Hot PLS—a framework for hierarchically ordered taxonomic classification by partial least squares

• Classification in a fixed hierarchy
• Utilization of replicate measurements for improved robustness through majority voting
• Automatic model building and complexity estimation from taxonomic information
• Detection of outliers and samples representing new classes absent in calibration

A novel framework for classification by partial least squares in a fixed hierarchy is presented. The hierarchical approach ensures flexible local modelling with varying complexity. It results in an intuitive classification path from the highest taxonomic levels down to species and beyond. Results are presented as phylogenetic trees with local diagnostic information to gain maximum information about the classification and help the researcher to focus on interesting phenomena.Information on sample replicates is included in the classification to increase performance and avoid misclassifications due to low quality measurements. Detection of samples coming from previously unobserved classes is enabled by estimating cut-off distances from the calibration data classes. To further increase flexibility and improve customization the canonical powered partial least squares algorithm is used for modelling and classification together with linear discriminant analysis. This opens up for additional sample response information and forced sharpening of focus on important variables. The presented framework is not limited to biological taxonomy, but was first developed for this purpose.