Reliable water quality monitoring is critical for safeguarding public health and ensuring the sustainability of ecosystems, especially in regions facing growing environmental and industrial pressures. This paper presents a novel edge-intelligent framework that combines Graph Neural Networks with real-time IoT sensor deployments to predict and monitor multiple water quality parameters. Leveraging a two-stage spatio-temporal graph construction process grounded in Euclidean and correlation-based criteria, we model the spatial relationships and temporal dynamics of diverse water parameters. Our models achieve strong predictive performance across 14 critical parameters, including temperature, dissolved oxygen, conductivity, and microbial indicators, with R² scores as high as 0.92. Deployed on low-cost Raspberry Pi-based edge devices, our system enables real-time inference and energy-efficient operations without reliance on cloud connectivity. This work bridges the gap between deep learning and in-situ environmental monitoring, demonstrating an IoT approach for data-driven water governance. The proposed solution holds significant promise for policy-makers, researchers, and communities aiming to decentralize environmental sensing and address the global challenge of clean water access.

Session-based cyberbullying detection on social media platforms presents significant challenges owing to the scarcity of labeled data and the complex, heterogeneous nature of social media sessions and interactions between their elements. In this study, we introduce BullyHCL, an innovative unsupervised heterogeneous graph contrastive learning framework specifically designed for session-based cyberbullying detection. This framework utilizes contrastive learning by training a discriminator with binary cross-entropy to maximize the mutual information between node and summary pairs from the original graph while minimizing it for perturbed pairs from augmented views. This approach encourages the encoder to learn representations that capture the essential structural and semantic information from the original graph while being robust to perturbations. An analysis of augmentation strategies and model parameters provides valuable insights into improving unsupervised cyberbullying detection. Experiments conducted on the Instagram and the Vine datasets reveal that the BullyHCL model consistently surpasses existing unsupervised baselines and achieves performance levels comparable to those of supervised models, underscoring its practical value as a label-free alternative.

In this paper, we propose a method for detecting fine-scale surface changes, such as scratches and dents, by combining 3D Gaussian Splatting (3DGS) with Vision-Language Models (VLMs). While 3DGS excels at rendering high-fidelity novel views that make even subtle changes visually detectable, identifying such changes automatically remains challenging. Our approach addresses this by precisely aligning two 3DGS models of the same object, before and after potential damage, and using VLMs to analyze the rendered views for localized differences. Experiments show that our method outperforms a traditional photogrammetry-based method in detecting fine-grained surface alterations.

In this paper, we aim to reduce the number of nodes from Graph Neural Networks (GNNs), thereby simplifying models and reducing computational costs. GNNs are highly effective for various tasks, such as prediction, classification, and clustering, due to their ability to learn node and edge attributes and relationships, and they have been utilized for intelligent transportation systems recently by converting sensor networks into graph structures. Deep spatio-temporal neural networks, including Spatio-Temporal Graph Convolutional Networks (STGCNs), capture spatial and temporal dependencies, making them suitable for traffic speed forecasting, traffic demand prediction, and travel time estimation. Despite their success, GNNs face challenges in industrial applications due to significant memory usage and time consumption. In this paper, we propose a new approach to node reduction that outperforms existing methods in computational efficiency. Our experiments on two real-world traffic datasets demonstrate that using the heuristic and edge information to reduce nodes can cut computation time of optimization up to 95% and, by eliminating noise, can even enhance prediction accuracy.

Wireless sensor networks are widely valued for their effectiveness in real-time data collection. As the amount of data exchanged within such networks grows, designing a robust network topology that maximizes area coverage with minimal sensors has become a critical challenge. The choice of topology impacts key network metrics, including sensor coverage, communication range, connectivity, inference, and installation and management costsIn this paper, we address the Wireless Sensor Network Planning Problem with Multiple Sources/Destinations, presenting an optimization approach based on deep reinforcement learning. This problem is noteworthy, as sensors in various applications are often required to share data within distinct destinationsWe leverage deep reinforcement learning to effectively address the complex task of selecting optimal sensor locations. Our reinforcement learning agent dynamically learns network structure by iteratively adding and removing sensors, optimizing both sensor coverage and the total number of sensors used. Experiment across diverse scenarios demonstrate the effectiveness of our method for network planning problems of varying scales, achieving full coverage with fewer sensors than traditional approaches. Additionally, our approach also produce solutions for large instances where Mixed Integer Programming solvers were not able to. Overall, our method was able to reduce the number of sensors used by up to 22.3% compared to other methods.

