FORECASTING FUTURE TRENDS: A GENERATIVE AI APPROACH TO DYNAMIC TREND PREDICTION

Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4886 FORECASTING FUTURE TRENDS: A GENERATIVE AI APPROACH TO DYNAMIC TREND PREDICTION VINODKUMAR REDDY SURASANI1, *SARVANI ANANDARAO2, NAGARAJU DEVARAKONDA3 1Sr Software Engineer, RBC Wealth Management, Minneapolis, MN 55401, USA 2School of Computer Science and Engineering, SRM University-AP, Amaravathi, AP, India 3School of Computer Science and Engineering, VIT-AP University, Amaravathi, AP, India [email protected], [email protected], [email protected], Corresponding Author: Sarvani Anandarao ABSTRACT In the rapidly evolving digital landscape, trend forecasting has become a critical task for decision-makers across industries. Traditional methods struggle with adaptability, scalability, and real-time trend identification. This paper presents a novel framework that integrates Generative AI with the Proposed Guided Remora Optimization Algorithm (PGROA) to enhance trend prediction accuracy while maintaining robustness across dynamic and multimodal datasets. The framework leverages transformer-based architectures for feature extraction, adaptive learning mechanisms for real-time updates, and cross-domain generalization techniques to ensure scalability. Additionally, interpretability methods such as SHAP values and attention mechanisms provide transparency in model predictions. The proposed system is evaluated on diverse datasets, demonstrating superior performance with an accuracy of 94.8%, an F1-score of 93.8%, and a significantly reduced RMSE of 0.072, outperforming existing deep learning and hybrid models. This research establishes a scalable and interpretable AI-driven approach to trend prediction, equipping decisionmakers with actionable insights for dynamic environments. Keywords: Generative AI, Trend Prediction, Adaptive Learning, Remora Optimization, Cross-Domain Generalization. 1. INTRODUCTION As the world becomes increasingly connected, the ability to track emerging trends has never been more critical. Traditional methods of trend detection, such as market research or expert analysis, often struggle to keep pace with the rapid evolution of technologies, social behaviors, and global events. However, with the exponential growth of digital data, new tools have emerged to analyze and predict these trends with greater precision. Generative AI, a branch of artificial intelligence that focuses on creating models capable of generating data based on patterns learned from existing information, is being explored as a powerful tool for trend forecasting. By analyzing vast amounts of historical data, generative AI can simulate future scenarios, identifying emerging topics before they gain widespread attention. This predictive capability has the potential to revolutionize various industries, enabling businesses, governments, and individuals to make more informed decisions about future developments. Generative AI enables the precise identification of trends by uncovering intricate patterns in large datasets, distinguishing significant developments from noise. It employs advanced natural language processing (NLP) and machine learning techniques to analyze structured and unstructured data sources, such as news articles, social media posts, and research papers. One critical aspect of effective trend tracking is feature extraction—identifying the most relevant elements that signal emerging topics. Traditional methods often relied on predefined statistical measures, whereas generative AI autonomously learns complex relationships within the data. Hybrid approaches, which combine generative AI with classical machine learning or deep learning techniques, have been shown to enhance trend prediction. For example, combining Generative Pretrained Transformers (GPT) with clustering algorithms such as k-means can group related topics, while integrating generative models with decision trees enables more interpretable insights. Similarly, frameworks like GPT+LSTM leverage generative Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4887 AI for language understanding and LSTM for capturing sequential dependencies in time-series data. These hybrid strategies improve accuracy by leveraging the strengths of each method, helping to identify subtle shifts in topics over time and distinguishing genuine trends from transient phenomena. By utilizing these advanced techniques, generative AI provides a more reliable foundation for trend prediction across diverse domains. To classify trends effectively and identify significant patterns, generative AI models often integrate with advanced techniques for feature extraction and categorization. For instance, unsupervised methods like clustering algorithms or supervised approaches like decision trees can be used to group and interpret patterns from raw data. In 2020, researchers demonstrated how multimodal frameworks, such as combining GPT models with visual analysis tools, can enhance the accuracy of predictions by utilizing diverse data sources like textual and visual content. However, relying solely on classical models or standalone generative techniques may fail to capture the deeper semantic relationships required for precise trend forecasting. Generative AI excels in extracting these semantic links by leveraging contextual comprehension and latent representation learning. Nevertheless, challenges such as overfitting, a common issue in classical machine learning methods, must be addressed when employing generative AI models. Overfitting occurs when a model learns the noise or specific patterns of training data rather than general trends, leading to poor performance on unseen data. To mitigate this, hybrid models that combine generative AI with regularization techniques or ensemble approaches have proven effective in ensuring accurate and robust trend prediction across dynamic datasets. A model that suffers from overfitting performs exceptionally well on training data but struggles to generalize to unseen data, leading to poor performance on fresh datasets. In the context of trend prediction, this challenge can hinder the ability of generative AI models to accurately forecast future developments. To address overfitting, the integration of neural networks with other advanced techniques has become a prevalent approach. Neural networks, known for their ability to interpret complex patterns, are enhanced through such combinations, ensuring better generalization and robustness. In the context of generative AI, various hybrid techniques have been employed to enhance trend prediction and mitigate challenges like overfitting. For instance, generative AI models like GPT have been integrated with convolutional neural networks (CNNs) for feature extraction and pattern recognition. Additionally, frameworks combining generative models with classical machine learning techniques, such as GPT+SVM or GPT+Random Forest, have been utilized to leverage the strengths of both paradigms. These approaches enable the extraction of semantic relationships and intricate patterns, ensuring robust and accurate predictions of trends across diverse datasets. Despite the significant advancements in leveraging generative AI for trend tracking, several limitations remain in existing methodologies. Current models often face challenges in handling highly dynamic and noisy datasets, which can result in inaccurate predictions. Additionally, many approaches struggle to generalize across diverse domains, as they rely heavily on domain-specific fine-tuning. The interpretability of generative AI models is another critical issue, making it difficult to understand the reasoning behind the predicted trends. Furthermore, computational complexity and the need for extensive labelled datasets can limit their scalability and applicability in real-time scenarios. This paper aims to address these limitations by proposing a robust and scalable framework that integrates generative AI with adaptive learning mechanisms to improve trend prediction accuracy across diverse datasets. By incorporating techniques for dynamic data handling, cross-domain generalization, and enhanced model interpretability, the proposed approach seeks to overcome the existing challenges and establish a comprehensive solution for tracking trends with generative AI. The need for this study arises from the increasing demand for real-time, robust, and interpretable trend prediction across diverse and dynamic domains. Traditional approaches fail to offer adaptability and explainability in multimodal environments. This study addresses the gap by proposing a generative AI-driven framework that integrates adaptive learning and optimization for better generalization. Literature was screened based on relevance to generative AI, multimodal trend prediction, hybrid deep learning architectures, and optimization techniques published between 2015 and 2024. Key selection criteria included methodological novelty, reported performance metrics, and domain applicability. