Motivation
In the current modern age of technological advancements, listening to music has become an important aspect of daily routine. Various music streaming platforms like Spotify, Apple Music and Google Music enhanced the music listening experience for the user.
The challenges of music recommendation arise due to the limited amount of implicit feedback and lack of consistent meta-data or content. Playlists, list of sequential tracks seem to be one of the approach to handle the task of music recommendation. Automatic Playlist Continuation is a problem that puts the problem into perspective.
Related Work
Previous research work in the field of music recommendation has been done in the aspect of Automatic Playlist Generation and song recommendation but very little has been explored in the area of Automatic Playlist Continuation. No commercial tools are being available in this scenario to handle this task. So, this novel task of Automatic Playlist Continuation which is also part of Spotify RecSys Challenge 2018.
Proposed Methods
A lot of experimentation and exploration into the dataset was done before finally analysizing and working out through the following methods. A subset of 50,000 playlists from Million Playlist Data (MPD) from Spotify is being used for training purposes due to scaling and computational constraints. The methodologies tried out are:
Collaborative Filtering model based on Artist Similarity
We find the presence of artists in each playlist and compare the similarity based on the occurrences of various artists in each playlist. The similarity metric used is cosine similarity in finding the songs for making the recommendations.
Metadata Information Retrieval - Bag of Words
We use the bag of words model to extract the metadata from the title of the playlist and description of the playlist and various similarity metrics are used to model the feature vectors for each playlist in making the predictions.
Evaluation & Analysis
The methodologies proposed have been evaluated by calculating the metrics R-precision and NDCG by Spotify against 500 recommendations made on the basis of Artist similarity. The formulation of the metrics are as follows:
The analysis obtained for the above metrics are as follows:
| # | Metric Name | Value |
|---|---|---|
| 1 | R-precision | 0.378790 |
| 2 | R-precision Artists | 0.572721 |
| 3 | NDCG | 0.408286 |
| 4 | NDCG Artists | 0.564970 |
The results of the analysis over the plots obtained for various values of seeds are:
R-precision plots
NDCG plots
