#实验结果

TEST_SVM_ON_VIPER.md

Training phase

20000 training instances The number of support vectors is around 15000, which means that most of the training data are near the separating hyper-plane. The training accuracy is around 80%

Test on full data

• Accuracy = 52.3734% (662/1264) (classification)
• Accuracy = 62.5791% (791/1264) (classification)
• Accuracy = 61.0759% (772/1264) (classification)
• Accuracy = 60.9968% (771/1264) (classification)
• Accuracy = 63.5285% (803/1264) (classification)
• Accuracy = 61.9462% (783/1264) (classification)
• Accuracy = 57.1203% (722/1264) (classification)
• Accuracy = 50% (632/1264) (classification)

Test on positive data

• Accuracy = 69.7785% (441/632) (classification)
• Accuracy = 92.5633% (585/632) (classification)
• Accuracy = 91.9304% (581/632) (classification)
• Accuracy = 91.6139% (579/632) (classification)
• Accuracy = 87.3418% (552/632) (classification)
• Accuracy = 79.7468% (504/632) (classification)
• Accuracy = 84.6519% (535/632) (classification)
• Accuracy = 88.9241% (562/632) (classification)

Test on negative data

• Accuracy = 34.9684% (221/632) (classification)
• Accuracy = 32.5949% (206/632) (classification)
• Accuracy = 30.2215% (191/632) (classification)
• Accuracy = 30.3797% (192/632) (classification)
• Accuracy = 39.7152% (251/632) (classification)
• Accuracy = 44.1456% (279/632) (classification)
• Accuracy = 29.5886% (187/632) (classification)
• Accuracy = 11.0759% (70/632) (classification)

分析

训练数据接近3/4都是支持向量，说明了数据本身的可分性不强，训练的结果更像是over-fitting的结果。 测试结果很差，且模型倾向于将结果分为正类别。 另外，可以注意到第8块的结果最差，可能包含的信息很少。