Executive Summary

Problem Summary

        With the advent of smartphones and the consumer economy, there has been a revolution in the ways that people consume products and content. At the same time, digital music and digital music distribution platforms have become some of the most widely accessible and highly consumed product markets in the world. Yet with this deluge of digital music content comes a challenge: how do users find new content that they enjoy, and how do digital music platforms enable music discovery by users? These challenges are exacerbated by the fact that in the modern fast-paced world, people are often time or attention limited, there are other platforms competing for user attention, and digital content-based company's revenue often relies on the time consumers spend on, or interact with, its platform. These companies need to be able to figure out what kind of content is needed to increase customer time spent on their platform, the amount of interaction had with their platform, and the overall satisfaction with a user’s experience on the platform. The key challenge for companies is in figuring out what kind of content their users are most likely to consume. Spotify is one such music content provider with a huge market base across the world. With the ever-increasing volume of streaming music becoming available, finding new music of interest has become a tedious task in and of itself. Spotify has grown significantly in the market because of its ability to make highly personalized music recommendations to every user of its’ platform based on a huge preference database gathered over time - millions of customers and billions of songs. This is done by using smart recommendation systems that can recommend songs based on users’ likes/dislikes, incorporating both content-based and latent features for song recommendations. However, the recommendation system used by Spotify and its hyperparameter settings have remained a proprietary, closely guarded secret. Here, I build a recommendation system to provide a top 10 of personalized song recommendations to a user that the user is most likely to enjoy/like/interact-with based on that users’ personal musical preferences.

Solution Summary

       In total, I explored six recommendation systems using the 1 million songs dataset1 as part of the solution design for this project: popularity-based, user-user collaborative filtering, item-item collaborative filtering, matrix factorization, cluster-based, and content-based. To evaluate the different models analyzed here, I relied on the F1_score, model predictions of user/song interactions, and the top 10 recommended songs by each model. These metrics clearly demonstrate a higher level of performance by three of the six models: matrix factorization (Singular Value Decomposition with default settings, with a 70% weight applied), user-user collaborative filtering (KNN basic algorithm with msd distance, max cluster size of 50, minimum cluster size of 9, and 30% weight applied), and the content-based model (tf-idf encoding and cosign-similarity distance on album title, artist name, and release year). Therefore, I propose a Hybrid-Based Recommendation System, built by combining these three models with the specified hyperparameters, be adopted for personalized recommendations of music with maximum user interaction potential. The proposed hybrid recommendation system boasts fast computational times, high prediction accuracy of user preferences, and a balanced recommendation of familiar and diverse new music for users. However, the model is subject to limitations such as a 'cold start' problem when making predictions for new users with few listens, and with the ever-increasing volume of songs becoming available and users adopting the platform this will lead to longer and longer computation times (and presumably less user satisfaction). Additionally, the model generally favors popular artists and songs, which could have an impact on the music industry by making it more difficult for new artists to break into the scene using this platform.
       Importantly, to reduce the computational resources required to test, train, and evaluate the solution design for this project, it was necessary to reduce the size of the dataset to make it more computationally tractable. The resulting filtered dataset reduced the original '1 million songs' dataset (with 2 million records) to 121,900 records and increased the accuracy of predicted user interactions by reducing high degree of imbalance in the data from infrequent users and unpopular songs. While these results do not address the 'cold start' issue, they demonstrate the importance of a filtering step for making fast and relevant recommendations. Therefore, I suggest the application of a pre-recommendation filter as part of a production design. Additionally, the content-based part of the recommendation system could also be improved by incorporating further features such as genre, lyrics, danceability, and encoded recordings. With these additions to the content portion of the recommendation system, the weight of the content-based model in the final hybrid recommendation system might be increased in future versions of the product. I recommend that these limitations and proposals be considered in future releases of the recommendation system for improved user personalization and increased user satisfaction.

==> NOTE <==
For the full code check out the GitHub Link at the bottom of the page

The objective:

Build a recommendation system to propose the top 10 songs for a user based on the likelihood of listening to those songs.

