CN104008138A - Music recommendation method based on social network - Google Patents

Abstract

The invention provides an individualized music recommendation scheme. In order to solve the problem that traditional music recommendation is blind and mechanical, a data mining technology is utilized, users are segmented in a stratified mode, and various different requirements are analyzed according to historical records of the users. The users are clustered by the combination of the social network and according to music favored by the users and hated by the users, an interest tendency label is generated for each user, the users are recommended to add adjacent users to be friends, and the friends are synchronously changed according to the change of user interests. Thus, friend circles serve as a data set, recommendation is conducted through an association rule, and recommendation is more accurate and more efficient. A cloud computing technology is utilized, mass music information and user data are acquired, and good expansibility is achieved.

Description

A kind of music recommend method based on social networks

Technical field

The present invention is a kind of individualized music suggested design.Blindness and the mechanical problem of for traditional music, recommending, utilize data mining technology, stratified by subscriber segmentation, and combine with social networks, makes to recommend more accurately efficient.Use cloud computing technology, to solve music information and the user data of magnanimity simultaneously.Belong to data mining and cloud computing field.

Background technology

Along with developing rapidly of Internet technology, arise at the historic moment in music recommend miscellaneous website, and network digital music also becomes important component part indispensable in people's studying and living gradually.Brand-new music carrier, when providing business opportunity for music site, has also proposed new challenge.How in network world, to attract new user, keep old user here, become the main task of music site.On the other hand, user, at the digital music in the face of magnanimity, therefrom find own requirement also just like looking for a needle in a haystack.Therefore, we need a comprehensive understanding user's of energy music tendentiousness, reflect user interest comprehensively, and can excavate the music recommend system of user's potential demand.

Music series products has passed through three development epoch, and first epoch are program request epoch, and user comes program request oneself to want the song of listening by song title and singer, as dried shrimp; Second epoch is to carry out data mining by single dimension, introduces algorithm, goes the song that judges that user likes, as bean cotyledon; The 3rd epoch are to recommend and data mining by social networks, by introducing social collection of illustrative plates and interest graph, excavate the interested music of user.

Summary of the invention

Technical matters: music recommend and text are recommended and other information recommendations make a big difference.One, music recommend has suitable subjectivity, can be subject to the impact of the many factors such as surrounding environment.Two, there is available time, can be along with broadcasting time and popular time change.Three, there is ambiguity, music recommend does not often need to be accurate to certain first song, and more should focus on the excavation of potential interest than the recommendation of similar song.Therefore, only use basic data digging method, the rigid classification of the labels such as utilization " special edition ", " singer ", " style " recommends or according to historical record association mining, going out suggested design is not well positioned to meet actual needs.

Technical scheme: blindness and the mechanical problem of recommending for traditional music, the present invention relates to a kind of new individualized music suggested design.Utilize data mining technology, stratified by subscriber segmentation, according to user's historical record, analyze various demand.Combine with social networks, the music of liking according to user and disagreeable music are carried out cluster to user simultaneously, and each user is generated to interest tendency label, recommend user that neighboring user is added as a friend.Then take friend circle as data set, adopt correlation rule to recommend, make to recommend more accurately efficient.

The operational scheme of a music recommend system of taking this programme is as follows:

1. the music recommend method based on social networks, is characterized in that this recommend method step is as follows:

Step 1. prompting user logins, and inquires about this user's historical record in background data base; If this user logins for the first time, perform step 2～step 4, otherwise execution step 5～step 7, the step 8 of finally seeking unity of action;

Step 2. is sorted according to the music number of times in current Qu Ku by music recommend system, selects the most popular song and recommends as initialization playlist;

This commending system of step 3. is when waiting for that user listens to, and prompting user likes, dislikes and skip the selection of option and these evaluation results are carried out to classification analysis every first song, calculates wherein, i represents certain user, and n represents to like the sum of song, and m represents the sum of disagreeable song, r _i,krepresent whether user i likes k song, t _i,krepresent the whether disagreeable k song of user i;

Step 4. simultaneously, is carried out cluster analysis by the selection result of collecting in step 3, calculates any two user i, the similarity sim between j ( _i,j), find adjacent friend circle, the interest tendency label of prompting neighboring user, the information of favorite song, special edition, recommendation user add is good friend; Then perform step 8;

