CV

Zhijin Guo

Email: zhijin.guo97@gmail.com
Website: https://eng.ox.ac.uk/people/zhijin-guo/
Personal Site: https://zhijinguo.github.io/

Visa Status: Global Talent Visa holder (no sponsorship required) — Available to start work immediately

Summary

Postdoctoral Researcher at the University of Oxford with 4+ years’ experience in fine-tuning large language models, embedding relational data, constructing knowledge graphs, developing retrieval-augmented generation systems, and interpreting model behavior. Proven ability to adapt machine-learning and NLP techniques from social-network and recommender domains into systematic-trading signal frameworks and bespoke quality-control metrics for quantitative finance.

Work Experience

• Present: Postdoctoral Researcher Natural Language Processing, University of Oxford
  • Supervisor: Janet B. Pierrehumbert, Xiaowen Dong
• Present: Quantitative Research Consultant LLMs for Finance, Predictive Labs Ltd
• 2023 – 2024: Research Assistant Twitter Data Analysis, University of Bristol
• 2021 – 2023: Teaching Assistant Introduction to Data Analytics (NLP), University of Bristol

Skills

• Programming & Data Science: Python (6+ years), Java (4 years)
• Machine Learning / NLP / LLM: Training & fine-tuning Transformers and encoder-decoder models; open-source LLMs (LLaMA, SBERT, DeepSeek, ALBERT)
• Knowledge Graph & RAG: Designing Retrieval-Augmented Generation workflows with embedding models, vector search and Knowledge Graph construction
• Libraries & Frameworks: Hugging Face Transformers; SBERT; spaCy; NLTK; Pytorch; DGL
• Unstructured Data Processing: Twitter, Reddit, emails and other relevant text/graph data structure
• API Application: Data extraction via Twitter API, Reddit API, and various databases
• Version Control & Data Access: Git workflows and handling structured (SQL) and semi-structured (JSON/CSV/TXT) data

Education

• 2021 – 2024: Ph.D. Computer Science & Natural Language Processing, University of Bristol
  • Supervisor: Nello Cristianini, Martha Lewis, Edwin Simpson
  • Thesis: Embedding Relational Data: Analysis, Methods, Applications
• 2019 – 2020: M.Sc. Advanced Computer Science, The University of Sheffield
  • Thesis: A Tool to Correct Errors Made by Grammar Inference Algorithm.
• 2015 – 2019: B.Sc. Software Engineering, Northeastern University
  • Thesis: Design and Implementation of ERP System for Iron and Steel Enterprises.

Selected Projects

2025 QuantNLP: Turning Financial Text into Trading Signals

• Real-time news aggregation from APIs like Yahoo Finance, Reuters, Polygon, etc.
• Alpha generation via automated feature discovery and predictive modeling, with a strong focus on NLP for sentiment analysis, event detection, and topic modeling.
• Market backtesting with real data.

2024-2025 Bridging Social Structure and Discourse

• Challenge: Reddit communities were exhibiting escalating polarization, making proactive moderation impossible and increasing user churn.
• Action: Built a hybrid RAG pipeline that (1) retrieves user-language features via embeddings, (2) fine-tunes LLAMA and BERT models (using Hugging Face Transformers) for sentiment and persuasion-signal prediction, and (3) trains a continuous-time GNN to model interaction dynamics and assign continuously-changing polarization weights.
• Result: Forecast community fragmentation weeks ahead with 70% accuracy, analyzed and identified the top linguistic and interaction factors driving polarization, enabling moderators to implement targeted interventions.
• Finance Application: Emulated rolling-window retraining and signal forecasting workflows common in quantitative trading, with QA metrics like rolling MAE and drift detection to ensure robust next-period market-prediction signals.

Quantifying Compositionality in Data Embedding

• Challenge: State-of-the-art Transformer and graph models lacked measurable limits on context-driven meaning shifts, undermining trust in novel expression generalization.
• Action: Designed a two-step evaluation: (1) applied Canonical Correlation Analysis to quantify linear alignment between known entity attributes and their embeddings; (2) reconstructed embeddings for unseen attribute combinations across SBERT, GPT, LLAMA, and Knowledge Graph embeddings (TransE/DistMult), and measured L2 loss, cosine similarity, and retrieval accuracy.
• Result: Demonstrated additive compositionality during training, with improved generalization in Multi-BERT and a 1.5× rise in attribute–embedding correlation in graph models, and tracked that deeper transformer layers peaked at 94% compositional signal before tail-off, guiding optimal layer selection for downstream tasks.
• Finance Application: Developed rigorous embedding-quality metrics analogous to quant-finance signal validation, guiding model-selection and improving reliability of text-driven systematic strategy inputs.

2023-2024 Medical Influencer Social Network Analysis

• Challenge: Health misinformation spread unchecked during COVID-19, complicating public-trust efforts by authorities.
• Action: Compiled tweets from the top 100 medical influencers by “Influencer Score,” then built a few-shot, multi-label classifier (50 classes) by fine-tuning ALBERT; used BERTopic for thematic extraction.
• Result: Achieved 70% F1 and a 15% boost in multi-label classification, uncovered the top misinformation topics and incorporated them into a dynamic knowledge graph to reveal influencer–audience discourse networks.
• Finance Application: Applied LLM-augmented knowledge-graph and RAG pipelines to financial texts, enabling discovery of hidden supply-chain links, co-investor networks, and merger signals for event-driven trading strategies.

2022 EXTRACT: Explainable Transparent Control of Bias in Embeddings

• Challenge: Knowledge-graph embeddings risked leaking protected attributes (e.g., gender, age, occupation) from behavioral data, raising privacy and fairness concerns.
• Action: Engineered a vector-based knowledge base by constructing and embedding a knowledge graphs. Developed EXTRACT, a suite that applies CCA to pinpoint bias-leakage sources, decomposes embeddings into private-attribute vectors via linear-system solving, and integrates four transparent mitigation methods to strip unwanted signals without degrading model utility.
• Result: Demonstrating robust recommending performance alongside bias control, highlighting the trade-off between accuracy and privacy.
• Finance Application: Adapted bias-mitigation techniques to financial-entity embeddings, ensuring fair, regulatory-compliant representations for risk models and portfolio construction without sacrificing signal fidelity.

Selected Publications

1. Z. Guo, E. Simpson, and R. Bernardi, “Medfluencer: A network representation of medical influencers’ identities and discourse on social media,” in epiDAMIK 2024: The 7th International Workshop on Epidemiology meets Data Mining and Knowledge Discovery at KDD 2024.

2. Z. Guo, Z. Li, B. Tyler, X. Dong, and J. Pierrehumbert, “Bridging social structure and discourse,” Preparing to submit to the ACL Rolling Review (ARR), 2025.

3. Z. Guo, C. Xue, Z. Xu, et al., “Quantifying compositionality in classic and state-of-the-art embeddings,” Under review by the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025.

4. Z. Guo, Z. Xu, M. Lewis, and N. Cristianini, “Extract: Explainable transparent control of bias in embeddings,” in AEQUITAS 2023: AEQUITAS 2023 First AEQUITAS Workshop on Fairness and Bias in AI co-located with ECAI 2023, 2023.

5. Z. Xu, Z. Guo, and N. Cristianini, “On compositionality in data embedding,” in International Symposium on Intelligent Data Analysis, Springer, 2023, pp. 484–496.