Machine learning and data science code repo on GitHub: machine-learning

G2Net Gravitational Wave Detection

Mission: This 2021 Kaggle competition aims to detect GW signals from the mergers of binary black holes, by building machine learning models to analyze simulated GW time-series data from 3 earth-based detectors.

Key features: High-pass filtered signal samples.

Model: 1-D CNN with batch normalization, GeM pooling and Silu activation layers.

GitHub: /machine-learning/tree/master/03_G2Net_gravitational_wave_detection

Exploratory Data Analysis Notebook: [G2Net_gravitational_wave_detection_EDA.ipynb]

Modeling Notebook: [G2Net_CNN1D_GeM_SGDW_Highpass.ipynb]






Web Traffic Time Series Forecasting

Mission: This 2017 Kaggle competition aims at the problem of forecasting multiple time series, the future web traffic for approximately 145,000 Wikipedia articles.

Key features: Seasonality; log1p transformed visits; The medians of target lags.

Model: DNN with bagging. Wavenet.

GitHub: /machine-learning/tree/master/02_web_traffic_time_series_forecasting

Exploratory Data Analysis Notebook: [web_traffic_forecasting_EDA.ipynb]

Modeling Notebook: [web_traffic_forecasting_DNN.ipynb]






M5 Forecasting - Accuracy

Mission: The 2020 Kaggle competition uses hierarchical sales data from Walmart, the largest company by revenue, to forecast daily sales for the next 28 days.

Key features: Grouped explanatory variables; Rolling statistics of target lags.

Model: Ensemble of gradient boosting regression trees for each local ID group.

GitHub: /machine-learning/tree/master/01_m5-forecasting-accuracy

Notebook: [m5_sales_accuracy_ensemble_gbt.ipynb]





Catch Joe

Mission: Given a dataset of users activities of an online system, build a machine learning model that can identify a specific user with user_id=0 (codename Joe).

Key features: Vectorized site visiting records, with TF-IDF N-Gram normalized to keep each user's site visiting order; User demographic features and client systems.

Model: Random forest, gradient boosting, DNN.

GitHub: /machine-learning/tree/master/00_catch_joe

Notebook:





Bootstrap Price Time Series

Mission: Explores time series bootstrapping methods to synthesize financial market price/return paths, which are useful for quantitative strategy development with robustness simulation and backtesting.

GitHub: https://github.com/0liu/bootstrap-price-time-series

Notebook: [Bootstrapping Price Time Series.ipynb]