GST Analytics Hackathon: Developing a Predictive Model in GST

Overview

This repository contains the implementation of a predictive model for the GSTN Hackathon analytics challenge. The goal was to construct a model Fθ(X) → Ypred that accurately estimates the target variable Yi for unseen inputs Xi. This project focused on advanced data preprocessing, handling imbalanced datasets, and developing an optimal neural network solution.

---

Problem Statement

Dataset Provided:

• Training Data (Dtrain): Matrix of dimension R(m × n)
• Test Data (test): Matrix of dimension R(m1 × n)
• Target Variables:
  • Ytrain: Dimension R(m × 1)
  • Ytest: Dimension R(m1 × 1)

Objective: Develop a model that accurately predicts Ypred for unseen test data.

---

Data Preprocessing

1. Dropping Unnecessary Columns:

   • Removed ID (irrelevant for predictions).
   • Dropped Column9 due to excessive missing values.

2. Selective Row Removal:

   • Dropped rows where the target variable was 0 and contained null values in key columns.

3. Handling Missing Values:

   • Applied K-Nearest Neighbors (KNN) Imputer to fill missing values based on similar data points.

4. Balancing the Dataset:

   • Used random sampling to create a balanced training set, addressing class imbalance for better model performance.

---

Model Training

1. Neural Network with MLPClassifier

• Architecture:
  • Hyperparameters tuned using Grid Search.
  • Explored configurations of hidden layers, activation functions, and solvers.
• Results on Test Data:
  • Accuracy: 97.58%
  • Precision: 0.83
  • Recall: 0.93
  • F1-Score: 0.88
  • AUC-ROC: 0.99

2. Cross-Validation with Stratified K-Fold

• Ensured balanced class distribution in each fold.

• Mean Results on Training Data:

  • Accuracy: 98.20%
  • Precision: 0.92
  • Recall: 0.95
  • F1-Score: 0.93

• Test Data Results:

  • Accuracy: 97.58%
  • Precision: 0.83
  • Recall: 0.93
  • F1-Score: 0.88
  • AUC-ROC: 0.99

3. Grid Search Optimization

• Fine-tuned hyperparameters including hidden layer sizes, activation functions, solvers, and learning rates.
• Improved Results on Test Data:
  • Accuracy: 97.60%
  • Precision: 0.84
  • Recall: 0.93
  • F1-Score: 0.89
  • AUC-ROC: 0.99

---

Key Evaluation Metrics

1. Accuracy (97.58%)

   • The model correctly classified 97.58% of test instances.

2. Precision (0.84)

   • 84% of instances predicted as positive were true positives, reducing false positives.

3. Recall (0.93)

   • 93% of actual positive cases were correctly detected, minimizing false negatives.

4. F1-Score (0.89)

   • Balances precision and recall, reflecting overall performance.

5. AUC-ROC (0.99)

   • Indicates excellent model capability in distinguishing between positive and negative classes.

---

Conclusion

• The Neural Network model, enhanced with Stratified K-Fold Cross-Validation and Grid Search, outperformed other approaches (e.g., linear and logistic regression).
• Achieved high performance metrics, making it robust for deployment in GST-related predictive analytics.

---

How to Run

1. Clone this repository:

   git clone https://github.com/raged-pineapple/GST-Analytics-Hackathon.git