Data Science | Hyfrankl

Human Brain Acitivity Detection (Ongoing)

Electroencephalography (EEG) is a non-invasive method of monitoring and recording electrical activity in the brain, which plays a crucial role in diagnosing and treating various brain-related disorders, particularly in critically ill patients. However, the manual interpretation of EEG data remains a major bottleneck in neurocritical care, as it is time-consuming, expensive, and prone to fatigue-related errors and inter-rater reliability issues. To address these challenges, there is a pressing need to develop automated methods for EEG analysis. The objective of this research is to develop a robust model trained on EEG signals to automatically detect and classify seizures and other types of harmful brain activity, aiming to assist doctors and brain researchers in providing faster and more accurate diagnoses and treatments. This work has significant implications for improving patient care in neurocritical settings, advancing epilepsy research, and supporting the development of new therapeutic interventions.

Frank (Haoyang) Ling, Sijun Tao, Yifei Zhang

Human Brain Acitivity Detection (Ongoing)

NPHardEval: Benchmarking Reasoning Ability of Large Language Models via Complexity Classes

NPHardEval serves as a comprehensive benchmark for assessing the reasoning abilities of large language models (LLMs) through the lens of computational complexity classes. This repository contains datasets, data generation scripts, and experimental procedures designed to evaluate LLMs in various reasoning tasks. The benchmark offers several advantages compared with current benchmarks:

Data construction grounded in the established computational complexity hierarchy
Automatic checking mechanisms
Automatic generation of datapoints
Complete focus on reasoning while exclude numerical computation

Lizhou Fan, Wenyue Hua, Lingyao Li, Frank (Haoyang) Ling, Yongfeng Zhang

NPHardEval: Benchmarking Reasoning Ability of Large Language Models via Complexity Classes

Movie IR System

In the modern entertainment landscape, finding the perfect movie across numerous streaming platforms has become a challenging task. Current recommendation systems often fall short, relying on generic algorithms or user ratings that overlook individual preferences. Our project addresses this challenge by redefining the movie discovery experience. Unlike traditional approaches, our innovative pipeline combines exact matching with semantic understanding and refined re-ranking. This approach not only achieves a notable increase of 0.11 in nDCG@10 over the baseline BM25 but also aligns more closely with user intent and content relevance.

Frank (Haoyang) Ling, Jia Zhu

Diamond Prices Prediction

This project aims to build accurate predictive models for diamond prices using regression models. It is crucial for diamond sellers and buyers to determine competitive prices and make informed decisions, respectively. Factors such as carat weight, cut quality, color grade, clarity, and physical dimensions impact a diamond’s price. The project uses a dataset from Kaggle, with data on these characteristics for 53,909 diamonds. The analysis encompasses three regression models and handles multi-collinearity among predictors and prevents model overfitting. A range of techniques like visualization, exploratory data analysis, model diagnostics, and train-test split are used to ensure the models’ reliability. The selected model achieves 0.9848 in R-squared.

Frank (Haoyang) Ling, Ruiqi Xu, Zhemin Qu

Machine-Generated Text Detection for ChatGPT

Discriminated statistical information of machine-generated text to propose an explainable classifier, achieving comparable predictability to BERT-based models.
Performed a thorough analysis of data augmentations on response length, identifying that truncated sentences can decrease the model performance by around 5%

Frank (Haoyang) Ling, Haoquan Zhou, Kexuan Huang

Machine-Generated Text Detection for ChatGPT

Yelp User Analysis

With the rise of deep learning and machine learning, many retailers adopt recommendation systems to increase their competitive ability in the market. Yelp platform has published an extensive dataset about its user and business profiles (around 9 GB). Many researchers have explored the dataset, but few of them focus on friend recommendations with users. In this project, the k-hop sub-graph or the ego-net of one specific user will be analyzed to provide diverse recommendations

Contrastive Learning on Graph Representation

Augmented graph data with random masking and self-attention after comparative analyses
to enhance the model’s robustness, surpassing baseline models in 5 out of 9 datasets.

Housing Price Prediction in Illinois

This project involves working with a dataset from the Cook County Assessor’s Office in Illinois, which contains over 500,000 records describing houses sold in the area in recent years. The dataset is split into training and test sets.

In Part 1, Exploratory Data Analysis (EDA) ais performed. In Part 2, the focus is on advanced prediction with machine learning. The criterion for evaluation is L2 loss, and the baseline model is ridge regression.