Data Analysis on Ames Housing Sale Price Prediction
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
The purpose of this project was to build an accurate model to predict sale prices for houses in the city of Ames, Iowa (Fig. 1). Using a dataset provided from a Kaggle competition, our team, SurrealEstates, conducted an in-depth exploratory data analysis (EDA), data pre-processing, and model building to accurately predict the sale price of a house based on certain variables. The data set contained about 81 features that accounted for various house characteristics, garage characteristics, and many other attributes that may influence the sale price of a home in Ames, Iowa.
We generated a complete process of machine learning from EDA to prediction. Using various Python modules and scikit-learn, we solidified our understanding of data manipulation, model selection, and model improvement. The source code is available in our Github repository.
In this specific project we sought to optimize for root mean square error (RMSE), as shown in Fig. 2. In order to select the correct inputs, we focused our research along three paths: data manipulation/EDA (exploratory data analysis), model selection/search, and refining of our models. Each of these three paths allowed us to better understand and adjust our model to reduce the RMSE.
The EDA phase allowed us to not only understand the composition of the data, but also to understand how the data related to each other. We looked at missing data and imputed it depending on both the type of missingness and the type of data. We worked through different methodologies, as well, when refining our model. Those iterations allowed us to see how the different types of imputation changed the RMSE.
The RMSE was also significantly impacted by our model selection and model search. We used a combination of grid search and ensembling to refine, test, and validate our results. We worked through multiple hyperparameters with each of our models. When the best model was chosen, we used ensembling to weight those models and determine the best overall score for the prediction. Multiple pass throughs and changes in our tuning methodologies allowed us to continue to decrease our RMSE.
We began the process of machine learning by exploring the linear relationships between sale price and the continuous input variables in the dataset. We produced a correlation matrix (see Fig. 3) to give a visual snapshot of the magnitude of these relationships.
We observed strong linear relationships between our outcome, sale price, and overall quality, great living area, number of cars a garage can manage, the area of the garage, total basement square feet, first floor square feet, number of full baths, number of rooms above ground, and year built. The relationships were further analyzed with scatter plots to further observe the observations in each variable and look at possible outliers that may be biasing correlation values.
Many outliers were observed within the continuous variables. Instead of removing just based on the scatterplot visualizations, outliers were systematically removed by using z-scores. The z-score method identifies outliers by finding the relationship with the standard deviation and mean of a given group of data points and scales them to follow a normal distribution where the standard deviation = 1 and mean = 0. Outliers are considered to be the observations furthest from the mean. In practice, the threshold for removal of outliers is if a data point is greater than 3 standard deviations away from the mean or less than -3 standard deviation of the mean.
For example, two outliers were removed from the great living area feature, shown in Figure 4.
Once the exploration was complete, the train and test data from Kaggle were merged into a larger set for further engineering. The variable ‘Saleprice’ was saved as its own separate data frame and removed from the larger dataset.
The merged dataset showed an abundance of missing values, as shown in Figure 5.
Many variables had more than 50% of their values missing. These values were imputed based on the nature of the variable. Some variables, such as PoolQC, MiscFeature, Alley, Fence, and FireplaceQu that contained a large NaN count, had meaning in the NAN values. For these variables, the NaNs indicated the lack of the feature at a particular house for sale. These NaN values were important to the analysis and were imputed with 'None.' Other imputation methods included: imputing with zeros (float type variables), the median (float type variables), and the mode (categorical variables).
Reduce skewness, Outcome
After imputation was complete, we set off to normalize all of the continuous variables to be included in the machine learning model to reduce any possible bias in the analysis.
Skewness is asymmetry in the distribution of numerical distribution in which the curve appears distorted either to the left or to the right. Skewness can be quantified to define the extent to which a distribution differs from a normal distribution. Sale price was the first to be normalized. A density plot of sale price showed that the outcome was skewed to the right and had some deviations from the line on the probability plot, as shown in Figure 6. The extremely skewed data would introduce bias in the results of our model. So we decided to use the logarithm of sale price to train our mode in order to reduce bias. In Fig. 7, the curve after taking logarithm is much closer to a normal distribution and shows little deviation from the line in the probability plot.
Reduce skewness, Inputs
Many of the continuous features were skewed in the data set. Instead of taking the logarithm of all of these features, the box cox method of normalization made all the difference in the success of our model.
