andremarco/house-price-advanced-regression-techniques

Kaggle competition

House-price-Advanced-Regression-Techniques

Kaggle competition

Project made by:

• Davide Aureli
• Valerio Guarrasi
• Andrea Marcocchia

Kaggle account -- The PODS --

Cleaning Data-1: reformatted character variables to numeric, without grouping them. So each observations has its respective value. Cleaning Data-2: drop off all the variables that have been numeric and are no longer needed. Cleaning Data-3: replace NA values using:

• 0 -> if NA meant that the feature doesn’t exist
• mode -> if the feature is discrete and NA is a lack of information
• mice package with classification and regression trees method (cart)

Feature engineering-1: created new features based on correlation between variables in train dataset Feature engineering-2: removed outliers following GrLivArea.

All the operations above are done for train and test dataset.

Model selection-1: Linear Model:

• standardize all the variables except dicotomical variables and SalePrice
• apply logarithm to SalePrice
• estimate the linear model using “backward elimination” algorithm to choose the variables to removeà score = 0.12372

Model selection-2: XGBoost:

• transform the train and test dataframes in Sparse Matrix Object
• estimate the XGBoost model with parameter declared in paramà score = 0.12047

Model selection-3: LASSO:

• use the cross-validation method to estimate the parameters
• estimate the LASSO model with parameter estimated beforeà score = 0.12106 Model selection-4: Evaluated weighted mean between the three previous methods with the respective weights: 10%, 68% and 22%.

Kaggle score: 0.11624