In this notebook, I use the following approach to generate predictions on the Kaggle House Price Dataset:

  • Build a simple Sequential model in Keras/Tensorflow
  • Use the Weights and Biases (WandB) platform to select optimal hyperparameters and record experiments.
    • Experiments are evaluated using K-Fold Cross Validation. Mean RMSE across folds for each experiment are custom logged in WandB.
  • Make predictions.
  • Blend predictions from the previous post (XGBoost, LightGBM) and this post (Neural Net).
    • By taking the mean of predictions, taking the weighted mean of predictions, and defining a meta-model.
    • Predictions are evaluated on a holdout dataset, kept separate from the training set since the beginning of the project.