CN106126615B - A kind of method and system that point of interest is recommended - Google Patents
A kind of method and system that point of interest is recommended Download PDFInfo
- Publication number
- CN106126615B CN106126615B CN201610457222.5A CN201610457222A CN106126615B CN 106126615 B CN106126615 B CN 106126615B CN 201610457222 A CN201610457222 A CN 201610457222A CN 106126615 B CN106126615 B CN 106126615B
- Authority
- CN
- China
- Prior art keywords
- parameter
- comment text
- theme
- matrix
- interest
- 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.)
- Active
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
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9537—Spatial or temporal dependent retrieval, e.g. spatiotemporal queries
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of methods that point of interest is recommended, comprising: according to GeoMF algorithm and TopicMF algorithm is fused in matrix decomposition and obtains objective function geography information, comment information and information of registering;Update is registered number information parameter, comment text collection parameter, converts coefficient of kurtosis, theme number parameter, User Activity matrix of areas parameter;Predesignated subscriber is calculated to the preference degree of predetermined interest point using the objective function according to updated parameters;This method merges number of registering, the point of interest recommended method of three dimension factors of geographical location information and comment information, and the more efficient point of interest of Lai Shixian is recommended;The invention also discloses the systems that a kind of point of interest is recommended, and have said effect.
Description
Technical field
The present invention relates to Data Analysis Services field, in particular to a kind of method and system of point of interest recommendation.
Background technique
With the fast development of mobile device and internet, social networks becomes more and more mature.Meanwhile the prevalence of GPS
The generation of location-based social networks, such as Foursquare, Gowalla, Facebook Places etc. are promoted.This
A little location-based social networking services allow user's exploration interest point, by declaring their physical bit to registering for point of interest
It sets, and user can share their the access experiences about point of interest.Point of interest recommend for user and point of interest owner all
Huge value can be generated.Therefore, personalized point of interest recommendation has become an important task.
Point of interest recommendation be different from traditional recommendation because point of interest recommendation in geographical location information play it is critically important
Role.Mention in the first geographical law that Tobler is proposed: " every thing object and other all things are all related, but close
Things it is more even closer than remote connection ".This demonstrate a space clustering phenomenons: user is instinctively tending near access
Point of interest.Some researchers present different modes in terms of portraying this space clustering phenomenon.GeoMF(geo matrix
It factorization) is the state-of-the-art method that geographical location factor and matrix decomposition are combined.It expands in matrix decomposition
Decomposition vector has been opened up, User Activity region vector extensions have been implied vector into user, and point of interest influence area vector extensions
It is implied in vector into point of interest.One User Activity region vector shows the historical behavior of the access region of this user, one
Point of interest influence area vector then describes the point of interest that the accessed history based on this point of interest is obtained in different areas
Influence degree in domain.However, there is no consider influence of comment of the user for point of interest to recommendation for this model.
Have in conventional recommendation some on how to the research for the comment for utilizing commodity.Some articles propose TopicMF
(topic matrix factorization) method comment text and user's scoring to combine.McAuley etc. is proposed
A method of it being HFT (hidden factors and hidden topics), it utilizes LDA (Latent Dirichlet
Allocation) model is scored to handle comment text and be handled user using matrix decomposition model.The method passes through one
A relationships of indices implies between theme in the implicit vector sum of user as obtained in comment establishes connection.However, it does not have
Have and geographical location information is taken into account to provide more effective point of interest and recommend.
Therefore, number of registering how is merged, the point of interest of three dimension factors of geographical location information and comment information is recommended
Method, the more efficient point of interest of Lai Shixian are recommended, and are those skilled in the art's technical issues that need to address.
Summary of the invention
The object of the present invention is to provide the method and system that a kind of point of interest is recommended, by merging number of registering, geographical position
The point of interest recommended method of confidence breath and three dimension factors of comment information, the more efficient point of interest of Lai Shixian are recommended.
In order to solve the above technical problems, the present invention provides a kind of method that point of interest is recommended, comprising:
Geography information, comment information and information of registering are fused to matrix decomposition according to GeoMF algorithm and TopicMF algorithm
In obtain objective function;
The theme number parameter for making comment text concentrate each word is constant, minimizes the mesh using gradient descent algorithm
Scalar functions are to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis;
The comment text collection parameter is remained unchanged, each word position sampling master of comment text collection described in iteration is passed through
Topic number is to update theme number parameter;
Make the number information parameter of registering, the comment text collection parameter and the conversion coefficient of kurtosis are constant, utilize
Gradient descent algorithm updates User Activity matrix of areas parameter;
According to the updated number information parameter of registering, the comment text collection parameter, the conversion coefficient of kurtosis,
Predesignated subscriber is calculated to pre- Dingxing using the objective function in theme number parameter and User Activity matrix of areas parameter
The preference degree of interest point.
