Current location - Music Encyclopedia - QQ Music - A brief discussion on the music recommendation algorithm: Listening to music is to find BGM for my current mood.
A brief discussion on the music recommendation algorithm: Listening to music is to find BGM for my current mood.

There are two scenarios for listening to music. For example, I just watched "Wreck-It Ralph 2". The theme song of the movie is very nice. I plan to listen to it on my way to and from get off work, so I open the music app. , enter the keyword "Wreck-It Ralph 2" to search, find "Zero" by Imagine Dragons, and listen to it repeatedly until you get tired of it. The other is to lie on the sofa on a rainy day and find some music to match my current mood. As for which one to listen to, I don’t know, but the need to listen to music does exist and is strong.

The two scenarios are completely different. The former has clear needs and users actively search for resources; the latter has hidden needs and needs to be mined. In most cases, they are in the second state, that is, they have no clear purpose and just want to listen to music. I don’t care which singer the music in my ears comes from, country or folk, Chinese or English, or even listen carefully, but it is indispensable, like playing a suitable BGM for my life.

Tencent recently released a music app --- MOO. Before coming into contact with MOO, I basically used QQ Music and NetEase Cloud Music. I carefully counted my own usage habits and found that many times, Basically, only the songs in the favorite list will be played in sequence, and these songs come from the scenario described above - "Inspire--Search--Listen--Add Like". QQ Music and NetEase Cloud also put a lot of effort into creating "discovery" sections, including lists, authors, topics, etc., but I rarely use them because I have no desire to explore. You need to listen to music to find out whether you like it. Clicking on lists, authors, and topics to listen to music you like is too costly and not very effective. Therefore, apps such as QQ Music and NetEase Cloud Music are music search and playback tools for me. The first time I used it, my eyes lit up. The highlights are as follows:

1. The immersive full-screen mode entered after launching the App, the playback animation is cool and visually impactful;

2 . Almost most icon annotations have been removed from the App, making the interface very simple;

3. Swipe up to switch songs, swipe right to change tabs, and slide horizontally at the bottom to adjust the playback progress;

4 . 20 songs will be recommended every day. No need to search, just open it and listen.

Overall, it is a very young product for the post-95s and post-00s generation. But what attracts me more is not the above, but that Moo has changed my music listening habits, or it has redefined my music listening scene. Recently, I listen to Moo with headphones on my way to get off work almost every day. The 20 songs are enough to satisfy the novelty of the day, and I don’t have to think about the question of "what to listen to." It’s that simple to collect what you like and swipe up what you don’t like. The operation cost is extremely low, making listening to music a pure process of exploration and appreciation.

The key to determining the quality of this user experience lies in: the accuracy of music recommendations.

The ideal state in my eyes is: it can get to know me better and better, know what kind of music I like to listen to, and even guess what music I want to listen to at a certain moment, and then play it!

In fact, Spotify (a genuine streaming music service platform, launched in October 2008) has already been doing this, mainly using the following three recommendation strategies:

1 . Collaborative filtering, looking for users with similar hobbies and interests, and then recommending songs on the playlist to one party;

2. NLP, crawling information about music on the Internet, and analyzing the specific discussions of users Artist or song content, such as what adjectives are used or which words are used most, through statistical analysis to obtain "cultural vectors" and "top terms". Through cultural vectors and high-frequency phrases, we can There is a high probability of finding music with a similar style.

3. The original audio model, convolutional neural network (CNN), analyzes audio information (rhythm, pitch, timbre) through computers, and then recommends songs with similar information data to the user. Mainly used to deal with the cold start problem of new songs.

Regarding music recommendation, I personally think that there are still several problems that have not been well solved:

1. Similarity in the dimensions of rhythm, tone, and timbre may not necessarily represent Similarity in the dimensions of emotion and style;

We can use computer technology to analyze the rhythm, tone, timbre, and even user comments, singer information, etc., but this information can express the emotion and emotion of the music. Style? The consumption of information news is objective and emotionless, but music is not. On the contrary, the consumption of music is largely determined by the emotions conveyed by the listener and the music itself. Therefore, emotional matching is critical.

2. It can obtain the user portrait of the listener, but it may not be able to obtain the listener's current emotional portrait;

Music is an expression of emotion, and emotion is different from interest, which lasts for a long time. is stable over a period of time, even if the user's potential interests are discovered, the basic interest profile is stable. The emotions are different. One moment I am laughing, and the next moment I may be worried about something. The mood will change and it will be fleeting. Therefore, for recommendation algorithms, how to capture, parse, store, and process users' emotional portraits in a timely manner is a big challenge.

3. What is the repeatability of music?

We all know that for news information, users basically only need to read it once, and few people will read it again. But music is different. Music has the attribute of repeated consumption. The difficulty is: how do you know when a user is tired of hearing the song? When did you suddenly want to listen to that old song that you thought you were tired of hearing?

--------------------------

The above opinions are purely personal, welcome God corrects and communicates.