This paper addresses the regression modeling of local environmental pollution levels for electric power industry needs, which is fundamental for the proper design and maintenance of high-voltage transmission lines and insulators in order to prevent various hazards, such as accidental flashovers due to pollution and the resultant power outages. The primary goal of our study was to increase the precision of regression models for this application area by exploiting additional input attributes extracted from satellite imagery and adjusting the modeling methodology. Given that thousands of different attributes can be extracted from satellite images, of which only a few are likely to contain useful information, we also explored suitable feature selection procedures. We show that a suitable combination of attribute selection methods (relief, FSRF-Test, and forward selection), regression models (random forest models and M5P regression trees), and modeling methodology (estimating field-measured values of target variables rather than their upper bounds) can significantly increase the total modeling accuracy, measured by the correlation between the estimated and the true values of target variables. Specifically, the accuracies of our regression models dramatically rose from 0.12–0.23 to 0.40–0.64, while their relative absolute errors were conversely reduced (e.g., from 1.04 to 0.764 for the best model).

In this paper we perform density-based clustering of discussion participants from online editions of two major Slovak national newspapers, Sme.sk and Cas.sk. We use language-independent statistical attributes characterising their communication patterns and the content of their posts. In each newspaper, we separately analyse two categories of news (domestic and international). A large majority of participants in each dataset was found to belong to one stable and dominant cluster present in all our datasets. We interpret it as comprising the “standard” or “average” discussion participants. The remaining participants could be viewed as various kinds of “anomalies” or “departures from normal” (not necessarily negative) and were assigned to several minor clusters. The shape and position of some minor clusters generalized well across the datasets. Overall, we found significant structural similarities between the four datasets in terms of histograms of attributes, the existence of one stable and dominant cluster, and the similar shape and location of several minor clusters. This is a significant result given that the four datasets were largely independent and the two newspapers adopted radically different policies for dealing with karma and foul language. The proposed approach therefore looks very promising as a means of identifying anomalous behaviour on diverse online discussion platforms.

This article helps establish reliable baselines for document-level sentiment analysis in highly inflected languages like Czech and Slovak. We revisit an earlier study representing the first comprehensive formulation of such baselines in Czech and show that some of its reported results need to be significantly revised. More specifically, we show that its online product review dataset contained more than 18% of non-trivial duplicates, which incorrectly inflated its macro F1-measure results by more than 19 percentage points. We also establish that part-of-speech-related features have no damaging effect on machine learning algorithms (contrary to the claim made in the study) and rehabilitate the Chi-squared metric for feature selection as being on par with the best performing metrics such as Information Gain. We demonstrate that in feature selection experiments with Information Gain and Chi-squared metrics, the top 10% of ranked unigram and bigram features suffice for the best results regarding online product and movie reviews, while the top 5% of ranked unigram and bigram features are optimal for the Facebook dataset. Finally, we reiterate an important but often ignored warning by George Forman and Martin Scholz that different possible ways of averaging the F1-measure in cross-validation studies of highly unbalanced datasets can lead to results differing by more than 10 percentage points. This can invalidate the comparisons of F1-measure results across different studies if incompatible ways of averaging F1 are used.

We propose a text classification method for the purpose of creating a language model for automatic recognition of spontaneous spoken speech. Transcripts from our departmental speech database served as spontaneous spoken texts. Using supervised machine learning methods, we have created multiple classification models (including neural networks), that were able to distinguish them from written texts with high accuracy. We subsequently verified the accuracy of our trained models on a database of texts containing direct speech extracted from newspaper articles.