After testing with the data, the results we got were not what we expected. The test AUC is approximately 50 percent, which means that it is almost the same as random predicting.
So after some discussion, we decided to change the way of deciding the negative examples.
And there is also some thought that maybe we should control the percentage of cold start problem, but I donot think it is a good way, any way....
also, it might be a good idea to see what happens on the global dataset.
We will try them both and see what happens to the final results.