabling timely and preventive support.
Microservices pose challenges for automated fault resolution due to their distributed and complex nature. We present SysResolve, a framework that automates the entire resolution pipeline by combining multi-modal Root Cause Analysis (RCA) with Large Language Models (LLMs). RCA outputs are converted to natural language and passed through a Retrieval-Augmented Generation (RAG) pipeline to produce executable scripts. We evaluated and experimented on two microservices applications with three LLM (LlaMa3-70B, GPT-4, Claude 3.7). Our analysis highlights significant gains of current LLMs generation power from few-shot learning, with SysResolve achieving expert-level remediation while reducing recovery time.
People could use various types of subtle metaphorical language forms to express individual opinions or negative emotion in daily life, such as sarcasm, satire or irony. Sarcasm employs contradictory and incongruous elements to convey the difference between reality and expectation. Detecting precisely whether comments, opinions or conversations have sarcastic intention is crucial for understanding the talkers' feeling and attitude. Since the complexity of sarcasm requires certain contextual information to catch on ironic meaning correctly, how to leverage multimodal information to grasp sarcastic remarks has become a hot research topic. A Multimodal Sarcasm Detection method based on Contrastive Learning (MSDCL) is proposed in this paper to exploit multimodal information and incongruity knowledge for improving performance. MSDCL extracts fine-grained features from text and image and captures contradictory semantics between text and image by using the multi-head self-attention mechanism. A supervised contrastive learning is implemented to better learn intra- and inter-class relationships among sarcastic and non-sarcastic data, where the classification loss and contrastive learning loss are integrated to guide the multimodal data embedding. The experiment results show MSDCL outperforms the compared methods.
Detecting manipulations in digital images is critical for ensuring their authenticity and integrity. This study evaluated the impact of frequency descriptors (DCT/FFT) derived from Error Level Analysis (ELA) maps on manipulation detection performance. Using Random Forest, SVM, XGBoost, and LightGBM on IMD2020, we compared models trained with baseline original image features against those augmented with ELA map features. The results show that ELA-derived features significantly improved the tree-based models. XGBoost and LightGBM yielded the best performance (F1-score ≈ 0.81 validation; XGBoost 0.815 test), demonstrating strong generalization. This study highlights the informative value of combining ELA maps and original image frequency analysis for effective classical machine learning-based detection.
Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) are two prominent generative models that excel in different aspects of generative tasks of creating new content, such as images, text, or music using data and patterns. VAEs are known for their capability to learn smooth and probabilistic latent representations, which enable structured generation, but they often produce blurry and unrealistic outputs due to their inherent loss function. On the other hand, GANs generate sharp and visually appealing results by utilizing a discriminator to refine output but suffer from mode collapse, limiting their diversity. This paper proposes a novel hybrid architecture, the Variational Cyclic Generative Adversarial Network (VarCGAN), to transfer the musical style of a song from one genre to another. VarCGAN combines the latent space modeling capabilities of VAEs with the adversarial optimization of GANs to overcome the limitations of each approach. The model introduces a cyclic adversarial loss, which ensures consistency and realism in style-transferred outputs while preserving the original song's musical essence. Furthermore, the hybrid design enables the generation of diverse variations of genre-transferred songs, capturing subtle stylistic features of the target genre. The proposed approach is evaluated on the GTZAN dataset, focusing on style transfers between classical, jazz, hip-hop, and rock genres. To the best of our knowledge, VarCGAN is the first framework to utilize this hybrid methodology for music genre style transfer, presenting a significant advancement in music composition and genre transformations.
Large language models (LLMs) have made significant strides in text-to-SQL performance, transforming natural language queries into SQL queries. The BIRD benchmark, a cross-domain dataset with 12,751 question-SQL pairs across 95 databases, is currently the most challenging benchmark in the field. However, its leaderboard is dominated by solutions that rely on closed-source LLMs, creating a financial barrier for researchers. This paper focuses on leveraging open-source models and presents an incremental validation methodology, called Wrecking SQL, that in six steps, incrementally modifies the schema of datasets by replacing meaningful column and table names with meaningless ones-a realworld problem found in legacy SQL systems. We explore how meaningless names affect LLM accuracy and demonstrate that inferring meaningful names improves translation accuracy.
The rise of big data and social networks necessitates using multiple data models, including NoSQL, leading to complex database management due to their schemaless nature and ambiguous structures. Graph data models offer unique flexibility to facilitate. This paper proposes a graph-centric multi-model DBMS leveraging a unified schema (U-Schema) to integrate heterogeneous data into a cohesive graph. This polystore architecture provides a transparent interface for querying as a unified graph, seamlessly supporting data migration and dynamic data placement. Qualitative and quantitative evaluations show transparent query processing and significant impact of data migration on execution speed, indicating optimization potential.
Emotional Support Conversation (ESC) aims to ease a helpseeker' psychological distress through dialogue with a supporter. We propose CasDecNet (Cascading Fusion-Guided Decoding Network), a novel response generation model that enhances empathy in three stages: understanding context, acquiring cognitive and affective empathy, and incorporating strategic support. Our model uses multi-dimensional external knowledge to better infer the user's situation and implement empathy effectively. Experiments on the ESConv dataset show that CasDecNet outperforms baselines in both automatic and LLM-based evaluations, especially in terms of response diversity and empathy, indicating its promise for human-like emotional support.
