Which features can better reflect data-sets with missing values?
Standard statistical modelling approaches can not deal with missing values. However, ANFIS work well fin such situation as it impues missing values..
thanks mr. Mahmoud, what is ANFIS? can you introduce an article to help me get more information?
ANFIS stands for Adaptive Neuro-Fuzzy Inference System. There is a Toolbox in Matlab with lots of help sand demos.
For example start running a demo program called gasdemo.m This demo teaches how fuel consumption (MPG) for automobiles, using data from previously recorded observations, can be predicted using ANFIS.
A nice tutorial is provided by Heikki Koivo (2000): ANFIS (Adaptive Neuro-Fuzzy Inference System).
The link is:
ftp://190.90.112.131/Facultades/FIET/DEIC/Materias/Control Inteligente/Parte III/clase 23 int/FVAnfis2.pdf
you may also learn about the structure of missing values first, see
Advances in Data Analysis and Classification
April 2012, Volume 6, Issue 1, pp 29-47
Exploring incomplete data using visualization techniques
One possiblity is also to replace (impute) the missing values before you apply learning methods, for example with the R package VIM or VIMGUI.
The best method for learning in presence of missing values reminds me of Hop Field Neural networks[HNN] which is unsupervised recurrent machine learning algorithm using Hebbian theory. It works on the concept of auto associative memory just like human brain where a significant part of the input data is enough to recall the input pattern.
For example :
If you train a HNN with human faces and during testing if you provide a face with goggles(missing or replaced values for eyes), still the HNN recalls the training pattern.
Refer to the document for more details
Good Luck
I recommend distance-based methods like k nearest neighbors method. They cope well with missing features. When some features are missing you simple calculate distance omitting them. And you construct local classifier/aproximator based on the nearest points.
Neural networks, neuro-fuzzy systems, linear regression, Partial Least Squares regression, LASSO and elastic net need complete input information (all features) in my opinion.
- Badges
- Science topic
- Similar topics
- Computer Science
- Data Mining