Table 1 Collaborative Filtering hyper-parameter tuning: Tabulated top 5 results from training multiple collaborative filtering models using neighborhood methods and matrix factorization methods
From: Implicit-descriptor ligand-based virtual screening by means of collaborative filtering
Distance | N. Threshold | AUC | BEDROC20 | EF1% |
|---|---|---|---|---|
Neighborhood-based collaborative filtering | ||||
 Pearson | 10−2 | 0.791 | 0.723 | 4.225 |
 Pearson | 10−5 | 0.792 | 0.723 | 4.225 |
 Pearson | 10−4 | 0.791 | 0.723 | 4.225 |
 Jaccard | 10−2 | 0.648 | 0.647 | 3.670 |
 Cosine | 10−5 | 0.644 | 0.647 | 3.670 |
Num. Factors | SGD step size | AUC | BEDROC20 | EF1% |
|---|---|---|---|---|
Matrix factorization-based collaborative filtering | ||||
 50 | 10−3 | 0.891 | 0.929 | 5.476 |
 32 | 10−4 | 0.860 | 0.889 | 5.028 |
 32 | 10−2 | 0.899 | 0.872 | 4.994 |
 25 | 10−4 | 0.868 | 0.867 | 4.765 |
 25 | 10−2 | 0.892 | 0.847 | 4.547 |