wenmin-wu

Utilizes LightGBM's tweedie objective for accurate zero-inflated count forecasting, enhancing predictions for retail demand.

Denoises 1D time series data using wavelet decomposition and thresholding, enhancing trend extraction without lagging the signal.

Trains LightGBM on combined sparse and dense data, preserving categorical features for effective model performance.

Implements walk-forward validation for ARIMA forecasting, ensuring accurate error distribution by refitting on historical data at each step.

Implements advanced simulated annealing with multiple move operators for effective combinatorial optimization.

Utilizes Shapely's STRtree for efficient polygon overlap detection, improving performance from O(n^2) to O(n log n).

Transforms sparse TF-IDF text vectors into dense components for efficient use in gradient boosting decision trees.

Enables day-by-day predictions using Kaggle's iter_test API while managing a rolling history buffer for lag feature computation.

Builds interaction features for retail forecasting by analyzing SNAP event impacts on sales and revenue across states.

Parses structured text fields into separate numeric columns, useful for analyzing sports rosters and inventory data.

Implements a phase-based strategy for game AI, enabling adaptive behavior through cycling between aggressive and economic phases.

Normalizes spatial coordinates in datasets to ensure consistent play direction, enhancing analysis in sports and robotics.

Generates polynomial and interaction features from numeric data to enhance model performance by capturing nonlinear relationships.

Enhances recommendation systems by filling gaps with popular items, improving user experience for cold-start scenarios.

Wraps Kaggle competitive game environments as OpenAI Gym environments for training PPO agents using stable-baselines3.

This skill corrects class imbalance in predictions by rescaling ensemble probabilities, enhancing model calibration for better accuracy.

Rebalances training data by oversampling the majority class to align with test-set class prior, enhancing model calibration.

Enhances model training by using high-confidence predictions as pseudo labels, improving AUC in semi-supervised learning for tabular data.

Ensembles predictions from multiple models by converting to ranks, averaging, and normalizing, enhancing prediction robustness.

Blends predictions from multiple models using rank calibration to ensure accurate and monotonic probability outputs.