No abstract

No abstract

No abstract

In this article we examine automated language‐independent authorship verification using text examples in several representative Indo‐European languages, in cases when the examined texts belong to an open set of authors, that is, the author is unknown. We showcase the set of developed language‐dependent and language‐independent features, the model of training examples, consisting of pairs of equal features for known and unknown texts, and the appropriate method of authorship verification. An authorship verification accuracy greater than 90% was accomplished via the application of stylometric methods on four different languages (English, Greek, Spanish, and Dutch, while the verification for Dutch is slightly lower). For the multilingual case, the highest authorship verification accuracy using basic machine‐learning methods, over 90%, was achieved by the application of the kNN and SVM‐SMO methods, using the feature selection method SVM‐RFE. The improvement in authorship verification accuracy in multilingual cases, over 94%, was accomplished via ensemble learning methods, with the MultiboostAB method being a bit more accurate, but Random Forest is generally more appropriate.

Abstract:Machine learning methods used for decision support must achieve (a) high accuracy of decisions they recommend, and (b) deep understanding of decisions, so decision makers could trust them. Methods for learning implicit, non-symbolic knowledge provide better predictive accuracy. Methods for learning explicit, symbolic knowledge produce more comprehensible models. Hybrid machine learning models combine strengths of both knowledge representation model types. In this paper we compare predictive accuracy and comprehensibility of explicit, implicit, and hybrid machine learning models for several standard medical diagnostics, electronic commerce, e-marketing and financial decision making problems. Their applicability in different environments -desktop, mobile and cloud computing is briefly analyzed. Machine learning methods from Weka and R/Revolution environments are used.