I'm not an expert on this field, but take a look at imputation methods (for instance http://www.stat.columbia.edu/~gelman/arm/missing.pdf ).  These will 'predict' the missing values (or, at least will try to approximate these) and, therefore, you can build the tree without any missing value. The most well known is EM (take a look at http://www.uvm.edu/~dhowell/StatPages/More_Stuff/Missing_Data/MissingDataNorm.html ). Hope it helps!

I've to say that this is a quite challenging problem.

have you try some data mining and machine learning software like Orange ? You can try it, it is very powerfull with included decision tree module.

http://orange.biolab.si/

I'm not an expert on this field, but take a look at imputation methods (for instance http://www.stat.columbia.edu/~gelman/arm/missing.pdf ).  These will 'predict' the missing values (or, at least will try to approximate these) and, therefore, you can build the tree without any missing value. The most well known is EM (take a look at http://www.uvm.edu/~dhowell/StatPages/More_Stuff/Missing_Data/MissingDataNorm.html ). Hope it helps!

I've to say that this is a quite challenging problem.

i have developed some classification methods (like BFC, BNC and recently DNC: http://www.sciencedirect.com/science/article/pii/S1877050914010813 )... in such case of missing values, usually i ignore this instance ... but with 30% of data containing missing value, i don't this that is a good idea :/ ... Another alternative is possible: you have to try affecting a value (like "none") to the missing data and considerate this "none" as an attribute value ;) ... 

http://www.sciencedirect.com/science/article/pii/S1877050914010813

Conference Paper Adaptive Learning of Nominal Concepts for Supervised Classification

Conference Paper Boosting Formal Concepts to Discover Classification Rules

I recommend dependency networks (http://research.microsoft.com/apps/pubs/default.aspx?id=64334), which are quite simple. Build one decision tree for each of your variables/columns. As training data, use the cases (rows) where the column value is defined. Now, at test time, when you want to fill in missing values, run the decision trees that correspond to each missing value. If you have more than one missing value, you can do Gibbs sampling: order the missing values, and then run the decision trees one-at-a-time, using the output of the other decision trees as inputs to the next. You can run around the Gibbs sampling loop several times.

If you already have decision tree code, it's a simple scripting wrap. Otherwise, you can use Max Chickering's WinMine toolkit (http://research.microsoft.com/en-us/um/people/dmax/WinMine/Tooldoc.htm)

If you have adequate numbers of complete data you can use machine learning approach using decision tree to predict missing fields !

Article Machine Learning Based Missing Value Imputation Method for C...

Dear Ahmed,

This paper can be useful for your needs:

Decision Trees for Uncertain Data:  http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=5210098&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D5210098

Kind regards,

Olivier PARISOT

pl see excellent article

Decision Tree: Review of Techniques for Missing Values at ... - UKSim

uksim.info/aims2015/CD/data/8675a122.pdf

by S Gavankar -

https://www.google.co.in/search?q=decision+tree+to+fill+in+the+missing+values+of+a+data+set%3F&rlz=1C1CHBF_enIN752IN752&oq=decision+tree+to+fill+in+the+missing+values+of+a+data+set%3F&aqs=chrome..69i57&sourceid=chrome&ie=UTF-8

thanks dr for recommendation

I'm working upon this right now; firstly, I must say that it's debatable whether use of decision trees for imputation is the most effective method. It will depend upon type of data, size of datasets and some other factors. However, you might apply tree algorithms like C 4.5 for imputation.

The logic works like this:

1) split your datasets into 2 subsets: one containing missing values(Dm) and the other one which doesn't(Dc).

2) use Dm to construct your decision tree model

3) the feature containing missing values might need to be discretized, based upon the distribution diagram of Dc

4) Apply your decision tree model upon Dm; class of missing values should be obtained

5) use the class range based upon your discretization to infer a potential missing value(a bit similar to random hot deck) or just use mean or most frequent values