Journal of Soft Computing and Computational Intelligence (p-ISSN: 3107-4855, e-ISSN: 3048-6610)

AharixAI – AI Based Personalized Diet Planner

2026-02-28T11:02:55+00:00

The limitations of generic dietary advice add to an increasing number of people with diet-related health problems, as people struggle to translate generic dietary advice to their own individualized physiological needs. This paper presents AharixAI, a novel AI-based personalized diet planning system with the intention to address this gap. AharixAI consists of a multi-tier architecture and a Flutter mobile client and a Flask backend. Its hybrid AI methodology is used to achieve a mix of K-Means clustering algorithm for original dietary profile matching and Rule-Based Filtering Engine for detailed personalization based on user preferences and restrictions. The K-Means algorithm works by clustering a large food data set into different nutritional profiles, and the system correlates where someone's health data falls to the centroid of the appropriate cluster. One of the major innovations is incorporating the explainable AI Coach Gemma large language model to offer natural language justification for its recommendation methodology, to ensure user trust and adherence to the recommendations. Hypothetical performance evaluation under different user-profiles (e.g. Healthy Weight Gain, General Wellness) led to F1-Scores that give an indication of the system's potential to give balanced and effective recommendations. With the combined power of effective personalization and explainability, AharixAI can help provide a promising framework for users to proactively manage their dietary habits and enhance their overall well-being.

Multimodal Face Recognition Using Motion Sensors and Voice Input

2026-03-16T11:23:26+00:00

The increasing demand for secure and reliable authentication systems has led to the development of multimodal biometric technologies that combine multiple sources of identity verification. Traditional unimodal systems, such as standalone face recognition, are often vulnerable to spoofing attacks, environmental variations, and high false acceptance or rejection rates. To address these limitations, this study proposes a Multimodal Face Recognition System Using Motion Sensors and Voice Input that integrates physiological and behavioral biometrics into a unified authentication framework. The proposed system combines three complementary modalities, facial recognition, motion-based behavioral analysis, and speaker verification. Facial features are extracted using deep convolutional neural networks (CNNs) trained on large-scale datasets to ensure high recognition accuracy under varying illumination and pose conditions. Motion data captured through embedded inertial sensors, including accelerometers and gyroscopes, is used to analyze dynamic head movement patterns during authentication, providing an additional behavioral biometric layer. Voice input is processed using Mel-Frequency Cepstral Coefficients (MFCCs) and classified using machine learning algorithms such as Support Vector Machines (SVM) or deep neural networks for speaker identification.

To enhance system performance, feature-level and decision-level fusion strategies are implemented to combine multimodal data effectively. Experimental evaluation demonstrates that the proposed system achieves higher accuracy, improved robustness against spoofing attacks, and lower False Acceptance Rate (FAR) and False Rejection Rate (FRR) compared to unimodal biometric systems. Furthermore, the integration of motion sensor data strengthens liveness detection, reducing vulnerability to photo, video, or replay-based attacks. The results indicate that multimodal biometric fusion significantly enhances security, reliability, and user trust. The proposed framework is particularly suitable for deployment in smartphones, IoT-enabled devices, banking systems, and high-security access control applications. This research contributes to the advancement of intelligent, sensor-assisted authentication systems for next-generation secure environments.

A Data-Driven Automata-Theoretic Framework for Emotional State Recognition Through Music Interaction Dynamics

2026-04-06T13:15:18+00:00

Music is an effective channel of expression of emotions, and the patterns of interaction between users and the music platforms have provided a richer source of behavioural indicators of emotional state. The present paper presents a data-driven Automata-Theoretic Framework of emotional state recognition via music interaction dynamics, a machine learning-based approach to formal automata theory to obtain structurally, interpretably, and adaptively modelled emotions. The framework relies on the data of user-music interaction, including the duration of listening, skip patterns, number of replays, preferred tempo, and genre changeover to derive behavioural characteristics which are of significance. These features are learned with data-driven learning models to give probabilistic emotion indicators that are then mapped to discrete emotional states with the help of a finite automaton. The automata-theoretic layer formalises emotion changes, which guarantees temporal consistency and logical correctness in emotion changes, and overcomes the weakness of the purely statistical approach that is frequently difficult to interpret. The given framework defines emotions as states, and interaction-driven cues as transition symbols; therefore, by doing so, it will be able to capture the levels of short-term affective reactions, as well as long-term emotional tendencies. Through experimental analysis, the hybrid approach proves to be more robust and less noisy (predictive) and more explainable than standalone machine learning models. The suggested framework can be used in music recommendation systems, affect-sensitive human-computer interaction, and mental health monitoring and adaptive entertainment systems. In general, this paper demonstrates the opportunities of the data-based intelligence and automata theory integration to provide a stable and interpretable emotional state recognition.

