1. Pravičnost gručenja k-meansPatrik Praprotnik, 2025, master's thesis Abstract: Naloga obravnava problem pravičnosti v algoritmu gručenja k-means in njegovih prilagojenih različicah. Raziskane so različne strategije za zmanjševanje vpliva občutljivih spremenljivk ter uravnoteženje skupin v gručenju. Eksperimenti na sintetičnih in realnih podatkovnih množicah prikazujejo vpliv predlaganih pristopov na kakovost gručenja in pravičnost rezultatov. Rezultati kažejo, da je iskanje kompromisa med pravičnostjo in kakovostjo gručenja precejšnji izziv.
Keywords: strojno učenje, pravičnost, gručenje, občutljive spremenljivke, k-means
Published in DKUM: 22.12.2025; Views: 0; Downloads: 5
 Full text (5,22 MB) |
2. Pristop za izboljšanje pravičnosti v strojnem učenju z uporabo arhitekture učitelj-študent in učenjem z učnim načrtomPeter Trdin, 2025, master's thesis Abstract: V magistrskem delu raziskujemo, kako lahko arhitektura učitelj-študent v kombinaciji z učenjem po učnem načrtu pripomore k zmanjševanju nepravičnosti v modelih strojnega učenja. Razvili smo več pristopov za prenos znanja, pri čemer smo uporabili obteževanje po senzitivnih skupinah in strukturirano inkrementalno učenje. Rezultati eksperimentov na zbirki Adult Income kažejo, da ti pristopi pomembno izboljšajo pravičnost napovedi, čeprav pogosto z rahlim zmanjšanjem točnosti. Magistrsko delo prispeva k razumevanju vpliva strukturiranega učenja in prenosa znanja na pravičnost napovednih modelov.
Keywords: strojno učenje, pravičnost, arhitektura učitelj-študent, učenje z učnim načrtom, destilacija znanja
Published in DKUM: 22.12.2025; Views: 0; Downloads: 4
Full text (2,47 MB) |
3. Empirična primerjava znanstvenih objav s semantičnimi vektorjiBojan Jan Javornik, 2025, master's thesis Abstract: V magistrski nalogi obravnavamo možnost uporabe semantičnih vektorskih reprezentacij kot orodja za primerjavo vsebine znanstvenih objav in spremljanje sprememb v raziskovalnih tematikah skozi čas. Menimo, da lahko z vektorizacijo povzetkov in merami kvantitativne razdalje zanesljivo sledimo podobnostim in razlikam v znanstveni produkciji. Posebej smo vključili tudi obdobje COVID-19, saj smo predvidevali, da bodo takratne vsebinske spremembe dober preizkus učinkovitosti metode.
V raziskavi smo iz baze Scopus pridobili objave izbranih znanstvenih revij za obdobje 2017–2022, povzetke pretvorili v vektorje s prednaučenim SBERT modelom “all-MiniLM-L6-v2” ter zgradili predstavitve revij po letih. Za merjenje razdalj med vektorji smo uporabili kosinusno razdaljo, hipoteze pa preverili z Wilcoxonovim testom ter podprli rezultate z vizualizacijami in modeliranjem tematik z BERTopic.
Rezultati potrjujejo, da obstajajo izrazite razlike med revijami, saj se vektorski centri različnih revij v prostoru jasno ločijo, medtem ko so objave znotraj posamezne revije bolj povezane. Prav tako smo zaznali tematske premike med zaporednimi leti iste revije, kar kaže na postopno, a zaznavno evolucijo raziskovalnih vsebin. Analiza obdobja pred in po letu 2020 je pokazala večje spremembe, vendar teh razlik ni mogoče neposredno pripisati COVID-19. Raziskava tako potrjuje, da je semantična vektorizacija skupaj s kosinusno razdaljo učinkovito orodje za empirično primerjavo vsebine znanstvenih objav.
