{ "id": 6181, "title": "Difficulties understand data format", "url": "/competitions/yandex-personalized-web-search-challenge/discussion/6181", "author_name": "", "post_date": "2013-10-31T19:52:50.677Z", "votes": null, "comment_count": 5, "views": 3462, "content": "

Hi,

\n

I have trouble understanding the following  statement (excerpt from this page):

\n

For each user from the test period, we take all her queries from the test period with at least one click with the dwell time not less than 50 time units (so, the clicked document is relevant or highly relevant according to our definition of personal relevance, see Evaluation).
From this set of queries we filter out all queries with clicks  performed at the same unit of time. Finally, from the resulting set of queries we uniformly sample only one query and consider it to be a test  query. 
If the sampled query does not have any short-term context (it is the first one in the session) and the user that asked this query has no search sessions in he training period, we remove this query from the test set (since, it has neither short nor long-term context useful for personalization).

\n

We do not disclose any user actions made after the test query. However, the user's actions performed in the same session before the test query are provided (if any).

\n

Could you please enlighten me about this?

\n

Thank you so much!

\n

X

", "messages": [ { "id": "32968", "postDate": "10/31/2013 19:52:50", "content": "

Hi,

\n

I have trouble understanding the following  statement (excerpt from this page):

\n

For each user from the test period, we take all her queries from the test period with at least one click with the dwell time not less than 50 time units (so, the clicked document is relevant or highly relevant according to our definition of personal relevance, see Evaluation).
From this set of queries we filter out all queries with clicks  performed at the same unit of time. Finally, from the resulting set of queries we uniformly sample only one query and consider it to be a test  query. 
If the sampled query does not have any short-term context (it is the first one in the session) and the user that asked this query has no search sessions in he training period, we remove this query from the test set (since, it has neither short nor long-term context useful for personalization).

\n

We do not disclose any user actions made after the test query. However, the user's actions performed in the same session before the test query are provided (if any).

\n

Could you please enlighten me about this?

\n

Thank you so much!

\n

X

", "rawMarkdown": "", "votes": null }, { "id": "33190", "postDate": "11/04/2013 17:07:18", "content": "

Hello,

\n

Sorry for the late reply.

\n

This text describes how the test set is generated. This algorithm can be useful for the participants who want to build their own validation/evaluation sets from the training data. Informally, the described algorithm achieves the following requirements for each test query:

\n
    \n
  1. at least one non-zero label is available;
    2. \n
  2. dwell time is uniquely defined (this can be problematic in case of two clicks at the same time unit);
    3. \n
  3. one test query per user;
    4. \n
  4. the test query can sampled from any part of the session (e.g., it can be the first, the second, ..., the last);
    5. \n
  5. we want to test the personalization algorithms only on the queries with some context available;
    6. \n
  6. no information about any events occurred after the test query is provided (as it is unrealistic in real-life scenario).
    7. \n
\n

Hope this helps.

", "rawMarkdown": "", "votes": null }, { "id": "33502", "postDate": "11/10/2013 09:22:16", "content": "

I still don't understand that "If the sampled query does not have any short-term context (it is the first one in the session)..."

\n

What does means "sort-term context"? Is it the first query in a session? 

", "rawMarkdown": "", "votes": null }, { "id": "33544", "postDate": "11/11/2013 07:01:03", "content": "

We adopt the following terminology. The test query's short-term context includes all actions performed in the same session, but before the test query itself. All sessions of the same user that took place before the session with the test query are called the long-term context.

", "rawMarkdown": "", "votes": null }, { "id": "37624", "postDate": "01/20/2014 05:23:24", "content": "


To,
Guocong,
I am student of computer science. I want to learn how to apply machine learning techniques to
solve practical problems. I think Kaggle is great platform for it.
I am complete newbie to this field. Can you please suggest courses and other references to start
with.
Regards,
Adwait

\n

p.s. sorry for unrelated post to this thread, i didnt find any other way to contact specific kaggler

", "rawMarkdown": "", "votes": null }, { "id": "37646", "postDate": "01/20/2014 17:30:23", "content": "

If you click username on the profile there is a Contact tab where you can email a kaggler. Also, under each post there is the 'Email user' link.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 33190, "author_name": "eugene1751", "author_url": "", "post_date": "11/04/2013 17:07:18", "content": "

Hello,

\n

Sorry for the late reply.

\n

This text describes how the test set is generated. This algorithm can be useful for the participants who want to build their own validation/evaluation sets from the training data. Informally, the described algorithm achieves the following requirements for each test query:

\n
    \n
  1. at least one non-zero label is available;
    2. \n
  2. dwell time is uniquely defined (this can be problematic in case of two clicks at the same time unit);
    3. \n
  3. one test query per user;
    4. \n
  4. the test query can sampled from any part of the session (e.g., it can be the first, the second, ..., the last);
    5. \n
  5. we want to test the personalization algorithms only on the queries with some context available;
    6. \n
  6. no information about any events occurred after the test query is provided (as it is unrealistic in real-life scenario).
    7. \n
\n

Hope this helps.

", "votes": null, "replies": [] }, { "id": 33502, "author_name": "songgc", "author_url": "", "post_date": "11/10/2013 09:22:16", "content": "

I still don't understand that "If the sampled query does not have any short-term context (it is the first one in the session)..."

\n

What does means "sort-term context"? Is it the first query in a session? 

", "votes": null, "replies": [] }, { "id": 33544, "author_name": "eugene1751", "author_url": "", "post_date": "11/11/2013 07:01:03", "content": "

We adopt the following terminology. The test query's short-term context includes all actions performed in the same session, but before the test query itself. All sessions of the same user that took place before the session with the test query are called the long-term context.

", "votes": null, "replies": [] }, { "id": 37624, "author_name": "adwaitpathak", "author_url": "", "post_date": "01/20/2014 05:23:24", "content": "


To,
Guocong,
I am student of computer science. I want to learn how to apply machine learning techniques to
solve practical problems. I think Kaggle is great platform for it.
I am complete newbie to this field. Can you please suggest courses and other references to start
with.
Regards,
Adwait

\n

p.s. sorry for unrelated post to this thread, i didnt find any other way to contact specific kaggler

", "votes": null, "replies": [] }, { "id": 37646, "author_name": "denissavenkov", "author_url": "", "post_date": "01/20/2014 17:30:23", "content": "

If you click username on the profile there is a Contact tab where you can email a kaggler. Also, under each post there is the 'Email user' link.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "32968": "", "33190": "", "33502": "", "33544": "", "37624": "", "37646": "" }, "source": "meta" }