Author: Meherab Hossain Shafin
Project Type: Deterministic Dataset Validation System for Machine Learning Readiness Assessment
Architecture: Python Audit Engine + React Single-Page Interface
Scope: Rule-based dataset integrity diagnostics with transparent readiness scoring
The Automated Dataset Reliability and Modeling Readiness Inspector is a deterministic dataset auditing system designed to evaluate structural integrity and predictive modeling suitability prior to statistical analysis or machine learning experimentation.
The system accepts a CSV dataset as input and produces a structured audit covering schema validity, completeness, anomalies, leakage signals, and modeling readiness using transparent rule-based diagnostics.
Unlike exploratory notebooks or visualization dashboards, this tool functions as a reusable pre-modeling validation layer that supports reproducible data science workflows.
The system performs automated dataset inspection across multiple reliability dimensions:
- Schema diagnostics and datatype inference
- Dataset completeness analysis (missingness profiling)
- Exact duplicate row detection
- Numeric outlier identification using IQR methodology
- Leakage-risk keyword detection
- Target-column imbalance diagnostics (when specified)
- Deterministic modeling readiness scoring (0–100)
- Structured plain-text audit report generation
All checks are deterministic and interpretable. No machine learning models are used for scoring.
The system has been evaluated using a real epidemiological dengue surveillance dataset to demonstrate robustness under partially incomplete administrative metadata conditions.
Observed audit findings:
- 17.6% dataset-wide missingness
- Three columns completely empty (100% null)
- Numeric outliers detected in
dengue_total - Leakage-risk column detected (
UUID) - No duplicate rows detected
- Final readiness score: 65 / 100 (Needs preprocessing)
Example dataset: example/dengue dataset.csv
Generated audit report: example/audit_report (2).txt
Demonstrates structured CSV ingestion and preview pipeline prior to audit execution.
Provides a consolidated overview of dataset diagnostics across six inspection dimensions.
Displays transparent penalty attribution with interpretable scoring logic.
The readiness scoring system begins with: score = 100
Penalties are applied deterministically across the following reliability dimensions:
| Dimension | Purpose |
|---|---|
| Schema Quality | Structural dataset validation |
| Missingness | Completeness evaluation |
| Duplicates | Sampling integrity verification |
| Datatype Integrity | Parsing consistency |
| Outliers | Distribution stability inspection |
| Class Imbalance | Target distribution fairness |
| Leakage Signals | Target contamination risk |
Score interpretation bands:
| Score Range | Interpretation |
|---|---|
| ≥ 85 | Ready |
| 70–84 | Acceptable with preprocessing |
| 50–69 | Needs preprocessing |
| < 50 | High risk |
Columns are flagged using deterministic substring heuristics if their names contain indicators such as:
completed, discharge, final, future, id, label, outcome, post, prediction, result, status, target.
Identifier-style variables (e.g., UUID) are flagged because they may introduce record-level leakage or non-generalizable signals if retained during predictive modeling workflows.
Numeric anomaly detection is implemented using the Interquartile Range (IQR) method:
$IQR = Q3 - Q1$ $Lower Bound = Q1 - 1.5 \times IQR$ $Upper Bound = Q3 + 1.5 \times IQR$
Columns exceeding a 10% outlier rate trigger readiness penalties.
Each dataset audit produces a structured plain-text report containing:
- DATASET OVERVIEW
- SCHEMA DIAGNOSTICS
- MISSINGNESS SUMMARY
- DUPLICATE ANALYSIS
- NUMERIC OUTLIER SUMMARY
- CLASS IMBALANCE SUMMARY
- LEAKAGE-RISK INDICATORS
- MODELING READINESS SCORE
- RECOMMENDATIONS
Reports are designed for portability and reproducibility across modeling environments.
backend/
├── audit_engine.py
├── server.py
└── requirements.txt
frontend/
├── src/
├── public/
├── package.json
├── tailwind.config.js
└── craco.config.js
docs/
├── upload_preview.jpeg
├── audit_summary.jpeg
└── model_readiness.png.jpeg
example/
├── dengue dataset.csv
└── audit_report (2).txt
tests/
README.md
To reproduce the demonstration audit and verify system determinism:
- Upload: Navigate to the upload interface and select
example/dengue dataset.csv. - Execute: Run the dataset audit to trigger the diagnostic pipeline.
- Verify: Compare the generated output with the reference file:
example/audit_report (2).txt.
- Language: Python
- Engine: Deterministic rule-based statistical diagnostics
- Output: Structured audit report generator
- Framework: React
- Styling: TailwindCSS
- Interface: Single-page inspection (SPA) architecture
CSV → Parser → Audit Engine → Scoring Engine → Report Generator
To ensure maximum transparency, interpretability, and deterministic reproducibility, the system intentionally excludes:
- Authentication Systems: No user login requirements to ensure immediate utility.
- Database Persistence: Operates as a stateless service to protect data privacy.
- Excel Ingestion: Strictly optimized for CSV structures.
- PDF Export: Prioritizes machine-readable and portable plain-text reporting.
- Cloud Storage Integration: Eliminates external dependencies and egress risks.
- ML-Based Scoring: Avoids "black-box" weights; all penalties are rule-derived.
- Fuzzy Matching Heuristics: Maintains strict data integrity standards.
- Workflow Automation: Focuses purely on diagnostic auditing rather than remediation.
This tool serves as a reusable validation framework for high-stakes data environments:
- Machine Learning Pipelines: Pre-modeling data quality gatekeeping.
- Epidemiological Research: Validating surveillance datasets (e.g., Dengue/Public Health).
- Forecasting Workflows: Ensuring temporal and structural consistency in time-series data.
- Statistical Auditing: Verifying tabular research datasets prior to peer review.
- Data Governance: Standardizing reproducible quality-screening protocols.
Meherab Hossain Shafin Designed and implemented as a deterministic dataset validation subsystem supporting reproducible pre-modeling reliability assessment workflows.