Abdul Wahid

Assessing the influence of scientific citations is critical to understanding the impact of scholarly research beyond the raw citation counts. Recent neural models achieve strong performance in citation classification but typically function as black boxes, limiting their suitability for transparent bibliometric evaluation. This study proposes an intrinsically interpretable framework for binary classification of in-text citation importance. The framework integrates a SciBERT encoder with a sparsity-controlled HardKuma-L₀ rationale selector. This selector identifies a minimal yet contextually informative set of tokens without requiring additional annotations, enabling the model to generate faithful, human-readable rationales that directly influence predictions. Experiments on the ACL-ARC and SciCite datasets show that the proposed model achieves state-of-the-art F₁ scores (0.95 and 0.92, respectively), significantly outperforming strong baselines (p < 0.01, McNemar’s test). Evaluation of Token-level rationale using ERASER metrics demonstrates improved sufficiency and comprehensiveness compared to attention-based methods, LIME and SHAP. A detailed ablation study confirms that structural cues and sparse rationale extraction contribute substantially to predictive performance. Finally, a web-based visualization interface was developed to support human-in-the-loop verification of model output. This work provides a transparent and statistically validated alternative to black-box citation classifiers, advancing interpretable bibliometric analysis.

Multi-access edge computing (MEC) has emerged as a promising paradigm to reduce communication overhead and computational burden on mobile users (MUs) by offloading resource-intensive applications to nearby multi-access edge servers (MESs). This strategy significantly lowers energy consumption and execution latency. However, most existing studies on MEC assume an ideal scenario in which MESs are already optimally deployed, typically co-located with every base station (BS) in fifth-generation (5G) networks. While deploying an MES at each BS ensures low-latency services and high quality of service (QoS), it introduces several practical limitations. These include high capital and operational expenditures for service providers, underutilised MESs in sparsely populated regions, and elevated service costs for end users due to continuous MES activation regardless of demand. To address these challenges, this work proposes a novel Delay-based Categorisation and Adaptive MES-Activation (DCAMA) approach aimed at optimising MES activation in beyond-5G (B5G) wireless networks. Rather than assuming static and full MES deployment, the DCAMA strategy dynamically determines the minimum number of MESs required based on current network workload and task urgency. This energy-aware and cost-efficient method enables scalable MES activation without compromising service performance .

The rapid advancement of Generative Artificial Intelligence (AI) has enabled the creation of highly realistic synthetic audio, presenting significant challenges to digital security, media forensics, and public confidence. While deepfake detection has been extensively explored for high-resource languages like English, low-resource languages remain critically underexamined. This paper introduces a systematic benchmark of deepfake audio detection methods for Urdu, a language characterized by its rich morphology and phonetic complexity. We evaluate convolutional and transformer-based architectures, including LCNN, CNN–LSTM with Attention, and Whisper variants, employing Mel-Frequency Cepstral Coefficients and Linear Frequency Cepstral Coefficients as front-end features. To address data scarcity, we construct three dataset configurations, baseline, augmented, and extended derived from an existing Urdu deepfake corpus and enhanced through controlled augmentation and additional recordings. Our experiments, conducted with multiple random seeds and statistical validation, demonstrate that MFCC-based models, particularly Whisper-small, achieve strong performance with an Equal Error Rate as low as 0.50%. Robustness tests under noise, pitch, and tempo perturbations highlight the limitations of lightweight CNNs and underscore the advantages of transformer embeddings for handling Urdu’s linguistic variability. This study represents the first structured benchmark of deepfake audio detection techniques for Urdu, offering empirical insights into how language characteristics influence model performance and generalization. The findings emphasize the importance of multilingual evaluation in the development of trustworthy speech forensics systems.