Wherein, geography information, comment information and information of registering are fused to by square according to GeoMF algorithm and TopicMF algorithm
Battle array obtains objective function in decomposing, comprising:
Geography information is fused to first object function obtained in matrix decomposition according to GeoMF algorithm
Comment information and information of registering are fused to the second objective function obtained in matrix decomposition according to TopicMF algorithm
It is added the first object function and second objective function to obtain objective function
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix,
X is User Activity matrix of areas parameter, and Y is point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | |
||FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter,The probability of each theme is concentrated for comment text,The probability for concentrating each word to belong to theme k for comment text.
Wherein, the theme number parameter for making comment text concentrate each word is constant, is minimized using gradient descent algorithm
The objective function is to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis, comprising:
Comment text is set to concentrate the theme number parameter z of each wordd,nIt is constant, it utilizesTo update the number information parameter Θ that registers, comment text
This collection parameter Φ and conversion coefficient of kurtosis κ.
Wherein, the comment text collection parameter is remained unchanged, each word position of comment text collection described in iteration is passed through
Theme number is sampled to update theme number parameter, comprising:
Make in the comment text collection parameter ΦWithIt remains unchanged, passes throughThe each word position of comment text collection described in iteration is adopted
Sample theme is numbered to update theme number parameter zd,n。
Wherein, making the number information parameter of registering, the comment text collection parameter and the conversion coefficient of kurtosis are constant,
User Activity matrix of areas parameter is updated using gradient descent algorithm, comprising:
Keep register the number information parameter Θ, the comment text collection parameter Φ and the conversion coefficient of kurtosis κ constant;
User Activity matrix of areas parameter X is initialized as zero matrix, then the zone of action x of user uuObjective function
L(xu) gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Utilize formulaUpdate xu;
It utilizesBy vectorProject to nonnegative quadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix,
Middle diagonal line valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient,
P+(xl) it is that vector x is mapped to nonnegative quadrantFunction.
The present invention also provides the systems that a kind of point of interest is recommended, comprising:
Objective function module, for by geography information, comment information and being registered according to GeoMF algorithm and TopicMF algorithm
Information, which is fused in matrix decomposition, obtains objective function;
First update module, the theme number parameter for making comment text concentrate each word is constant, using under gradient
Objective function described in drop algorithmic minimizing is to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis;
Second update module passes through comment text collection described in iteration for remaining unchanged the comment text collection parameter
Each word position sampling theme number is to update theme number parameter;
Third update module, for making register number information parameter, the comment text collection parameter and the conversion
Coefficient of kurtosis is constant, updates User Activity matrix of areas parameter using gradient descent algorithm;
Preference degree computing module, for according to the updated number information parameter of registering, the comment text collection ginseng
Number, the conversion coefficient of kurtosis, theme number parameter and User Activity matrix of areas parameter are calculated using the objective function
Predesignated subscriber is obtained to the preference degree of predetermined interest point.
Wherein, the objective function module includes:
First object function unit, for geography information to be fused to obtained in matrix decomposition according to GeoMF algorithm
One objective function
Second objective function unit, for comment information and information of registering to be fused to matrix point according to TopicMF algorithm
Second objective function obtained in solution
Objective function unit, for being added first object function and the second objective function to obtain objective function
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix,
X is User Activity matrix of areas parameter, and Y is point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | |
||FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter,The probability of each theme is concentrated for comment text,The probability for concentrating each word to belong to theme k for comment text.
Wherein, first update module is specially the theme number parameter z for making comment text concentrate each wordd,nNo
Become, utilizesTo update the number information parameter Θ that registers,
Comment text collection parameter Φ and the module for converting coefficient of kurtosis κ.
Wherein, second update module is specially and makes in the comment text collection parameter ΦWithIt keeps
It is constant, pass throughThe each list of comment text collection described in iteration
Lexeme sets sampling theme number to update theme number parameter zd,nModule.
Wherein, the third update module includes:
Gradient computing unit, for making the number information parameter Θ that registers, the comment text collection parameter Φ and described
Coefficient of kurtosis κ is constant that User Activity matrix of areas parameter X is initialized as zero matrix for conversion, then the zone of action x of user uu
Objective function L (xu) gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Updating unit, for utilizing formulaUpdate xu;
Projecting cell, for utilizingBy vectorIt projects to
Nonnegative quadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix,
Middle diagonal line valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient,
P+(xl) it is that vector x is mapped to nonnegative quadrantFunction.
