20.500.12556/ReVIS-14068 Uporaba umetne inteligence za detekcijo neželenih SMS sporočil: analiza metod in eksperimentalna validacija magistrsko delo študijskega programa druge bolonjske stopnje Spletna znanost in tehnologije V magistrskem delu je raziskana uporaba metod umetne inteligence za detekcijo neželenih SMS sporočil, pri čemer so primerjane tradicionalne tehnike, algoritmi strojnega učenja in modeli globokega učenja. V teoretičnem delu je opravljen sistematičen pregled literature, ki vključuje metode na osnovi pravil, algoritme strojnega in globokega učenja ter hibridne modele. Empirični del je temeljil na UCI SMS Spam Collection zbirki, ki je bila predhodno očiščena, normalizirana, tokenizirana in podvržena vektorizaciji oziroma vdelavi besed. Izvedeni so bili poskusi z metodami Multinomial Naive Bayes, Support Vector Machine, Random Forest, Convolutional Neural Network in Bidirectional LSTM, pri čemer sta MNB ter SVM dosegla najvišji F1-score, algoritma globokega učenja pa sta pokazala konkurenčno natančnost. Tradicionalne metode so bile najhitrejše, globoki modeli so zahtevali največje računske vire, pri čemer je BiLSTM izkazal najboljšo stabilnost rezultatov med modeli globokega učenja. Glavne ugotovitve iz literature kažejo, da globoki modeli prinašajo višjo povprečno natančnost, medtem ko ostajajo tradicionalni algoritmi strojnega učenja zelo učinkoviti. Zaradi specifičnih lastnosti SMS sporočil in širokega nabora pregledane literature nismo povsem uspeli ponoviti ugotovitev iz literature. Čeprav so naši rezultati pokazali, da so modeli strojnega učenja nekoliko bolj učinkoviti, smo lahko potrdili visoko učinkovitost tako modelov strojnega kot tudi globokega učenja. In the master thesis we explored the use of artificial intelligence methods for detecting unwanted SMS messages, comparing traditional techniques, machine learning algorithms, and deep learning models. The theoretical part provides a systematic literature review, covering rule based methods, machine learning algorithms, deep learning approaches. The empirical part was based on the UCI SMS Spam Collection, which was preprocessed through cleaning, normalization, tokenization, and vectorization or word embedding. Experiments were conducted with Multinomial Naive Bayes, Support Vector Machine, Random Forest, Convolutional Neural Network, and BiLSTM methods. MNB and SVM achieved the highest F1-scores, while the deep learning algorithms demonstrated competitive accuracy. Machine learning modles proved to be the fastest, while deep learning models required the most computational resources, with BiLSTM showing the most stable results among the deep models. The main findings from the literature indicate that deep models generally achieve higher average accuracy, while machine learning algorithms remain highly effective. Due to the specific characteristics of SMS messages and the wide range of reviewed literature, we were not able to reproduce the findings from previous studies. Although our results showed that machine learning models were a little more effective, we were able to confirm the high efficiency of both machine learning and deep learning approaches. umetna inteligenca SMS detekcija strojno učenje globoko učenje artificial intelligence SMS detection machine learning deep learning true false false Ž. Ikovic Slovenski jezik Angleški jezik Magistrsko delo/naloga 2026-06-19 11:26:40 2026-06-19 11:26:40 2026-06-20 03:10:57 0000-00-00 00:00:00 2025 0 Maribor 2025 Maribor 74 str., [40] f. pril. 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.8:621.395(043.5) 282234115 Ikovic_Ziga_md_2025.pdf Ikovic_Ziga_md_2025.pdf 1 32C195A43B73C6DD120875055EB3623D 99513e4a76c38d40ea8a79775b0b46b937b4f62057b8fc69ebc7015b78776b63 3bd02e90-6bb0-11f1-be3d-001a4af901a5 https://revis.openscience.si/Dokument.php?lang=slv&id=17055 Univerza Alma Mater Europaea 0 0 0