3
$\begingroup$

I have video viewing data (length of session, nb of videos, etc), as well as if the user clicked on the like button. We can use the like button as a confirmation that the user had a positive viewing experience, however, only 0.1% of users click on this button. I would like to find a way to find users that have a similar data to those who liked the video without having them explicitly click the like button.

I thought about having the like variable be the response variable in a binary classification problem, however, not liking the video does not mean negative experience.

I also thought of maybe treating it as an unsupervised task, where I look if the liked sessions fall naturally inside a specific cluster.

Edit: I did not make it clear, but the service is similar to Youtube, where we are trying to figure out if a user had a positive viewing experience after clicking on a video. Right now, there is no recommendation engine and this is the first part in building one.

Edit: After the answers, I am leaning more towards approaching this task as an unsupervised learning task, rather than supervised.

Any thoughts how to approach this problem? Thanks

Improve this question
$\endgroup$

2 Answers 2

Reset to default
0
$\begingroup$

This is not a data science solution, but wouldn't the fact that the user watched the video in the first place (took the time to find and locate the video then started the video) imply that the user have some interest in that particular video over users that never interacted with the video? I would consider the fact that the entry exists as a positive response, the lack of a "like" as a neutral response, and (maybe?) the lack of watching it as a negative response. It, of course, depends heavily on the source and assumptions of your dataset.

Nonetheless, I agree that clustering the videos with likes is a good starting point in this situation, and that would definitely be my first gut instinct thing to do.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks for the response. I did not make it clear, but the service is similar to Youtube, where we are trying to figure out if a user had a positive viewing experience after clicking on a video. Right now, there is no recommendation engine and this is the first part in building one. In this case, the fact that the user clicked on a video, does not mean that he liked it. Eg: if he exited 5 seconds in. $\endgroup$
    – VincFort
    Sep 20, 2019 at 16:58
  • $\begingroup$ For many video sites, even the fact that user clicked the link to a video (without watching it) implies that the user was interested in the premise of the video. That alone gives valuable information about the potential interests of an user. If I'm interested in basketball, for example, I would be more likely to click on a link to a thumbnail related to sports over a beauty vlog. While it does not deserve as much weight as a "like", I am just suggesting this type of data may be surprisingly useful, especially if you do not have much information about those particular users in the first place. $\endgroup$
    – Jason K Lai
    Sep 20, 2019 at 17:38
0
$\begingroup$

Liking a video is not the only signal that says that a viewer would like to see more of the videos of this kind.

Take a like as just another feature of your dataset. Maybe, give it more weight. But their are other features that are important too. Those can be - the amount of time a user watched the video, the fraction of the total time that was watched, which video was watched before this one, which one was watched next, the text in the subtitles of the video converted in numerical vectors, the text in title and the description of the video and many more.

Viewerships can be attributed to titles and descriptions as well, because they clearly convey the content of the video. If 7/10 people watched a video describing gun laws, after watching a video on mass shootings, you know what to recommend.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Yes, I understand that, however, I am looking for a way to predict positive experiences other than likes. I know it would include the features you described, however, I am unsure how to formally create a model including those features. $\endgroup$
    – VincFort
    Sep 22, 2019 at 20:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.