CN105279289A - Personalized music recommendation sorting method based on exponential decay window - Google Patents

Abstract

The invention provides a personalized music recommendation sorting method based on an exponential decay window. The personalized music recommendation sorting method comprises the steps that similarity rule transition probability matrixes of all pieces of music in twos in a music library are acquired; different weights are given to data of corresponding rows of the similarity rule transition probability matrixes corresponding to user playlists, and then operation is performed, so that a row of user recommendation transition probabilities are obtained; according to the user recommendation transition probabilities, the transformation probabilities of corresponding user music recommendation lists are obtained; according to the transformation probabilities, the music in the user music recommendation lists is sorted. The personalized music recommendation sorting method solves the cold start problem of new music and can guarantee high personalized music recommendation accuracy.

Description

Individualized music based on exponential damping window recommends sort method

Technical field

The present invention relates to data processing technique, particularly relate to a kind of music recommend method, is exactly that a kind of individualized music based on exponential damping window recommends sort method specifically.

Background technology

Along with the fast development of music internet, people can obtain the music sources of magnanimity gradually by internet.But present while magnanimity information, There is no way to begin to make user feel when choosing music on the one hand, make a large amount of music cannot be found by the potential user liked on the other hand, and very few people interested to make this part music.Individualized music recommends between user and music, to set up a binary relation, goes out the potential interested music of each user by the relation excavation between music and user, then carries out personalized music recommend.In fact, individualized music recommends there is important application prospect on the internet, is the effective way solving internet information overload.

The personalized recommendation of music promotes the adhesion relation between website and user due to it in Web Hosting, and therefore become 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.At present, ripe general music recommend method has: Collaborative Filtering Recommendation Algorithm, content-based recommendation algorithm, based on the proposed algorithm of figure and rule-based proposed algorithm.But content-based recommendation algorithm can cause the accuracy of recommendation very low, greatly reduces the meaning of personalized recommendation; There is the problem meeting with cold start-up problem that new music brings and large-scale calculations in various degree in other several proposed algorithms.

Just because of the existence of above problem, many researchers attempt content-based recommendation algorithm and other proposed algorithms to mix, and form mixing proposed algorithm.But existing a large amount of mixing proposed algorithm just specifically selects some small feature in song content, such as: song label, singer etc., be then brought in other algorithmic formulas by one of them feature, improve formula.Improvement small like this, only can make the scope of application of mixing proposed algorithm reduce, and the individualized music that can not be applicable under different situation is recommended.

Therefore, those skilled in the art need a kind of music recommend method badly, can solve the cold start-up problem of the stylish music of music personalized recommendation, can ensure again higher personalized recommendation precision simultaneously.

Summary of the invention

In view of this, the technical issues that need to address of the present invention are to provide a kind of individualized music based on exponential damping window to recommend sort method, solve the problem that existing music recommend personalization level is not high.

In order to solve the problems of the technologies described above, an embodiment of the present invention provides a kind of individualized music based on exponential damping window to recommend sort method, comprising: obtain all music rule of similarity transition probability matrix between any two in music libraries; Give computing after different weight to the rule of similarity transition probability matrix corresponding line data that user's playlist is corresponding, obtain a line user and recommend transition probability; Transition probability is recommended to obtain the transition probability of respective user music recommend list according to described user; According to described transition probability, the music in the list of user's music recommend is sorted.

Known based on the above-mentioned embodiment of the present invention, individualized music based on exponential damping window recommends sort method at least to have following beneficial effect or feature: content-based recommendation algorithm and rule-based proposed algorithm are combined neatly when music recommend, as much as possible play the advantage of two kinds of algorithms, overcome the problem that new music cold start-up runs into, music in the list of utilization index decay window model treatment user music recommend, music in the list of user's music recommend is sorted by transition probability, thus can ensure that higher individualized music recommends precision.

It is to be understood that above-mentioned general description and following embodiment are only exemplary and illustrative, its can not limit the present invention for advocate scope.

Accompanying drawing explanation

Appended accompanying drawing is below a part for instructions of the present invention, and it depicts example embodiment of the present invention, and appended accompanying drawing is used for principle of the present invention is described together with the description of instructions.

The process flow diagram of the embodiment one of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 1 provides for the specific embodiment of the invention;

The process flow diagram of the embodiment two of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 2 provides for the specific embodiment of the invention;

The process flow diagram of the embodiment three of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 3 provides for the specific embodiment of the invention;

The process flow diagram of the embodiment four of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 4 provides for the specific embodiment of the invention;

The process flow diagram of the embodiment five of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 5 provides for the specific embodiment of the invention;

Fig. 6 is the weights change curve in time utilizing decaying exponential function of the present invention to illustrate single music;

Fig. 7 is that user's playlist contrasts schematic diagram with rule of similarity set;

Fig. 8 is rule of similarity product process figure of the present invention;

The particular flow sheet of the individualized music recommend method that Fig. 9 provides for the specific embodiment of the invention;

The music redirect that Figure 10 provides for the specific embodiment of the invention and redirect frequency schematic diagram.

