Decision tree ensembles for automatic spectroscopic classification of tidal disruption events

The principal objective of this study was to develop a reliable model for the automatic classification of tidal disruption events (TDEs) using spectroscopic data. A total of 147 TDE spectra and 3626 spectra of various supernova types and AGNs were included in the data, sourced from PESSTO-SSDR1-4. An ensemble learning approach was employed using bagging with decision trees as base learners, optimized through cost-sensitive analysis and Bayesian hyperparameter tuning. A high test accuracy of 97.67 per cent, with balanced precision and recall, was achieved by the optimized model. To enhance TDE detection, a dynamic threshold adjustment was applied, prioritizing recall, which increased from 47.22 per cent to 83.33 per cent. Most TDEs were correctly identified due to this adjustment, with a reduction in precision from 85.00 per cent to 22.22 per cent and a decrease in overall accuracy from 97.67 per cent to 88.23 per cent, reflecting the prioritization of recall over precision. Relative to their occurrence in our data set, SN IIn, SN IIP, SN II, and AGNs are the most likely objects to be misclassified as TDEs. The effectiveness of the proposed methodology in accurately classifying TDEs while managing the rate of false positives is demonstrated by these results. This approach is particularly valuable in TDE detection, where minimizing false negatives is crucial to ensuring these rare events are not missed. The potential of ensemble learning, combined with cost-sensitive analysis and threshold optimization, in handling data sets in astrophysical research is highlighted by the study, offering a robust tool for future TDE classifications. The proposed method could be particularly beneficial for upcoming large-scale surveys.