Projects

Fake news is a huge problem, especially on social media platforms like Twitter and Weibo. To tackle it, I helped create GraMuFeN, a model that combines text and image data to spot fake news more effectively. GraMuFeN leverages Graph Convolutional Networks (GCNs) for text and Convolutional Neural Networks (CNNs) for images. By training it with both text and images, we achieved a lightweight, high-performance model with a 10% improvement in accuracy and half the trainable parameters compared to existing approaches. The model aligns textual and visual content to detect mismatches that could indicate falsification, showing it’s possible to make fake news detection both powerful and efficient.

This project implements a neural network to train AI agents, enabling them to learn the best strategies for survival using NEAT (NeuroEvolution of Augmenting Topologies). The AI leverages NEAT-Python to evolve networks over generations, optimizing the agent's ability to dodge falling obstacles.

Classifying text when classes are unevenly represented is a tough job, so I developed a model that treats text as a graph to address this imbalance. We used Graph Convolutional Networks (GCN) with LSTMs to capture relationships in text and added adversarial weight-generating networks that assign higher importance to underrepresented classes. This balancing technique directs the model’s attention to minority classes by adjusting weights dynamically as it learns. Tested on Yelp reviews, our model outperformed traditional approaches with a 4% accuracy boost, effectively handling real-world imbalances while improving F1-scores for minority classes.

Developed an AI algorithm to analyze Open-High-Low-Close (OHLC) candlestick patterns in financial data, aiming to detect recurring trading signals. The algorithm identifies similar candlestick patterns by analyzing OHLC characteristics, helping uncover actionable trends. Integrated with a decision-making neural network, this solution offers buy/sell recommendations based on pattern insights, optimizing trading strategies across our operational cycle. This integration enables real-time analysis and advice, enhancing decision accuracy in market environments.