Embodiment

Clearly understand for making the object of the embodiment of the present invention, technical scheme and advantage, below by with accompanying drawing and describe the spirit clearly demonstrating disclosed content in detail, any art technician is after the embodiment understanding content of the present invention, when can by the technology of content institute of the present invention teaching, be changed and modify, it does not depart from spirit and the scope of content of the present invention.

Schematic description and description of the present invention is for explaining the present invention, but not as a limitation of the invention.In addition, in drawings and the embodiments use element/component that is identical or like numerals will to be used to represent identical or similar portions.

About " first " used herein, " second " ... Deng, the not special meaning of censuring order or cis-position, is also not used to limit the present invention, and it is only in order to distinguish the element or operation that describe with constructed term.

About direction used herein term, such as: upper and lower, left and right, front or rear etc., be only the direction with reference to accompanying drawing.Therefore, the direction term of use is used to illustrate and is not used for limiting this creation.

About " comprising " used herein, " comprising ", " having ", " containing " etc., be open term, namely mean including but not limited to.

About used herein " and/or ", comprise the arbitrary of described things or all combine.

About term used herein " roughly ", " about " etc., in order to modify any can the quantity of microvariations or error, but this slight variations or error can't change its essence.Generally speaking, the scope of the microvariations that this type of term is modified or error can be 20% in some embodiments, can be 10% in some embodiments, can be 5% or other numerical value in some embodiments.It will be understood by those skilled in the art that the aforementioned numerical value mentioned can adjust according to actual demand, not as limit.

Some in order to the word that describes the application by lower or discuss in the other places of this instructions, to provide those skilled in the art about guiding extra in the description of the application.

The process flow diagram of the embodiment one of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 1 provides for the specific embodiment of the invention, as shown in Figure 1, when recommending music to user from musical database, proposed algorithm based on music content and the proposed algorithm based on rule of similarity are combined neatly, to solve in prior art and cannot carry out the problem of recommending for new music, and solve along with passage of time, the Music Appreciation grade of user changes, and can not carry out the problem of music recommend based on this change.

This accompanying drawing embodiment comprises:

Step 101: obtain all music rule of similarity transition probability matrix between any two in music libraries.If have 100 songs in music libraries, and this 100 song is order arrangement, order can not change, so rule of similarity transition probability matrix is exactly the matrix of 100 × 100, this matrix shows the transition probability between any two songs, it should be noted that the rule of similarity transition probability homography shifted from m song to the n-th song, m is capable, numerical value (1≤m≤100 of the n-th row, 1≤n≤100), and n-th line the rule of similarity transition probability homography shifted from the n-th song to m song, the numerical value of m row, two numerical value are not identical.

Step 102: give computing after different weight to the rule of similarity transition probability matrix corresponding line data that user's playlist is corresponding, obtain a line user and recommend transition probability.If have three songs in user's playlist, clooating sequence in music libraries is respectively 10,30,88, the rule of similarity transition probability matrix corresponding line data that so user's playlist is corresponding are exactly the 10th row, the 30th row, the 88th row in the matrix of 100 × 100, each row of data all show the transition probability of other music in this music redirect music libraries, wherein, definition a piece of music jumps to the transition probability of oneself is 0.After giving different weights then to respectively the 10th row, the 30th row, the 88th row data, by the computing of respective column phase, obtain the data line that columns is 100, these row data are exactly that user recommends transition probability.

Step 103: recommend transition probability to obtain the transition probability of respective user music recommend list according to described user.Music in the list of user's music recommend is also all music in musical database, if, music in the list of user's music recommend is totally 4 head, the order of this 4 song in music libraries is respectively 15,45,55,70, so, just recommends to get the 15th row, the 45th row, the 55th row, the 70th columns value transition probability data line from user, assuming that the 15th columns value is 0.25,45th columns value is the 0.70,55th columns value be the 0.50,70th columns value is 0.82.

Step 104: the music in the list of user's music recommend is sorted according to described transition probability.According to the size of the 15th row recommending from user to get in transition probability, the 45th row, the 55th row, the 70th columns value, the 15th first, the 45th first, the 55th first, the 70th song are sorted, namely the recommendation order to user is the 70th, the 45th, the 55th, the 15th song, if only recommend song to user, so this song is exactly the 70th head.

In the specific embodiment of the invention, described weight can be exponential damping window weights W _i, described exponential damping window weights W _icomputing formula be:

W _i＝(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 is the order of current music in user's playlist, and i is more than or equal to the integer that 1 is less than or equal to L.

Described user recommends transition probability computing formula NEXT to be specially:

$NEXT=\underset{i=1}{\overset{L}{\Σ}}{e}_i{W}_i$

Wherein, e _ifor the rule of similarity transition probability matrix corresponding line data that user's playlist is corresponding.Because a piece of music every in user's playlist all participates in recommending user, use exponential damping window, make the music in user's playlist to the impact of user's recommendation results along with the time smoothly reduces gradually; User can be clicked the data stream of adding music and regarding constantly input as, by songs all in data stream recommend next time regard a level and smooth accumulated value as, the weight wherein adopted constantly decays, therefore, song occurs more in a stream, and its corresponding weight is also less.And, the present invention adopts exponential damping window to carry out recording musical and recommends ability over time, in computation process, only preserve the final weights that exponential damping window calculation obtains, so just need not recommend all to recalculate historical data at every turn, improve treatment effeciency.

