feisun/bert4rec

BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer

bert4rec's Issues

Why the size of ml-20m.zip is only 51M?

I see on this website https://grouplens.org/datasets/movielens/ ml-20m.zip is 196MB.

TF 2.0

Can you upgrade your code to TensorFlow 2.0?

Questions about Benchmarks

预测速度很慢的原因是啥？怎么解决？

bert4rec在我们的场景下使用的时候，预测速度非常慢。
这个原因是不是拿了所有物料做softmax导致的。
但是作为cloze任务，这个又无法避免，想问下有什么解决办法吗

Code on Ml-20m can't achieve the same performance as in paper

my result:
hit@1:0.2399， ndcg@5:0.393, hit@5:0.536, ndcg@10:0.4379=, hit@10:0.6718, ap:0.3794

result from paper:
hit@1:0.3440 ， ndcg@5:0.4967, hit@5:0.6323, ndcg@10:0.5340, hit@10:0.7473, ap:0.4785

PyTorch Implementation

Hi, I'm wondering if you also have a PyTorch implementation of the BERT4REC available. Thank you very much!

cannot duplicate result

I used the original code as well as the ml-1m data listed in the repo. But the loss didn't drop since 10000 iteration, final loss remained on about 5.12. Is there any preprocessing do I need to do before training? Thanks!

How to use User features(such as age, demographics) to construct the Sequential Recommendation System ?

I want to train a next-procedure prediction for the medical dataset. After checking the data processing pipeline, what I have figured out is: the emphasis is done on the sequence of item interaction for users along with the timestamp in the loaded data. But there is no consideration of user features.
But, what I know is user features and item features also play a role, In my case, user features are very critical. How I can use such features to design the sequential recommendation system? Thank you!

pytorch implementation

I'm wondering if you also have a PyTorch implementation of the BERT4REC available. Thank you very much!

Where does the code reflect the l2 regularizer

Why is the result of NDCG not stored in the eval_result.txt file?

After running run_beauty.sh,I got eval_result.txt.
The eval_result.txt only included masked_lm_accuracy and masked_lm_loss.
I tried and found that NDCG and other evaluation metrics only outputed in the info log.
Could you please tell me how can I get these evaluation metrics results and store them in the result txt file?
I found the source codes as followed:
` #tf.logging.info('special eval ops:', special_eval_ops)
result = estimator.evaluate(
input_fn=eval_input_fn,
steps=None,
hooks=[EvalHooks()])

    output_eval_file = os.path.join(FLAGS.checkpointDir,
                                    "eval_results.txt")
    with tf.gfile.GFile(output_eval_file, "w") as writer:
        tf.logging.info("***** Eval results *****")
        tf.logging.info(bert_config.to_json_string())
        writer.write(bert_config.to_json_string()+'\n')
        for key in sorted(result.keys()):
            tf.logging.info("  %s = %s", key, str(result[key]))
            writer.write("%s = %s\n" % (key, str(result[key])))`

It seems to exist data leakage issue in data preprocessing

The data used in test may exist in the training set.

Data generation not deterministic

BERT4Rec/gen_data_fin.py

Line 279 in 615eaf2

masked_lm_prob, max_predictions_per_seq, vocab, random.Random(random.randint(1,10000)),

Your data generation code is not deterministic, hence making it difficult to reproduce your result.

As shown in the referenced code, create_instances_threading() receive random.Random(random.randint(1, 10000)) as rng, which makes it undeterministic.

Please reply.

can you share other people git hub repos with different code for session based res sys ?

any updates for 4 years?
can you share other people git hub repos with different code for session based res sys ?

Code isn't aviable

This paper refer me to this repo. I don't see any code in this repo. Can you update this repo? Thank you

Learning rate

I modified lr=1e-3 (1e-4 in your paper) and found that the convergence speed is much faster. Similar phenomena also appeared in other code reproduction repositories, e.g, cr_1 and cr_2.

Data preprocessing?

Hello. Thanks for uploading the code.

But can you also upload how you preprocessed the data?

I would like to reproduce your results from raw data.

Thank you.