Car-Evaluation-Analysis

Business Understanding

• Car Evaluation for different features becomes an important tool for recommendation to the users as per their requirements, and needs to be designed as a tool that can help in developing closer relationships between the automobile industry and its customers.
• It is useful in identifying cusotmer bases and loyalties as well as can be used for customer segmentation as per needs and income classes.

Data Understanding

• The data contains 1728 rows and 7 columns with descriptions as follows:
  • buying: Buying Level or Capacity of the customer (Very High:vhigh, High:high, Low:low, Medium:med)
  • maint: Maintainence Level (Very High:vhigh, High:high, Low:low, Medium:med)
  • doors: Number of doors in the car (2, 3, 4 and 5 or more)
  • lug_boot: Luggage Boot Size in the car (small, med, big)
  • safety: Safety Level of Car (high, med, low)
  • class: Unacceptable, Acceptable, VeryGood, Good

Data Preparation

• The data is a classic multi-class classification task with multiple classes of cars available as the final evaluation
• The data does not have any missing values
• The data does not have repetitive columns
• The data is fully categorical with even the final target variable is also in form of a category
• Label Encoding and mapping is used to make continous labels for the majorly ordinal data

Modelling and Evaluation

• Random Forest Classifier tuned with Synthetic Minority Oversampling Technique or SMOTE is used to balance the classes as per need.
• The bias in model is removed and boosts the accuracy to over 96% percent
• Grid Search is performed to tune up with Random Forest to determine suitable number of estimators.
• Logistic Regression fails to converge to a good solution with 81 percent accuracy.
• MLP Classifier also converges to a less good solution
• Decision tree is successful with minimum 10-11 estimators, and fails on over 12 estimators.
• Accuracy Score is chosen as the evaluation metric
• The feature importances for the dataset is as follows with respect to Random Forest Model at 96.5% accuracy:

Deployment

• The app is deployed as a Streamlit Web Application with the following options:

  • Exploratory Data Analysis:
    • 5 point summary
    • Countplot of classes
    • Column Selection
  • Prediction:
  • Logistic Regression
  • MLP Classifier
  • Random Forest Classifier
  • Custom Data Entry

• To run the web app follow the instructions:
  git clone <repo-URL>

  
  

  streamlit run app.py

  • It is also deployed on Heroku for a while at Link