Xinyi's Odyssey

Natural Language Process

In this project, I have focused on harnessing cutting-edge natural language processing and machine learning technologies, such as Pretrained BERT, LoRA, LangChain, Word2Vec, GNNs, GraphSAGE, RNNs, to address a wide range of real-world challenges, ranging from building citation recommendations to building dynamic question-answering systems, analyzing customer sentiment, and detecting fake news. Each project presents unique challenges and goals, but they share a common purpose: significantly improving the accuracy and accessibility of information to influence and enhance decision-making processes.

Reinforcement Learning

I have developed and implemented a range of reinforcement learning algorithms designed for autonomous mastery in various game environments. My work has primarily focused on leveraging techniques such as Dueling Double Deep Q-Networks, which achieved a score of 367 in the Breakout game. Additionally, I have applied Monte Carlo methods and Temporal Difference learning to improve decision-making in simulations like Blackjack and Cliff Walking, as well as dynamic programming techniques such as policy evaluation and value iteration for markov decision processes. These advanced algorithms are designed to enhance decision-making, strategy formulation, and problem-solving in complex environments.

Information Retrieval

In my journey of developing recommender systems, I considered the realm of item-based collaborative filtering and matrix factorization. It was designed using various similarity metrics to optimize recommendations. This technique revolves around the concept of understanding and leveraging the relationships between different items based on user interactions and preferences.

Taking the system a notch higher, I ventured into the integration of advanced deep learning methods. This included the implementation of Neural Collaborative Filtering (NCF) and Transformer models. The incorporation of NCF allowed me to blend the classic matrix factorization approach with deep neural networks. In addition to utilizing Transformer models in my recommender system, where their capacity to handle complex dependencies in natural language processing, I am exploring further applications.