CN105279289B - Individualized music based on exponential damping window recommends sort method - Google Patents
Individualized music based on exponential damping window recommends sort method Download PDFInfo
- Publication number
- CN105279289B CN105279289B CN201510885733.2A CN201510885733A CN105279289B CN 105279289 B CN105279289 B CN 105279289B CN 201510885733 A CN201510885733 A CN 201510885733A CN 105279289 B CN105279289 B CN 105279289B
- Authority
- CN
- China
- Prior art keywords
- music
- rule
- transition probability
- user
- similarity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
- G06F16/63—Querying
- G06F16/635—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
- G06F16/63—Querying
- G06F16/638—Presentation of query results
- G06F16/639—Presentation of query results using playlists
Abstract
The present invention provides a kind of, and the individualized music based on exponential damping window recommends sort method, wherein this method comprises: obtaining the rule of similarity transition probability matrix of all music between any two in music libraries；To operation after the different weights of the corresponding rule of similarity transition probability matrix corresponding line data imparting of user's playlist, obtains a line user and recommend transition probability；Recommend the transition probability of the corresponding user's music recommendation list of transition probability acquisition according to the user；The music in user's music recommendation list is ranked up according to the transition probability.The present invention solves the problems, such as new music cold start-up, can guarantee that higher individualized music recommends precision.
Description
Technical field
It is specifically exactly that one kind is based on the present invention relates to data processing technique more particularly to a kind of music recommended method
The individualized music of exponential damping window recommends sort method.
Background technique
With the fast development of music internet, people gradually pass through the music sources of the available magnanimity in internet.So
And presented while massive information, on the one hand make user feel that There is no way to begin when choosing music, on the other hand makes a large amount of
Music can not be found by the potential user that likes, and make this part music very few people interested.Individualized music is recommended can be with
A binary crelation is established between user and music, it is potential to go out each user by the relation excavation between music and user
Then interested music carries out personalized music and recommends.In fact, individualized music recommendation have on the internet it is important
Application prospect, be solve internet information overload effective way.
The personalized recommendation of music since it is promoting the adhesion relation between website and user in Web Hosting, at
For a popular research field, a large amount of experts and scholars have done a large amount of research work to this both at home and abroad.Currently, mature general
Music recommended method have: Collaborative Filtering Recommendation Algorithm, content-based recommendation algorithm, the proposed algorithm based on figure and based on rule
Proposed algorithm then.But the accuracy that content-based recommendation algorithm will cause recommendation is very low, substantially reduces personalized recommendation
Meaning；There is meeting with asking for new music bring cold start-up and large-scale calculations in various degree in other several proposed algorithms
Topic.
Just because of the presence of problem above, many researchers attempt to calculate content-based recommendation algorithm and other recommendations
Method mixes, and forms mixing proposed algorithm.However, existing a large amount of mixing proposed algorithms only specifically select in song
Certain small features in appearance, such as: then one of feature is brought into other algorithms public affairs by song label, singer etc.
In formula, formula is improved.Improvement small so only can be such that the scope of application of mixing proposed algorithm reduces, can not
Recommend suitable for the individualized music under different situations.
Therefore, those skilled in the art need a kind of music recommended method, are able to solve the stylish sound of music personalized recommendation
Happy cold start-up problem, while can guarantee higher personalized recommendation precision again.
Summary of the invention
In view of this, the technical problem to be solved in the invention is to provide a kind of personalization based on exponential damping window
Music recommends sort method, solves the problems, such as that existing music recommends personalization level not high.
In order to solve the above-mentioned technical problem, a specific embodiment of the invention provides a kind of based on exponential damping window
Individualized music recommends sort method, comprising: between any two based on all music in playlist rule of similarity calculating music libraries
The first rule of similarity transition probability matrix；Between any two second similar of all music in music libraries is calculated based on music content
Regular transition probability matrix；And superposition the first rule of similarity transition probability and the second rule of similarity transition probability obtain
The rule of similarity transition probability of all music between any two into music libraries；Give user's playlist corresponding rule of similarity transfer
Probability matrix corresponding line data assign operation after different weights, obtain a line user and recommend transition probability；It is pushed away according to the user
Recommend the transition probability that transition probability obtains corresponding user's music recommendation list；User's music is recommended to arrange according to the transition probability
Music in table is ranked up；
The weight is exponential damping window weight Wi, the exponential damping window weight WiCalculation formula are as follows:
Wi=(1-c)L-i
Wherein, WiFor exponential damping window weight；C is decay factor；L is the total quantity of music in user's playlist；i
For sequence of the current music in user's playlist, i is the integer for being less than or equal to L more than or equal to 1.
Based on the above-mentioned specific embodiment of the present invention it is found that the individualized music based on exponential damping window recommends sequence side
Method at least has the advantages that or feature: when music is recommended by content-based recommendation algorithm and rule-based recommendation
Algorithm neatly combines, and as much as possible plays the advantage of two kinds of algorithms, and overcome that the cold start-up of new music encounters asks
It inscribes, the music in utilization index decay window model treatment user's music recommendation list, to the sound in user's music recommendation list
Pleasure is ranked up by transition probability, so as to guarantee that higher individualized music recommends precision.
It is to be understood that above-mentioned general description and following specific embodiments are merely illustrative and illustrative, not
The range of the invention to be advocated can be limited.
