Cold Start Problem

[Martagdvalle]

Our cold start problem was to provide a leverage for new artists that are not well known in the industry by recommending it to users based on the genre and type of artist they are listening to.

This idea has been put down as we want to focus on building a robust recommender model that covers our main two points, as the cold start problem requires a high level of time and investment that we can not compromise.

[barto-official]

The problem also stems from the data — to create a real, state-of-the-art recommender for a cold start in the music industry, we need an updated dataset (take data from Spotify API directly). The dataset that we use are years old, which means recommending "niche" artist is not a valid task anymore because it's an outdated dataset for this problem. Secondly, taking data directly from API has a limit per hour on how much data you can request. That would make a dataset very small and the recommender for cold start problem was designed with CNNs which are data-intensive.

[talphaidze]

It is important to mention that the cold start problem was in the end indirectly tackeled by implementing implicit Bayesian Personalized Ranking recommendation system. This is as the BPR algorithm takes into account songs that are known to the user as well as songs that are not known to the user. For more info please refer to the paper: https://arxiv.org/ftp/arxiv/papers/1205/1205.2618.pdf