Jinyuan Kang

Driven Data Analyst and Math/Computer Science Tutor with a proven track record at Guosen Securities and NZ Academy Learning Centre. Excelled in automating data processes, reducing costs, and enhancing student understanding through technology integration. Skilled in Python programming and team collaboration, effectively leveraging machine learning and data visualization to drive decisions and foster learning environments. Detailed Data Analyst Intern well-versed in various statistical methods such as descriptive and influential statistics. Independent and versatile individual skillful in identifying data issues and monitoring existing metrics. Offering [Number] years of experience collecting data, establishing facts and drawing valid conclusions. Proficient in [Software] paired with first-rate report writing abilities. Hardworking and passionate job seeker with strong organizational skills eager to secure entry-level [Job Title] position. Ready to help team achieve company goals.

1). Course Planning and Recommendation System

• Developed a course planning and recommendation system to assist new students by summarizing prerequisites and constraints of school courses.
• Acted as a project leader, interfacing with clients to gather requirements, dividing tasks, and optimizing the project workflow.
• Built the overall framework of the application, ensuring seamless integration of components.
• Managed user-related data collection, performing data cleaning and pre-processing using JupyterLab and Pandas, and imported data into a MySQL database.
• Collaborated with back-end developers to write the Django framework app and developed the REST API interface.

2) Sentiment Analysis of Restaurant Reviews

• Utilized TensorFlow to implement LSTM and Bi-LSTM architectures for model training.
• Collected and analyzed restaurant reviews from popular Chinese review websites.
• Conducted sentiment analysis to accurately determine and rate customer sentiment.

3) Question Answering System using NLP

• Implemented a question answering system based on SQuAD 2.0 using pre-trained models BERT and DistilBERT.
• Trained models with a fixed learning rate initially, then fine-tuned hyperparameters to enhance performance, including transitioning to a linear learning rate scheduler and incorporating weight decay along with beta1 and beta2 parameters to mitigate overfitting.
• Integrated a cosine learning rate scheduler to evaluate its impact on model performance.
• Conducted training with a batch size of sixteen over three epochs.
• Achieved notable results: BERT outperformed DistilBERT overall, while DistilBERT with a cosine scheduler achieved the highest performance metrics (Exact Match score of 48.91 and F1 score of 51.37).
• Demonstrated that DistilBERT is a more effective approach under time constraints and limited computational resources.