The key questions:

The Data

The core data is the Taste Profile Subset released by the Echo Nest as part of the Million Song Dataset. There are two files in this dataset. The first file contains the details about the song id, titles, release, artist name, and the year of release. The second file contains the user id, song id, and the play count of users.

Data Source

http://millionsongdataset.com/ (1)

The dataset is split into two .csv files I load here. The two dataframes and their features are:

song_data

user_id song_id play_count
0 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOAKIMP12A8C130995 1
1 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOBBMDR12A8C13253B 2
2 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOBXHDL12A81C204C0 1
3 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOBYHAJ12A6701BF1D 1
4 b80344d063b5ccb3212f76538f3d9e43d87dca9e SODACBL12A8C13C273 1
5 b80344d063b5ccb3212f76538f3d9e43d87dca9e SODDNQT12A6D4F5F7E 5
6 b80344d063b5ccb3212f76538f3d9e43d87dca9e SODXRTY12AB0180F3B 1
7 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOFGUAY12AB017B0A8 1
8 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOFRQTD12A81C233C0 1
9 b80344d063b5ccb3212f76538f3d9e43d87dca9e SOHQWYZ12A6D4FA701 1

count_data

song_id title release artist_name year
0 SOQMMHC12AB0180CB8 Silent Night Monster Ballads X-Mas Faster Pussy cat 2003
1 SOVFVAK12A8C1350D9 Tanssi vaan Karkuteillä Karkkiautomaatti 1995
2 SOGTUKN12AB017F4F1 No One Could Ever Butter Hudson Mohawke 2006
3 SOBNYVR12A8C13558C Si Vos Querés De Culo Yerba Brava 2003
4 SOHSBXH12A8C13B0DF Tangle Of Aspens Rene Ablaze Presents Winter Sessions Der Mystic 0
5 SOZVAPQ12A8C13B63C Symphony No. 1 G minor "Sinfonie Serieuse"/All... Berwald: Symphonies Nos. 1/2/3/4 David Montgomery 0
6 SOQVRHI12A6D4FB2D7 We Have Got Love Strictly The Best Vol. 34 Sasha / Turbulence 0
7 SOEYRFT12AB018936C 2 Da Beat Ch'yall Da Bomb Kris Kross 1993
8 SOPMIYT12A6D4F851E Goodbye Danny Boy Joseph Locke 0
9 SOJCFMH12A8C13B0C2 Mama_ mama can't you see ? March to cadence with the US marines The Sun Harbor's Chorus-Documentary Recordings 0

Observations and Insights:

==> NOTE <==

Exploratory Data Analysis (EDA):

Lets take a look at the distribution of ‘play_count’, the feature I use as a proxy for user ‘rating’:

png

From this distribution plot, we can see that the number of songs played by each user (lets call it user interactions) is heavily right skewed. There are many (thousands) users who have only listened to a few songs, so any matrix I build from this data will be extremely sparse. Next I reduce the skew in the play_counts by filtering out users who have less than a minimum (I settled on 90) number of total plays. These are users that there is very little preference data for. Plotting the distriution of ‘play_count’ after filtering:

png

<
Column Non-Null Count Dtype
0 user_id 1224498 non-null object
1 song_id 1224498 non-null object
2 title 1224498 non-null object
3 release 1224498 non-null object
4 artist_name 1224498 non-null object
5 year 1224498 non-null int64
6 play_count1224498 non-null int64

Filtering out less active users has decreased the imbalance somewhat. The distribution is still heavily right skewed in the above plot, but the class imbalance has been reduced by a factor of 5 (y limit is 1200 instead of 7000). This has also decreased the size of the dataset to 1.2 million records down from 2 million.

This is still not good enough, so I continue to decrease the sparcity and imbalance of the data by filtering out any user/song records that have a play count less than 6. There are many more songs that users have only listened to 1 or a few times. I want to recommend highly rated songs, so I am going to get rid of these songs with low interactions and assume they are uninteracted with ‘not-liked’.

