/Cog-State-Classification

Classification solver to determine the present cognitive state of an astronaut based on biosensor data, and predict their future cognitive state(in 3s). TopCoder-NASA Competition

Primary LanguageJupyter Notebook

Cog-State-Classification

Running Instructions

  • please refer to the README.md within /code folder for default running procedures. -v for external data is expected.

Feature Engineering

Missing value handling

  • all sensordata columns are first convert to numeric values and then filled with -9999.9 to eliminating outliers/missing values
  • -9999.9 is then auto handled by the XGBoost model

Data Compose

  • data within 1s timeframe are combined via taking a mean(exluding default values)
  • data are loading by chunks for RAM concerns, duplicates between diffferent chunks have further aggregated

Model Selection

  • mainly applied XGBoost tree model for the multi-class classification
  • model parameters configuration can be found in the src/sample/submission/config.py file
  • T-3s is found always stay the same for given training data, so same prediction model is used
  • Overall Top-Three-Feature is observed and written to the solution file

Possible future work

  • could have investigate the model performance on full training data without merging them, although the merge is able to fillup missing info within the second if it's present eleswhere.