🤖 Production-Grade AutoML System for Classification & Regression

A comprehensive, clean-architecture Automated Machine Learning (AutoML) platform designed to automate end-to-end machine learning pipelines. Built with a modular Python backend and an interactive Streamlit frontend, this system handles messy, real-world datasets with extreme edge-case resilience and zero-config deployment.

🏗️ System Architecture

Pipeline Flow

flowchart TD
    U(["👤 User"]):::user

    subgraph UI["🖥️ Streamlit Frontend  (app.py)"]
        direction TB
        T1["📤 Upload Tab"]
        T2["🔍 EDA & Issues Tab"]
        T3["⚙️ Preprocessing Tab"]
        T4["🎯 Training Tab"]
        T5["📈 Results Tab"]
        T6["💾 Export Tab"]
    end

    subgraph INGEST["📥 Data Layer  (automl/data/)"]
        direction TB
        DI["DataIngestor\n─────────────────\n• Encoding detection\n• Delimiter heuristics\n• Compression handling\n• BOM & header cleanup"]
        DP["DataProfiler\n─────────────────\n• Summary statistics\n• Missing value ratios\n• Outlier detection (IQR)\n• Class distribution"]
        DV["ColumnValidator\n─────────────────\n• Currency & unit stripping\n• Boolean inference\n• Mixed-type coercion"]
    end

    subgraph PREP["⚙️ Preprocessing Layer  (automl/preprocessing/)"]
        direction TB
        CL["Cleaners\n─────────────────\n• MissingValueHandler\n• OutlierHandler (IQR/Z)\n• ConstantFeatureRemover\n• HighCardinalityHandler\n• RareCategoryCollapser"]
        EN["Encoders\n─────────────────\n• CategoricalEncoder (OHE/Ordinal)\n• NumericScaler (Standard/MinMax)\n• FeatureEngineer (Polynomial)"]
        PB["PipelineBuilder\n─────────────────\n• sklearn.Pipeline orchestration\n• Leakage column removal\n• Duplicate row removal\n• fit_transform / transform"]
    end

    subgraph MODELS["🎯 Model Layer  (automl/models/)"]
        direction TB
        MR["ModelRegistry\n─────────────────\n• Logistic / Linear Regression\n• Decision Tree / Random Forest\n• KNN / Naive Bayes / SVM\n• GradBoost / AdaBoost\n• XGBoost / LightGBM"]
        MT["ModelTrainer\n─────────────────\n• RandomizedSearchCV tuning\n• StratifiedKFold CV\n• Timeout & resource guardrails\n• Best-model selection vs Dummy"]
    end

    subgraph REPORT["📊 Output Layer  (automl/reports/)"]
        RG["ReportGenerator\n─────────────────\n• HTML executive summary\n• EDA findings\n• Preprocessing steps log\n• Model comparison table\n• Next-steps & code snippet"]
    end

    subgraph CFG["🔧 Cross-Cutting  (automl/config/ & utils/)"]
        direction LR
        S["AutoMLConfig\n(settings.py)"]
        E["ErrorHandlers\n(error_handlers.py)"]
        L["Logger\n(logger.py)"]
    end

    U --> T1
    T1 -->|CSV upload| DI
    DI -->|clean DataFrame| DP
    DP -->|profile dict| DV
    DV -->|validated df| T2
    T2 -->|issues + approval| T3
    T3 -->|config choices| CL
    CL --> EN
    EN --> PB
    PB -->|X_processed, y| T4
    T4 --> MR
    MR -->|model configs + param grids| MT
    MT -->|all_results, best_model| T5
    T5 --> T6
    T6 --> RG
    RG -->|automl_report.html| U

    CFG -. "settings & error boundaries" .-> INGEST
    CFG -. "settings & error boundaries" .-> PREP
    CFG -. "settings & error boundaries" .-> MODELS

    classDef user fill:#6366f1,color:#fff,stroke:none,rx:8
    classDef layer fill:#1e293b,color:#94a3b8,stroke:#334155
    classDef mod fill:#0f172a,color:#e2e8f0,stroke:#1e293b,rx:6

Loading

Module Dependency Map

graph LR
    APP["app.py"]

    APP --> DI2["data/ingestion.py"]
    APP --> DP2["data/profiling.py"]
    APP --> PB2["preprocessing/pipeline_builder.py"]
    APP --> MT2["models/trainer.py"]
    APP --> RG2["reports/report_generator.py"]

    PB2 --> CL2["preprocessing/cleaners.py"]
    PB2 --> EN2["preprocessing/encoders.py"]
    PB2 --> DP2
    PB2 --> DV2["data/validation.py"]

    MT2 --> MR2["models/model_registry.py"]

    DI2 --> CFG2["config/settings.py"]
    DP2 --> CFG2
    PB2 --> CFG2
    MT2 --> CFG2
    RG2 --> CFG2

    DI2 --> EH["utils/error_handlers.py"]
    DP2 --> EH
    PB2 --> EH
    MT2 --> EH
    RG2 --> EH

    style APP fill:#6366f1,color:#fff,stroke:none
    style CFG2 fill:#0ea5e9,color:#fff,stroke:none
    style EH fill:#f59e0b,color:#fff,stroke:none