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4888 1.1 Motivation The ability to predict future trends plays a pivotal role in shaping strategic decision-making across various fields, such as healthcare, technology, finance, and social sciences. Despite significant advancements, existing methodologies often fall short in addressing the complexities associated with dynamic, real-world datasets. For example, in the healthcare sector, predicting the progression of diseases like Alzheimer's Disease (AD) from medical imaging data is vital for early diagnosis and timely interventions. However, challenges such as data noise, domain dependency, and the need for explainability hinder the effectiveness of current AI models. Similarly, in the technology sector, identifying emerging trends based on noisy and rapidly evolving data sources like news articles, social media, or research publications is an equally challenging task. Motivated by the success of generative AI in capturing intricate patterns, this study aims to develop a scalable and interpretable framework capable of addressing these challenges. The proposed framework aspires to bridge the gap between theoretical advancements and practical implementations, ensuring that trend prediction models remain reliable, adaptive, and actionable. The paper is structured as follows: Section 2 provides a comprehensive review of related works in trend prediction and generative AI methodologies, highlighting advancements and identifying existing gaps that motivate this study. Section 3 outlines the proposed framework, focusing on the integration of generative AI with adaptive learning mechanisms to enhance trend prediction. This section also details preprocessing techniques for managing dynamic datasets and presents strategies for achieving crossdomain generalization and model interpretability. Section 4 discusses the experimental results, offering a comparative analysis of the proposed approach against existing models using a variety of performance metrics. Finally, Section 5 concludes the paper with key insights, explores potential realworld applications, and suggests directions for future research to further advance trend prediction methodologies. Unlike earlier works that focused on statistical modeling or single-modality learning, this research leverages a multi-modal, generative AI-based framework enhanced with optimization and interpretability mechanisms. While prior efforts like GAN-based synthesis or Transformer models have shown promise, they often lack adaptability, transparency, and cross-domain performance. Our work differs by introducing a novel combination of PGRO optimization, interpretability through SHAP and attention, and domain-adversarial training to address the evolving nature of trend prediction tasks. 1.2 Research Contribution  Proposed an innovative Generative AIdriven framework that integrates the Proposed Guided Remora Optimization Algorithm (PGROA) for enhancing trend forecasting accuracy in dynamic environments.  Developed PGROA, an improved version of the Remora Optimization Algorithm, which enhances exploration-exploitation balance, ensuring faster convergence and better performance in optimization tasks.  Utilized transformer-based architectures (BERT, Vision Transformer (ViT), Temporal Convolutional Transformer (TCT)) for extracting features from textual, visual, and temporal data sources, improving trend prediction accuracy.  Implemented incremental learning with elastic weight consolidation (EWC) and Replay Buffer Systems to prevent catastrophic forgetting and maintain accuracy in real-time trend updates.  Applied Domain-Adversarial Neural Networks (DANN) to enable the model to generalize across multiple domains, ensuring robustness to domain shifts and varying data distributions. In addition to proposing a new trend prediction framework, this study introduces original knowledge in the integration of PGRO for adaptive optimization, interpretable multi-modal feature fusion, and domain-adversarial learning. These elements address significant limitations in existing works, particularly in adaptability and transparency. This contributes not only technical advancement but also provides a scalable blueprint for real-world deployment of generative AI systems in volatile trend environments. 1.3 Problem Statement and Research Questions Despite advances in trend prediction, current methods fall short in adaptability, scalability, and interpretability when applied to dynamic, multimodal data. Generative AI models often struggle with generalization and lack integration Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4889 with real-time learning mechanisms. This study aims to address the following research questions:  How can generative AI frameworks be designed to adapt dynamically to new data across domains?  What optimization strategies enable realtime learning without catastrophic forgetting?  How can interpretability techniques be effectively integrated with deep generative models for trend prediction? 2. RELATED WORK Early Foundations (2015-2018) Zhang and colleagues (2015) pioneered the integration of deep learning with time series prediction, proposing a basic LSTM architecture for financial trend forecasting. While their work demonstrated superior accuracy compared to traditional statistical methods, the model struggled with scalability across different data domains and required extensive domain-specific feature engineering. Building upon these limitations, Chen et al. (2016) introduced an attention-based mechanism to improve model adaptability across different time series data. Their framework showed a 15% improvement in prediction accuracy but was computationally intensive and required large training datasets, making it impractical for real-time applications. Wang and Liu (2017) addressed the computational constraints by developing a lightweight neural architecture for trend prediction. While achieving faster training times, their approach sacrificed accuracy for efficiency and showed inconsistent performance on non-stationary data. Integration of Generative Models (2019-2021) A significant paradigm shift occurred when Kumar and Martinez (2019) introduced the first GAN-based framework for trend prediction. Their approach generated synthetic training data to improve model robustness, achieving a 20% improvement in accuracy for limited dataset scenarios. However, the framework suffered from mode collapse and training instability issues. Lee et al. (2020) tackled the stability problems by developing a modified GAN architecture with regularization techniques. While successfully addressing training stability, their solution introduced additional hyperparameters requiring manual tuning, limiting its practical applicability. Rodriguez and Kim (2020) proposed a hybrid approach combining transformers with GANs for improved feature extraction. Their work showed promising results on financial datasets but failed to generalize well across different domains. Park and Thompson (2021) introduced an adaptive learning rate mechanism to improve training convergence. While their approach showed better stability, it still required significant computational resources and expert knowledge for optimal performance. Advanced Architectures and Scalability (20222024) Wilson et al. (2022) developed a scalable framework using distributed computing and federated learning. Their approach improved computational efficiency but struggled with maintaining prediction accuracy across heterogeneous data sources. Chang and Patel (2023) introduced a novel selfattention mechanism specifically designed for trend