TSI logo 25 years
LV EN
Home > About Us > Structure and Government > Abstracts

DEGREE

PROGRAMME

FACULTY

YEAR

LANGUAGE

KEYWORDS

Improvement of machine leaning algorithms performance by data set dimensionality reduction using cellular automata

A significant challenge in Machine Learning is dealing with high-dimensional data. Complexity knowns as the "curse of dimensionality" results in deterioration оf Machine Learning algorithms performance as the dimensionality and dataset size increases. Cellular automata are a dynamical discrete computational system with mathematical functions knows as rules that result in complex global behaviour. We used one-dimensional elementary cellular automata as a tool for dataset size. Model variables were selected for initial status vector generation and its further transformation to format that is suitable for cellular automata rules application known in cellular automata theory as configuration. Then model iterated through all possible cellular automata rules and various epochs variations were applied. Model performance for reduced dataset was compared with benchmark results of original dataset after standard dimensionality reduction technics used. It was concluded that applied cellular automata rules can be used as alternative methods for dataset size reduction without deteriorating model performance.

Author: Alexey Kuchvalskiy

Supervisor: Dmitry Pavlyuk

Degree: Master

Year: 2024

Work Language: English

Study programme: Computer Sciences

More...

Unsupervised machine learning approach for hierarchical graph-based representation of natural language text collections.

Managing big data efficiently is important in various fields, much so when data consists of human-written documents. Recent advances in Natural Language Processing (NLP), particularly LLMs, allowed to solve many task in this domain, despite the high demand for labelled data, compute resources and specialized skills.To tackle these limitations, current study proposed a NLP pipeline to identify topic hierarchies in collections of scientific publications. The work focused on evaluation of available unsupervised machine learning methods and quality metrics in NLP, and development of visualization techniques to build a prototype of the pipeline.Proposed solution is based on the hARTM approach optimized for interpretability. It demonstrated the capacity to infer human-interpretable topic hierarchies from collections of scientific texts and construct meaningful hierarchy of topic-based document representations. The visualization approaches rely on MDS to present inter-document similarity and Sankey plots to show document cluster relatedness within topic hierarchy.Utility was demonstrated on two datasets, focusing on interpretability and meaning of the topic hierarchy and associated topic definitions. Potential application areas include personal education and scientific writing.

Author: Jevgenijs Bodrenko

Supervisor: Irina Jackiva

Degree: Master

Year: 2024

Work Language: English

Study programme: Computer Sciences

More...

Application of machine learning in decision support system

The aim of the work is to improve the accuracy of predicting wait times in an existing queue management system using machine learning. Client-provided data was analyzed, and models were trained using various machine learning algorithms. Performance measures of the models were collected, and the best one was selected. Additionally, software and a database were developed to manage the training process and evaluate the quality of the models. The quality of the software was assessed using industry-standard methodologies and tested.

Author: Jevgēnijs Nikolajevs

Supervisor: Jeļena Kijonoka

Degree: Bachelor

Year: 2024

Work Language: Latvian

Study programme: Computer Science

More...

Boosting Algorithms for Credit Card Fraud Detection Across Varied Datasets

Manual reviews and rule-based systems, as well as data mining techniques such as clustering and classification algorithms, are crucial for identifying credit card fraud since they help identify fraudulent transactions. Despite obstacles in gathering training data, more data has lately been available, however, a complete comparison of current machine learning approaches has yet to be conducted. Algorithms like XGBoost, AdaBoost, and Gradient Boosting Machine frequently outperform older approaches. This study compares boosting algorithms to traditional approaches using three different credit card transaction datasets: synthetic, balanced with 50% fraudulent transactions, and very unbalanced with only 0.17% fraudulent transactions. The genuine transaction datasets contained 28 anonymized parameters such as time and location. Each method was evaluated using the F1 score, accuracy, precision, and recall. This study makes recommendations on which algorithms to use in real-world scenarios, giving important insights for future research and practical use in credit card fraud detection.

Author: Justs Vīdušs

Supervisor: Nadežda Spiridovska

Degree: Master

Year: 2024

Work Language: English

Study programme: Computer Sciences

More...

Enhancement Strategies for Retrieval-Augmented Generation Systems

This thesis systematically explores the enhancement of Retrieval-Augmented Generation (RAG) systems within Large Language Models, emphasizing optimization of retrieval parameters and generation accuracy. We investigate optimal configurations in RAG systems, including chunk size and overlap percentages, top-k selection, query transformations, different retrieval methods, different LLMs, namely GPT-3.5-Turbo and GPT-4, discovering that a chunk size of 500 tokens generally offers the best performance. Vector search using cosine similarity emerges as the most effective retrieval method, significantly enhancing both context precision and recall across various tasks and knowledge bases. Experimentation within the CRUD-RAG framework demonstrates its applicability in diverse tasks from content creation to knowledge refinement. Our findings indicate that enhancements in retrieval settings can markedly improve the performance of RAG systems, making them more efficient and adaptable for complex information synthesis and retrieval tasks. These results affirm the potential of systematic enhancements to improve AI-driven language models in practical applications, contributing significant insights and practical approaches to the evolving landscape of RAG system research.

Author: Sigita Lapiņa

Supervisor: Dmitry Pavlyuk

Degree: Master

Year: 2024

Work Language: English

Study programme: Computer Sciences

More...

Table View
Text View