Claude Skills by wenmin-wu

Routes QA predictions through an answer-type classifier to emit boolean answers, extractive spans, or null based on type logits.

Trains a classifier to distinguish train from test data, detecting distribution shift and identifying leaked features.

Detects sentinel anomaly values in numeric columns, creates a boolean flag feature, then replaces the sentinel with NaN for proper imputation.

Probe an anonymized regression target by testing whether simple invertible transforms (2**y, exp(y), log(y), affine rescale) produce a distribution with recognizable structure — round numbers, integer histograms, or a familiar finance/retail range — and use the recovered semantics to motivate features and loss choices the host's bland description would never suggest

Split a transaction table by a binary status flag (authorized vs. declined, paid vs. refunded) into two parallel sub-tables, then build the same aggregate feature pipeline on each — the declined-transaction features are usually as predictive as the authorized ones because they encode risk and friction the authorized stream alone hides

Train a PyTorch autoencoder on time-series summary statistics to produce dense encoded features for downstream GBDT models

Class-balanced log loss that weights each class by the inverse of its sample count, equalizing the contribution of minority and majority classes.

Predict where a projectile will land using kinematic equations with estimated gravity to intercept aerial passes in game AI simulations

Recover missing categorical values by matching words in a related text field against a known vocabulary built from the full dataset

CatBoostRegressor with MultiRMSE loss for native multi-output regression, predicting all targets in a single model without per-target loops.

Predict correlated targets sequentially, using earlier target predictions as input features for subsequent targets to exploit inter-target dependencies

Streams large HDF5 files in fixed-size row chunks to compute summary statistics without loading the full dataset into memory.

Recommends items frequently purchased together with a customer's recent items using pre-computed pair dictionaries.

Removes redundant features by iterating pairwise Pearson correlations and dropping one member of each pair exceeding a threshold.

Augments imbalanced tabular data by independently shuffling each feature column within a class, creating synthetic samples that preserve per-column marginal distributions.

Hard-clips predicted probabilities to 0 or 1 when they exceed high-confidence thresholds, reducing log loss on near-certain predictions.

Encodes categorical groups by their target rate scaled by a log-confidence factor, smoothing unreliable rates from low-frequency groups toward zero.

Precomputes item/content difficulty as historical mean accuracy, merged as a static feature for user-item prediction tasks.

Minimize axis-aligned bounding box side length by finding the optimal rotation angle over convex hull vertices using bounded scalar optimization

Build user-level behavioral features (avg listing duration, relisting frequency, total items) by joining auxiliary activity tables that share user_id but not item_id with train/test

Model cumulative distribution via softmax output layer and CRPS loss — for probabilistic regression over discrete bins

Encodes cyclical features (hour, month, day-of-week) using sine/cosine transforms to preserve circular distance.

Config-driven feature factory that generates groupby aggregation features from a declarative spec list, supporting count, mean, var, nunique, cumcount, and custom lambdas.

Reshapes model predictions to match the known label distribution from training data using rank-based mapping.

Predefines minimal unsigned integer dtypes before CSV loading to cut DataFrame memory usage by 2-4x without any data loss.

Use sklearn FeatureUnion with closure-based preprocessors to apply different vectorizers to different DataFrame columns in a single fit_transform call

Adds each feature's value-count frequency as a new column, enabling tree models to split on how common or rare a value is.

Encode a 2D game board into a normalized multi-channel feature tensor with log-scaled resources, signed unit counts, and directional features for RL agents

Fits a Gaussian Mixture Model on the joint feature-target space and samples synthetic data pairs to augment small tabular datasets.

Constructs a customer similarity graph via KNN on mixed features, then trains a GraphSAGE GNN for node classification. Captures relational patterns that tree and linear models miss, adding ensemble diversity.

Uses GroupKFold to prevent data leakage when multiple rows share a common entity (e.g., same user, question, or document).

Custom evaluation metric that computes log of per-group MAE then averages, penalizing uniformly bad groups.

Splits train/validation using GroupShuffleSplit so that related samples (forks, families, sessions) never span both sets.

Generates geographically proximate candidate pairs for entity matching using KNN with haversine distance, optionally partitioned by country.

Three-level polygon overlap test — AABB early exit, then point-in-polygon ray casting, then segment intersection — for fast non-convex collision detection

Two-level group-then-pattern dispatch for game AI agents where groups filter by game state and ordered patterns within a group fire the first matching action

Alternating Least Squares matrix factorization on sparse user-item interaction matrices for implicit feedback recommendations.

Computes leak-free target encoding statistics (mean, std, min, max) using nested inner KFold within each outer CV fold, preventing target leakage that occurs with naive groupby-based encoding.

Multi-round pseudo labeling with progressively confident test predictions merged into training plus OOF-based train label correction

Extract and visualize per-subgroup feature coefficient signs from L1-regularized models as an interaction heatmap for EDA

Compute first-order difference between last and second-to-last rows per entity in panel data to capture recent trend direction and magnitude

Iterates through rows chronologically to accumulate user statistics, fetching current state before updating to prevent future data leakage.

Use LightGBM DART boosting (dropout on trees) with aggressive feature and bagging fractions to reduce overfitting on high-dimensional tabular data

Average model predictions across CV folds in log-odds space rather than probability space for better-calibrated ensemble outputs

Applies logit transformation to base model probabilities before fitting a logistic regression meta-learner, enabling principled linear combination in log-odds space.

Rerank session candidates using log-spaced recency weights multiplied by interaction-type multipliers

Blend multiple submission CSVs by row-wise majority voting on discrete predictions to produce a more robust final output

Converts molecular SMILES strings to fixed-length Morgan fingerprint bit vectors using RDKit for use as tabular ML features.

Keras multi-input model with separate embedding layers for categoricals, GRU for text sequences, and dense layers for numerics, all concatenated into a shared regression trunk

Neural network with multiple output heads for main target plus auxiliary targets, improving representation learning via shared layers.