Music on the phone

Users use mobile phones and other communication terminals to WAP, WEB, APP and other access to access music-themed content-related services, this service is called wireless music. It includes ring tones, wireless music clubs, and mobile client software businesses. Smartphone era, mobile client music has gradually become an important way for users to enjoy time.

With the advent of the smart phone era, some of the biggest mobile music client companies in the wireless music industry have millions of users or even tens of millions of users.

The continuous development of mobile client music and the continuous growth of user groups have led to the emergence of a large amount of wireless music data. These data seem disorganized, redundant, but hidden a lot of secrets. If these data can be effectively organized and managed, and the relevant technologies can be used to excavate and analyze the results, the results of a company's decision-making after implementation can be revealed. It finds that there are major problems existing in the company and finds potential high-value businesses Or needs, these businesses or needs are likely to provide strategic guidance for the development of the company.

The following to a famous domestic mobile music company client wireless music data, for example, we still follow the discovery, problem solving, verification of these three aspects to illustrate the organization and application of wireless music data.

problem found

Through the analysis of the data mining, we found the following questions.

(1) users, songs have long tail effect

From the data we found that users have two kinds of behavior, one is to download, one is audition,

For each of these activities, we found that there was a "long tail" phenomenon for both users and songs. Most users only tried or downloaded a few songs in the system, while most of the songs were out of status. Specific information as shown below

Long tail music problem

NOTE: In the figure, the horizontal axis of the left subgraph indicates the user's song, and the vertical axis indicates the proportion of the corresponding user. The abscissa of the right sub-graph indicates how many people have heard the song, and the ordinate indicates the proportion of the corresponding song. Causes of this may be due to: large amount of data, serious information overload users can not find their favorite songs.

At this point most users directly to the pop chart or song songs list, it will result in the system more popular hot songs, popular songs more popular phenomenon.

(2) low song coverage

From the data we also found that the coverage of the song is very low, the song in the entire sample data

Coverage is only 2.01%. The vast majority of songs were never heard or downloaded by the user, which not only caused huge waste of system resources, but also caused innocent loss of company funds (because each song had to pay a copyright fee, and 98% of the songs in the system were wasted status). The cumulative coverage of songs as shown below.

Song coverage chart

Explanation: The abscissa in the figure indicates the number of songs listened to by the song (weight loss), and the ordinate is the proportion of the songs no less than this number.

Reasons for this may be: a large number of songs in a cold start, sparse data. As a cold-start composition, the system does not know how to push him to the right user, and the user can not find him in an effective way, making the darkness of such songs processing systems less visible.

(3) Users listen to songs every day was intermittent distribution

In a given sample of data, we found that the user's listening behavior is not evenly distributed, but rather intermittently, ie the user's listening concentration is different at different times. In order to better see the effect, we divide the day into 8 time periods, each time period includes 3 hours. During each time period, the user's song activity is as shown below.

User active time chart

Explanation: In the figure, the horizontal axis represents the time period and the vertical axis represents the user's activity ratio in this time period.

Causes of this may be due to: get off work, rest, sleepy tired time

Users listen to songs on the infinite side of the model or tend to leisure and entertainment, mainly based on the rest of debris time.

(4) Different users have different dependencies on the attributes of songs

In the sample data, songs have album and singer two properties. We analyze the user's listening behavior from the aspects of the user's long-range relevance and the short-range relevance, and analyze the specific results as follows:

Description: In the figure, the Strong null model, the Weak null mode, and the Temporal null model represent the similarity values ​​of all the broadcasts in the system, the similarity values ​​of all the songs, and the similarity values ​​of the adjacent broadcasts respectively. Album said the album, Artist said singer. This may be caused by the fact that users tend to listen to the same singer's song as the album

(5) Different users listen to different songs

Our analysis from the data also shows that different active users listen to different songs. In the analysis, we analyze the songs of different active users from three aspects: novelty of songs, similarities of songs in albums, similarities of songs in singers.

Specific information as shown below

Three dimensional analysis of songs

Explanation: In the figure, the abscissa indicates the activity value of the user, the ordinate indicates the novelty value corresponding to the song listened to by the active user, the similarity value of the song on the album, the similarity value of the song on the singer

The reason for this may be: users may be clustered

Less active users may be ordinary users, these users according to their hobbies to choose the songs you want to hear. Highly active users may be professional users, such users according to their own professional needs to choose the songs you want to hear.

solution

From the discussion in the above section, we already know several issues that may be hidden in big data at the wireless music side:

① users, songs have long tail effect

② low song coverage

③ users listen to songs every day was intermittent distribution

④ different users depend on the properties of the songs different

⑤ different users listen to different songs

When a company in the face of these problems should be what kind of solution to solve or

Improving the current situation is another important issue. Especially the above problems ①, ②, if not handled properly, may affect the entire company is functioning properly, and even affect the company's development.

Therefore, this section from the wireless music data, put forward several suitable solutions.

(1) users, songs have long tail effects, we can use the following technology

With information filtering technology, a way to categorize songs and map different users to different song categories. Another way is to personalize recommended techniques, the system automatically analyzes the user's preferences for different users to filter the appropriate songs.

(2) low song coverage, we can use the following technology

Low song coverage is mainly because the user can not find the music, causing this problem for two main reasons: ① the music itself is not enough information, ② music has information, but the user can not find the music.

So on the one hand we can label the music, use the label information to express the specific attribute of the song; on the other hand, we can use the recommended technology to personalized the song recommendation.

(3) Users listen to songs every day intermittent distribution, we can use the following technology

At different times, we set different theme songs to suit different listening scenes, such as relaxing soothing, smooth songs, morning rock, heavy metal and the like.

Of course, the specific scenario needs to be obtained through further data mining. This article merely presents a method for not elaborating on specific technologies.

(4) Different users have different dependencies on the attributes of songs, we use the following techniques

Through historical data analysis to obtain the user's dependence on the song attributes, from which we can know what kind of attributes the user more dependent on. When we find that users are more dependent on the genre, we can play the songs for them according to the genre. When the user is found to be interested in the singer, I can play the songs for them according to the singer.

(5) Different users listen to different acts, we can use the following technologies

Group users based on user characteristics, which divides users into different groups. Different groups for different songs we play to their different, such as ordinary users can play hot songs, and for professional singers, we are a variety of songs to play for them.

Result verification

In order to further illustrate the effectiveness of the above solution, here we only use the recommended algorithm to explain when the system is used in the system, some significant changes in the system, the specific changes are as follows:

l users find it easier to find their favorite songs

Users find it easier to find the song

Currently, the music website uses song list (GRM) to organize songs. In this way, the probability that a user finds his favorite song is about one in a thousand. When we use three kinds of recommendation methods (OCF, HC, MD respectively) After that, he found that the user's probability of finding his favorite song increased significantly, and the accuracy of the MD algorithm was improved by 10 times.

System long tail changes

Before using the recommended algorithm

After using the recommended algorithm

Long tail effect to improve

Obviously from the figure above, we can see that there is a significant change in the long tail effect of the system. Such a result should be the result that the company most wants to see, not only greatly reduces the unnecessary waste of the company, but also provides the user with a better user experience.