Full Papers

Predicting Customer Default Payments with Key Features Using Data Mining Techniques

Dara Tourt, Simon Jin, and Queen Booker – Metropolitan State University

Credit card default prediction poses a significant challenge for the financial industry, affecting lenders and consumers alike. This study aims to tackle this challenge by utilizing a dataset from Taiwanese credit card users to explore the predictive power of various machine learning models. Employing Recursive Feature Elimination with Cross-Validation, we identified key predictors of default and assessed the performance of seven different models. Our findings reveal that feature selection can substantially improve model accuracy, with Random Forest models showing the most significant improvement. This research provides valuable insights into effective predictors of credit card default, offering a solid foundation for financial institutions to enhance their risk management protocols. Despite encountering data quality issues, the methodologies applied herein demonstrate the robust potential of machine learning in financial risk analysis.

A Natural Language Processing Indexing Algorithm in Python to find Future Recommendations in Journal Databases

Hector Ojeda and Michael Hart – Minnesota State University, Mankato

Machine learning continues to help improve processes such as indexing. Researchers can benefit from these improvements by leveraging techniques that reduce the time it takes to complete exhaustive reviews of literature. To optimize searches for future recommendations in scholarly journal databases, the authors develop an improved index algorithm using natural language processing. Following the design science research method, a prototype is developed using the Python programming language to test the index in several academic databases. Experimental results are positive when using natural language processing techniques to find future research recommendations. Similar indexing algorithms can increase the rate of finding keywords requiring specialized string matching. Therefore, natural language processing should be considered when developing modern indexing tools for researchers.

The Influence of Age and Webpage Prototypicality on Human-Computer Interaction (HCI) Using Modern Linear Regression Assumptions

Lyndsey Bowers and Michael Hart – Minnesota State University, Mankato

This study focuses on an age group that could be underrepresented specific to Human Computer Interaction (HCI) and website design. Review of literature found that people over the age of 60 are not commonly represented when designing websites. More intentional inclusion in HCI standards may foster higher levels of interaction for these website users. To explore this, the authors conducted a study that investigates webpage prototypicality on demographic data that includes age, education level, gender, and occupation. The research subscribes to a canonical action research design to test whether demographics influence prototypicality. Although there is a significant difference between age and the webpage prototypicality factors of pre-use usability and trustworthiness, there is insufficient variance in the sample. Goodness of fit tests indicates future statistical analysis is needed to identify subsequent factors that could contribute to how varying age groups rank factors such as pre-use usability, trustworthiness, and
visual aesthetics.

Snow Classification Using Prompt-Based Generated Images

Ricardo de Jeijn and Rajeev Bukralia – Minnesota State University, Mankato

This study aims to assess the effectiveness of synthetic images for training a snow
detection model compared to real-world images, addressing data scarcity. The
emergence of generative neural network models has revolutionized fields like medical imaging and environmental monitoring by enabling synthetic image creation. Generative models, especially those utilizing generative adversarial networks (GANs) and attention-based architectures, have significantly advanced image synthesis. While early GANs struggled with abstract reasoning, subsequent integrations of latent space encoding and attention mechanisms improved their performance, with large language models further enhancing text-to-image synthesis for photorealistic images. In this experiment, a snowy sidewalk detection network, trained on 98 real-world images, was tested on both real and synthetic datasets using VGG-19 and ResNet-50 architectures, with predictions combined through a weighted average ensemble. Synthetic images were generated using Imagen v3, yielding 225 photorealistic images of snowy and clear sidewalks. Results, evaluated through accuracy, recall, and F1-score, showed that ResNet-50 outperformed VGG-19 on synthetic data due to its advanced feature extraction
capabilities; however, both models demonstrated lower performance on synthetic data compared to real-world data, highlighting the challenges of accurately representing complex visual phenomena like snow and ice. The ensemble model performed worse than the individual models on synthetic data, emphasizing the need for high-quality synthetic data. This study underscores both the potential and the limitations of using synthetic images in training machine learning models for snow detection, advocating for continued advancements in image generation techniques.

Prompt Engineering for Database Design

Saumya Gautam – Minnesota State University, Mankato

With the increased complexity of data management there has been a great need for an efficient way for database design. This paper shows how prompt engineering can be used to help with it using large language models (LLMs) like ChatGPT. Prompt engineering can be used in the design process to increase the capabilities of the LLM and help it better understand the data structure during the design process. The objective of the research is to find effective prompting strategies for LLMs to get the most relevant database entities, attributes, relationships and diagrams, while addressing normalization requirements. We will review pre-existing prompting techniques and demonstrate their application to database design. Our goal is to identify effective prompting techniques that can help us generate meaningful database design. This research has important implications for database design and shows us new ways of using LLMs in data management tasks.

PointNet model for 3D Object Detection using Deep Learning Techniques

Abdilahi Jama, Boubacary Bocoum, Mansi Bhavsar, and Chandra Jaiswa – Minnesota State University, Mankato and North Carolina Agricultural and Technical State University

Point Clouds are unique due to their irregular format which is voluminous in nature. Usually, it is a set of points in 3D space that have no ordering can be shuffled and will still have the same exact same point cloud. This provides an advantage since it captures spatial structures and encodes them as features in lower dimensional space. In addition, object detection has incorporated the use of deep learning techniques to accurately detect and monitor surrounding objects in real-time to address safety while driving. LiDAR (Light Detection and Ranging) sensors mounted on vehicles collect point cloud data that enhance 3D Object Detection. To overcome that we propose a CNN-based PointNet which
is a unified approach to handling point clouds directly as inputs and produce either class labels for the entire input or per point label as output for each point of the input. Classifying an input point cloud, performing semantic and even part segmentation is part of PointNet’s robust features. It can extract both global and local features of a point cloud irrespective of orientation which are features in separate heads for classification or segmentation tasks. We evaluate and compare results using two datasets: the raw KITTI dataset and the nuScenes mini dataset. Our findings demonstrate that the proposed CNN-PointNet model outperforms well in terms of accuracy and computational efficiency. Specifically, our experiments indicate a significant improvement in detection accuracy on both datasets, with CNN PointNet achieving an average accuracy of 97.29% on the
KITTI dataset and 73.27% on the nuScenes mini dataset.

Automation and Operator Algebras for RPA in RavaTTT: A Comparative Study with IFTTT and Existing RPA Frameworks

Bheemaiah Anil K – UMN Twin Cities

This paper explores the necessity of advanced automation frameworks like RavaTTT, an event programming language invented by the author, in comparison to traditional platforms such as IFTTT and existing Robotic Process Automation (RPA) tools, using a real-world conservation scenario. We demonstrate how RavaTTT leverages operator algebras to achieve superior performance in managing multi-sensor fusion, real-time state management, and dynamic automation. A case study on the conservation of an endangered plant species is presented, utilizing sensors for temperature, humidity, and lighting conditions, along with actions such as lamp and humidity controls. We prove the superiority of operator algebraic approaches over threshold-based conditional logic, particularly in non-linear environments where multiple factors must be dynamically coordinated.

A Deep Learning Approach for Brachial Plexus Identification

Neha Sure, Aishwarya Matam, Deepika Velanati, Naseef Mansoor – Minnesota State University, Mankato

Surgical pain can be intolerable and agonizing for patients. While general anesthesia is used to reduce pain, it acts on the entire body and causes many side effects, such as nausea, vomiting, and other complications. Since the safety of the patient comes first, alternate pain mitigation methods like regional anesthetics have become a customary approach. In this paper, we proposed machine learning methods for correctly identifying the brachial plexus nerve block for performing regional anesthesia safely. The method numbs only the surgical area, allowing the patient to stay awake and reducing the risk of systemic side effects from general anesthesia. This research hypothesizes that a hybrid CNN-U-Net model will achieve superior segmentation performance compared to standalone U-Net and Resnet based U-Net model for brachial plexus nerve identification ultrasound images, balancing accuracy, precision, and computational efficiency.

A Review of Two Stochastic Methods Used in Deblurring Images

Mehdi Hakim-Hashemi

In this paper I briefly review two stochastic methods used in deblurring (denoising) pictures. Both methods assume the same premise and use random injection of noise to blur several pictures and after learning the parameters involved they use the process to deblur a given picture. The first method uses
stochastic differential equations and the second, known as DDPM, uses properties of Markov processes to do the job. Both models have been used in physical diffusion processes successfully.

Review based Group Recommender System using Attention Mechanism

Adith Santosh Thaniserikaran, Kashi Sharma, Dr. Sriram G Sanjeevi – New York University and National Institute of Technology Warangal, Telangana

In this paper, we propose a new approach to group recommendation that integrates both user ratings and reviews to improve the relevance and accuracy of suggestions. Traditional recommender systems often rely solely on user ratings for generating recommendations, which can lead to a loss of valuable information contained in user reviews. These reviews frequently provide context and insights that go beyond a simple rating, yet they are usually tailored for individual users. Our model addresses this limitation by using reviews alongside ratings to generate recommendations for groups of users, not just individuals. To achieve this, we process user reviews through a CNN-based text processor and apply an attention mechanism to emphasize the most relevant aspects of these reviews. The resulting predicted ratings are used to model individual user preferences, which are then further refined to predict group preferences, creating a recommendation system that captures a broader range of user feedback and improves the overall quality of group recommendations.

Work In Progress Papers

Collatz and a Ramsey Weight: Correlation Spotted, Explanation Sought

Mojtaba Moniri – Normandale Community College

We consider two functions on binary strings of length at most 363. Our weight function w, defined via edge-labeled ternary trees of depth 5, takes integer values between 1-8 (usually 3-5). We have algorithms designed to make the value low or high. The other is the famous Collatz function c taking real number values, here with a mean of about 6.95 and standard deviation about 0.63. We investigated their possible correlation and confirmed that with statistical hypothesis testing applied to our 96,000,000 generated inputs. It would be nice if a theoretical reasoning could be found for such correlations, which would potentially contribute to understanding the intriguing function c.

Facial Recognition using Deep Learning Algorithms for Authentication

Aaron Bottner and Rushit Dave – Minnesota State University, Mankato

Constant achievements in the field of machine learning (ML) and deep learning (DL) have caused an ability to create robust, smooth and accurate authentication systems. Facial recognition models can be lightweight while remaining accurate. These models can run anywhere from a lowly Raspberry Pi to the most powerful NVIDIA H100s. Currently, authentication suffers from the issue of decision-making. Humans are not great and creating, random data if you’ve ever seen a list of the top 10,000,000 passwords you’ll see that “randomness” is almost always consistent, meaning cracking a password you thought of as “random” is in actuality a password someone can learn to crack in hours. Thus, one of the counter measures against would be facial recognition. To adequately stop the threat, researchers in the past have created DL models to turn an image of a face into embeddings that describe your face with incredible accuracy, based on ML/DL techniques like Convolutional Neural Networks (CNN). This paper aims to investigate different methodologies with intent to achieve an efficient, low-cost model with a higher accuracy than other approaches by using different types of the datasets, which is to address the of the versatility dataset.

Assessing Credit Risk Through Transparent Machine Learning Algorithms

David Omondi Onyango, Shoumik Hasan, Mouhamed N Traore, Noah Haile, Naseef Mansoor – Minnesota State University, Mankato

To ensure transparency and auditability, balancing between credit risks and regulatory demands is essential for modern credit scoring models. With such a goal, this research investigates approaches for improving the predictive accuracy of simple and transparent models, attempting to achieve performance comparable to advanced machine learning techniques while preserving auditability. In this work, we experiment with logistic regression and SVM model for credit risk assessment. Our experiment demonstrates 71% accuracy for the logistic regression and 72% accuracy for the SVM model.

Research in Progress: Predicting demand in shelter beds for homeless in Minnesota

Rolande Umuhoza – Minnesota State University, Mankato

The Department of Housing and Urban Development (HUD) assists individuals or families affected by homelessness. Minnesota’s Homeless Management Information System(HMIS) defines homeless as a person, unaccompanied youth, or family sleeping on the streets, living in a car, or an emergency shelter without a permanent place to live. The dataset used in this research ranges from the year 2007 to 2023. The Prophet model will predict Minnesota’s demand for shelter beds by accurately predicting the capacity so that Minnesota’s community programs and HUD can better allocate resources for homeless services. Although still in progress, the future steps include implementing and validating the Prophet model for
forecasting.

Entity-Aware Caption Generation for Medical Images

Wengel Tsegaselassie, Tram Nguyen, Syeda Lamima Farhat, Naseef Mansoor

The radiology department plays a crucial role in healthcare by utilizing imaging techniques to monitor and diagnose medical conditions, providing detailed insights into the body’s internal state. Medical image captioning is a valuable tool in this domain, which can automate the interpretation of clinical images, streamlining the diagnostic process. This technique enhances the availability and accessibility of data, making it more efficient for healthcare professionals to assess and manage diseases. This study presents an entity-aware caption generation model for chest Xray images that combines Convolutional Neural Networks (CNN) for visual feature extraction
with Transformer models for text generation.

A review of Forward-Forward Algorithms in Convolutional Neural Networks

Spencer Connolly and Naseef Mansoor – Minnesota State University, Mankato

The Forward-Forward algorithm (FF) is yet another novel invention by Geoffrey Hinton, the creator of the famous backpropagation algorithm (BP). Since its proposal, many papers have been published exploring its potential, and good progress has been made. However, it continually falls just short of the standards set by BP. Despite this, the sentiment among researchers remains that FF has the potential to dethrone BP as the superior algorithm. In this paper, we present a literature review for FF algorithms applied to Convolution Neural Networks (CNN) for image classification tasks. For our
research, we focus specifically on papers which extend FF to CNNs due to the algorithm’s unexplored nature in this context.