With the increasing demand for sustainable energy solutions, accurate biomass forecasting has become essential for optimizing renewable energy utilization and supporting system-level predictive maintenance and operational decision-making. However, the complexity and high dimensionality of biomass energy data pose significant challenges for conventional forecasting approaches. To address these challenges, this study proposes a framework integrating Spatio-Temporal Graph Convolutional Networks (STGCN) with metaheuristic optimization techniques. The proposed approach improves forecasting performance by capturing spatio-temporal dependencies while optimizing feature selection and model hyperparameters. Initially, the baseline STGCN model achieved an MSE of 0.0025, RMSE of 0.0500, and R^2 of 0.8317. Feature selection using the Binary Comment Feedback Optimization Algorithm (bCFOA) reduced redundancy and improved the forecasting results to an MSE of 0.0018, RMSE of 0.04243, and R^2 of 0.912. Further improvement through CFOA-based hyperparameter optimization produced the best performance, achieving an MSE of 0.000554 ± 0.000012 , RMSE of 0.02354 ± 0.00028 , MAE of 0.00410 ± 0.00009 , and R^2 of 0.981 ± 0.002 . The results demonstrate that combining graph-based spatio-temporal learning with evolutionary optimization significantly improves forecasting accuracy and model generalization. Moreover, the proposed CFOA-STGCN framework provides a scalable, interpretable, and computationally efficient solution suitable for biomass energy forecasting and operational decision-support systems.
A structured computational framework for conjecture generation in number theory is presented, in which a sequential pipeline is formally defined from data acquisition to hypothesis construction. In contrast to heuristic driven approaches, the proposed model is based on an explicit transformation process in which curated numerical data are encoded into stable embedding representations, followed by extraction of latent statistical patterns and subsequent formulation of candidate conjectures. The embedding stage is designed to preserve arithmetic and structural invariants, while statistical learning procedures, including clustering and symbolic regression, are applied to identify regularities across high dimensional representations. A rigorous experimental design is implemented, in which datasets are partitioned into training, validation, and testing subsets, and multiple random seeds are used to ensure robustness and reproducibility. Statistical significance of the extracted patterns is evaluated through controlled hypothesis testing, and multiple comparison effects are addressed using false discovery rate correction. The resulting conjectures are assessed both quantitatively, via goodness of fit measures and out of sample validation, and qualitatively, through consistency with known theoretical results. The findings indicate that the proposed pipeline yields statistically reliable and mathematically meaningful conjectures, while reducing bias associated with purely intuition driven methods. This study establishes a reproducible and interpretable bridge between machine learning based pattern discovery and formal mathematical reasoning, and provides a scalable framework for systematic exploration in pure mathematics.
Location-based social networks (LBSNs) are important platforms for business applications due to the rich combination of user relationships and geographical information. One of the key challenges in this context is identifying influential users by estimating social influence, which plays a central role in targeted advertising and behavior prediction. Prior studies suggest that user similarity is a significant factor contributing to social influence. Traditional methods typically compute similarity using structural metrics, while more recent approaches adopt embedding techniques to capture user proximity in vector space. However, these methods often reflect only limited aspects of user characteristics, overlooking the synergy between network structure and spatial behavior. To address this, our research introduces and evaluates two progressive methodological contributions. First, we propose the NSH (Node Embedding Spatial Homophily) framework, a robust and interpretable hybrid method that linearly combines embedding-based structural similarity with spatial homophily. This unsupervised fusion provides a strong baseline for analyzing the relationship between similarity and influence. Building upon this, our core innovation is the GeoSocial2Vec framework, a novel, end-to-end unified embedding model. It learns a single representation for each user by intrinsically merging social ties and spatial co-visitation patterns within a heterogeneous graph via an adaptive random walk mechanism, advancing the state-of-the-art in joint socio-spatial representation learning. We evaluate the relationship between user similarity and social influence through Pearson correlation and regression analysis on geographically filtered segments of three real-world LBSNs: Gowalla, Brightkite, and Foursquare. Our results demonstrate that both proposed methods—where spatial homophily is integrated with similarity measures—produce a significantly stronger correlation with observed influence patterns compared to classical approaches. The GeoSocial2Vec framework, in particular, achieves superior performance, highlighting the critical role of intrinsically coupled socio-spatial learning. Furthermore, while GeoSocial2Vec offers the highest accuracy, the NSH framework maintains advantages in interpretability and computational simplicity, offering a practical tool for various application scenarios.
