Recommendation System Application Research: Music Radio

The recommendation system rarely has a separate product form, many of which are combined with other products, and play an auxiliary use effect. such as personal use more recommended system has input method of Word association, book site books recommended, music radio guess users like Music and Google reader recommended entries.

For watercress music channel use less, although there is no research on music, and will not buy an album or listen to concerts, but very much like listening to music, work while listening to music habits. Watercress Radio has been listening to 13,670, like 136 and do not like 666, is the heavy users of the radio bar, the individual operating behavior are:

as background music for working hours, in the background running state. Sometimes the work is entranced, as long as it is not particularly annoying or particularly like songs will not go to visit the Radio Web page. When leaving the station, the headphones do not turn off the music, and at least 30% of the more than 10,000 songs don't really listen. Hear the ads and the "Dog Gang" and other noisy concerts to find the station's Web page, choose not to like, can not stop the ads directly click off the radio, because it is easy to cause distractions affect work. Hear like music 60% probability will find the radio page click "Like". When you find a particular favorite or wish to recommend a song, click "Like." Listening to less-than-favorite music will also choose to click on the "Next song."

Watercress Radio recommended new songs more powerful than shrimp nets, operation is more convenient, do not need to be like the initialization of the shrimp network to choose like singers and choose to like music after the selection of labels, more intelligent, reduce the operation steps. Shrimp's song library seems not enough, many are recommended popular songs, but the recommended form more rich and tricky than watercress.

Watercress does not support by singers and albums to listen to songs, if three consecutive clicks "Do not like" did not hear like music or often play hate yuannv love songs, may choose to listen to shrimp Nets singer album. If you choose to listen to more than 100 songs that you have chosen to love, you may not switch to another product.

Watercress Radio usually as a browser bookmarks exist alone, usually direct access to watercress and not from the Watercress home page to visit the watercress, from the radio back to the Watercress homepage can only manually modify the URL or click the album link to see the album Back to the home page. The love of the song removed, does not mean that the user no longer like the song and to jump to the next song.

Watercress before starting to do, already have a relatively perfect initialization of data, most albums have labels and ratings, such as the album "We Are May days" of the relevant metadata:

The average score of 9592 people was 9.3. Tags are: May days (2450), Taiwan (701), MAYDAY (644), we are May days (374), Rock (338), we are:, May days (264), Rocky (201), Pop (160). The number of each song collection.

From the above tags can be seen, singers, regions and music style is the user's common label, you can assume that these three tags are affecting the user's choice of the most important metadata. English mayday, May days, we are May days and we are: the May days are equivalent tags, rock and roll are equivalent tags, there is no such as the Google pick song of the mutual exclusion tag (such as the rhythm of the soothing and strong). From this example, we can see that the user language is ambiguous and it is necessary to use the controlled thesaurus to control the different language expressions of the synonyms.

Watercress's user base is relatively small, the added label quality is relatively high, the use of tags can simply organize the link between songs, the use of ratings, the number of collectors and the number of labels can determine the recommended priority.

If the user chooses on the radio to like the album "We Are May days" in the song "Hug", that may be more reliable according to the singer's recommendation, users choose more than May days of the song, that can be identified as the May days of fans, song library added the May days of new songs, since when the first recommended.

Compared with the style, the impact of the region is lower, but does not exclude some users have this preference, such as only listening to Mandarin songs, which need to go through the decision tree and so on several methods recommended to prove. Select the style "Rock", you can recommend the label for Rock and score a higher singer, gradually confirm the user's music preferences model. Before you strongly recommend a song, you can first complete the collection of users like songs. In the uncertain circumstances, you can play more music you already like or like the singer's popular songs.

The user likes the singer to have many, but likes the style not to be many, the same style has many singers, the user likes the style to be more complex and fuzzy than the singer. The singer is a side-by-side label, and the style will form a distinct bias. As time changes, so does the style.

In addition, the label should be more consistent with the song, heard the radio recommended murmur more than May days concert songs, will choose not to like or skip to the next one. The system may mistakenly assume that the user does not like the song, and that low quality songs can affect the accuracy of the system's recommendations.

Source: http://daichuanqing.com/index.php/archives/1880