Multiclass relevance units machine: benchmark evaluation and application to small ncRNA discovery

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Statistics

Background:

• Multiclass classification is essential in bioinformatics but challenging due to computational costs of direct binary algorithm extensions.
• Existing indirect methods using binary decomposition often struggle with accurate posterior probability retrieval.

Purpose of the Study:

• To extend the Classification Relevance Units Machine (CRUM) to a multiclass setting, creating the Multiclass Relevance Units Machine (McRUM).
• To develop a computationally efficient decoding algorithm for multiclass probabilistic outputs.

Main Methods:

• Extended the probabilistic kernel-based CRUM algorithm to handle multiclass problems using error-correcting output codes.
• Proposed a linear-time decoding algorithm (Naïve decoding) as an alternative to the computationally intensive GBT algorithm.

Main Results:

• McRUM with Naïve decoding demonstrated comparable predictive accuracy to GBT but with significantly improved computational efficiency on benchmark datasets.
• On large-scale ncRNA classification, Naïve McRUM showed competitive performance against SVMs, particularly in achieving high true positive rates with slightly increased false positive rates.
• McRUM exhibited superior computational efficiency compared to SVM, making it suitable for large-scale bioinformatics analyses.

Conclusions:

• McRUM provides an effective multiclass extension of CRUM, suitable for large-scale problems.
• The Naïve decoding algorithm enhances McRUM's computational efficiency, making it a viable alternative to SVMs in bioinformatics and other fields.