Data Science Commons^™

Random Graph Models For Dual Graphs, Anne Friedman

Scripps Senior Theses

This paper aims to better characterize dual graphs derived from state districting maps by developing random graph models that replicate their structural properties. Dual graphs provide a simplified way to represent districting maps, making it computationally feasible to analyze their structure. These representations enable researchers, legislators, and courts to assess district compactness, detect signs of gerrymandering, and generate alternative districting plans. A deeper understanding of the structural patterns of these dual graphs can help researchers choose or design more effective algorithms for redistricting analysis. The random graph models developed in this study serve as testbeds for evaluating algorithmic approaches to …

A Proof Of Np-Completeness For The K-Means Clustering Algorithm, Brooke C. Feinberg

Scripps Senior Theses

The k-means clustering algorithm is one of the most widely used clustering techniques in data analysis and machine learning, yet its exact computational complexity remains subject to ongoing theoretical investiga- tion. This work establishes the NP-completeness of k-means by proving (1) it is NP-hard and (2) it lies in NP. To demonstrate NP-hardness, we construct a series of polynomial-time reductions from well-known NP-complete problems. Specifically, we reduce 3sat to Vertex Cover, and then reduce Vertex Cover to k-means, thereby establishing the computational hardness of the k-means clustering problem. We then prove k-means is in NP, and thus conclude it is …

Empirical Analysis Of Political Districting Splitability Via Uniform Spanning Trees In Polynomial Time, Brooke C. Feinberg

Scripps Senior Theses

This work expands a recently proven conjecture that a polynomial fraction of all uniform spanning trees (USTs) are splittable into k balanced partitions on grid graphs to real-world political districting plans. We investigate whether similar structural properties hold for the planar dual graphs of U.S. counties (cnty) and tracts (t), using Wilson’s algorithm to generate uniform random spanning trees and Breadth- First Search (BFS) to check for splitability into balanced partitions. Our empirical findings suggest that real-world districting plans can be split into 2-balanced, connected partitions in a fraction of polynomial time. This result highlights the potential for scalable redistricting …

Prediction Of Audit Findings Using Deep Learning With Financial And Non-Financial Data: A Case Study In Province X, Fery Yohan Setiawan, Eko Mulyanto Yuniarno, Reza Fuad Rachmadi

Knowledge Engineering and Data Science

The implementation of the audit from the local government financial statements by The Audit Board of The Republic of Indonesia (BPK RI), especially for the Province X representative, are frequently faced by the various limitations, one of them being the required audit time. At this moment, the BPK RI representative of Province X doesn’t have the tools that are able to help the accurate of sample determination for the pick test, which resulted in this study proposing the application of multi-label classification to predict the findings of financial statement (Laporan Keuangan, LK) audits based on financial and non …

Deep Learning For Irish Garden Bird Identification: Exploring The Role Of Cnn-Lstm In Video-Based Recognition, Antonina Dolynenko

ICT

Bird populations are widely used as indicators of ecosystem health, but traditional monitoring based on manual observation is labour-intensive and difficult to scale. Recent advances in deep learning and low-cost edge hardware offer new opportunities for automated, real-time bird identification in gardens and other local habitats. This thesis investigates whether video-based deep learning models can reliably classify common Irish garden birds from short motion-triggered clips and how temporal modelling compares to image-based models.

A primary dataset of 20-second clips was collected in a private garden in Ireland using a Raspberry Pi with a high-resolution camera and a YOLO-based trigger to …

Premier League Results Predictions, Laura Consuegra

ICT

This project applies machine learning techniques to predict outcomes in the English Premier League, one of the most prestigious and widely followed football competitions worldwide. By analysing historical and real-time match data, including performance metrics such as goals scored, shots on target, and cards received, predictive models are developed to forecast match results with higher accuracy. The study evaluates the effectiveness of these models and explores the influence of key statistical features on team performance. The findings aim to provide strategic insights for fans, bookmakers, and coaching staff, supporting performance evaluation, tactical decision-making, and a deeper understanding of the factors …

A Comparative Analysis Of The Performance Of Customised And Pre-Trained Convolutional Neural Networks In The Classification Of Facial Expressions., Guilherme Soares Da Costa

ICT

Facial expression recognition (FER) is a highly relevant problem in the field of computer vision, with applications in various areas. Although this topic has made significant progress in recent years with deep learning, facial expression recognition continues to face significant challenges when faced with a noisy, imbalanced database, and low-resolution images, as is the case with the widely used FER2013 database. This work investigates the usability of different approaches based on Convolutional Neural Networks (CNNs) for the task of facial expression classification.

First, a custom CNN was developed and trained from scratch, using FER2013 as database. Subsequently, three well-known pre-trained …

Clustering And Predictive Modelling Of Cryptocurrencies: An Empirical Study Of Lead–Lag Dynamics And Forecasting Performance., Viktor Varga

ICT

Cryptocurrency markets pose analytical challenges due to their volatility, non-stationarity, and complex interdependencies. Traditional econometric models have struggled with these features, while machine learning models often face overfitting and limited interpretability. This thesis has developed an integrated framework combining clustering, lead–lag detection, and predictive modelling to examine whether structural and temporal dependencies in cryptocurrency markets can improve forecasting performance.

