ML Playground, Auto‑Detect, Clean & Future‑Proof

An interactive Streamlit + Scikit‑Learn web application that allows anyone to upload a dataset, automatically identify whether the problem is classification or regression, train multiple ML models interactively, visualize the results, and export the trained model, all from a single browser interface.

---

Table of Contents

1. Overview
2. Architecture
3. Key Features
4. Data Flow Diagram
5. Design Decisions
6. Preprocessing Pipeline
7. Supported Models
8. Metrics Explained
9. User Guide
10. Deployment Guide
11. Troubleshooting
12. Performance Tips
13. Roadmap

---

Overview

ML Playground provides a practical and educational environment for machine learning experimentation. It bridges the gap between Jupyter notebooks and full production pipelines, ideal for data analysts, ML students, educators, and developers who want to quickly prototype or demonstrate machine learning behavior on real data.

Goals:

• Automate tedious parts of ML workflow (data cleaning, encoding, scaling)
• Detect task type and prevent invalid configurations (classifier on continuous targets)
• Provide instant, intuitive visual feedback
• Ensure full forward and backward compatibility with future Streamlit and Scikit‑Learn versions

---

Architecture

System Flow

flowchart TD
A[CSV Upload / Sample Data] --> B[Preprocessing Layer]
B --> C[Task Detection Engine]
C --> D[Model Selector]
D --> E[Training Engine]
E --> F[Metrics & Visualization]
F --> G[Export / Download]

Loading

Component Breakdown

Component	Purpose
Frontend (Streamlit)	Renders user interface, handles file uploads, parameter selection, and visualization.
Task Detection Engine	Analyzes target column distribution to classify problem type (classification vs regression).
Preprocessing Layer	Builds ColumnTransformer with numeric scaling and categorical one-hot encoding.
Training Engine	Constructs scikit-learn Pipeline, trains model, evaluates metrics.
Visualization Engine	Uses Plotly for interactive confusion matrix, ROC, and residual plots.
Export Layer	Serializes the trained pipeline using Joblib for reuse in production scripts.

---

Key Features

Auto Task Detection

Automatically determines if the target column represents a classification or regression task using data type and unique value heuristics.

Model Management

Select from built‑in algorithms:

• Classification: Logistic Regression, Random Forest, XGBoost (optional)
• Regression: Linear Regression, Random Forest, XGBoost (optional)

Visual Insights

• Real‑time metrics panel
• Confusion matrix & ROC for classification
• Residual and importance plots for regression

Model Persistence

Save the full pipeline as .joblib for production deployment.

Backward Compatible

Supports all versions of scikit‑learn 0.24+ by auto‑detecting encoder parameters (sparse_output vs sparse).

---

Data Flow Diagram

sequenceDiagram
    participant User
    participant StreamlitApp
    participant Preprocessor
    participant Model
    participant PlotlyCharts

    User->>StreamlitApp: Upload CSV / Choose sample
    StreamlitApp->>Preprocessor: Analyze data types
    Preprocessor->>Model: Provide processed tensors
    Model->>StreamlitApp: Return predictions + metrics
    StreamlitApp->>PlotlyCharts: Render visuals interactively
    User->>StreamlitApp: Download model.joblib / predictions.csv

Loading

---

Design Decisions

Decision	Rationale
ColumnTransformer-based preprocessing	Guarantees clean numeric and categorical handling without data leakage.
Streamlit caching (@st.cache_data)	Improves performance on repeated dataset loads.
Pipeline encapsulation	Ensures model reproducibility and allows joblib export.
Responsive Plotly charts	Future-proof visualization compatible with Streamlit 2025+.
Auto-hyperparameter UI	Simplifies model experimentation for non-programmers.

---

Preprocessing Pipeline

Numeric features:

• Standard scaling (zero mean, unit variance)

Categorical features:

• OneHotEncoder (sparse_output for sklearn ≥1.2 or sparse=False otherwise)

Pipeline Example:

preprocess = ColumnTransformer(
    transformers=[
        ("num", StandardScaler(), numeric_features),
        ("cat", OneHotEncoder(handle_unknown="ignore"), categorical_features)
    ]
)
model = RandomForestClassifier()
pipe = Pipeline([("preprocess", preprocess), ("model", model)])

---

Supported Models

Classification

Model	When to Use	Key Parameters
Logistic Regression	Simple linear separable data	C, max_iter
RandomForestClassifier	Nonlinear problems, tabular data	n_estimators, max_depth
XGBClassifier	Large-scale data or high-dimensional	learning_rate, n_estimators, max_depth

Regression

Model	When to Use	Key Parameters
Linear Regression	Simple linear relationships	None
RandomForestRegressor	Nonlinear regression	n_estimators, max_depth
XGBRegressor	Complex relationships, feature interactions	learning_rate, max_depth

---

Metrics Explained

Metric	Formula	Task	Interpretation
Accuracy	Correct / Total	Classification	Higher is better; best for balanced classes
F1 Score	2 × (Precision × Recall) / (Precision + Recall)	Classification	Balances precision vs recall
ROC‑AUC	∫TPR d(FPR)	Classification (binary)	Measures separability; 0.5 = random
R²	1 - SSR/SST	Regression	Proportion of variance explained
MAE	Σ	y−ŷ	/ n
RMSE	√(Σ(y−ŷ)²/n)	Regression	Penalizes large errors more

---

User Guide

Load Data

Upload your CSV or use the built-in samples.

Target Selection

Choose which column to predict, the system auto‑detects task type.

Model Training

Click Train Model, adjust hyperparameters (if needed).

Visualization

Interactive Plotly charts display metrics, confusion matrices, ROC, or residuals.

Export

Download model_pipeline.joblib and predictions.csv for reuse.

---

Deployment Guide

Local Run

streamlit run app_streamlit_clean.py

Streamlit Cloud

Push your repository to GitHub, then deploy via share.streamlit.io.

Docker Deployment

FROM python:3.10-slim
WORKDIR /app
COPY . .
RUN pip install -r requirements.txt
EXPOSE 8501
CMD ["streamlit", "run", "app_streamlit_clean.py", "--server.port=8501", "--server.address=0.0.0.0"]

Hugging Face Spaces

• Select Streamlit template.
• Upload repository → done!

---

Troubleshooting

Issue	Cause	Solution
OneHotEncoder got unexpected keyword 'sparse'	scikit‑learn ≥1.2 uses sparse_output	Auto‑detected; upgrade sklearn if needed.
Unknown label type: continuous	Chose classifier for numeric target	Auto‑detect now prevents this.
mean_squared_error got unexpected kw 'squared'	sklearn <1.0	Automatically handled.
DeprecationWarning: use_container_width	Streamlit 2025 update	Already replaced with width='stretch'.

---

Performance Tips

• For large CSVs, use preprocessed subsets.
• Use RandomForest for fast, robust tabular modeling.
• Enable XGBoost only if you have sufficient memory.
• Limit categorical cardinality to avoid excessive OHE expansion.

---

Roadmap

• Add cross‑validation & Optuna tuning
• Integrate SHAP explainability
• Support time‑series & NLP preprocessing
• Multi‑page dashboards & reporting
• REST API for live predictions
• Continuous Integration with GitHub Actions