Forecasting air pollution is crucial for understanding the phenomenological and contextual variety of mechanisms underlying pollution in a particular area or region. Analyzing high-dimensional data with spatial and temporal dependencies is pointed out as a major challenge for traditional machine learning approaches for air pollution forecasting. The unprecedented advances in deep learning employed on massive quantities of IoT sensor data raise high hopes for the future of air pollution forecasting. Drawing on past and current experiences and solutions advanced in a growing body of research on the topic, we proposed and evaluated four encoder-decoder architectures with attention for forecasting particulate matter (PM) levels, with a general applicability that is both location- and season-independent. The attention mechanism was employed to better learn representation of time series data. By selectively focusing the attention on elements that contribute the most, superior performance gains have been obtained by the proposed models. Sensor measurements obtained in the past 7 or 14 days as well a set of external factors known to affect the pollutant levels, such as weather conditions, location and time frame (e.g., season, time of day) were taken into account. In this research, we carried out a case study on PM 2.5 forecasting for the city of Skopje, allowing us to discuss the relevance of the results obtained by the proposed solutions. The ability to draw valid inferences from data has critical importance when forecasting air pollution. In this respect, any air pollution forecasting model based on deep learning would require a component for generation of realistic data samples to augment the time series dataset. We address the challenges in missing data by proposing and evaluating two adversarial networks for data augmentation. We have conducted a number of experiments to investigate the performance of the predictive models with and without augmenting the training datasets and obtain superior performance gains by the proposed adversarial models for data augmentation. The deep neural architectures advanced in this research are general enough to be used for predictive and generative tasks for other pollutants; moreover, they could be adopted for related tasks handling time series data in other domains.
The rapid increase in scientific publications has made it challenging to keep up with the latest advancements. Conducting systematic reviews using traditional methods is both time-consuming and difficult. To address this, new review formats like rapid and scoping reviews have been introduced, reflecting an urgent need for efficient information retrieval. This challenge extends beyond academia to many organizations where numerous documents must be reviewed in relation to specific user queries. This paper focuses on improving document ranking to enhance the retrieval of relevant articles, thereby reducing the time and effort required by researchers. By applying a range of natural language processing (NLP) techniques, including rule-based matching, statistical text analysis, word embeddings, and transformer- and LLM-based approaches like Mistral LLM, we assess the article’s similarities to user-specific inputs and prioritize them according to relevance. We propose a novel methodology, Weighted Semantic Matching (WSM) + MiniLM, combining the strengths of the different methodologies. For validation, we employ global metrics such as precision at K, recall at K, average rank, median rank, and pairwise comparison metrics, including higher rank count, average rank difference, and median rank difference. Our proposed algorithm achieves optimal performance, with an average recall at 1000 of 95% and an average median rank of 185 for selected articles across the five datasets evaluated. These findings give promising results in pinpointing the relevant articles and reducing the manual work.
With the exponential growth of data, extracting actionable insights becomes resource-intensive. In many organizations, normalized relational databases store a significant portion of this data, where tables are interconnected through some relations. This paper explores relational learning, which involves joining and merging database tables, often normalized in the third normal form. The subsequent processing includes extracting features and utilizing them in machine learning (ML) models. In this paper, we experiment with the propositionalization algorithm (i.e., Wordification) for feature engineering. Next, we compare the algorithms PropDRM and PropStar, which are designed explicitly for multi-relational data mining, to traditional machine learning algorithms. Based on the performed experiments, we concluded that Gradient Boost, compared to PropDRM, achieves similar performance (F1 score, accuracy, and AUC) on multiple datasets. PropStar consistently underperformed on some datasets while being comparable to the other algorithms on others. In summary, the propositionalization algorithm for feature extraction makes it feasible to apply traditional ML algorithms for relational learning directly. In contrast, approaches tailored specifically for relational learning still face challenges in scalability, interpretability, and efficiency. These findings have a practical impact that can help speed up the adoption of machine learning in business contexts where data is stored in relational format without requiring domain-specific feature extraction.