The process flow diagram of the embodiment two of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 2 provides for the specific embodiment of the invention, as shown in Figure 2, when recommending music to user from musical database, proposed algorithm based on music content and the proposed algorithm based on rule of similarity are combined neatly, gives full play to the strong point of two kinds of algorithms.

This accompanying drawing embodiment comprises:

Step 1011: calculate all music the first rule of similarity transition probability matrix between any two in music libraries based on playlist rule of similarity.Specifically, playlist rule of similarity obtains according to per song redirect frequency in music libraries.For a piece of music in music libraries, other any a piece of music in music libraries may be jumped to, namely need the redirect frequency of trying to achieve between two between music, in order to accurate recommendation, therefore need to add up the redirect frequency between two between music for all user's playlists.

Step 1012: calculate all music the second rule of similarity transition probability matrix between any two in music libraries based on music content.For new music, music is recommended just to there will be deviation based on playlist completely, if but consider the relation between music content, just can solve the cold start-up problem of new music, such as, increase a piece of music in music libraries, the music similar to it can be found by the lyrics, and then by this new music recommend to the user listening to new music similar music with this.

Step 1013: superpose described first rule of similarity transition probability and described second rule of similarity transition probability obtains all music rule of similarity transition probability between any two in music libraries.These two kinds of transition probabilities are combined and obtains music final transition probability between any two, utilize final rule of similarity transition probability accurately can recommend music to user.

See Fig. 2, the present invention recommends by the time series of music in playlist, can reflect that the interest of user changes along with the change of time; Music all in user's playlist all participates in recommending user, and it reduces along with the change of time gradually to the ability that user recommends, and also can embody user and change along with the change of time interest in music; Pass through rule of similarity, by the similar music calculated based on music content with by user's playlist calculate to similar music unite, perfect can apply on the one hand in the algorithm frame of the present invention's proposition, also accomplish the unification of rule-based recommendation and content-based recommendation on the other hand.

The process flow diagram of the embodiment three of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 3 provides for the specific embodiment of the invention, as shown in Figure 3, being neglected by the numerical value being less than predetermined value in first rule of similarity transition probability and the second rule of similarity transition probability, is 0 by this part rule of similarity transition probability.

In this accompanying drawing embodiment, calculate the step of all music the second rule of similarity transition probability matrix between any two in music libraries based on music content after, the method also comprises:

Step 1012-1: ignore the described first rule of similarity transition probability being less than first threshold.The described first rule of similarity transition probability being less than first threshold is set to zero, and first threshold can be arbitrary numerical value of 0.1 ~ 0.2, specifically can need setting according to user.

Step 1012-2: ignore the described second rule of similarity transition probability being less than Second Threshold.The described second rule of similarity transition probability being less than Second Threshold is set to zero, and Second Threshold can be arbitrary numerical value of 0.1 ~ 0.2, and concrete numerical value can need setting according to user.

See Fig. 3, user also exists randomness to a certain degree in the process playing music, so it may not be large especially for having the song correlativity be connected before and after some in user's playlist, even likely style difference is also very large.Due to the existence of this reason, for the music site of large-scale consumer amount, by the historical data of user's playlist calculate based on there is the very weak rule of similarity transition probability of a large amount of intensity in the rule of similarity transition probability of playlist.The existence that the very weak regular transition probability of these intensity is a large amount of, the operational efficiency of system can be had a strong impact on and the storage space of at substantial, by setting threshold value (such as, first threshold and Second Threshold are less than or equal to 0.2) the regular transition probability very weak to intensity prune, high-strength regular transition probability is retained, while improving search efficiency, also improve the accuracy of recommendation.

The process flow diagram of the embodiment four of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 4 provides for the specific embodiment of the invention, as shown in Figure 4, assign weight to respectively the first rule of similarity transition probability and the second rule of similarity transition probability, then sue for peace, obtain the rule of similarity transition probability between two between music in music libraries.

In this accompanying drawing embodiment, superpose described first rule of similarity transition probability and described second rule of similarity transition probability obtains all music rule of similarity transition probability between any two in music libraries, specifically comprise:

Step 10131: described first rule of similarity transition probability is multiplied by the first weights and obtains the first transition probability.The setting of the first weights is main relevant with the quantity of the new music that music libraries increases, if the quantity of new music is little, the first weights can be large, otherwise the first weights can be less, and such as, the first weights can be 0.5.Main only have old user just to have playlist because the first rule of similarity transition probability mainly considers the correlativity of playlist, and the first rule of similarity transition probability accurately can reflect the Music Appreciation tendency of user.