A mission to explore a distant planet is being planned by the Pakistan Space and Upper Atmosphere Research Commission. For this mission to be successful, they must select the most suitable nanomaterials for various parts of the spacecraft, including the hull, solar panels, and energy storage systems. For this evaluation, we propose a novel decision making model based on a fuzzy neural network under the fuzzy credibility information to select the optimal nanomaterial for these parts. To do this, we collect the information about the nanomaterials in spacecraft from three specialists. The hidden layer of the fuzzy neural network is calculated by combine the input data with their corresponding weights using the Frank aggregation operator. The importance of each criterion is determined by experts using the Shannon entropy method. The Criteria Importance through Intercriteria Correlation method is then applied to calculate the hidden layer weights, and these weights are used again with the Frank aggregation operator to combine the hidden layer information. Next, we calculate the output scores and apply activation functions to get the final results of the fuzzy neural network. We also check how sensitive the model is to changes in the Frank parameter and compare the results of our proposed model with other existing decision-making models. The comparison demonstrates that the proposed approach is useful and reliable as a decision support system.
Industrial Internet of Things (IIoT) is an application of IoT, which generates high volume, high velocity, and heterogeneous telemetry streams from geographically distributed edge devices in supply-chain networks. Federated learning (FL) has the advantage of not sending sensitive operational data to a central server to achieve privacy-preserving intrusion detection, but the performance of FL is sensitive to non-independent and identically distributed traffic, significant class imbalance, communication limitations, and malicious updates from clients. This study aims to provide a threat-aware self-healing federated intrusion detection system (TSHF-IDS) for secure and scalable IIoT supply-chain networks. The framework combines Independent Component Analysis (ICA) for feature reduction with a light-weight contrastive representation learning module to perform rapid screening of anomalies at the edge. The suspicious traffic is then directed to a Capsule Temporal Threat Detection Network (CTTD-Net) for temporal attack dependency and hierarchical behaviour patterns. This study uses a trust-weighted byzantine-resilient aggregation method, based on trimmed mean aggregation, which assesses the client’s update based on the gradient deviation, historical consistency, behavioral similarity, and anomaly likelihood. A self-healing adaptation layer automatically diminishes the impact of suspicious clients, quarantines malicious updates and recovers validated global checkpoints if the performance degradation due to poisoning is above a certain limit. Model updates are encrypted using AES-GCM authenticated encryption, thus ensuring confidentiality and integrity. Different ratio of malicious clients is used for targeted and untargeted poisoning attacks. The framework is trained on BoT-IoT, and tested on the TON_IoT, Edge-IIoTset and CIC-IoT2023 datasets with stratified and non-IID client partitions. Experimental results on BoT-IoT, TON_IoT, Edge-IIoTset, and CIC-IoT2023 demonstrate that TSHF-IDS achieves an average accuracy of 98.97
Network Function Virtualization (NFV) in the context of Mobile Edge Computing (MEC) enables flexible and modular network services, where Virtual Network Functions (VNFs) are executed sequentially within Service Function Chains (SFCs). However, the sequential execution of SFCs often increases service latency for latency-sensitive applications. While Network Function Parallelism (NFP) reduces latency by enabling concurrent processing across multiple VNFs, the combined effects of SFC partitioning and resource availability have largely been overlooked. To address this gap, we propose a Graph-Based Federated Reinforcement Learning (GFRL) framework for parallel SFC placement, featuring three key capabilities: (1) capturing dependencies among VNFs using graph representations to create partially parallelized structures, (2) partitioning SFCs into multiple sub-SFCs with integrated NFP modules, and (3) deploying sub-SFCs in parallel through training a federated deep reinforcement learning model based on graph neural networks. Meanwhile, GFRL improves resource utilization and load balancing by incorporating live migration. GFRL is formulated as a Markov Decision Process (MDP) aimed at maximizing long-term cumulative rewards. Simulations across diverse scenarios demonstrate that GFRL significantly reduces end-to-end latency and improves resource utilization compared with state-of-the-art methods. Specifically, the proposed GFRL framework achieves, on average, a 4.1
