Project URL As mentioend on piazza, we don't need to deploy the project
Video URL https://youtu.be/5e2C-uLXFgs
The data science problem addressed in this project is the exploration and understanding of the evolution of music over time using a curated dataset from the Interdisciplinary Contest in Modeling (ICM) for Problem D in 2021. The dataset encompasses various characteristics of music, such as acousticness, energy, instrumentalness, loudness, tempo, explicitness, and frequency of musical keys, among others. The objective is to gain insights into music trends, artist characteristics, and the influence of past music on new compositions.
The Streamlit application developed for this purpose consists of four main sections: Data by Year, Data by Artist, Data by Song, and Music Influence. In the Data by Year section, informative visualizations are presented to highlight trends and changes in music characteristics over different decades, accompanied by explanations of significant events influencing these trends. The Data by Artists section provides a characteristic overview of each artist, allows comparison between artists, and ranks artists by greatest and smallest popularity. The Data by Song section allows users to explore the dataset comprehensively through three tabs, offering an overview of the dataset, ranking songs based on popularity, and comparing individual songs across various attributes. The Music Influence section delves into influencer-follower relationships among artists, employing innovative techniques such as the "pivot-melt" approach for constructing stacked distribution charts and incorporating interactive network visualizations.
Our solution effectively addresses the problem by providing a user-friendly interface that facilitates in-depth exploration of the dataset, enabling users to uncover patterns, correlations, and influential factors in the realm of music. The incorporation of diverse visualizations and interactivity enhances the user experience and contributes to a comprehensive understanding of the multifaceted aspects of music over time.
Music has been an integral part of human societies for centuries. As part of an effort to understand the role of music in the collective human experience, we used the curated dataset from the Interdisciplinary Contest in Modeling (ICM) for Problem D in 2021 and developed a Streamlit application to visualize music evolution. There are many influence factors for music creation, such as the artist's own personal experience, social events, or access to new pieces of technology. Our goal is to understand the evolution of music from multiple perspectives, including time series, artist characteristics, song characteristics, and measure the influence of previously produced music on new music and artists.
For this application's Data by Year page, we incorporated information that aims to provide some context for trends in the data. We researched the history of Western music in the 20th century, as Western music makes up the bulk of this dataset. The sources we used for contextual information to explain these trends are cited at the bottom of each page if any sources were used. The trajectory of our research was dictated by first inspecting the trends in the Data by Year dataset, picking out notable changes in the trends (e.g., a large change in the values for acoustics starting in the 1950s), and using that as a starting point for events to look into.
This page inspects the Data by Year dataset, which aggregates the music characteristic values of all songs in the dataset released in the same year. Below, we discuss the content shown on this page and the layout design.
This page was meant to be more of an informative/educational one rather than one focused on interactive visualizations as this particular dataset was limited in how much granularity it contained. We focused on the strongest trends we observed in this dataset through EDA and dedicated a visualization to each trend followed by an explanation of the events or developments in certain time periods that may have contributed to those trend changes. We discussed the rise of electronic music, dominant keys, changes in tempo and loudness, and a new feature we created from the full dataset called "explicitness," which measures the proportion of songs from each year, which contain explicit lyrics.
This page contains six different visualizations, each with an accompanying "Data" tab so that the data plotted in each visualization is available to the user if they wish to use it without cluttering the page with tables.
The first visualization illustrates the rapid and sustained changes in the features acoustics, energy, and instrumentalness in the decades between the 1950s and 1980s. The plot is followed by our hypothesis of what caused this trend. This format of plot-explanation pairings are repeated for the next four visualizations that display loudness, tempo, explicitness, and frequency of musical keys.
The last visualization on this page is an interactive one where the user can select a number of the features in this dataset to display on one chart in the time range of their choice. To standardize the scale of this plot, we limit the choices of features to those with a range of 0 to 1.
This section summarizes details of the design choices, techniques, and algorithms we developed for the aggregated data on artists in the full music dataset.
In the data_by_artist.csv file, while all the values in the artist_id columns were distinct, the associated column values were the same. That is, the same artist was assigned multiple artist_id values. This duplication was discovered by running a profiler and inspecting the report. Subsequently, the 41 artists for whom the data was duplicated were cleaned to keep the first data instance. This changed the number of data points from 5854 rows × 16 columns to 5810 rows × 16 columns. Please look into the EDA_data_by_artist.ipynb notebook to review the entire EDA process with profiling, cleaning, and visualizations. The cleaned data file was saved as cleaned_data_by_artist.csv.
