Sequence model with any loss other than adaptive_hinge recommends same items

[mokarakaya]

I created the model as follows similar to the tests;

model = ImplicitSequenceModel(loss='pointwise',
                                  representation='lstm',
                                  batch_size=8,
                                  learning_rate=1e-2,
                                  l2=1e-3,
                                  n_iter=10,
                                  use_cuda=CUDA,
                                  random_state=RANDOM_STATE)

And I recommend items to all sequences in test_seq. There are 229 sequences. So, in the loop below, I recommend k * 229 items.
However, all recommendations are same for any sequence in test_seq.
In order words, len(aggregate_diversity) = k for losses = pointwise or bpr

adaptive_hinge works well.

I've tried k =10 and k =20 and got the same results.
I've used Movielens 100K dataset.

Do I miss something?

        sequences = test_seq.sequences[:, :-k]
        aggregate_diversity = set()
        for sequence in sequences:
            predictions = -model.predict(sequence)
            if exclude_preceding:
                predictions[sequences] = FLOAT_MAX

            predictions = predictions.argsort()[:k]
            aggregate_diversity.update(predictions)
            
        return len(aggregate_diversity)

[maciejkula]

It looks like your L2 regularization is very high. Can you try with a lower value?

[mokarakaya]

Thanks for the quick reply.

Low L2 and high n_iter solved my issue;
l2=1e-10, n_iter=100,

The loss starts to decrease after around 60th epoch.
So, I think it was an underfitting problem.