A Simplified Evaluation Method Used for Unsupervised Learning Algorithms

Abstract

Professional evaluations for unsupervised learning algorithms are a crystal important and challenging topic due to its input-output specialty. There are no output labels available in unsupervised learning models to work as an evaluation criterion to check and compare the predicted outputs with the actual outputs. In other words, it is relatively difficult to effectively estimate or validate the correctness or efficiency of any unsupervised learning algorithms with regular methods used in supervised learning studies.
In this study, we developed a simplified evaluation method by using average or mean values for testing observations and comparing them with the mean values of clustered observations to identify their correctness of their clustering. This method can be used to effectively perform some evaluations for unsupervised learning models. A comparison between our method with some traditional evaluation methods, such as Silhouette Scores, is performed and the comparison result shows that our method is better than that traditional method.
The experimental study confirmed the effectiveness of our evaluation methods used for unsupervised learning algorithms, and the correct evaluation rate is up to 97%.