1. Analiza uporabe umetne inteligence v slovenskih medijihŽiga Kapun, 2024, undergraduate thesis Abstract: Uporaba umetne inteligence je prisotna že na nešteto področjih, kjer se uporablja za najrazličnejše namene. Najdemo jo lahko v optimizaciji procesov, pregledovanju velikih količin podatkov ter tudi ustvarjanju vsebin. Podobno kot na drugih področjih, se je umetna inteligenca, ali AI, uveljavila tudi na področju medijev. Tako se diplomska naloga posveča prav uporabi AI na področju medijev v Sloveniji. Naloga le-te je predstaviti vrste AI in možnosti, ki jih slednja omogoča medijem, prav tako pa z analizo ankete predstavi dejansko stanje uporabe AI v Sloveniji.
Keywords: Umetna inteligenca - AI, mediji, generativna AI, strojno učenje, globoko učenje
Published in DKUM: 14.11.2024; Views: 0; Downloads: 21
 Full text (2,03 MB) |
2. Učinkovitost avtomatiziranega oblikovanja testnih primerov s pomočjo velikih jezikovnih modelovJovana Murdjeva, 2024, master's thesis Abstract: V magistrskem delu je bila raziskana uporabo ChatGPT-ja kot veliki jezikovni model za avtomatizirano oblikovanje testnih primerov v primerjavi s tradicionalnimi metodami, ki jih uporabljajo strokovnjaki za testiranje programske opreme. Delo se je osredotočilo na vpliv natančno opredeljenih pozivov (inženiring pozivov) na kakovost, pokritost kode in učinkovitost pri odkrivanju napak. Rezultati raziskave kažejo, da ChatGPT z ustrezno oblikovanimi vnosnimi zahtevami dosega primerljivo ali celo boljšo uspešnost kot ročno oblikovani testi, kar pomeni velik potencial za optimizacijo procesov testiranja programske opreme.
Keywords: avtomatizirano testiranje, veliki jezikovni modeli, inženiring pozivov, ChatGPT, kakovost testnih primerov
Published in DKUM: 22.10.2024; Views: 0; Downloads: 14
Full text (1,31 MB) |
3. Strategije prilagajanja izhodov velikih jezikovnih modelovBard Grujič, 2024, master's thesis Abstract: Namen magistrskega dela je sistematično predstaviti, raziskati in analizirati delovanje velikih jezikovnih modelov, s posebnim poudarkom na modelu transformatorja, ter raziskati, kako prilagoditi izhode teh modelov za specifične potrebe organizacij. V praksi bomo preučili, kako ta prilagoditev deluje, tako da bomo razvili in demonstrirali aplikacijo za iskanje ključnih besed po dokumentih v PDF formatu z uporabo velikega jezikovnega modela GPT-4 podjetja OpenAI.
Keywords: veliki jezikovni modeli, modeli obdelave naravnega jezika, umetna inteligenca, inženiring spodbud
Published in DKUM: 19.09.2024; Views: 0; Downloads: 14
Full text (2,79 MB) |
4. Primerjava pristopov gručenja z algoritmi po vzorih iz naraveDavid Mikek, 2024, master's thesis Abstract: V tem delu smo se lotili gručenja s petimi različnimi algoritmi po vzoru iz narave. V ta namen smo razvili štiri različne pristope za njihovo uporabo pri reševanju problema gručenja. Njihovo učinkovitost smo preverili z eksperimentom nad šestimi različnimi podatkovnimi seti in na koncu izvedli primerjavo. Ugotovili smo, da lahko algoritmi po vzoru iz narave učinkovito rešujejo problem gručenja, vendar na rezultate in čas izvajanja močno vpliva izbira pristopa in algoritma.