Step 10132: described second rule of similarity transition probability is multiplied by the second weights and obtains the second transition probability.The setting of the second weights is also main relevant with the quantity of the new music that music libraries increases, if the quantity of new music is little, the second weights can be less, otherwise the second weights can be large, and such as, the second weights can be 0.5.Main because the second rule of similarity transition probability mainly considers the correlativity of music content, the cold start-up problem of new music can be solved very well.

Step 10133: described first transition probability and described second transition probability are added and obtain rule of similarity transition probability.Wherein, described first weights and described second weights sum are 1, and usual first weights and the second weights are 0.5.In other specific embodiment of the present invention, along with user listens to the growth of music number, the first weights increase gradually, and the second weights reduce gradually, thus improve the accuracy of recommending further.

The process flow diagram of the embodiment five of a kind of individualized music recommend method based on exponential damping window and rule of similarity that Fig. 5 provides for the specific embodiment of the invention, as shown in Figure 5, the recommendation list of all music in user's playlist is obtained according to rule of similarity transition probability matrix, ask the union of these recommendation list, and music existing in user's playlist is also concentrated eliminating from music, the list of user's music recommend can be obtained.

In this accompanying drawing embodiment, recommend music based on described final rule of similarity transition probability and user's playlist to user, specifically comprise:

Step 1031: the recommendation list obtaining each music in user's playlist based on described rule of similarity transition probability matrix.Per song has one to recommend set.

Step 1032: ask the union of all described recommendation list to obtain recommendation list set.Ask the union of these recommendation list, just can obtain the recommendation list set of user.

Step 1033: remove already present music in user's playlist from described recommendation list set, thus obtain the list of user's music recommend.Because music already present in user's playlist does not need to recommend to user, if therefore there is already present music in user's playlist in recommendation list set, need to get rid of these music again, ensure the accuracy recommended.

The recommended models based on exponential damping window in the present invention, is specifically described as follows:

Exponential damping window is introduced into calculate the level and smooth aggregate-value of in data stream (stream of songs) one, and current data (current song) is all retained with the information in historical data (historical song).Therefore, along with the continuous inflow of data stream, the weight of historical data stream is constantly decayed, and song occurs more early in stream of songs, and its weight is also less, also less on the impact of result.As shown in Figure 6, the time dependent curve of the weight for each song, in figure, W _i=(1-c) ^l-i, the W wherein in formula _ifor exponential damping window weight; C is decay factor; L is the total quantity of music in user's playlist; I is the order of current music in user's playlist, and i is more than or equal to the integer that 1 is less than or equal to L.Wherein, stream of songs is made to be { e ₁, e ₂..., e ₃, wherein e ₁be first song play, e _nfor the song of current broadcasting.C is decay factor, then the exponential damping window (exponentiallydecayingwindow) of stream of songs is defined as:

$\underset{i=0}{\overset{n-1}{\Σ}}{e}_{n-i}{(1-c)}^i$

Although all take part in user-customized recommended based on all songs in the proposed algorithm playlist of figure, but because each song having resource in playlist all to be participated in the distribution resource with equal status, so can not reflect the change that the interest of user occurs along with the change of time.Markov model has time series, can reflect user interest over time.When but this algorithm model is recommended, only recommend according to last song, just equal to have ignored other all songs in user's playlist except last song to the impact of the song that user will listen to future, a large amount of information will inevitably be caused like this to be left in the basket.Therefore, by improving Markov chain, the a part of information in the historical record of playlist is retained by exponential damping window, then temporally sequence gives each song with different resource allocations, distributes the result obtained carry out personalized recommendation to user by all songs.

As shown in Figure 7, a rule of similarity set can all be obtained according to each song of user's playlist.This rule of similarity set is exactly song t _ithe rule of similarity set of (1≤i≤m).Song in these rule of similarity set is all the song that there is certain incidence relation with song in user's playlist.Therefore, by association index exponential damping window model, the recommendation results of last recommendation results with songs all in playlist can be combined, user often listens to a song, and its history recommendation results weight just becomes original (1-c) doubly.

Based on the rule of similarity generation method introduction of playlist in the present invention one specific embodiment:

To a certain user u _k, the sequence of songs Playlist listened to _k{ t ₁, t ₂..., t _m, u _kthat listens to often listens a song t _xthere is a rule of similarity set the recommendation set of the last item song that then user listens to is user u _kthe set recommended be:

$\begin{array}{cc}{\mathrm{Next}}_{u_k}=\underset{i=0}{\overset{n}{\∪}}({\mathrm{Next}}_{t_{m-i}}-{\mathrm{Playlist}}_k),where& n\≤m\end{array}$

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; Set can be recommended get rid of the music in user's playlist.

For the song t of in user's playlist _i, need to calculate the song collection of weight and be:

${\mathrm{Compute}}_{t_i}={\mathrm{Next}}_{t_i}\∩{\mathrm{Next}}_{u_k}$

Wherein, u _kfor user; for song t _ithe set recommended.

