Jianing (Oscar) Xie

Supervisors: Alan Wang, Wenqi Shen, Dan Ke

• Developed machine-learning-based NLP methods for 50 stylometric features grounded in the Writeprints framework, covering lexical, syntactic, structural, and semantic dimensions. Refined syntactic measurement through dependency parsing and improved named entity recognition of time, location, and event expressions.
• Introduced Hill numbers from ecology to measure linguistic diversity on a single effective-number scale. Unlike KL and JS divergence, which only report how far two distributions drift apart, Hill numbers of different orders reveal whether diversity is lost among rare categories or among dominant ones. Built an evenness ratio on Hill numbers to isolate distributional balance, showing collapse erodes common categories faster than dominant ones.
• Built a 20-generation recursive training pipeline that fine-tunes Llama-3.1-8B and Qwen3-8B with LoRA on Wikipedia and Reddit corpora using NVIDIA A100 GPUs. Each generation trains on a 1:1 mix of synthetic text from the previous generation and fresh human text, and every setting is replicated across 10 random seeds.
• Applied a two-way fixed effects (TWFE) model with generation and seed effects to test feature redundancy on residuals, removing the shared collapse trend that inflates raw correlations.

Supervisors: Richard A. Hunt, Joseph Simpson

• Conducted a science-of-science integrative review of how NHST (Null Hypothesis Significance Testing) became an institutionalized epistemic rule in management research rather than a purely statistical technique.
• Designed a reproducible search and screening protocol on Web of Science in Python; predefined inclusion criteria narrowed 492 candidate articles from FT50 and leading review journals to a final sample of 165 peer-reviewed papers.
• Scored each article's stance toward NHST through LLM-based polarity analysis and mapped papers onto management sub-fields; built the sentiment heat maps and publication-trend figures in the review.
• Proposed one of the review's six core insights: the risk of epistemic inversion under generative AI, where AI workflows can reinforce p-hacking by reasoning backward from detectable patterns to seemingly theory-driven claims.

Supervisors: Rong Du, David M. Townsend

• Used a quasi-natural experiment with Difference-in-Differences (DID) to study the impact of AI applications (recommendation, review summarization, knowledge Q&A) on consumer purchasing behavior and perceived quality.
• Developed web-scraping solutions with the Scrapy framework, extracting more than 100,000 user reviews.
• Conducted sentiment analysis with LLM APIs and NLP techniques (word2vec, BERT) and linguistic analysis with LIWC to identify key features in user reviews.
• Cleaned and processed over 1 million transaction records with ETL pipelines using NumPy, Pandas, and DuckDB.
• Theoretically, first to combine cue-utilization theory with generative AI in e-commerce, analyzing long-tail vs. winner-take-all dynamics; introduced Baidu Index as an objective brand-awareness metric with real sales data.

Supervisors: Rong Du, Richard A. Hunt

• Pioneered a diagnostic framework for AI-Generated Content (AIGC) trustworthiness by integrating Nomology theory with a dual cognitive perspective — a multi-level, nine-dimension framework built via Grounded Theory and validated through large-scale text mining.
• Conducted semi-structured interviews with over 20 users, including domestic and international undergraduates, doctoral students, and industry professionals.
• Performed text mining and sentiment analysis on social-media data using topic modeling (e.g., BERTopic) to validate the framework and quantify trust across the nine dimensions.
• Diagnosed a core "representation–substance imbalance": positive perceptions of AIGC's surface experience are undermined by distrust in its core substance (information quality, transparency).

Supervisors: Rong Du, Andrew Burton-Jones

• Used a quasi-natural experiment with DID to study how product traceability information (traditional and blockchain-based) affects consumer purchase behavior, with brand reputation and eWOM as moderators.
• Applied LDA topic modeling and text mining to 14,000 product reviews, constructing an attention index for user topic focus.

Rank 1 / 93

• B.Mgmt. in Big Data Management and Application (Xidian) — GPA 3.9 / 4.0
  Minor: Artificial Intelligence & Large Language Model Application (Xidian)
• B.S. in Business, Management: Entrepreneurship, Innovation & Technology Management (Virginia Tech) — GPA 3.78 / 4.0

Selected core courses

• Business Intelligence Analysis (Data Mining) — A−
• Introduction to Data Science — A
• Management Information System — A−
• Database Principle and Application — A
• Business Statistics, Analytics & Modeling — A
• Statistical Machine Learning & Text Mining — A
• Natural Language Processing — A
• Principles of Large Models and Industry Applications — A