Loading

---

🌟 Core Features & Engineering Highlights

1. 📥 Robust Data Ingestion

• Automatic Encoding Detection: Leverages chardet alongside multi-encoding attempts (utf-8, latin-1, etc.) to read files without Unicode errors.
• Smart Delimiter Detection: Samples file lines and uses statistical consistency (mean/std ratio) to automatically detect commas, semicolons, tabs, and pipes.
• Compression & Format Support: Seamlessly extracts and reads .zip and .gzip files, and automatically detects if an Excel or HTML file is incorrectly disguised with a .csv extension.
• Pre-Clean Safeguards: Detects and strips Byte Order Marks (BOM), removes duplicate/empty column names, and discards repeated headers.

2. 🔍 Automated Profiling & Data Quality Checks

• Exploratory Data Analysis (EDA): Automatic calculation of dataset statistics, missing value ratios, and outlier ratios.
• Visualizations: Matplotlib & Seaborn-powered correlation matrix heatmaps, numerical distribution histograms, and categorical frequency bar charts.
• Intelligent Issue Detection: Identifies critical dataset concerns:
  • Dataset Feasibility: High-dimensionality warnings ($features > samples$) and small sample counts.
  • Class Imbalance: Flags skew in class distribution.
  • Outliers & Cardinality: Detects extreme values and flags high-cardinality columns.
• User-in-the-Loop Fixes: A dedicated Streamlit interface allowing users to review and approve suggested preprocessing fixes.

3. ⚙️ Preprocessing Pipeline

• Dynamic Imputation: Custom strategies for mean, median, mode, or constant imputation.
• Currency & Unit Cleaning: Automatically detects currency symbols and units of measurement (e.g., "kg", "ft", "$") in string columns, parses them, and cleans them into floating point values.
• Categorical Encoders: Smart One-Hot and Ordinal encoding with automatic rare category collapsing.
• Feature Leakage Prevention: Identifies and automatically drops identifier columns (e.g., id, uuid, index).
• Feature Scaling: Configurable Standard and Min-Max scaling.

4. 🎯 Model Registry & Trainer Suite

• Comprehensive Algorithm Support: Fits up to 9 classification and 11 regression algorithms:
  • Linear/Logistic Regression, Random Forests, Gradient Boosting (AdaBoost, XGBoost, LightGBM), K-Nearest Neighbors, SVMs, and Naive Bayes.
  • Baseline Dummy classifiers and regressors are used as control variables to measure exact value add.
• Randomized Hyperparameter Tuning: Automatically optimizes models via randomized grid search cross-validation.
• Resource Guardrails: Enforces time limits (MAX_TRAINING_TIME_SECONDS), limits training rows on large datasets, and manages parallel job counts to prevent memory overflows.

5. 📊 Evaluation Dashboard & Reporting

• Metric Dashboards: Shows sortable comparison tables of training time, accuracy, precision, recall, F1-score, RMSE, MAE, and $R^2$.
• Interactive Visuals: Confusion matrices, class distribution plots, and correlation matrices.
• Automated Report Generation: Generates standalone HTML executive summaries detailing the data quality profile, preprocessing steps, training configurations, and next steps for loading the model.

---

📁 Project Directory Structure

AutoML/
├── app.py                      # Interactive Streamlit Web Interface
├── requirements.txt            # Python Dependencies
├── README.md                   # System Documentation
│
├── automl/                     # Core Backend Framework
│   ├── config/
│   │   └── settings.py        # Centralized system settings and thresholds
│   ├── data/
│   │   ├── ingestion.py       # Robust file ingestion and format resolution
│   │   ├── profiling.py       # Statistical profiling and EDA
│   │   └── validation.py      # Column-level validation & unit/currency cleaner
│   ├── preprocessing/
│   │   ├── cleaners.py        # Missing value, outlier, and variance cleaners
│   │   ├── encoders.py        # Scalers, categorical encoders, and feature engineering
│   │   └── pipeline_builder.py# Scikit-learn Pipeline orchestrator
│   ├── models/
│   │   ├── trainer.py         # Cross-validation, tuning, and evaluation
│   │   └── model_registry.py  # Model configurations and parameter spaces
│   ├── reports/
│   │   └── report_generator.py# HTML report generator engine
│   └── utils/
│       ├── error_handlers.py  # Customized exceptions and error boundaries
│       └── logger.py          # Unified logger settings
│
└── sample_data/               # Pre-bundled datasets (Iris, Titanic, Wine)

---

🚀 Getting Started

1. Installation

Clone the repository:

git clone https://github.com/HaseebUllahButt/AutoML.git
cd AutoML

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate  # On Windows use: .venv\Scripts\activate

Install dependencies:

pip install -r requirements.txt

2. Run the Web Interface

Launch the Streamlit app:

streamlit run app.py

The interface will automatically load at http://localhost:8501.

---

👥 Contributors

• Haseeb Ullah Butt
• Ali Mubashir

---

📝 License

This project is created for educational purposes as part of CS-245 Machine Learning course.