Identifying Judicially-Challenged Patents Held by Chinese Listed Companies: A Machine Learning Benchmark with Partial Judicial Labels and Positive-Unlabelled Re-evaluation

We release a tabular machine-learning benchmark for predicting which patents held by Chinese A-share listed companies are vulnerable to judicial invalidation by the China National Intellectual Property Administration. Following Elkan and Noto (2008) and du Plessis et al. (2014), we frame the task as positive-unlabelled (PU) learning: 545 patents have been actually challenged and invalidated, while the remaining 1.17 million unlabelled patents mix genuinely safe patents with challengeable patents that have not yet been challenged. The benchmark integrates four feature families: 12 patent-level structured features after removing two post-grant cumulative-citation variables identified as ex-post leaky, 100 firm-year innovation indicators, 4 event-study cumulative abnormal returns, and 32 PCA-reduced sentence-transformer embeddings of patent claim text. Seven tabular classifiers (Logistic Regression, Random Forest, MLP, XGBoost, CatBoost, LightGBM, and a Frank-Hall ordinal ensemble) are benchmarked under a leakage-aware StratifiedGroupKFold-by-firm protocol, complemented by paired-bootstrap confidence intervals on AUC differences, five-seed multi-seed stability, three K-fold settings, an imbalance-handling sensitivity, a CAR estimation-window robustness, and ordinal-axiom variants. Headline findings: naive K-fold inflates patent-level AUC by an average of 0.151 across the seven models; Random Forest is the strongest tabular baseline at patent-level AUC 0.7081 and survives every seed and every robustness check; at the firm-year level, logistic regression matches LightGBM at AUC 0.855, with the top-decile invalidation rate roughly 5.3 times the firm-year base rate of 0.93%. The release ZIP contains the parquet-formatted analysis data, the upstream Stata pipeline that rebuilds the analysis sample from CSMAR raw inputs, the Python ML benchmark, all out-of-fold predictions and metric JSON files, the three figures used in the manuscript, and the Knowledge-Based Systems submission package.