RECOMMENDATIONS SYSTEMS FOR LASTFM

Introduction:
Web-based music streaming services such as Spotify, Wynk Music, LastFM etc provide their users with
many opportunities to discover new music, whether in the form of specific pieces of music the user hadn’t
heard before or in the form of musical artists to which the user hasn’t previously been exposed. These
systems make use of sevaral Recommendation Tools as part of their efforts to further engage their user
base. Given the widespread use of such methodologies for enabling the discovery of new music, today’s
web-based streaming environment offers ample opportunity for those interested in exploring both the
methods typically used for constructing recommender systems and how such systems can effectively be
applied to enable the discovery of novel content.

Objective:
The goal of this project is to provide artists recommendation to LastFM users based on several
recommendation systems and choose the best recommendation system by evaluating different models
based on quantitative and qualitative techniques.

Data Collection:
The data set to be used is comprised of music listening information for a set of 1,892 users of
the Last.fm online music system. A total of 17,632 distinct musical artists are represented within the data
set. 11946 Tags/genres were included in this data.

Types of Recommendation Systems Developed:

• Collaborative filtering recommendation system.
• Content Based Recommendation System.
• Hybrid Recommendation Systems.

1. Discussion of the transformations you use in step 1.

Collaborative Filtering:
For Collaborative filtering, we started with the given data Artists.dat. The data consisted of UserID, ArtistID
and weight. Weight refers to the number of times the user has listened to a specific artist.
On further evaluation of the given raw data, we found that the minimum value of weight is 1 while
maximum value is 352698. The average turning out to be 745.24. With such a huge range, we decided to
further check the skewness of the weights. On plotting the data values on a normal function, we found
high skewness in the data.
With such a high skewness, the recommendation matrix will not be right, as the available categories are
wide, and the recommendation results would be inaccurate. Thus, to remove the skewness of the data,
we did some mathematical transformation of the weights.

We took a logarithm with base 10 of the weights in the first step. This was not enough as the weights with
1 will have a log value of 0. Hence as a second step we treated the following log weights with linear model
to obtain the intercept/constant value. After completing the transformations, the constant value was
5.389e+00. This was then added to the logarithm values weight, the resultant weights when plotted on a
histogram showed a uniform distribution ranging from ratings 5.389 to 18.162.

‘’ user_artists$log_weight = log(user_artists$weight) + 5.389e+00 ‘’

We had to subset the data as the given data was taking long to get processed, it was better for us to
train our models on a subset of the entire data. The main question being, out of 17000+ artists, which
one to be chosen for the subset. With some brainstorming, we used the MOST Popular artists for the
dataset. We defined the most popular artists as the ones who are very frequently heard by user/users
and have been heard by more than a specific number of users and not only followed by one or two
users. This clearly defines that most popular artists would be better for our modeling as the most
popular artists recommendation would be more beneficial for the company. With such a subset, we
could capitalize the most sorted artists, thus in turn increasing the quality of data while reducing the
quantity.

For our case, we filtered the artists with log weight greater than or equal to 10 and number of unique
users heard to the artist as more than or equal to 5. With these constraints, we brought down the data
set with 1943 unique artists and 1876 unique users.

CONTENT BASED RECOMMENDATIONS SYSTEMS:

User Tagged Artists and Tags Files were imported into R.

The user_taggedartists.dat file contains a listing of each instance in which a last.fm user has assigned a
musical genre label (a.k.a., a “tag”) to an artist. Though the file also contains the date (day, month, and
year) of the “tagging”, those attributes will be ignored for purposes of this project.
The file is found to contain a total of 186,479 total instances of last.fm users having applied a genre tag
to an artist. All 1,892 users are represented within the file, and a total of 12,523 distinct artists have
been tagged with at least one genre name.
The tags.dat file contains a list of musical genres that last.fm users have used to categorize the various
musical artists represented within the last.fm online music streaming platform. Each genre is assigned a
unique identifier, or “tagID”. A count of the unique tagID’s reveals the presence of 11,946 distinct genre
names:

Selection of Final Tags:

We aggregated the number of tags per user and we took only those Top 200 tags which are listened by
the greatest number of distinct users.

Selection of Final Artists:

We only kept those Artists which are used in the Collaborative Filter Recommendation Systems. To get
a match between both the RS.

