Car Price Prediction by Using K-nearest-neighborhood (KNN) Regression Model

• Objective: Explore the best numeric feature subset and K-value for car price prediction.

Dataset is from 1985 Ward's Automotive Yearbook
Available from: https://archive.ics.uci.edu/dataset/10/automobile

Here I implemented:

1. Data Cleansing and Data Transformation
2. Univariate KNN model with single K-value
3. Univariate KNN model with multiple K-values (Hyperparameter Tuning)
4. Multivariate KNN model with single K-Value
5. Multivariate KNN model with multiple K-Values (Hyperparameter Tuning & K-Fold Cross-Validation)

Conclusion

• Best Feature Subset: ['city-mpg', 'wheel-base', 'curb-weight', 'highway-mpg', 'peak-rpm']
  Best k Value: 1
  Best Average Accurcy: 86.43%

Future Step

• With the multivariate KNN hyperparameter tuning, we used f_regression scoring function for best feature subset selection.
  However, f_regression only examines the linear relationship between features and target, and return p-value.
  
• We can further apply Mutual Information function from Scikit-learn for feature selection, to see if non-linear results can make the prediction better.
• We only used numerical features for prediction. It is interesting to explore how combinations of categorical and numeric features ​​can achieve better accuracy by implementing One-Hot Encoding.