Automated contrastive optimization of class-based feature distribution for noncontinuous dataset

Abstract Discrete data and categorical data classification is an essential task across domains such as machine learning, data mining, decision-making, and statistical analysis, aimed at categorizing observations into distinct classes based on their attributes, which can be numerical, categorical, or a combination of both. One characteristic of such data is its uncertainty and ambiguity, stemming from the limitations of information provided by a sparse data space. The Transposed Feature Class Granular (TFCG) distribution has effectively classified this kind of data, highlighting the significance of integrating fuzzy sets, granular computing, and principles of artificial neural networks. Properly tuned hyperparameters using the aforementioned method can greatly improve classification performance. However, manually tuning hyperparameters presents several challenges, including being time-consuming, subjective, and facing scalability issues, as well as the risk of overfitting and yielding inconsistent results. With the assistance of contrastive learning, which enables the model to distinguish between similar (positive) samples and dissimilar (negative) samples, this paper presents an automated hyperparameter optimization method designed to identify effective TFCG distribution classifiers for noncontinuous data. Our proposed approach contributes to simplifying the process of finding optimal hyperparameters for TFCG classification, significantly reducing time complexity. It’s well known that manually tuning parameters is difficult and time-consuming. Our method eliminates the need for manual hyperparameter tuning and produces better results. We address the issue by defining a series of contrastive loss functions for noncontinuous attributes and applying them to derive an optimized TFCG classifier.