Creating an Artist - Genre Matrix:

As part of our content-based recommendation efforts for this project, we would like to be able to
provide last.fm users with the ability to receive a “Top N” list of suggested musical artists who are likely
to be like an artist selected by the user. Furthermore, we’d like to be able to provide a user with a “Top
N” list of suggested of artists for a user-selected musical genre. A path to the achievement of each of
these objectives can be found through the creation of an artist-genre matrix.
Earlier were able to reduce our users-taggedartists data by limiting it to the top 200 genres as
determined by the number of times each genre has been applied within the data set. Furthermore, we
were able to calculate the number of times each artist had been labeled with a genre tag. We now use
the results of those calculations as the basis of an artist-genre matrix.

A portion of the artist-genre matrix is shown below. The row names within the matrix represent the
unique last.fm artist ID’s (which are present in CF Matrix as well) while the column names represent the
unique last.fm genre tag ID’s we extracted from the tags.dat (Top 200 as explained above).The Matrix
was binarized.

Artist-genre matrix was used for purposes of generating an artist similarity matrix. Though an argument
could be made in favor of making use of the raw genre tagging counts for each artist as the basis of a
similarity matrix, an equally strong argument can be made in favor of treating all such tags as binary
indications of whether last.fm users consider an artist to belong to a given genre. Therefore, we made
use of a binary version of the artist-genre matrix for purposes of generating an artist similarity matrix.

2. Creation of Functions:

USER BASED CF – Pearson correlation

With the prepared dataset, we started modeling the UBCF. With the possibility to choose the similarity
function among Pearson correlation and cosine, we defined the function for Pearson correlation using
the mathematical formula

The above code snippet shows the Pearson correlation function definition for the user-based CF for
multiple users. Similarly, Pearson correlation was defined for single user
ITEM BASED CF
After User Based CF, we started modeling the recommendation system based on item-item similarity or
IBCF. For this, we again defined the similarity function with Pearson correlation and determined the
similarity between the items/artists based on the log weight. With some relevant updates and
modification in the previous Pearson correlation function, the new similarity function was defined for
Item Based CF. The code snippet above shows the formulation of the same. For recommendation, the
reference function was used to find the recommendation on based of item similarities.

Evaluation Metrics - MAE

After defining the function for UBCF and IBCF, it was time to define the functions which could calculate
the accuracy of the models. Not solely accuracy, but also the performance was gauged by evaluation
metrics like RSME, MAE and classification accuracy.

The following code snippet shows the function definition for MAE
Evaluation Metrics – F1
For defining the function for F1 metrics, we used the existing classification function and defined the F1
formula as the below snippet.

3. Argumentation on using specific recommendation technique, evaluation

systems and hybridization technique.

Recommendation technique:

Collaborative filtering:

We used collaborative filtering to start with as this is a very relevant solution for the given
problem statement. This commonly used recommendation system fits well with the dataset we
have. With around 1900 unique artists, we can recommend every type of user a different artist
according to user-user and artist-artist similarity.

Within collaborative filtering, we used three different techniques like Userbased, Item based
and Userbased techniques to check the results with different techniques with evaluation metrics
and suggest the best one along them.

Content based:

In this Project as we deal with various Artists and genre of music, its highly likely that if there is a
recommendation on Artists like a given artist or recommendation on genre based on genre /
artists becomes very helpful to the users in the real-world music websites. Hence, we choose
the Content based Recommendation systems with Artists and Genre / Tags.

Hybrid Based:

In recommendation systems always, it is better to use a hybrid, by combining the two

recommendation systems that give better accuracies and reduces sensitivity of the data.

In our case we used two hybrid techniques for comparison

Cluster based + Content Based:

In a website like LastFm the Clustering based Recommendation system helping in understanding
various types of similar users and if we recommend content to each class of user defined by the
cluster it would be helpful, so we tried this recommendation system in our case study.

Userbased + Content Based:

In this approach based on nearest neighbor of users the content would be recommended. For
example, user listens to artist A and user B is close to user so recommend him the artists heard
by user and the artists close to those artists. The algorithms calculate the better between these
2 and recommends finally the artists.

Evaluation Metrics:

We used different evaluation metrics such as prediction accuracy, ranking accuracy and
classification accuracy to compare the results of different recommendation techniques.