The "Data by Artist" interactive page allows you to explore the artists in the dataset and the characteristics of their music. This page features three tabs for users to interact with and a tab to view the underlying data. The first tab provides a simple overview of the details of a single artist based on the user’s choice. The dropdown & search bar lets the user understand the musical attributes influencing a particular artist's songs. The second tab allows users to compare these musical attributes for multiple artists, up to 8 at a time. It also visually captures their popularity ranking to help users understand what characteristics of music tend to be audience favorites, leading to a better popularity score. The third tab lets users query the dataset's top k and bottom k artists based on their popularity scores.
The pair plot visualization featuring the bar and radar charts for top k and bottom k artists uncovers the relationship between the characteristics of music that reflect a high popularity score. The top artists focus on gearing their music towards high valence, energy, and danceability, while the unpopular artists tend to have high values of acouticness, instrumentalness, and valence.
This section summarizes our steps for data cleaning, design choices, and developing a Streamlit application for the Data by Song page.
The dataset we used in this section is the full music dataset that contains information for 98,340 songs released between 1921 and 2020. Since we obtained the dataset from ICM, it has already been sufficiently cleaned. To adapt the data for use in our Streamlit application, we made minimal changes to the data, such as converting datatypes of specific columns.
To effectively and thoroughly understand the data, we decided to use three tabs within this page: an overview tab, a song ranking tab, and a song comparison tab.
First, the overview tab is a static tab that shows the user the original dataset, statistics of numerical features of the original dataset, such as min, max, and standard deviation, and a correlation matrix of the numerical features in the dataset. We used this tab as an introduction, allowing the user to examine the data directly. We hope this static introduction will help the users appreciate the interactivities in other tabs more.
Next, the song ranking tab will allow the user to see the most and least popular songs in the dataset. The user can filter songs by date, choose how many songs to display, and set the minimum and maximum popularity of songs to be displayed. We used Altair to create two horizontal bar plots, plotting the top and bottom simultaneously and color coding the popularity, allowing the user to easily visualize the difference in popularity.
Lastly, the song comparison tab will allow the user to freely choose up to 6 songs in the dataset and compare their characteristics. The reason we decided on 6 songs are for aesthetic reasons, as more songs in the visualization make it harder to distinguish. The visualization we decided to use is a polar graph with a plot, which can easily allow the comparison of multiple entries with multiple numeric features. For the convenience of song searching, we sorted the dataset by song title and concatenated the artist names to the end of the song title so that users can search songs by title or artist.
We used Streamlit as our main framework for creating the application. In the Data by Song page, we also used Python libraries, including pandas, for data processing, matplotlib, Altair, and plotly for creating visualizations. More specifically, matplotlib was used to create the static correlation matrix in the overview tab, Altair was used to create the dynamic song ranking bar plot and plotly was used to create the dynamic song comparison polar plot.
The influence dataset proves to be pretty neat without null values or duplicates. There are two unique attributes of this dataset. Firstly, it contains 42770 influencer-follower relationships. Secondly, this dataset has a "main genre" feature associated with each artist, which is found in no other datasets. To best leverage both, we analyze the data at the " genre " and "influencer" levels. When computing statistics at the genre level, "influencers" and "followers" are "equivalent"- both are artists. So, we prepare a "full_artists" data frame that vertically stacks influencer and follower data. When focusing on the influencer level, we aim to answer questions like "Who is the most impactful artist of all time?" Thus, we prepare another "influencer" data frame, extracting information only regarding the influencers. With these datasets prepared, we efficiently pave the way for the following steps.
Among various data manipulation techniques, the most creative must be the "pivot-melt" approach for constructing the "Artist Population In Top N Genres Over Time", a stacked distribution chart. Pivot table is an efficient technique to re-structure our data frame so that one of its column contents may be pivoted as the row, making subtotal/total summary extra convenient. After we obtain this summary pivot table, we "melt" it to conform to the input standard of Altair, reformating a wide data frame to a long one. These transformations make our analysis flexible and efficient.
One highlight of this section is our incorporation of network visualizations. A network is a graph consisting of nodes and edges. Our research shows various ways to visualize a network via libraries such as "networkx", "pyvis", etc. To cater to the interactive nature of our app, however, we choose a library that 1) is compatible with Streamlit 2) is fun to interact with. "streamlit-agraph" perfectly meets our criteria. It allows users to add nodes and edges with customized attributes and provides an interactive interface where users can "drag" the network around to inspect the details. We implemented networks for both genres and artists for the audience to fully explore the dynamics.
The time series in this dataset highlight trend changes in certain music characteristics over the years. We present two interesting insights from this section below.
