在Sklearn中将概率估计与支持向量机相结合

Question

I'm currently using a support vector machine to predict which item a user will buy given demographic data. The data set also includes how many users of a certain age group bought each item. It looked something like this:

   items a b c
age 
15-20    10 3 10
20-25    1 5 6
25-30    2 5 6

I am unsure how to incorporate this into the training data, as the only way I can think of to incorporate this is to include a set of probability values of the user buying the item, but this is very unwieldy. Another idea I had was to use an ensemble learning method, and combine the svm with maybe a Naive Bayes classifier. I am using sklearn to build my model.

Answer 1

When you want to introduce weightings for data points, SVM is no longer so attractive. The underlying algebra doesn't work as well when identical or very close data points have differing classifications. From the data you give above, I do expect that Naive Bayes will give a faster computation and cleaner results.

That said, what SVM algorithm are you using? If it's one that weights the vectors in some fashion -- use each point exactly once, or pick a random point for each iteration of a gradient descent approach -- then you can certainly handle this by adding each point to your training set the given number of times. For instance, you'd have 10 rows stating that teens bought item a.

On the other hand, Naive Bayes would give you weightings for a statistically accurate model. Instead of predicting almost unilaterally that 20-somethings will buy item c (which is actually a large minority of the purchases), you'd have a model that could tell you that 48% of people in their late 20s will buy item c, and almost as many will buy item b.

Does this discussion help?