rivolli / mfe Goto Github PK

View Code? Open in Web Editor NEW

26.0 5.0 7.0 407 KB

Meta-Feature Extractor

License: Other

R 100.00%

machine-learning meta-learning meta-features characterization-measures mtl

mfe's People

Contributors

Stargazers

Watchers

Forkers

jehadaldahdooh lpfgarcia yuancz brunocordeiro180 jotakas danielsilva83 codinglifev

mfe's Issues

Problem with landmarking when there are a column named with '0'

when the dataset has a column named with '0' there are a problem in the landmarking.
Specifically in the function replace.nominal.columns

Infotheo measures

TODO: attribute_entropy, class_entropy is the normalized version
TODO: irrelevant.attributes
TODO describe discretization method

Problem when extracting landmarking meta-features

If one of the landmarkers fails, an error happens and there is no result returned by the function mfe::mf.landmarking().

In my case, I debugged the the function call, and 6 different values should be returned. However, it failed forth one (knn), due to some dataset's characteristics. In theory, the first three algorithms worked, but with the error, no previous result is provided.

I handled that using tryCatch statement and imputing NAs. However, the whole vector has NAs values, but the correct would be just the missing descriptor.

$landmarking
       decision.stumps elite.nearest.neighbor    linear.discriminant 
                    NA                     NA                     NA 
           naive.bayes       nearest.neighbor             worst.node 
                    NA                     NA                     NA

Is there any smartest way I could use from my side? Calling the functions individually?

landmarking fold problems

The mfe throws an error when the dataset has fewer instances in one class than the number of folds.
I remember that we solved this problem by doing min(10, min(table(y))

Statistical measures wLambda breaks up if there is an empty level in the class

If the dataset target attribute has a level without no sample, the measure will raise an exception and stop the code unexpected

Quantile problems

Using quantile to summarize data can generate problems when NA is produced in the measures

Range returns the min and max

It should be better if the summarization function range works like iqr and returns a single value

landmarking problems with columnames

landmarking brokes using this dataset, after I rename the columns it works.

colnames(xdata)
   [1] "-"                     "/"                     "0"                    
   [4] "00"                    "04"                    "0;"                   
   [7] "0cm"                   "1"                     "1-1/2"                
  [10] "1-1/2-year"            "1-month"               "1-year"               
  [13] "1-year-1-month"        "1-year-4-month"        "1-year-5-month"       
  [16] "1/2"                   "1/2/01"                "10"                   
  [19] "10-month"              "10-year"               "101"                  
  [22] "102"                   "103"                   "104"                  
  [25] "105"                   "11"                    "11-1/2-year"          
  [28] "11-month"              "11-month-16-day"       "11-year"              
  [31] "11-year-5-month"       "12"                    "12-day"               
  [34] "12-month"              "12-year"               "12-year-7-month"      
  [37] "13"                    "13-1/2-year"           "13-day"               
  [40] "13-month"              "13-year"               "134"                  
  [43] "14"                    "14-day"                "14-year"              
  [46] "15-month"              "15-year"               "15-year-6-month"      
  [49] "16"                    "16-month"              "16-year"              
  [52] "17-day"                "17-year"               "18"                   
  [55] "18-month"              "18-year"               "19-month"             
  [58] "19-year"               "2"                     "2-1/2-year"           
  [61] "2-3"                   "2-month"               "2-year"               
  [64] "2-year-3-month"        "2-year-4-month"        "2-year-9-month"       
  [67] "20-day"                "20-year"               "2001"                 
  [70] "21"                    "21-day"                "21-month"             
  [73] "21-year"               "22"                    "22-day"               
  [76] "27"                    "28"                    "2;"                   
  [79] "2nd"                   "3"                     "3-4"                  
  [82] "3-month"               "3-year"                "3-year-10-month"      
  [85] "3-year-5-month"        "38"                    "3mm"                  
  [88] "3rd"                   "4"                     "4-1/2-year"           
  [91] "4-month"               "4-year"                "4-year-11-month"      
  [94] "4-year-3-month"        "4-year-4-month"        "4-year-8-month"       
  [97] "5"                     "5-1/2-year"            "5-6"                  
 [100] "5-month"               "5-month-19-day"        "5-year"               
 [103] "591"                   "5th"                   "6"                    
 [106] "6-month"               "6-month-6-day"         "6-year"               
 [109] "6-year-11-month"       "6-year-4-month"        "7"                    
 [112] "7-1/2-year"            "7-month"               "7-year"               
 [115] "7-year-4-month"        "8"                     "8-month"              
 [118] "8-month-13-day"        "8-year"                "8-year-2-month"       
 [121] "8-year-9-month"        "85-month"              "9"                    
 [124] "9-day"                 "9-month"               "9-month-24-day"       
 [127] "9-month-30-day"        "9-year"                "9-year-1-month"       
 [130] "90s"                   "90th"                  "9;"                   
 [133] "abdomen"               "abdominal"             "abnormal"

rivolli / mfe Goto Github PK

mfe's People

Contributors

Stargazers

Watchers

Forkers

mfe's Issues

Recommend Projects

Recommend Topics

Recommend Org

Jobs