No-Code ML Dataset Evaluator: Empowering Experts with Intelligent Model Selection Using Web-Based Statistical Analysis

Abstract

This paper presents a comprehensive web-based system that automates machine learning dataset quality assessment and algorithm recommendation for non-expert practitioners. The platform implements a novel six-component trainability scoring framework (0-100 scale) that evaluates data completeness, feature quality, sample size adequacy, duplicate detection, outlier prevalence, and class balance to quantify dataset readiness. The system performs dual-layer statistical analysis: descriptive metrics (mean, median, standard deviation, skewness, kurtosis) for distribution characterization, and inferential tests (Pearson correlation, normality assessment) for relationship identification and transformation guidance. A multi-task evaluation engine independently scores datasets for classification, regression, and clustering suitability, while a context-aware recommendation algorithm matches datasets to specific models based on detected quality signatures— suggesting XGBoost with class weighting for imbalanced data or Huber loss regressors for outlier-heavy datasets. Users receive three to five prioritized improvement recommendations with quantified impact predictions, comparing baseline versus projected performance metrics. Interactive visualizations render statistical insights through histograms, correlation heatmaps, and quality dashboards. Implemented entirely client-side using JavaScript, Chart.js, and PapaParse, the system ensures data privacy while providing instant feedback, successfully.