Final Project (M.Sc. Data Science, HIT). Predicting the bounties of One Piece characters — a quantitative threat score — from wiki text combined with a character-relationship network. Data collected independently from the One Piece Fandom Wiki API; no Kaggle or pre-built dataset.
- Independent Data Collection: 1,734 character pages scraped via the public Fandom API — 218 characters with a known bounty (the target) plus hundreds more for the relationship graph.
- Text Features: TF-IDF + Truncated SVD on cleaned page text — bounty figures scrubbed before modeling, and TF-IDF/SVD fit on the training fold only within each CV split (leakage control).
- Network Analysis: Character-relationship graph with structural features, visualized via a layered k-core layout (~1,000 nodes).
- Error Analysis: Systematic errors investigated (e.g. weak characters on strong crews over-predicted), plus six visualizations.
Three-way comparison with CatBoost — combined model R² = 0.42 vs. 0.37 text-only and 0.13 network-only (median baseline: −0.05): text and network carry complementary signal, and their combination beats either alone.
one_piece_nlp_project.ipynb: Full end-to-end notebook (collection, features, modeling, evaluation).data_collection.py/data_collection.ipynb: Fandom-API scraper producingonepiece_raw.csv.text_features.py: Text cleaning and TF-IDF/SVD extraction.network_plots.py: Graph construction and visualization.features.py: Feature assembly.modeling.py: CatBoost training, cross-validation and evaluation.onepiece_raw.csv: Raw collected dataset (~27MB, 1,734 rows) — regenerate withdata_collection.py.one_piece_report.pdf/one_piece_bounty_presentation.pdf: Full report and presentation.Project_Proposal.pdf/Final_Project_Instructions.pdf: Approved proposal and course guidelines.DATA.md/ETHICS.md/REFLECTION.md/AI_USAGE.md: Data provenance, ethics, reflection, and AI-usage documentation.