Supervised and Unsupervised Machine Learning Approaches in Predicting Startup Success

Abstract

The alarming rates of failure that startups are facing highlight the need to promptly discover the crucial factors that determine success. This study investigates the forecasting of start-up success using both supervised and unsupervised machine learning techniques. Unsupervised instance filters were employed to address missing values and excessive standard deviations. In addition, the issue of data imbalance was addressed by implementing the Synthetic Minority Oversampling Technique (SMOTE). The investigation, conducted using the correlation and Info Gain attribute evaluator, revealed that relationships and milestones have the highest correlation with startup success. In order to forecast this achievement, we employed classifiers such as NaïveBayes, Functions.Logistics, Lazy.lBk, and Trees.J48. Out of all the options, the Trees.J48 model had the highest accuracy rate of 94.3%, with a confidence factor of 0.75. The accuracy of the Lazy.lBK (k-NN) variations decreases from 87.1% to 80.9% when the k-NN values increase from 3 to 7. Trees.J48 consistently exhibited strong prediction ability across different confidence levels in comparison to the other classifiers.