Detailed description of the invention
Following appended attached drawing is part of specification of the invention, depicts example embodiments of the present invention, institute
Attached drawing is used to illustrate the principle of the present invention together with the description of specification.
Fig. 1 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment one of happy recommended method；
Fig. 2 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment two of happy recommended method；
Fig. 3 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment three of happy recommended method；
Fig. 4 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the example IV of happy recommended method；
Fig. 5 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment five of happy recommended method；
Fig. 6 is to change over time curve graph using the weight that decaying exponential function of the present invention is painted single music；
Fig. 7 compares schematic diagram with rule of similarity set for user's playlist；
Fig. 8 is rule of similarity product process figure of the invention；
Fig. 9 is the specific flow chart for the individualized music recommended method that the specific embodiment of the invention provides；
Figure 10 is that the music that the specific embodiment of the invention provides jumps and jump frequency schematic diagram.
Specific embodiment
Understand in order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below will with attached drawing and in detail
Narration clearly illustrates the spirit of disclosed content, and any skilled artisan is understanding the content of present invention
After embodiment, when the technology that can be taught by the content of present invention, it is changed and modifies, without departing from the essence of the content of present invention
Mind and range.
The illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but not as a limitation of the invention.
In addition, in the drawings and embodiments the use of element/component of same or like label is for representing same or like portion
Point.
About " first " used herein, " second " ... etc., not especially censure the meaning of order or cis-position,
It is non-to limit the present invention, only for distinguish with same technique term description element or operation.
About direction term used herein, such as: upper and lower, left and right, front or rear etc. are only the sides with reference to attached drawing
To.Therefore, the direction term used is intended to be illustrative and not intended to limit this creation.
It is open term, i.e., about "comprising" used herein, " comprising ", " having ", " containing " etc.
Mean including but not limited to.
About it is used herein " and/or ", including any of the things or all combination.
About term used herein " substantially ", " about " etc., to modify it is any can be with the quantity or mistake of microvariations
Difference, but this slight variations or error can't change its essence.In general, microvariations that such term is modified or error
Range in some embodiments can be 20%, in some embodiments can be 10%, can be in some embodiments 5% or its
His numerical value.It will be understood by those skilled in the art that the aforementioned numerical value referred to can be adjusted according to actual demand, it is not limited thereto.
It is certain to describe the word of the application by lower or discuss in the other places of this specification, to provide art technology
Personnel's guidance additional in relation to the description of the present application.
Fig. 1 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment one of happy recommended method, as shown in Figure 1, when recommending music to user from musical database, by base
Proposed algorithm in music content and the proposed algorithm based on rule of similarity neatly combine, solve in the prior art without
Method solves as time goes by aiming at the problem that new music is recommended, and the Music Appreciation grade of user changes, and
The problem of cannot carrying out music recommendation based on the variation.
The attached drawing specific embodiment includes:
Step 101: obtaining the rule of similarity transition probability matrix of all music between any two in music libraries.If music libraries
In share 100 songs, and this 100 song is that sequence arranges, and will not sequentially be changed, then rule of similarity transfer is general
Rate matrix is exactly one 100 × 100 matrix, this matrix shows the transition probability between any two song, needs to infuse
What is anticipated is the numerical value of m row, the n-th column into the rule of similarity transition probability homography that the n-th song shifts from m song
(1≤m≤100,1≤n≤100), and the rule of similarity transition probability homography shifted from the n-th song to m song
The numerical value of middle line n, m column, two values are not identical.
Step 102: assigning different power to the corresponding rule of similarity transition probability matrix corresponding line data of user's playlist
Operation after weight obtains a line user and recommends transition probability.If three songs are shared in user's playlist, in music libraries
Collating sequence is respectively 10,30,88, then the corresponding rule of similarity transition probability matrix corresponding line data of user's playlist are just
Be 100 × 100 matrix in the 10th row, the 30th row, the 88th row, each row of data all shows the music and jumps its in music libraries
The transition probability of its music, wherein defining a piece of music and jumping to the transition probability of oneself is 0.Then respectively to the 10th row, the
After 30 rows, the 88th row data assign different weights, by respective column phase operation, the data line that columns is 100, this line number are obtained
According to be exactly user recommend transition probability.
Step 103: recommending the transition probability of the corresponding user's music recommendation list of transition probability acquisition according to the user.With
Music in the music recommendation list of family is also all music in musical database, if, music in user's music recommendation list totally 4
Head, this sequence of 4 song in music libraries is respectively 15,45,55,70, then, just recommend one line number of transition probability from user
The 15th column, the 45th column, the 55th column, the 70th columns value are taken in, it is assumed that the 15th columns value is that the 0.25, the 45th columns value is 0.70,
55th columns value is that the 0.50, the 70th columns value is 0.82.
Step 104: the music in user's music recommendation list being ranked up according to the transition probability.According to from user
The 15th taken in recommendation transition probability arranges, the size of the 45th column, the 55th column, the 70th columns value is first to the 15th, the 45th head, the 55th
First, the 70th song is ranked up, i.e., is the 70th, the 45th, the 55th, the 15th song to the recommendation order of user, if only to
User recommends song, then this song is exactly the 70th head.
