Table 4 Model comparison: distributed versus centralised and local for every test.

From: Evaluating distributed-learning on real-world obstetrics data: comparing distributed, centralized and local models

  

Distributed > local

Distributed = local

Distributed < local

Row total

SGD

Distributed > centralised

72 (7.0)

14 (1.4)

9 (0.8)

95 (9.3)

Distributed = centralised

14 (1.4)

17 (1.7)

6 (0.6)

37 (3.6)

Distributed < centralised

11 (1.1)

11 (1.1)

17 (1.7)

39 (3.8)

NN

Distributed > centralised

44 (4.3)

44 (4.3)

7 (0.7)

95 (9.3)

Distributed = centralised

2 (0.2)

33 (3.2)

2 (0.2)

37 (3.6)

Distributed < centralised

0 (0)

17 (1.7)

22 (2.1)

39 (3.8)

KNN

Distributed > centralised

16 (1.6)

0 (0)

1 (0.1)

17 (1.7)

Distributed = centralised

10 (1)

2 (0.2)

1 (0.1)

13 (1.3)

Distributed < centralised

72 (7)

28 (2.7)

41 (4)

141 (13.7)

ADA

Distributed > centralised

64 (6.2)

25 (2.4)

22 (2.1)

111 (10.8)

Distributed = centralised

5 (0.5)

12 (1.2)

10 (1)

27(2.6)

Distributed < centralised

10 (1)

6 (0.6)

17 (1.7)

33 (3.2)

NB

Distributed > centralised

51 (5)

19 (1.9)

34 (3.3)

104 (10.1)

Distributed = centralised

5 (0.5)

19 (1.9)

12 (1.2)

36 (3.5)

Distributed < centralised

3 (0.3)

4 (0.4)

24 (2.3)

31 (3)

DT

Distributed > centralised

27 (2.6)

0 (0)

1 (0.1)

28 (2.7)

Distributed = centralised

8 (0.8)

0 (0)

0 (0)

8 (0.8)

Distributed < centralised

97 (9.5)

12 (1.2)

26 (2.5)

135 (13.2)

Total

 

511 (49.8)

263 (25.6)

252 (24.6)

1026 (100)

  1. Each cell is the total of distributed model when compared with centralised model (row) and local model (column) across different silos and outcome variable. (> for better, = for non significance and < for worse). The first example is 72 which means that 72 iterations of the distributed SGD was better than the centralised and local. Comparison was done with 2-sample T-test with a \(\alpha\) of 0.05 (% in parenthesis).
  2. SGD stochastic gradient descent, NN neural network, KNN K-nearest neighbors, ADA AdaBoost, NB naive Bayes, DT decision tree.