The method that point of interest provided by the present invention is recommended, comprising: will be geographical according to GeoMF algorithm and TopicMF algorithm
Information, comment information and information of registering are fused in matrix decomposition and obtain objective function;Comment text is set to concentrate each word
Theme number parameter is constant, minimizes the objective function using gradient descent algorithm to update number information parameter of registering, comments
By text set parameter and conversion coefficient of kurtosis;The comment text collection parameter is remained unchanged, comment text described in iteration is passed through
Collect each word position sampling theme number to update theme number parameter;Make the number information parameter of registering, the comment
Text set parameter and the conversion coefficient of kurtosis are constant, update User Activity matrix of areas parameter using gradient descent algorithm;Root
According to the updated number information parameter of registering, the comment text collection parameter, the conversion coefficient of kurtosis, theme number ginseng
Predesignated subscriber is calculated to the hobby of predetermined interest point using the objective function in several and User Activity matrix of areas parameter
Degree;
This method all decomposes three factors in matrix, and optimizes update to relevant parameter, by updated ginseng
Number is brought objective function into and is calculated, and comes to recommend point of interest to user eventually by the value for the preference degree being calculated;Melt
Number of registering, the point of interest recommended method of three dimension factors of geographical location information and comment information are closed, Lai Shixian is more efficient
Point of interest recommend;The present invention also provides the systems that a kind of point of interest is recommended, and have said effect, details are not described herein.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is the flow chart for the method that point of interest provided by the embodiment of the present invention is recommended;
Fig. 2 is the structural block diagram for the system that point of interest provided by the embodiment of the present invention is recommended.
Specific embodiment
Core of the invention is to provide a kind of method and system that point of interest is recommended, by merging number of registering, geographical position
The point of interest recommended method of confidence breath and three dimension factors of comment information, the more efficient point of interest of Lai Shixian are recommended.
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
Every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
Referring to FIG. 1, Fig. 1 is the flow chart for the method that point of interest provided by the embodiment of the present invention is recommended;This method can
To include:
S100, geography information, comment information and information of registering are fused to by square according to GeoMF algorithm and TopicMF algorithm
Battle array obtains objective function in decomposing;
Where it is assumed that there is M user U={ u1,u2...ui,...uMAnd N number of point of interest V={ v1,v2...vj,...vN}。It is rating matrix, R is degree matrix of registering, wherein Ri,jIt is register number of the user i to point of interest j.Ri,jValue
Indicate that user i never registered point of interest j for 0.The problem of point of interest is recommended is to recommend suitable point of interest to user.
Collaborative filtering method (such as PMF) based on model is used to assist to recommend.Due to there is a kind of collaborative filtering in POI recommendation
(One Class Collaborative Filtering), solution are one weight matrixs of increase in matrix decomposition,
Referred to as weight matrix decomposes (Weighted Matrix Factorization).
Firstly, point of interest recommendation needs to consider geographic factor.Here it can be described by X, Y-direction amount in point of interest recommendation
The influence of geographic factor.Secondly, user can leave their comments about the point of interest accessed in Access Interest point.di,j
It is comment text of the user i to point of interest j, usually along with behavior of once registering.In conventional recommendation, TopicMF method
(such as CTR, HFT) combines comment text in recommendation and achieves good effect.
Due to limitation of the SVD singular value decomposition when decomposing sparse matrix in terms of accuracy and scalability, Koren is proposed
A kind of method of the new matrix decomposition for recommendation.The method can pass through formulaTo predict i pairs of user
The scoring r of commodity ji,j.This present matrix disassembling method (MF) has solved the problems, such as basic recommended method when scoring.I
Rating matrix is resolved into two low-rank matrixes, the purpose is to find implicit user characteristics matrixWith commodity spy
Levy matrixWherein K is the number (K < < M, N) of implicit vector.Can be learnt by following formula out this two
A matrix:
For score information, the factor wherein implied is found with matrix decomposition.And for text information, then with LDA come
Explore the implicit theme contained in text set.Every article d, that is, comment text collection d has a theme distribution θd, which show texts
The probability of each theme included in chapter d.Similar, each theme k has a word distribution phik, which show each lists
Word belongs to the probability of theme k.The likelihood function for generating text set M is expressed as follows:
Firstly, geography information is fused in matrix decomposition into matrix decomposition according to GeoMF algorithm by fusion geography information
Obtained first object functionDetailed process can be with
It is:
Due to the presence of geographic factor in point of interest recommendation, geographic factor is influenced here and matrix decomposition combines.?
During point of interest is recommended, the behavior of registering shows the frequency of user's Access Interest point.Which results in only exist positive example to face one
Class collaborative filtering problem (One Class Collaborative Filtering).Its solution be to each user with
Machine samples some negative examples and bears the small weight of example distribution ratio positive example to these.(Weighted is decomposed based on weight matrix
Matrix Factorization), objective function is then:
Wherein, it is 0/1 that the R in the formula, which is value,
Rating matrix, indicates whether user u accessed point of interest i.
Further, in order to describe geographical location preference, two matrixes X, Y are provided with.It is drawn the zone of action of all users
It is divided into L sub-regions.Each user has a zone of action vector x, each of these value xlIndicate that this user is likely to occur
Probability in subregion l.Similar, each point of interest has an influence area vector y, each of these value ylIt indicates
Disturbance degree of this point of interest in subregion l.The objective function of this model are as follows:
Wherein, in the formula
It is predicted scoring.Disturbance degree of the point of interest i in subregion l beWherein K () is standard normal
It is distributed and σ is standard deviation.