One of the strongest trends we observed in this dataset is how sharply the feature acousticness dropped starting around the 1950s. Acousticness is defined as the creation of sound solely through non-electronic instruments, so the first hypothesis for this trend change that we investigated was when the first eletronic musical instruments were developed. We found that while the advent of electronic synthesizers occurred in the 1940s, the 1950s onwards was when these new electronic means of creating music began to catch on and permeate throughout the music industry. This seemed to strongly affirm our hypothesis, and we also looked into features like energy and instrumentalness that we thought would also be impacted by this new technology. While the trend changes in these features were not as strong as the trend change in acousticness, they were definitely present.
Another interesting and drastic trend we noticed was how rapidly and steeply the proportion of songs containing explicitness increased over the last three to four decades. The numbers displayed are actually likely an underestimation since the dataset flags each song with a boolean value of 1 if explicitness is detected and 0 if no explicitness is detected or explicitness is uncertain; because of this uncertainty, underestimation of these proportions is very likely. We hypothesize that this change occurred due to a combination of factors, the most impactful of which could be Western culture's increasing acceptance of more graphic topics. However, from listening to older music, we observe that older music utilizes euphemisms more frequently than outright explicitness; we postulate then that cultural liberalism plays a significant role in the stark increase in explicitness that we see in this data in addition to strong shifts in topics in music.
This tab summarizes the prevalent musical characteristics observed in the songs of the artist (user choice).
This tab allows users to compare the musical characteristics of multiple artists (upto 8 at a time) simultaneously and understand their relationship to popularity.
This tab allows users to visualize the top k and bottom k artists in the dataset according to user input (max. 25). It also attempts to visually capture what attributes of music characteristics are similar among the popular and unpopular artists.
The visualizations we produced on this page help with exploring the relationship of different numerical characteristics of different songs.
This is the correlation matrix shown in the overview tab. This graph, although static, contains condensed information that clearly shows the relationship between different features in the entire dataset. This visualization can answer the preliminary question of what music factors are closely correlated.
This is an example of the visualization created in the song ranking tab for top songs ranking. The date range for this graph is from 2000-1-11 to 2020-6-16, with a k of 15 and a maximum popularity of 94.
This is an example of the visualization created in the song ranking tab for the bottom songs ranking. The date range for this graph is from 2000-1-11 to 2020-6-16, with a k of 15 and a minimum popularity of 10.
The song rankings graphs allow the users to have a clear sense of what the popular music is. The graph also dynamically displays the relevant information including popularity and year for each song, if the user hovers over the bar for the song. Having the top-ranking and bottom-ranking songs side by side allows the user to investigate the difference between popular and unpopular songs. It allows the user to discover the trends themselves.
This tab allows the user to dive deep into the differences in characteristics of individual songs. It allows the user to select songs and compare their detailed characteristics. The application will also dynamically display the popularity of the hovered song. This allows the user to investigate how songs with different popularity differ exactly in terms of their characteristics, which allows the user to explore how features affect a song's popularity.
This is an example of the visualization created in the song comparison tab. The two songs chosen are "Stuck with U (with Justin Bieber) by ['Ariana Grande', 'Justin Bieber']" and "I Know You Care by ['Ellie Goulding']". The reason for choosing the two songs because they have distinct characteristics and show up very differently in song ranking.
The visualizations your system produces and any data to help evaluate your approach. For example, you might describe a case study illustrating how your visualization(s) or algorithm(s) address your chosen problem.
This graph answers the question, "what is the mainstream genre during a particular period of time?" To find the answer, use the slide bar to fix a period of time and observe the majority of genres in the pie chart. For instance, we could see that "Pop/Rock" has been an overall mainstream genre of music.
This bar chart gives us a more quantitative overview of top n genre distribution. With it, we could answer questions such as "Is there an overwhelming genre? If so, how often is it larger than the second popular genre?" During 1930-2010, for instance, we could observe that "Pop/Rock" is such a majority class that it has more than 5 times more artists than the second popular genre, "R&B".
This stacked distribution chart helps us answer questions such as "What is the development trend of a genre? And when did it reach a peak?" Take "Pop/Rock," for instance; it accelerated the increase from 1950 to 1960, reached a plateau from 1960 to 1970, and surged to a peak around 1990. It then displays a decreasing trend yet still maintains the dominant genre.
This network graph helps us answer the question, "How does one genre influence the other?" For instance, when we focus on the top 7 genres, the mainstream genre "Pop/Rock" influences every other genre. While a less popular genre, "Electronic", influences every other genre except "Vocal", "Jazz", and "Country".