In the specific embodiment of the invention, the weight can be exponential damping window weight Wi, the exponential damping window
Mouth weight WiCalculation formula are as follows:
Wi=(1-c)L-i
Wherein, W is exponential damping window weight；C is decay factor, and in the specific embodiment of the invention, decay factor is
0.6；L is the total quantity of music in user's playlist；I be sequence of the current music in user's playlist, i be greater than etc.
In the integer that 1 is less than or equal to L.
The user recommends transition probability calculation formula NEXT specifically:
Wherein, eiFor the corresponding rule of similarity transition probability matrix corresponding line data of user's playlist.Since user broadcasts
It emplaces every a piece of music in table and is involved in and user is recommended, using exponential damping window, so that in user's playlist
Influence of the music to user's recommendation results gradually smoothly reduces with the time；User's click addition music can be regarded as constantly defeated
The data flow entered regards the recommendation next time of songs all in data flow as a smooth accumulated value, wherein the weight used
Constantly decaying, therefore, song occurs more early in a stream, and corresponding weight is also smaller.Also, the present invention uses index
Decay window recommends ability to change with time to record music, and exponential damping window calculation is only saved in calculating process and is obtained
Final weight, need not thus recommend all to recalculate historical data every time, improve treatment effeciency.
Fig. 2 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment two of happy recommended method, as shown in Fig. 2, when recommending music to user from musical database, by base
Proposed algorithm in music content and the proposed algorithm based on rule of similarity neatly combine, and give full play to two kinds of algorithms
Strong point.
The attached drawing specific embodiment includes:
Step 1011: the first similar rule of all music between any two in music libraries are calculated based on playlist rule of similarity
Then transition probability matrix.Specifically, playlist rule of similarity is to jump frequency according to per song in music libraries to obtain.
For a piece of music in music libraries, other any a piece of musics in music libraries may be jumped to, that is, need to acquire two-by-two
Frequency is jumped between music, in order to accurately recommend, it is therefore desirable to statistics for all user's playlists music two-by-two it
Between jump frequency.
Step 1012: it is general that the second rule of similarity transfer of all music between any two in music libraries being calculated based on music content
Rate matrix.For new music, recommend music just to will appear deviation based entirely on playlist, but if it is considered that music content it
Between relationship, just can solve the cold start-up problem of new music, for example, increasing a piece of music in music libraries, can be looked for by the lyrics
To music similar with its, and then the new music is recommended to the user listened to the new music similar music.
Step 1013: being superimposed the first rule of similarity transition probability and the second rule of similarity transition probability obtains sound
The rule of similarity transition probability of all music between any two in music storehouse.Both transition probabilities are combined to obtain music two-by-two
Between final transition probability, using final rule of similarity transition probability can accurately to user recommend music.
Referring to fig. 2, the present invention is recommended by the time series of music in playlist, can reflect the interest of user with
The variation of time and change；All music, which are involved in, in user's playlist recommends user, recommends user
Ability be gradually reduced with the variation of time, can also embody user and interest in music is changed with the variation of time；
It is by rule of similarity, the similar music being calculated based on music content is similar to arriving of being calculated by user's playlist
Music is united, on the one hand can perfection apply in algorithm frame proposed by the present invention, on the other hand also accomplished to be based on
The recommendation of rule and the unification of content-based recommendation.
Fig. 3 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment three of happy recommended method, as shown in figure 3, the first rule of similarity transition probability and the second rule of similarity are turned
It moves the numerical value in probability less than predetermined value to neglect, i.e., is 0 by the part rule of similarity transition probability.
In the attached drawing specific embodiment, based on music content calculate music libraries in all music between any two second
After the step of rule of similarity transition probability matrix, this method further include:
Step 1012-1: ignore the first rule of similarity transition probability less than first threshold.First threshold will be less than
The first rule of similarity transition probability be set to zero, first threshold can be 0.1~0.2 any value, specifically can root
It needs to set according to user.
Step 1012-2: ignore the second rule of similarity transition probability less than second threshold.Second threshold will be less than
The second rule of similarity transition probability be set to zero, second threshold can be 0.1~0.2 any value, and specific value can
To need to set according to user.
Referring to Fig. 3, there is a degree of randomness during playing music by user, so user's playlist
In to have the connected song correlation in some front and backs may not be especially big, it could even be possible to style difference is also very big.By
In the presence of this reason, for the music website of large-scale consumer amount, calculated by the historical data of user's playlist
The rule of similarity transition probability very weak there is a large amount of intensity in the obtained rule of similarity transition probability based on playlist.
The very weak regular transition probability of these intensity largely exists, and can seriously affect the operational efficiency of system and expend and largely deposit
Space is stored up, it is general by given threshold (for example, first threshold and second threshold are less than or equal to 0.2) the rule transfer very weak to intensity
Rate is trimmed, so that high-strength rule transition probability is retained, while improving search efficiency, also improves the accurate of recommendation
Degree.
Fig. 4 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the example IV of happy recommended method, as shown in figure 4, giving the first rule of similarity transition probability and the second similar rule respectively
Then transition probability distributes weight, then sums, and obtains in music libraries the rule of similarity transition probability between music two-by-two.
