Re: DM: missing attribute values in classification trees

[A. Famili]

Dear Tjen-Sien Lim:


An important issue in simulating missing attributes or in general
working with the data that has missing attributes is the reason for the
the data to be missing. You cannot simply decide on 5%, 10% or ...


There could be several reasons for missing attributes:


1. Human error, lack of entry, lack of transmission, lack of conversion
...


2. Sensor failure, and failure in reading, conversion, transformation,
etc.


3. Data transformation and transmission problems, ...


4. Software error in any stages of data acquisition.


5. Even some data could be missed during data warehousing process,
filtering, parsing, etc.


I have worked in the areas of electrochemical milling, wafer
fabrication/
semiconductor manufacturing and aerospace. I have seen anywhere from
~2%-20% missing attributes. You may also relate the reasons for missing
attributes with the size of your data (number of attributes and number
of cases).


Best wishes,



Fazel Famili


PS. you may refer to the first article in volume 1(1) of IDA at:


http://www.elsevier.com/locate/ida