prediction. Their work showed exceptional performance on structured data but had limitations handling multimodal inputs and real-time updates. Recent work by Singh and colleagues (2024) focused on developing lightweight, efficient architectures suitable for edge deployment. While achieving impressive efficiency gains, their approach showed reduced accuracy compared to larger models. Emerging Methodologies (2022-2024) Wilson et al. (2022) developed a scalable framework using distributed computing and federated learning. Their approach improved computational efficiency but struggled with maintaining prediction accuracy across heterogeneous data sources. Yamamoto and Chen (2022) proposed a novel quantum-inspired generative model for trend prediction. While showing promising results for complex pattern recognition, their approach required specialized hardware and faced scaling limitations. Anderson et al. (2022) introduced a multi-task learning framework combining trend prediction with anomaly detection. Though innovative, their Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4890 solution showed decreased performance when handling imbalanced datasets. Brown and colleagues (2023) developed a selfsupervised learning approach for trend prediction in unlabeled datasets. While reducing the need for labeled data, their method struggled with real-time adaptation. Chang and Patel (2023) introduced a novel self-attention mechanism specifically designed for trend prediction. Their work showed exceptional performance on structured data but had limitations handling multimodal inputs. Li and Thompson (2023) proposed a robust ensemble method combining multiple generative models. Though achieving high accuracy, their approach faced significant computational overhead and complex deployment requirements. Hassan et al. (2023) developed a privacy-preserving generative framework for sensitive trend data. While ensuring data privacy, their method showed reduced prediction accuracy compared to non-private alternatives. Kim and Park (2023) introduced an interpretable deep learning framework for trend analysis. Their approach provided clear explanations but sacrificed some accuracy for interpretability. Zhang et al. (2023) proposed a hybrid approach combining statistical methods with deep learning. While showing good performance on traditional datasets, their method struggled with highly nonlinear patterns. Patel and Rodriguez (2024) developed an automated architecture search method for optimal model selection. Though innovative, their approach required extensive computational resources during the search phase. Singh and colleagues (2024) focused on developing lightweight, efficient architectures suitable for edge deployment. While achieving impressive efficiency gains, their approach showed reduced accuracy compared to larger models. Fischer and Lee (2024) introduced a novel cross-domain adaptation technique for trend prediction. Their method showed promise in transfer learning but required significant fine-tuning for each new domain. Our review highlights a lack of comprehensive frameworks that effectively combine dynamic learning, cross-domain adaptability, and model interpretability. Most existing methods are either domain-specific or computationally intensive, lacking real-time adaptability. This study fills these gaps by proposing a generative AI framework supported by PGRO optimization and interpretability tools like SHAP and attention visualization, which are currently underexplored in this context. Tabular Summary of Key Works Year Techniqu e Key Contribu tion Primary Limitation Chandra et al. (2021) Deep learning models (LSTM, GRU) for multi-step time series prediction Evaluated multiple architectu res for long-term forecastin g Limited interpretabi lity and requires extensive hyperpara meter tuning Hollis et al. (2018) Comparis on of LSTM and attention mechanis ms for financial forecastin g Demonstr ated the effectiven ess of attention for capturing temporal dependen cies Lacks exploration of multimodal data Li et al. (2020) Neural architectu re search (AutoST) for spatiotemporal prediction Automate d model selection for timeseries forecastin g High computatio nal cost for model selection Liu et al. (2023) GANbased classificat ion for financial time series volatility prediction Improved trend detection using generative models Stability issues in GAN training Smith & Smith (2020) Condition al GAN for timeseries generatio n Enhanced realism in synthetic data for forecastin g Requires fine-tuning for domain adaptation Shu et al. (2024) Transfor mer-GAN hybrid for precipitati on prediction Improved spatiotemporal accuracy with Computati onal complexity remains a challenge Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4891 generative models Zhao et al. (2023) Bidirectio nal Transfor mer GAN for human motion prediction Strong performan ce in longterm sequence modelling High training cost and resource demands Liu et al. (2024) Trend detectionbased autoscaling for highconcurren cy systems Improved scalability and resource managem ent Limited application to financial and social trends Alcazar et al. (2024) Classical and quantum generativ e models for optimizati on Hybrid approach improves combinat orial optimizati on Quantum methods are still experiment al Zhao et al. (2015) Multitask learning for spatiotemporal event forecastin g Improved predictive power across tasks Lacks interpretabi lity mechanism s Zhang et al. (2024) Selfsupervise d learning for timeseries analysis Categoriz ed selfsupervise d technique s for forecastin g Requires large labeled datasets for downstrea m tasks Fahim et al. (2021) Hybrid LSTM with selfattention for scientific research forecastin g Improved accuracy with attentionenhanced recurrent models High training complexity Yang et al. (2022) Generativ e ensemble regression Physicsinformed deep generative Limited adaptabilit y to nonfor particle dynamics models for timeseries prediction physical domains Chen et al. (2024) Generativ e ML methods for ensemble postproce ssing Enhanced calibratio n of ensemble prediction s Needs extensive historical data for effective training Yao et al. (2024) Interpreta ble trend analysis neural networks Trend analysis with explainabl e deep learning Still lacks domainspecific customizati on Wikle & Zammit - Mangio n (2022) Statistical deep learning for spatial and temporal data Bridging statistical methods with deep learning Requires strong statistical expertise for effective use Alsharef et al. (2022) Review of ML and AutoML for timeseries forecastin g Provided a comprehe nsive survey of forecastin g models Lacks realworld experiment al validation George et al. (2023) Survey on Edge Computin g and its future in cloud computin g Explored computati onal offloading for timeseries tasks Edge-based models still require optimizatio n Jia et al. (2025) Contrasti ve representa tion domain adaptatio n for industrial time series Improved crossdomain generaliza tion for industrial forecastin g Requires large-scale industrial datasets for training Yazdani et al. (2023) Robust optimizati on for timedependent systems Evaluated timeaware optimizati on strategies Optimizati on techniques may not generalize Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4892 across domains Nott et al. (2023) Bayesian inference for generativ e models Improved uncertaint y quantifica tion in trend prediction High computatio nal cost for inference Jin et al. (2024) Multilayer temporal graph neural networks for social media trends Captured complex temporal relationsh ips in social media Requires fine-tuning for different datasets Fu et al. (2022) Reinforce ment learningbased model combinati on for timeseries forecastin g Adaptive model selection for varying trend dynamics Computati onally expensive training process Ji et al. (2017) Trend-intrend research design for causal inference Proposed a framewor k for causal analysis in timeseries data Requires expert domain knowledge Iwata & Kumaga i (2020) Few-shot learning for timeseries forecastin g Improved generaliza tion with limited training data Struggles with noisy or unstructure d data Gu et al. (2021) Transfer learning, active learning, and metric learning for timeseries Integrated multiple learning paradigms for prediction Requires domain adaptation for different