Electric Vehicle (EV) communication in Internet of Electric Vehicles (IoEV) ecosystems is highly dynamic, distributed, and privacy-sensitive, which makes scalable real-time intrusion detection and secure collaborative model training particularly challenging. Existing approaches often do not adequately address energy constraints, context heterogeneity, and explainability requirements, especially in federated environments involving On-Board Units (OBUs), roadside infrastructure, and vehicular cloud services. To address these challenges, we propose HAMID (Hierarchical Aggregation for Mobility-aware Intrusion Detection), a hierarchical federated Intrusion Detection System (IDS) for IoEV systems. It combines lightweight Gated Recurrent Unit (GRU)-based detectors, Deep Reinforcement Learning (DRL)-based participant selection, and privacy-preserving federated learning across a three-tier IoEV infrastructure. HAMID runs in a loop where each Charging-Hub Edge (CHE) node smartly picks energy-efficient and high-quality EVs for training based on real-time data, such as signal quality, CPU capacity, state-of-charge, and charging phase. Some cars locally preprocess telemetry data, break it up into 1-second windows, and use a GRU intrusion detection model that changes as new Vehicle-to-Everything (V2X) threats, like spoofing or message manipulation, arise. Model changes are quantized and sent to CHEs, who then combine them into better global weights that are sent to all nodes. This improves privacy because no raw data leaves EVs. To make decisions clearer, HAMID uses Integrated Gradients (IG) to describe things on the device. HAMID achieves 99.2
In the context of modern big data environments, traditional data quality (DQ) assessment methods (such as business rule-based methods, threshold-based validation, constraint checking, or DAMA) face significant limitations in terms of automation, adaptability and scalability. Also, the literature remains limited in terms of complete methodologies that guide the practical application of intelligent methods for DQ assessment. To face this limitations, this paper proposes a structured and replicable methodology for assessing data quality using machine learning (ML), addressing multiple DQ dimensions such as accuracy, completeness, timeliness, consistency, relevance, and credibility. The approach is built upon a modular pipeline that includes data collection from structured, semi-structured, and unstructured sources, enriched preprocessing phase with domain-aware rules, which refers to context-specific transformations and validations tailored to the semantics of the dataset. Model training uses various ML algorithms, and multi-metric evaluation aligned with DQ dimensions. A novel aspect of our work is the integration of progressive learning, with built-in versioning and traceability mechanisms that allow the system to track model evolution and compare different iterations over time, enabling incremental refinement of models. Experimental validation on five public datasets from diverse domains demonstrates the effectiveness of the proposed framework in enhancing data quality assessment precision, reducing manual effort, and improving model generalizability. This research sets the stage for a new generation of intelligent, self-improving data quality management systems adapted to dynamic and heterogeneous data ecosystems.
Abstract This paper presents ERI-AIC, an Arabic image captioning dataset designed to better reflect Arab cultural contexts and the informational needs of visually impaired individuals. To the best of our knowledge, ERI-AIC is the first dataset to provide original, human-written Arabic captions created from scratch rather than translated from English, while also being guided by a user-study-based understanding of the image description needs of visually impaired users. The dataset was constructed in two main stages. First, we conducted a questionnaire study to identify the interests, preferences, and essential description elements required by people with visual impairments. Second, we developed ERI-AIC by collecting 4,000 images and 12,000 original Arabic captions, with three captions per image, written directly in Arabic to describe culturally relevant scenes and accessibility-oriented details. To evaluate the usefulness of ERI-AIC, we benchmarked baseline captioning models and compared their performance with prior Arabic image captioning resources. Within the reported experimental setting, models trained on ERI-AIC generally achieved higher captioning scores than the corresponding models trained on Arabic Flickr8k and Arabic MS-COCO, with the Transformer model achieving the highest ERI-AIC performance (BLEU-1: 65.1, BLEU-4: 21.4, METEOR: 0.43, CIDEr: 0.38, ROUGE: 0.39). Although the Encoder-Decoder architecture reports unusually high scores on Arabic MS-COCO, this advantage does not transfer to ERI-AIC, suggesting that performance differences are influenced by dataset characteristics, caption style, and model architecture. Overall, the findings indicate that ERI-AIC can serve as a useful training and evaluation resource for culturally grounded and accessibility-oriented Arabic image captioning, while broader claims about generalization require further systematic cross-dataset quantitative evaluation.