Estimating the remaining useful life (RUL) of critical industrial assets is of crucial importance for optimizing maintenance strategies, enabling proactive planning of repair tasks, enhanced reliability, and reduced downtime in prognostic health management (PHM). Deep learning-based data-driven approaches have made RUL prediction a lot better, but traditional methods often do not look at the similarities and differences in the data, which lowers the accuracy of the estimates. Previous attempts to use Long Short-Term Memory (LSTM) networks for RUL prediction have failed because they depend on learned features for regression at the very end of the time step. The single objective function for estimation also constrains the learned representations, which has an impact on RUL estimation. The goal of this study is to find out how to predict the RUL of mechanical systems using complex sensor data. To do this, we present a data-driven framework called temporal convolution, along with a recurrent skip component and an attention mechanism network called TCRSCANet. It uses a combination of temporal convolution, recurrent skip parts, and an attention mechanism to make RUL estimation more accurate. The recurrent skip component finds long-term patterns in time series data, while temporal convolution pulls out high-level features from longer sequences. Finding hidden representations and degradation-development interactions between features at each window position in the input matrix is what the attention layer does to focus on the most important information for RUL estimation. The proposed methodology is tested and validated against the well-established C-MAPSS dataset, which focuses on aircraft degradation. The TCRSCANet model is better at predicting RUL as compared to other state-of-the-art methods because it uses the root mean square error (RMSE) and a scoring function to measure performance. The results of this study demonstrate the importance of the recurrent skip component and attention mechanisms for determining how long an industrial asset will be valuable.

The research proposes a detailed study aimed at forecasting stock and currency prices using state-of-the-art Machine Learning (ML) techniques, such as Transformers, LSTM, Simple RNN, N-HEATS, and N-BITS. The article provides an in-depth overview of t...

The proliferation of sensing technologies such as sensors has resulted in vast amounts of time-series data being produced by machines in industrial plants and factories. There is much information available that can be used to predict machine breakdown and degradation in a given factory. The downtime of industrial equipment accounts for heavy losses in revenue that can be reduced by making accurate failure predictions using the sensor data. Internet of Things (IoT) technologies have made it possible to collect sensor data in real time. We found that hybrid modelling can result in efficient predictions as they are capable of capturing the abstract features which facilitate better predictions. In addition, developing effective optimization strategy is difficult because of the complex nature of different sensor data in real time scenarios. This work proposes a method for multivariate time-series forecasting for predictive maintenance (PdM) based on a combination of convolutional neural networks and long short term memory with skip connection (CNN-LSTM). We experiment with CNN, LSTM, and CNN-LSTM forecasting models one by one for the prediction of machine failures. The data used in this experiment are from Microsoft’s case study. The dataset provides information about the failure history, maintenance history, error conditions, and machine features and telemetry, which consists of information such as voltage, pressure, vibration, and rotation sensor values recorded between 2015 and 2016. The proposed hybrid CNN-LSTM framework is a two-stage end-to-end model in which the LSTM is leveraged to analyze the relationships among different time-series data variables through its memory function, and 1-D CNNs are responsible for effective extraction of high-level features from the data. Our method learns the long-term patterns of the time series by extracting the short-term dependency patterns of different time-series variables. In our evaluation, CNN-LSTM provided the most reliable and highest prediction accuracy.

Navigation without GPS is a practical problem in indoor environment. The received signal strength indication known as RSSI can be used to compute the approximate distance and direction between an access point and a user mobile. Three or more multiple APs are normally cooperated to estimate the location of the mobile. However, the RSSI measurements are non-stationary and their mean and variance slowly changes over time and space. The RSSI propagation model derived from such measurements are not reliable. Hence, in this paper, an adaptive system using a neural, fuzzy inference system is introduced to effectively overcome the changes in indoor environment and improve the accuracy of mobile position estimation. The proposed system can reduce the erroneous mapping of the physical distances better than the conventional methods based on RSSI propagation modeling. The results of experiments show that the proposed system scales well in terms of the number of sensor nodes and provides an enhanced localization performance.

Image restoration and reconstruction from blurry and noisy images have proved to be challenging problem. Noise removal plays any important role in preserving the meaningful and useful information in images. Our paper is based on a denoising technique known as total variation (TV). Over the years, high quality images and videos have become a trend. However, noise has remained an integral part in images and videos. Many denoising techniques have been developed over the time to remove noise from images and videos. Linear filters, Non-Linear filters, median filters and their modifications, have been a great success in noise removal. However, they resulted in blurring of images. We came up with a directional total variation algorithm for denoising. Over the past, most of the denoising methods have been used with noisy images. In our study, we make use of sequential 1D total variation on the pixel sequence procured in various positions along with horizontal, vertical and zig-zag. The evaluation of our proposed approach is performed based on standard test images and the nature of the denoised images is measured by making use of objective matrices like visual signal to noise ratio (VSNR), and peak signal to noise ratio. Numerous experiments proved that our proposed method yields promising results.