Here, the transition probability of association index exponential damping window and Markov model, user's playlist historical record, definition user playlist playlist _kin the i-th song t _iby Resourse Distribute to song t _jcapacity of water be following formula:

Wherein, t _ifor the song of in user's playlist; M is total number of song in user's playlist; Obtain user u _kset to be recommended in all songs Resources allocation ability after, need to calculate set to be recommended in each song Resources allocation to the probability of particular songs (i.e. rule of similarity transition probability):

$g\_p_{t_i\→t_j}=\frac{\underset{i=1}{\overset{m}{\Σ}}g\_{\mathrm{frequency}}_{t_i\→t_j}}{\underset{t_x\∈{\mathrm{Next}}_{t_i}}{\Σ}g\_{\mathrm{frequency}}_{t_i\→t_x}}$

Wherein, represent the i-th song t in user's playlist _iby Resourse Distribute to jth song t _jability; represent the i-th song t in user's playlist _iby Resourse Distribute to xth song t _xability; M is total number of music in user's playlist; I is positive integer.

Based on the rule of similarity generation method introduction of song content in the present invention:

The content of song can be divided into: the lyrics, song audio files, song metadata (comprising singer, song classification etc.) etc., these song contents can as the foundation of Similarity measures between song.

First, for the lyrics, need to proceed as follows successively: participle, stem extract (English), remove stop words, feature selecting, characteristic weighing and Similarity Measure.The high efficiency method that extensive text set carries out feature selecting is document frequency (DF).DF is by setting threshold value, and make the word lower than this threshold value be considered to low frequency words, the word higher than this threshold value is considered to high frequency vocabulary.Low frequency words is removed from feature space, and then reduces the dimension of feature space.Conventional characteristic weighing algorithm is TF-IDF algorithm, and its formula is as follows:

$TF-IDF(d,w)=TF(d,w)\×IDF(d,w)=TF(d,w)\×\log \left(\frac{\left|D\right|}{DF\left(w\right)}\right)$

Wherein, TF is word frequency in the text, | D| is text sum, and DF is the frequency that in all texts, certain word occurs.So just can obtain the vector space of the word obtained according to the lyrics.

Song files content commonly uses each frame proper vector that Mel frequency cepstral coefficient (MFCC) extracts song audio, the characteristic vector space of composition song audio.

Method according to the similarity between characteristic vector space calculating song comprises usually: Euclidean distance, cos similarity, minihash method, SimRank etc.Wherein cos similarity is considered to one of result of calculation the best way.For a proper vector t=(v ₁, v ₂..., v _n), cos similarity formula is specific as follows:

$cos(t_i,t_j)=\frac{{\mathrm{\Σ}}_{k=1}^n{v}_k\left(t_i\right)\×v_k\left(t_j\right)}{\sqrt{\left({\mathrm{\Σ}}_{k=1}^n{v}_k^2\right(t_i\left)\right)\×\left({\mathrm{\Σ}}_{k=1}^n{v}_k^2\right(t_j\left)\right)}}$

Then to song t _iall similar songs be normalized, obtain song in the user's playlist based on song content like regular transition probability (namely second like regular transition probability):

${\mathrm{sim}}_{t_i\→t_j}=\frac{cos(t_i,t_j)}{{\mathrm{\Σ}}_{k=1}^{n}cos(t_i,t_k)}$

Wherein, n is all with the i-th song t in music libraries _ithe number of similar music; K is positive integer.

In actual application of the present invention, user also exists randomness to a certain degree in the process of played songs, so having the song be connected before and after some in user's playlist may correlativity be not large especially, even likely style difference is also very large.Due to the existence of this reason, for the song website of large-scale consumer amount, by the historical data of user's playlist calculate based on there is the very weak rule of a large amount of intensity in the rule of similarity of playlist.A large amount of existence of the very weak rule of these intensity, can have a strong impact on the operational efficiency of system and the storage space of at substantial.Consider in a system there is a large amount of cold song or new song, if use support to prune, the rule trimming that this part song generates will be fallen.So we propose to be pruned weak rule by degree of confidence, make strong rule be retained.Make support support (t _i→ t _j) for obtain in training set from song t _ibrowse to song t _jfrequency, that is:

$Support(t_i\→t_j)={\mathrm{frequency}}_{t_i\→t_j}$

That is:

$Support\left(t_i\right)=\underset{t_x\∈{\mathrm{Next}}_{t_i}}{\Σ}Support(t_i\→t_x)$

Regard song every in user record as a node, then degree of confidence confidence (t _i→ t _j) be that in all user's playlists records, father node is t _iwhen, child node is t _jnumber percent, namely father node is t _iconditional probability under condition, its computing formula is as follows:

$Confidence(N_i\→N_j)=\frac{Support(N_i\→N_j)}{Support\left(N_i\right)}$

In like manner, when calculating content-based rule of similarity, also there will be the rule that a large amount of regular weights are relatively very little, if these rule trimmings, not only consume a large amount of storage spaces, and have a strong impact on search efficiency when algorithm performs, so identical method also should be used to prune.

