I helped my students understand the decision tree classifier in sklearn recently. Maybe they think I helped too much. But I think it was good for them. We did an interesting little exercise, too, writing a program that writes a program that represents a decision tree. Maybe it will be useful to someone else as well:
def print_tree(t, root=0, depth=1):
if depth == 1:
print 'def predict(X_i):'
indent = ' '*depth
print indent + '# node %s: impurity = %.2f' % (str(root), t.impurity[root])
left_child = t.children_left[root]
right_child = t.children_right[root]
if left_child == sklearn.tree._tree.TREE_LEAF:
print indent + 'return %s # (node %d)' % (str(t.value[root]), root)
print indent + 'if X_i[%d] < %.2f: # (node %d)' % (t.feature[root], t.threshold[root], root)
print_tree(t, root=left_child, depth=depth+1)
print indent + 'else:'
See it in action here.
Did I do this for MILK a few years ago? I’m becoming an absent-minded professor ahead of my time.
Doctors love decision trees, computer scientists love recursion, so maybe that’s why decision trees have been coming up so much in the Artificial Intelligence for Health Metricians class I’m teaching this quarter. We’ve been very sklearn-focused in our labs so far, but I thought my students might like to see how to build their own decision tree learner from scratch. So I put together this little notebook for them. Unfortunately, it is a little too complicated to make them do it themselves in a quarter-long class with no prerequisites on programming.
In a recent post, I confessed my interest in a recent National Academy Press report on teaching methods. The tough thing for me about using this discipline-based education research (DBER) approach is not the name or the acronym, but coming up with the misunderstood concepts from the discipline that students benefit from learning actively. In the report examples, it seems like they are articulated by geniuses dedicated to teaching after years of student observation. I don’t know if I’ll get there one day, but I’m certainly not there now.
But I had a great idea, or at least one that I think is great: see what people are confused by online. I tried this out for my lecture last week on cross-validation, using the stats.stackexchange site: http://stats.stackexchange.com/questions/tagged/cross-validation?sort=votes&pageSize=50
After reading a ton of these, I decided that if my students know when they need test/train/validation splits and when they can get aways with test/train splits then they’ve really figured things out. Now I can’t find the question that I thought distilled this best, though.
To complement that ASA address about what is statistics that I read last week, here is the abstract KDD address about what is data mining: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=754923
Does the talk exist somewhere?
I gave a talk on automatic methods to map from verbal autopsy interview results to underlying causes of death last August, and I like the slides so much that I’m going put them online here. Cheers to the new IHME themed templates for Power Point!
One exciting announcement that I got to make at the Verbal Autopsy Congress last October is that the PHMRC gold standard verbal autopsy validation data is now available for all researchers. You can find it in the Global Health Data Exchange: Population Health Metrics Research Consortium Gold Standard Verbal Autopsy Data 2005-2011.
Insert an example of doing something with it here.