Class imbalance and overlapping instances problems have long been recognized as one of the major causes of the performance deterioration of the classification model. Moreover, the majority class may have an irrelevant and noisy instance that shifts the decision boundary of the classification far away from the ideal one. We propose a framework for balancing the class distribution and mitigating the class overlap problem in a dataset. The key feature of our framework is its ability to detect the overlapping instances between classes and then remove the problematic instances from the majority class. Thus, it will have more precise information for the oversampling method to generate the synthetic minority instances. We evaluated the proposed framework using the Lending club and ten other datasets from the KEEL repository. We demonstrate the implementations of our framework using Tomek and Edited Nearest Neighbor for removing the overlapping instances from the majority class and SWIM-MD for generating the synthetic minority instances. Also, we used eight well-known classifiers to show that our proposed framework can improve the performance of various classifiers. Lastly, we present a detailed analysis of the experimental result that shows the superiority of our proposed framework. Our proposed framework outperformed the state-of-the-art methods in terms of geometry mean classification performance metric.

The performance of a machine learning algorithm is highly dependent on its hyperparameters. However, hyperparameter optimization is not a trivial task as it is problem-specific. The difficulty rises when coupled with a larger number of hyperparameters resulting in high search space dimensions. The common understanding seems to be that the optimization is only done on limited hyperparameters. Indeed, a larger number of hyperparameters have not been commonly utilized in hyperparameter optimization. This study investigates the role of hyperparameters by using a genetic algorithm (GA) as the main optimization method for a convolutional neural network (CNN). The novelty of this study is two-fold. Firstly, we defined 20 hyperparameters and their ranges, specifically for text classification. Secondly, we conducted experiments with different numbers of hyperparameters and different numbers of optimized hyperparameters. GA-CNN was evaluated using a disaster tweets dataset and compared to the other methods, i.e., grid search, random search, TPE, TPOT, CNN, LSTM, CNN-LSTM, and BERT. The experimental results demonstrated that the proposed method shows better performance over other methods. The results also showed that a larger number of hyperparameters and layer-specific hyperparameter values are indeed important.

Opinion mining from mobile app reviews has grown exponentially during the last decade. Most studies in this area, however, have focused on a sentiment analysis. In this study, we consider review mining from another perspective, that is, capturing user justifications behind whatever actions are explicitly stated in their app reviews, e.g., the reason behind user purchases. This study highlights how different app features can promote different user decisions, which in turn can be beneficial for software developers to gain valuable data-driven requirements for the planning and development of application updates. We collected, used, and shared our manually annotated 46k mobile app reviews from 12 different app categories in the Google Play Store. We designed three classification problems to filter reviews containing both arguments and decisions from non-argumentative reviews. We extracted three features, namely, structural, lexical, and contextual, from the body of review sentences. Four classifiers (naive Bayes, logistic regression, support vector machine, and random forest) and different feature combinations were trained on the dataset and evaluated to examine if such features can allow us to classify user arguments and decisions. The results show an improved performance over previous studies and show the efficacy of the proposed approach compared to human-assessments.

Reviews are used in generating explainable recommendation. However, the use of reviews has so far not been adequately addressed. In this paper, we examine methods that make use of reviews effectively. There is a trade-off between the number and quality of reviews to use, that is, we should like to use reviews as many as possible to generate explainable recommendation, however in a large number of reviews there can be low quality ones, which can cause low quality explainable recommendation generation. We discuss new methods that use not only reviews written by a user but also those utilized by the user to generate good explainable recommendation. Our methods can be applied to different explainable recommender approaches, which is shown by adopting two state-of-the-art explainable recommender approaches in this paper. Experimental results demonstrate that our methods can be of benefit to existing explainable recommender approaches as regards both recommendation and its explanation qualities.