This network graph allows us to answer the question, "Who has been the most dynamic influencer?" According to the network, The Beatles has been the central influence on many top artists such as David Bowie, Jimi Hendrix, etc. On the other hand, Hank Williams is a relatively "isolated" artist focusing on Country music, only influencing Bob Dylan among the top 10 artists.
This project addresses the challenge of understanding music's evolution and artists' influencer-follower dynamics over time. The organization of our Streamlit application into four main sections (Data by Year, Data by Artist, Data by Song, and Artist Influence) reflects a comprehensive approach to exploring various aspects of the music dataset. The "Data by Year" section provides historical context and explanations for significant trends observed in the dataset. This approach visualizes the data and educates the user about the potential influences on music characteristics over different decades. The inclusion of interactive visualizations and the ability for users to select features and time ranges adds depth to the exploration. The "Data by Artist" page captures the information about the different artists in the dataset, summarizing the song characteristics of a single artist, comparing the musical characteristics of several artists, and visualizing the top and bottom k artists along with the impact of their prevalent musical characteristics on popularity scores. The "Data by Song" section showcases a well-structured and user-friendly interface with three tabs for an overview, song ranking, and song comparison. Including statistics, correlation matrices, and dynamic visualizations using Altair and Plotly enhances the user experience. The "Artist Influence" section leverages influencer-follower relationships and incorporates interactive network visualizations. We intended to capture the depth of artist-music influence dynamics by including genre and artist networks.
In terms of what the audience has learned from our work, the application provides insights into trends in music characteristics over time, the popularity of artists and songs, and the dynamics of influencer-follower relationships in the music industry. Users can gain a deeper understanding of how different factors contribute to the evolution of music and explore the interconnectedness of artists and genres.
Incorporating various Python libraries such as pandas, matplotlib, Altair, Plotly, and "streamlit-agraph" demonstrates a diverse technical approach. The documentation of data cleaning, design choices, and development steps adds transparency to our methodology. To gather more formal feedback and insights about our system, we would consider conducting user surveys or interviews to understand how users interact with your application, what they find most valuable, and future improvement areas. Additionally, monitoring user engagement metrics within the application upon public deployment, such as the most accessed sections or features, can provide valuable insights into user preferences.
Overall, our project was intended to be a comprehensive and insightful exploration of the evolution and influence of music, offering users a rich and interactive experience to uncover patterns, correlations, and influential factors between artists in the world of music.
In the future, we will improve our application in the following directions. For the data by-year dataset, social & historical background would be incorporated with current time series analysis to show the relationship between social events and major music revolutions. For the data by artist dataset, we discovered that the same artist could correspond to multiple main genres and start years. Could this indicate an artist's style shift throughout his/her career? We could look closer to visualize and analyze who has been the most dynamic artist with the most versatile styles. We could cluster the songs by their principal attributes and inspect the clusters for the data-by-song dataset to find high-level patterns. What makes a cluster distinct? Are the songs in one cluster sharing the time period, artist, or style? Would this cluster overlap with the artist genre classes? For the music influence data, we could extend by adding artist images in the network nodes. To do this, we could either search for a comprehensive database with artist-image url data or scrape the web for artist images. In addition, we could find some ways to incorporate genre into the artist network. For instance, using differen colors for the nodes to denote the genres. This would merge genre and artist information to give a more holistic view.
Below is a list of references that includes our data source as well as websites we referenced for domain information.
Get tracks’ audio features. Web API Reference | Spotify for Developers. (n.d.). https://developer.spotify.com/documentation/web-api/reference/get-several-audio-features
2021 ICM Problem D - Mathmodels.org. Mathmodels.org. (n.d.). https://mathmodels.org/Problems/2021/ICM-D/2021_ICM_Problem_D.pdf
Wikimedia Foundation. (2023, November 19). Synthesizer. Wikipedia. https://en.wikipedia.org/wiki/Synthesizer
ChrisDelClea. (n.d.). GitHub - ChrisDelClea/streamlit-agraph: A Streamlit Graph Vis. GitHub. https://github.com/ChrisDelClea/streamlit-agraph
pandas.pivot_table — pandas 2.1.3 documentation. (n.d.). https://pandas.pydata.org/docs/reference/api/pandas.pivot_table.html
pandas.DataFrame.melt — pandas 2.1.3 documentation. (n.d.). https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.melt.html#pandas.DataFrame.melt
altair.Row — Vega-Altair 5.2.0 documentation. (n.d.). https://altair-viz.github.io/user_guide/generated/channels/altair.Row.html
plotly.express.pie — 5.18.0 documentation. (n.d.). https://plotly.com/python-api-reference/generated/plotly.express.pie