CN103488705B - The user interest model increment updating method of personalized recommendation system - Google Patents

Abstract

The invention discloses the user interest model increment updating method of a kind of personalized recommendation system, the method basic thought is the intermediate object program that storage generates in the calculating process of current user interest model, when updating user interest model, this intermediate object program basis carries out incremental computations。The present invention, under guaranteeing the premise that renewal process does not lose interest information, can meet the requirement that user interest model also can constantly quickly update when data volume is huge, improves commending system performance, provides the user higher-quality service。

Description

The user interest model increment updating method of personalized recommendation system

Technical field

The present invention relates to Computer Applied Technology field, particularly the user interest model increment updating method of a kind of personalized recommendation system。

Background technology

Personalized recommendation system is by setting up the binary crelation between user and recommended, existing selection course or similarity relationships is utilized to excavate the potential object interested of each user, and then carry out personalized recommendation (Liu Jianguo, Zhou Tao, Wang Binghong. the progress [J] of personalized recommendation system. natural science is in progress, 2009,19 (1), 1-15.)。Along with the variation of user's request, personalized recommendation system application becomes more extensive, is applied not only to ecommerce, is also used for recommending webpage, document etc.。Need often to consult substantial amounts of data document for copy writer and researcher。The document of interest content read by collecting and analyze user based on the personalized recommendation system of document content information is understood the reading interest of user and sets up user interest model, by comparing the matching degree of document content and user interest model, recommend, to user, the document that matching degree is high。Personalized recommendation system based on document content information has three important modules: user interest MBM, recommended MBM, proposed algorithm module, this system model is as shown in Figure 1。

Based in the commending system of document content information, user interest MBM is the module of one of them core, and its effect is that extraction user interest model the change according to user interest realize interest model renewal from the document interested that user read。For realizing high-precision recommendation, user interest model allows for the current interest of accurate description user, and the renewal of interest model allows for quickly following the tracks of the change of user interest。

The renewal of current user interest model mainly has two kinds of methods, time window method and forgetting function method, time window method is to utilize time slip-window to filter out-of-date interest, forgetting function method is to utilize the weight (Fei Hongxiao of forgetting function decay interest, Dai Yi, solemn etc. based on the user interest drift method [J] optimizing time window. computer engineering, 2008,34 (16), 210-214.)。Document (SHINH., CHOS..NeighborhoodPropertyBasedPatternSelectionforSuppor tVectorMachines [J] .NeuralComputation, 2007,19 (3), 816-855.) middle employing time window method renewal user interest model, the method utilizes time slip-window to filter out-of-date interest。Document (KEERTHIS.S., SHEVADES.K., BHATTACHARYYA., etal.AFastIterativeNearestPointAlgorithmforSupportVector MachineClassifierDesign [J] .IEEETransactionsonNeuralNetworks, 2000,11 (1), 124-136.) middle employing forgetting function method renewal user interest model, the method utilizes the weight of forgetting function decay interest。Single Rong (single Rong. the renewal of user interest model and Forgetting Mechanism research [J]. microcomputer is applied, 2011,27 (7), 10-11,69) updating interest model according to the surfing of the feature of html document and user, the weight in conjunction with forgetting factor correction Feature Words carrys out forgeing of implementation model。Document (Li Feng, Pei Jun, the ocean of trip. based on the adaptive user interest model [J] of implicit feedback. computer engineering and application, 2008,44 (9), 76-79.) user interest is divided into short-term interest and Long-term Interest, short-term interest adopts time window update mechanism, and Long-term Interest adopts the more New Policy of time-based forgetting function。

How existing user interest model update method it is emphasised that reject the document of deviation user interest in the middle of user's document interested, and increase new document of interest, make the document for building user interest model more can reflect user's current interest, and have ignored the computational efficiency problem that user interest model updates。Along with the increase of user's reading documents quantity, the number of documents interested of its labelling also can increase, and the computational efficiency problem that user interest model updates highlights gradually, causes model modification speed too low and can not meet the adverse consequences of user's request。

Summary of the invention

The technical problem to be solved is, not enough for prior art, the user interest model increment updating method of a kind of personalized recommendation system is provided, under guaranteeing the premise that renewal process does not lose interest information, improve the computational efficiency that user interest model updates, meet the requirement that user interest model also can constantly quickly update when data volume is huge, improve personalized recommendation system performance, provide the user higher-quality service。