The final rule of similarity of the present invention is that the rule of similarity generated based on playlist superposes with the final of rule of similarity generated based on song content.Its computing formula is as follows:

$R_{t_i\→t_j}=w\·g\_p_{t_i\→t_j}+(1-w)\·{\mathrm{sim}}_{t_i\→t_j},(0\≤w\≤1)$

Here it should be noted that due to with the result obtained after being all through normalization, so its similarity is irreversible, i.e. sim (t _i, t _j) and sim (t _j, t _i) result be unequal, in like manner with result be also unequal.

First the present invention will generate the ordering rule pair of redirect mutually between song (music) according to the broadcasting record of user, as shown in Figure 8, the generation step flow process of rule of similarity is specific as follows for the flow process of concrete generation method:

The first step: the frequency of adding up the music appearance that in all playlists, temporally sequence is adjacent, before namely in playlist, a piece of music is father node, and a follow-up node is child node, and this sequence node is irreversible;

Second step: according to the hop frequencies between the frequency computing node of the first step, this frequency is the rule of similarity based on playlist;

3rd step: prune the rule of similarity that second step obtains according to degree of confidence, removes weak rule;

4th step: extract MFCC proper vector according to audio file or extract weighted feature term vector according to lyrics file, generating the proper vector of audio frequency;

5th step: calculate the COS similarity between audio frequency according to the proper vector of audio frequency;

6th step: the similar audio set of the some audio frequency calculated is normalized, so just obtains content-based rule of similarity;

7th step: the rule of similarity that the 3rd step and the 6th step generate is weighted, obtains final rule of similarity.

The overall flow figure of algorithm of the present invention as shown in Figure 9, is algorithm frame overall flow below:

The first step: find out all music similar with the music in user list, composition can recommend collection of music;

Second step: class recommends the music initial weight in collection of music to be all 0, then from user's playlist First music, from the rule of similarity set of this song, find all similar music, and rule of similarity weights are added to the music can recommended in collection of music; Then start to search the second song in playlist, first the rule of similarity weights of music all in playlist are multiplied by (1-c), then the rule of similarity weights of music in the rule of similarity set of the second song are added to the music can recommended in collection of music.Music all in playlist afterwards all repeats the operation of the second song, until last song;

3rd step: last for the recommended music peace in second the cumulative weights obtained is sorted;

4th step: according to real needs, the topN removed wherein forms recommendation results.

A specific embodiment of the present invention:

(1) instance analysis of rule of similarity generation

Example: the input of algorithm comprises music playlist and the audio content data of user.As table 1 be 6 users playlists of music that oneself is selected on certain music site to example:

Table 1. user playlist

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

This example calculation step is as follows:

(1) music in the playlist of user to be separated in order between two, such as playlist playList _u1{ t ₁, t ₃, t ₅r (t can be partitioned into ₁, t ₃, 1) and R (t ₃, t ₅, 1) and (R (t _i, t _j, t n) _irepresent father node, t _jrepresent child node, n represents redirect frequency between father node and child node).The playlist of all users in statistical form, namely obtains redirect frequency chart between music, as shown in Figure 10, such as, and t ₁jump to t ₅number of times be 1, t ₁jump to t ₄number of times be 2, t ₁jump to t ₃number of times be 2, t ₁jump to t ₂number of times be 0, t ₁jump to t ₁number of times be 0.

(2) I={t is made ₁, t ₂, t ₃, t ₄, t ₅be all song collection, then based on the weight computing result of the music rule of similarity of playlist as shown in following display:

$g\_p(I^T\×I)={(g\_p(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0& 0.4& 0.4& 0.2\\ 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1\\ 0.5& 0& 0.5& 0& 0\\ 0.25& 0.75& 0& 0& 0\end{array}\right\}$

Wherein, t ₁the total degree jumping to other music is 5, for above-mentioned display the first row, uses t respectively ₁jump to t ₁, t ₂, t ₃, t ₄, t ₅number of times be 0,0,2,2,1 divided by total degree 5, other row.

(3) at this, make degree of confidence be 0.2, then, after pruning rule, the rule of similarity based on playlist is:

$g\_p(I^T\×I)={(g\_p(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0& 0.4& 0.4& 0\\ 1& 0& 0& 0& 0\\ 0& 0& 0& 0& 1\\ 0.5& 0& 0.5& 0& 0\\ 0.25& 0.75& 0& 0& 0\end{array}\right\}$

Here, assuming that arrived the proper vector of every song by content obtaining, its each proper vector weight is as shown in table 2 below:

Table 2. song content proper vector value

(4) according to the combination of eigenvectors in table 2, can calculate audio frequency cos similarity between any two, the concrete outcome calculated is as shown in following matrix:

$\cos (I^T\×I)={\left(\cos \right(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0.019& 0.394& 0.418& 0.137\\ 0.019& 0& 0.09& 0.02& 0.137\\ 0.394& 0.09& 0& 0.377& 0.096\\ 0.418& 0.02& 0.377& 0& 0.209\\ 0.137& 0.137& 0.096& 0.209& 0\end{array}\right\}$