In the attached drawing specific embodiment, it is superimposed the first rule of similarity transition probability and second rule of similarity
Transition probability obtains the rule of similarity transition probability of all music between any two in music libraries, specifically includes:
Step 10131: the first rule of similarity transition probability is obtained into the first transition probability multiplied by the first weight.First
The setting of weight is mainly related with the quantity of the increased new music of music libraries, if the quantity of new music is seldom, the first weight meeting
It is big, conversely, the first weight can be smaller, for example, the first weight can be 0.5.It is primarily due to the first rule of similarity transition probability
The main correlation for considering playlist, only old user just have playlist, and the first rule of similarity transition probability can be quasi-
Really reflect the Music Appreciation tendency of user.
Step 10132: the second rule of similarity transition probability is obtained into the second transition probability multiplied by the second weight.Second
The setting of weight is also mainly related with the quantity of the increased new music of music libraries, if the quantity of new music is seldom, the second weight
Can be smaller, conversely, the second weight can more greatly, for example, the second weight can be 0.5.It is general to be primarily due to the transfer of the second rule of similarity
Rate mainly considers the correlation of music content, can solve the problems, such as the cold start-up of new music very well.
Step 10133: first transition probability being added to obtain rule of similarity transfer generally with second transition probability
Rate.Wherein, first weight and second weights sum are 1, and usual first weight and the second weight are 0.5.At this
Invention other specific embodiments in, as user listens to the growth of music number, the first weight is gradually increased, the second weight by
It is decrescence small, to further increase the accuracy of recommendation.
Fig. 5 is a kind of personalized sound based on exponential damping window and rule of similarity that the specific embodiment of the invention provides
The flow chart of the embodiment five of happy recommended method, as shown in figure 5, obtaining user's played column according to rule of similarity transition probability matrix
The recommendation list of all music in table, seeks the union of these recommendation lists, and by music existing in user's playlist from sound
Find pleasure in and concentrate exclusion, can be obtained user's music recommendation list.
In the attached drawing specific embodiment, based on the final rule of similarity transition probability and user's playlist to
Music is recommended at family, is specifically included:
Step 1031: the recommendation of each music in user's playlist is obtained based on the rule of similarity transition probability matrix
List.Per song has a recommendation set.
Step 1032: the union of all recommendation lists being asked to obtain recommendation list set.Ask these recommendation lists and
Collection, so that it may obtain the recommendation list set of user.
Step 1033: from already present music in user's playlist is removed in the recommendation list set, to obtain
User's music recommendation list.Since already present music does not need to recommend to user again in user's playlist, so if
There are already present music in user's playlist in recommendation list set, need to exclude these music, it is accurate to guarantee to recommend
Degree.
The recommended models based on exponential damping window in the present invention, are specifically described as follows:
Exponential damping window is introduced for calculating a smooth aggregate-value in data flow (song stream), so that current number
All retained according to (current song) and the information in historical data (historical song).Therefore, with the continuous inflow of data flow,
The weight of historical data stream is constantly decayed, and song occurs more early in song stream, and weight is also smaller, the influence to result
Also smaller.As shown in fig. 6, for the curve that changes over time of weight of each song, in figure, Wi=(1-c)L-i, wherein in formula
WiFor exponential damping window weight；C is decay factor；L is the total quantity of music in user's playlist；I is current music
Sequence in user's playlist, i are the integer for being less than or equal to L more than or equal to 1.Wherein, enabling song stream is { e1,e2,...,
e3, wherein e1The song played for first, enFor currently playing song.C is decay factor, then the index of song stream declines
Subtract window (exponentially decaying window) is defined as:
Although all songs in the proposed algorithm playlist based on figure have been involved in user-customized recommended,
It is all participated in the distribution resource with same status due to possessing each song of resource in playlist, so not can reflect
The variation that the interest of user occurs with the variation of time.Markov model has time series, is able to reflect out user
Interest changes with time.But the algorithm model is recommended only according to last song, is equal to when being recommended
Have ignored the shadow for the song that other all songs in user's playlist other than last song will listen to user future
It rings, necessarily will cause a large amount of information in this way and be ignored.Therefore, it by being improved to Markov Chain, is declined by index
Subtract window to retain a part of information in the historical record of playlist, gives each song in temporal sequence then with not
Same resource allocation, the result distributed with all songs carry out personalized recommendation to user.
As shown in fig. 7, can all obtain a rule of similarity set according to each song of user's playlist.This phase
Being like regular collection is exactly song tiThe rule of similarity set of (1≤i≤m).Song in these rule of similarity set, be all with
There is the songs of certain incidence relation for song in user's playlist.Therefore, by association index exponential damping window model,
Last recommendation results can be combined with the recommendation results of songs all in playlist, user often listens to a first song
Song, history recommendation results weight just become original (1-c) times.
Rule of similarity generation method introduction based on playlist in a specific embodiment of the invention:
To a certain user uk, the sequence of songs Playlist that listens tok{t1,t2,...,tm, ukEvery tin listened to a song
txThere is a rule of similarity setThe recommendation collection for the last item song that then user listens to is combined intoUser uk
Recommend gather are as follows:
Wherein, m is the sum of rule of similarity set music；K is number of users；N is the number of music in user's playlist；
It can recommend to gatherThe music in user's playlist is excluded.
For a song t in user's playlisti, need to calculate the song collection of weight are as follows:
Wherein, ukFor user；For song tiRecommend gather.
