Price sale for housing regressor

The objective of this project is to create an accurate sale price calculation tool for a Real State company, developing in the process a full Data Science Project

Overview

- Exploratory Data Analysis of housing data (Insights):

• Missing values handling
• All numerical variables
• Distribution of numerical variables
• Categorical variables
• Cardinatily of categorical variables
• Outliers
• Relationships between dependent and independent features

- Feature engineering framework:

• Splitting data into train and test set with stratified sampling tecnique to preserve distributions found in the housing data
• Creation of an advanced custom transformation pipeline with helpfull tools such as: median imputer, log normalization, custom column modifiers, rare category handler, NaN dropper, categoric variable encoder and variables scaler. (to clean and transform in a fast way future raw sets)

Note: I could have used ScikitLearn tools such as imputer or Ordinal Encoder, however to add more value to the project and have more control I defined my own cleaning functions

- Feature selection framework:

-Use filter algorithms simple methods such as Chi square and Correlation matrix methods to have an initial visualization.

• Different regression embeded methods will be used to determine feature importance and select the best features to feed our models:
• Linear Regression
• Decision Tree Regression
• Random Forest Regression
• XgBoost Regression
• Permutation Regression

I propose 2 ways of selecting best features, one is simply selecting the top occurrences total sum of the counts, the second one is selecting the the features that appeared as a top in at least 2 different models

- Model selection, training and tunning:

Diferent algorithms to find the one with the best metrics and results for this problem

Models proposed:

• Linear Regressor
• Decision Tree Regressor
• Random Forest Regressor
• Support vector machine regressor
• KG-Boost Regressor
• K-n Neighbor regressor
• Lasso Regressor

- Results summary

A regressor tool for estimating housing sale price was developed from scrath with excelent results, reaching a 88% accuracy in test set. All steps and assumptions were important to reach this result