After replacing the ml-1m.txt with the original data, the ndcg@10 is close to 0.8

[zuoy3964]

hi, I download the original ml-1m dataset from https://grouplens.org/datasets/movielens/, and replace the "ml-1m.txt",

fpath = '/ml-1m/ratings.dat'

def read_data(fpath):
    data = []
    with open(fpath, 'r') as f:
        for line in f:
            parts = line.strip().split('::')
            if len(parts) >= 2:
                uid, itemid = parts[0], parts[1]
                data.append((uid, itemid))
    data = pd.DataFrame(data, columns=['uid', 'itemid'])
    return data

def convert_to_txt(data, output_path):
    with open(output_path, 'w') as f:
        for _, row in data.iterrows():
            f.write(f"{row['uid']}\t{row['itemid']}\n")


data = read_data(fpath)
data.head()
output_path = 'ml-1m.txt'
convert_to_txt(data, output_path)

Then run the code you provided in README.md.

python main.py --dataset=ml-1m --train_dir=default --maxlen=200 --dropout_rate=0.2 --device=cuda

I found that after just 20 epochs run, the ndcg@10 is close to 0.78.
Why is so different?

[pmixer]

@zuoy3964 well, as you found the issue, you maybe the 1st one need to explain it actually.

If you mean https://raw.githubusercontent.com/RuidongZ/Recurrent_Recommender_Networks/refs/heads/master/data/ml-1m/ratings.dat this dataset, it should be sorted according to the timestamp which is absent in your code piece.

I don't think ml1m is a good dataset actually, predicting which movie to going to be rated nextly is bit wried,
have you tried newer datasets like https://github.com/otto-de/recsys-dataset ?

Lastly, how would you introduce your DDualSE: Decoupled Dual-head Squeeze and Excitation Attention for Sequential Recommendation work, if within 200 words limit, I'm just interested if you are trying sth new following this work.