Firstly, fusion comment information, and registers information into matrix decomposition, according to TopicMF algorithm by comment information and
Information of registering is fused to the second objective function obtained in matrix decompositionDetailed process may is that
The scoring of user provides recessive feedback for recommendation.Meanwhile user also provides the comment text of commodity
Some implicit informations.It will be than individually considering that one of them is more efficient if carrying out recommendation using both information simultaneously.
On the one hand commodity preference is hidden in scoring, this information on the other hand can be also obtained from comment.Respectively by matrix decomposition and
LDA come find score and comment among imply preference.HFT (hidden factors and hidden topics) method exists
It is generated under this background.The method combines scoring and comment information by following conversion:
Wherein parameter κ is used to control " kurtosis " of conversion.Due to the presence of index, θj,k
Value always positive number.Above equation establishes VjAnd θjBetween connection so that they can simultaneously increase or reduce.This
A relationship is very reasonable because of the product features that user centainly and in scoring is hidden by the feature of the commodity of comment description
It is similar.
HFT method combines scoring and comment information by minimizing following objective function:
Wherein, the first part of this equation is prediction
The error of scoring, second part are then the likelihood functions of comment text.Parameter lambda is used to balance the contribution of the two.
Finally it is added the first object function and second objective function to obtain objective function
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix,
X is User Activity matrix of areas parameter, and Y is point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | |
||FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter,The probability of each theme is concentrated for comment text,The probability for concentrating each word to belong to theme k for comment text.Detailed process may is that
By above description, by being solved the problems, such as in conjunction with the implicit comment theme of implicit scoring vector sum.Pass through minimum
Change objective function below to combine number of registering, geographical location information and comment text:
WhereinWith | | x | |1It is the regular terms in order to avoid over-fitting, parameter μ is used to balance whole
The contribution degree of comment text collection in a method.The first part of formula is that prediction is registered the error of number.This conventional recommendation is most
Big is not both that this formula combines geographical location influence in matrix decomposition, and the extension of matrix X, Y in formula embodies this
One is different.The second part of formula is the likelihood function of comment text.
Wherein, with universal and GPS the development of mobile device, location-based social networks has attracted millions of users
Share their position.Point of interest is recommended to select user attractive point of interest to be highly important, point of interest recommendation
All it is related with the factor of many dimensions, such as number of registering, geographical location information and comment information.Existing method is good
Geographical location influence factor is integrated into matrix decomposition, but it does not consider that the comment information of user pushes away point of interest
The influence recommended.The step merges number information of registering, geographical location information and comment information using mixed method, leads to
Crossing influences geographical location to incorporate matrix decomposition to combine register number and geographical location information;Meanwhile it can be become by one
It brings connection and implies the number factor and from LDA (Latent Dirichlet of registering as obtained in matrix decomposition
Allocation the implicit comment theme obtained in) combines register number and user comment information with this.
S110, the theme number parameter for making comment text concentrate each word are constant, are minimized using gradient descent algorithm
The objective function is to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis;
S120, the comment text collection parameter is remained unchanged, passes through each word position of comment text collection described in iteration
Theme number is sampled to update theme number parameter;
S130, making the number information parameter of registering, the comment text collection parameter and the conversion coefficient of kurtosis are constant,
User Activity matrix of areas parameter is updated using gradient descent algorithm;
Wherein, while related number information parameter Θ={ U, the V } that register of optimization, related comment text collection parameter Φ=θ,
φ }, User Activity matrix of areas parameter X, conversion coefficient of kurtosis κ, theme number parameter zd,n, and then utilize the parameter after optimization
Minimize objective function
It was found that parameter Θ and parameter Φ are associated (due to above-mentioned conversion formulas).So independent it cannot be optimized
, learning parameter Θ and X can be come with gradient descent method here, using gibbs sampler (Gibbs sampling) come Optimal Parameters
Φ。
Detailed process may is that
Comment text is set to concentrate the theme number parameter z of each wordd,nIt is constant, it utilizesTo update the number information parameter Θ that registers, comment text
This collection parameter Φ and conversion coefficient of kurtosis κ.
Wherein, comment text is kept to concentrate the theme number z of each wordd,nIt is constant, then more by gradient descent method
New parameter Θ, Φ and κ.The non-of multiple variables is solved using the Quasi-Newton iterative method (L-BFGS) for being similar to the decline of common gradient
Linear optimization problem, because the method is very convenient when in use.When there are many parameters, it can easily calculate ladder
Degree.
Make in the comment text collection parameter ΦWithIt remains unchanged, passes throughThe each word position of comment text collection described in iteration is adopted
Sample theme is numbered to update theme number parameter zd,n。
Wherein, the parameter in relation to θ and φ remains unchanged, then the word position all by the entire comment text set of iteration
It sets to sample theme number zd,n.Z is given as LDAd,nInitialization one 1 to the random value between k and according toProbability distribution sample zd,n.It is not by Di Li Cray that this step and LDA, which are not uniquely both theme distribution θ,
Gradient decline optimization process before profile samples obtain, but the sub-step before passing through obtains i.e. obtains.