For solving above-mentioned technical problem, the technical solution adopted in the present invention is: the user interest model increment updating method of a kind of personalized recommendation system, and the method is:

1) the user interest vector space model U based on document content is built₀；

2) described user interest vector space model U is set up₀User document of interest collection D₀={ d₀₁,d₀₂,...,d_0m, make D={d₁,d₂,...,d_nFor document sets to be recommended, wherein document d_iCharacteristic vector beWherein, d_0eRepresent described user document of interest collection D₀In document, e=1,2 ..., m, m is described user document of interest collection D₀In total number of documents；T_ikRepresent document d_iKth item Feature Words；W_ikRepresent document d_iThe weight of kth item Feature Words；I=1,2 ..., n；K=1,2 ..., a；A represents document d_iThe total item of Feature Words；Here, document sets to be recommended is generally collected from network and is obtained or obtain from documents and materials；

3), when recommending document, calculate that all file characteristics in described document sets D to be recommended are vectorial and described user interest vector space model U₀Similarity r, it is recommended that going out the similarity r document more than threshold alpha, and feed back new document interested, described new collection of document isThe span of threshold alpha is between 0 to 1, needs to regulate α size according to user, and when user intentionally gets more recommendation results, the value of α is closer to 0, and when user intentionally gets recommendation results more accurately, the value of α is closer to 1；When selecting out-of-date in user's document of interest set or document that is that deviate user interest, set of computations D respectively₀In each file characteristics is vectorial and described user interest vector space model U₀Similarity r', select r' less than the document of threshold alpha as out-of-date or deviation user interest document, described out-of-date or deviation user interest collection of document be It is the document in D' for described new collection of document, f=1,2 ..., q, q is the total number of documents in described new collection of document D'；For described out-of-date or deviation user interest collection of document D " in document, h=1,2 ..., c, c be described out-of-date or deviation user interest collection of document D " in total number of documents；

4), to when increasing user's document of interest set, described new collection of document D' is added described user document of interest collection D₀In, constitute new first user document of interest collection D₁；When rejecting the document of out-of-date in user's document of interest set or deviation user interest, by described out-of-date or deviation user interest collection of document D " from described user document of interest collection D₀Middle rejecting, constitutes the second new user document of interest collection D₂；

5) described new first user document of interest collection D is calculated according to following formula₁Center vector

$\stackrel{\‾}{W_{D_1}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q}=\frac{m\stackrel{\‾}{W_{D_0}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q};$

Wherein,For described user document of interest collection D₀In the characteristic vector of e document；For the characteristic vector of f document in described new collection of document D'；Q is the total number of documents in described new collection of document D'；For described user document of interest collection D₀Center vector；M is described user document of interest collection D₀In total number of documents；E=1,2 ..., m；F=1,2 ..., q；

The second new user document of interest collection D is calculated according to following formula₂Center vector

$\stackrel{\‾}{W_{D_2}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c}=\frac{m\stackrel{\‾}{W_{D_0}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c};$

Wherein,For described user document of interest collection D₀In the characteristic vector of h document；For out-of-date or deviation user interest collection of document D " in the characteristic vector of document；C be out-of-date or deviation user interest collection of document D " in total number of documents；For described user document of interest collection D₀Center vector；M is described user document of interest collection D₀Middle total number of documents；H=1,2 ..., c；

6) willOrEach dimension sorts from big to small by weights, selectsOrFront N dimension build new user interest vector space model U₁Or U₂, handle simultaneouslyOrIt is stored in personalized recommendation system；Wherein, N less thanOrDimension；With described new user interest vector space model U₁Or U₂Replace step 1) in U₀Carry out new round recommendation。

Described step 1) in, build the user interest vector space model U based on document content₀Specifically comprise the following steps that

1) document that all users are interested is carried out Feature Words selection and term weight function calculates；File characteristics word selection and term weight function can by the keyword extraction gain-of-functions of ICTCLAS Chinese word segmenting software (http://ictclas.nlpir.org/), or the Feature Words system of selection based on word frequency obtains；

2) extract the characteristic vector of all users document interested, constitute file characteristics vector set D₃；

3) described file characteristics vector set D is calculated₃Center vector, by described file characteristics vector set D₃Center vector sort from big to small by the weight of each dimension, choose front M tie up as user interest vector space model U₀；Wherein M is less than described file characteristics vector set D₃The dimension of center vector。

File characteristics vector set D₃={ d₃₁,d₃₂,...,d_3xCenter vectorComputing formula be:

$\stackrel{\‾}{W_{D_3}}=\frac{\underset{y=1}{\overset{x}{\Σ}}{W}_{d_{3y}}}{x};$

Wherein, x is described file characteristics vector set D₃The number of middle element；For described file characteristics vector set D₃The characteristic vector of middle y-th document；Y=1,2 ..., x。

Document d in document sets D to be recommended_iCharacteristic vector and described user interest vector space model U₀The computing formula of similarity r be:

$r=cos(W_{d_i},U_0)=\frac{W_{d_i}\·U_0}{\left|\right|W_{d_i}|{|}_2\×|\left|U_0\right|{|}_2};$ Wherein, | | | |₂Represent two norms。

User document of interest collection D₀In the e file characteristics be vectorial and described user interest vector space model U₀The computing formula of similarity r' be:

$r^{\′}=cos(W_{d_{0e}},U_0)=\frac{W_{d_{0e}}\·U_0}{\left|\right|W_{d_{0e}}|{|}_2\×|\left|U_0\right|{|}_2}.$

The basic thought of the user interest model increment updating method that the present invention proposes is the intermediate object program that storage generates in the calculating process of current user interest model, when updating user interest model, carries out incremental computations on this intermediate object program basis。

Compared with prior art, the present invention is had the beneficial effect that to the present invention is directed to the efficiency that the user interest model of the commending system based on document content information updates, under ensureing the premise that user profile is complete, the update method of the present invention decreases amount of calculation when user interest model updates, make the user interest model can quick frequent updating, improve the performance of personalized recommendation system, can quickly realize User interest tracking, to adapt to the change of user interest, provide the user higher-quality service。

Accompanying drawing explanation

Fig. 1 is the commending system based on document content information；

Fig. 2 is the structure flow process of user interest model of the present invention。

Detailed description of the invention

The present invention builds the flow process of the user interest vector space model based on document content as shown in Figure 1, first document interested in user carries out Feature Words selection and term weight function calculates, and obtains a file characteristics vector being made up of a stack features word and weight thereof。File characteristics vector extracting method can utilize the Feature Words abstraction function of ICTCLAS Chinese word segmenting software (http://ictclas.nlpir.org/), or the Feature Words system of selection based on word frequency obtains。Multiple file characteristics vectors constitute file characteristics vector set。After calculating the center vector obtaining file characteristics vector set, center vector is respectively tieed up and sorts from big to small by weight, choose the front N dimension interest model vector as this user。

The center vector computational methods of file characteristics vector set are as follows:

Collection of document D₃={ d₃₁,d₃₂,...,d_3x, document d_2iCharacteristic vector beWherein, t_3ikRepresent document d_3iKth item Feature Words, w_3ikRepresent document d_3iThe weight of kth item Feature Words, then center vectorComputing formula is:

$\stackrel{\‾}{W_{D_3}}=\frac{\underset{y=1}{\overset{x}{\Σ}}{W}_{d_{3y}}}{x}---\left(1\right)$

In this formula, file characteristics vector is sued for peace by mating often one-dimensional Feature Words, and the identical then corresponding weights of Feature Words are added。This center vector is respectively tieed up the front M item after by weight sequencing and is interest model U, the M dimension less than center vector of this user, is generally determined by training sample empirical value。

Assume that user's document of interest is { d₁,d₂,d₃, set up the process of user interest model in Table 1。

Table 1 user interest model sets up process

Center vector in tableCalculated gained by formula (1), select front 5 characteristic items of this center vector as user interest model U herein。

The increment updating method that the present invention proposes to implement step as follows:

If U₀For the user interest model that user currently has built up, the user's document of interest setting up this user interest model integrates as D₀={ d₀₁,d₀₂,...,d_0m}。Collection of document D={d₁,d₂,...,d_nFor document to be recommended, document d_iCharacteristic vector be $W_{d_i}=\{(t_{i1},w_{i1}),(t_{i2},w_{i2}),...,(t_{\mathrm{ia}},w_{\mathrm{ia}})\}.$

(1) when recommending document, by all file characteristics vector in cosine angle formulae set of computations D and user model U₀Similarity r, it is recommended that going out the similarity r document more than threshold alpha, user browses the new document that backward system feedback is interested, if the document set isWhen selecting out-of-date in user's document of interest set or document that is that deviate user interest, set of computations D respectively₀In each file characteristics is vectorial and described user interest vector space model U₀Similarity r', select r' less than the document of threshold alpha as out-of-date or deviation user interest document, described out-of-date or deviation user interest collection of document be $D^{\′\′}=\{d_{b_1},d_{b_2},...,d_{b_c}\};$