The significance of early vulnerability identification in ensuring security during software development cannot be denied. In this research, we introduce CWEpredBELL, a unique automated vulnerability prediction method that makes use of a modified pre-trained language model derived from CodeBERT. With a binary classification layer, an improved optimizer, and a fine-tuned loss function to boost model performance, our method is especially tailored for identifying vulnerabilities in source code. We used cross-validation techniques and the Local Interpretable Model-Agnostic Explanations (LIME) approach to identify particular lines of error in the source code. The experimental comparison demonstrates that CWEpredBELL is an effective method of automatically identifying vulnerabilities.
Predicting the Remaining Useful Life (RUL) in maintenance often encounters challenges such as high dimensionality, feature redundancy, and limited explainability. This paper presents a novel approach that combines Interpretable Divisive Feature Clustering (IDFC) with Long Short-Term Memory (LSTM) networks. The IDFC algorithm leverages the strengths of variable clustering methods (VARCLUS) and the Clustering of Variables around Latent Components (CLV) to identify significant features and non-orthogonal latent components. This method enables effective dimensionality reduction by selecting key features rather than combining them. Integrating IDFC with a single-layer LSTM and Shapley Additive Explanations (SHAP) results in a robust and interpretable framework for RUL prediction, achieving a balance between accuracy and transparency. Experimental results on a bearing dataset show that the IDFC + LSTM model outperforms traditional methods while enhancing interpretability through the identification of key energy-related features which influence the RUL prediction more.
Food waste is a serious problem in modern society. A specific aspect of food waste concerns meat consumption in gastronomy, where typically only prime cuts of meat are used in the kitchen. To facilitate the usage of all parts of animals and thereby reducing food waste, we present Onfoods, a system that recommends alternative meat cuts in recipes and integrates inventory data to help with the creation of menus. Onfoods uses an ontology and a knowledge graph to model recipes, meat cuts and the relationships between the two, similarity measures to find candidates for alternative meat cuts, and inventory data to track the availability of different meat cuts. An intuitive user interface allows the user on one hand to update the knowledge graph and inventory data, and on the other hand to navigate through recipes and choose alternative meat cuts.
Successful data analytics implementation requires seamless access to both data and related metadata. In many organizations, analytics challenges arise from Data Silos, which impede cross-functional access to data and knowledge sharing across the organization. This article presents practical insights from a data architecture transformation project conducted at a large institution with over 1,400 employees and overseeing over 2,000 market entities. The organization faced significant analytical and operational challenges due to the presence of Data Silos–isolated repositories associated with specific business areas. To address these limitations, the institution initiated a transition to a Data Mesh architecture to improve data availability and enhance analytical capabilities. This article explains the rationale behind the persistence of silos, evaluates alternative architectural models, and justifies the choice of Data Mesh based on organizational context. Key elements of the transformation include developing a data management framework, implementing a data catalog, creating a data lake to provide data input flexibility, and establishing a common analytics platform based on Data Domains. While the project is still ongoing, the paper describes the methods being implemented and shares early results, key learnings, and practical recommendations for institutions undertaking similar architectural transitions.
Highway monitoring is crucial to ensuring the safety, efficiency, and maintenance of road networks. This process allows for the identification of dangerous traffic, rapid detection of accidents, management of vehicle flow, and route optimization, among other benefits. For public agencies, monitoring facilitates the collection of traffic data, which is essential for complying with laws and preventing incidents. However, the large volume and speed of data can make processing difficult, impacting decision-making. In this context, parallel processing stands out as an effective solution, especially in image analysis, through the division of subtasks and the use of GPU. Using virtual machines with different memory and processing resources, the delays in image processing were analyzed, using queueing theory to propose machine configurations that allow processing close to real time to minimize the delays generated during the processing of these images. The analysis showed that the use of GPU significantly reduces processing time; however, the use of machines with CPU obtained satisfactory performance from the orchestration of different machines in parallel. It is concluded that parallel processing, combined with queueing theory, optimizes highway monitoring, balancing performance and resource costs.
Unstructured data, such as textual data, is prevalent on digital platforms such as social media, with users posting content that sometimes includes sensitive information. For instance, in March of 2025, a press story by journalist Jeffrey Goldberg indicated that secret Trump administration plans for an air strike on Yemen had inadvertently been leaked to him ( U.S. national-security leaders included me in a group chat about upcoming military strikes in Yemen... ). Despite regulations, and the existence of access control policies data leaks such as this are persistent. One of the reasons for this is that contextually analysing unstructured data to determine whether or not sensitive information exists therein, is a challenging problem. Existing approaches identify individual elements of sensitive information such as Personally Identifiable Information (PII), but few works look at the issue of document sensitivity based on the occurrence of sensitive information. In this paper we propose a novel approach to classifying documents as either public or private, based on the occurrence of, and contextual interpretation of sensitive information within the document. We employ an ensemble model composed of transformer-encoder models and standard machine learning models to support the classification process. Our empirical study, conducted on a curated dataset composed of ENRON emails and Tweets of U.S. Congress members, indicates that the Random Forest algorithm when combined with the BERT model classifies all public documents correctly (i.e. 100
