TGSRec model is very nice and fancy, but I have some questions about your open code as follows:

1. Here, I think left should add one more, because neighbors_idx[left] must small than neighbors_idx, right? https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/graph.py#L87
2. I think this will sample some neightbors more than one time when len(ngh_idx) is smaller than num_neighbors, and I think this will lead bad influence... https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/graph.py#L108
3. why not directly use inner product but affinity_score here where I think it is not match your paper, right? https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/model.py#L478
4. there is only node_hist_embed without time embedding parameter like your paper said? https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/run_TGREC.py#L181-L184
5. it didn't use sin but only cos, but use both of them in your paper? https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/model.py#L249
6. why the matric of SYSRec、BERT4Rec etc. is so low where I think it doesnot match the ranking in https://paperswithcode.com/sota/collaborative-filtering-on-movielens-1m althogh it uses 1m not 100k dataset?
7. I think off_set_l[src_idx+1] here may happen index out of range when src_idx is max. https://github.com/DyGRec/TGSRec/blob/9b61480db2a18a752a79f16df7bdd17c749cdd70/graph.py#L66 help me please!!!

Hi,I am a student interested in your paper and code,but found that your code is actually run on Torch but not TensorFlow. But in your readme.MD, it is any problem with TensorFlow 1.1.2? Pardon me for my poor English! Looking forward to your reply!

Hi I'm a student and in attempting to understand your code I cloned the repository and ran the sample command at the bottom of the readme. I got the following error. I understand that this CSV file is created in the process.py file but even that file refers to './processed/ml{}.csv'. which is not present in the repository. I was wondering if I missed a critical step. Sorry if this comes across as a silly question

data.py", line 11, in init g_df = pd.readcsv('./processed/ml{}.csv'.format(DATASET)) FileNotFoundError: [Errno 2] No such file or directory: './processed/ml_ml-100k.csv'

Thanks

Hi, there

I carefully read the paper, and there is one thing that really makes me confused: what's the key difference between TGSRec and TGAN?

After reading the public codes line by line, it seems that the major difference is whether to use the edge features or not?

Could you kindly provide more comments to help me better interpret the insights of TGSRec, please?

Thanks

I went through the code and one thing is bothering me. I think there is a major bug in the implementation. It is possible that I don't understand something, so please correct me if I'm wrong, but as of my current understanding this code trains and validates using the information "from the future" .

If you examine values in the code below you will see that there negative values for the delta. https://github.com/DyGRec/TGSRec/blob/0c7ba17b1b787648ac0af0b57dfc7b91f2f00654/model.py#L557

I can see that mask is created only for the 0 values so negatives values are still used. https://github.com/DyGRec/TGSRec/blob/0c7ba17b1b787648ac0af0b57dfc7b91f2f00654/model.py#L575

Data here is sorted by edge_ids not timestamps, so the possible fix for that would by sorting by x[2], instead of x[1] https://github.com/DyGRec/TGSRec/blob/0c7ba17b1b787648ac0af0b57dfc7b91f2f00654/graph.py#L34-L39

If you look at the: TGSRec/datasets/ml-100k/u.data

data is not sorted, by the timestamp and there is no point in your codebase, where this sorting happens (I guess).

I tried to run experiments for ml-100 for both scenarios: your original implementation and with the sorted input data and the results I got are significantly worse, at least for the early stages of training. I haven't run it for 200 epochs, so maybe the final results are closer to each other, but firstly I would like to see if my assumption is correct.

Results afters 20 epochs: Without sorting:

valid acc: 0.7069337926425662
valid auc: 0.8038448618385599
valid f1: 0.7070636805233961
valid ap: 0.8172432697828477

With sorting:

valid acc: 0.5271334211112526
valid auc: 0.7374971517076878
valid f1: 0.6789490018391757
valid ap: 0.7001478155001294