Herein, the transition probability of association index exponential damping window and Markov model, user's playlist history
Record defines user's playlist playlistkIn the i-th song tiAllocate resources to song tjCapacity of water be such as
Lower formula:
Wherein, tiFor a song in user's playlist；M is the total number of song in user's playlist；It obtains
User ukSet to be recommendedIn all songs distribution resource ability after, need to calculate set to be recommended
In each song allocate resources to the probability (i.e. rule of similarity transition probability) of particular songs:
Wherein,Indicate the i-th song t in user's playlistiAllocate resources to jth head sound
Happy tjAbility；Indicate the i-th song t in user's playlistiAllocate resources to xth head sound
Happy txAbility；M is the total number of music in user's playlist；I is positive integer.
Rule of similarity generation method introduction in the present invention based on song content:
The content of song can be divided into: the lyrics, song audio files, song metadata (including singer, song classification etc.)
Deng these song contents all can serve as the foundation of Similarity measures between song.
Firstly, for the lyrics, need successively to proceed as follows: participle, stem extract (English), remove stop words, feature
Selection, characteristic weighing and similarity calculation.The high efficiency method that feature selecting is carried out on extensive text set is document frequency (DF).
DF is by one threshold value of setting, so that the word lower than this threshold value is considered as low frequency words, the word higher than this threshold value is considered as
High frequency vocabulary.Low frequency words are removed from feature space, and then reduce the dimension of feature space.Common characteristic weighing algorithm
For TF-IDF algorithm, formula is as follows:
Wherein, TF is the frequency of word in the text, | D | it is text sum, DF is the frequency that some word occurs in all texts
Number.The vector space of the word obtained according to the lyrics can thus be obtained.
Song files content often extracts each frame feature vector of song audio with Mel frequency cepstral coefficient (MFCC),
Form the characteristic vector space of song audio.
Generally comprised according to the method that characteristic vector space calculates the similarity between song: Euclidean distance, cos are similar
Degree, minihash method, SimRank etc..Wherein cos similarity is considered as one of calculated result the best way.For one
A feature vector t=(v1,v2,...,vn), cos similarity formula is specific as follows:
Then to song tiAll similar songs be normalized, obtain in user's playlist based on song content
The seemingly regular transition probability of song(i.e. second like regular transition probability):
Wherein, n is all in music libraries and the i-th song tiThe number of similar music；K is positive integer.
In actual application of the present invention, there is a degree of randomness, institutes during playing song by user
It is not especially greatly, it could even be possible to style phase to have the possible correlation of the connected song in some front and backs in user's playlist
Difference is also very big.Presence due to this passes through user's playlist for the song website of large-scale consumer amount
There is the rules that a large amount of intensity is very weak in the rule of similarity based on playlist that historical data is calculated.These intensity
A large amount of presence of very weak rule, can seriously affect the operational efficiency of system and expend a large amount of memory space.In view of one
Can have a large amount of cold songs or new song in system will be generated this part song if trimmed using support
Rule trimming fall.So it is proposed that being trimmed by confidence level to weak rule, so that strong rule is retained.It enables
Support support (ti→tj) it is obtained in training set from song tiBrowse to song tjFrequency, it may be assumed that
That is:
Regard song every in user record as a node, then confidence level confidence (ti→tj) it is that institute is useful
Father node is t in the playlists records of familyiIn the case where, child node is tjPercentage, i.e. father node is tiUnder the conditions of condition
Probability, calculation formula are as follows:
Similarly, when calculating the rule of similarity based on content, it also will appear the comparatively small rule of a large amount of regular weights
Then, if these rule trimmings, a large amount of memory space is not only consumed, but also seriously affects search efficiency when algorithm executes,
So should also be as being trimmed using identical method.
The final rule of similarity of the present invention is the rule of similarity generated based on playlist and is generated based on song content
The final superposition of rule of similarity.Its calculation formula is as follows:
At this point, it should be noted that due toWithIt is all to be obtained after normalization as a result, institute
It is irreversible with its similitude, i.e. sim (ti,tj) and sim (tj,ti) the result is that unequal, similarlyWithResult be also unequal.
The present invention first has to record the ordering rule pair for generating and mutually jumping between song (music) according to the broadcasting of user,
The process of specific generation method is as shown in figure 8, the generation step process of rule of similarity is specific as follows:
Step 1: counting the frequency that music adjacent in temporal sequence in all playlists occurs, i.e., in playlist
Preceding a piece of music is father node, and a subsequent node is child node, and this sequence node is irreversible；
Step 2: this frequency is to be based on playlist according to the hop frequencies between the frequency calculate node of the first step
Rule of similarity；
Step 3: trimming according to confidence level to the rule of similarity that second step obtains, weak rule is removed；
Step 4: according to audio file extract MFCC feature vector or according to lyrics file extract weighted feature word to
Amount, generates the feature vector of audio；
Step 5: calculating the COS similarity between audio according to the feature vector of audio；
Step 6: the similar audio set to some audio being calculated is normalized, base has thus been obtained
In the rule of similarity of content；
Step 7: the rule of similarity that third step and the 6th step generate is weighted, final rule of similarity is obtained.
