Scalable stellar evolution forecasting - Deep learning emulation versus hierarchical nearest-neighbor interpolation

Open Access

Table 3

Ranking of the predictive models Hierarchical Nearest Neighbor Interpolation (HNNI), feedforward neural network (ffNN), random forest (RF) and k-nearest neighbors (KNN) regressors, according to the performance scores outlined in Sect. 2.4 to assess predictive accuracy of stellar observables.

Score	HNNI	ffNN	RF	KNN
Validation data set
	6.57E-05	−4.46E-05	2.612E-04^*	2.506E-04
	−4.94E-06	1.82E-04^*	−2.00E-05	−1.758E-05
	−9.22E-05	4.16E-04	−4.77E-04^*	−4.55E-04
	0.102	0.145	0.210	0.217^*
	0.014	0.032	0.093	0.095^*
	0.169	0.165	1.05	1.08^*
	−0.115	−0.108	−0.700	−0.721^*
	−0.011	−0.016	−0.0378^*	−0.0375
	−0.143	−0.191	−0.286	−0.294^*
MSE	1.11E-05	2.01E-05	2.39E-04	5.79E-04^*
MAE	0.00041	0.00193	0.00479^*	0.00270
Test data set
	0.0166	0.0176	0.0319^*	0.0237
	0.0225	0.0283	0.0442^*	0.0283

Notes. The best performance is marked in bold, the worst with a “^*” tag. The manually tuned ffNN outperforms the grid search hyperparameter optimized RF and KNN models according to all scores except . HNNI outperforms ffNN as assessed by all scores, except and .

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.