In order to normalize all numerical features, we applied box-cox method built in Python on the features with skewness greater than 0.75. The equation of box-cox is shown in Fig. 8. The ideal lambda value of each numerical feature was obtained by using the built in function in Python. Fig. 9 shows the distribution of garage area before and after box-cox transform.
Add additional features
At the end of our data pre-processing, the categorical features were dummified in order to pass them through the machine learning model, and a couple of new features were introduced into the data. Two variables, ‘TotalSF’ and ‘Total_Bath’, were engineered by aggregating all of the features that quantify the square feet of a home and features containing the number of bathrooms in a house. Furthermore, ‘hasbasement’ and ‘has2ndfloor’ were engineered by turning the ‘2ndFlrSF’ feature and ‘BsmntSF’ feature into binary variables. All features used to make the new variables were dropped from the larger data set to reduce the dimensionality of the data.
Fig. 10 shows the workflow of model fitting. After finishing the preliminary data preprocessing, we first fitted the basic linear models (Ridge, Lasso, ElasticNet). We used the grid search cross-validation method to find the best model and parameters with full data. Because tuning hyperparameters of basic models were fast compared to high-end models, it allowed us to refine the preprocessing part with less time. We then split the data into training and testing data set and compared the root mean square error between them. This procedure helped us have a better idea of how models performed when they faced unknown data.
We tried different combinations of processes in the preprocessing part and fitted models to examine which combination gave us the smallest test RMSE and Kaggle score. We found that reducing the skewness of the numerical feature played a critical role in getting a smaller RMSE. This results showed that normalized numerical features would improve the performance of the linear models.
After refining the preprocessing procedure with basic linear models, we used the grid search method to tune the hyperparameters of the gradient boost model (GBM), XGboost (XGB), and support vector regressor (SVR). In the end, we used averaged models by putting different weight on the predictions to improve the housing sale price predictions further, as shown in Fig. 11.
Fig. 12 displays the results of each model. The GBM and XGB had higher R^2 values (score_grid), and lower train RMSE compared to other models. However, they had the highest test RMSE among all models. These results clearly showed that these two models overfitted the data. As a result, they had higher Kaggle RMSE score compared to other models.
As can be seen in Fig. 12, the simple linear models performed better than those "high-end" models based on the Kaggle score. There are fewer hyperparameters of simple linear models, so it is easier to find their best model. We might need to tune hyperparameters of those more complex models better since there are more hyperparameters for the GBM or XGB. However, this requires more study on the influence on different hyperparameters, which can be done in the future.
As mentioned in the previous session, we used the averaged model to improve the model predictions further. The best combination of weight of each model is 0.25*Ridge + 0.2*Lasso + 0.2*ElNet + 0.15*XGB + 0.2 *GBM, which gave us the best Kaggle score 0.11653. This score put us at the top 16% among all teams in this Kaggle competition.
Figure 13 demonstrates the top ten crucial features from the GBM and XGB models. The feature importance can help us understand which variables have more influence (either positive or negative) on the housing sale price.
GBM showed that the living space, the quality of the house, year built, garage area, and the numbers of bathrooms were essential features. This result was consistent with the top ten correlated variables with the sale price. However, there are some discrepancies with the top ten critical features between XGB and GBM. XGB shows that the size of the garage is the most vital feature, which is unexpected. This result might indicate that more hyperparameter tuning is required for the XGB to obtain a more accurate prediction.
We continued to refine our models until we reached a peak score of 0.11653, which placed us at top 16% in this competition. This score was due to the choice not only of the individual models and hyperparameters, but also on the chosen weights for ensembling our models. We determined a few key factors which reduced our RMSE.
The most important areas we determined for reducing the model consisted of: method of imputation, removal of outliers, data normalization, feature engineering (adding/removing), and hyperparameter tuning. While we did ensemble our different models, our best stand alone model was ridge.
There are many ways in which we would choose to further explore this data and change our models in order to continue to reduce our RMSE. We could continue to tune our models, add more models, or even change our methodologies on imputation and normalization. Another option would be to do additional feature engineering.
The additional model tuning could be approached in a few different ways. We would like to add additional models outside of the ones we have already tested. Since we did not find the best hyperparameters for XGB and GBM, we would continue to iterate with our grid search to find higher performing values. Adding additional values from outside sources such as the Census Bureau, US Fed, and other economic indicators would also allow us to predict the purchase cost of homes.