This paper aims to examine the impact of seasonality on a monthly basis in the context of forecasting PM10 levels using machine learning methods based on weather conditions. Predicting air pollution concentrations is crucial for public health. In this study, these pollutants are analyzed using predictive modeling. The research hypothesis of this paper is that seasonality influences the prediction accuracy of machine learning models forecasting air pollution levels. The study explores selected applications of machine learning methods for predicting air pollution levels and analyzes seasonal variations in pollutant concentrations. Data for this analysis were obtained from measurement stations in a highly urbanized region of Central Europe. The study demonstrated that, when forecasting air pollutant concentrations based on atmospheric conditions, the models achieved higher R² values in winter than in summer, yet exhibited greater prediction errors during colder months compared to warmer periods. Further analysis requires the inclusion of additional environmental factors to better reflect their influence on the accuracy of predictions. The results underscore the importance of seasonality in forecasting air pollution based on weather conditions, demonstrate its impact on the performance of machine learning models, and suggest directions for future research aimed at improving predictive accuracy and model generalization.

In this paper, we address the problem of classifying medical cases into appropriate operating rooms. Each case is provided in textual form and must be assigned to one of two distinct operating rooms. To solve this problem, we begin with anonymization and preprocessing of the raw data. We then apply three different natural language processing (NLP) pipelines to transform the data into a structured tabular format suitable for classification. Common classification methods, including both shallow and deep learning approaches, are evaluated, and the results are presented using well-established classification metrics such as accuracy, precision, and recall. Numerical experiments conducted on real-world medical cases described in Polish demonstrate the effectiveness of the proposed method, which can serve as a decision support tool for surgical scheduling and operating room assignment.

Recently, sports performance forecasting has become a topic of great interest among both researchers and practitioners. The goal of such forecasts is not only to better understand the meanderings and strategies of sports competitions, but also to be able to use these forecasts for commercial purposes, such as sports betting. These forecasts are based on the analysis of historical data and characteristics that affect the outcome of the games. One important aspect of this process is the selection of features that are used to create predictive models. The selection of appropriate features has a significant impact on the quality of the predictions, as well as on the comprehensibility and simplicity of the model. In this article, we present a dimensionality reduction approach to create a reduced set of features for effective classification of sports performance. The article presents the results of sports performance prediction experiments using the English Premier League and German Bundesliga as examples. The aim of the study was to create a reduced set of features that would allow equally effective classification using a heterogeneous set of classifiers compared to a full list of features. Experimental results indicate that it is possible to reduce the feature list without significant loss of classification quality. In the case of the German Bundesliga, the best results were obtained with the ’Sequential-back’ approach, which consists of fea-tures:’PositionHT’, ’MatchesHT’, ’WinsHT’, ’GoalDiferenceHT’, ’WinsVT’, ’PointsVT’ i ’Diference’. For the English Premier League, the ’Importance’ approach, which includes features, performed best: ’PositionHT’, ’WinsHT’, ’DrawsHT’, ’LossesHT’, ’PositionVT’, ’DrawsVT’ i ’LossesHT’. In both cases, the reduced set of features allowed for comparable results with the full set of features. The results have potential applications in predicting the performance of less popular football leagues, where access to statistics is limited. The shortened list of features can also reduce the time needed for model development and prediction.

Sports events attract a lot of attention and are an exciting playground for testing various machine learning methods. However, the non-consistency of data, the necessity of including a chronology of events, and correlation among the data lead to a situation where complex algorithms are not in favor of deriving good classification results. This article focuses on extending our previous works and deriving new voting schemas for the ensemble of classifiers. The ensemble is built based on well-known algorithms from literature and can be considered a heterogeneous approach. Especially, we put a particular focus on improving the poor results achieved for selected decision classes. To do so, we first identify the weak spots for existing methods, extend our approach to the new voting schemas, and eventually improve the classification results. However, in the presented approach, we observe the problem of the trade-of between improving some results while decreasing the overall good results of the remaining decision classes. The whole experiments are performed on the real-world data obtained from websites, preprocessed, and transformed to ft into the format used by the popular classification methods. Obtained results indicate that the analyzed problem is not trivial, and improvements for some results lead to decreasing quality of some others.

We present a novel approach based on the original idea of the Ant Colony Decision Tree (ACDT) algorithm used in the problem of building the decision trees. One of the crucial limitations of the canonical ACDT algorithm was its link to strict decision rules. In this paper we transform the algorithm in such way, that it is capable to manage complex association rules. Research is conducted on the various sets of financial data closely related with the swiss frank currency. Evaluation of results was possible on the basis of accuracy measure as well as the proposed fuzzy accuracy. These preliminary studies show, that the proposed algorithm is capable to maintain its effectiveness even in the problems with large number of attribute values.