application s Pöppelb aum et Contrasti ve learning - Improved feature representa Still lacks interpretabi lity for al. (2022) based selfsupervise d timeseries analysis tion for forecastin g decisionmaking Zhao et al. (2023) Metalearning for increment al stock trend forecastin g Adaptive learning approach for financial markets High sensitivity to initial training conditions Shen et al. (2024) Uncertain ty quantifica tion for oil well trend prediction Addresse d uncertaint y in energy sector forecastin g Requires extensive field data for effective modeling He et al. (2022) Machine learning for crude oil price trend prediction Integrated multimod al features for improved accuracy Requires additional real-time market inputs Research Gaps Despite significant advancements in time-series forecasting, trend prediction, and interpretability, existing methodologies still face critical limitations that hinder their effectiveness in dynamic and multimodal environments. Traditional statistical models, such as ARIMA and Random Forest, struggle with capturing non-stationary trends and rapidly evolving data distributions, making them unsuitable for real-time forecasting. While deep learning approaches, including LSTMs and GRUs, offer improved sequence modeling capabilities, they suffer from catastrophic forgetting and require frequent retraining to adapt to new patterns. Furthermore, cross-domain generalization remains a challenge, as many models are highly domainspecific, necessitating extensive fine-tuning to perform across diverse datasets. Current domain adaptation methods, such as contrastive learning and few-shot learning, show promise but often fail to handle high-dimensional and multimodal data effectively. Another significant gap in existing research is the lack of interpretability in deep learning-based trend Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4893 forecasting. Many high-performing models, including transformer-based architectures and generative adversarial networks (GANs), excel in predictive accuracy but offer little transparency in their decision-making processes. Although techniques such as SHAP values, attention visualization, and saliency maps have been introduced for explainability, their integration with trend prediction remains underexplored. Moreover, computational complexity poses a major limitation, as hybrid models, such as CNN-LSTM and Transformer-GAN, demand extensive resources, making them impractical for real-time applications. While optimization algorithms like PSO and ACO enhance model tuning, they introduce high computational latency, limiting their scalability. To address these challenges, our research proposes a Generative AI-driven trend forecasting framework that integrates the Proposed Guided Remora Optimization Algorithm (PGROA) for efficient hyperparameter tuning, transformer-based architectures for robust cross-domain generalization, and advanced interpretability techniques such as SHAP analysis, attention visualization, and counterfactual explanations. Additionally, our framework employs incremental learning mechanisms to ensure adaptability without frequent retraining while maintaining computational efficiency. By balancing accuracy, scalability, and interpretability, our proposed approach offers a novel, high-performing, and transparent solution for dynamic trend forecasting across multiple domains. 3. REMORA OPTIMIZATION ALGORITHM AND PROPOSED GUIDED REMORA OPTIMIZATION ALGORITHM This section provides an overview of the Remora Optimization Algorithm, a bio-inspired optimization technique that emulates the mutualistic relationship between remoras and their host organisms. It then introduces the Proposed Guided Remora Optimization Algorithm, which incorporates tailored guidance strategies to refine the search process, improve solution quality, and address the limitations of the original algorithm in complex problem domains. 3.1 Remora Optimization Algorithm The Remora Optimization Algorithm is an innovative nature-inspired optimization technique that draws inspiration from the symbiotic relationship between remora fish and their host species, such as sharks or whales. Just as remoras benefit from the protection and food sources provided by larger marine animals, this algorithm seeks to enhance optimization processes by utilizing a cooperative search strategy. The Remora Optimization Algorithm operates by simulating a population of agents, or "remoras," that explore the solution space in search of optimal solutions to complex problems. Each remora evaluates its position based on a fitness function, which quantifies the quality of its solution relative to others. By continuously updating their positions and exchanging information about the best-found solutions, the remoras collaboratively navigate the search landscape, effectively balancing exploration and exploitation. Below provide the mathematical models for "Free travel" and "Eat thoughtfully." 3.3.1 Initialization Let 𝑅, represents the remora position which is given by the Eq.1 𝑅=(𝑅,𝑅,………𝑅) (1) Where 𝑅 represent current position of remora, 𝑑 denotes the dimension and 𝑖 the number of the remora 𝑅, represents the optimal solution which is given by the Eq.2 𝑅 =(𝑅,𝑅,………𝑅) (2) Evaluation of candidate solutions using fitness function is given by the Eq.3 𝑓(𝑅)=𝑓[(𝑅,𝑅,………𝑅)] (3) 3.1.2 Free Travel (Exploration) Eq.4 defines the shift in remora's location when swordfish serve as hosts. 𝑅 =𝑅 −(𝑟𝑎𝑛𝑑(0,1)∗   −𝑅  (4) as 't' stands for the number of the iteration and 𝑅 for the random location. A few tiny steps around the host is represented with 𝑅 and its calculation is shown in the Eq.5. This is helpful in determining whether to switch hosts and change in the position. 𝑅 =𝑅 +𝑅 +𝑅∗𝑟𝑎𝑛𝑑𝑛 (5) where Rpre stands for the previous position, "𝑟𝑎𝑛𝑑𝑛 " indicates random integer, 𝑅 is the current position. Let 𝑓(𝑅 ) denotes fitness value of the current position of remora and 𝑓(𝑅) denotes the fitness value of the remora’s tentative step. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4894 𝑖𝑓 𝑓(𝑅 )>𝑓(𝑅) remora continues feeding on the same host else remora shifts to another host which can be whale or swordfish. The host is chosen by the Eq.6 𝐻(𝑖)=𝑟𝑜𝑢𝑛𝑑(𝑟𝑎𝑛𝑑) (6) The remora shifts to the whale when H(i) equals 1, and shifts to the swordfish when H(i) equals 0. 3.1.3 Eat Thoughtfully (Exploitation) Eq.7, Eq.8, Eq.9, Eq.10 describes how the position of the remora changes when whale is the host. 𝑅 =𝐷∗ 𝑒∗cos(2𝜋𝛼)+𝑅 (7) 𝛼 =𝑟𝑎𝑛𝑑(0,1)∗(𝑎−1)+1 (8) 𝑎 = −(1+  ) (9) 𝐷 = |𝑅 −𝑅| (10) where "𝑎" decreases linearly between [-2,-1], "𝐷" is the current optimal solution distance from remora present on whale to prey, and α is a random value between [-1, 1]. In host feeding around the body of whale. Eq.11, Eq.12, Eq.13, Eq.14 describes the host feeding around the body of whale. 𝑅 =𝑅 +𝐴 (11) 𝐴 =𝐵∗(𝑅 −𝐶∗𝑅) (12) 𝐵 =2∗𝑉∗𝑟𝑎𝑛𝑑(0,1)−𝑉 (13) 𝑉 =2∗(1−  _) (14) where "C" denotes a parameter that is utilized to regulate the location of the remora its ranging from 0 to 1. "A" symbolizes the movement of the remora over the host, or tiny steps. The volume of the host is represented by B. The remora's volume is represented using V. 3.2 Proposed Guided Remora Optimization Algorithm The next position in the classic Remora Optimization Algorithm is selected at random. This may lead the algorithm move away from the best location and possibly toward a worse one as a result of this randomness, which could result in suboptimal convergence or a local minimum being reached. To address this issue we proposed a Guided Remora Optimization Algorithm (PGROA) that updates the subsequent location depending on the previous best position. The PGROA computes the ratio of the best-known position (𝑅) to that of the next location (𝑅). This ratio indicates that the new position (𝑅) is allowed if ratio falls between 0.7 and 1. The optimal location is then updated by computing the fitness value of this new position. Iteratively, this procedure continues until the optimal position does not improve. 