Question to People:
Any Ideas about how to rate the trust level/score?

What is the trust rating for exactly?

• Is it for the video in question?
  • is the video trustworthy?
    • What makes it trustworthy?
    • Does it even need to be rated to be trustworthy, what if it isn't a video that needs a trustworthy score? (Such as an animation video for fun)
• Is it for the creator?
  • Does the creator produce and publish content that requires a trustworthy score?
    • E.g. tutorials videos or videos featuring downloads
  • In the screenshot above, you have a "Creator Reported" status, could this be open to abuse if you were to allow users to report creators? (See below about users dislike bombing as this could also happen to lead to report bombing as well)
• Is it for the dislike themselves?
  • Are the dislikes reflective of the video or the creator?
  • Are the dislikes warranted?
    • Are actual users disliking the videos because they dislike the video or are they dislike bombing the video because of an outside factor?
      • Such as a controversy involving the creator or because someone decided to be funny and tell their fanbase to dislike someone's (or their own) video(s)
  • Are these actual users who are clicking dislike or is it a bot/script?
    • How would we tell if it was? User agent checking? Require an account? Use captchas? Check YouTube account APIs and use a scoring system to determine if this is a legitimate user or a bot?

My interpretation is that the trust rating is for when YouTube removes dislikes from their API on December 13th. As stated in the FAQ of this project, at that time they will be switching over to an estimated dislike count based on the view count and like ratio, and this estimation is provided to the end user as it is now currently. That estimation can thus have a confidence rating.

I believe the best way to handle this estimation is not through a one size fits all equation, but using a model which estimates dynamically based on key performance indicators.

Example:

• Videos with a large number of views might have a different like/dislike ratio than videos with less views.
• Videos by creators with larger subscriber counts might have more engagement in likes/dislikes, causing the ratio to change.
• Subscriber count ratio to view count might also be important, as like/dislike engagement might be a curve where low and high subscriber communities have low like/dislike engagement as a ratio against views compared to medium size communities.
• The category the video is a part of might influence the rate at which people engage in liking/disliking. For instance trending videos might get a lot of views, but less engagement than a tutorial video, again, requiring a different formula.

The above are just examples of possible metrics that might be used, but would need to be tested for correlation (re: confidence).

How to do this then? We use their database to train and test, using different statistical/machine learning packages, and find the important key performance indicators that best correlate dislike estimation to actual. These models return confidence ratings in their result set, which would then be returned to the end user as described by @sy-b.

I am personally very interested in having access to their database to explore this.

@FM1337
Trust level means the gathered data's trust level.
e.g. high trust level if the data is fresh, reported by user
& low trust level if the data is shown from archived database, is old

Bot scripts only make sense if return-youtube-dislike starts collecting data from users. And if it does, then discussion on #109 might (I can't confirm) take care of that.

@FM1337 Trust level means the gathered data's trust level. e.g. high trust level if the data is fresh, reported by user & low trust level if the data is shown from archived database, is old

Okay cool, thank you for the clarification

The above are just examples of possible metrics that might be used, but would need to be tested for correlation (re: confidence).

How to do this then? We use their database to train and test, using different statistical/machine learning packages, and find the important key performance indicators that best correlate dislike estimation to actual. These models return confidence ratings in their result set, which would then be returned to the end user as described by @sy-b.

Absolutely this!

Furthermore, we could try to visualize the probable dislike ratios by simply drawing a probability distribution.