How abrar2652/qamplifynet works
Project type and purpose
• This is a research‐oriented, multi‐pipeline machine‐learning framework for supply chain backorder prediction.
• It orchestrates data ingestion, preprocessing, benchmarking of classical models, a deep Q‐learning agent, quantum‐classical hybrid networks (QAmplifyNet), quantum ensembles, and pure quantum neural networks.
File structure
• Top‐level modules:
• Preprocessing – data cleaning, feature‐selection and outlier‐detection notebooks and results
• Classical – 10 classical ML model experiments (ANN, CatBoost, RF, XGBoost, etc.)
• DDQN_RL – deep Q‐learning approach
• QAmplifyNet – the hybrid quantum‐classical neural network core
• QEnsembles – stacking ensembles combining quantum circuits with classical meta‐learners
• QNN – end‐to‐end quantum neural network experiments (CNN+encoder+QNN, MERA, RY‐CNOT)
• Each module contains Jupyter notebooks, artifacts (plots, checkpoints), and subdirectories for detailed experiments.
README for architectural cues
• The README calls out 4 benchmark categories: classical, QAmplifyNet, Q‐ensembles, RL.
a. Main components
• Data Ingestion & Storage: Kaggle backorder dataset (CSV files)
• Preprocessing Module: notebooks implementing outlier detection (IFLOF, IQR), scaling, encoding, VIF analysis
• Modelling Pipelines:
– Classical Models (10 algorithms)
– Deep Reinforcement Learning (DDQN)
– Hybrid Quantum‐Classical (QAmplifyNet core)
– Quantum Ensembles (stacking of VQC, qSVM, etc.)
– Pure Quantum Neural Networks (CNN+QNN, MERA, RY‐CNOT)
• Evaluation & Reporting: common metrics calculators, confusion matrices, ROC/PR plots, LIME/SHAP explanations
• Artifacts Storage: generated plots, model checkpoints, tensorboard logs
b. Relationships & interactions
– Raw Data flows into Preprocessing
– Preprocessed data is branched to all modelling pipelines in parallel
– Each pipeline produces evaluation artifacts and logs
– A final benchmarking/aggregation step compares results across pipelines
c. Architectural patterns & principles
– Modular pipeline per experiment type (separation of concerns)
– Reusable preprocessing steps feeding multiple models
– Experiment isolation: each model lives in its own subdirectory with artifacts
– Hybrid quantum‐classical integration via PennyLane or Qiskit (in QAmplifyNet)
d. Technologies & frameworks
– Python, Jupyter notebooks
– Scikit‐learn (classical models, preprocessing)
– LightGBM, XGBoost, CatBoost
– TensorFlow/PyTorch (possibly for QNN encoder/CNN)
– PennyLane/Qiskit for quantum circuits (QAmplifyNet, QEnsembles, QNN)
– OpenAI Gym or custom code for DDQN
– LIME/SHAP for interpretability
– TensorBoard for training visualization