The Year Is 2015, Where Are the Flying Cars?

In our humble opinion, one reason why there are so few successful real-world recommender systems for movies is that similarity is highly subjective and would require very precise features to describe the manifold topics of movies and even in case of collaborative approaches, an enormous amount of data is required to train good models, but is not sufficient to capture all relevant aspects. Furthermore, matrix factorizations -that are often used- are transductive which means we cannot extract features for unknown, new, movies which is a serious limitation for real-word systems where each day some new movies are added. In a nutshell, if we have enough data, ratings, the computational complexity of accurate systems is usually too high which means that we need a trade-off between accuracy and runtime.

The idea of Deep Learning, to start with simple features which are then combined into more complex ones, is also useful for the domain of movies. At the lowest level, there are keywords and the keywords can be used to describe themes which then lead to sub-genres which finally lead to genres. However, usually those descriptions of movies are hand-crafted and thus prone to errors and up to a certain degree very subjective. Therefore, it cannot be guaranteed that all relevant topics are properly encoded.

We could borrow for the domain of vision because an image provides all the information that is required to precisely describe the content. Of course it is hard to extract all data from raw pixels but there is not much room for interpretation. A cat is a cat, regardless of the shape, color and orientation. As we mentioned in earlier posts, the problem with video is that we now have a sequence of images and thus, a context we need to understand. With the current technology, we cannot condense the story of movie accurately based on a stream of images.

Bottom line, movie recommender systems which are deployed these days might be pretty clever or not, but the depth of such systems is obviously limited. They might do a good job to assist the user to find interesting movies, which is pretty cool, but on the other hand, they are often also very disappointing when it comes to relating movies. For obvious clusters like Star Wars or X-Men movies it works very good, but too often, the results of “similar movies” look pretty scary. Plus, our mission, to build a personalized EPG system, slowly fades away because these days, everybody uses video on-demand or commercial video streaming services to watch “TV”. Nevertheless, we will continue and thanks to Turing, our software is not limited to DVB-{S,T,C}.


Leave a Reply

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

You are commenting using your 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