Amidst the expanding landscape of the Internet of Things (IoT), integrating intelligent surveillance systems within smart environments is crucial for enhancing public safety and security. This study presents a transformer-based activity recognition (TAR) system that leverages IoT enabled surveillance to autonomously detect and localize suspicious activities in real time. The system utilizes the ResNet 50 architecture for feature extraction and the Detection Transformer (DETR) for precise activity detection. The system was developed using a custom dataset of 6684 images, collected from diverse surveillance environments and augmented to improve training. The dataset encompasses both suspicious and non-suspicious activities, ensuring broad representation. Key performance metrics, including precision, recall, mean average precision (mAP), class loss, box loss, and generalized intersection over union (GIoU), were evaluated to assess the system’s efficacy. The model achieves an accuracy of 91.07
What constitutes an unsupervised approach varies across the time series anomaly detection literature, as methods described as unsupervised often rely on labelled data for threshold selection or hyperparameter tuning. These require a labelled data subset, which is often not available in the real world. Though manually and indiscriminately labelling all recorded data would be possible, it would also be associated with a significant cost, and, especially in big data settings, is unlikely to be feasible. To address this, active learning can be used to select a representative subset of samples. Specifically, this work integrates active learning with unsupervised anomaly detection by querying labels for multivariate time series to more accurately refine the threshold selection process. To achieve this, we introduce a novel query strategy called the dissimilarity-based query strategy (DQS). DQS aims to maximise the diversity of queried samples by evaluating the similarity between anomaly scores using dynamic time warping. We assess the detection performance of DQS and explore the impact of mislabelling, a topic that is underexplored in the literature. Our findings indicate that, for the dataset used, DQS performs best in small-budget scenarios, though the other methods appear to be more robust when faced with mislabelling. While no longer truly unsupervised, all query strategies outperform the unsupervised threshold even in the presence of mislabelling. Thus, whenever it is feasible to query an oracle, employing an active learning-based threshold is recommended.
This work surveys the application of artificial intelligence (AI) towards researching the correlation between two non-communicable diseases, Type 2 Diabetes (T2D) and dementia. These are two prominent non-communicable diseases that affect over 500 million people a year. As prior studies have presented evidence that T2D may strongly increase the likelihood of developing dementia, there has been more attention and research on exploring the link between the two diseases. AI and machine learning (ML) are noted as uniquely well-suited for this research due to their ability to interpret and process large amounts of data. In this context, this survey reviewed studies from 2015 to 2025 and the different modalities in which AI has been applied to exploring this connection. The reviewed works are summarized along with their application of AI or ML and their results. The most common modalities found were those that used clinical, genomic, and radiomics data. Additionally, two distinct perspectives on the relationship between T2D and dementia were observed. These perspectives either assumed the link between T2D and dementia to be established or viewed it as inconclusive. In terms of AI and ML applications, simpler models such as Random Forest (RF) were used most frequently and exhibited overall good performance. Research gaps were also observed which may provide the opportunity for future studies to uncover more evidence of the relationship between these two diseases. Examples include the potential use of LLMs for NLP tasks and methods for addressing imbalanced datasets, such as applying data sampling techniques or utilizing more suitable metrics.
Colorectal cancer (CRC) is one of the most prevalent and life-threatening cancers worldwide, where early detection is crucial for improving patient outcomes. Image segmentation plays a fundamental role in medical analysis by enabling precise polyp detection, yet traditional methods face computational challenges in multilevel thresholding, an NP-hard problem that requires efficient optimization. To address this, this study proposes CoINFOSC, a centroid opposition-based weighted mean of vectors (INFO) algorithm enhanced with harmonic oscillation, for robust multi-threshold segmentation of CRC pathology images, enhancing diagnostic precision and efficiency. The proposed method leverages Kapur entropy as the objective function to determine optimal thresholds, enabling precise segmentation of histopathology images. CoINFOSC integrates centroid opposition and harmonic oscillation strategies to improve the exploration–exploitation balance, enhance population diversity, and prevent premature convergence, thereby achieving highly accurate solutions. The optimization performance of CoINFOSC is rigorously validated through extensive experiments on twenty-five unimodal and multimodal benchmark functions and further tested on the IEEE CEC 2017 for dimensions 30 and 50, and on the IEEE CEC 2019 test suites, demonstrating its superior convergence accuracy and robustness compared to state-of-the-art algorithms. Extensive experiments were conducted to evaluate CoINFOSC’s performance, with segmentation results assessed using six metrics: PSNR (27.72862), SSIM (0.81629), FSIM (0.93167), UIQI (0.17803), QILV (0.97781), and HPSI (0.62943). Furthermore, region-based clinical segmentation metrics, including the Dice coefficient and Jaccard index, were evaluated against expert-annotated ground truth masks. The results demonstrate that CoINFOSC outperforms state-of-the-art algorithms across segmentation accuracy, robustness, and convergence speed. The high-quality segmented images produced by CoINFOSC highlight its effectiveness in handling the complexities of CRC pathology images.
