predictionio-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pat Ferrel <...@occamsmachete.com>
Subject Re: PredictionIO Universal Recommender user rating
Date Mon, 09 Oct 2017 17:42:52 GMT
Yes, this is a very important point. We have found that the % of video viewed is indeed a very
important factor but rather than sending some fraction to indicate the length viewed we have
taken the approach before to determine the % that indicates the user liked the video.

This we do by triggering a “veiw-10”, “view-25”, “view-95” etc for different viewing
times. We found that for different content types there were different % of viewing that best
predicts what the user will like. We found that for “newsy” videos “view-10” was the
best indicator. This make sense because people often do not need all the details to understand
a videos content. But for movies a “view-10” indicated a dislike. The User started a movie,
hated it and stopped it. We used “view-95” as the best indicator.

1) You know your content, do you think you have multiple types of content like “newsy”
and “stories/movies”? You may need different indicators of a user “like” corresponding
to different % of watch based on the type
2) Gather the viewing experience as % and create categories like  “veiw-10”, “view-25”,
“view-95” etc. Ingest each event for any given user. Run cross-validation tests to see
which gives the best results for each type on content you have. If you have only one type
you will find the best % to gather.
3) the problem with simply sending in the % is that for one type of content 10% is a like
(newsy) and for another type 10% alone is a dislike (long-form movies) This leads us to using
the categorical method for defining indicators to give the best result instead of using the
% of video raw, which may yield confusing of wrong results.

The extra step of testing the indicators in #2 can make a significant difference in performance.


BTW if you are able to find an indicator of dislike, this may be useful  to predict likes:
https://developer.ibm.com/dwblog/2017/mahout-spark-correlated-cross-occurences/ <https://developer.ibm.com/dwblog/2017/mahout-spark-correlated-cross-occurences/>


On Oct 9, 2017, at 10:23 AM, Daniel Tirdea <dan.tirdea@gmail.com> wrote:

Hi, 


I know there were a lot of question on this matter, I've looked everywhere but didn't find
a good answer.

I'm using the Universal Recommender to make a recommendation system for a video sharing website.
I have a lot of details in terms of user behavior but the most important one ( at least that's
what I'm now ) is the amount of seconds consumed by a visitor. A ration between the video
length in seconds and the seconds the visitor actually has seen from it.

Let's say that a visitor reached a landing page with a video with total length of 60 seconds.
If the user actually sees 60 seconds ( the video player reports that the video played the
entire 60 seconds ) I think I can assume that the visitor gave an implicit score of 10 out
of 10 for this video.

Is there a way I can include this value in the prediction system ? Or, order the returned
items by this value?

Thanks for reading this, any thought will be greatly appreciated.


Thanks,
Dan


Mime
View raw message