Wasit Journal of Computer and Mathematics Science

With the rapid spread of fake news on social media platforms, it has become essential to develop
effective detection mechanisms that overcome the limitations of content-only or propagation-only approaches. This
study aims to design a unified framework that jointly models textual semantics and diffusion characteristics for
improved misinformation detection. To achieve this, we propose GRAFT-FND, a Graph-Aware Recurrent Fusion
Deep Ensemble architecture that integrates contextual word embeddings (Word2Vec, BERT, and BERTweet) with
recurrent neural networks (RNN/GRU/LSTM/BiLSTM) and graph-based node embedding methods (Node2Vec and
DeepWalk) within a fusion-aware learning module. Extensive experiments conducted on the Twitter15 and Twitter16
benchmark datasets using 10-fold cross-validation demonstrate that the proposed framework consistently outperforms
baseline and recent state-of-the-art models, with the fusion mechanism and propagation-aware representations
contributing significantly to performance improvement. The results indicate that jointly modeling semantic and
structural information enhances the ability to capture complex misinformation patterns and improves generalization
across datasets. In conclusion, the proposed framework provides a robust and scalable solution for fake news detection
in social media environments. Future work is recommended to investigate adversarial robustness, real-time
deployment, and the integration of multimodal data sources.

Academic attrition in the Iraqi higher educational system is a significant problem with socio-economic
consequences. The purpose of this study is to address the problem through the development of a high-fidelity
predictive model using existing educational data sets. The approach is to develop a hybrid model based on the
Knowledge Discovery in Databases (KDD) process and Artificial Neural Networks (ANN), with a newly proposed
Weighted Euclidean Distance metric. The approach was validated using a dataset of 5,000 Iraqi university students
and texture analysis, which was inspired by image mining techniques. The analysis was performed to find the optimal
weight for socioeconomic and academic features. The KDD + ANN model demonstrated considerable ability to grasp
the intricacies and nonlinearity of behavioral patterns of students at the university level. Therefore, it achieved
remarkable accuracy of 94.21%, precision of 93.15%, and F1-score of 93.32%. Moreover, feature importance analysis
showed that Major Subject GPA had the highest weight of 0.48, followed by attendance in the first six weeks with a
weight of 0.32. Moreover, it is important to point out the accuracy with which the weighted metric has been identified
as an essential latent risk factor in the Iraqi scenario as well. Conclusion: The weighted proximity metric in neural
architecture is an effective approach in dealing with the complexities of educational data sets, thereby reducing the
chances of incorrect classification. Recommendations: This strategic early warning system needs to be incorporated
in the administration of universities in Iraq to bring about a shift in the approach from generalized to more targeted
advisories.

How to construct strong queuing schemes in a network logistics setting using mathematical
programming will be considered in this paper. As global supply chains get increasingly complex, Effective
management of queues at logistics hubs is crucial for business performance. This work provides a unifying structure
which combines queueing models and mathematical programming methods to minimize waiting time, maximize
capacity utilization, and optimizing the allocation of resources (e.g., servers) in an unknown demand environment.
We employ both analytical models (e.g. M/M/c queueing systems) and simulation techniques developed by means of
WinQSP software to explore performance measures under varying circumstances. For illustration, sortation systems
of a realistic distribution center network have been placed in simulated. The computational results also show that our
technology can reduce the average waiting time by 42% and increase the system\'s utilization rate by 28% compared
to traditional methods by applying mathematical programming for optimal server scaling. The findings of the study
will enable logistics managers to develop stable queuing systems that are responsive to demand fluctuation and
service performance. This study bridges the gap between queuing theory and practical logistics applications, providing
managers with actionable insights.

Emotion recognition in human-robot interaction (HRI) is important in order to develop socially aware
and responsive robotic systems. In this work, a near real-time emotion recognition architecture is introduced which
combines deep AI models such as CNN, and recurrent architectures (LSTM/ GRU) for audio and visual-based
emotion detection. The system is tested on the benchmark datasets like FER-2013 and RAVDESS. The goal is to
provide a strong and scalable approach that can serve as a step toward the full integration of emotionally intelligent
robots into everyday life, to encourage empathic, adaptive, and rewarding human-robot interaction. The proposed
deep AI model was tested on a dataset comprising 5,000 multimodal samples of human emotional expressions
collected in controlled and real-world human-robot interaction scenarios. The dataset included five emotional
categories: Happy, Sad, Angry, Neutral, and Surprise. Experimental results show the effectiveness of the proposed
system based on the public datasets, and its practical use in the simulated human-robot interaction (HRI) scenario.
Also, the proposed approach provides high accuracy and low inference delay, which can support robotic agents to
have effective emotion-adaptive behaviors in live-interaction environments.

This paper presents a hybrid optimization propose an algorithm that combines Augmented Lagrangian
(AL) framework and Limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) method to solve large-scale
constrained optimization problems effectively. In the AL formulation, quadratic penalty terms along with the linear
Lagrange multiplier terms are imposed on the constraints, and the solutions to the resulting unconstrained Lagrangian
for each given Lagrange multiplier determine a constrained problem sequence. The L-BFGS algorithm solves these
sub problems by approximating the search direction with a two-loop recursion that utilizes a small history of pairs of
gradient and displacement vectors. It is highly scalable to large-scale problems since this approach only demands a
small poly logarithmic number of vector operations per iteration as well as linear in memory storage, and completely
obviates explicit computation or storage for the Hessian matrix. This led to a new accelerated L-BFGS method that
speeds-up convergence in the inner minimization loop and a systematic outer loop for updating the Lagrange
multipliers and adjusting the penalty parameter based on the primal and dual residual norms. The proposed approach
integrates these components to produce a simultaneous iterative scheme that is both efficient and robust to maintaining
equality-inequality constraints simultaneously.