The overall flow figure of inventive algorithm as shown in figure 9, be algorithm frame overall flow below:
Step 1: finding out all music similar with music in user list, composition can recommend collection of music；
Step 2: it is all 0 that class, which recommends the music initial weight in collection of music, the then First from user's playlist
Music starts, and all similar musics is found from the rule of similarity set of this song, and rule of similarity weight is added to can
Recommend the music in collection of music；Then the second song in playlist is begun looking for, first by sounds all in playlist
Happy rule of similarity weight is multiplied by (1-c), then by the rule of similarity weight of music in the rule of similarity set of the second song
It is added to the music that can recommend in collection of music.Music all in playlist all repeats the operation of the second song later,
Until last a piece of music；
Step 3: by recommending the last cumulative obtained weights of music peace to be ranked up in second；
Step 4: removing topN therein according to specific requirements and forming recommendation results.
The specific embodiment of the present invention:
(1) instance analysis that rule of similarity generates
Example: the input of algorithm includes the music playlist and audio content data of user.As table 1 to example be 6
The playlist of a user music that oneself is selected on some music website:
1. user's playlist of table
| User | Playlist |
| U1 | t1,t3,t5 |
| U2 | t3,t5,t2,t1,t4 |
| U3 | t1,t3,t5,t2 |
| U4 | t4,t1,t5,t2 |
| U5 | t1,t4,t3 |
| U6 | t3,t5 |
| U7 | t5,t1 |
Steps are as follows for the example calculation:
(1) music in the playlist of user is separated two-by-two in order, such as playlistR (t can be divided into1,t3, 1) and R (t3,t5,1)(R(ti,tj, n) in tiRepresent father node, tjGeneration
Table child node, n, which is represented, jumps frequency between father node and child node).In statistical form the playlist of all users to get arrive sound
Frequency chart is jumped between pleasure, as shown in Figure 10, for example, t1Jump to t5Number be 1, t1Jump to t4Number be 2, t1It jumps
To t3Number be 2, t1Jump to t2Number be 0, t1Jump to t1Number be 0.
(2) I={ t is enabled1,t2,t3,t4,t5It is all song collections, then music rule of similarity based on playlist
Weight computing result is as shown in following display:
Wherein, t1The total degree for jumping to other music is 5, for above-mentioned display the first row, uses t respectively1Jump to t1、
t2、t3、t4、t5Number be 0,0,2,2,1 divided by total degree 5, other rows.
(3) here, enabling confidence level is 0.2, then after trimming rule, the rule of similarity based on playlist are as follows:
Here, it is assumed that the feature vector of every song has been got by content, each feature vector weight such as the following table 2
It is shown:
2. song content feature vector value of table
(4) according to the combination of eigenvectors in table 2, the cos similarity of audio between any two can be calculated, it is calculated
Concrete outcome is as shown in following matrixes:
(5) then cos similarity is normalized, obtains the rule of similarity power of the rule of similarity based on content later
Value is as follows after normalizing:
(6) rule of similarity is trimmed, deletes rule of the rule of similarity weight less than 0.2, obtains the phase based on content
Like the weight results of rule are as follows:
(7) weight w=0.5 for enabling two kinds of rule of similarities be superimposed, the then weight for generating final rule of similarity are as follows:
(2) personalized recommendation
It is assumed that decay factor c=0.6, and to user U7Personalized recommendation is carried out, recommendation results are 1.It is specific to calculate step
It is rapid as follows:
(1) determination can be recommended to gather.The playlist of this user ist5Recommendation collection be combined intot1Recommendation collection be combined intoCollection can be then recommended to be combined into
(2) by first song t in playlist5The weight of rule of similarity be assigned toIt can recommend the song in set
Song, i.e.,Carry out next time calculate before, should first byThe power of middle song
Value is multiplied by (1-c)=0.4, i.e.,Then by next song t1Rule of similarity
Weight tax be added toIn song, i.e.,
(3) rightIt is ranked up according to calculated weight
(4) when recommendation results are one, by the t of maximum weight4Recommend user.
The present invention also at least has the advantages that
1. the present invention is recommended by the time series of music in playlist, user interest can reflect with the time
Change and changes.
User is recommended 2. all music are involved in user's playlist of the present invention, the energy that user is recommended
Power is gradually reduced with the variation of time, can also embody variation of the user to interest in music.
Ability is recommended to change with time 3. the present invention records music using exponential damping window, in this way in the mistake of calculating
The final weight that exponential damping window calculation obtains only is saved in journey, need not thus recommend all to recalculate history number every time
According to.
4. the present invention by the similar music being calculated based on content and passes through user's played column meter by rule of similarity
Calculate to similar music unite, on the one hand can perfection apply in algorithm frame proposed by the present invention, on the other hand
Also the unification between rule-based recommendation and content-based recommendation has been accomplished.
5. the present invention is based on content and rule-based mixed recommendation, in the case where being able to maintain higher recommendation precision
Solve the problems, such as that new projects' bring is cold-started.
6. the present invention is greatly reduced by trimming to orderly rule of similarity in the case where losing very little accuracy rate
Time of the algorithm in actual motion.
7. the personalized recommendation environment that the present invention can be suitable for all user behaviors and content, it is only necessary to based on content
Similitude individually consider after can use the method that mention of the present invention, use is more flexible.
8. the present invention in the case where not calculating the rule of similarity based on content, can obtain than based on figure proposed algorithm and
The higher accuracy rate of Hidden Markov Model and recall rate can have very in the case where calculating the rule of similarity based on content
Good recommendation performance.
9. the present invention uses exponential damping window, so that music recommends ability to change over time and gradually subtract in playlist
It is small.
