Ensemble Approach for Zero-Shot Learning

Project Overview

This repository contains the implementation of my thesis project, which focuses on enhancing zero-shot learning models through an innovative ensemble approach. This methodology aims to improve model generalization to unseen classes, a common challenge in machine learning, particularly in real-world applications.

Detailed Methodology

1. Base Model Training (Model A)

• Objective: Train Model A with all available training data.
• Process:
  • Identify instances predicted correctly (pos) and incorrectly (neg).

2. Weakness Focused Training (Model B)

• Objective: Train Model B specifically on the neg instances from Model A's predictions to address its weaknesses.
• Process:
  • Focus on instances where Model A's predictions were incorrect.

3. Transductive Inference

• Objective: Use both Model A and Model B to predict labels for all data, including unseen data.
• Process:
  • Evaluate pseudo-labels from both models, considering higher confidence scores as more reliable.

4. Discriminator Training

• Objective: Train a discriminator to choose between Model A and Model B based on reliability.
• Process:
  • Virtual labels are assigned based on which model predicts correctly:
    • Label A if only Model A is correct.
    • Label B if only Model B is correct.
  • Ignore instances where both models agree on incorrect predictions.

5. Instance Evaluation and Partitioning

• Objective: Use the discriminator's evaluations to partition data for further refinement.
• Process:
  • Assign instances to Model A or Model B based on the discriminator's decision.

6. Model Refinement

• Objective: Retrain Models A and B on newly partitioned data sets to enhance predictive accuracy.
• Process:
  • Continuously improve the models based on discriminator feedback.

Transductive Discriminator

• Objective: Mitigate bias and domain shift problems, particularly with unseen data.
• Process:
  • Use reliability metrics for each pseudo-label to assess confidence.
  • Assign virtual labels based on which model's pseudo label is deemed more reliable.

Files and Usage

• ensemble.py: Contains the full implementation of the zero-shot learning models and the transductive approach.
• To run the project, ensure that you have Python 3.x installed along with the necessary libraries listed in requirements.txt.

Requirements

• numpy
• pandas
• sklearn
• imblearn
• tensorflow

Please install the required Python packages using:

pip install -r requirements.txt