The increasing prevalence of marine pollution and the escalating impacts of climate change have intensified the need for advanced underwater monitoring systems capable of operating reliably in harsh subsea environments. However, underwater environmental monitoring remains a challenging task due to factors such as high hydrostatic pressure, limited visibility, dynamic water conditions, and the inherent constraints of underwater communication channels. To address these challenges, this study proposes a Hybrid Underwater Wireless Sensor Network (UHWSN) that integrates acoustic and optical sensing modalities with advanced deep learning techniques to enable efficient underwater debris detection, tracking, and classification. The proposed framework combines mean-shift-based object tracking, lossless arithmetic coding, Convolutional Neural Networks (CNNs), and Deep Belief Networks (DBNs) within a hierarchical processing architecture. At the sensor-node level, mean-shift tracking is employed for region-of-interest (ROI) localization, while arithmetic coding provides efficient lossless data compression, achieving a data-volume reduction of approximately 55–60
Illegal marketplaces have increasingly migrated to the deep/dark web and high-risk social platforms, facilitating the anonymous trade of drugs, weapons, and stolen credentials. Detecting and categorizing such content remains challenging due to scarce labeled data, rapidly evolving language, diverse structural characteristics, and the limitations of existing approaches in handling noisy, cross-platform illicit content. To address these challenges, we propose a novel two-stage hierarchical semi-supervised framework that integrates domain-adapted ModernBERT embeddings, manually engineered structural features, and an entropy-based ensemble learning strategy for illicit marketplace detection and classification. ModernBERT is fine-tuned on domain-specific data to better capture specialized jargon, obfuscated expressions, and long-context linguistic patterns commonly found in illicit marketplaces. These embeddings are combined with layout, pattern-specific, and metadata features to enrich document representation across heterogeneous platforms. In the first stage, sales-related documents are identified using an ensemble of XGBoost, Random Forest, and SVM classifiers within a self-training framework enhanced by a novel entropy-based weighted voting mechanism that dynamically adjusts classifier contributions based on prediction confidence. In the second stage, three specialized semi-supervised XGBoost classifiers categorize detected sales content into drug, weapon, and stolen credential sales. To develop and evaluate the proposed framework, a 21,575-sample multi-source corpus comprising 1,575 labeled and 20,000 unlabeled samples is collected from the deep/dark web, Telegram, Reddit, and Pastebin. Experimental results demonstrate superior performance, achieving macro-averaged scores of 0.96489 accuracy, 0.93467 F1, and 0.95388 TMCC. Also, the proposed framework consistently improves over baseline models, including ModernBERT, BERT, Longformer, ALBERT, BigBird, and DarkBERT, with accuracy gains of 3.1-6.7
Abstract This article introduces an enhanced version of the Newton-Raphson-based optimizer (I-NRBO) with a complex optimization technique termed the Local Escaping Operator (LEO). This method effectively balances solution variety with convergence speed, enhancing optimization effectiveness and reducing the likelihood of local optima. The research presents an advanced and efficient model that utilizes a pre-trained deep learning architecture, MobileNet, with the I-NRBO algorithm to enhance the classification accuracy of brain tumors. The I-NRBO method is evaluated over CEC-2022, the results showed that I-NRBO outperformed the traditional NRBO and other leading algorithms in terms of statistical convergence and various criteria. In the subsequent study, the proposed I-NRBO is applied to optimize hyperparameters in the MobileNet model, specifically for brain tumor classification. The I-NRBO-MobileNet model is assessed against many previous research and alternative deep learning architectures with comparable parameters, including VGG16, ResNet50, InceptionV3, and MobileNet as baseline models. The experimental findings reveal that the I-NRBO-MobileNet model outstanding performance, with an accuracy of 99.24%, a precision of 99.25%, a recall of 99.24%, an AUC of 99.95%, and a specificity of 99.70%. Results demonstrate that the suggested I-NRBO-MobileNet model can accurately classify brain cancers.
