Predict Billboard Top 10 Hits Using RStudio, H2O and Amazon Athena
Success in the popular music industry is typically measured in terms of the number of Top 10 hits artists have to their credit. The music industry is a highly competitive multi-billion dollar business, and record labels incur various costs in exchange for a percentage of the profits from sales and concert tickets.
Predicting the success of an artist’s release in the popular music industry can be difficult. One release may be extremely popular, resulting in widespread play on TV, radio and social media, while another single may turn out quite unpopular, and therefore unprofitable. Record labels need to be selective in their decision making, and predictive analytics can help them with decision making around the type of songs and artists they need to promote.
In this walkthrough, you leverage H2O.ai, Amazon Athena, and RStudio to make predictions on whether a song might make it to the Top 10 Billboard charts. You explore the GLM, GBM, and deep learning modeling techniques using H2O’s rapid, distributed and easy-to-use open source parallel processing engine. RStudio is a popular IDE, licensed either commercially or under AGPLv3, for working with R. This is ideal if you don’t want to connect to a server via SSH and use code editors such as vi to do analytics. RStudio is available in a desktop version, or a server version that allows you to access R via a web browser. RStudio’s Notebooks feature is used to demonstrate the execution of code and output. In addition, this post showcases how you can leverage Athena for query and interactive analysis during the modeling phase. A working knowledge of statistics and machine learning would be helpful to interpret the analysis being performed in this post.
Your goal is to predict whether a song will make it to the Top 10 Billboard charts. For this purpose, you will be using multiple modeling techniques―namely GLM, GBM and deep learning―and choose the model that is the best fit.
This solution involves the following steps:
- Install and configure RStudio with Athena
- Log in to RStudio
- Install R packages
- Connect to Athena
- Create a dataset
- Create models
Install and configure RStudio with Athena
Use the following AWS CloudFormation stack to install, configure, and connect RStudio on an Amazon EC2 instance with Athena.
Launching this stack creates all required resources and prerequisites:
- Amazon EC2 instance with Amazon Linux (minimum size of t2.large is recommended)
- Provisioning of the EC2 instance in an existing VPC and public subnet
- Installation of Java 8
- Assignment of an IAM role to the EC2 instance with the required permissions for accessing Athena and Amazon S3
- Security group allowing access to the RStudio and SSH ports from the internet (I recommend restricting access to these ports)
- S3 staging bucket required for Athena (referenced within RStudio as ATHENABUCKET)
- RStudio password (Note: username is rstudio)
- Setup logs in Amazon CloudWatch Logs (if needed for additional troubleshooting)
- Amazon EC2 Systems Manager agent, which makes it easy to manage and patch
All AWS resources are created in the US-East-1 Region. To avoid cross-region data transfer fees, launch the CloudFormation stack in the same region. To check the availability of Athena in other regions, see Region Table.
Log in to RStudio
The instance security group has been automatically configured to allow incoming connections on the RStudio port 8787 from any source internet address. You can edit the security group to restrict source IP access. If you have trouble connecting, ensure that port 8787 isn’t blocked by subnet network ACLS or by your outgoing proxy/firewall.
- In the CloudFormation stack, choose Outputs, Value, and then open the RStudio URL. You might need to wait for a few minutes until the instance has been launched.
- Log in to RStudio with the and password you provided during setup.
Install R packages
Next, install the required R packages from the RStudio console. You can download the R notebook file containing just the code.
Connect to Athena
Next, establish a connection to Athena from RStudio, using an IAM role associated with your EC2 instance. Use ATHENABUCKET to specify the S3 staging directory.
Verify the connection. The results returned depend on your specific Athena setup.
Create a dataset
For this analysis, you use a sample dataset combining information from Billboard and Wikipedia with Echo Nest data in the Million Songs Dataset. Upload this dataset into your own S3 bucket. The table below provides a description of the fields used in this dataset.
|Year that song was released
|Title of the song
|Name of the song artist
|Unique identifier for the song
|Unique identifier for the song artist
|Variable estimating the time signature of the song
|Confidence in the estimate for the timesignature
|Continuous variable indicating the average amplitude of the audio in decibels
|Variable indicating the estimated beats per minute of the song
|Confidence in the estimate for tempo
|Variable with twelve levels indicating the estimated key of the song (C, C#, B)
|Confidence in the estimate for key
|Variable that represents the overall acoustic energy of the song, using a mix of features such as loudness
|Continuous variable that indicates the pitch of the song
|timbre_0_min thru timbre_11_min
|Variables that indicate the minimum values over all segments for each of the twelve values in the timbre vector
|timbre_0_max thru timbre_11_max
|Variables that indicate the maximum values over all segments for each of the twelve values in the timbre vector
|Indicator for whether or not the song made it to the Top 10 of the Billboard charts (1 if it was in the top 10, and 0 if not)
Create an Athena table based on the dataset
In the Athena console, select the default database, sampled, or create a new database.