Keywords: semantična vektorizacija, znanstvene objave, podobnost vsebine, SBERT, kosinusna razdalja
Published in DKUM: 22.12.2025; Views: 0; Downloads: 2
Full text (39,90 MB) |
4. Razlaga odločitev umetne inteligence v obliki pogovorov z velikimi jezikovnimi modeliJan Sernec, 2025, master's thesis Abstract: Razlaga odločitev, sprejetih s pomočjo umetne inteligence, postaja vse pomembnejša, saj umetna inteligenca danes aktivno sodeluje v sistemih, ki se uporabljajo v kritičnih področjih, kot so medicina, energetika, bančništvo. Take odločitve imajo pomemben vpliv na naše življenje, zato je ključnega pomena, da so uporabnikom razumljive in transparentne. Čeprav so statične razlage kot na primer, SHAP in LIME, v praksi še vedno najpogosteje uporabljene, se pogosto izkaže, da ne naslovijo vseh vprašanj in dvomov uporabnikov. Zaradi tega postajajo vse bolj relevantne interaktivne razlage, ki omogočajo dialog z velikim jezikovnim modelom in s tem prilagojeno, pojasnjevalno izkušnjo. V okviru magistrskega dela smo izvedli eksperiment, v katerem smo primerjali vpliv statičnih in interaktivnih razlag na uporabniško izkušnjo. Za ta namen smo razvili spletno orodje, ki omogoča predstavitev obeh tipov razlag. Uporabniško razumevanje odločitev smo merili s pomočjo vedenjskih nalog, oblikovanih na podlagi metod za merjenje mentalnih modelov. Rezultati eksperimenta nudijo pomemben vpogled v to, kako uporabniki zaznavajo in vrednotijo različne oblike razlag ter razkrivajo njihove preference pri razumevanju odločitev umetne inteligence.
Keywords: veliki jezikovni modeli, interaktivne razlage, statične razlage, razlaga odločitev, razložljiva umetna inteligenca
Published in DKUM: 22.12.2025; Views: 0; Downloads: 15
Full text (2,10 MB) |
5. A personalized approach to understanding food cravings and intake : a study protocolSaša Zorjan, Sašo Karakatič, Marina Horvat, Satja Mulej Bratec, Živa Krajnc, 2025, original scientific article Abstract: Background: Studies on food craving and consumption often overlook the interconnectedness of risk factors, assuming uniform mechanisms that drive individuals to (over)consume food. This project seeks to address this gap by leveraging a precision health framework to explore whether multimodal clustering can predict weight and eating outcomes after six months, providing a more nuanced understanding of individual variability. Methods: The project will include a longitudinal study, encompassing several sub-studies where self-report, electrophysiological, and time series dynamic data will be collected at three time points. At baseline, participants will complete comprehensive assessments, including an electroencephalography (EEG) experiment and a one-week experience sampling study (ESM). Machine learning techniques will be employed to uncover distinct participant clusters, characterized by unique patterns of food consumption and weight changes over six months. Markers that best differentiate these profiles will be identified with explainable AI techniques, which aim to make machine learning model outputs understandable by highlighting the key features or patterns driving predictions, enabling personalized insights into key factors contributing to eating behaviors and weight management. Discussion: By exploring the variability of mechanisms influencing food consumption, eating regulation, and weight gain, we aim to uncover subgroups of individuals who are most affected by specific influences, such as stress, emotion regulation difficulties, or sleep deprivation. This project will advance theoretical understanding by integrating multimodal data and emphasizing idiographic methods to capture individual variability. Findings will provide a foundation for future research on precision approaches to eating behaviors and may offer insights into personalized strategies for prevention and management of both normative and disordered eating patterns.
Keywords: food cue reactivity, EEG, experienxe sampling methodology, personalized medicine, achine learning, explainable artificial inteligence
Published in DKUM: 19.12.2025; Views: 0; Downloads: 0
Full text (1,60 MB)
This document has many files! More... |
6. Detection of malicious software using large language models : master's degree thesisMartina Tivadar, 2025, master's thesis Abstract: This thesis examines the success rate of large language models (LLM) in detecting macOS malware through Endpoint Security logs. A literature review and 144 experiments with three ChatGPT variants and six prompt types evaluated accuracy, precision, recall, specificity, and F1-score. Results show that prompt wording is crucial: zero-shot and chain-of-thought prompts performed best, while conservative prompts minimized false positives but missed threats. GPT-4o and o1 outperformed o4-mini but showed similar results. Findings suggest LLMs can support, but not replace, traditional detection, with prompt design proving as important as model choice.
Keywords: malware, large language models, detection
Published in DKUM: 03.11.2025; Views: 0; Downloads: 17
Full text (29,72 MB) |
7. Automated speech analysis in depressive disorder: enhancing diagnosis and monitoring : magistrsko deloAljaž Neuberg, 2025, master's thesis Abstract: This paper investigates the automatic recognition of depression by integrating acoustic, linguistic and emotional features extracted from clinical interviews in the DAIC-WOZ dataset. A total of six classical machine learning classifiers such as Decision Tree, Random Forest, SVM, Gradient Boosting, AdaBoost and XGBoost were systematically evaluated under different class balancing methods (such as SMOTE, SMOTETomek and Random Undersampling) and feature selection strategies. The best model, a decision tree classifier with SMOTE-based balancing and a feature selection technique, achieved a weighted F1 score and accuracy of 0.78 with only eight selected features. These features included all three modalities, demonstrating the
added benefit of a multimodal approach. The results suggest that even relatively simple models, when supported by careful preprocessing and dimensionality reduction, can provide accurate and interpretable predictions. This work emphasizes the importance of feature engineering and balancing techniques in clinical machine learning tasks and lays the foundation for future research on scalable and explainable depression detection systems.
