msnews / msnews.github.io Goto Github PK

Hey~
The definition of impressions in MIND paper is:

An example is shown as below:

I am wondering if this impressions sequence follows the order that user has seen? (eg, user U131 see N4 earlier than N34). Or it has been re-ordered by whether it was clicked (I noticed clicked news always ahead of non-clicked news)?

Many Thanks.
Qin

https://msnews.github.io/competition.html#competition

In the description of the dataset it reads:

History. The news click history (ID list of clicked news) of this user before this impression. The clicked news articles are ordered by time.

are the oldest or most recent news ids appearing first? For example, given the line:

N37780 N8541 N81937 N15638 N47169 N23596 N86567 N73301

is N37780 the oldest or the newest click?

Thanks!

Thanks for your nice work.

I noticed that the codalab website can only show the scores on the 10% test set. However, the result table in the orginal MIND paper shows the scores on the complete test set.
I want to know how to access the complete test score.

Thank you very much.

For example, 1st sample from MINDsmall_train/news.tsv, the URL is https://assets.msn.com/labs/mind/AAGH0ET.html , accessing the URL returns the following xml:

<Error>
<Code>ResourceNotFound</Code>
<Message>The specified resource does not exist. RequestId:dd08bc8b-101e-0038-13ca-4fb0f6000000 Time:2023-03-06T01:25:44.5985962Z</Message>
</Error>

As described in the paper, There are 2,186,683 samples in the training set, 365,200 samples in the validation set, and 2,341,619 samples in the test set.
But after I downloaded the dataset and analyzed it, I noticed that there are 223,274,8 samples in the training set, 376,471 samples in the validation set, and 237,072,7 samples in the test set.

I'm wondering whether there were some changes in the dataset?
Thank you so much.

Hi
I'm very interested in this news recommendation competition.
I cannot open the codalab competition website via the Chinese IP, even use the VPN proxy .
Does it have an IP blacklist?

The dataset provides the vec file of KG entities and relations, they are very useful. However, the triplets are more important when using KG. So, could you share the basic triplets(wikidata) of MIND?

Thank you very much.

I noticed the 'behaviors.tsv' just has four columns while it was described as five columns in 'introduction.md'. Could you fix the missing "Impression ID" column? Thanks.

Hello!

If I could create my own news dataset from today's news, how do I generate the Title and Abstract entities? It was mentioned in the MIND paper that it came from using an internal NER and entity linking tool but no further explanation about it. I am interested in knowing how it was processed.

I have some ideas for getting the keys except Confidence and OccurenceOffsets.

Is there a MIND util function for generating entities? Or is there any code I can refer to to generate entities?

I would appreciate any help. Thanks!

Hey, Thanks for collecting this awesome news dataset.

I am wondering if news clicks follow the ascending time order, or they are random sequences?

For example. we assume that User U131 has historic news clicks: N11 N21 N103. Is the time of clicked N11 is earlier than the time of clicked N21.

Thanks in advanced.

This is Ying from MS News :)

Hi,

I wonder what is the type of entity in the dataset.

The description tells us just "the type of entity in wikidata", but I couldn't figure it out.

In wikidata, the type of entity is either "item" or "property", but in this dataset, the type of entity is a single character ranging from "A" to "Z.

Is there anyone who understands this meaning? Just a single URL related to this one will be helpful :)

Thanks,

Hello.

Is there a reason you only provide timestamps for impressions and not for each click in the user's history?

Since temporal information also plays an important role in recommendation (with many recent research proving so), not providing historical timestamps seems like a huge loss in terms of dataset quality.

MIND dataset is one of few reliable public datasets in news recommendation research. It would be much appreciated if you could boost the research in this area by providing those extra features.

Thanks for the otherwise wonderful dataset.

Hi,
The dataset description said that the 'userClickHistory' is all the 'news' that the user has clicked.
So, why it is not updated sequentially ,
all the 'userClickHistory' by the same user(if the user has more than on row) are all the same.
thanks