(2), to when increasing user's document of interest set, described new collection of document D' is added described user document of interest collection D₀In, constitute new user document of interest collection D₁；When rejecting the document of out-of-date in user's document of interest set or deviation user interest, by described out-of-date or deviation user interest collection of document D " from described user document of interest collection D₀Middle rejecting, constitutes new user document of interest collection D₂；

(3) for complete reservation user interest, it is to avoid double counting, improving algorithm performance, system has prestored calculating user interest model U₀Time collection of document D₀Center vectorFormula (1) is deformed into formula (2) and calculates the center vector of the new interest model after increasing new document:

$\stackrel{\‾}{W_{D_1}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q}=\frac{m\stackrel{\‾}{W_{D_0}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q}---\left(2\right)$

Formula (2) is deformed into formula (3) and calculates the center vector of the new interest model after rejecting out-of-date or the document of deviation user interest:

$\stackrel{\‾}{W_{D_2}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c}=\frac{m\stackrel{\‾}{W_{D_0}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c}---\left(3\right)$

(4) willEach dimension sorts from big to small by weights, and before selecting, N dimension builds new user interest model U₁(U₂), handle simultaneouslyIt is stored in system。With the new user interest model U obtained₁(U₂) replace the U in step (1)₀Carry out new stage recommendation。

From formula (2) and formula (3) it can be seen that center vectorAppear in the two formula。Center vectorBeing the front intermediate object program once calculating user interest model, the core of the present invention all preserves this center vector when updating user interest model exactly every timeMake to recalculate this partial content when updating, thus improving renewal efficiency next time。

Example in table 2, the user interest model described in his-and-hers watches 2 is increasing document d₄During renewal, if d₄={ { { insurance, 3.6}, { domestic, 2.5}, { amount of increase, 2.0}}, at center vector for automobile, 4.0}Basis on update, for Feature Words " automobile ", its weight w₁Calculate as shown in formula (4),

$w_1=\frac{3.2*3+4.0}{3+1}=3.4---\left(4\right)$

Reject document d₁During renewal, for Feature Words " automobile ", its weight w₂Calculate as shown in formula (5),

$w_2=\frac{3.2*3-5.3}{3-1}=2.15---\left(5\right)$

Obtain new user interest model center vector by that analogy, update result in Table 2。This example only carries out incremental computations on the basis that user's document of interest number is 3, so computational efficiency improves and inconspicuous in this example。This example is only used for Incremental Updating Algorithm is described。In practical application, the document of interest quantity of user's labelling can compare many, and the number of files of increase or proposition is relatively fewer, and at this time the efficiency of Incremental Updating Algorithm can become apparent from。

Table 2 user interest model renewal process

User interest model in contrast table 1 extract and and table 2 in the incremental update process that proposes of the present invention, it is found that center vectorAs an intermediate object program in last user interest model establishment or renewal process, the present invention carries out incremental update on the basis of this intermediate object program, thus avoiding substantial amounts of vector summation work；And it is seen that the center vector that the increment updating method that the present invention proposes obtains is identical with what directly extract from the collection of document after renewal。In general, the document for building new user interest model has two parts to constitute, and Part I is the document interested newly increased；Part II is to reject remaining part after the document of deviation current user interest in original document of interest, and this partial document quantity accounts for the overwhelming majority。The meaning of the incremental update mode that the present invention proposes in avoiding the double counting work of Part II document, thus effectively reducing user interest model to update amount of calculation。

Claims (5)

1. the user interest model increment updating method of a personalized recommendation system, it is characterised in that the method is:

1) the user interest vector space model U based on document content is built₀；

2) described user interest vector space model U is set up₀User document of interest collection D₀={ d₀₁,d₀₂,...,d_0m, make D={d₁,d₂,...,d_nFor document sets to be recommended, wherein document d_iCharacteristic vector beWherein, d_0eRepresent described user document of interest collection D₀In document, e=1,2 ..., m, m is described user document of interest collection D₀In total number of documents；T_ikRepresent document d_iKth item Feature Words；W_ikRepresent document d_iThe weight of kth item Feature Words；I=1,2 ..., n；K=1,2 ..., a；A represents document d_iThe total item of Feature Words；