Keywords: Depression, Classification, Machine Learning, Data Balancing, Feature Selection
Published in DKUM: 03.11.2025; Views: 0; Downloads: 12
Full text (4,28 MB) |
8. Influences on and prevention of self-harm behavior among the most at-risk adolescents : study protocol for the SH-MARA prospective longitudinal cohort studyLana Sernec Podnar, Petra Tomažič, Anja Tomašević Kramer, Barbara Plemeniti Tololeski, Gorjan Tasevski, Žiga Rosenstein, Simona Klemenčič, Tadej Battelino, Blaž Vrhovšek, Tadej Lahovnik, Jernej Kovač, Carla Sharp, Barbara Jenko Bizjan, Sašo Karakatič, Maja Drobnič Radobuljac, 2025, original scientific article Abstract: Background Both suicidal and non-suicidal self-injuring behaviors (NSSI) are common during adolescence In Slovenia, adolescent suicide rates are high, making suicide the leading cause of death in the year 2022 in this age group. These behaviors are influenced by a complex interplay of environmental, psychological, and genetic factors. Previous research has identified risk and protective factors mainly for suicidal behavior in adults, a notable gap in understanding these factors in adolescents remains, especially for NSSI. Notably there is an important lack of effective clinical tools or psychometric assessment methods to reliably assess the risk for either suicidal or NSSI behaviors in acutely hospitalized adolescents. Methods and analysis The proposed study uses a mixed-method observational design consisting of a prospective longitudinal cohort component involving adolescents hospitalized for high risk of DSH, and a cross-sectional comparison with a control group of healthy adolescents recruited from primary care settings. It is aimed at identifying genetic, psychosocial, and clinical factors associated with suicidal behaviors and NSSI in adolescents. The study group is recruited from adolescents aged 12–19, admitted to the Intensive Child and Adolescent Psychiatry Unit in Ljubljana due to severe self-harm risk. Exclusion criteria include involuntary treatment, acute psychotic disorders, intellectual disability, severe physical or central nervous system illnesses and acute intoxication. The control group comprises adolescents of comparable age, recruited through regular scheduled health check-ups in Slovenia. Exclusion criteria include suicidality, severe mental disorder, a history of self-harm behavior in a first-degree relative, intellectual disability, severe physical or central nervous system illnesses and acute intoxication. Enrollment runs from February 1, 2023, to December 31, 2025. Participation is voluntary, requiring parental or guardian consent for those 14 or younger
Keywords: adolescents, deliberate self-harm, non-suicidal self-injury, suicidal behavior, intensive psychiatry, personality disorder, traumatic experience, genetics, epigenetics
Published in DKUM: 17.10.2025; Views: 0; Downloads: 6
Full text (1,10 MB) |
9. |
10. Impact of developer queries on the effectiveness of conversational large language models in programmingViktor Taneski, Sašo Karakatič, Patrik Rek, Gregor Jošt, 2025, original scientific article Abstract: This study investigates the effects of LLM-based coding assistance on web application development by students using a frontend framework. Rather than comparing different models, it focuses on how students interact with LLM tools to isolate the impact of query type on coding success. To this end, participants were instructed to rely exclusively on LLMs for writing code, based on a given set of specifications, and their queries were categorized into seven types: Error Fixing (EF), Feature Implementation (FI), Code Optimization (CO), Code Understanding (CU), Best Practices (BP), Documentation (DOC), and Concept Clarification (CC). The results reveal that students who queried LLMs for error fixing (EF) were statistically more likely to have runnable code, regardless of prior knowledge. Additionally, students seeking code understanding (CU) and error fixing performed better, even when normalizing for previous coding ability. These findings suggest that the nature of the queries made to LLMs influences the success of programming tasks and provides insights into how AI tools can assist learning in software development.
Keywords: large language models, LLMs, prompt engineering, query type analysis, AI-assisted programming, educational software development
Published in DKUM: 23.06.2025; Views: 0; Downloads: 12
Full text (994,41 KB)
This document has many files! More... |
