No matter who you are, you can create your own music channel and have your voice to be heard by the world via YouTube. The spirit of opening and sharing has made YouTube the largest online music resource. However, this openness is a double-edged sword. It makes YouTube music data not as clean and well organized as major music platforms like iTune, Pandora and Spotify. If you search for "Blank Space" of Taylor Swift, many different videos of this song will show up. This duplication may affect user experience, and jeopardize the quality of recommendation engine.
MusicGalaxy is designed as a clean and neat music platform based on the YouTube music data universe. At the heart of MusicGalaxy is a machine, trained to understand the semantic meaning of EVERY YouTube music video titles. This machine learning process tranforms noisy un-structured YouTube text data into a structured music database. You can jump to the Tech View for more details on the algorithms.
Many music lovers always keep an eye on the Billboard Hot 100 music lists to follow the most recent music trend. MusicGalaxy provides a YouTube version Billboard Top 100. At this time, there are two rank lists available: one for song track, and the other for artist across all music genre. Each rank list is ordered by the weekly music trend: the weekly cumulative view growth of all YouTube music videos featuring a song or an artist. Both rank lists are made from a sample of over 2 million YouTube music videos, and updated every week.
In MusicGalaxy, you can start by choosing music from the top 100 rank list or search for any music video you like directly. For example, you can click on Sia - Chandelier on the top 100 rank list. While this music video is loading, you may see the artist name "Sia" and song name "Chandelier" has been recognized algorithmically and listed by the video window. These names are there for recommendation purpose.
Two types of recommendations are provided. The first type follows the "artist to song" rule. You can click on artist name Sia for other songs of her. The second type follows the "song to artist" rule. You can click on song name "Chandelier" for different versions of this song by other artists. The second type of recommendation not only gives music lovers a unique music experience, but also provides great opportunities for newbie singers. Here is why.
We know that newbie singers usually need to start by singing cover versions of songs of famous singers. From MusicGalaxy's recommendation, you may notice that Sia's "Chandelier" has been covered by many other singers, such as Jean Kelley and Jasmine Thompson. However, artists like Jean Kelley and Jasmine Thompson are much less well-known than Sia, and the YouTube recommendation engine rarelly recommends their cover songs when you are watching Sia's "Chandelier". It usually recommends other songs by Sia. MusicGalaxy's recommendation provides greater chance for the voice of newbies to be heard, and their talents to be discovered.
So what is the 'magic' supporting at the backend for music service like top 100 rank and recommendation? A simple answer is a Named-entity Recognition algorithm.
As we know a YouTube video can be uploaded by anyone, with any language background. The uploader of a music video will describe the materials of the music in video title in their own way. To build a structured music database, we need to know the artist names and song names of every YouTube music video. This is the starting point of building a good music platform. Due to the huge amount of YouTube music videos, we need to train a machine to recognize artist names and song names from multilingual video title text automatically.
MusicGalaxy takes the state of art named-entity detection technique to perform this task. A machine is trained by hidden markov model to assign tags to each word in YouTube music video titles. Those parts with tags "Artist" and "Song" are recognized as artist name and song name respectively.
The most chanlleging part of the MusicGalaxy project is getting training data for the hidden markov model. Thanks to Musicbrainz's open source music data, the emission probability for words in "Artist" and "Song" tags can be estimated. For the transition probability, 500 human tagged YouTube music video titles are provided for estimation. We can treat these estimation as initial parameters in the hidden markov model, and use it to process the total of 2 million videos. The results can be used as the new training data. Due to the big training data size, the parameters can be re-estimated, and performce may improve. This re-estimation process is repeated till the estimation becomes stable.
Another chanllege is the multilingual environment of YouTube video titles. It is very common to see multiple language mixed together in the same video title. A special tokenizer is designed to process multilingual text in a reasonable way.