𝑅=𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑜𝑓 𝑡ℎ𝑟 𝑟𝑒𝑚𝑜𝑟𝑎 𝑎𝑡 𝑡𝑖𝑚𝑒 𝑡 𝑅 = 𝐵𝑒𝑠𝑡 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑓𝑜𝑢𝑛𝑑 𝑏𝑦 𝑡ℎ𝑒 𝑟𝑒𝑚𝑜𝑟𝑎 𝑠𝑜 𝑓𝑎𝑟 𝑤ℎ𝑖𝑐ℎ 𝑖𝑠 𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑 𝑏𝑦 𝑓𝑖𝑡𝑛𝑒𝑠𝑠 𝑅 =𝑁𝑒𝑤 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑓𝑜𝑟 𝑡ℎ𝑒 𝑟𝑒𝑚𝑜𝑟𝑎 𝐹𝑖𝑡𝑛𝑒𝑠𝑠(𝑅)=𝐹𝑖𝑡𝑛𝑒𝑠𝑠 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑅 𝛼 =𝑅𝑎𝑡𝑖𝑜𝑛 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑,𝑠𝑒𝑡 𝑡𝑜 0.7 Initialize 𝑅=𝐼𝑛𝑡𝑖𝑎𝑙 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑅 = 𝑅 Iterate Generate New Position 𝑅 =𝑅𝑎𝑛𝑑𝑜𝑚 𝑁𝑒𝑤 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 Calculate Ratio 𝑅𝑎𝑡𝑖𝑜 = 𝑅 𝑅 Check Ratio and Update Position: If 0.7≤ratio<1 then 𝑅 = 𝑅 [The next position is 𝑅] Else 𝑅 =𝐴𝑛𝑜𝑡ℎ𝑒𝑟 𝑅𝑎𝑛𝑑𝑜𝑚 𝑁𝑒𝑤 𝑃𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑎𝑔𝑎𝑖𝑛 𝑐ℎ𝑒𝑐𝑘 𝑡ℎ𝑒 𝑟𝑎𝑡𝑖𝑜𝑛 𝑎𝑛𝑑 𝑢𝑝𝑑𝑎𝑡𝑒 3. Fitness Calculation and Update Best Position 𝑖𝑓 𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑅)>𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑅) 𝑡ℎ𝑒𝑛 𝑅 = 𝑅 The above procedure continues, until there is no change in the best position. 4. PROPOSED METHODOLOGY 4.1 Data Collection and Pre-Processing In this work, data is collected from various sources such as news websites (Google News, BBC), social media platforms (Twitter, Reddit), e-commerce platforms (Amazon, eBay). These sources are chosen because they provide rich, diverse, and dynamic data reflecting global events, user behaviour, and market trends. For preprocessing, several techniques are applied to clean and prepare the data for analysis. First, irrelevant information, duplicate entries, HTML tags, and special characters are removed using regular expressions to ensure the dataset retains only meaningful text. Missing values are handled through imputation techniques like mean replacement for numerical data or forward filling for time-series data to maintain dataset integrity. Text normalization is performed by converting text to lowercase, removing stopwords, and applying stemming or lemmatization, which reduces Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4901 Compared to traditional optimization algorithms like Adam or SGD, which require careful tuning of hyperparameters and often struggle with highdimensional or noisy data, GROA provides a more adaptive and guided approach. While Adam excels in computational efficiency, it is prone to overfitting in dynamic data environments. GROA, on the other hand, integrates global search with guided exploitation, making it more robust to dynamic changes and better suited for continual learning scenarios. By integrating GROA into the training pipeline, the model achieves superior performance, stability, and adaptability. Coupled with techniques like EWC and Cosine Annealing for learning rate scheduling, the training process ensures the model avoids catastrophic forgetting while retaining robust generalization capabilities. This cohesive approach is designed to handle complex, evolving datasets and deliver high performance in dynamic environments. 4.8 Trend Prediction and Reporting The final stage of the framework focuses on forecasting future trends and presenting insights in an actionable and interpretable format. After training the model on diverse and dynamic data streams, the prediction pipeline uses its learned representations to identify patterns and temporal relationships across domains. For example, trends in technology adoption are predicted by analyzing patterns extracted from news articles, research publications, or social media data. The system quantifies these predictions with a confidence interval, ensuring a high degree of reliability. Only trends surpassing a 95% confidence threshold are flagged for reporting, guaranteeing both precision and robustness. The system incorporates advanced techniques like Guided Remora Optimization Algorithm (GROA) to fine-tune the model's parameters during training, enhancing its ability to generalize across evolving datasets. Additionally, interpretability mechanisms such as SHAP (Shapley Additive Explanations) provide insights into the model's decision-making process, highlighting key features or temporal relationships that influence predictions. Compared to traditional forecasting methods like ARIMA or machine learning models such as Random Forests, the proposed framework excels in handling complex, noisy, and multi-modal data, ensuring superior performance in dynamic environments. Reports generated by the system include trend predictions alongside explanations and visualizations. For example, a report predicting the rise of a technology might include time-series graphs, heatmaps, and a textual analysis of contributing factors, such as market demands or research activities. This ensures that decisionmakers not only receive accurate forecasts but also gain insights into the underlying dynamics driving the trends. The system also supports periodic updates, enabling real-time monitoring and adaptability to changing datasets. To ensure that our model provides accurate predictions of future trends, we rigorously validate its performance using multiple evaluation metrics, including Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), on historical data with known outcomes. Cross-validation is employed across diverse datasets to assess the generalizability of predictions. Furthermore, the model's ability to consistently achieve high confidence levels and align closely with real-world trends demonstrates its reliability and robustness. In conclusion, this framework represents a comprehensive solution for dynamic trend prediction and reporting. By combining robust optimization techniques, advanced interpretability methods, and adaptive learning mechanisms, it not only forecasts trends accurately but also ensures that stakeholders have access to actionable and interpretable insights. This holistic approach equips decision-makers with the tools to anticipate future developments effectively, making the framework an invaluable asset in addressing complex, evolving challenges across domains. The Figure 1 shows the workflow of the proposed system. Figure 1: Proposed System Workflow Algorithm: Complete Proposed System Input:  Data Sources D: News Websites, social media, e-commerce platforms. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4902  Pre-trained model parameters 𝜃.  Initial hyperparameters: learning rate α – 0.001, batch size B = 32.  Replay buffer R, Fisher information Matrix F.  Confidence Threshold τ = 0.75. Output  Selected features and predicted trends. 1: Intialize parameter 𝑡,𝑁,𝛽,𝐶 2: Collect data D and preprocess using: 2.1: Remove duplicated, HTML tags and special characters using regex. 2.2: Normalize text by converting to lowercase and removing stopwords. 2.3 apply tokenization and numerical scaling using Min-Max Scaling. 3: Feature Extraction using generative AI models: 3.1: Textual Features: use BERT and GPT to extract semantic relationships: 𝐸 = 𝑇𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑟(𝑋) 3.2: Visual Features: Use Visison Transformers(ViT) to capture spatial patterns: 𝐸 = 𝑇𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑟(𝑋) 3.3 Temporal Features: Use Temporal Convolutional Transformer(TCT) to retain sequential dependencies: 𝐸 = 𝑇𝐶𝑇(𝑋) 4: While 𝑡 ≤ 𝑡 4.1: Update the adaptive factor α using the PGRO Optimization: 𝛼= 𝛼 .𝐺𝑅𝑂(𝜃,𝑑) 4.2: For each feature in the dataset: 4.2.1: Update archive using the Penalty-Based Boundary Intersection(PBI) scalarization method: 𝑃𝐵𝐼(𝑥)= 𝑑+ 𝜃.𝑑 where 𝑑= ‖𝑥−𝑧‖, 𝑑 is the distance to the boundary, and 𝑧 is the reference point. 4.2.2 Sort archive members in descending order based on 𝑃𝐵𝐼(𝑥). 4.3: Merge 𝑃= 𝑃+ 𝑃. 4.4: Perform non-dominated sorting of features to refine 𝑃. 4.5: If 𝑟 < 0.5: 4.5.1: Update positions using Guided HBA: 𝑋 = 𝑋+ 𝛽.(𝑋 − 𝑋)+ 𝛼 .(𝑋−𝑋) where 𝛽 controls the influence of the best solution, 𝛼 is the adaptive factor, and 𝑋 is a randomly selected neighbor. 4.6: Else: 4.6.1: Update positions using a global search: 𝑋 = 𝑋+ 𝛼 .𝑋− 𝑋, where 𝑋 is the global best position. 4.7: Perform selection using NSGA-III 4.8: Apply crossover and mutation: 4.8.1: Crossover: 𝑋 = 𝜆 .𝑋 +(1−𝜆).𝑋 where 𝜆 is a random scalar. 4.8.2: Mutation: 𝑋 = 𝑋+ 𝛿 .𝑟𝑎𝑛𝑑(−1,1) where 𝛿 controls mutation strength. 5: Adaptive learning mechanism: 5.1: Elastic weight consolidation (EWC) for regularization: 𝜃 = 𝜃− 𝜆 .𝐹.(𝜃− 𝜃), where F is the Fisher information Matrix 5.2: Update replay buffer R: Combine prior data with new batches. 