Finally, making register the number information parameter Θ, the comment text collection parameter Φ and the conversion coefficient of kurtosis κ
It is constant;
User Activity matrix of areas parameter X is initialized as zero matrix, then the zone of action x of user uuObjective function
L(xu) gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Utilize formulaUpdate xu;
It utilizesBy vectorProject to nonnegative quadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix,
Middle diagonal line valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient,
P+(xl) it is that vector x is mapped to nonnegative quadrantFunction.
Wherein, in this step, learn the zone of action matrix X of user.Parameter Θ is kept, Φ, κ is constant, then about X's
Objective function is similar to the minimization problem of a non-negative weight.Undated parameter is come by using gradient descent algorithm.Simultaneously more
New all parameters are very difficult, it is possible to individually learn the zone of action vector of each user.For the activity of user
Region X is initialized as a zero matrix.Objective function L (x about xu) gradient be:
▽L(xu)=YTWu(Yxu-(ru-Vuu))+λ2
Then we update x by following formulau:
Wherein α is learning rate, P+(xl) a vectorIts nonnegative quadrant is projected toOn.
S140, according to the updated number information parameter of registering, the comment text collection parameter, the conversion kurtosis
Predesignated subscriber is calculated to pre- using the objective function in coefficient, theme number parameter and User Activity matrix of areas parameter
Determine the preference degree of point of interest.
Wherein, updated parameter is brought into objective function, it is availablePass throughI couples of available user
The preference degree of point of interest j.Here hobby angle value can obtain and can think by scoringResult be user i to interest
The scoring of point j.It is known that user to the fancy grade of a point of interest, can recommend preference degree highest according to the height of scoring
Point of interest to user, the interest-degree of the predetermined number of preference degree from high to low can also be recommended to user.
Based on the above-mentioned technical proposal, the method that point of interest provided in an embodiment of the present invention is recommended, can merge number of registering,
The point of interest recommended method of three dimension factors of geographical location information and comment information, the more efficient point of interest of Lai Shixian push away
It recommends.
The embodiment of the invention provides the methods that point of interest is recommended, and merge number of registering, geographical location information and comment letter
The point of interest recommended method of three dimension factors is ceased, the more efficient point of interest of Lai Shixian is recommended.
The system recommended below point of interest provided in an embodiment of the present invention is introduced, and point of interest described below is recommended
System can correspond to each other reference with the method that above-described point of interest is recommended.
Referring to FIG. 2, Fig. 2 is the structural block diagram for the system that point of interest provided by the embodiment of the present invention is recommended;The system
May include:
Objective function module 100, for according to GeoMF algorithm and TopicMF algorithm by geography information, comment information and label
It is fused in matrix decomposition to information and obtains objective function;
First update module 200, the theme number parameter for making comment text concentrate each word is constant, utilizes gradient
Descent algorithm minimizes the objective function to update number information parameter of registering, comment text collection parameter and conversion kurtosis system
Number;
Second update module 300 passes through comment text described in iteration for remaining unchanged the comment text collection parameter
Collect each word position sampling theme number to update theme number parameter;
Third update module 400, for making the number information parameter of registering, the comment text collection parameter and described turn
It is constant to change coefficient of kurtosis, updates User Activity matrix of areas parameter using gradient descent algorithm;
Preference degree computing module 500, for according to the updated number information parameter of registering, the comment text collection
Parameter, the conversion coefficient of kurtosis, theme number parameter and User Activity matrix of areas parameter utilize the objective function, meter
Calculation obtains predesignated subscriber to the preference degree of predetermined interest point.
Optionally, the objective function module 100 includes:
First object function unit, for geography information to be fused to obtained in matrix decomposition according to GeoMF algorithm
One objective function
Second objective function unit, for comment information and information of registering to be fused to matrix point according to TopicMF algorithm
Second objective function obtained in solution
Objective function unit, for being added first object function and the second objective function to obtain objective function
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix,
X is User Activity matrix of areas parameter, and Y is point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | |
||FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter,The probability of each theme is concentrated for comment text,The probability for concentrating each word to belong to theme k for comment text.
Optionally, first update module 200 is specially the theme number parameter for making comment text concentrate each word
zd,nIt is constant, it utilizesTo update number information ginseng of registering
Number Θ, comment text collection parameter Φ and the module for converting coefficient of kurtosis κ.
Optionally, second update module 300 is specially and makes in the comment text collection parameter ΦWith
It remains unchanged, passes throughComment text collection described in iteration is every
A word position sampling theme number is to update theme number parameter zd,nModule.
