Design of an Improved Method for Clustering Using Variational Autoencoders, DBSCAN, and Genetic Algorithms

Abstract

With the increase in data complexity and volume, the demand for more accurate clustering methods in data analysis has grown to an importance bordering on the critical. In many applications, existing clustering methods perform poorly on high-dimensional data, in the presence of noise, and for the identification of arbitrary-shaped clusters. In this setting, the current study develops a novel framework that integrates density-based clustering with deep learning, rule mining, and genetic algorithms to improve clustering accuracy. Traditional clustering algorithms, such as k-means and hierarchical clustering, are limited in their ability to handle complex data distributions and noise. Predefined cluster shapes drive these methods and work suboptimally with high-dimensional data. Our approach can overcome such challenges by leveraging the variability of Variational Autoencoders—DBSCAN, the Apriori algorithm with decision trees, and an Adaptive Genetic Algorithm for parameter optimization. This is the realm into which VAEs, in conjunction with DBSCAN, finally outperform traditional clustering methods. VAEs, on the other hand, can model complex data distributions and reduce their dimensionality, thereby making the data more amenable to clustering. Subsequently, DBSCAN is applied to the lower-dimensional latent representations produced by VAEs to identify clusters of arbitrary shapes that are robust to noise. This combination resulted in high clustering accuracy, with an Adjusted Rand Index of 0.85 and a significant reduction in the impact of noise on sets. We use the Apriori algorithm and decision trees for cluster interpretation. The Apriori algorithm finds frequent itemsets in each.