Wine Dataset Classification Pipeline

This project focuses on solving a classification problem with 3 classes using the Wine dataset. The pipeline includes data preprocessing, model building, and performance evaluation steps.

Dataset Description

• Source: UCI Machine Learning Repository
• Problem: Classification with 3 classes.
• Preprocessing: Min-Max scaling and replacing NaN values with the mean value of each class.

Pipeline Steps

1. Data Preprocessing:

   • Min-Max scaling: Normalize the feature values.
   • Replace NaN values: Replace missing values with the mean value of each class.

2. Data Splitting:

   • Divide the data into 10% test and 90% train using stratified sampling.

3. One-Hot Encoding:

   • Encode each class using one-hot encoding:
     • Class 1: [1, 0, 0]
     • Class 2: [0, 1, 0]
     • Class 3: [0, 0, 1]

4. Model Building:

   • Use a Softmax activation in the last layer to train the network to recognize the one-hot encoding.
   • Employ the multi-layer perceptron (MLP) network using TensorFlow and Keras.
   • Perform parameter exploration using tenfold cross-validation design with stratified sampling:
     • Number of layers: [1, 2, 3]
     • Neurons per layer: [32, 64, 128]
     • Learning rate: [0.001, 0.01, 0.1]

5. Performance Evaluation:

   • Use F1 score for performance evaluation.
   • Explore the best parameters for the dataset while reducing bias towards a specific data splitting using cross-validation.

6. Pipeline Implementation:

   • Create a custom pipeline by passing the output of one library function to the next.
   • Add control flow to organize the tenfold cross-validation experiments.

References

• NOMIDL: What is the difference between Sigmoid and Softmax activation function?
• Scikit-learn: Cross-validation