<
Column Non-Null Count Dtype
0 user_id 185694 non-null object
1 song_id 185694 non-null object
2 title 185694 non-null object
3 release 185694 non-null object
4 artist_name 185694 non-null object
5 year 185694 non-null int64
6 play_count185694 non-null int64

This last filtering step dramatically reduced the size of the dataset by 85%, this will speed up processing times for the models and also make the predictions more accurate by reducing the class imbalance. Finally, I filter out all songs that have less than 20 user interactions:

png

<
Column Non-Null Count Dtype
0 user_id 124147 non-null object
1 song_id 124147 non-null object
2 title 124147 non-null object
3 release 124147 non-null object
4 artist_name 124147 non-null object
5 year 124147 non-null int64
6 play_count124147 non-null int64

With these filtering steps, I have reduced the class imbalance, decreased the size of the dataset to make models and gridsearch more tracteable, and I have decreased the extreme sparcity of our resulting recommendations matrices. Now I am going to apply a threshhold limit to further reduce class imblance and the play_count range, and then apply a min max scalar to standardize the play_counts so I can effectively use them as a proxy for a 1-10 rating.

Clipping and Scaling play_count:

Because there are very few users who have listened to a song more than 25 times, I will set a threshold at 25 plays. I dont want to drop records with more than 25 plays because this is important information on users likes, So I will clip anything > 25 to 25 and then apply a MinMaxScalar function from the sklearn package to scale the playcounts from 1-10.
Distribution plot after filtering:

png

Distribution plot after scaling:

png

My final play_counts, filtered, clipped, and scaled look pretty good. The class imbalance between the left side of the X axis and the higher play_counts has been reduced, The play_counts are scaled from 1-10, and I have kept the records containing the highest ratings as 10’s.

(121900, 8)

The final dataset has 121,900 records. This is a much more maneagable number of records for training and testing models. After previous iterations of testing the filtering, I had dropped over 90% of the songs so the final recommendations that were being made were not diverse and nearly the same for every user. So Im going to continue EDA and take a look at the filtered dataset to make sure the user/song diversity is still good…

Exploratory Data Analysis Continued…

Checking the total number of unique users, songs, artists in the data

Total number of unique user id

Number of unique USERS =  19212

Total number of unique song id

Number of unique SONGS =  2210

Total number of unique artists

Number of unique artists =  1194

Observations and Insights:

Let’s find out about the most interacted songs and interacted users

Most interacted songs

song_id play_count
138 SOBONKR12A58A7A7E0 1466
80 SOAUWYT12A81C206F1 1445
1624 SOSXLTC12AF72A7F54 1186
478 SOFRQTD12A81C233C0 1133
88 SOAXGDH12A8C13F8A1 1024
357 SOEGIYH12A6D4FC0E3 993
1188 SONYKOW12AB01849C9 831
1901 SOWCKVR12A8C142411 748
1343 SOPUCYA12A8C13A694 744
647 SOHTKMO12AB01843B0 616

Most interacted users

user_id play_count
5683 4be305e02f4e72dad1b8ac78e630403543bab994 106
766 0a4c3c6999c74af7d8a44e96b44bf64e513c0f8b 82
827 0b19fe0fad7ca85693846f7dad047c449784647e 81
8262 6d625c6557df84b60d90426c0116138b617b9449 74
16334 da3890400751de76f0f05ef0e93aa1cd898e7dbc 69
7930 695179610d0b1fbb9d66267a3bd24946617af7fb 67
17477 e9a7dba8248ced646ea192016660e3c9056c0d03 66
2996 283882c3d18ff2ad0e17124002ec02b847d06e9a 65
5405 48567d388c6a7dda0e9d0a7b6648bdb42440475c 65
10463 8c78c69701072e204f4340ca4d6ee44fe39e40cc 64

Observations and Insights:

Songs played in a year

0 49 48 47 46 43 45 50 41 42 ... 17 21 7 6 3 5 12 1 2 4
year 0 2009 2008 2007 2006 2003 2005 2010 2001 2002 ... 1977 1981 1967 1966 1962 1965 1972 1958 1960 1963
play_count 26580 12782 10616 7499 6693 5718 4861 4617 4316 4147 ... 151 145 137 134 113 110 77 66 62 61

2 rows × 51 columns

Text(0, 0.5, 'number of releases')

png

Observations and Insights:

Now I apply different algorithms to build a recommendation system.