Optionally, the third update module 400 includes:
Gradient computing unit, for making the number information parameter Θ that registers, the comment text collection parameter Φ and described
Coefficient of kurtosis κ is constant that User Activity matrix of areas parameter X is initialized as zero matrix for conversion, then the zone of action x of user uu
Objective function L (xu) gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Updating unit, for utilizing formulaUpdate xu;
Projecting cell, for utilizingBy vectorIt projects to
Nonnegative quadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix,
Middle diagonal line valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient,
P+(xl) it is that vector x is mapped to nonnegative quadrantFunction.
Each embodiment is described in a progressive manner in specification, the highlights of each of the examples are with other realities
The difference of example is applied, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment
Speech, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place is referring to method part illustration
?.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
The step of method described in conjunction with the examples disclosed in this document or algorithm, can directly be held with hardware, processor
The combination of capable software module or the two is implemented.Software module can be placed in random access memory (RAM), memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
The method and system recommended above point of interest provided by the present invention are described in detail.It is used herein
A specific example illustrates the principle and implementation of the invention, and the above embodiments are only used to help understand originally
The method and its core concept of invention.It should be pointed out that for those skilled in the art, not departing from this hair
, can be with several improvements and modifications are made to the present invention under the premise of bright principle, these improvement and modification also fall into power of the present invention
In the protection scope that benefit requires.
Claims (8)
1. a kind of method that point of interest is recommended characterized by comprising
Geography information is fused to first object function obtained in matrix decomposition according to GeoMF algorithm
Comment information and information of registering are fused to the second objective function obtained in matrix decomposition according to TopicMF algorithm
It is added the first object function and second objective function to obtain objective function
The theme number parameter for making comment text concentrate each word is constant, minimizes the target letter using gradient descent algorithm
Number is to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis;
The comment text collection parameter is remained unchanged, theme is sampled by each word position of comment text collection described in iteration and is compiled
Number to update theme number parameter;
Make the number information parameter of registering, the comment text collection parameter and the conversion coefficient of kurtosis are constant, utilize gradient
Descent algorithm updates User Activity matrix of areas parameter;
According to the updated number information parameter of registering, the comment text collection parameter, the conversion coefficient of kurtosis, theme
Predesignated subscriber is calculated to predetermined interest point using the objective function in number parameter and User Activity matrix of areas parameter
Preference degree;
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix, and X is
User Activity matrix of areas parameter, Y are point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | | |
|FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter, and z is the theme,Each theme is concentrated for comment text
Probability,The probability for concentrating each word to belong to theme k for comment text.
2. the method that point of interest as described in claim 1 is recommended, which is characterized in that comment text is made to concentrate the master of each word
It is constant to inscribe number parameter, minimizes the objective function using gradient descent algorithm to update number information parameter of registering, comments on
Text set parameter and conversion coefficient of kurtosis, comprising:
Comment text is set to concentrate the theme number parameter z of each wordd,nIt is constant, it utilizesTo update the number information parameter Θ that registers, comment text
This collection parameter Φ and conversion coefficient of kurtosis κ;Wherein, ΘnewFor new number information parameter of registering, ΦnewFor new comment text
Collect parameter, κnewFor new conversion coefficient of kurtosis, zoldIt is renewal function for old theme number parameter and update.
3. the method that point of interest as claimed in claim 2 is recommended, which is characterized in that keep the comment text collection parameter not
Become, theme number sampled to update theme number parameter by each word position of comment text collection described in iteration, comprising:
Make in the comment text collection parameter ΦWithIt remains unchanged, passes throughThe each word position of comment text collection described in iteration is adopted
Sample theme is numbered to update theme number parameter zd,nAnd update the probability that comment text concentrates each word to belong to theme k
Wherein, sample is sampling function,For new theme number parameter,Each word is concentrated for new comment text
Belong to the probability of theme k.
4. the method that point of interest as claimed in claim 3 is recommended, which is characterized in that make the number information parameter of registering, institute
It states comment text collection parameter and the conversion coefficient of kurtosis is constant, update User Activity matrix of areas using gradient descent algorithm and join
Number, comprising:
Keep register the number information parameter Θ, the comment text collection parameter Φ and the conversion coefficient of kurtosis κ constant;
User Activity matrix of areas parameter X is initialized as zero matrix, then the zone of action x of user uuObjective function L (xu)
Gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Utilize formulaUpdate xu;
It utilizesBy vectorProject to nonnegative quadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix, wherein right
Linea angulata valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient, P+
(xl) it is that vector x is mapped to nonnegative quadrantFunction, 1 be subregion 1, L be all users zone of action sum.
