99.9% percent of sample predicted label are -1

[zxsimple]

I trained a DeepFM model with AUC 0.84, while predicting on the training data, 99.9% and more samples predicted as -1.

[wangcaihua]

can you post you training hyper-parameter?

the reason could be that the regularization parameters is so large that the model is under trained

I gauss that the true positive/negative rate are also abnormal.

[zxsimple]

The FM model has this problem, after I predicting on WAD model, the prediction data seems much reasonable. The Pos/Neg ratio is 1:3 on both models.

Here is my FM model parameters:

... ...
--ml.model.size 9257 \
--ml.feature.index.range 9257 \
--ml.data.type libsvm \
--ml.learn.rate 0.1 \
--ml.reg.l2 0.2 \
--ml.fm.field.num 61 \
--ml.fm.rank 8 |
--ml.inputlayer.optimizer ftrl \
--ml.epoch.num 10 \
--ml.data.label.trans.class PosNegTrans \
-- ml.data.label.trans.threshold 0.5 \
... ...