10. the present invention is to use orderly rule of similarity as the foundation of the similar recommendation of per song.
11. the present invention orderly music adjacent in playlist is considered as it is relevant, user select playlist be
In selection similar music, so calculating the rule of similarity based on playlist by playlist.
12. the similar music based on content is normalized the present invention, it is considered as the rule of similarity based on content.
13. the present invention realizes mixed recommendation by the superposition based on content and rule-based rule of similarity weight.
14. the present invention calculate rule of similarity during respectively to based on content and based on the rule of similarity of playlist into
Row trimming recommends the similar music that weight is low in set to reduce per song, and then improves the search efficiency of collection of programs.
15. the present invention calculates rule of similarity weight by the transition probability between music in playlist.
Rule of similarity is the rule of similarity based on broadcasting and weight between the rule of similarity based on content in 16 present invention
Superposition is realized and is based on content and rule-based mixed recommendation.
17. the present invention when calculating separately based on content and rule-based rule of similarity weight, carries out rule
Memory space is saved in trimming, is improved and is recommended efficiency.
The present invention provides a kind of individualized music recommendation sort method based on exponential damping window, when music is recommended,
Content-based recommendation algorithm and rule-based proposed algorithm are neatly combined, all music in music libraries are obtained
Rule of similarity transition probability matrix between any two as much as possible plays the advantage of two kinds of algorithms；According in playlist
All music obtain user and recommend transition probability, can guarantee that higher individualized music recommends precision.
The above-mentioned embodiment of the present invention can be implemented in various hardware, Software Coding or both combination.For example, this hair
Bright embodiment can also be in the middle execution executed of data signal processor (Digital Signal Processor, DSP)
State the program code of program.The present invention, which can also refer to computer processor, digital signal processor, microprocessor or scene, to be compiled
The multiple functions that journey gate array (Field Programmable Gate Array, FPGA) executes.It can configure according to the present invention
State processor and execute particular task, by execute define the machine-readable software code of ad hoc approach that the present invention discloses or
Firmware code is completed.Software code or firmware code can be developed as different program languages and different formats or form.
It can also be in order not to same target platform composing software code.However, executing the software code and other classes of task according to the present invention
Different code pattern, type and the language of type configuration code do not depart from spirit and scope of the invention.
The foregoing is merely the schematical specific embodiments of the present invention, before not departing from conceptions and principles of the invention
It puts, the equivalent changes and modifications that any those skilled in the art is made should belong to the scope of protection of the invention.
Claims (8)
1. a kind of individualized music based on exponential damping window recommends sort method, which is characterized in that this method comprises:
The the first rule of similarity transition probability square of all music between any two in music libraries is calculated based on playlist rule of similarity
Battle array；
The the second rule of similarity transition probability matrix of all music between any two in music libraries is calculated based on music content；And
It is superimposed the first rule of similarity transition probability and the second rule of similarity transition probability obtains all sounds in music libraries
Happy rule of similarity transition probability between any two；
To operation after the different weights of the corresponding rule of similarity transition probability matrix corresponding line data imparting of user's playlist, obtain
A line user recommends transition probability；
Recommend the transition probability of the corresponding user's music recommendation list of transition probability acquisition according to the user；And
The music in user's music recommendation list is ranked up according to the transition probability；
The weight is exponential damping window weight Wi, the exponential damping window weight WiCalculation formula are as follows:
Wi=(1-c)L-i
Wherein, WiFor exponential damping window weight；C is decay factor；L is the total quantity of music in user's playlist；I is to work as
Sequence of the preceding music in user's playlist, i are the integer for being less than or equal to L more than or equal to 1.
2. the individualized music based on exponential damping window recommends sort method as described in claim 1, which is characterized in that institute
It states user and recommends transition probability calculation formula NEXT specifically:
Wherein, eiFor the corresponding rule of similarity transition probability matrix corresponding line data of user's playlist.
3. the individualized music based on exponential damping window recommends sort method as described in claim 1, which is characterized in that institute
Stating decay factor is 0.6.
4. the individualized music based on exponential damping window recommends sort method as described in claim 1, which is characterized in that base
After the step of music content calculates all music the second rule of similarity transition probability matrix between any two in music libraries, also
Include:
Ignore the first rule of similarity transition probability less than first threshold；And
Ignore the second rule of similarity transition probability less than second threshold.
5. the individualized music based on exponential damping window recommends sort method as claimed in claim 4, which is characterized in that institute
It states first threshold and the second threshold is less than or equal to 0.2.
6. the individualized music based on exponential damping window recommends sort method as described in claim 1, which is characterized in that folded
The first rule of similarity transition probability and the second rule of similarity transition probability is added to obtain in music libraries all music two-by-two
Between rule of similarity transition probability, specifically include:
The first rule of similarity transition probability is obtained into the first transition probability multiplied by the first weight；
The second rule of similarity transition probability is obtained into the second transition probability multiplied by the second weight；And
First transition probability is added to obtain rule of similarity transition probability with second transition probability,
Wherein, first weight and second weights sum are 1.
7. the individualized music based on exponential damping window recommends sort method as claimed in claim 6, which is characterized in that institute
The first weight is stated to increase with increasing for music in user's playlist.
