Good classifier for related (or higher level) features? I'm trying to find good classifier to handle features that related to each other. For example:
Features: gender, age, weight, height.
Label: healthy or not.
(Here I made up the example but hope you understand the idea)
I'm using Decision Tree or Random Forest, but looks like training process can't see the relationship between features, e.g. weight and height. Performance is not good.
If I add a higher level feature: BMI (Body Mass Index), constructed from weight and height, then the decision tree works very well.
So my question is:

*

*Does decision tree-typed model works well with related features?

*What is best algorithm to deal with related data?

EDIT: adding an example to demonstrate
#### Prepare randomized: w,h; label is is_square
df = pd.DataFrame()
df['w'] = np.arange(1, 1000)
df['h'] = df.apply(lambda row: row.w if random.random() > 0.5 else random.randrange(1, 1000), axis = 1)
df['ratio'] = df.w / df.h # constructed feature
df['is_square'] = df.w == df.h

#### Train base on raw params: w,h
X = df[['w','h']]
y = df['is_square']
X_train, X_test, y_train, y_test = train_test_split(X,y,test_size = 0.3)

clf = tree.DecisionTreeClassifier()
clf.fit(X_train, y_train)
clf.score(X_test, y_test)
# score here is just: 0.91

#### Train base on constructed params: ratio=w/h
X = df[['ratio']]
y = df['is_square']
X_train, X_test, y_train, y_test = train_test_split(X,y,test_size = 0.3)

clf = tree.DecisionTreeClassifier()
clf.fit(X_train, y_train)
clf.score(X_test, y_test)
# score here is 1.00

You can see if I construct param 'ratio', the performance is better.
I guess the construction of 'ratio' required domain knowledge, and it help a lot. My concern is that despite of the data is clean, the decision tree failed to recognize the relationship between w and h (which is 'ratio' in this case)
 A: Decision trees are actually made for what you call "related features" (and, since random forests are just ensembles of decision trees, it holds true for them, too). Another name for "related features" would be "nested features", and trees are the very structures that are made for describing nesting. So, when you ask for recommended articles, I think any article on decision trees should make this clear.
Just think of a decision tree that first "decides" about the height of a person and then, nested, for each of those height decisions, it makes a decision about weight. And there you have your "related features".
But the question then is, why don't those models work for you? To find the reason, I would suggest debugging that gets you step by step from a trivial situation to your actual problem. First, apply random forests (RF) to some standard dataset that is known to work with RF; just use one from any of the many tutorials and examples online. Then you see whether your installation works at all. Then use a very simple version of your dataset with a very simple version of your RF so that you still can see what is going on. Then gradually increase the complexity of both the data and the model until you arrive at your current setting. This should help you isolate the problem.
