Content-Based Filtering

Contents

The following notebook illustrates our content filtering approach that uses track similarity (measured by cosine distance) to recommend tracks to playlists.

0. Motivation

In order to recommend songs to playlists, we want to recommend songs that share similar features with the existing songs in the playlists. One way to find songs that share similar features with the existing songs is to look at cosine similarity of their audio features.

The idea is that if they share similar audio features (thus a short cosine distance), they are similar songs.

Cosine similarity measures the orientation of two n-dimensional sample vectors irrespective to their magnitude. It is calculated by the dot product of two numeric vectors, and it is normalized by the product of the vector lengths. The output value ranges from 0 to 1, with 1 as the highest similarity.

We compute a similarity matrix for tracks by using sklearn pairwise distance method, with cosine similarity:

0.1 Strategy

To compute prediction set per playlist, our model does the following:

  1. The model looks at individual track of the playlist (from the training set) and find the cosine distance of that track and all other tracks in our training set.
  2. It recommends the top k similar songs per track until the recommendation list has reached its pre-determined limit (which is test-set size * 15)

1. Data Processing

1.1 Train-test split

We did a stratified splitting of the data (by Playlistid) into training and test set by 80-20. Stratified splitting ensures that we have the right proportion of train and test for playlists of different lengths.

1.2 Data cleaning

We drop some non-numeric features in order to calculate the cosine similarity matrix. Below shows the first few rows of our training data (with the audio features of individual tracks).

train_cleaned.head()
Track_Duration acousticness artist_popularity danceability energy instrumentalness key liveness loudness mode speechiness tempo time_signature valence
0 221939 0.72300 84 0.714 0.343 0.000005 10 0.1020 -7.641 1 0.3800 72.348 4 0.372
1 246213 0.02810 71 0.770 0.421 0.000000 0 0.1110 -8.255 0 0.0849 131.935 4 0.214
2 216120 0.01970 86 0.399 0.787 0.000061 1 0.0685 -2.880 1 0.0499 117.089 5 0.572
3 304200 0.04070 77 0.681 0.570 0.000000 7 0.1050 -5.829 0 0.1360 135.121 4 0.153
4 256946 0.00164 60 0.545 0.905 0.000002 7 0.0532 -10.732 1 0.0648 103.574 4 0.826

1.3 Create a cosine-similarity matrix

We create a cosine-similarity matrix of the individual tracks in the training set. The shape of the cosine matrix shows 2463 unique tracks (in 100 playlists) in the training set.

train_scaled_cos_matrix.shape

(2463, 2463)

Here’s the cosine similarity of the first 5 tracks in the training set. We expect the diagnoal of the matrix to be 1, because it is referring to the same song.

train_scaled_cos_matrix[:5]

array([[1.        , 0.70370129, 0.81424758, ..., 0.87263617, 0.79012625,
        0.76284813],
       [0.70370129, 1.        , 0.82541858, ..., 0.77956107, 0.89152418,
        0.82271377],
       [0.81424758, 0.82541858, 1.        , ..., 0.93027042, 0.81602446,
        0.72451838],
       [0.79433301, 0.93252096, 0.80163412, ..., 0.86770269, 0.90692581,
        0.89966543],
       [0.85392504, 0.74018692, 0.9374127 , ..., 0.96500088, 0.84458619,
        0.79398798]])

train.head()
Playlistid Trackid Artist_Name Track_Name Album_Name Track_Duration Artist_uri Track_uri Album_uri acousticness ... loudness mode speechiness tempo time_signature valence Playlist Album Track Artist
0 231023 4 Chance The Rapper Blessings Coloring Book 221939 spotify:artist:1anyVhU62p31KFi8MEzkbf spotify:track:2VQc9orzwE6a5qFfy54P6e spotify:album:71QyofYesSsRMwFOTafnhB 0.72300 ... -7.641 1 0.3800 72.348 4 0.372 Chance The Rapper 21 38 16
1 249966 51 Chief Keef Love Sosa Finally Rich 246213 spotify:artist:15iVAtD3s3FsQR4w1v6M0P spotify:track:4IowQDUOzUvNtp72HMDcKO spotify:album:6VML3yX1WR5pXtg6fmwdzZ 0.02810 ... -8.255 0 0.0849 131.935 4 0.214 Rap 37 53 28
2 2535 46 Sia Chandelier 1000 Forms Of Fear 216120 spotify:artist:5WUlDfRSoLAfcVSX1WnrxN spotify:track:4VrWlk8IQxevMvERoX08iC spotify:album:3xFSl9lIRaYXIYkIn3OIl9 0.01970 ... -2.880 1 0.0499 117.089 5 0.572 Feel Good 94 102 85
3 58121 26 Fetty Wap No Days Off (feat. Monty) Fetty Wap 304200 spotify:artist:6PXS4YHDkKvl1wkIl4V8DL spotify:track:3wr2fJ91iGjZlrcDLJ6bYp spotify:album:2gKQvajkEEaDtkqJ8FJ4uw 0.04070 ... -5.829 0 0.1360 135.121 4 0.153 Grindin 31 38 23
4 67941 1 Spin Doctors Two Princes Pocket Full Of Kryptonite (Anniversary Edition) 256946 spotify:artist:2PSiyldxmJze7xiqbz658m spotify:track:0cuq829zRNq4AEdmDJA4aw spotify:album:3jx5C8Z66FwdjNkqsSDBQh 0.00164 ... -10.732 1 0.0648 103.574 4 0.826 90's 26 28 21

5 rows × 28 columns

Just an example, we can look at the distribution of cosine distance of an individual track to all other tracks in the training set.

import matplotlib.pyplot as plt
plt.figure(figsize=(8,8))
plt.hist(train_scaled_cos_matrix[:1][0])
plt.xlabel("Cosine distance")
plt.ylabel("Frequency")
plt.title("Distribution of cosine distance of: Blessings by Chance The Rapper")
plt.show()

png

We can see a somewhat normal distribution of cosine distance, that most songs have >= 0.6 cosine distance to ‘Blessings’, with a peak at 0.80-0.85.

2. Model Performance

2.1 Metrics

We used the following metrics to evaluate our model, based on Spotify RecSys rules

2.2 Model Test-Set Performance

Data R-Precision NDCG Average of the two metrics
Test set of 100 playlists 0.045856 0.051839 0.048847
Test set of 1000 playlists 0.033373 0.042258 0.037815

3. Conclusion

In conclusion, content filtering outperforms our baseline model that uses collaborative filtering on co-occurrence.

Interestingly, we did not see an improvement in score when predicting on a larger dataset (with more unique tracks and playlists).

Source code: https://github.com/phoebewong/spotify-teamNPK/blob/master/src