Hourly OHLCV data for a selection of liquid cryptocurrencies from the Binance API over a thirteen-month period has been analysed. Dynamic Time Warping (DTW) has been used to identify asset clusters, while cross-correlation, Granger causality, and Vector Autoregression (VAR) have been …

Exogenous Variables In Time Series Forecasting During Economic Volatility: A Cross-Sector And Cross-Crisis Evaluation., Orla Kavanagh

ICT

Economic disruptions pose challenges for time series forecasting, but also offer opportunities to evaluate the value of exogenous information. This study investigates the role of external indicators during the COVID-19 pandemic and the 2025 U.S. tariff shock, using data from Dell (technology) and Boeing (aerospace/manufacturing). ARIMA, SARIMA, and LSTM models were tested with candidate exogenous variables. Results indicate that the benefit of external signals is highly context-dependent: sector, disruption type, and forecast horizon all influence effectiveness. LSTM models generally excel at longer horizons with exogenous inputs, while ARIMA-based models perform better short-term. Findings emphasize that careful variable selection, ongoing evaluation, …

Understanding Model Behaviour And Interpretability In Time Series Forecasting: A Deep Dive Into Lstm And Gru With Xai Techniques., Federico Ariton

ICT

Time-series forecasting is widely used in data analytics, yet the interpretability of deep learning models remains a key challenge. This study compares Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models using a dual interpretability framework that combines attention mechanisms and SHAP analysis. Two feature sets were evaluated across assets with different volatility regimes: Solana and Shiba Inu (high volatility), Bitcoin (moderate volatility), and Apple (low volatility). Results show that autoregressive features achieved the lowest forecasting errors, while volatility- and momentum-based indicators provided stronger interpretability. GRU performed best in moderate-volatility conditions, whereas LSTM demonstrated more consistent performance across varying …

Traditional Vs Deep Learning Approaches For Efficient Electricity Consumption Prediction., Lucas Sant’Ana

ICT

Accurate forecasting of electricity demand is critical for reliable energy planning, resource allocation, and policy design. Traditional statistical models, such as ARIMA, SARIMA, and ARIMAX, have been widely applied but remain constrained by linear assumptions, limited temporal memory, and inflexibility in handling multiple exogenous drivers. In this study, we systematically compare these approaches with multivariate Long Short-Term Memory (LSTM) networks, which are capable of capturing nonlinear dependencies, long-term temporal dynamics, and multivariate interactions. Historical electricity consumption data were combined with weather variables, including temperature, wind speed, and rainfall, and pre-processed through cleaning, scaling, and temporal alignment. Statistical baselines and deep …

Credit Card Fraud Detection., Sonia Ndonga

ICT

This capstone project applies machine learning to detect credit card fraud, addressing a critical financial threat to banks and payment providers. Using an anonymised dataset of 284,807 transactions, which is highly imbalanced with only 0.172% fraudulent cases, three models—Logistic Regression, Random Forest, and Gradient Boosting—were developed and evaluated. The pipeline incorporates data preprocessing, feature engineering, hyperparameter tuning, cross-validation, and interpretability analysis using SHAP values, SHAPASH, and permutation importance. Random Forest achieved the highest performance with an ROC AUC of 0.97 and Average Precision of 0.66. The study also considers fairness, threshold optimisation, and practical deployment strategies, providing a robust automated …

Comparison Of Convolutional Neural Network Architectures For The Classification Of Microscopic Images: Performance Evaluation Of Lightweight Models., Cristian Ricardo Gonzalez Donoso

ICT

This study evaluates lightweight convolutional neural networks (CNNs) for multi-class classification of microscopic fungi images in CPU-only environments, addressing resource-limited laboratories and educational settings. Four pre-trained models—EfficientNetV2-B0, MobileNetV3-Small, MobileNetV2, and NASNetMobile—were compared on the DeFungi dataset (9,114 images, five fungal classes) using stratified training, realistic data augmentation, and two-stage transfer learning. Performance was measured via accuracy, macro-F1, precision, recall, and efficiency metrics including training time, inference latency, model size, and FLOPs. EfficientNetV2-B0 achieved the highest accuracy, while MobileNetV3-Small provided the best macro-F1 and lowest computational cost, making it the optimal choice for CPU-only deployment. Findings highlight the trade-off between predictive …

Enhancing Fake News And Rumor Detection Using Metadata, Context, And User Interaction With A Hybrid Gcn-Transformer., Lucas Romulo Zuin Gigli