The user that step 5. once logged in has storage of history data P in database, and commending system reads this user's buddy list and friend circle and plays record from background data base;

The history that step 6. reads this user by commending system from background data base is again play record, if the frequent degree of listening in the recent period surpasses every day three times, surpasses half an hour at every turn, perform step 7, otherwise skips steps 7 directly performs step 8;

Step 7. commending system is the music of up-to-date interpolation in this user's commending friends circle;

Step 8. commending system carries out correlation rule recommendation according to playing record in this user's friend circle, for this user recommends music, and continues to record its evaluation result;

Step 9., at music recommend system run duration, can be monitored interest migration factor ε=p of this user always ²+ d * l, wherein p representation unit time skip song and the always ratio of recommendation song, d represents to delete the ratio that disagreeable song accounts for all recommendation songs, l represents that the deleted song of once liking accounts for all ratios of liking song, if meet acquiescence threshold values 0.6, perform step 10～step 12, otherwise directly perform step 13;

Step 10. couple user re-starts classification analysis, upgrades user's interest tendency label;

Step 11. couple user re-starts cluster analysis, upgrades friend recommendation list, adds new good friend, deletes old friendship friend;

Step 12. is carried out correlation rule recommendation according to playing record in the friend circle after upgrading;

Step 13. continues as that user provides music service until user exits this service.

Beneficial effect: the present invention proposes a kind of new individualized music suggested design, the main advantage of this scheme is:

One, segment user's request, avoided unexpected winner song cannot enter the defect of recommendation list;

Two, combine with social networks, dwindled the scope of data mining, blindness and the mechanicalness of having avoided traditional music to recommend, make to recommend more accurately efficient;

Three, proposed the concept of the interest migration factor, increased the ability of machine learning, the user's needs of more fitting, follow the trail of changes in demand, and can revise data mining error, avoid repeating recommending;

Four, use cloud computing technology, to process music information and the user data of magnanimity, there is good expansion.

Accompanying drawing explanation

Fig. 1 music recommend process flow diagram.

Fig. 2 customer relationship logical diagram.

Fig. 3 user interest transition graph.

Embodiment

The key that individualized music is recommended is, each user is body one by one, has only got each individual feature clear, just can put the contact between individuality in order, recommends efficiently.Therefore, user account is using the unique identification as discriminate individuals.

If log in for the first time, select 20 the most popular songs as initial list, in order to collect user's preliminary information, if be the very short visitor of listening period, the hottest song is also enough to satisfy the demands.

Interest tendency label: can select when user listens to like, disagreeable, the option such as skip, these selection results are classified, the music of liking and disagreeable music are classified respectively, according to the attribute of music itself, such as bent wind, special edition, singer etc., for user generates interest tendency label, comprise and like singer, music style, played songs, likes song etc. most at most.For example: (user A, rock and roll, May, the stubborn > > of favorite song < <)

With user, for body one by one, according to these results, carry out cluster analysis, the similarity between user is

${\mathrm{sim}}_{(i,j)}=\frac{{\mathrm{\&Sigma;}}_{k=1}^n{r}_{i,k}\&times;r_{j,k}}{\sqrt{{\mathrm{\&Sigma;}}_{k=1}^n{r}_{i,k}^2\&times;{\mathrm{\&Sigma;}}_{k=1}^n{r}_{j,k}^2}}\&times;\frac{{\mathrm{\&Sigma;}}_{k=1}^m{t}_{i,k}\&times;t_{j,k}}{\sqrt{{\mathrm{\&Sigma;}}_{k=1}^m{t}_{i,k}^2\&times;{\mathrm{\&Sigma;}}_{k=1}^n{t}_{j,k}^2}}$

Wherein, i, j represents two users, and n represents to like the sum of song, and m represents the sum of disagreeable song, r _i,k, r _j,krepresent user i, whether j likes k song, t _i,k, t _j,krepresent user i, the whether disagreeable k song of j.

According to the similarity of user and each friend circle central point, for user sorts out, find adjacent friend circle, the interest tendency label of prompting neighboring user, the information such as favorite song, special edition, recommendation user add is good friend, finally by user oneself, judged whether to add, avoid systematic error.

