tina-hoeflich/music-recommendation-system

Music Recommendation System

System Architecture Overview

Components

1. Spotify Web API

from the feature the system only uses valence and energy. 2. With this the Valence-Arousal Plane, a well known psychological models can be built, that can display all dimensins of mood. The model is it is widely adopted because it strikes a great balance between complexity and predictive power.

3. Vector Distance Model for recommendations:
   • 
   • Each track can be seen as a vector
   • The tracks can be connected through vectors as well
   • Tracks with similar vibe = connecting vecktor has lowest length
   • -> connect find all the vectors (arrows) connecting a given track with all the other tracks (t2-t1), apply sqrt(a²+b²) to find the norm and take the arrow with the lowest norm/length.
   • this can be implemented as: def distance(p1, p2): distance_x = p2[0]-p1[0] distance_y = p2[1]-p1[1] distance_vec = [distance_x, distance_y] norm = (distance_vec[0]**2 + distance_vec[1]2)(1/2) return norm or just use numpy.linalg.norm(p2-p1)

System

1. Authorization with Spotify Client ID and Client Secret.
2. Get all of Spotify’s 120 genre labels using <sp.recommendation_genre_seeds()> and set the number of recommendations per genre to 100
3. Fetch a dataset of songs from spotify. The system draws 100 tracks from over 120 genres, resulting in a track database of around 12000 tracks with music from various styles.
4. For each of these tracks, crawl metadata and audio information and store them in the <data_dict>
5. Transform the dictionary to a pandas dataframe, drop duplicate id’s and export the dataframe to working directory: df = pd.read_csv("valence_arousal_dataset.csv")
6. Read in the dataframe
7. Combine Valence and Mood to a vector
8. Use the above generated data to create a recommendation system:
   • Crawl the input track’s “valence” and “energy” values from the Spotify API
   • Compute the distances between the input track and ALL other tracks in our reference dataset.
   • Sort the reference tracks from lowest to highest distance.
   • Return the n least distant tracks.

Packages used

1. Tekore for authorization
2. pandas
3. tqdm