8. the individualized music based on exponential damping window recommends sort method as described in claim 1, which is characterized in that institute
User's music recommendation list is stated to obtain by following steps:
The recommendation list of per song in user's playlist is obtained based on the rule of similarity transition probability matrix；
The union of all recommendation lists is asked to obtain recommendation list set；And
From already present music in user's playlist is removed in the recommendation list set, so that obtaining user's music recommends column
Table.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510885733.2A CN105279289B (en) | 2015-12-04 | 2015-12-04 | Individualized music based on exponential damping window recommends sort method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510885733.2A CN105279289B (en) | 2015-12-04 | 2015-12-04 | Individualized music based on exponential damping window recommends sort method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105279289A CN105279289A (en) | 2016-01-27 |
| CN105279289B true CN105279289B (en) | 2019-03-22 |
Family
ID=55148303
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510885733.2A Expired - Fee Related CN105279289B (en) | 2015-12-04 | 2015-12-04 | Individualized music based on exponential damping window recommends sort method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105279289B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107835441B (en) * | 2017-10-10 | 2020-01-03 | 武汉斗鱼网络科技有限公司 | Live broadcast recommendation method, storage medium, device and system based on path prediction |
| CN107818483B (en) * | 2017-11-27 | 2021-08-24 | 微梦创科网络科技（中国）有限公司 | Network card and ticket recommendation method and system |
| CN109961329A (en) * | 2017-12-14 | 2019-07-02 | 北京京东尚科信息技术有限公司 | Articles handling method and device, storage medium and electronic equipment |
| CN108153898A (en) * | 2018-01-10 | 2018-06-12 | 上海展扬通信技术有限公司 | Audio frequency playing method, terminal and computer readable storage medium |
| CN111488484B (en) * | 2020-04-16 | 2023-04-28 | 北京雷石天地电子技术有限公司 | Song recommendation method, medium and device based on class index distribution and TF-IDF characteristics |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102263992A (en) * | 2011-08-08 | 2011-11-30 | 上海文广互动电视有限公司 | Program recommendation degree estimation method based on user viewing record |
| CN103544212A (en) * | 2013-09-09 | 2014-01-29 | Tcl集团股份有限公司 | Content recommending method and system |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103365842B (en) * | 2012-03-26 | 2016-08-24 | 阿里巴巴集团控股有限公司 | A kind of page browsing recommends method and device |
| CN102982131B (en) * | 2012-11-16 | 2015-12-23 | 杭州东信北邮信息技术有限公司 | A kind of based on markovian book recommendation method |
| CN105022760B (en) * | 2014-04-30 | 2019-06-25 | 深圳市腾讯计算机系统有限公司 | A kind of news recommended method and device |
| CN105095279B (en) * | 2014-05-13 | 2019-05-03 | 深圳市腾讯计算机系统有限公司 | File recommendation method and device |
-
2015
- 2015-12-04 CN CN201510885733.2A patent/CN105279289B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102263992A (en) * | 2011-08-08 | 2011-11-30 | 上海文广互动电视有限公司 | Program recommendation degree estimation method based on user viewing record |
| CN103544212A (en) * | 2013-09-09 | 2014-01-29 | Tcl集团股份有限公司 | Content recommending method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105279289A (en) | 2016-01-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Interiano et al. | Musical trends and predictability of success in contemporary songs in and out of the top charts | |
| Kaminskas et al. | Location-aware music recommendation using auto-tagging and hybrid matching | |
| CN105279289B (en) | Individualized music based on exponential damping window recommends sort method | |
| Kenter et al. | Ad hoc monitoring of vocabulary shifts over time | |
| CN102024058B (en) | Music recommendation method and system | |
| CN110532479A (en) | A kind of information recommendation method, device and equipment | |
| CN106919575A (en) | application program searching method and device | |
| Germain et al. | Spotify me: Facebook-assisted automatic playlist generation | |
| Monti et al. | An ensemble approach of recurrent neural networks using pre-trained embeddings for playlist completion | |
| Darshna | Music recommendation based on content and collaborative approach & reducing cold start problem | |
| CN113813609A (en) | Game music style classification method and device, readable medium and electronic equipment | |
| Korzeniowski et al. | Mood classification using listening data | |
| Ariyaratne et al. | A novel automatic hierachical approach to music genre classification | |
| CN105488195A (en) | Search result diversification ordering method based on hierarchical structure subtopic | |
| Jiang et al. | Unveiling music genre structure through common-interest communities | |
| Zhang et al. | A personalized next-song recommendation system using community detection and markov model | |
| Chen et al. | Music recommendation based on multiple contextual similarity information | |
| CN109471951A (en) | Lyrics generation method, device, equipment and storage medium neural network based | |
| CN108717445A (en) | A kind of online social platform user interest recommendation method based on historical data | |
| CN106294563B (en) | A kind for the treatment of method and apparatus of multi-medium data | |
| Schindler | Multi-modal music information retrieval: augmenting audio-analysis with visual computing for improved music video analysis | |
| CN109992694A (en) | A kind of music intelligent recommendation method and system | |
| Paulus | Improving Markov Model Based Music Piece Structure Labelling with Acoustic Information. | |
| Zhao et al. | Trailmix: An ensemble recommender system for playlist curation and continuation | |
| Matsumoto et al. | Context-aware network analysis of music streaming services for popularity estimation of artists |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190322 Termination date: 20211204 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |