Automobile dataset - https://archive.ics.uci.edu/ml/datasets/Automobile
-
Listing as many use cases for the dataset as possible.
- Predicting Prices for the given Car/Truck
- Classifying Symboling for the given Car/Truck : To figure out the Car/Truck is safe or not
- Finding the Normalized losses for any given Car/Truck, i.e., average loss on each car every year
- Optimizating the best configuration of cars to yield minimum loss
-
Picking one of the use cases listed above and describing how building a statistical model based on the dataset could best be used to improve it - I have picked up the Symboling classification problem
- The assigned symbol to the car has been partitioned in 2 categories, 0 and 1.
- 0 for cars with riskiness (i.e., Symboling value greater than 1), and 1 for cars with lesser riskiness (i.e., Symboling value lesser or equal to 0)
- Various features given have been segmented as continuous and categorical variables, and are used for the classification model
- XGBoost model has been used finally for classification approach
-
Implementing the model described above in Python. The code is retrieving the data, train and test a statistical model, and reporting relevant performance criteria.
- Model described above has been implemented in Python, in 3 separate files
-
Explaining each and every design choices (e.g., preprocessing, model selection, hyper-parameters, evaluation criteria). Comparing and contrasting choices with alternative methodologies.
- All the steps have been explained in detail along with the code. Below is a snapshot of the same
- The code has been divided in 3 files :
a) Data_Preprocessing_1.ipynb : - Data Loading - Segmentation of Continuous and Categorical variables - Imputation for Continuous and Categorical variables - Categorical variables split into multiple cols - Train data and Target variable are saved
b) Feature_Selection_2.ipynb : - VIF checks are made for all variables - Variable Importance is recorded for all variables - ANOVA F-value for Continuous variables, and Chi2 P-value for Categorical variables used to select variables - Above stats for all variables are stored in auto_data.csv (or, auto_data_xls.xls) file, and variables are shortlisted - Shortlisted Continuous and Categorical variables are saved
c) Model_Tuning_Minimizing_CV_LogLoss_3.ipynb : - Above saved Train, Target, Shortlisted Categorical, Shortlisted Continuous variables are loaded - XGBoost model with default parameters is fitted on the above dataset - Model parameters are tuned with aim of Minimizing LogLoss - all hyper parameters are stored simultaneously in separate files - Final model with best hyper-parameters is saved - Model is tested on Training data itself, and shows promising results across all the top decile
- Improving the model made above if you had more time.
- Using other classification models, like Logisitic, RandomForest, Ensemble Approach
- Creating more features for the given set of Car/Truck
- Using better techniques for the imputation of the Normalized Losses, Price and other variables