The Role of Regression Analysis in Data-Driven Research: Foundation and Real-Life Applications

2026-04-13T04:11:33+00:00

Regression analysis is one of the most fundamental and widely applied statistical techniques in engineering and data science. This literature review synthesizes findings from ten research papers spanning theoretical foundations, methodological developments, and practical applications across diverse domains, including healthcare, education, real estate, transportation safety, and environmental science. The review demonstrates that regression methods—ranging from simple linear regression to advanced logistic and multivariate techniques—serve as essential tools for prediction, optimisation, and decision-making in engineering practice. Key findings reveal that while linear regression remains highly effective for continuous outcome prediction, logistic regression extends these capabilities to classification problems, and advanced techniques like regularisation and ensemble methods address challenges such as overfitting and multicollinearity. This review is structured to provide first-year engineering students with a clear understanding of regression fundamentals, awareness of common pitfalls, and appreciation for the breadth of real-world applications. The evidence base demonstrates that mastery of regression analysis is critical for modern engineering practice, enabling data-driven solutions to complex problems across multiple disciplines.

ABC-SVM: A Novel Artificial Bee Colony Optimization Framework for Optimal Feature Selection in Breast Cancer Diagnosis

2026-04-13T06:53:29+00:00

Background: Feature selection constitutes a critical pre-processing step in medical diagnosis because high-dimensional datasets frequently contain irrelevant or redundant features that degrade classifier performance and increase computational complexity. The artificial bee colony algorithm offers a powerful metaheuristic approach for solving complex feature selection problems, while support vector machines provide robust classification with strong theoretical foundations.

Objective: This paper proposes ABC-SVM, a novel hybrid framework that integrates artificial bee colony optimization with support vector machines for optimal feature selection in breast cancer diagnosis. The framework simultaneously optimizes feature subsets to maximize classification accuracy while minimizing the number of selected features.

Methods: The proposed ABC-SVM framework employs binary-encoded food sources that represent feature subsets. A multi-objective fitness function balances the accuracy obtained from five-fold cross-validated SVM against feature parsimony. The ABC algorithm iterates through employed bee, onlooker bee, and scout bee phases to evolve optimal feature subsets. The framework was evaluated on four benchmark medical datasets, including Wisconsin Breast Cancer, PIMA Indian Diabetes, Hepatitis, and Mammographic Mass. The performance was compared against genetic algorithm and particle swarm optimization using ten-fold cross-validation with five repeats.

Results: ABC-SVM achieved 98.71% accuracy on the Wisconsin dataset with 66.7% feature reduction, selecting 3.0 features from the original 9 features, thereby outperforming GA-SVM, which achieved 97.94% accuracy with 55.6% reduction and PSO-SVM, which achieved 98.21% accuracy with 61.1% reduction. On the PIMA dataset, ABC-SVM achieved 86.78% accuracy with 60.0% feature reduction, compared to GA-SVM at 84.56% and PSO-SVM at 85.12%. On the Hepatitis dataset, ABC-SVM achieved 87.93% accuracy with 62.6% reduction. The algorithm converged within 50 to 80 iterations, demonstrating an efficient exploration-exploitation balance. The selected feature subsets aligned with clinical knowledge, including bare nuclei, clump thickness, and uniform cell size for breast cancer, glucose and BMI for diabetes, and liver function tests for hepatitis.

Conclusion: ABC-SVM provides an effective framework for optimal feature selection in medical diagnosis, achieving superior feature reduction and improved classification accuracy compared to standard SVM and competing metaheuristics. The multi-objective fitness function successfully balances accuracy and parsimony, producing clinically interpretable feature subsets. The framework’s consistent performance across diverse medical datasets demonstrates its broad applicability for developing parsimonious, accurate, and interpretable clinical decision support systems.