Building a baseline popularity-based recommendation system

In this basic recommendation system, I take the count and sum of play counts of the songs and build the popularity recommendation system based on the sum of play counts (For the full code check out the Github Link at the bottom of the page). The rank based recommender function is:

Here is the output after running the popularity-based RS on the filtered dataset:

title artist_name
0 221 keller williams
1 Call It Off (Album Version) Tegan And Sara
2 Clara meets Slope - Hard To Say Clara Hill
3 Kelma Rachid Taha
4 Numb (Album Version) Disturbed
5 Voices On A String (Album Version) Thursday
6 What If I Do? Foo Fighters
7 Encore Break Pearl Jam
8 Reign Of The Tyrants Jag Panzer
9 Dance_ Dance Fall Out Boy

Collaborative Filtering, Matrix Factorization, and Clustering based recommendation sytems

Before running the following recommendation systems, I developed several functions for calculating metrics to evaluate the models. The metrics I used are Root Mean Squared Error, Mean Average Error, Precision, Recall, and F1.

I then split the data into a train and test set (Link to full code is at the bottom of the page). Also, to build the user-user-similarity-based and subsequent models I used the “surprise” library from Python.

User User Similarity-Based Collaborative Filtering Metrics

The code for implementing the user-user similarity matrix with base settings is:

Result:

RMSE: 3.1852
MAE:  2.5755
  algorithm      rmse       mae  precision  recall  f1_score  popularity
0  KNNbasic  3.185239  2.575525      0.682   0.772     0.724        24.3

Observations and Insights:
For the untuned User-user similarity-based model,

predictions using knn.KNNBasic for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Next, I tuned the model using GridSearchCV to try to improve the model performance.

The metrics after tuning are:

2.8805243519619927
{'k': 50, 'min_k': 9, 'sim_options': {'name': 'msd', 'user_based': True}}

RMSE: 2.8309
MAE:  2.2550
        algorithm      rmse       mae  precision  recall  f1_score  popularity
0        KNNbasic  3.185239  2.575525      0.682   0.772     0.724        24.3
1  KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745        64.9

Observations and Insights:
After tuning the user-user model,

predictions using knn.KNNBasic tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Finally, because more ‘popular’ songs are more likely to be ‘liked’, I adjust the tuned recommendations by using a custom script to weight recommendations by the number of plays. The final, weighted, recommednations from the User-User similarity-based recommendation system for user ‘6ccd111af9b4baa497aacd6d1863cbf5a141acc6’ are:

title artist_name count_plays predicted_interaction corrected_ratings
0 Clara meets Slope - Hard To Say Clara Hill 89 9.205088 9.099088
1 Numb (Album Version) Disturbed 68 8.825430 8.704162
2 When You're Gone Avril Lavigne 76 8.568666 8.453958
3 #40 DAVE MATTHEWS BAND 72 8.503340 8.385488
4 Speechless (Album Version) The Veronicas 45 8.512847 8.363776
5 XRDS Covenant 51 8.475145 8.335117
6 (iii) The Gerbils 88 8.352223 8.245623
7 Modern world Modern Lovers 51 8.274458 8.134430
8 The Memory Remains Metallica / Marianne Faithfull 94 8.209485 8.106343
9 Sunburn Muse 15 8.156890 7.898691

Observations and Insights:

Item-Item Similarity-based collaborative filtering recommendation system

Metrics of the item-item model compared to user-user

RMSE: 2.9990
MAE:  2.2550
        algorithm      rmse       mae  precision  recall  f1_score  popularity
0        KNNbasic  3.185239  2.575525      0.682   0.772     0.724        24.3
1  KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745        64.9
2   KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695        25.1

Observations and Insights:

predictions using knn.KNNBasic_item for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Next, I ran GridsearchCV to to search for the optimal hyperparameters to tune the item-item model with. The optimal hyperparameter settings after running GridSearch are:

2.9243151529965794
{'k': 50, 'min_k': 3, 'sim_options': {'name': 'cosine', 'user_based': False}}

Model Metrics for item-item model after rerunning with the tuned paramters

RMSE: 2.9040
MAE:  2.3017
             algorithm      rmse       mae  precision  recall  f1_score  \
0             KNNbasic  3.185239  2.575525      0.682   0.772     0.724   
1       KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745   
2        KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695   
3  KNNbasic_item_tuned  2.903997  2.301706      0.688   0.785     0.733

predictions using knn.KNNBasic_item tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

The final, weighted, recommendations from the Item-Item similarity-based recommendation system for user ‘6ccd111af9b4baa497aacd6d1863cbf5a141acc6’ are:

title artist_name count_plays predicted_interaction corrected_ratings
0 Ni Tú Ni Nadie (Versión Demo) Alaska Y Dinarama 31 10.000000 9.820395
1 My Perfect Cousin The Undertones 21 9.842105 9.623887
2 Walk On Water Or Drown (Album) Mayday Parade 24 9.684211 9.480086
3 Docking Bay 94 The Alter Boys 25 9.661287 9.461287
4 Waters Of Nazareth (album version) Justice 29 9.210526 9.024831
5 Valentine Justice 24 8.934211 8.730086
6 Please_ Before I Go Derek Webb 32 8.905551 8.728775
7 Hitsville U.K. The Clash 25 8.421053 8.221053
8 Go Places The New Pornographers 20 8.342105 8.118498
9 Uptown Drake / Bun B / Lil Wayne 24 8.061607 7.857483

Observations and Insights:

Model Based Collaborative Filtering - Matrix Factorization

Model-based Collaborative Filtering is a personalized recommendation system, the recommendations are based on the past behavior of the user and it is not dependent on any additional information. It uses latent features to find recommendations for each user. Here I am using Singular Value Decomposition (SVD) method from the Suprise library, and calculate the metrics after running the base model with default settings:

RMSE: 2.7215
MAE:  2.1682
             algorithm      rmse       mae  precision  recall  f1_score  \
0             KNNbasic  3.185239  2.575525      0.682   0.772     0.724   
1       KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745   
2        KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695   
3  KNNbasic_item_tuned  2.903997  2.301706      0.688   0.785     0.733   
4                  SVD  2.721453  2.168153      0.696   0.798     0.744

predictions using SVD tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Improving matrix factorization based recommendation system by tuning its hyperhyperparameters

To tune the SVD model, first I run a factor checking function that plots the RMSE based on a range of latent features.

Here is the plot for this data run for 100 features:

png

According to the figure, there is a decreasing trend of better performance with higher k. The lowest RMSE is achieved when k is 80. However, it is worth mentioning that k = 52 and >84 are also good. The result suggests a range of values which can be used in GridSearchCV()for hyperparameter tunning. Next I ran GridSearchCV to find the optimal hyperparameter settings. The hyperparameter settings for the model that reduced RMSE the most are:

2.7097656150739744
{'n_epochs': 30, 'lr_all': 0.01, 'reg_all': 0.4, 'n_factors': 80}

Model Metrics for Matrix Factorization (SVD) method after rerunning with the tuned paramters