Keywords: Gručenje, Algoritmi po vzoru iz narave
Published in DKUM: 19.09.2024; Views: 0; Downloads: 7
Full text (5,16 MB) |
5. Kako poštena so klasifikacijska odločitvena drevesa?Andrej Kostić, 2024, master's thesis Abstract: Poštenost klasifikacijskih odločitvenih dreves je na področju strojnega učenja postala kritično vprašanje. Klasifikacijska in regresijska drevesa (CART) so znana po svoji preprostosti in učinkovitosti pri obravnavanju klasifikacijskih in regresijskih nalog. Vendar lahko ti modeli nehote ohranijo ali celo povečajo pristranskost, prisotno v podatkih, kar vodi do nepoštenih odločitev, ki nesorazmerno prizadenejo določene skupine. To magistrsko delo raziskuje poštenost modelov CART z implementacijo metode FairCART, ki vključuje omejitve poštenosti med postopkom oblikovanja dreves. V delu je ocenjena učinkovitost metode FairCART pri zmanjševanju pristranskosti ob hkratnem ohranjanju kakovosti odločitev, kar omogoča vpogled v kompromise med poštenostjo in točnostjo. Implementacija in rezultati eksperimenta kažejo, da lahko metoda FairCART zmerno zmanjša pristranskost in ohrani splošno kakovost odločitvenega drevesa.
Keywords: klasifikacijska in regresijska drevesa, poštenost v strojnem učenju, CART, FairCART
Published in DKUM: 19.09.2024; Views: 0; Downloads: 10
Full text (3,47 MB) |
6. Implementacija gručenja k-means z genetskim algoritmomAlen Šaruga, 2024, undergraduate thesis Abstract: K-means algoritem je eden najpopularnejših in najučinkovitejših algoritmov gručenja podatkov. Kljub temu algoritem predstavlja izziv, saj je občutljiv na začetno postavitev centroidov gruč. Zato lahko algoritem stremi k lokalnemu optimumu in ne h globalno optimalni rešitvi. Namen diplomskega dela je implementacija optimiziranega k-means algoritma, manj občutljivega na začetne centroide gruč, z uporabo genetskega algoritma. Delo se osredotoča na postopek gručenja in genetski algoritem. Implementacija je izvedena v programskem jeziku Python s knjižnico NiaPy. Na koncu so predstavljeni rezultati eksperimentov, kjer je izvedena primerjava standardnega in optimiziranega k-means algoritma na različnih podatkovnih množicah.
Keywords: gručenje, k-means, genetski algoritem, centroidi
Published in DKUM: 19.09.2024; Views: 0; Downloads: 11
Full text (1,34 MB) |
7. Commit-level software change intent classification using a pre-trained transformer-based code modelTjaša Heričko, Boštjan Šumak, Sašo Karakatič, 2024, original scientific article Abstract: Software evolution is driven by changes made during software development and maintenance. While source control systems effectively manage these changes at the commit level, the intent behind them are often inadequately documented, making understanding their rationale challenging. Existing commit intent classification approaches, largely reliant on commit messages, only partially capture the underlying intent, predominantly due to the messages’ inadequate content and neglect of the semantic nuances in code changes. This paper presents a novel method for extracting semantic features from commits based on modifications in the source code, where each commit is represented by one or more fine-grained conjoint code changes, e.g., file-level or hunk-level changes. To address the unstructured nature of code, the method leverages a pre-trained transformer-based code model, further trained through task-adaptive pre-training and fine-tuning on the downstream task of intent classification. This fine-tuned task-adapted pre-trained code model is then utilized to embed fine-grained conjoint changes in a commit, which are aggregated into a unified commit-level vector representation. The proposed method was evaluated using two BERT-based code models, i.e., CodeBERT and GraphCodeBERT, and various aggregation techniques on data from open-source Java software projects. The results show that the proposed method can be used to effectively extract commit embeddings as features for commit intent classification and outperform current state-of-the-art methods of code commit representation for intent categorization in terms of software maintenance activities undertaken by commits.
Keywords: software maintenance, code commit, mining software repositories, adaptive pre-training, fine-tuning, semantic code embedding, CodeBERT, GraphCodeBERT, classification, code intelligence
Published in DKUM: 14.08.2024; Views: 87; Downloads: 9
Full text (1,65 MB) |
8. FairBoost: Boosting supervised learning for learning on multiple sensitive featuresIvona Colakovic, Sašo Karakatič, 2023, original scientific article Abstract: The vast majority of machine learning research focuses on improving the correctness of the outcomes (i.e., accuracy, error-rate, and other metrics). However, the negative impact of machine learning outcomes can be substantial if the consequences marginalize certain groups of data, especially if certain groups of people are the ones being discriminated against. Thus, recent papers try to tackle the unfair treatment of certain groups of data (humans), but mostly focus on only one sensitive feature with binary values. In this paper, we propose an ensemble boosting FairBoost that takes into consideration fairness as well as accuracy to mitigate unfairness in classification tasks during the model training process. This method tries to close the gap between proposed approaches and real-world applications, where there is often more than one sensitive feature that contains multiple categories. The proposed approach checks the bias and corrects it through the iteration of building the boosted ensemble. The proposed FairBoost is tested within the experimental setting and compared to similar existing algorithms. The results on different datasets and settings show no significant changes in the overall quality of classification, while the fairness of the outcomes is vastly improved.
