Drawbacks of Recommendations Systems

The problem with just Collaborative or Content Based Filtering models is that they push users into a bubble. They're constantly recommended content similar to the ones they're watching - leaving no scope for discovery or new genres and content.

The watch-recommend processes ends up creating a closed feedback loop that is hard to get out from. Some studies state that over 70% of users feel frustrated that the algorithms on streaming services like Netflix keep recommending the same content to them.

Here's an example of the loop with the case of Netflix:

This is why BaRT deliberately tries to introduce you to new music.

Another known problem is the 'cold start problem'.

Since collaborative filtering and NLP recommend songs based on what people have already listened to, it puts new artists at a disadvantage. A new independent artist could have released a song that the user would like on Spotify one day, but chances are that it will not be surfaced to that user because another artist receives tons of streams for a song that the user wouldn't like as much as the indie artist's song.

Spotify considers streaming a song as a success metric for recommendation. But chances are that the user streamed a particular song, but ended up disliking it. The problem lies in the fact that the algorithm has no means to tell:

• whether the user liked the song they streamed

• and whether the user dislikes the song they haven't streamed yet