I'm facing a challenge while building a classification model. Much thanks in advance to everyone who would like to help!
I need to develop a classification model based on a large amount of data. My dataset consists of:
- about 2 thousand reliable positive examples - P;
- about 2 thousand reliable negative examples - N;
- about 1.6 million unlabeled examples - U.
My challenge is, the P&N data were labeled based on following logic:
- Some professionals analyzed the whole data set;
- select a small portion of examples based on 2 certain influencing factors, let's say A & B, like A>1000 and B in certain values. That is to say, the P&N above are examples that they believe most likely to be negative;
- investigate the examples and decide each of them is positive or negative, thus getting the P & N examples above.
- The rest, not investigated, are U examples.
However, according to my prior examples, the A & B features are the most possible influencing factors that can make one example positive or negative, which means, if I train my model on P&N examples, my model is not likely to work on the rest data, and cannot classify the unlabeled examples correctly.
Anyone knows how to make use of small and selected P & N examples and large U examples?
