Table 7 Detailed ablation study results.

From: Multi-source heterogeneous data fusion and intelligent prediction modeling for chemical engineering construction projects based on improved transformer architecture

Configuration

Progress Accuracy (%)

Quality F1-Score

Risk AUC

Overall Performance

Full Model

91.8

0.907

0.924

100% (baseline)

w/o Multi-scale Attention

87.6 (− 4.2)

0.876 (− 0.031)

0.901 (− 0.023)

95.4%

w/o Cross-modal Alignment

89.3 (− 2.5)

0.889 (− 0.018)

0.915 (− 0.009)

97.2%

w/o Adaptive Weighting

88.1 (− 3.7)

0.883 (− 0.024)

0.908 (− 0.016)

96.1%

w/o Enhanced Positional Encoding

90.4 (− 1.4)

0.901 (− 0.006)

0.920 (− 0.004)

98.5%

Single-head Attention

86.2 (− 5.6)

0.862 (− 0.045)

0.889 (− 0.035)

93.9%