Run the following create table statement.
Inspect the table definition for the ‘billboard’ table that you have created. If you chose a database other than sampledb, replace that value with your choice.
Run a sample query
Next, run a sample query to obtain a list of all songs from Janet Jackson that made it to the Billboard Top 10 charts.
Determine how many songs in this dataset are specifically from the year 2010.
The sample dataset provides certain song properties of interest that can be analyzed to gauge the impact to the song’s overall popularity. Look at one such property, timesignature, and determine the value that is the most frequent among songs in the database. Timesignature is a measure of the number of beats and the type of note involved.
Running the query directly may result in an error, as shown in the commented lines below. This error is a result of trying to retrieve a large result set over a JDBC connection, which can cause out-of-memory issues at the client level. To address this, reduce the fetch size and run again.
From the results, observe that 6787 songs have a timesignature of 4.
Next, determine the song with the highest tempo.
Create the training dataset
Your model needs to be trained such that it can learn and make accurate predictions. Split the data into training and test datasets, and create the training dataset first. This dataset contains all observations from the year 2009 and earlier. You may face the same JDBC connection issue pointed out earlier, so this query uses a fetch size.
Create the test dataset
Convert the training and test datasets into H2O dataframes
Inspect the column names in your H2O dataframes.
You need to designate the independent and dependent variables prior to applying your modeling algorithms. Because you’re trying to predict the ‘top10’ field, this would be your dependent variable and everything else would be independent.
Create your first model using GLM. Because GLM works best with numeric data, you create your model by dropping non-numeric variables. You only use the variables in the dataset that describe the numerical attributes of the song in the logistic regression model. You won’t use these variables: “year”, “songtitle”, “artistname”, “songid”, or “artistid”.
Create Model 1: All numeric variables
Create Model 1 with the training dataset, using GLM as the modeling algorithm and H2O’s built-in h2o.glm function.
Measure the performance of Model 1, using H2O’s built-in performance function.
The AUC metric provides insight into how well the classifier is able to separate the two classes. In this case, the value of 0.8431394 indicates that the classification is good. (A value of 0.5 indicates a worthless test, while a value of 1.0 indicates a perfect test.)
Next, inspect the coefficients of the variables in the dataset.
Typically, songs with heavier instrumentation tend to be louder (have higher values in the variable “loudness”) and more energetic (have higher values in the variable “energy”). This knowledge is helpful for interpreting the modeling results.
You can make the following observations from the results:
- The coefficient estimates for the confidence values associated with the time signature, key, and tempo variables are positive. This suggests that higher confidence leads to a higher predicted probability of a Top 10 hit.
- The coefficient estimate for loudness is positive, meaning that mainstream listeners prefer louder songs with heavier instrumentation.
- The coefficient estimate for energy is negative, meaning that mainstream listeners prefer songs that are less energetic, which are those songs with light instrumentation.
These coefficients lead to contradictory conclusions for Model 1. This could be due to multicollinearity issues. Inspect the correlation between the variables “loudness” and “energy” in the training set.
This number indicates that these two variables are highly correlated, and Model 1 does indeed suffer from multicollinearity. Typically, you associate a value of -1.0 to -0.5 or 1.0 to 0.5 to indicate strong correlation, and a value of 0.1 to 0.1 to indicate weak correlation. To avoid this correlation issue, omit one of these two variables and re-create the models.
You build two variations of the original model:
- Model 2, in which you keep “energy” and omit “loudness”
- Model 3, in which you keep “loudness” and omit “energy”
You compare these two models and choose the model with a better fit for this use case.
Create Model 2: Keep energy and omit loudness
Measure the performance of Model 2.
You can make the following observations:
- The AUC metric is 0.8431933.
- Inspecting the coefficient of the variable energy, Model 2 suggests that songs with high energy levels tend to be more popular. This is as per expectation.
- As H2O orders variables by significance, the variable energy is not significant in this model.
You can conclude that Model 2 is not ideal for this use , as energy is not significant.
CreateModel 3: Keep loudness but omit energy
You can make the following observations:
- The AUC metric is 0.8492389.
- From the confusion matrix, the model correctly predicts that 33 songs will be top 10 hits (true positives). However, it has 26 false positives (songs that the model predicted would be Top 10 hits, but ended up not being Top 10 hits).
- Loudness has a positive coefficient estimate, meaning that this model predicts that songs with heavier instrumentation tend to be more popular. This is the same conclusion from Model 2.
- Loudness is significant in this model.
Overall, Model 3 predicts a higher number of top 10 hits with an accuracy rate that is acceptable. To choose the best fit for production runs, record labels should consider the following factors:
- Desired model accuracy at a given threshold
- Number of correct predictions for top10 hits
- Tolerable number of false positives or false negatives
Next, make predictions using Model 3 on the test dataset.