While a lot of useful information can be found in SNS, false information also diffuses through it, thereby confusing many people sometimes. In this paper, we predict a tendency of tweets to be well retweeted and consider applying the tendency to false information detection. The tendency prediction can be implemented with simple features of tweets. We examine the effect of the tendency when it is used in false information detection empirically. Our experimental results indicate that it would be valuable to take the tendency into account for the detection. We also discuss findings when applying them to tweets in Japanese.

As the number of users of location based social networks (LBSNs) grows, a large volume of valuable data including check-in data have been stored in them and available to us. In this paper, we focus on bursts of visits and show a feasibility study of exploiting them for point-of-interest (POI) recommendation on LBSN. We extract bursts of visits from time series check-in data stored in LBSN and suppose them as a signal that events happened. In other words, the bursts could indicate the locations and durations of some events. We make use of the signals and visited locations of users whose check-in data are similar for simple POI recommendation. Our experimental results indicate that it would be valuable to take bursts of visits into account for POI recommendation.

Semantic text similarity (STS) uses specific test collections as its performance evaluation measurement. The test collections consist of text pairs with the same meaning even though in different text form. The existence is scarce compared with information retrieval (IR) test collections. This paper investigates the possibility to reuse IR test collections for STS tasks. Text pairs are derived from the relevant pair of IR test collections. Latent semantic analysis (LSA) and explicit semantic analysis (ESA) evaluate Glasgow's test collections, which are provided by ACM SIGIR community. Jaccard index measures the lexical similarity. Recall metric measures retrievability of recycling test collection with two existing test collections, Microsoft research paraphrase corpus and Microsoft research video description corpus, as evaluation baselines. Evaluation yields a promising outcome; the evaluated test collections have low Jaccard index and their recall values between the two baselines.

In this paper, we propose a system which allows users to attach annotations to access controls in wikis. When managing documents in wikis, they need to have access controls for preventing the documents from being accessed inappropriately. The access controls are usually modified dynamically according to changes of needs on wiki documents. It is generally difficult to keep access controls consistent beyond the modifications. Annotations are introduced in this paper for this problem. We focus on a situation where a set of access controls that was supposed to be stable is recovered using backups, discuss contradictions between access controls potentially occurred in the situation, and propose what kind of annotations will help solve the contradictions.

It is desired to have a system that can recommend news articles according to interests of a user, which would change with time. In this paper, we attempt to improve a news recommendation system with supervised classification by integrating a clustering method into the system in order to adapt flexibly to the variety of interests of a user. To follow the changes of interests with time, we need to make not only the classifier but also the clustering module of the system be easily updatable. We construct clusters in the feature space from combining one-dimensional clusters to make the size of storage to hold for updating clusters be constant. We let the data distribution in each one-dimensional space be influenced by the clustering results from another one-dimensional space, thereby taking into account of data distribution in the original multiple dimensional space. Main contribution of this paper is to propose a method that can achieve both of the two goals: to improve the performance of a news recommendation system and to make the amount of data to hold for updating clusters be constant. Some experimental results are shown and the effective and weak points of our proposal are discussed.

Quick recovery from a failure is required essentially for distributed stream processing systems. We focus on single-node fail-stop failures occurred in high availability stream processing systems in this paper. One of high availability mechanisms is to provide a backup node for a processing node in the systems. We propose exploitation of backup nodes for reducing recovery cost in such an environment. We report some simulation results to show the effectiveness of our proposal.

In this paper, we propose an index allocation method for data access over multiple wireless channels. Our method first derives external index information from the scheduled data, and then allocates it over upper channels. Moreover, local exponential indexes with different parameters are built within each channel for local data search. Experiments are performed to compare the effectiveness of our approach with an existing approach. The results show that our method outperforms the existing method.