contentwise / contentwise-impressions Goto Github PK

Hi there,

Thanks for sharing the dataset. I have read through the corresponding paper and got a few questions that I hope could be clarified.

In the impression with direct link dataset, each row is associated with a recommendation id. May I check if this id is the identifier of the recommendation algorithm used to generate the row or is it purely just treated as a key to merge the impression with direct link dataset with the interaction dataset? If the latter is the case, can I ask if there is anyway to distinguish impressions generated by different recommendation algorithms?

Thanks for in advance for your help. :)

Best wishes,
Helen

I receive an exception at line# 48 when trying to run <run_generate_splits.py> and read the dataset. The exception is raised from the apache parquet. I have installed pyarrow-0.17.1. Below is the stacktrace:

File "contentwise-impressions-master/run_generate_splits.py", line 48, in
dataset.read_dataset()
File "contentwise-impressions-master\Utils\decorator.py", line 21, in timed
result = method(*args, **kwargs)
File "contentwise-impressions-master\Utils\dataset.py", line 229, in read_dataset
self._local_interactions_file_path)
File "contentwise-impressions-master\Utils\dataset.py", line 213, in _read_parquet
return ddf.read_parquet(path=local_file_path, engine="pyarrow")
File "C:\Program Files\Anaconda3\lib\site-packages\dask\dataframe\io\parquet.py", line 1154, in read_parquet
categories=categories, index=index, infer_divisions=infer_divisions)
File "C:\Program Files\Anaconda3\lib\site-packages\dask\dataframe\io\parquet.py", line 706, in _read_pyarrow
pf = piece.get_metadata(_open)
TypeError: get_metadata() takes 1 positional argument but 2 were given

Hello,
when i used this datase, i had a problem as follows.
The situation 'when a impressions_without_direct_link is presented, a user interacts with a item which not appear in the impressions ' is what situation??
is it contained in rec_id=-1 in interactions.csv?

Hope your reply, thank you very much!

Can you add a timestamp for impressions without direct link? Without that, any time-based split is not possible.

Hello, Thanks for your amazing work!

I'm now checking the baseline you provided, before starting my research. And I found that 'ItemKNNRecommender.py' utilizes the item content matrix to compute similarity. But the paper and the dataset you published don't contain item information except for series length and type of the episodes. And there are no parts to make the ICM matrix in 'dataset.py'. My main questions are as follows.

1. Are the results you reported on the paper as an Item KNN based on the user rating matrix?
2. If not, how can I get the item content matrix you used to report results?

Once again, I'm really thankful for providing a great dataset!

It looks like a bug in dataset.py:

        self.URM = {
            "train": train_split,
            "test": validation_split,
            "validation": test_split
        }

validation_split has name "test" and test_split has name "validation".

Hi ContentWise team,

Thanks for this great work! I have a question about reconstructing the grid of recommended items.

Since we have impressions-*direct-link datasets that contain row positions of recommended series lists. I wonder if it is possible to reconstruct the grid of recommendation items that user viewed before?

For example,

1. Find [user1, rec1, row=1] from impressions-direct-link dataset (user identifier is retrieved from interactions);
2. Find [user1, row={2,3,4,5}] from impressions-none-direct-link dataset;
3. Combine them together as a 5-row grid of recommendations where user1 interacted with the first row only.

Looking forward to hearing your suggestions about this idea. Is it possible (since we don't know the timestamps in impressions-none-direct-link)?