JIST2015 Data Challenge

• 中文 stop words

• Apply Filters for Attributes
• Change all properties to nominal using regex pattern
• String regex = "^(https?|ftp|file)://[-a-zA-Z0-9+&@#/%?=~_|!:,.;]*[-a-zA-Z0-9+&@#/%=~_|]";
• Replace all {0} or {1} to {0,1} in train.arff and test.arff files to let them be compatible.
• Normalize all numeric attributes into [0,1] interval

• Save Predictions to a File
• [Visualize classifier errors] - > save as a .arff file
• remove all headers and import it into a .csv file
• final two columns show the "predicted classes" and "actual classes"

• Parameter Selection
• Use CVParameterSelection as the classifier
• Choose a classifier to optimize the parameters
• Add a parameter need to be tuned in CVParameters
• e.g., I 10 250 25 for Random Forest will test 10, 20 ... 250 (25 steps) for the [number of trees] parameter
• It will show the best option as output: "Classifier Options: -I 90 -K 0 -S 1"

• Feature Selection
• Fast feature (attribute) selection using ranking (using searching is not feasible due to the number of attributes is large)
• GainRatioAttributeEval with Rank 
• Set threshold < 0.1, < 0.2, < 0.3, < 0.4 to run experiments to see which threshold is working best -> 0.2

• Consider Index (Can it improve the performance?)
• Used formula 1-index/size
• Answer: