Fraud Detection & EDA on STEG Dataset (Unbalanced Data)

Project Overview

Abstract:

The main goal of this project is to make you feel familiar with Supervised Learning Algorithms like Regression and Gradient Boosting algorithms along with practicing more Exploratory Data Analysis techniques and developing your ML intuition.

Description:

The Tunisian Company of Electricity and Gas (STEG) is a public and non-administrative company, it is responsible for delivering electricity and gas across Tunisia. The company suffered tremendous losses in the order of 200 million Tunisian Dinars due to fraudulent manipulations of meters by consumers. Using the client’s billing history, the challenge aims to detect and recognize clients involved in fraudulent activities. The solution will enhance the company’s revenues and reduce the losses caused by such fraudulent activities.

Data Problem:

Building a model that will help classify which customer is likely to commit fraud by that saves the company from making losses.

Dataset

Data Issues:

• Imbalanced Dataset with around 20% for class 1 and 80% for class 0
• Training set and test set each have a different distribution
• The dataset is not normally distributed
• Training and test sets have different amounts of each class in it as training includes around 8% of class 1 and 92% of class 0 while the test set includes 50% for each class

Data Preprocessing

• Removing mistaken values, null values, and useless features like client ID, etc
• Normalizing the following features as they’re the ones that need normalization:
  • Consommation_level_1
  • Consommation_level_2
  • Consommation_level_3
  • Consommation_level_4
  • Old_index
  • New_index
• Merge training and testing sets and then split them again to solve different distribution issues
• Apply undersampling to solve imbalance on the whole dataset issue
• Applying label smoothing to improve model performance potentially.

Results

We tested a list of different model and different types, the models we tried were:

• Gradient Boosting Classifier
• Decision Tree Classifier
• Logistic Regression
• Cat Boost Classifier
• Ada Boost Classifier

The results determined that the best model was Decision Tree Classifier, it achieved a results as follow:

• Precision 0.849
• Recall 0.852
• Accuracy 0.850