Identification Of Algorithm from the Given Dataset Using AI/ML Techniques

Abstract: Selecting a suitable machine learning algorithm for a given dataset remains a persistent challenge due to the wide range of available algorithms and the varying statistical characteristics of real-world data. This paper presents a meta-learning based approach that recommends a suitable algorithm family for a given tabular dataset by analyzing its statistical meta-features, such as feature correlation, class imbalance, skewness, and dimensionality. Five candidate algorithms - Linear/Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, and K-Nearest Neighbors - were evaluated on sixteen benchmark datasets (five real-world datasets and eleven synthetically generated datasets covering diverse characteristics such as class imbalance, noise, and high dimensionality), and a Random Forest meta-classifier was trained to map dataset meta-features to the best-performing algorithm. Using Leave-One-Out cross- validation, the system achieved a Top-1 recommendation accuracy of 31.25% and a Top-2 accuracy of 50.00%, compared to a naive baseline of 37.50%. While the small meta-dataset size limits definitive conclusions, the results and their limitations offer useful insight into the data requirements of meta-learning based algorithm recommendation systems, and feature importance analysis identified feature correlation and class imbalance as the strongest predictors of algorithm performance.

Keywords: Meta-learning, Algorithm Selection, Machine Learning, AutoML, Meta-features, Random Forest, Baseline Comparison