(5) then to be normalized cos similarity, obtain the rule of similarity weights of content-based rule of similarity afterwards, as follows after normalization:

$\mathrm{si}m(I^T\×I)={\left(sim\right(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0.02& 0.407& 0.432& 0.141\\ 0.074& 0& 0.337& 0.073& 0.514\\ 0.412& 0.094& 0& 0.394& 0.1\\ 0.409& 0.019& 0.368& 0& 0.204\\ 0.237& 0.237& 0.166& 0.36& 0\end{array}\right\}$

(6) prune rule of similarity, delete the rule that rule of similarity weights are less than 0.2, the weight results obtaining content-based rule of similarity is:

$sim(I^T\×I)={\left(sim\right(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0& 0.407& 0.432& 0\\ 0& 0& 0.337& 0& 0.514\\ 0.412& 0& 0& 0.394& 0\\ 0.409& 0& 0.368& 0& 0.204\\ 0.237& 0.237& 0& 0.36& 0\end{array}\right\}$

(7) make weight w=0.5 of two kinds of rule of similarity superpositions, then the weights generating final rule of similarity are as follows:

$R(I^T\×I)={\left(R\right(t_i\→t_j\left)\right)}_{5\×5}=\left\{\begin{array}{ccccc}0& 0& 0.404& 0.416& 0\\ 0.537& 0& 0.268& 0& 0.207\\ 0.206& 0& 0& 0.197& 0.55\\ 0.454& 0& 0.434& 0& 0.102\\ 0.243& 0.493& 0& 0.18& 0\end{array}\right\}$

(2) personalized recommendation

Assuming that decay factor c=0.6, and to user U ₇carry out personalized recommendation, recommendation results is 1.Concrete calculation procedure is as follows:

(1) determine to recommend set.The playlist of this user is t ₅recommendation set be t ₁recommendation set be set then can be recommended to be

(2) by first song t in playlist ₅the weights of rule of similarity be assigned to the song in gathering can be recommended, namely on carrying out once calculate before, should first by the weights of middle song are multiplied by (1-c)=0.4, namely then by next song t ₁rule of similarity weights compose be added to in song, namely ${\mathrm{Next}}_{u_6}\{t_2=0.176,t_3=0.404,t_4=0.488\}.$

(3) right weights according to calculating sort ${\mathrm{Next}}_{u_6}\{t_4=0.488,t_3=0.404,t_2=0.176\}.$

(4) when recommendation results is one, by the t of maximum weight ₄recommend user.

The present invention also at least has following beneficial effect:

1. the present invention recommends by the time series of music in playlist, can reflect that user interest changes along with the change of time.

2. in user's playlist of the present invention, all music all participate in recommending user, and it reduces along with the change of time gradually to the ability that user recommends, and also can embody the change of user to interest in music.

3. the present invention adopts exponential damping window to carry out recording musical and recommends ability over time, only preserves the final weights that exponential damping window calculation obtains like this, so just need not recommend all to recalculate historical data at every turn in the process calculated.

4. the present invention passes through rule of similarity, by the content-based similar music calculated with by user's playlist calculate to similar music unite, perfect can apply on the one hand in the algorithm frame of the present invention's proposition, also accomplish the unification between rule-based recommendation and content-based recommendation on the other hand.

5. the present invention is that content-based and rule-based mixing is recommended, and can keep solving when higher recommendation precision the cold start-up problem that new projects bring.

6. the present invention is by pruning orderly rule of similarity, time when drastically reduce the area algorithm at actual motion when loss very little accuracy rate.

7. the present invention can be applicable to the personalized recommendation environment of all user behaviors and content, only needs the method that the present invention just can be used to mention after considering separately content-based similarity, uses comparatively flexible.

8. the present invention is not when calculating content-based rule of similarity, and can obtain, than based on the proposed algorithm of figure and the higher accuracy rate of Hidden Markov Model (HMM) and recall rate, when calculating content-based rule of similarity, can have and well recommend performance.

9. the present invention uses exponential damping window, music recommend ability in playlist is changed in time and reduces gradually.

10. the present invention is by the foundation of orderly rule of similarity as the similar recommendation of per song.

Orderly music adjacent in playlist is considered as being correlated with by 11. the present invention, and user selects playlist to be namely at selection similar music, so calculate the rule of similarity based on playlist by playlist.

Content-based similar music is normalized by 12. the present invention, is considered as content-based rule of similarity.

13. the present invention realize mixing by the superposition of content-based and rule-based rule of similarity weight and recommend.

14. the present invention calculate in the process of rule of similarity and prune content-based and based on playlist rule of similarity respectively, recommend the similar music that in set, weight is low, and then improve the search efficiency of collection of programs to reduce per song.

15. the present invention calculate rule of similarity weight by the transition probability between music in playlist.

In 16 the present invention, rule of similarity is the superposition based on weight between the rule of similarity play and content-based rule of similarity, realizes content-based and rule-based mixing and recommends.

