Some question about bugs in train.py

[chao-SFD]

In the run function in train.py, there is the following code:

declare_constraints = get_declare(config["dataset"], templates=config["template"])

not_found_constraints = set(log.case_id.unique()) - set(
        declare_constraints.case_id.unique()
    )

In my testing case with the sepsis dataset, the case_id in declare_constraints are index values (0, 1, 2, ...) while the case_id in the original log are strings (e.g., 'A', 'AA', ...).

I am not sure whether the case_id values are being altered in process_log.py or train.py. I look forward to your reply, thank you.

[raseidi]

Hi there, thanks for reaching out.

Indeed at this point they have different indexes. Sorry for the messy code, let me try to clarify it to you. In a nutshell, in as_dataset.py I am merging both by their case_id columns when building the pytorch dataset.

For instance, at line 214 the tensors variable is a nested dictionary containing the cases; here, each feature cat_{feature} will be assigned to its respective case_id as a new array. Further, at line 298 the respective constraints will be assigned accordingly.

Please let me know if there's any other question.

[chao-SFD]

Thank you very much for your reply! It seems that during the preprocessing stage, we need to restore the original case_id when extracting the constraints. I added two lines of code in preprocess_log.py:
cases = log["case:concept:name"].unique() # Line 46
encoded['case_id'] = encoded['case_id'].apply(lambda x: list(cases)[x]) # Line 72
I'm not sure if this aligns with your preprocessing approach, so I'd appreciate it if you could check it for me.

Additionally, in the sepsis dataset, according to the provided preprocess_log.py, 919 cases are filtered out, but in the cache_log dataset you provided, there are 951 cases. I debugged and found that it seems that the _common_preprocessing function in pre_process.py doesn't process step 3: drop incomplete traces based on the last activity frequency. I'm not sure if it's a version issue with preprocess.py.

Thank you again for your reply!

[raseidi]

Hi there,

Sorry for the late reply, I went through a tiny surgery this week.

I investigated and there is indeed an issue about the case ids. I remember the Declare4py library had a few bugs and I modified them locally to get them fix and to return the case IDs as well. That's why on line 71 I simply renamed the index column back to case_id. Furthermore, my cached files (you can find them on the README file) also have the correct case id column.

That was my mistake, I wrote and ran the preprocess script once and never touched it again, sorry for that. I should've. However, if you just use the cached constraints from my google drive you won't have any further issues reproducing the repository.

By the way, have you doubled checked if your solution preserves the correct order of cases? If so, feel free to open a PR if you want, I'd gladly merge your fix :)

Regarding the number of cases:

1. the cache_log dataset has all the original cases. The preprocess_log return less cases indeed and I don't know exactly the reason, but it is due to some internal rules from the Declare4Py library. That's why here I drop the cases that did not have any constraint extracted before.
2. Especifically for the spesis dataset, no trace is dropped indeed since none ends with activities that satisfy the rule (activitiy frequency less that 5%). However, this step works for other datasets.

I apprenciate very much your contact. Let me know if you have any further question.