ICT

The increasing influence of social media has significantly impacted how news spreads among users, making the detection of misinformation, including fake news and rumours, a critical task. Previous research has explored content-based and metadata-driven approaches, while recent advancements have leveraged Graph Neural Networks (GNNs) such as GCNs, GATs, and SAGE to analyse user interactions and propagation patterns. This thesis investigates the effectiveness of a Hybrid Graph-Transformer Convolutional Neural Network (GCN-Transformer) for fake news detection, combining graph-based learning with Transformer attention mechanisms. Through in-depth data analysis and advanced detection techniques, this model aims to enhance predictive performance and mitigate the spread …

Analysis Of Customer Churn In The Banking Sector And Its Mitigating Factors., Fiachra O Callaghan

ICT

This project investigates customer churn within the banking sector, analysing a dataset of 10,000 customers to identify patterns and potential drivers of attrition. By leveraging data analytics and visualization techniques, we aim to understand the demographic and behavioural factors influencing customer decisions to leave for competing banks. The study applies the CRISP-DM framework to structure the analysis, enabling iterative refinement of objectives and hypotheses as insights emerge. Findings indicate that certain customer segments, such as those aged 40–60, are more prone to churn, highlighting opportunities for targeted retention strategies, including fee reductions and tailored loyalty programmes. By integrating academic research …

Credit Card Default Prediction Using Machine Learning., Yassine Zohair

ICT

This project investigates the use of machine learning to predict credit card payment defaults, aiming to help banks and credit card companies mitigate financial losses. Using historical customer data, including demographics, income, education, and previous payment behaviour, three machine learning algorithms were implemented to forecast the likelihood of default. Techniques such as cross-validation, hyperparameter tuning, and SHAPASH were applied to improve model performance and interpretability. Accurate prediction of potential defaulters enables financial institutions to take proactive measures, such as adjusting credit limits or providing targeted financial guidance, thereby enhancing risk management and customer retention.

An Enhanced Deep Learning Framework For Crop Disease Detection Using Gan-Based Data Augmentation., Andrew Mc Guinn

ICT

Plant disease detection is a critical challenge in agriculture, where limited annotated datasets often hinder the training of robust deep learning models. This thesis explores the application of generative adversarial networks for data augmentation in the classification of six classes of sorghum diseases, addressing the scarcity of real training data. A lightweight convolutional neural network  (CNN) was employed as the classifier, trained under three experimental conditions: using only real images, a stratified mix of real and DCGAN-generated samples, and a stratified mix of real and WGAN-GP-generated samples.

The baseline CNN, trained on 279 real images, achieved an accuracy of 89.7%, …

Discovering Latent Themes: Mixed-Methods Comparative Analysis Of Topic Extraction And Clustering., Laura Byrne

ICT

This study investigates whether clustering and topic modeling can uncover themes within the 167 verses of the King James Version of the Book of Esther. A standardized preprocessing pipeline was applied, and TF-IDF and sentence-embedding feature spaces were used to evaluate topic extraction (NMF, GSDMM, BERTopic, Top2Vec) and clustering (HDBSCAN, DBSCAN, Gaussian Mixtures, Agglomerative) using coherence, cluster validity, lexical distinctiveness, and cross-model similarity metrics. NMF provided the most interpretable topics, GSDMM yielded compact low-overlap topics, BERTopic offered entity-centered groupings, and Top2Vec highlighted broad thematic regions. Cross-model analysis revealed overlapping motifs, notably decree and banquet themes, but clusters and topics did …

Leaf Classification Using Convolutional Neural Networks And Vision Transformers, Louis Wilkie

ICT

This study investigates deep learning techniques for plant image classification using four publicly available datasets: Folio, Flavia, Swedish, and Sugarcane. Three modelling approaches—Convolutional Neural Networks (CNN), Vision Transformers (ViT), and hybrid CNN-ViT ensembles—are evaluated. Models leverage PyTorch implementations and pretrained weights from the Hugging Face hub, with hyperparameter tuning applied to optimize accuracy. Transfer learning with pretrained ViT models achieved the highest performance, attaining 100% accuracy on Folio, Flavia, and Swedish datasets, and 94.07% on Sugarcane. CNN and CNN-ViT models showed comparable results, but pretrained CNN-ViTs were more resource-efficient. The findings highlight the importance of evaluating multiple pretrained architectures to …

Improving Fairness In Convolutional Neural Networks For Demographic Face Classification., Leandro Andrade

ICT

This study examines racial bias mitigation in Convolutional Neural Networks (CNNs) for demographic face classification using the FairFace dataset. Three architectures—ResNet50, VGG19, and InceptionV3—are evaluated, with dataset balancing strategies including undersampling and class weighting. Results indicate that InceptionV3 with class weighting achieves the most consistent performance across racial groups, with improved F1-scores and generalization through hyperparameter optimization and data augmentation. Challenges remain in distinguishing visually similar groups, highlighting the need for equitable datasets and fairness-aware training. These insights are critical for ensuring accuracy and fairness in applications such as law enforcement, healthcare, and human–computer interaction.