RMSE, Recall, precision and F1 are the most commonly used evaluation metrics in
recommendation systems. In order to evaluate prediction accuracy, Root mean square error and
mean absolute error was calculated. Further, we wanted to evaluate product relevancy and thus
we calculated the classification accuracy metrics Recall, Precision and F1. Finally, we also
calculated ranking accuracy like AUC for userbased and item based collaborative filtering to
evaluate based on the ranking of the user.
 4. A final comparative discussion of the recommendation systems
identifying the best one in step 3.

Please find the below results of several recommendation systems.

Own functions:

Evaluation & USER ITEM BASED Cluster Content Hybrid Hybrid

Classification BASED CF CF Based CF Based (Cluster (Content +
CF + Userbased
Content collaborative)
based)
RMSE 0.09884 1.4935 0.1762 1.20905 0.8892 0.602
MAE 0.01027 0.2001 - 0.153425 0.10588 0.07
Precision 0.8 0.666667 0.805081 1 0.99627 1
Recall 0.4 0.1 0.908086 0.33287 0.43274 0.9316
F1 0.53333 0.173913 0.853487 0.49948 0.60339 0.964
AUC 0.5476 0.5033 - - - -

Recommenders lab:

Prediction POPULAR ITEM BASED CF USER BASED CF HYBRID(SVD)

accuracy
RMSE 0.87536 0.93223 0.88048 0.86481
MSE 0.76626 0.86906 0.77529 0.74800
MAE 0.64196 0.72331 0.644638 0.63433

The above results show us that the RMSE and MAE values for User based is the best as it has lowest
along with a good value for precision. This shows us that userbased collaborative filtering is the optimal
quantitative results for this case Study. We can also say that User based is the best as we have more
number of items than the users.

Below are the top 3 artists predictions for few users using userbased collaborative filtering

Users V1 V2 V3
5 65 1375 81
8 306 466 913
12 930 718 1854
14 227 5446 2893
15 868 220 1412
20 603 2605 2677
23 1014 1713 58
30 298 681 294
31 1713 1019 1014
But for Qualitative results we found that the Content Based Recommendation systems was second best
performing close to real life and recommending artists like the artists and recommending top artists for
a genre as shown below:

Qualitative Results:
Display a List of 4 Recommended Artists like a Given Artist

The artist similarity matrix allows Last.fm users to find musical artists that are like one they specify. To
simulate how this might work, we can randomly select an artist ID and display a list of 5 recommended
similar artists:

Generate a Top N Artist List by Genre:

We made use of the non-binary version of the artist-genre matrix to provide a user with a “Top N” list of
suggested of artists from a user-selected musical genre. Recall that the artist-genre matrix is comprised
of counts of how often a given artist has been labeled as belonging to a genre. This metric can serve as a
proxy for how strongly the last.fm user community feels that an artist belongs to a genre. Therefore, we
can use the tag counts to rank artists within genres: The more often they’ve been tagged with a genre
label, the higher they rank within the genre.
 5. A discussion of advantages and disadvantages of the best
recommendation system(s) and some thoughts for improvement.

Advantages:

Userbased:
• The main advantage of using user-based system is it is most appropriate collaborative
technique when there are less users than items and gives good results.

Content Based:
• Comparison between the Items is possible. And User gets recommendation to similar
types of music on his past playlist.
• Content based recommended system overcomes the challenges of collaborative filtering
techniques. They have the ability to recommend even if there are no ratings provided by
users for new items. It has the capacity to adjust its recommendations in a short span of
time if the user preferences change. Content based filtering technique can also provide
explanations on how recommendations are generated to users.

Disadvantages:

Userbased:
• Cold start problem, where a new user/item with no ratings is added, is extremely
difficult to deal with using these methods. This issue is actually quite recurrent when
designing recommendation systems so keep it in mind!
• As we use a user's own ratings to make a prediction. If we want to predict User1 rating
for Rock Music, the item-based approach will look at User1 ratings for Related features
to Rock Music, (thus very similar to the target Music). Meanwhile, with the user-based
approach, we look at similar users who might have overlapping but different interests
than User1 which may result in a decrease in accuracy.

Content Based:
• Overspecialization is one of the main concerns generally faced.

Case Study By: Priya Varadarajan,Harshit Paliwal,Nitesh Shanbhag, Manju NR, Ruturaj Mokashi