Training Data Generation

As mentioned previously, you will build your model on implicit feedback from the user. You will look at how a user interacts with media recommendations to generate positive and negative training examples.

Generating training examples

One way of interpreting user actions as positive and negative training examples is based on the duration for which the user watched a particular show/movie. You take positive examples as ones where the user ended up watching most of a recommended movie/show, i.e., watched 80% or more. You take negative examples, again where we are confident that the user ignored a movie/show, i.e., watched 10% or less.

If the percentage of a movie/show watched by the user falls between 10% and 80%, you will put it in the uncertainty bucket. This percentage is not clearly indicative of a user’s like or dislike, so you ignore such examples. For instance, let’s say a user watched 55% of a movie. This could be considered a positive example considering that they liked it enough to watch it midway. However, it could be that a lot of people had recommended it to them, so they wanted to see what all the hype was about by at least watching it halfway through. However, they, ultimately, decided that it was not according to their liking.

Hence, to avoid these kinds of misinterpretations, you label examples as positive and negative only when you are certain about it to a higher degree.

Balancing positive and negative training examples

Each time a user logs in, Netflix provides a lot of recommendations. The user cannot watch all of them. Yes, people do binge-watch on Netflix, but still, this does not improve the positive to negative training examples ratio significantly. Therefore, you have a lot more negative training examples than positive ones. To balance the ratio of positive and negative training samples, you can randomly downsample the negative examples.

📝 We balance the negative and positive ...