5.3: Dynamically adjust thresholds τ: 5.3.1: If Precision or Recall < 0.75: 𝜏 = 𝜏−0.05 Else: 𝜏 = 𝜏+0.05 6: Train Temporal model on combined data (replay buffer + new) 7: Evaluate model performance: 7.1: Use metrics: Accuracy(A), Root Mean Square Error(RMSE), F1-Score 7.2 Save updated parameters if performance satisfies: 𝐴 >0.9 𝑎𝑛𝑑 𝑅𝑀𝑆𝐸 < 0.05 8: Trend prediction and reporting: 8.1: Predict trends surpassing a 95% confidence threshold. 8.2: Generate visualizations and explanations using SHAP values. 𝜙= 𝜈 ( 𝑆 𝑈 {𝑖})− 𝜈 (𝑆) where 𝜙 is the contribution of feature 𝑖, S is a subset of features, and 𝜈 (𝑆) is the model output for S. 9: Stop when the criteria are satisfied. 10: Return the selected features and predicted trends. 5. PERFORMANCE ANALYSIS AND COMPARISON Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4903 The Feature Extraction Performance highlights the effectiveness of the proposed framework in handling diverse data modalities, with BERT-base, Vision Transformer (ViT-base), and Temporal Convolutional Transformer (TCT) outperforming competing models in the Table 1. For textual data, BERT-base (Proposed) achieves the highest accuracy (94.5%) by capturing nuanced contextual dependencies, significantly surpassing traditional models like TF-IDF (78.3%) and Word2Vec (84.7%). While XLNet performs well (92.7%) due to its permutational attention mechanism, it lags behind BERT-base in both accuracy and efficiency. For visual data, ViT-base (Proposed) excels in spatial feature extraction with an accuracy of 92.8%, outperforming EfficientNet-B4 (90.4%) and traditional convolutional models like ResNet-50 (88.2%). While EfficientNet-B4 demonstrates lower latency (35 ms), its accuracy is slightly compromised compared to the transformer-based approach. MobileNetV2, though efficient, delivers the lowest accuracy (85.4%), making it less suitable for high-resolution trend prediction. In the temporal domain, TCT (Proposed) emerges as the most reliable model, achieving an accuracy of 91.6% by effectively capturing sequential dependencies and long-term patterns. It outperforms both Transformer XL (89.5%) and traditional recurrent models like LSTM (87.3%) and GRU (88.1%), which struggle with complex temporal relationships. Despite marginally higher latency, TCT demonstrates a strong balance between performance and computational feasibility. Overall, the proposed framework demonstrates superior adaptability and performance across modalities, leveraging advanced transformer architectures to achieve state-of-the-art results while maintaining reasonable latency. This makes it an ideal choice for applications requiring robust, multimodal data processing. For text, BERT-base outperforms traditional methods like TF-IDF and Word2Vec, showcasing the necessity of context-aware mechanisms in capturing nuanced relationships. In visual data, ViTbase sets a new standard by surpassing convolutional models like ResNet-50, emphasizing the shift towards attention-based architectures for spatial understanding. Similarly, in temporal data, TCT demonstrates superior accuracy compared to LSTM and GRU by effectively handling long-term dependencies without the gradient issues of recurrent models. While traditional techniques remain useful in resource-constrained settings, the table underscores that transformer-based approaches are indispensable for achieving robust, highprecision results in dynamic, multi-modal environments. Table 1: Feature Extraction Performance comparing proposed framework with other related techniques. Data Type Model/Tec hnique Featu res Extra cted Accur acy (%) Late ncy (ms) Text BERT-base (Proposed) 768 94.5 45 XLNet 768 92.7 45 TF - IDF 200 78.3 25 Word2Vec 300 84.7 30 Imag es Vision Transforme r (ViTbase) (Proposed) 768 92.8 50 EfficientNe t - B4 1792 90.4 35 ResNet - 50 2048 88.2 40 MobileNet V2 1280 85.4 30 Temp oral Temporal Convolutio nal Transforme r (TCT) (Proposed) 512 91.6 60 Transforme r XL 1024 89.5 70 LSTM 256 87.3 50 GRU 256 88.1 45 The results in the Table 2 demonstrate that the Proposed PGRO framework is the most effective optimization technique, achieving the highest accuracy (94.5%) and F1-Score (93.2%), while also converging in the shortest time (20 epochs). These outcomes align directly with the aim of the proposed work, "A Generative AI Approach to Dynamic Trend Prediction," by ensuring that the model can adapt quickly to evolving data and generate precise trend predictions in real-time. The high F1-Score indicates a strong balance between precision and recall, making PGRO especially valuable for dynamic and imbalanced datasets, where identifying nuanced trends is critical. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4904 In contrast, other techniques, such as Bayesian Optimization, while competitive with a high accuracy of 93.0% and F1-Score of 91.9%, are computationally heavy and less suitable for the fastpaced nature of trend prediction. Traditional optimizers like SGD, RMSProp, and Momentum exhibit lower accuracy (88.7%–90.5%) and F1Scores, reflecting their inability to adapt effectively to complex, multi-modal data distributions. Similarly, population-based techniques like Particle Swarm Optimization and Genetic Algorithm perform reasonably well but converge slower, which limits their practicality for dynamic, real-time applications. The PGRO framework's superior performance directly supports the aim of dynamic trend prediction by enabling quick, accurate, and reliable insights across multi-modal data streams. Its ability to efficiently adapt to changing patterns and maintain robust generalization ensures that emerging trends can be identified and acted upon with precision, making it an indispensable tool for real-world applications in industries such as market analysis, social media monitoring, and technology forecasting. Table 2: Proposed Optimization algorithm (PGRO) Method With 10 Other Optimization Techniques. Technique Accurac y (%) F1Scor e (%) Convergenc e Time (epochs) Proposed (PGRO) 94.5 93.2 20 Adam Optimizer 91.8 90.5 25 SGD 88.7 87.0 40 RMSProp 89.2 87.9 35 AdaGrad 87.3 85.5 45 Momentum 90.5 89.3 30 Particle Swarm Optimizatio n 92.0 91.1 30 Genetic Algorithm 91.5 90.2 35 Hyperband 89.8 88.6 40 Bayesian Optimizatio n 93.0 91.9 25 The graph in the Figure 2 highlights the exceptional performance of the Proposed PGRO method, which achieves the lowest training loss (0.015) and validation loss (0.020) among all techniques. This indicates its ability to learn effectively from the training data while maintaining excellent generalization on unseen validation data. In contrast, traditional methods like SGD and AdaGrad show higher loss values, reflecting their slower adaptation to dynamic and complex datasets. Bayesian Optimization and Particle Swarm Optimization perform competitively, but their slightly higher validation losses indicate less efficient generalization compared to PGRO. The increasing gap between training and validation losses for methods like Grid Search highlights their tendency to overfit, making them less suitable for dynamic trend prediction tasks. The consistent low loss values for PGRO align directly with the aim of dynamic trend prediction, as they ensure accurate and robust learning even in rapidly evolving environments. By converging efficiently with minimal loss, PGRO demonstrates its superiority in handling multi-modal data and achieving real-time predictions. The Mean Squared Error (MSE) was employed to calculate both the training loss and validation loss. Training Loss was computed on 80% of the dataset, which was used for training the model and Validation Loss was calculated on the remaining 20% of the dataset, ensuring that the validation data was not used during training to provide an unbiased assessment of model generalization. Figure 2: Comparation of training loss and validation loss across various techniques The Trend Prediction Confidence Table 3 highlights the exceptional reliability and precision of the Proposed PGRO framework, achieving the highest confidence score (96.5%) and Predicted vs Actual Match (95.3%). These results demonstrate the model's ability to accurately predict dynamic trends, such as AI adoption, while maintaining high Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4905 certainty in its outputs. PGRO’s performance showcases its adaptability to evolving patterns and its robustness in generating actionable insights for real-world scenarios. In comparison, models like Bayesian Optimization and Particle Swarm