Unlocking Home Value
Leveraging Kitchen Quality and Neighborhood Insights for Strategic Renovations
According to NerdWallet, one of the top ways you can increase the value of your home is with a kitchen renovation. For this project, we cosplayed as a kitchen renovation company in Ames, Iowa during the years 2006-2010. Given the nature of that business, it's crucial for us to find neighborhoods where homebuyers value kitchen quality. A higher demand for high-quality kitchens will likely result in more investment in upgrading kitchens, in turn, increasing property values. In this blog post, we will discuss how we utilized the Ames Iowa housing dataset, a random forest model, and SHAP values to identify the impact of kitchen quality on house prices and identify potential neighborhoods for our services. We will also explore the questions of whether certain neighborhoods value kitchen quality more than others, and if this preference is reflected in the sale prices of the houses.
Data and Methodology
The Ames Iowa housing dataset comprises 2,580 properties with 81 features, including the KitchenQual feature that describes kitchen quality. The provided dataset pertains to the sale of residential properties in Ames, Iowa spanning from 2006 to 2010. The dataset comprises 2930 records and encompasses a vast range of explanatory variables, including 23 nominal, 23 ordinal, 14 discrete, and 20 continuous variables, all of which are crucial in evaluating the worth of homes.
The original dataset contained missing data in 27 columns. Some were missing just one value, and others could be missing over 2,000 values. To handle these missing values, we first filled in nominal data gaps for features like Garage Condition or Basement Quality with strings like "No_garage" or "No_Basement". For the MiscFeature column, which represents additional items such as sheds, we subtracted the corresponding MiscValue from the SalePrice and then removed the MiscValue column.
We also addressed columns with highly imbalanced distributions by retaining only those that added value to our analysis, such as KitchenAbvGr (number of kitchens not in the basement). In total, we dropped three observations: two outlier houses with over 4,000 sq ft and one house in the Landmark neighborhood.
Feature Engineering and Collinearity
To enhance our analysis, we converted two numeric columns (MSSubClass and Month) to categorical variables and combined similar features to create new aggregated features, such as OutdoorSF and TotalBath. We also added Age and RemodAge features to represent the age of the house when sold and the number of years since it was last remodeled at the time of sale, respectively.
Ordinal data represented as string values (Such as “Poor” or “Fair”) was encoded with numeric values (Such as 0 or 2) to facilitate analysis. However, to maintain ordinality, we made sure that they were encoded with values that maintained the implied direction of the strings For instance, we converted the values from variables “KitchenQual”, “ExterQual”, and “HeatingQC” to the below values:
- “Poor” → 0
- “Fair” → 1
- “Average” → 2
- “Good” → 3
- “Excellent” → 4
To ensure equal representation of neighborhoods in the train/test datasets, we stratified by neighborhood. We also removed highly correlated columns, such as GarageArea, to minimize multicollinearity and focus on the most relevant features for our model.
Model Selection and Tuning:
We experimented with various linear and tree-based models to predict house prices. To get a baseline score for each mode, we performed some baseline model tuning including data normalization, dummification, and lambda value tuning. After evaluating the performance of different models, including simple linear regression, multiple linear regression, Ridge, Lasso, ElasticNet, Gradient Boosting, and Random Forests, we quickly realized that linear based models were not as good a fit for these data because they had trouble capturing non-linear relationships and handling outliers. Many Features in this dataset broke the assumptions of linearity; consequently, these models struggled with performance.
In contrast, the tree-based models such as Gradient Boosting and Random Forests show more potential in their scores. Tree based models have a strong ability to handle non-linear relationships, higher dimensional data, and outliers. Instead of relying on a linear relationship between features, these two models create decision trees that recursively split the data into smaller subsets based on the most significant features, until a stopping criterion is reached.
We ultimately determined that tree-based models were more suited to our dataset and moved forward with a Random Forest Regressor, which achieved a performance score of 91.86% after hyperparameter tuning.
Using the Random Forest Model with SHAP
SHAP (SHapley Additive exPlanations) is a powerful method for interpreting the output of machine learning models, particularly useful for complex models like random forests. SHAP values help us understand the contribution of each input feature to the final prediction of sale prices by fairly distributing the prediction among the features.
In the context of our analysis, positive SHAP values indicate that a feature has increased the predicted sale price, while negative values imply a decrease. It is important to note that SHAP values are model-agnostic and locally accurate, meaning they give precise explanations for each individual prediction made by the model.
When interpreting individual SHAP values, it's essential to exercise caution as they only provide an estimation of the feature's true importance. These values serve as a useful guide but may not capture the full complexity of the relationships between features and their contributions to the model's predictions. However, by using SHAP values as a tool to understand the impact of various features on the model's output, we can gain valuable insights into the factors that drive house prices and the role of kitchen quality in different neighborhoods.
To help visualize the contribution of each feature to the final prediction for a specific instance, we used SHAP's waterfall plot. This plot displays the SHAP values for each feature, giving a clear picture of how each feature contributes to the prediction for a given instance.
Results and Insights:
Our analysis of SHAP values revealed that high-quality kitchens had a significant positive impact on house prices, indicating that homebuyers in the Ames Iowa dataset valued upgraded kitchens.
By identifying neighborhoods where the KitchenQual feature has a substantial influence on property values, we can target our kitchen renovation services more effectively and help homeowners maximize their return on investment.
We also found that features in our dataset were not independent of each other, and changing one feature affects the contribution of others. This interdependence is common in most datasets, where relationships between features can be non-linear or interdependent. For example, upgrading a kitchen might reduce the negative impact of a home's age on the sale price, as buyers might perceive the house as more up-to-date and well-maintained despite its age.
Case Study: Kitchen Renovation in Different Neighborhoods
We analyzed two examples of houses in different neighborhoods to demonstrate the impact of kitchen quality on sale prices. In Northridge Heights, a neighborhood where kitchen quality is highly valued, upgrading a kitchen from average to excellent quality increased the overall predicted sale price by $34,795.
In contrast, in Sawyer West, a neighborhood where kitchen quality is not valued as much, upgrading the kitchen only increased the predicted sale price by $306. This highlights the importance of focusing on neighborhoods where kitchen quality is valued to maximize returns on renovation investments.
Utilizing the Ames Iowa housing dataset, a Random Forest model, and SHAP values, we were able to identify the impact of kitchen quality on house prices and highlight the importance of targeting neighborhoods where homebuyers value high-quality kitchens. This approach allows our kitchen renovation company to better focus its services on areas with the highest potential for return on investment, ensuring that homeowners can reap the benefits of their upgrades.
Neighborhoods like Northridge Heights had the potential to increase sales price up to 8% while neighborhoods like Sawyer only saw a maximum of around a 1% increase in sales price.
The key takeaway from our analysis is the importance of the neighborhood in determining the value of a kitchen renovation. We found that certain neighborhoods place a higher value on kitchen quality, implying that future buyers in these areas are likely to appreciate and prioritize high-quality kitchens. This understanding allows us to focus on finding the right houses to renovate, maximizing the return on investment for both our company and the homeowners we serve.
To further enhance the accuracy and applicability of our model, we plan to incorporate data on recent kitchen renovation permits and the latest house sale figures. This additional information will allow us to validate our model's predictions and gain a deeper understanding of the return on investment for kitchen renovations in different neighborhoods.
Additionally, future work could involve exploring the relationship between the importance of kitchen quality and the characteristics of homebuyers in specific neighborhoods. It would be interesting to investigate whether certain factors, such as income levels or lifestyle preferences, influence the value placed on kitchen quality. For example, do wealthier individuals prioritize high-quality kitchens because they prefer to avoid dealing with renovations when moving in, or do people with limited budgets consider kitchen quality less important when making purchasing decisions? By asking the right questions and delving into these relationships, we can refine our understanding of the market and better serve our customers, ensuring that they make well-informed decisions when it comes to upgrading their kitchens and maximizing their property value.
Thanks to our mentors for this project Jonathan Presley, Kyle Gallatin, Vinod Chugani, and everyone at NYCDSA