• WHAT? The problem is to understand the mood of a listener to recommend tracks based on their current mood

• WHY? We want to enhance the listener’s user experience. So, imagine if an angry person gets suggestion to listen to party music, it may lead to a bad user experience.

• HOW? Develop a predictive analytical product that can predict the mood of a person, provided we have data about his/her recently played songs.

Folder that contains data we collected from the API: https://github.com/doyeunlim87/CIS-8392-Course-Project/

The above mentioned data files contain the information that we have used for our analysis

Recap of our proposal:

We were able to acheive what we have proposed with an accuracy better than what was initially expected.

• Taking into account the time at which a user listens to a particular genre: No time stamps

• Are there more attributes available? If yes, could you prioritize among all the available attributes to make better decisions?: Yes available. But we don’t need to consider them as we are able to get same results with the default 5 attributes.

• Do you collect the mood data from the same APIs as other 5 features? : No, we labelled the clusters with critical human thinking

• How can you use this to boost your profit? Rule 1: Know your customer better to gain the revenue.

• The data used in this project was collected by downloading a list of 3,000 artists from the following github repository

• This list of artists are simply the top 3,000 most popular artists in the United States

• These songs have 5 musical attributes as follows:

1. Acousticness
2. Danceability
3. Liveliness
4. Loudness
5. Speechiness

• No other data was pulled outside of the API.
• Sample size in your collected data were more than 20,000 instances.
• R package to make API call-Spotify https://cran.r-project.org/web/packages/spotifyr
• Some limitations on the number, frequency, and speed of your API calls were addressed by proper planning.

• Each audio feature describes the song using numeric values.
• Every feature has a scale which varies from 0 to 1 except loudness.
• Loudness is on a scale from -60db to 0db and was thus rescaled to vary from 0 to 1.
• Only required pre-processing (normalizing). The data was normalized so that all variables are on the same scale.

• The collected data was visualized using several methods.
• The first method included plotting a distance matrix.
• Therefore, it seems reasonable to use some type of distance metric and classify the set of songs using some type of clustering algorithm.
• Use K-meand Elbow Method to find optimal number of clusters. K-means Elbow Method results indicate optimal number of clusters = 5.

• Add cluster labels to original dataset
• We choose to use K-means clustering since it is a simple, intuitive and time-proven method for clustering. To determine the optimal number of clusters
• Split the data into training and testing dataset
• Start training supervised classification algorithms on the training set.

• We have tried SVM and Random forest.
• We have acheived 0.98 F1 score with SVM and stood out to be the best
• Other algorithms performed well too.
• Key take-aways from your findings: Making use of this important data to monitor a person’s mood on real time basis is a sophisticated process for Spotify. However, we believe that the reason how these insights translate into money in a business setting is worth the investment of time and money to implement this anlytical tool.