Chasing Your Own Tail

With the siamese networks we have a very powerful tool to learn a similarity measure for our features. However, at the same time we are doomed, because now we need to express a similarity of movie pairs to learn something at all. A very naive idea is to use the genre information because if two movies share the same genre, they have at least something in common. In case of top-level genres, like action or music, it is obvious that such labels are not optimal because even a simple genre like ‘action’ can contain many different concepts.

For instance, most people would agree that the movies ‘Die Hard’ and ‘The Transporter’ are both action movies, but with very different topics. Therefore, a training with genre labels would allow to separate action movies from drama, but it would probsably fail to learn topics that are present in a particular genre. When a teacher is present, we could ask him if two movies are similar or not, but this would involve a lot of interaction and is very time consuming.

One possible solution would be to develop a hierarchy for genres to further split top-level genres like action, into finer categories. On the one hand, even if this helps to improve the model, at some point the hierarchy is still too flat to differ between concepts. On other hand, the split of action into finer categories like {thriller, drama, sci-fi, crime} is definitely useful to model the similarity of pair-wise movies within the same genre.

At the end, it comes down to the philosophical question “What does equal mean?” and the answer is, in the context of movies, that there is no consent about it. And for it means, we need to continue our adventure trip…

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s