Table 4 Feature selection process driven by performance of XGBoost on different feature sets.

short description of the feature set	number of features	depth of trees	number of trees	number of predictions	AUC
short description of the feature set	number of features	depth of trees	number of trees	number of predictions	average	std dev
GO w/o ancestors, with ageing GOs	16820	6	20	20	0.8787	0.0061
GO w/o ancestors	16800	6	20	20	0.8729	0.0050
GO	21000	6	20	20	0.9086	0.0049
GO XGBoost one pass filter	373	6	20	20	0.9187	0.0042
GO XGBoost two pass filter	65	6	20	20	0.9219	0.0033
GO XGBoost two pass filter UniNet, CoExp	79	6	20	20	0.9294	0.0034
GO XGBoost two pass filter, UniNet	78	6	20	20	0.9293	0.0036
GO XGBoost two pass filter, degree	66	6	20	20	0.9283	0.0027
GO XGBoost two pass filter, ageing_n	66	6	20	20	0.9314	0.0029
GO XGBoost three pass filter, ageing_n	32	1	50	20	0.9322	0.0011

Performance of different feature sets, from weakest down to strongest, by comparing classification performance of 20 prediction each. Default settings for Gene Ontology (GO) features are “without ageing GOs but with GO ancestors”; we marked when used otherwise. For each feature set description (row), we list the number of features, the depth and number of trees in the model and the average and standard deviation of AUC values generated by 20 predictions of 5-fold cross-validation. “UniNet” means the set of network features (including degree, ageing_n, and the remaining network features), “CoExp” means the co-expression feature.

Quick links

Search