بحوث سكوبس — علي سليم حليم جابر

Supply chain management depends on good demand estimates, even if conventional techniques battle complicated, nonlinear demand trends. This work aims to enhance the precision of demand estimates by developing a data-driven optimization approach that utilizes Long Short-Term Memory (LSTM) networks. The proposed method, which blends LSTM-based forecasting with historical sales data, employs a novel optimization technique to fine-tune the model parameters. Using RMSE and MAE, among other performance evaluation tools, the model was trained and evaluated on real supply chain datasets under time-series preprocessing. Experimental results indicate that the LSTM-based optimization algorithm outperforms conventional forecasting methods in lowering prediction errors and enhancing trend identification. The model may also withstand fluctuations in demand across various product categories and throughout the year. Ultimately, the proposed method offers a consistent and adaptable solution to the challenge of supply chain demand forecasting. © 2025 IEEE.

Effective and real-time facial expression detection is becoming increasingly crucial in affective computing, intelligent user interfaces, and mental health screening to comprehend human emotions. Lighting, face angle, and occlusion changes typically challenge standard emotion recognition algorithms in real-life, ever-changing scenarios. This paper introduces EmotionSenseNet, a novel Emotion Sensing Network with Spatiotemporal and Attention Fusion, a deep learning network that combines spatial, temporal, and attentional data, improving emotion identification and targeting these issues. The design uses a lightweight CNN based on MobileNetV2 to extract spatial information and an LSTM network to define facial motion temporal correlations. The approach is simplified with a focus module highlighting essential face traits and periods. Face landmark alignment preprocessing provides input normalization across face pictures. The suggested technique has minimal latency and 91.3% accuracy on benchmark datasets like FER-2013 and CK+, making it suited for real-time edge device applications. Despite practical constraints, EmotionSenseNet can identify emotions reliably and scalable. Therefore, future research will focus on establishing more extensive emotional classifications and energy-efficient embedded system deployment techniques. © 2025 IEEE.

The exponential growth of digital news material has called for the creation of smart algorithms able to offer brief summaries to enable the fast absorption of information. Presented in this study is DeepDigest, a creative text summary tool built on deep learning meant to automatically generate news digests. The proposed model for semantic understanding uses a hybrid architecture combining a Pointer-Generator Network with Bidirectional Encoder Representations from Transformers (BERT). This enables both extractive and abstractive summarizing to be done. A large-scale dataset of news articles published in several languages was gathered and processed to train and assess the algorithm's performance. DeepDigest scores very well on ROUGE-1, ROUGE-2, and ROUGE-L, confirming its value in preserving coherence and contextual significance. Experimental findings show that DeepDigest outperforms current baselines like TextRank and LSTM-based algorithms. Qualitative studies have also demonstrated that DeepDigest can provide more interesting and readable summaries. The survey results show that DeepDigest offers a means for the automatic summary of real-time news that is both efficient and scalable. This is important for content recommendation systems, digital journalism, and tailored news feeds. Future studies aiming at greatly enhancing the quality of summarization and its adaptability throughout a broad spectrum of material domains will mainly rely on the inclusion of reinforcement learning. © 2025 IEEE.

Though the growth of online forums has caused a revolution in digital communication, it has also led to more negative behavior exhibited by members. Because of the many characteristics of toxicity, which can show itself as insults, threats, vulgarity, or hate speech often happening simultaneously inside a single comment, it is a difficult process to find and categorize toxic comments. This complicates the process of spotting and classifying harmful comments. Being created to identify and classify various toxicity in online communication precisely, TOX-MULTINET is a ground-breaking multi-label classification tool. This paper presents TOX-MULTINET. A neural ensemble architecture is included in the framework proposed after creation. Comprising Convolutional Neural Networks (CNNs) for the extraction of contextual features, Bidirectional Long Short-Term Memory (BiLSTM) for sequential dependency learning, and an attention mechanism for enhancing interpretability, this architecture consists of When compared to other models thought to be at the cutting edge of technology, TOX-MULTINET outperforms them in precision, recall, and F1-score. All of this is achieved using training and assessment procedures utilizing benchmark datasets available to the general public. The findings of the studies show that the algorithm is strong enough to manage overlapping hazardous categories, therefore improving the accuracy of detection in situations that exist in the real world. The model offered offers a method for filtering user-generated content that is both scalable and adaptable. This approach is meant to help create online communities more supportive of good living. © 2025 IEEE.

Predicting the secondary structure of proteins continues to be a significant hurdle in the field of bioinformatics. This anticipation plays a crucial role as an intermediary stage in addressing the challenge of predicting the tertiary structure of proteins, which is instrumental in determining their functions. This prediction holds the potential to facilitate drug development and contribute to the identification of viral diseases. One can forecast the secondary structure of a protein by examining its primary components, including the amino acid sequence and various additional factors. Through the examination of established sequences and recognized protein types, it becomes feasible to anticipate unfamiliar sequences. The objective of this article is to enhance the forecast accuracy of protein secondary structure by adjusting the current code, aiming to reach an 80% accuracy rate. Copyright: ©2024 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license.