RMSE: 2.6736
MAE:  2.1434
             algorithm      rmse       mae  precision  recall  f1_score  \
0             KNNbasic  3.185239  2.575525      0.682   0.772     0.724   
1       KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745   
2        KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695   
3  KNNbasic_item_tuned  2.903997  2.301706      0.688   0.785     0.733   
4                  SVD  2.721453  2.168153      0.696   0.798     0.744   
5            SVD_tuned  2.673590  2.143426      0.694   0.799     0.743

predictions using SVD tuned tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Since the untuned SVD model had predictions that were the closest to the users actually play counts, Im going to look at the weighted recommendations for both the default and tuned SVD algorithms. The final, weighted, recommendations from default SVD matrix factorization algorithm for user ‘6ccd111af9b4baa497aacd6d1863cbf5a141acc6’ are:

title artist_name count_plays predicted_interaction corrected_ratings
0 Catch You Baby (Steve Pitron & Max Sanna Radio... Lonnie Gordon 616 10.000000 9.959709
1 Clara meets Slope - Hard To Say Clara Hill 89 9.879379 9.773379
2 Make Her Say Kid Cudi / Kanye West / Common 156 9.186673 9.106609
3 Something (Album Version) Jaci Velasquez 95 8.532769 8.430171
4 Electric Feel MGMT 179 8.380396 8.305653
5 He's A Pirate Klaus Badelt 20 8.514556 8.290949
6 221 keller williams 51 8.379492 8.239464
7 Gunn Clapp O.G.C. 86 8.313128 8.205295
8 Call It Off (Album Version) Tegan And Sara 66 8.098702 7.975610
9 Girls Death In Vegas 25 8.072738 7.872738

And the Tuned recommendations:

title artist_name count_plays predicted_interaction corrected_ratings
0 False Pretense The Red Jumpsuit Apparatus 21 7.612781 7.394563
1 Sorrow (1997 Digital Remaster) David Bowie 77 6.734795 6.620835
2 I'd Hate To Be You When People Find Out What T... Mayday Parade 21 6.831190 6.612972
3 Recado Falado (Metrô Da Saudade) Alceu Valença 143 6.648431 6.564807
4 Q-Ball Brotha Lynch Hung 33 6.705204 6.531126
5 Underground Eminem 24 6.631778 6.427654
6 Cold Blooded (Acid Cleanse) The fFormula 38 6.565968 6.403747
7 Walk Through Hell (featuring Max Bemis Acousti... Say Anything 31 6.578550 6.398944
8 Night Village Deep Forest 20 6.575026 6.351420
9 Drive Savatage 24 6.539166 6.335042

Observations and Insights:

Cluster Based Recommendation System

In clustering-based recommendation systems, we explore the similarities and differences in people’s tastes in songs based on how they rate different songs. We cluster similar users together and recommend songs to a user based on play_counts from other users in the same cluster. After running the Coclustering method with default settings, the metrics compared to the other models are:

RMSE: 2.9591
MAE:  2.2116
             algorithm      rmse       mae  precision  recall  f1_score  \
0             KNNbasic  3.185239  2.575525      0.682   0.772     0.724   
1       KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745   
2        KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695   
3  KNNbasic_item_tuned  2.903997  2.301706      0.688   0.785     0.733   
4                  SVD  2.721453  2.168153      0.696   0.798     0.744   
5            SVD_tuned  2.673590  2.143426      0.694   0.799     0.743   
6         CoClustering  2.959111  2.211564      0.622   0.666     0.643

predictions using SVD tuned tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

Next, I run GridsearchCV on the clustering-based recommendation system to tune its hyper-hyperparameters. The optimal hyperhyperparameters suggested by the search are:

3.0094114249385258
{'n_cltr_u': 3, 'n_cltr_i': 3, 'n_epochs': 40}

The model metrics after rerunning coclustering with the tuned hyperparamters

RMSE: 2.9613
MAE:  2.2139
             algorithm      rmse       mae  precision  recall  f1_score  \
0             KNNbasic  3.185239  2.575525      0.682   0.772     0.724   
1       KNNbasic_tuned  2.830942  2.254989      0.698   0.799     0.745   
2        KNNbasic_item  2.998977  2.254952      0.658   0.736     0.695   
3  KNNbasic_item_tuned  2.903997  2.301706      0.688   0.785     0.733   
4                  SVD  2.721453  2.168153      0.696   0.798     0.744   
5            SVD_tuned  2.673590  2.143426      0.694   0.799     0.743   
6         CoClustering  2.959111  2.211564      0.622   0.666     0.643   
7   CoClustering_tuned  2.961292  2.213922      0.622   0.666     0.643

predictions using SVD tuned tuned for user 6ccd111af9b4baa497aacd6d1863cbf5a141acc6:

Observations and Insights:

The final, weighted, recommendations from the tuned coclustering algorithm for user ‘6ccd111af9b4baa497aacd6d1863cbf5a141acc6’ are:

title artist_name count_plays predicted_interaction corrected_ratings
0 Love Is Gone (Original Mix) David Guetta - Joachim Garraud - Chris Willis 20 8.561523 8.337917
1 Cold Blooded (Acid Cleanse) The fFormula 38 8.275809 8.113588
2 The World Is Mine David Guetta 21 8.056761 7.838544
3 What Is Light? Where Is Laughter? (Album Version) Twin Atlantic 20 7.872448 7.648841
4 Kelma Rachid Taha 58 7.642269 7.510962
5 Love Is Not A Fight Warren Barfield 36 7.561523 7.394857
6 I Wonder Why Dion & The Belmonts 41 7.517873 7.361699
7 Hasta siempre Varios 20 7.552595 7.328988
8 Numb (Album Version) Disturbed 68 7.418666 7.297398
9 221 keller williams 51 7.411147 7.271119

Observations and Insights:

Content Based Recommendation System

So far I have only used the play_count of songs to find recommendations but there are other information/features on songs as well, and these features can be used to increase the personalization of the recommendation system. For example, we can use the artist name and album title to recommend songs to users from artists/albums they like but songs they have not heard yet. We can also include the year the song was released and recommend music from the same time period. Here are the main functions I used for the content model:

Before Running any of the content based models, I preprocessed the text data by tolkenizing it with natural language processing methods that come standard in the NLTK package. First, I coded the year column into text:

song_id artist_name release year year_text
0 SODGVGW12AC9075A8D Justin Bieber My Worlds 2010 twothousandtens
1 SODGVGW12AC9075A8D Justin Bieber My World 2.0 2010 twothousandtens
2 SOEPZQS12A8C1436C7 Deadmau5 Ghosts 'n' Stuff 2009 twothousandzeros
3 SOGDDKR12A6701E8FA Eminem / Hailie Jade The Eminem Show 2006 twothousandzeros
4 SOWGXOP12A6701E93A Eminem Without Me 2002 twothousandzeros

Then I combined all of the text features into a single column ‘text’:

year_text text
song_id
SODGVGW12AC9075A8D twothousandtens Justin Bieber My Worlds twothousandtens
SODGVGW12AC9075A8D twothousandtens Justin Bieber My World 2.0 twothousandtens
SOEPZQS12A8C1436C7 twothousandzeros Deadmau5 Ghosts 'n' Stuff twothousandzeros
SOGDDKR12A6701E8FA twothousandzeros Eminem / Hailie Jade The Eminem Show twothousa...
SOWGXOP12A6701E93A twothousandzeros Eminem Without Me twothousandzeros

I then calculated a tf-idf matrix and calculated the cosign similarity.