The growth in data volume and complexity of decision-making create the need for new data analytics methods, such as machine learning (ML) and deep learning (DL), which are critical for improving various tasks in supply chain management (SCM). Recent studies have highlighted the benefits of integrating DL/ML in SCM. These efforts have been applied to measures, such as accuracy, precision, sensitivity, recall, and F1 score which can be used to evaluate the results and outputs of the ML/DL techniques; however, they cannot reflect stakeholders’ viewpoints on their benefits and effects on organizations. Therefore, this work aims to introduce a Single-Valued Trapezoidal Neutrosophic Number Weighted Arithmetic Average (SVTNNWAA)-based method for assessing the effect of integrating ML/DL with SCM tasks to benefit organizations. The proposed method integrates the full consistency method (FUCOM) to provide consistent weights and a robust tool, with fewer comparisons. The proposed method is applied with a practical case study, and the results show that it can enable the decision-makers to assess and visualize the benefit rates of ML/DL which indicate the most suitable method for more effective SCM. Additionally, the results indicate that the proposed method can evaluate the effect of ML/DL in SCM from stakeholders’ viewpoints towards their benefits under neutrosophic environment which can handle uncertainty and indeterminacy.
Knowledge-graph-enhanced recommendation can improve personalized retrieval, but existing methods often suffer from noise accumulation in multi-hop propagation and increased computational cost. To address these issues, we propose Quatnet, an interest propagation recommendation model based on quaternion knowledge graph embedding. Quatnet combines quaternion representation and convolutional feature extraction to model multi-relational triplets, and it performs lightweight recommendation using 1-hop interest propagation. In this way, the model strengthens semantic representation while reducing the influence of distant noisy entities. Experiments on MovieLens-1M and Last.FM show that Quatnet achieves competitive recommendation performance. On MovieLens-1M, its AUC and ACC are improved by 0.7
Information retrieval (IR) systems generally start with a user-provided query that guides the search algorithm. The internal information structure across matching documents is used for relevance ranking. IR engines have been developed and optimized for several types of information sources, such as structured, semi-structured, and unstructured documents. These engines typically focus on the content within the documents being indexed and the structure of the overall information. The user’s context (the purpose of the query) is not an active consideration for the IR engine, making the user try various adjustments to the keywords in order to find the information they need. In real-world applications, (i) context manifests as an awareness of aspects surrounding the query such as geography, roles, capabilities, and other component attributes, and (ii) unfortunately often is not directly embedded in the metadata of the information within the documents being indexed. For this to be achieved, an IR engine needs three capabilities: an ability to index the information effectively, an ability to comprehend the search query, and the ability to understand the context of the search query. However, many recent techniques are only capable of deriving intrinsic context from domain-specific corpora and have yet to be integrated with extrinsic context such as a query’s location, or any other aspect. Moreover, existing search techniques are often general and sometimes discrete in nature. We present here a novel end-to-end Context-Aware Search Engine, named CASE, which provides three key features: (i) automating the indexing of documents by using a state-of-the-art contextual keyphrase extraction framework; (ii) understanding a query with both the intrinsic and extrinsic contexts via meta models and query-based knowledge graphs; and (iii) achieving robustness using a combined relevance ranking score. The results of our experiments demonstrate that CASE outperforms other baseline methods when compared to the machine reading comprehension (MS MARCO) dataset.
Developing a recommender system for multi-day tourist route planning that aligns with user preferences presents challenges in optimization. Many previous studies within the scope of Tourist Trip Design Problem (TTDP) have developed systems that recommend multi-day routes by utilizing the analogy of solving the Team Orienteering Problem (TOP). In addition, prior studies have not accommodated the user’s desire to visit a combination of several tourist categories, such as nature, shopping, and culinary, during a multi-day tour. We refer to this problem as a multi-day mixed destination tour. Therefore, in this research, we formulate the multi-day mixed destination tour problem as a Capacitated Team Orienteering Problem with Time Windows (CTOPTW) and introduce a new algorithm, XHABC, a hybridized version of Artificial Bee Colony (ABC) and Harris Hawk Optimization (HHO), to solve the problem. This study specifically focuses on combining two categories of Points of Interest (POIs), i.e., culinary and non-culinary POIs. To ensure the optimality of per-day routes, we develop an algorithm called DaySplitter based on the greedy strategy. In addition, we consider multi-attribute user preferences, which are integrated using the utility function in Multi-Attribute Utility Theory (MAUT). Based on the experimental results, XHABC consistently outperformed the comparison algorithms in both primary and secondary metrics, while additional evaluations further demonstrated its effectiveness, stability, and competitiveness in solving CTOPTW.