Optimization deliver competitive results (93.0% and 92.7% match scores, respectively) but fall short of PGRO in terms of precision and computational efficiency. Traditional optimizers like Adam and Momentum exhibit moderate reliability, while SGD and AdaGrad struggle with lower confidence and accuracy, making them less effective for rapidly changing trends. Techniques such as Grid Search and Hyperband lag significantly, as they fail to handle complex datasets efficiently, further validating the need for a more adaptive framework like PGRO. The analysis confirms that PGRO’s superior prediction accuracy and high confidence levels align perfectly with the goal of dynamic trend prediction, making it the optimal choice for scenarios requiring fast, accurate, and reliable insights. Table 3: Trend Prediction (Trend 1 (AI Adoption)) Confidence Comparing The Proposed Model With 10 Other Models. Trend Name Model Predic ted Value Confid ence Score (%) Predic ted vs Actua l Match (%) Trend 1 (AI Adopti on) Propose d (PGRO) 0.85 96.5 95.3 Adam Optimiz er 0.82 92.3 91.0 SGD 0.78 89.0 86.5 RMSPro p 0.79 89.5 87.0 AdaGra d 0.75 87.0 84.2 Moment um 0.80 91.0 89.4 Particle Swarm Optimiz ation 0.83 94.5 92.7 Genetic Algorith m 0.81 93.0 91.5 Hyperba nd 0.79 90.0 87.8 Bayesia n Optimiz ation 0.84 95.2 93.0 Grid Search 0.76 88.5 85.0 Table 4: Trend Prediction (Trend 2 (Electric Vehicles)) Confidence Comparing The Proposed Model With 10 Other Models. Tren d Name Model Predic ted Value Confid ence Score (%) Predic ted vs Actual Match (%) Tren d 2 (Elect ric Vehic les) Propose d (PGRO) 0.88 95.8 94.7 Adam Optimiz er 0.84 91.5 90.2 SGD 0.80 88.5 85.0 RMSPro p 0.81 89.8 86.7 AdaGrad 0.78 87.2 84.5 Moment um 0.82 91.8 89.7 Particle Swarm Optimiz ation 0.86 94.2 92.5 Genetic Algorith m 0.85 93.0 91.8 Hyperba nd 0.80 90.5 88.0 Bayesian Optimiz ation 0.87 95.0 93.2 Grid Search 0.79 89.0 85.7 Table 5: Trend Prediction (Trend 3 (Remote Work)) Confidence comparing the proposed model with 10 other models. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4906 Tren d Nam e Model Predic ted Value Confide nce Score (%) Predic ted vs Actual Match (%) Trend 3 (Rem ote Work ) Proposed (PGRO) 0.87 96.2 94.9 Adam Optimize r 0.83 92.0 90.5 SGD 0.79 88.7 86.0 RMSPro p 0.80 89.2 86.4 AdaGrad 0.76 87.3 84.8 Moment um 0.82 91.5 89.2 Particle Swarm Optimiza tion 0.85 94.2 92.5 Genetic Algorith m 0.83 93.2 91.3 Hyperba nd 0.81 89.5 87.0 Bayesian Optimiza tion 0.86 95.0 93.2 Grid Search 0.78 88.0 85.0 Table 6: Trend Prediction (Trend 4 (5G Technology)) Confidence Comparing The Proposed Model With 10 Other Models. Trend Name Model Predi cted Value Confid ence Score (%) Predi cted vs Actua l Matc h (%) Trend 4 (5G Technol ogy) Propose d (PGRO) 0.89 96.7 95.6 Adam Optimiz er 0.85 92.8 91.4 SGD 0.81 89.3 86.7 RMSPr op 0.82 89.8 87.3 AdaGra d 0.79 87.5 84.7 Moment um 0.84 91.9 89.5 Particle Swarm Optimiz ation 0.87 94.8 93.0 Genetic Algorith m 0.86 93.5 91.8 Hyperba nd 0.82 90.7 88.3 Bayesia n Optimiz ation 0.88 95.5 93.8 Grid Search 0.80 88.8 85.5 Table 7: Trend Prediction (Trend 5 (E-Commerce Growth)) Confidence Comparing The Proposed Model With 10 Other Models. Trend Name Model Predic ted Value Confid ence Score (%) Predic ted vs Actua l Match (%) Trend 5 (Ecomm erce Growt h) Propose d (PGRO) 0.92 97.0 96.0 Adam Optimiz er 0.87 93.2 92.0 SGD 0.83 89.7 86.8 RMSPro p 0.84 90.0 87.5 AdaGra d 0.80 88.5 85.0 Moment um 0.86 92.5 90.7 Particle Swarm Optimiz ation 0.89 95.0 93.5 Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4907 Genetic Algorith m 0.88 94.2 92.8 Hyperba nd 0.84 91.2 89.0 Bayesia n Optimiz ation 0.91 96.0 94.5 Grid Search 0.82 89.0 85.3 The analysis of the Trend Prediction Confidence Table 4, Table 5, Table 6, Table 7 reveals that the Proposed PGRO framework consistently delivers superior performance across all five trends, achieving the highest confidence scores and Predicted vs Actual Match percentages. This demonstrates the robustness and adaptability of PGRO in accurately capturing dynamic and evolving trends. Techniques such as Bayesian Optimization and Particle Swarm Optimization perform competitively but fall short in terms of computational efficiency and precision. Traditional methods like SGD and AdaGrad struggle with lower accuracy and confidence, making them less effective for fast-changing trends such as AI Adoption and Ecommerce Growth. The ability of PGRO to achieve consistent high confidence and accuracy directly supports the aim of the proposed work: “A Generative AI Approach to Dynamic Trend Prediction.” By effectively generalizing across diverse data patterns and evolving scenarios, PGRO ensures timely and reliable trend predictions, crucial for real-world applications such as market forecasting, consumer behaviour analysis, and technological advancements. The framework’s capability to outperform competing techniques underscores its potential to revolutionize trend prediction by combining accuracy, adaptability, and computational efficiency. The Figure 3 shows the point graph for Epochs vs Accuracy highlights the performance of different algorithms in terms of accuracy relative to the number of training epochs. The Proposed PGRO framework achieves the highest accuracy (94.5%) in only 20 epochs, demonstrating superior efficiency and precision compared to all other models. Algorithms like Bayesian Optimization and PSO show competitive accuracy (93.0% and 92.0%, respectively) but require more epochs (25 and 30), indicating slower convergence. Traditional methods like SGD and AdaGrad underperform significantly, with lower accuracy (88.7% and 87.3%) and longer training durations (40+ epochs). This analysis emphasizes that the Proposed PGRO framework achieves the best trade-off between accuracy and efficiency, making it an optimal choice for tasks requiring both high performance and faster convergence, aligning directly with the goals of dynamic trend prediction. Figure 3: Epochs vs Accuracy The Figure 4 illustrates the relationship between the number of epochs and the time required per epoch for various optimization techniques. Our proposed model, PGRO, demonstrates a distinct advantage with the lowest time per epoch, even with competitive epoch counts. This indicates a significant improvement in computational efficiency, making it a highly scalable and resourcefriendly approach. In contrast, other techniques like Grid Search and AdaGrad, while achieving similar epoch numbers, incur much higher time costs, making them less suitable for dynamic trend prediction tasks. Techniques like Bayesian Optimization and Adam Optimizer also perform relatively well but still lag behind PGRO in efficiency. This performance is directly aligned with the aim of the proposed work, which focuses on dynamic trend prediction. By reducing computation time, PGRO enables faster adaptability to trends, thus addressing real-time prediction needs effectively. This balance between reduced time and competitive performance reinforces PGRO’s suitability for generative AI applications in dynamic environments. Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4908 Figure 4: Epochs Vs Time The Figure 5 illustrates the relationship between epochs and Mean Squared Error (MSE) across various models. Our proposed PGRO model achieves the lowest MSE at 0.015 with minimal epochs, demonstrating superior accuracy and efficient training. This highlights PGRO’s ability to converge quickly with minimal error, making it well-suited for precise trend prediction. Other models, such as Grid Search and AdaGrad, exhibit higher MSE despite longer training times, reflecting inefficiencies in error minimization. Techniques like Bayesian Optimization and Adam Optimizer perform moderately but are still outperformed by PGRO. This analysis reinforces that PGRO’s optimization process aligns with the goal of dynamic trend prediction by minimizing error effectively, ensuring reliable and real-time generative AI predictions. Figure 5: Epochs vs Mean Square Error (MSE) The generated graph in the Figure 6 highlights the performance progression of the Proposed (PGRO) model over 24 epochs. Initially, the accuracy increases consistently, indicating the model's ability to learn and adapt effectively during training. This growth continues up to epoch 20, where the accuracy reaches a maximum value of 94.5%, demonstrating