Keywords: fairness, boosting, machine learning, supervised learning
Published in DKUM: 11.06.2024; Views: 147; Downloads: 8
Full text (1,70 MB)
This document has many files! More... |
9. Adaptive boosting method for mitigating ethnicity and age group unfairnessIvona Colakovic, Sašo Karakatič, 2024, original scientific article Abstract: Machine learning algorithms make decisions in various fields, thus influencing people’s lives. However, despite their good quality, they can be unfair to certain demographic groups, perpetuating socially induced biases. Therefore, this paper deals with a common unfairness problem, unequal quality of service, that appears in classification when age and ethnicity groups are used. To tackle this issue, we propose an adaptive boosting algorithm that aims to mitigate the existing unfairness in data. The proposed method is based on the AdaBoost algorithm but incorporates fairness in the calculation of the instance’s weight with the goal of making the prediction as good as possible for all ages and ethnicities. The results show that the proposed method increases the fairness of age and ethnicity groups while maintaining good overall quality compared to traditional classification algorithms. The proposed method achieves the best accuracy in almost every sensitive feature group. Based on the extensive analysis of the results, we found that when it comes to ethnicity, interestingly, White people are likely to be incorrectly classified as not being heroin users, whereas other groups are likely to be incorrectly classified as heroin users.
Keywords: fairness, boosting, machine learning, classification
Published in DKUM: 24.05.2024; Views: 283; Downloads: 12
Full text (1,66 MB)
This document has many files! More... |
10. Attachment in close relationships and glycemic outcomes in children with type 1 diabetesSimona Klemenčič, Jasna Klara Lipovšek, Anja Turin, Klemen Dovč, Nataša Bratina, Yael Shmueli-Goetz, Katarina Trebušak Podkrajšek, Barbka Repič-Lampret, Barbara Jenko Bizjan, Sašo Karakatič, Tadej Battelino, Maja Drobnič Radobuljac, 2023, original scientific article Abstract: Background: Our aim was to determine whether child attachment to parents, parent attachment style, and morning cortisol levels were related to diabetes outcomes measured by average glycated hemoglobin (HbA1c), HbA1c variability over 4 years and time in range (TIR) in children with type 1 diabetes (T1D). Research design and methods: 101 children with T1D and one of their parents were assessed at baseline for child attachment (Child Attachment Interview; CAI) and parent attachment (Relationship Structures Questionnaire; ECR-RS). Serum samples were collected for cortisol measurements before the interviews. HbA1c levels were measured during a 4-year follow-up period at regular 3-monthly visits, and data for TIR were exported from blood glucose measuring devices. Multivariate linear regression models were constructed to identify independent predictors of glycemic outcomes. Results: More girls than boys exhibited secure attachment to their mothers. The results of the regression models showed that securely attached girls (CAI) had higher average HbA1c than did insecurely attached girls (B = -0.64, p = 0.03). In boys, the more insecure the parent's attachment style, the worse the child's glycemic outcome: the higher the average Hb1Ac (B = 0.51, p = 0.005), the higher the HbA1c variability (B = 0.017, p = 0.011), and the lower the TIR (B = -8.543, p = 0.002). Conclusions: Attachment in close relationships is associated with glycemic outcomes in children with T1D, and we observed significant differences between sexes. A sex- and attachment-specific approach is recommended when treating children with less favorable glycemic outcomes. Special attention and tailored support should be offered to securely attached girls in transferring responsibility for diabetes care and at least to male children of insecurely attached parents to prevent suboptimal glycemic control. Further studies in larger samples and more daily cortisol measurements may help us better understand the links between stress response, attachment and T1D.
Keywords: attachment, childhood and adolescence, cortisol, diabetes control, time in range
Published in DKUM: 21.05.2024; Views: 152; Downloads: 10
Full text (980,31 KB)
This document has many files! More... |