5. the system that a kind of point of interest is recommended characterized by comprising
Objective function module, for according to GeoMF algorithm and TopicMF algorithm by geography information, comment information and information of registering
It is fused in matrix decomposition and obtains objective function;
First update module, the theme number parameter for making comment text concentrate each word is constant, is declined using gradient and is calculated
Method minimizes the objective function to update number information parameter of registering, comment text collection parameter and conversion coefficient of kurtosis;
Second update module, it is each by comment text collection described in iteration for remaining unchanged the comment text collection parameter
Word position samples theme number to update theme number parameter;
Third update module, for making register number information parameter, the comment text collection parameter and the conversion kurtosis
Coefficient is constant, updates User Activity matrix of areas parameter using gradient descent algorithm;
Preference degree computing module, for according to the updated number information parameter of registering, the comment text collection parameter, institute
Conversion coefficient of kurtosis is stated, theme number parameter and User Activity matrix of areas parameter are calculated pre- using the objective function
User is determined to the preference degree of predetermined interest point;
Wherein, the objective function module includes:
First object function unit, for geography information to be fused to the first mesh obtained in matrix decomposition according to GeoMF algorithm
Scalar functions
Second objective function unit, for comment information and information of registering to be fused in matrix decomposition according to TopicMF algorithm
The second obtained objective function
Objective function unit, for being added first object function and the second objective function to obtain objective function
Wherein, R is degree matrix of registering,The rating matrix for being 0/1 for value, U are user's matrix, and V is interest dot matrix, and X is
User Activity matrix of areas parameter, Y are point of interest influence area matrix, and W is weight matrix of registering, μ, λ1, λ2For coefficient, | | |
|FFor norm, | | | |1For nucleus number, Ri,jIt is user i to the number of registering of point of interest j, d is comment text collection, NdFor N number of interest
The comment text collection of point, M are user's number,Θ is number information parameter of registering, and Φ is comment text
This collection parameter, κ are conversion coefficient of kurtosis, zd,nBe the theme number parameter, and z is the theme,Each theme is concentrated for comment text
Probability,The probability for concentrating each word to belong to theme k for comment text.
6. the system that point of interest as claimed in claim 5 is recommended, which is characterized in that first update module is specially to make to comment
Paper originally concentrates the theme number parameter z of each wordd,nIt is constant, it utilizesTo update the number information parameter Θ that registers, comment text
This collection parameter Φ and the module for converting coefficient of kurtosis κ;Wherein, ΘnewFor new number information parameter of registering, ΦnewIt is commented for new
By text set parameter, κnewFor new conversion coefficient of kurtosis, zoldIt is renewal function for old theme number parameter and update.
7. the system that point of interest as claimed in claim 6 is recommended, which is characterized in that second update module is specially to make institute
It states in comment text collection parameter ΦWithIt remains unchanged, passes throughThe each word position of comment text collection described in iteration is adopted
Sample theme is numbered to update theme number parameter zd,nModule and update comment text each word concentrated to belong to the general of theme k
RateWherein, sample is sampling function,For new theme number parameter,It is concentrated for new comment text
Each word belongs to the probability of theme k.
8. the system that point of interest as claimed in claim 7 is recommended, which is characterized in that the third update module includes:
Gradient computing unit, for making register number information parameter Θ, the comment text collection parameter Φ and the conversion
Coefficient of kurtosis κ is constant to be initialized as zero matrix for User Activity matrix of areas parameter X, then the zone of action x of user uuMesh
Scalar functions L (xu) gradient be ▽ L (xu)=YTWu(Yxu-(ru-Vuu))+λ2;
Updating unit, for utilizing formulaUpdate xu;
Projecting cell, for utilizingBy vectorIt projects to non-negative
QuadrantOn;
Wherein, α is learning rate, xlThe probability in subregion l, W possibly are present at for useruFor a diagonal matrix, wherein right
Linea angulata valueFor weight wu,i, xuFor the zone of action vector of user u, ruFor the column scoring vector of user u, λ2For coefficient, P+
(xl) it is that vector x is mapped to nonnegative quadrantFunction, 1 be subregion 1, L be all users zone of action sum.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610457222.5A CN106126615B (en) | 2016-06-22 | 2016-06-22 | A kind of method and system that point of interest is recommended |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610457222.5A CN106126615B (en) | 2016-06-22 | 2016-06-22 | A kind of method and system that point of interest is recommended |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106126615A CN106126615A (en) | 2016-11-16 |
CN106126615B true CN106126615B (en) | 2019-06-11 |
Family
ID=57269270