3), when recommending document, calculate that all file characteristics in described document sets D to be recommended are vectorial and described user interest vector space model U₀Similarity r, it is recommended that going out the similarity r document more than threshold alpha, and feed back new document interested, described new collection of document isWhen selecting out-of-date in user's document of interest set or document that is that deviate user interest, calculate described user document of interest collection D respectively₀In each file characteristics is vectorial and described user interest vector space model U₀Similarity r', select r' less than the document of threshold alpha as out-of-date or deviation user interest document, described out-of-date or deviation user interest collection of document beThe span of described threshold alpha is 0～1；It is the document in D' for described new collection of document, f=1,2 ..., q, q is the total number of documents in described new collection of document D'；For described out-of-date or deviation user interest collection of document D " in document, h=1,2 ..., c, c be described out-of-date or deviation user interest collection of document D " in total number of documents；

4), to when increasing user's document of interest set, described new collection of document D' is added described user document of interest collection D₀In, constitute new first user document of interest collection D₁；Or when rejecting the document of out-of-date in user's document of interest set or deviation user interest, by described out-of-date or deviation user interest collection of document D " from described user document of interest collection D₀Middle rejecting, constitutes the second new user document of interest collection D₂；

5) described new first user document of interest collection D is calculated according to following formula₁Center vector

$\stackrel{\‾}{W_{D_1}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q}=\frac{m\stackrel{\‾}{W_{D_0}}+\underset{f=1}{\overset{q}{\Σ}}{W}_{d_{pf}}}{m+q};$

Wherein,For described user document of interest collection D₀In the characteristic vector of e document；For the characteristic vector of f document in described new collection of document D'；For described user document of interest collection D₀Center vector；

Described the second new user document of interest collection D is calculated according to following formula₂Center vector

$\stackrel{\‾}{W_{D_2}}=\frac{\underset{e=1}{\overset{m}{\Σ}}{W}_{d_{0e}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c}=\frac{m\stackrel{\‾}{W_{D_0}}-\underset{h=1}{\overset{c}{\Σ}}{W}_{d_{bh}}}{m-c};$

Wherein,For out-of-date or deviation user interest collection of document D " in the characteristic vector of h document；

6) willOrEach dimension sorts from big to small by weights, selectsOrFront N dimension build new user interest vector space model U₁Or U₂, handle simultaneouslyOrIt is stored in personalized recommendation system；Wherein, N less thanOrDimension；With described new user interest vector space model U₁Or U₂Replace step 1) in U₀Carry out new round recommendation。

2. the user interest model increment updating method of personalized recommendation system according to claim 1, it is characterised in that described step 1) in, build the user interest vector space model U based on document content₀Specifically comprise the following steps that

1) document that all users are interested is carried out Feature Words selection and term weight function calculates；

2) extract the characteristic vector of all users document interested, constitute file characteristics vector set D₃；

3) described file characteristics vector set D is calculated₃Center vector, by described file characteristics vector set D₃Center vector sort from big to small by the weight of each dimension, choose front M tie up as user interest vector space model U₀；Wherein M is less than described file characteristics vector set D₃The dimension of center vector。

3. the user interest model increment updating method of personalized recommendation system according to claim 2, it is characterised in that described file characteristics vector set D₃={ d₃₁,d₃₂,...,d_3xCenter vectorComputing formula be:

$\stackrel{\‾}{W_{D_3}}=\frac{\underset{y=1}{\overset{x}{\Σ}}{W}_{d_{3y}}}{x};$

Wherein, x is described file characteristics vector set D₃The number of middle element；For described file characteristics vector set D₃Middle y-th document characteristic vector；Y=1,2 ..., x。

4. the user interest model increment updating method according to the personalized recommendation system one of claims 1 to 3 Suo Shu, it is characterised in that document d in described document sets D to be recommended_iCharacteristic vector and described user interest vector space model U₀The computing formula of similarity r be:

$r=cos(W_{d_i},U_0)=\frac{W_{d_i}\·U_0}{{||W_{d_i}||}_2\×{||U_0||}_2};$

Wherein, | | | |₂Represent two norms。

5. the user interest model increment updating method of personalized recommendation system according to claim 4, it is characterised in that described user document of interest collection D₀In the e file characteristics be vectorial and described user interest vector space model U₀The computing formula of similarity r' be:

$r^{\′}=cos(W_{d_{0e}},U_0)=\frac{W_{d_{0e}}\·U_0}{{||W_{d_{0e}}||}_2\×{||U_0||}_2}.$