The Recommendations using this method are:

title artist_name count_plays predicted_interaction corrected_ratings
0 Cats In The Cradle Ugly Kid Joe 30 7.789474 7.606899
1 Cold Blooded (Acid Cleanse) The fFormula 38 7.000000 6.837779
2 Hold Me_ Thrill Me_ Kiss Me_ Kill Me U2 33 6.770725 6.596648
3 Twilight Galaxy Metric 28 6.696697 6.507715
4 Last Nite The Strokes 21 6.724206 6.505988
5 The Calculation (Album Version) Regina Spektor 44 6.636842 6.486086
6 Mansard Roof (Album) Vampire Weekend 26 6.587426 6.391310
7 Face To Face Daft Punk 32 6.525241 6.348465
8 Injection Rise Against 23 6.485387 6.276872
9 Chinese Lily Allen 14 6.092105 5.824844

Observations and Insights:

hybrid recommendation system

I have explored five different recommender methods (six if you include the popularity-based method) for recommending songs to a user in a personalized way. This is all well and good, but how do we know which method is recommending songs the user will actualy listen to? As an avid listener of music, is it not also true that a users preferences can change, and may change drastically during the day or from day to day? It also seems that the models we have presented so far sometimes recommend similar songs, but other models are drastically different as in the item-item recommendations and content based recommendations. In this section, I build a hybrid recommender system which takes into account recommendations from multiple models, applys weights based on which models we want to be most represented in the results, and then provides a ‘hybrid’ recommendation.

In my hybrid system, I fit models and make predictions by combing two of the previously evaluated models:

I chose these two models because both had the best F1_scores metrics, both had fairly good predictions, and both recommended markedly different songs. I then add in a subset of recommendations from the content-based filtering method to increase the ‘familiarity’ for the user of the recommended songs/artists. This will hopefully provide a highly personalized recommendation for a user with a balance of new songs and familiar artists to give choices based on potential dynamic user preferences. I also evaluated the performance of different weight combinations using the hold-out set. For example, we can try different combinations of wA and wB for the SVD and User-User model combination, ranging from (0.1, 0.9) to (0.9, 0.1), and compute their respective RMSE scores:

RMSE: 2.7976
MAE:  2.2302
Weight combination (0.1, 0.9): RMSE = 2.7976, MAE = 2.2302
RMSE: 2.7495
MAE:  2.1971
Weight combination (0.3, 0.7): RMSE = 2.7495, MAE = 2.1971
RMSE: 2.7186
MAE:  2.1749
Weight combination (0.5, 0.5): RMSE = 2.7186, MAE = 2.1749
RMSE: 2.7057
MAE:  2.1634
Weight combination (0.7, 0.3): RMSE = 2.7057, MAE = 2.1634
RMSE: 2.7110
MAE:  2.1619
Weight combination (0.9, 0.1): RMSE = 2.7110, MAE = 2.1619

Observations and Insights:

Conclusion and Recommendations

1. Comparison of various techniques and their relative performance based on chosen Metric (Measure of success):

2. Refined insights:

3. Proposal for the final solution design:

Final 10 song recommendation for user 6d625c6557df84b60d90426c0116138b617b9449:

title artist_name count_plays predicted_interaction corrected_ratings
0 Clara meets Slope - Hard To Say Clara Hill 89 9.677092 9.571092
1 Catch You Baby (Steve Pitron & Max Sanna Radio... Lonnie Gordon 616 8.470891 8.430600
2 Make Her Say Kid Cudi / Kanye West / Common 156 7.772713 7.692649
3 Cats In The Cradle Ugly Kid Joe 30 7.789474 7.606899
4 Recado Falado (Metrô Da Saudade) Alceu Valença 143 7.642525 7.558901
5 #40 DAVE MATTHEWS BAND 72 7.661807 7.543956
6 Walk Through Hell (featuring Max Bemis Acousti... Say Anything 31 7.678929 7.499324
7 Eternal Flame (Single Version) Atomic Kitten 49 7.411352 7.268495
8 Cold Blooded (Acid Cleanse) The fFormula 38 7.000000 6.837779
9 Hold Me_ Thrill Me_ Kiss Me_ Kill Me U2 33 6.770725 6.596648

References

  1. Thierry Bertin-Mahieux, Daniel P.W. Ellis, Brian Whitman, and Paul Lamere. The Million Song Dataset. In Proceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR 2011), 2011.
Top of Page Raw Code