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610457222.5A Active CN106126615B (en) | 2016-06-22 | 2016-06-22 | A kind of method and system that point of interest is recommended |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106126615B (en) |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106960044B (en) * | 2017-03-30 | 2020-04-07 | 浙江鸿程计算机系统有限公司 | Time perception personalized POI recommendation method based on tensor decomposition and weighted HITS |
CN108346066A (en) * | 2018-01-19 | 2018-07-31 | 广东欧珀移动通信有限公司 | Advertisement sending method, device, server and storage medium |
CN110210691B (en) * | 2018-04-27 | 2024-02-06 | 腾讯科技(深圳)有限公司 | Resource recommendation method, device, storage medium and equipment |
CN109190053A (en) * | 2018-07-04 | 2019-01-11 | 南京邮电大学 | One kind being based on point of interest importance and the authoritative point of interest recommended method of user |
CN109508428B (en) * | 2019-01-21 | 2021-03-12 | 宿州学院 | Interest point recommendation method based on interest point true popularity and implicit trust mining |
CN109961183B (en) * | 2019-03-20 | 2023-06-23 | 重庆邮电大学 | Method for measuring influence of comment information on user sign-in |
CN110134885B (en) * | 2019-05-22 | 2022-12-20 | 广东工业大学 | Interest point recommendation method, device, equipment and computer storage medium |
CN111209475B (en) * | 2019-12-27 | 2022-03-15 | 武汉大学 | Interest point recommendation method and device based on space-time sequence and social embedded ranking |
CN111177565B (en) * | 2019-12-31 | 2021-08-31 | 杭州电子科技大学 | Interest point recommendation method based on correlation matrix and word vector model |
CN112650932B (en) * | 2021-01-04 | 2022-09-23 | 重庆邮电大学 | Social and position relationship fused neural collaborative filtering POI recommendation method |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103023977A (en) * | 2012-11-19 | 2013-04-03 | 华南理工大学 | Recommendation system and method based on positional information |
CN105574003A (en) * | 2014-10-10 | 2016-05-11 | 华东师范大学 | Comment text and score analysis-based information recommendation method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10235403B2 (en) * | 2014-07-08 | 2019-03-19 | Palo Alto Research Center Incorporated | Parallel collective matrix factorization framework for big data |
-
2016
- 2016-06-22 CN CN201610457222.5A patent/CN106126615B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103023977A (en) * | 2012-11-19 | 2013-04-03 | 华南理工大学 | Recommendation system and method based on positional information |
CN105574003A (en) * | 2014-10-10 | 2016-05-11 | 华东师范大学 | Comment text and score analysis-based information recommendation method |
Non-Patent Citations (1)
Title |
---|
一种融合情景和评论信息的位置社交网络兴趣点推荐模型;高榕,等;《计算机研究与发展》;20160430;第752-762页 * |
Also Published As
Publication number | Publication date |
---|---|
CN106126615A (en) | 2016-11-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106126615B (en) | A kind of method and system that point of interest is recommended | |
CN107239444B (en) | A kind of term vector training method and system merging part of speech and location information | |
Prashker et al. | Route choice models used in the stochastic user equilibrium problem: a review | |
Arndt-Lappe | Analogy in suffix rivalry: The case of English-ity and-ness | |
CN110188182A (en) | Model training method, dialogue generation method, device, equipment and medium | |
CN110119475B (en) | POI recommendation method and system | |
Xu et al. | Collaborative APIs recommendation for artificial intelligence of things with information fusion | |
CN111259133A (en) | Personalized recommendation method integrating multiple information | |
Tan et al. | Improving aspect-based sentiment analysis via aligning aspect embedding | |
Pei et al. | A modular task-oriented dialogue system using a neural mixture-of-experts | |
CN110209954A (en) | Group recommending method based on LDA topic model and deep learning | |
Zhang et al. | The optimally designed dynamic memory networks for targeted sentiment classification | |
Peng et al. | Incremental term representation learning for social network analysis | |
Li et al. | Learning daily activity patterns with probabilistic grammars | |
Liu et al. | Smart city moving target tracking algorithm based on quantum genetic and particle filter | |
Yang et al. | [Retracted] Research on Students’ Adaptive Learning System Based on Deep Learning Model | |
Shen et al. | A college student behavior analysis and management method based on machine learning technology | |
Zhang et al. | [Retracted] Application and Analysis of Artificial Intelligence in College Students’ Career Planning and Employment and Entrepreneurship Information Recommendation | |
Persia et al. | Improving orienteering-based tourist trip planning with social sensing | |
Fort et al. | Can a linguistic serial founder effect originating in Africa explain the worldwide phonemic cline? | |
CN109684561B (en) | Interest point recommendation method based on deep semantic analysis of user sign-in behavior change | |
Li et al. | [Retracted] Research on the Recommendation Algorithm of Rural Tourism Routes Based on the Fusion Model of Multiple Data Sources | |
CN110008411A (en) | It is a kind of to be registered the deep learning point of interest recommended method of sparse matrix based on user | |
Nawshin et al. | Modeling weather-aware prediction of user activities and future visits | |
CN106779181A (en) | Method is recommended by a kind of medical institutions based on linear regression factor Non-negative Matrix Factorization model |
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 | ||
TR01 | Transfer of patent right |
Effective date of registration: 20201217 Address after: Room 509, 5 / F, golden house building, 280 Dongping street, Suzhou Industrial Park, Suzhou, Jiangsu Province Patentee after: Shenxing Taibao Intelligent Technology (Suzhou) Co.,Ltd. Address before: 215123 No. 199 benevolence Road, Suzhou Industrial Park, Jiangsu, China Patentee before: Suzhou University |
|
TR01 | Transfer of patent right |