I have a dataset with about 15 feature columns and about 1000 rows that I'd like to use for supervised training.
Every row can be classified as "related" or "unrelated" to another row. About fifteen rows are "related" to one another out of the 1000.
I want to predict whether a row is related or unrelated based on the 15 feature columns.
But my issue is how to do this with classification. It's not really a multi-class problem. I don't want to predict whether something is a member of a specific class based on a featureset. Instead, I want to predict whether a row is "related" or "unrelated" to another row based on the feature set. Is it possible to do this with random forest or gbms?
It's not obvious to me how I'd do this because I can't just have a column called "related" with a binary variable. It seems like what I might need to do is to label each group with a specific class name, then use that as a response. But again, I don't want to predict class membership, just related/unrelated. So maybe I need to rewrite my random forest to search for class membership, but in the course of binary prediction?
I either end up with one 1000 row dataset with 15 "related" and 985 "unrelated" or I end up with 1000 rows, with 66 different classes. What I really want to think of it as is 66 different datasets with 1000 rows and 15 "related". But I don't know how to use online learning or update a classifier with Random Forest / gbm. So thinking I may need to do a custom net here. But looking for other ideas..