Skip to main content

Table 4 Prediction rates of transport modes in the Test sets before and after a posteriori homogenization

From: Application of machine learning to predict transport modes from GPS, accelerometer, and heart rate data

 

Prediction ratesa (%)

Before homogenization

1-min bandwidth

2-min bandwidth

3-min bandwidth

4-min bandwidth

5-min bandwidth

Overall

79 (57–88)

86 (68–93)

88 (66–94)

89 (67–95)

90 (67–95)

90 (67–95)

Overall transport

74 (41–89)

76 (39–91)

79 (42–92)

80 (42–94)

78 (40–95)

77 (35–96)

Activity place

80 (55–89)

89 (68–95)

90 (68–96)

91 (69–97)

91 (71–97)

92 (71–97)

Bike

92 (76–100)

93 (77–100)

95 (79–100)

95 (80–100)

95 (78–100)

95 (77–100)

Private motorized

69 (0–87)

73 (0–92)

77 (0–96)

78 (0–96)

79 (0–97)

79 (15–100)

Public transport

62 (0–90)

63 (0–95)

63 (0–95)

65 (0–100)

65 (0–100)

66 (0–100)

Walking

80 (28–95)

81 (29–97)

81 (23–97)

82 (22–97)

81 (21–98)

78 (17–99)

  1. aPrediction rates presented as median prediction rates from 126 RF models with 2.5th and 97.5th percentiles in the parentheses. The overall prediction rate is from a model (with heart rate data) that is unweighted by category size, while the mode-specific prediction rates are from a corrected model (with heart rate data) applying a weight. Correction for category size entails modifying the cut-offs for prediction to the observed proportions of the categories, at the forest prediction step (when aggregating information from all trees). For the same tree predictions, a higher “proportion of votes” is reached for rarer categories in the weighted vs. unweighted model