Different users is different for the demand of recommending, and is apparent that most, repeats to listen a small amount of concert to produce sense of fatigue, and the potential interest that good commending system should be able to digging user.Therefore the hot music that, only just constantly repeats to recommend user to like type is inadequate.If this user listens to very frequent, just should recommend compared with the song of unexpected winner, to excavate its potential interest in the song of up-to-date issue and circle of friends.

According to playing record in friend circle, carry out correlation rule recommendation, according to self historical record search rule, obtain recommendation results.Data set is dwindled within friend circle, reduced calculated amount, facilitate decimation rule, improved again the accuracy of recommending.

The interest migration factor: user's hobby is not unalterable, can change along with factors such as time and environment, and by the resulting result of data mining, also might not start just very accurate.So there will be the disagreeable music of continuous recommendation user, the music of frequently skipping, and user deletes the situations such as music of once liking.Therefore, add the concept of interest migration factor ε, learnt user's hobby, the error that the variation of reaction user interest and equilibrium criterion are excavated.

ε＝p ²+d×l

Wherein, p representation unit time skip song and total ratio of recommending song, d represents to delete the ratio that disagreeable song accounts for all recommendation songs, and l represents that the deleted song of once liking accounts for all ratios of liking song.

When ε meets certain threshold values, just need to revise proposed standard, user is re-started to classification analysis, upgrade user's interest tendency label, user is re-started to cluster analysis, upgrade friend recommendation list, add new good friend, delete old friendship friend.Determining of concrete threshold values need to be considered song storehouse sum, total number of users, and many practical factors such as average line duration, cannot simply determine a fixed value, should adjust at any time according to actual conditions, and acquiescence is made as 0.6.

Every first song can be commented on and be visible to the user in friend circle, lists file names with nearest part comment during recommendation, increases the surcharge of music, improves stickiness.

Claims (1)

1. the music recommend method based on social networks, is characterized in that this recommend method step is as follows:

Step 1. prompting user logins, and inquires about this user's historical record in background data base; If this user logins for the first time, perform step 2～step 4, otherwise execution step 5～step 7, the step 8 of finally seeking unity of action;

Step 2. is sorted according to the music number of times in current Qu Ku by music recommend system, selects the most popular song and recommends as initialization playlist;

This commending system of step 3. is when waiting for that user listens to, and prompting user likes, dislikes and skip the selection of option and these evaluation results are carried out to classification analysis every first song, calculates wherein, i represents certain user, and n represents to like the sum of song, and m represents the sum of disagreeable song, r _i,krepresent whether user i likes k song, t _i,krepresent the whether disagreeable k song of user i;

Step 4. simultaneously, is carried out cluster analysis by the selection result of collecting in step 3, calculates any two user i, the similarity sim between j ( _i,j), find adjacent friend circle, the interest tendency label of prompting neighboring user, the information of favorite song, special edition, recommendation user add is good friend; Then perform step 8;

The user that step 5. once logged in has storage of history data P in database, and commending system reads this user's buddy list and friend circle and plays record from background data base;

The history that step 6. reads this user by commending system from background data base is again play record, if the frequent degree of listening in the recent period surpasses every day three times, surpasses half an hour at every turn, perform step 7, otherwise skips steps 7 directly performs step 8;

Step 7. commending system is the music of up-to-date interpolation in this user's commending friends circle;

Step 8. commending system carries out correlation rule recommendation according to playing record in this user's friend circle, for this user recommends music, and continues to record its evaluation result;

Step 9., at music recommend system run duration, can be monitored interest migration factor ε=p of this user always ²+ d * l, wherein p representation unit time skip song and the always ratio of recommendation song, d represents to delete the ratio that disagreeable song accounts for all recommendation songs, l represents that the deleted song of once liking accounts for all ratios of liking song, if meet acquiescence threshold values 0.6, perform step 10～step 12, otherwise directly perform step 13;

Step 10. couple user re-starts classification analysis, upgrades user's interest tendency label;

Step 11. couple user re-starts cluster analysis, upgrades friend recommendation list, adds new good friend, deletes old friendship friend;

Step 12. is carried out correlation rule recommendation according to playing record in the friend circle after upgrading;

Step 13. continues as that user provides music service until user exits this service.