The first set of output results specifies the probabilities associated with each predicted observation. For example, observation 1 is 96.54739% likely to not be a Top 10 hit, and 3.4526052% likely to be a Top 10 hit (predict=1 indicates Top 10 hit and predict=0 indicates not a Top 10 hit). The second set of results list the actual predictions made. From the third set of results, this model predicts that 61 songs will be top 10 hits.
Compute the baseline accuracy, by assuming that the baseline predicts the most frequent outcome, which is that most songs are not Top 10 hits.
Now observe that the baseline model would get 314 observations correct, and 59 wrong, for an accuracy of 314/(314+59) = 0.8418231.
It seems that Model 3, with an accuracy of 0.8552, provides you with a small improvement over the baseline model. But is this model useful for record labels?
View the two models from an investment perspective:
- A production company is interested in investing in songs that are more likely to make it to the Top 10. The company’s objective is to minimize the risk of financial losses attributed to investing in songs that end up unpopular.
- How many songs does Model 3 correctly predict as a Top 10 hit in 2010? Looking at the confusion matrix, you see that it predicts 33 top 10 hits correctly at an optimal threshold, which is more than half the number
- It will be more useful to the record label if you can provide the production company with a list of songs that are highly likely to end up in the Top 10.
- The baseline model is not useful, as it simply does not label any song as a hit.
Considering the three models built so far, you can conclude that Model 3 proves to be the best investment choice for the record label.
H2O provides you with the ability to explore other learning models, such as GBM and deep learning. Explore building a model using the GBM technique, using the built-in h2o.gbm function.
Before you do this, you need to convert the target variable to a factor for multinomial classification techniques.
This model correctly predicts 43 top 10 hits, which is 10 more than the number predicted by Model 3. Moreover, the AUC metric is higher than the one obtained from Model 3.
As seen above, H2O’s API provides the ability to obtain key statistical measures required to analyze the models easily, using several built-in functions. The record label can experiment with different parameters to arrive at the model that predicts the maximum number of Top 10 hits at the desired level of accuracy and threshold.
H2O also allows you to experiment with deep learning models. Deep learning models have the ability to learn features implicitly, but can be more expensive computationally.
Now, create a deep learning model with the h2o.deeplearning function, using the same training and test datasets created before. The time taken to run this model depends on the type of EC2 instance chosen for this purpose. For models that require more computation, consider using accelerated computing instances such as the P2 instance type.
The AUC metric for this model is 0.7568822, which is less than what you got from the earlier models. I recommend further experimentation using different hyper parameters, such as the learning rate, epoch or the number of hidden layers.
H2O’s built-in functions provide many key statistical measures that can help measure model performance. Here are some of these key terms.
|Measures the proportion of positives that have been correctly identified. It is also called the true positive rate, or recall.
|Measures the proportion of negatives that have been correctly identified. It is also called the true negative rate.
|Cutoff point that maximizes specificity and sensitivity. While the model may not provide the highest prediction at this point, it would not be biased towards positives or negatives.
|The fraction of the documents retrieved that are relevant to the information needed, for example, how many of the positively classified are relevant
Provides insight into how well the classifier is able to separate the two classes. The implicit goal is to deal with situations where the sample distribution is highly skewed, with a tendency to overfit to a single class.
0.90 – 1 = excellent (A)
0.8 – 0.9 = good (B)
0.7 – 0.8 = fair (C)
.6 – 0.7 = poor (D)
0.5 – 0.5 = fail (F)
Here’s a summary of the metrics generated from H2O’s built-in functions for the three models that produced useful results.
|Deep Learning Model
Note: ‘t’ denotes threshold.
Your options at this point could be narrowed down to Model 3 and the GBM model, based on the AUC and accuracy metrics observed earlier. If the slightly lower accuracy of the GBM model is deemed acceptable, the record label can choose to go to production with the GBM model, as it can predict a higher number of Top 10 hits. The AUC metric for the GBM model is also higher than that of Model 3.
Record labels can experiment with different learning techniques and parameters before arriving at a model that proves to be the best fit for their business. Because deep learning models can be computationally expensive, record labels can choose more powerful EC2 instances on AWS to run their experiments faster.
In this post, I showed how the popular music industry can use analytics to predict the type of songs that make the Top 10 Billboard charts. By running H2O’s scalable machine learning platform on AWS, data scientists can easily experiment with multiple modeling techniques and interactively query the data using Amazon Athena, without having to manage the underlying infrastructure. This helps record labels make critical decisions on the type of artists and songs to promote in a timely fashion, thereby increasing sales and revenue.
If you have questions or suggestions, please comment below.
About the Authors
Gopal Wunnava is a Partner Solution Architect with the AWS GSI Team. He works with partners and customers on big data engagements, and is passionate about building analytical solutions that drive business capabilities and decision making. In his spare time, he loves all things sports and movies related and is fond of old classics like Asterix, Obelix comics and Hitchcock movies.
Bob Strahan, a Senior Consultant with AWS Professional Services, contributed to this post.