I want to make sure I correctly understand the training data, in particular, the relationship between 'invited', 'interested', and 'not interested'. These are the possible scenarios as I see them.
\r\n| Case | \r\nInvited | \r\nInterested | \r\nNot Interested | \r\nExplanation | \r\n
| A | \r\n1 | \r\n1 | \r\n0 | \r\nUser was invited, visited the page and clicked interested | \r\n
| B | \r\n1 | \r\n0 | \r\n1 | \r\nUser was invited, visited the page and clicked not interested | \r\n
| C | \r\n1 | \r\n0 | \r\n0 | \r\nUser was invited, but either didn't visit the event page, or visited the page, and clicked neither | \r\n
| D | \r\n0 | \r\n1 | \r\n0 | \r\nUser was not invited, but visited the event page and clicked interested | \r\n
| E | \r\n0 | \r\n0 | \r\n1 | \r\nUser was not invited, but visited the event page clicked not interested | \r\n
| F | \r\n0 | \r\n0 | \r\n0 | \r\nUser was not invited, but visited the event page, and clicked neither | \r\n
The cases I am not sure about are C and F. In my interpertation, every row in the training data corresponds to a user who visited an event page or was invited to do so. Once on the page, some users clicked interested, some clicked 'not interested', and some\r\n clicked neither. If a (user, event) pair (U,E) is missing from the data, we can assume user U was never invited to event E, and also never visited the event page.
\r\nIs this correct?
", "messages": [ { "id": "20071", "postDate": "02/05/2013 19:16:08", "content": "I want to make sure I correctly understand the training data, in particular, the relationship between 'invited', 'interested', and 'not interested'. These are the possible scenarios as I see them.
\r\n| Case | \r\nInvited | \r\nInterested | \r\nNot Interested | \r\nExplanation | \r\n
| A | \r\n1 | \r\n1 | \r\n0 | \r\nUser was invited, visited the page and clicked interested | \r\n
| B | \r\n1 | \r\n0 | \r\n1 | \r\nUser was invited, visited the page and clicked not interested | \r\n
| C | \r\n1 | \r\n0 | \r\n0 | \r\nUser was invited, but either didn't visit the event page, or visited the page, and clicked neither | \r\n
| D | \r\n0 | \r\n1 | \r\n0 | \r\nUser was not invited, but visited the event page and clicked interested | \r\n
| E | \r\n0 | \r\n0 | \r\n1 | \r\nUser was not invited, but visited the event page clicked not interested | \r\n
| F | \r\n0 | \r\n0 | \r\n0 | \r\nUser was not invited, but visited the event page, and clicked neither | \r\n
The cases I am not sure about are C and F. In my interpertation, every row in the training data corresponds to a user who visited an event page or was invited to do so. Once on the page, some users clicked interested, some clicked 'not interested', and some\r\n clicked neither. If a (user, event) pair (U,E) is missing from the data, we can assume user U was never invited to event E, and also never visited the event page.
\r\nIs this correct?
", "rawMarkdown": "", "votes": null }, { "id": "20135", "postDate": "02/07/2013 15:16:15", "content": "So it looks like the answer to my previous post is contained in the original description of the data; the above interpretation is correct. I would like to summarize and\r\n add some information from previous posts concerning event_attendees before asking my next question.
\r\nHere is what we know:
\r\n-None of the users in the test.csv are in train.csv. \r\nTherefore, train.csv alone cannot be used for collaborative filtering to predict the user preferences in test.csv
\r\n- About 52% of events in test.csv are in train.csv.
\r\n-Only .2% and .6% of the users in train and test respectively are in event_attendees.\r\nTherefore, for 99.8% of users in the test set, we have no information about their interest nor their attendance. The demographic information about users is therefore crucial.
\r\n-All the events in train, test are in event_attendees. So for each event in train and test, we know which users responded yes/maybe/no to the invitation.
\r\nAs events4u pointed out, this is orthogonal to their interest in the event, in the sense that for a given (user,event) pair any attendance value (yes/maybe/no) could correspond to any interest value (interested/not interested/no response).
\r\nHowever, as John Park mentioned, it is reasonable to assume there is some correlation between interest and attendance, otherwise the popularity benchmark (based on attendance) would not have performed above chance at predicting the interest\r\n level.
\r\nIt seems there are three options:
\r\n1) Assume attendance and interest measure the same thing, with a mapping such as: yes -> interested, maybe -> unknown, no -> not interested.Combine the training data and event attendance data into one rating matrix, and use it to predict the\r\n interest in the test data (together with user information and event metadata).
\r\n2) Use event_attendees only to measure event popularity, which can be used as a feature.
\r\n3) Ignore the event_attendee altogether.
\r\nCan anybody think of any other way to use the event_attendee data?
", "rawMarkdown": "", "votes": null }, { "id": "20227", "postDate": "02/09/2013 14:52:51", "content": "if you combine into a single dataset, you lose out on the timestamp the user saw the event - because event_attendees does not have that information
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 20135, "author_name": "ayy69366", "author_url": "", "post_date": "02/07/2013 15:16:15", "content": "So it looks like the answer to my previous post is contained in the original description of the data; the above interpretation is correct. I would like to summarize and\r\n add some information from previous posts concerning event_attendees before asking my next question.
\r\nHere is what we know:
\r\n-None of the users in the test.csv are in train.csv. \r\nTherefore, train.csv alone cannot be used for collaborative filtering to predict the user preferences in test.csv
\r\n- About 52% of events in test.csv are in train.csv.
\r\n-Only .2% and .6% of the users in train and test respectively are in event_attendees.\r\nTherefore, for 99.8% of users in the test set, we have no information about their interest nor their attendance. The demographic information about users is therefore crucial.
\r\n-All the events in train, test are in event_attendees. So for each event in train and test, we know which users responded yes/maybe/no to the invitation.
\r\nAs events4u pointed out, this is orthogonal to their interest in the event, in the sense that for a given (user,event) pair any attendance value (yes/maybe/no) could correspond to any interest value (interested/not interested/no response).
\r\nHowever, as John Park mentioned, it is reasonable to assume there is some correlation between interest and attendance, otherwise the popularity benchmark (based on attendance) would not have performed above chance at predicting the interest\r\n level.
\r\nIt seems there are three options:
\r\n1) Assume attendance and interest measure the same thing, with a mapping such as: yes -> interested, maybe -> unknown, no -> not interested.Combine the training data and event attendance data into one rating matrix, and use it to predict the\r\n interest in the test data (together with user information and event metadata).
\r\n2) Use event_attendees only to measure event popularity, which can be used as a feature.
\r\n3) Ignore the event_attendee altogether.
\r\nCan anybody think of any other way to use the event_attendee data?
", "votes": null, "replies": [] }, { "id": 20227, "author_name": "rkirana", "author_url": "", "post_date": "02/09/2013 14:52:51", "content": "if you combine into a single dataset, you lose out on the timestamp the user saw the event - because event_attendees does not have that information
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "20071": "", "20135": "", "20227": "" }, "source": "meta" }