the efficiency of the model in achieving high performance. Beyond epoch 20, the accuracy stabilizes at 94.5%, signifying that the model has converged and further training does not lead to any significant improvement. This stabilization suggests that the model has effectively captured the underlying patterns in the data, avoiding overfitting or unnecessary adjustments. This progression aligns with the aim of the proposed work, which focuses on dynamic trend prediction. The stable accuracy after 20 epochs ensures reliable and consistent predictions, validating the robustness and efficiency of the proposed approach. Figure 6: Accuracy Progression Across Epochs for Proposed (PGRO) The graphs shown in Figure 7 illustrate the training and testing accuracy, as well as the loss of the proposed method over 20 epochs. The accuracy graph on the left shows a steady increase, with training accuracy rising from approximately 60% to nearly 100%, while testing accuracy follows a similar trend, reaching above 90%. The gap between training and testing accuracy remains moderate, indicating effective learning with minimal overfitting. The loss graph on the right exhibits a sharp decline in both training and testing loss during the initial epochs, dropping from around 2.2 to below 0.5 by epoch 7 and stabilizing at approximately 0.1 after epoch 10. The similarity in training and testing loss curves suggests that the model generalizes well to unseen data. Overall, the Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4909 proposed method demonstrates strong learning capability, effective optimization, and good convergence behaviour. Figure 7: Training And Testing For Accuracy And Loss Of Proposed Method The time-series heatmap of trends shown in the Figure 8 provides a comprehensive visualization of prediction confidence across five key domains Ecommerce, Entertainment, Finance, Healthcare, and Technology over a defined period. The color gradient represents prediction confidence levels, where higher values (depicted in red) indicate strong certainty, while lower values (shown in blue) suggest weaker confidence. A detailed analysis reveals that E-commerce and Technology exhibit relatively stable and high confidence trends throughout the year, with peak values exceeding 0.9 in multiple instances. E-commerce, in particular, demonstrates strong predictive confidence during March to May 2023, whereas Technology maintains high confidence in March, September, and December 2023. Conversely, Finance displays noticeable fluctuations, oscillating between high (0.98, 0.97) and low confidence values (0.52, 0.53), indicating potential market volatility or external disruptions impacting predictive stability. Similarly, Entertainment and Healthcare exhibit substantial variability, with Entertainment showing lower confidence levels (<0.6) in several months, except for sharp increases in March and August 2023. Healthcare follows a similar pattern, despite an initial strong confidence level of 0.92 in January 2023, with subsequent inconsistencies. These fluctuations highlight the dynamic nature of certain domains, suggesting that external factors, data distribution shifts, or evolving market conditions contribute to variations in predictive certainty. This analysis underscores the importance of robust adaptive learning mechanisms to enhance the generalization and stability of predictive models in dynamic environments. Figure 8: Time-Series Heatmap Of Trends On Small Portion Of Data Present In Dataset Our proposed framework aims to predict emerging trends using multi-modal data (text, images, timeseries) while ensuring adaptability across different domains. The Radar Chart (Spider Chart) is crucial for evaluating the model’s generalization across diverse domains. The framework uses DomainAdversarial Neural Networks (DANN) for crossdomain generalization. A well-balanced radar shape indicates consistent performance across all domains. he radar chart shown in Figure 9 illustrates the crossdomain generalization performance of the model across multiple industries, including Technology, Healthcare, Finance, E-commerce, Education, Manufacturing, and Retail. The four key evaluation metrics Accuracy, Precision, Recall, and F1-Score demonstrate the model's effectiveness in maintaining consistent performance across diverse domains. The nearly uniform distribution of the plotted values, all closely clustered near the outer ring, suggests that the model achieves high predictive reliability across sectors, with minimal performance variation. Technology and Finance exhibit slightly higher scores, particularly in Precision and F1-Score, indicating the model’s strength in making accurate predictions while balancing recall and precision effectively. In contrast, Education and Retail show marginally lower values, suggesting potential challenges in generalizing predictions within these Journal of Theoretical and Applied Information Technology 15th June 2025. Vol.103. No.11 © Little Lion Scientific ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195 4910 domains, likely due to more dynamic or less structured data distributions. The chart confirms the model’s ability to generalize effectively across various industries, with consistently strong recall and precision scores. The balance between these metrics indicates robustness in identifying true positives while minimizing false positives and negatives. The minimal performance variance suggests that the applied optimization techniques and domain adaptation strategies effectively enhance generalization, making the model well-suited for dynamic, cross-domain applications Figure 9: Cross-Domain Generalization Performance Radar Chart A Temporal Dependency Visualization (Attention Map) is crucial in understanding how the model learns dependencies in time-series data. This insights into how different attention heads focus on various time steps when making predictions. The attention map visualization in the Figure 10 shows the temporal dependencies captured by different attention heads within the model. Each row represents an attention head, while the columns denote different time steps, providing insights into how the model distributes its attention across sequential inputs. The intensity of the color, as indicated by the scale on the right, corresponds to the attention weight assigned to each time step, with higher values (yellow) signifying stronger focus and lower values (purple) indicating diminished importance. From the visualization, it is evident that attention is dynamically allocated across different time steps. For instance, attention head 1 assigns a high weight (0.94) to the first time step, suggesting that initial data points significantly influence the model’s predictions. Similarly, attention head 4 exhibits peak attention at time steps 7 and 9, with weights reaching 1.00 and 0.98, respectively, indicating the model’s reliance on these points for key decision-making. Conversely, some regions, such as time step 7 in attention head 2, receive minimal attention (0.00), suggesting that this particular input holds little significance in the model's learned dependencies. The structured yet varied distribution of attention weights across different heads reflects the model’s ability to capture both short-term and long-term dependencies in sequential data. The presence of multiple attention peaks across different heads highlights the benefits of multi-head attention, ensuring that diverse temporal patterns are accounted for. This visualization confirms that the model effectively learns temporal relationships, making it well-suited for applications requiring sequential reasoning, such as financial forecasting, healthcare diagnostics, and time-series analysis. Figure 10: Attention Heads versus Time Steps Table 8 demonstrate the effectiveness of our proposed framework, we present a comparative analysis against various models and also the models discussed in the literature. While previous sections have detailed performance improvements at various stages of the pipeline, this final comparison provides a holistic evaluation of our model’s overall accuracy, recall, and F1-score relative to existing approaches. The comparison includes traditional statistical models, machine learning-based methods, deep learning architectures, and transformer-based