Where can I find the music of the heart?Music recommendation full module logic shallow analysis

Edit Introduction: There are thousands of songs, how do you guess your preferences in your song app? This requires disassembling the music recommendation logic of the music apps. In this article, the author combined with the existing music app to split and summarize music recommendation logic. Let's take a look.

Since the "Music Genome Project" launched by Internet Music Radio Pandora in 2000, it has been recommended by the recommended songs that users like by the algorithm for more than 20 years. However Potential interests and fully satisfy the most important propositions of algorithm push songs.

This article will try to disassemble the basic logic of existing music recommendation and analyze the existing functional strategy design through major APP instances.

1. Music recommendation logic disassembly

From a user perspective, users' demands for personalized songs are mainly to find a favorite and fresh songs that can be found hard, so for personalized push songs, this appeal can be simplified and disassembled into three parts to meet ,which is:

How to discover users' preferences; how to identify and understand music content; how to match interest and meet user needs. Next we analyze one by one:

1. How to discover the user's preference interest

There are two major categories of user interest paths that mainly actively express and passive expression. Active expression of interest preferences are more of the form of interest collection relying on functions, while passive expression is more based on analyzing the calculation of consumer behavior in the user station. Specifically Look at it:

1) Users actively express interest preferences

The main practice in the industry is to provide interest collection functions when logging in new users/silent users, and guide their selection preference category (as shown below) through major classification options, but this path rely on the user's active selection and is a real -time expression that users currently prefer. There are problems such as low penetration rate and insufficient confidence in choosing preferences. It is more about the use of the content display of the cold opening stage, and it is difficult to completely be used as the main basis for recommendation.

2) Swerry preferences based on user consumption behavior

Different from users' active expression of interest preferences, according to user consumption behavior, speculation preferences are based on the cause, and their degree of confidence is relatively high, which is the main basis for discovering user interest preferences. In terms of subdivision, it can be comprehensively considered by introducing external consumer signals and consumer behavior in the site.

① Introduce external consumer signals

The introduction of external signals can play a supplementary role in user portraits. It is determined that the user characteristics are determined to speculate on the characteristics and interest preferences of similar user characteristics.

For example, a long time ago, you can get the Applist through some channels. By analyzing the application type characteristics, you can basically determine that the user's social identity and similar preference can be basically determined, and then the user model speculates interest preferences.

Another common method is partial interconnection of the company's own APP. By reading the related consumer signals, it is easier to judge its potential interests. For example, YouTube provides user music -related content on the information of YouTube Music, so as to use it to step on the information. A large amount of recommendations are provided. Of course, the problem of contamination of coupling data between the two has always been the place where Volkswagen expects optimization.

② The consideration of consumer behavior in the site

When users start to accumulate consumer behavior in the station, recommending and optimizing through consumption behavior in the station is a more important way.

In addition to deep consumption behaviors such as praise, transfers and collection of various scenarios can clearly speculate on user preferences, other specific scenarios to collect user lightweight expression preferences through functional design are reflecting the product design ingenuity. The second module in this article will analyze through instance analysis. Expand analysis in detail.

2. How to identify and understand music content

Just as people pay attention to music genres, singers, and song tunes when they are looking for songs to make a judgment of whether they like it. The first step to match the user's interest is to allow the algorithm to fully extract the characteristics of music content from multi -dimensional. Content understanding.

Simplified content understanding: whether it can be explained into explanation into explicit label recognition and hidden content vector recognition.

1) Refining of explicit labels

Expressive label, as the name suggests, is the label content that can also be identified artificially. In addition to the labels of fixed attributes such as music genres and singers, it also contains non -structured labels co -created by users such as song keywords, such as music leaders boss Spotify once used Internet crawlers and NLP technology to refine the "keywords" such as describing expression, sentences, nouns such as sentences, and nouns when they were mentioned in user comments, and set different weights for these keywords to quantify in people's eyes Which songs are similar.

Of course, this non -structural label acquisition method has certain limitations. On the one hand, the songs covered by the explicit labels are limited and the granularity is thick; on the other hand, for the new songs or unpopular songs, the user's comment popularity and quantity With less, it is difficult for NLP to have a good effect. Based on this hidden content vector recognition, it came into being.

2) Illustration of implicit content vector recognition

From the subjective perspective of users, the music of various morphological characteristics actually affects user preferences. For example, the lyrical version of a song and the Remix version with heavy metals are extremely different. Calculating sound data information to obtain hidden content vectors is an extremely important part of music recommendation.

For example, the strategy adopted by the music software JOOX transforms the audio signal of music into a spectrum, and then uses the CNN (convolutional neural network) to classify the picture classification (see examples below) to represent the sound information of the music, which is used to recommend information learning. Essence

Note: Content understanding includes not only the above feature extraction, but also the construction of the basic information of content quality, content timeliness, etc., and it is not too much to say about it. Therefore, this article is not talking.

3. How to match interest to meet user needs

On the basis of the user's understanding and content understanding, the recommendation strategy can begin to play a role. Although the recommended algorithm is changing with a variety of days, it is not separated from its ancestors. Estimate the degree of interest, comprehensively consider users and business demands and then recall sort to show users.

There are many disassembly articles and books related to recommendation strategies. Interested students can read books such as "Recommended Algorithms" and other books.

As a result, we can simply summarize the role of the modules and corresponding strategic logic of music recommendation:

Example analysis

Then, by experiencing test test reverse disassembly, specifically analyzed how Netease Cloud Music, Soda Music, and QQ Music Music Recommendation Stream meets the recommendations of users to find songs and listen to songs through functional and strategic design.

Note: In order to measure the user experience, the author evaluates the effects of the recommendation of each app music from his own experience, and discusses the advantages and disadvantages from the perspective of accuracy, timelyness, and diversity, but the degree of activity of each app may affect the degree of activity differently. To a certain assessment letter, only reference here.

From the above analysis, we can see that under the circumstances of each basic function, a certain differentiation function is made based on their own functional positioning. A more obvious example, each of them has the idea of ​​playing List design ideas around their own positioning Different.

We will find that NetEase Cloud has canceled the playback List and only supports a song back. Combined with the private FM function positioning, we can speculate that such restrictions are mainly to educate users to cherish the current choice opportunities. "The system understands yourself faster and better. Of course, such a design can reduce the opportunity to listen to VIP songs by non -VIP users to a certain extent, or it may be a comprehensive consideration in combination with business monetization.

Based on the logic that the two users can find their favorite songs and can consume on the spot, the two users provide unique features such as playing List and providing unique functions such as deleting songs and songs. It can also bring more users with lightweight expression of recommendation results, and can also help the recommendation system recommendation more accurate.

Three, conclusion

Although we will see that the algorithm has a mature method logic through the above disassembly, it is more convenient for users to meet the demands of music search and music consumption, but it is also criticized by the public to have many problems such as cocoon houses. How to explore the potential needs of users, Growth demand is also an important issue facing various content recommendations now.

In response to how to solve this problem, the next article is analyzed with you ~

Author: Dundun Chongchong, WeChat public account: wow small fish

This article is published by @本 本 本 本 本 is the product manager. Reprinting is prohibited without permission.

The title map is from Unsplash, based on the CC0 protocol

- END -