17. the present invention, when calculating content-based and rule-based rule of similarity weight respectively, prune rule, save storage space, improve and recommend efficiency.

The invention provides a kind of individualized music based on exponential damping window and recommend sort method, when music recommend, content-based recommendation algorithm and rule-based proposed algorithm are combined neatly, obtain all music rule of similarity transition probability matrix between any two in music libraries, as much as possible play the advantage of two kinds of algorithms; Obtain user according to all music in playlist and recommend transition probability, can ensure that higher individualized music recommends precision.

The above-mentioned embodiment of the present invention can be implemented in various hardware, Software Coding or both combinations.Such as, embodiments of the invention also can be the program code of the execution said procedure performed in data signal processor (DigitalSignalProcessor, DSP).The present invention also can relate to the several functions that computer processor, digital signal processor, microprocessor or field programmable gate array (FieldProgrammableGateArray, FPGA) perform.Can configure above-mentioned processor according to the present invention and perform particular task, it has been come by the machine-readable software code or firmware code performing the ad hoc approach defining the present invention's announcement.Software code or firmware code can be developed into different program languages and different forms or form.Also can in order to different target platform composing software codes.But the different code pattern of the software code of executing the task according to the present invention and other types configuration code, type and language do not depart from spirit of the present invention and scope.

The foregoing is only the schematic embodiment of the present invention, under the prerequisite not departing from design of the present invention and principle, the equivalent variations that any those skilled in the art makes and amendment, all should belong to the scope of protection of the invention.

Claims (10)

1. the individualized music based on exponential damping window recommends a sort method, and it is characterized in that, the method comprises:

Obtain all music rule of similarity transition probability matrix between any two in music libraries;

Give computing after different weight to the rule of similarity transition probability matrix corresponding line data that user's playlist is corresponding, obtain a line user and recommend transition probability;

Transition probability is recommended to obtain the transition probability of respective user music recommend list according to described user; And

According to described transition probability, the music in the list of user's music recommend is sorted.

2. recommend sort method based on the individualized music of exponential damping window as claimed in claim 1, it is characterized in that, described weight is exponential damping window weights W _i, described exponential damping window weights W _icomputing formula be:

W _i＝(1-c) ^L-i

Wherein, W _ifor exponential damping window weight; C is decay factor; L is the total quantity of music in user's playlist; I is the order of current music in user's playlist, and i is more than or equal to the integer that 1 is less than or equal to L.

3. recommend sort method based on the individualized music of exponential damping window as claimed in claim 2, it is characterized in that, described user recommends transition probability computing formula NEXT to be specially:

$NEXT=\underset{i=1}{\overset{L}{\Σ}}{e}_i{W}_i$

Wherein, e _ifor the rule of similarity transition probability matrix corresponding line data that user's playlist is corresponding.

4. recommend sort method based on the individualized music of exponential damping window as claimed in claim 2, it is characterized in that, described decay factor is 0.6.

5. recommend sort method based on the individualized music of exponential damping window as claimed in claim 1, it is characterized in that, the step obtaining all music rule of similarity transition probability matrix between any two in music libraries specifically comprises:

All music the first rule of similarity transition probability matrix between any two in music libraries is calculated based on playlist rule of similarity;

All music the second rule of similarity transition probability matrix between any two in music libraries is calculated based on music content; And

Superpose described first rule of similarity transition probability and described second rule of similarity transition probability obtains all music rule of similarity transition probability between any two in music libraries.

6. recommend sort method based on the individualized music of exponential damping window as claimed in claim 5, it is characterized in that, calculate the step of all music the second rule of similarity transition probability matrix between any two in music libraries based on music content after, also comprise:

Ignore the described first rule of similarity transition probability being less than first threshold; And

Ignore the described second rule of similarity transition probability being less than Second Threshold.

7. recommend sort method based on the individualized music of exponential damping window as claimed in claim 6, it is characterized in that, described first threshold and described Second Threshold are less than or equal to 0.2.

8. recommend sort method based on the individualized music of exponential damping window as claimed in claim 5, it is characterized in that, superpose described first rule of similarity transition probability and described second rule of similarity transition probability obtains all music rule of similarity transition probability between any two in music libraries, specifically comprise:

Described first rule of similarity transition probability is multiplied by the first weights and obtains the first transition probability;

Described second rule of similarity transition probability is multiplied by the second weights and obtains the second transition probability; And

Described first transition probability and described second transition probability are added and obtain rule of similarity transition probability,

Wherein, described first weights and described second weights sum are 1.

9. recommend sort method based on the individualized music of exponential damping window as claimed in claim 8, it is characterized in that, described first weights increase along with increasing of music in user's playlist.

10. recommend sort method based on the individualized music of exponential damping window as claimed in claim 1, it is characterized in that, the list of described user's music recommend is obtained by following steps:

The recommendation list of per song in user's playlist is obtained based on described rule of similarity transition probability matrix;

The union of all described recommendation list is asked to obtain recommendation list set; And

From described recommendation list set, remove already present music in user's playlist, thus obtain the list of user's music recommend.