CN105894310A - Personalized recommendation method - Google Patents
Personalized recommendation method Download PDFInfo
- Publication number
- CN105894310A CN105894310A CN201410578957.4A CN201410578957A CN105894310A CN 105894310 A CN105894310 A CN 105894310A CN 201410578957 A CN201410578957 A CN 201410578957A CN 105894310 A CN105894310 A CN 105894310A
- Authority
- CN
- China
- Prior art keywords
- user
- article
- characteristic vector
- collection
- vector
- 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.)
- Pending
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a personalized recommendation method. The method comprises the steps of firstly setting initial values for feature vectors of a plurality of users and a plurality of articles, then updating the feature vector of each user who accesses to the articles by the feature vectors of the articles according to an article access record of each user, updating the feature vector of the article to which the user accesses by the feature vector of the user according to the article access record of the user, executing the above feature vector updating process repeatedly, and finally recommending a set quantity of articles to a specific user according to the similarity between the feature vector of the specific user and the feature vector of each article. The method can significantly reduce the computation complexity of a recommendation system, and can improve a predication accuracy index of the recommendation system. The method has very wide application prospects in the fields such as electronic commerce.
Description
Technical field
The present invention relates to internet arena, relate in particular to a kind of personalized recommendation method.
Background technology
Personalized recommendation is the Characteristic of Interest according to user and purchasing behavior, recommends its letter interested to user
Breath and commodity.Along with the continuous expansion of ecommerce scale, commodity number and kind quickly increase, Gu Kexu
Devote a tremendous amount of time and just can find the commodity oneself wanting to buy.Although user can be carried by search engine
High search efficiency, but, the product that a lot of users are interested, user in advance and is unaware of, therefore, it is impossible to
Search engine based on key word is used to obtain relevant information.In order to solve the problems referred to above, it is necessary to individual character
Change the help of commending system.Personalized recommendation system can either solve the problem of information overload of user, it is also possible to
User is helped to find him to be likely to information or the commodity liking the most also being unaware of.Well-known Internet firm Asia horse
Inferior (Amazon), Netflix, Google, Alibaba, Baidu and Tengxun etc., all have started to use personalization
Commending system recommends its information interested, commodity and application to user.
At present, in the industry cycle existing a lot of proposed algorithm comes into operation, and such as based on article collaborative filterings are calculated
Method (itemCF), collaborative filtering based on user (userCF), hidden semantic model (such as LFM) and based on
Model of figure etc..But, the shortcoming of these methods is that computation complexity is higher, arithmetic speed is slower.Ratio
As, the core proposed algorithm-ItemCF method that current a lot of large-scale websites all use, it calculates article and is correlated with
The time complexity of table is O (M (K/M)2), and UserCF method calculates the time complexity of user's correlation table
It is O (N (K/N)2), the time complexity of hidden semantic model LFM when article number is a lot of up to O (M N F),
Wherein M is user's number, and N is article number, and F is characterized number.Therefore, as article or user
When quantity increases, the performance of existing proposed algorithm can be remarkably decreased.This causes large-scale commending system to typically require
Huge calculating equipment realizes proposed algorithm, and is difficult to the real-time recommendation for article.
Summary of the invention
The problem existed in view of above-mentioned prior art, it is an object of the invention to provide a kind of personalized recommendation side
Method, reduces the computation complexity of commending system, and the index such as the accuracy rate improving commending system.
Purpose in accordance with the above, the present invention proposes a kind of personalized recommendation method, it is characterised in that
Described method is included in the server accessing the Internet and performs following steps:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is that the multiple article in described article collection I arrange characteristic vector initial value, and is described
User collects the multiple user setup characteristic vector initial values in U；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
Compared with prior art, the inventive method considerably reduces the computation complexity of commending system, for reality
Time article recommend provide new method.In addition, this patent method recommendation accuracy rate (precision),
The indexs such as recall rate (recall) and coverage rate (coverage) are significantly better than main flow proposed algorithm itemCF.
Accompanying drawing explanation
Fig. 1 is the characteristic vector method for expressing that user collects each user in U；
Fig. 2 is the characteristic vector method for expressing of each article in article collection I；
Fig. 3 is the algorithm flow chart of a kind of commending system；
Fig. 4 is a kind of idiographic flow that Fig. 3 method performs T iteration；
Fig. 5 is the algorithm flow chart of a kind of commending system；
Fig. 6 is a kind of idiographic flow that Fig. 5 method performs T iteration；
Fig. 7 is the concrete methods of realizing of step S4 of Fig. 3 and Fig. 5 method；
Fig. 8 is step S4 of Fig. 3 and Fig. 5 method*Concrete methods of realizing.
Detailed description of the invention
In conjunction with accompanying drawing, the inventive method is described in further detail.
First, illustrate that user collects U, article collection I, feature set K and the implication of set of relations R.Accessing mutually
In the server of networking, obtain and store the user being made up of multiple users (user) mark and collect U and by multiple
The article collection I of article (item) mark composition.Described ID is the unique identifier of user, such as user
Account, phone number, Email address or social networks mark etc..Described article mark be article only
One identification code.Article are interconnection user on the network's all entities interested, including the article mark on website,
Brand, the offer businessman of article, Internet advertising and Web page etc..If described user collects U
Containing M element, described article collection I contains N number of element.
In the server accessing the Internet, store and identified, by multiple features (feature), feature set K formed.
The plurality of feature is that described user collects the feature of each user in U, is again each in described article collection I
The feature of article.User uses identical feature set K with article.If user has musical features, illustrate to use
Family hobby music, and article have musical features, illustrate that article are relevant to musical theme.Feature set K contains
L element.Feature set K can include multiple character subset, such as K={K1, K2..., Kp, wherein Ki
Represent feature set K ith feature subset (i=1,2 ... p；p≥2).
In the server accessing the Internet, obtain and store the behavior record composition of article accessible by user
Set of relations R.When arbitrary article are accessed by any user, then one user behavior record of storage, should
Record is at least made up of ID and article mark.And this record is called an element of set of relations R.
Such as set of relations R={..., (m, n) ..., wherein (m n) represents that user m have accessed article n to element.User visits
Asking the mode of article, including browsing, put into shopping cart, buy and evaluation etc., therefore, set of relations R is also
User's access mode parameter can be included.Such as set of relations R={..., (m, n, b) ... }, wherein element (m, n, b)
Representing that user m have accessed article n, its access mode is b.Described access mode b is purchased for browsing, putting into
Thing car, buy and evaluate in one.The data of set of relations R are typically from syslog file (Log).
The method for expressing of the characteristic vector of user and article is described below.Described characteristic vector method for expressing with to
The vectorial expression method of quantity space model VSM is similar, i.e. using characteristic item as user characteristics or article characteristics
Ultimate unit.In this patent, using the set of the degree of association of user and each feature as the feature of user to
Amount, using the set of the degree of association of article and each feature as the characteristic vector of article.
Fig. 1 is the characteristic vector method for expressing that user collects each user in U.In user collects U any one
The characteristic vector of user m (m ∈ U) is set to (uwm1, uwm2..., uwmk..., uwmL), wherein said uwmkTable
Show the degree of association of described user m and feature k (k ∈ K).It addition, described user to be collected each user in U
Pool together with the degree of association of feature k, composition of vector (uw1k, uw2k..., uwMk), this vector is called user
Kth user's column vector of collection U, wherein k ∈ K.
Fig. 2 is the characteristic vector method for expressing of each article in article collection I.In article collection I any one
The characteristic vector of article n (n ∈ I) is set to (dwn1, dwn2..., dwnk..., dwnL), wherein said dwnkRepresent institute
State the degree of association of article n and described feature k (k ∈ K).It addition, by each article in described article collection I with
The degree of association of feature k pools together, composition of vector (dw1k, dw2k..., dwNk), this vector is called article collection
The kth columns of items vector of I, wherein k ∈ K.
The method to set up of the characteristic vector initial value of explanation user and article below.This patent method needs institute
State a part of user setup characteristic vector initial value that user collects in U, and in described article collection I
Part objects arranges characteristic vector initial value.The span of the characteristic vector initial value of user and article is usual
It is set to, for each i ∈ U, j ∈ I and k ∈ K, have uwik∈ [0,1] and dwjk∈ [0,1].If user or article
Characteristic vector be not provided with initial value, characteristic vector initial value is default is set to null vector for it.Below with user
The method to set up of characteristic vector initial value is described as a example by i and article j.
Example 1. arranges feature sum L=5, feature set K=(1,2,3,4,5), arrange the feature of described user i to
Amount (uwi1, uwi2, uwi3, uwi4, uwi5)=(0,0,0,1,0), i.e. user i is 1 with the degree of association of feature 4, with it
The degree of association of its feature is zero, in like manner, can arrange the characteristic vector (dw of described article jj1, dwj2, dwj3,
dwj4, dwj5) initial value.
Example 2. is submitted to one group of article set H={..., r by user i ... }The feature of article r (r ∈ H) to
Amount is (dwr1, dwr2..., dwrL), therefore for each feature k ∈ K, uw is setik=(σ/s) ∑(r∈H)dwrk,
Wherein s is the element number of described set H, and σ is preset constant.Use similar approach, it is also possible in institute
State user and collect the characteristic vector initial value selecting one group of user to calculate user i in U.
Whether each user and article are designed with each component of its characteristic vector and may be used in each iterative process
The mark being updated.From the point of view of such as the kth feature in the t time iteration, update and be masked as 1 expression
The kth component of the characteristic vector of user or article can be updated in the t time iteration, and renewal is masked as
The kth component of 0 characteristic vector representing user or article cannot be updated in the t time iteration.
Fig. 3 is the algorithm flow chart of a kind of commending system.Described algorithm is specifically included in the clothes accessing the Internet
In business device, execution following steps:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is that the multiple article in described article collection I arrange characteristic vector initial value, and is described
User collects the multiple user setup characteristic vector initial values in U；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
In method described in Fig. 3, after performing described step S4, described method also includes that judgement is described
User collects whether the characteristic vector of each user in U is updated, and to each user characteristics being updated
Vector implements the step of normalized.Assume that the characteristic vector of user m (m ∈ U) is for (uwm1, uwm2...,
uwmk..., uwmL), then as follows to the normalization processing method of the characteristic vector of user m.To each feature k
∈ K, calculates uwmk *=uwmk/∑K=1,2 ..., L uwmk, then after the characteristic vector normalization of user m,
Its numerical value becomes (uwm1 *, uwm2 *..., uwmk *..., uwmL *)。
In method described in Fig. 3, after performing described step S4, described method also includes each spy
Levy k ∈ K, user is collected the kth user's column vector in U and implements the step of normalized.Assume to use
Kth user's column vector of family collection U is (uw1k, uw2k..., uwik..., uwMk), then normalization processing method
As follows.I.e. to each user i ∈ U, calculate uwik *=uwik/∑J=1,2 ..., M uwjk, then kth user row
After vector normalization, its numerical value becomes (uw1k *, uw2k *..., uwik *..., uwMk *)。
In method described in Fig. 3, performing described step S4*Afterwards, described method also includes that judgement is described
Whether the characteristic vector of each article in article collection I is updated, and to each article characteristics being updated
Vector implements the step of normalized.If the characteristic vector of article n (n ∈ I) is (dwn1, dwn2..., dwnk,
..., dwnL), then as follows to the normalization processing method of the characteristic vector of article n.To each feature k ∈ K,
Calculate dwnk *=dwnk/∑K=1,2 ..., L dwnk, then, after the characteristic vector normalization of article n, its numerical value becomes
For (dwn1 *, dwn2 *..., dwnk *..., dwnL *)。
In method described in Fig. 3, performing described step S4*Afterwards, described method also includes each spy
Levy k ∈ K, the kth columns of items vector in article collection I is implemented the step of normalized.Assume article
The kth columns of items vector of collection I is (dw1k, dw2k..., dwik..., dwNk), then normalization processing method is as follows.
To each article i ∈ I, calculate dwik *=dwik/∑J=1,2 ..., N dwjk, then the kth columns of items of article collection I
After vector normalization, its numerical value becomes (dw1k *, dw2k *..., dwik *..., dwNk *)。
In method described in Fig. 3, the similarity calculating method of user and article is as follows.With user m and article
Illustrate as a example by n.
When feature set K is not grouped (p=1), (m n) can be set to similarity Sim of user m and article n
Sim (m, n)=[∑(k∈K)(uwmk·dwnk)]/{[∑(k∈K)(uwmk)2]1/2·[∑(k∈K)(dwnk)2]1/2}
And when feature set K contains p stack features (p >=2), first calculate user m and article n each
Sub-similarity under character subset, then goes out user m according to each sub-Similarity Measure similar with article n
Degree.Such as set feature set K={K1, K2..., Kp, wherein Ki(i=1,2 ..., p) represent the i-th of feature set K
Individual subset, and with Sim (m, n, Ki) represent that user m and article n is at character subset KiIn sub-similarity,
With Sim, (m n) represents user m and the similarity of article n, then has
Sim (m, n, Ki)=[∑(k∈Ki)(uwmk·dwnk)]/{[∑(k∈Ki)(uwmk)2]1/2·[∑(k∈Ki)(dwnk)2]1/2}
Sim (m, n)=∑I=1,2...., p{αiSim (m, n, Ki)}
The method of above-mentioned calculating similarity is cosine similarity method.It is similar to, it is also possible to similarity will be calculated
Method change other method calculating similarity such as Pearson came similarity (Pearson) into.It addition, above-mentioned calculating
The method of the similarity of user m and article n, have employed user m and article n under each character subset
The linear weighted function of sub-similarity, weight coefficient is respectively α1, α2..., αp.Nonlinear algorithm can also be used
Calculate user m and the similarity of article n, such as
Sim (m, n)=function (Sim (m, n, K1), Sim (m, n, K2) ..., Sim (m, n, Kp))
Wherein function (x1, x2..., xp) it is non-linear increasing function.
In method described in Fig. 3, the characteristic vector initial value method to set up of described step 3 includes three types:
(1) characteristic vector of the multiple article during article integrate I is as non-vanishing vector, and user collects each user's in U
Characteristic vector is null vector；(2) characteristic vector of each article in article collection I is null vector, and uses
Family integrates the characteristic vector of the multiple users in U as non-vanishing vector；(3) feature of the multiple article in article collection I
Vector is non-vanishing vector, and user integrates the characteristic vector of multiple users in U as non-vanishing vector.
In method described in Fig. 3, if performing step S4 and step S4*It is defined as an iteration,
So, described step S4 is being performed*Afterwards, described method also includes the step performing t-1 iteration again,
Wherein t >=2.And in each iterative process, it is also possible to the one in selection the following two kinds iterative manner:
Mode 1 is for performing step S4 and step S4*, mode 2 is not for perform step S4*Only carry out step S4.
Fig. 4 is a kind of idiographic flow that Fig. 3 method performs T iteration.The method achieve step S4 and
Step S4*T iteration.Often perform an iteration, be judged as whether iterations t is more than T, be to hold
Row step S5, otherwise returns step S4.
In method described in Fig. 4, by arranging iterations T and the iterative manner of each iteration, described side
Method includes following application example: (1) sets T=1 and selects iterative manner 2, then Exactly-once step S4；(2)
If T=1 and selection iterative manner 1, then perform step S4 and step S4*, (3) set T=2 and twice iteration is divided
Other selection mode 1 and mode 2, then perform step S4, step S4*With step S4, (4) set T=2 and twice
Iteration all selects mode 1, then perform step S4, step S4*, step S4 and step S4*, (5) set T=3 and
Priority selection mode 1, mode 1 and mode 2 in three iteration, then perform step S4, step S4*、
Step S4, step S4*With step S4, etc..Realize step S4 and step S4 according to this*Successive ignition.
Fig. 5 is the algorithm flow chart of a kind of commending system.Described algorithm is specifically included in the clothes accessing the Internet
In business device, execution following steps:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is that the multiple article in described article collection I arrange characteristic vector initial value, and is described
User collects the multiple user setup characteristic vector initial values in U；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
Fig. 5 method is on the basis of Fig. 3 method, exchanges step S4 and step S4 of Fig. 3*Execution suitable
Sequence and obtain.
In method described in Fig. 5, performing described step S4*Afterwards, described method also includes that judgement is described
Whether the characteristic vector of each article in article collection I is updated, and to each article characteristics being updated
Vector implements the step of normalized.Concrete methods of realizing is identical with Fig. 3 method.
In method described in Fig. 5, performing described step S4*Afterwards, described method also includes each spy
Levy k ∈ K, the kth columns of items vector in article collection I is implemented the step of normalized.Implement
Method is identical with Fig. 3 method.
In method described in Fig. 5, after performing described step S4, described method also includes that judgement is described
User collects whether the characteristic vector of each user in U is updated, and to each user characteristics being updated
Vector implements the step of normalized.Concrete methods of realizing is identical with Fig. 3 method.
In method described in Fig. 5, after performing described step S4, described method also includes each spy
Levy k ∈ K, user is collected the kth user's column vector in U and implements the step of normalized.Concrete real
Existing method is identical with Fig. 3 method.
In method described in Fig. 5, user and the similarity calculating method of article are identical with Fig. 3 method.
In method described in Fig. 5, the characteristic vector initial value method to set up of described step 3 includes three types:
(1) characteristic vector of the multiple article during article integrate I is as non-vanishing vector, and user collects each user's in U
Characteristic vector is null vector；(2) characteristic vector of each article in article collection I is null vector, and uses
Family integrates the characteristic vector of the multiple users in U as non-vanishing vector；(3) feature of the multiple article in article collection I
Vector is non-vanishing vector, and user integrates the characteristic vector of multiple users in U as non-vanishing vector.
In method described in Fig. 5, if performing step S4*It is defined as an iteration with step S4,
So, after performing described step S4, described method also includes the step performing t-1 iteration again,
Wherein t >=2.And in each iterative process, it is also possible to the one in selection the following two kinds iterative manner:
Mode 1 is for performing step S4*With step S4, mode 2 only carries out step S4 for not performing step S4*。
Fig. 6 is a kind of idiographic flow that Fig. 5 method performs T iteration.Described method achieves step S4*
T the iteration with step S4.Often perform an iteration, be judged as whether iterations t is more than T, be then
Perform step S5, otherwise return step S4*。
In method described in Fig. 6, by arranging iterations T and the iterative manner of each iteration, described side
Method includes following application example: (1) sets T=1 and selects iterative manner 2, then Exactly-once step S4*；(2)
If T=1 and selection iterative manner 1, then perform step S4*With step S4, (3) set T=2 and twice iteration is divided
Other selection mode 1 and mode 2, then perform step S4*, step S4 and step S4*, (4) set T=2 and two
Secondary iteration all selects mode 1, then perform step S4*, step S4, step S4*With step S4, (5) set T=3
And in three iteration successively selection mode 1, mode 1 and mode 2, then perform step S4*, step S4,
Step S4*, step S4 and step S4*, etc..Realize step S4 according to this*Successive ignition with step S4.
Fig. 7 is the concrete methods of realizing of step S4 of Fig. 3 and Fig. 5 method.Comprise the steps:
Step S41. reads first element in set of relations R；
Step S42., according to the described element read, obtains the mark of user therein and article；
Step S43. reads characteristic vector and the characteristic vector of described article of described user；
Step S44. uses the characteristic vector of described article to update the characteristic vector of described user；
Step S45. reads the next element in set of relations R；
Step S46. judges whether next element exists, and is then to return step S42, otherwise terminates.
Described step S44 of said method is meant that the characteristic vector of described user is the feature of described article
The function of vector.Assuming that described user is user m, described article are article n, the most described step S44
A kind of concrete methods of realizing is
Example 1. uses equation below to update the characteristic vector of user m.
uwmk=uwmk+λ1(n, m, k, t) f1(dwnk) (for each k ∈ K)
Wherein, t represents iterations, λ1(n, m, k t) represent and user m, article n, feature k and iteration
What number of times t was relevant affects coefficient, and λ1(n, m, k, t) >=0.Described f1(dwnk) it is described dwnkIncreasing function,
And f1(x) >=0, f1(0)=0.Such as f1(dwnk)=σ1·dwnkOr f1(dwnk)=σ2·(dwnk)a, wherein σ1, σ2
With a for presetting normal number.It addition, access the different modes of article for user, as bought, putting into and buy
Car, comment on and browse, it is assumed that the λ of its correspondence1(n, m, k, numerical value t) diminishes successively.
In method described in Fig. 7, if setting t1During secondary iteration, certain feature k of user m1Can not
It is updated, then for each article n, λ is set1(n, m, k1, t1)=0.If setting iterations t1≥T1Time
Certain feature k of user m1Can not be updated, then as iterations t1≥T1Time, for each article n,
λ is set1(n, m, k1, t1)=0, wherein T1For setting constant.
In method described in Fig. 6, if it is desired at t1Secondary iteration selects do not perform step S4, then a kind of
Implementation method is for each user m, article n and feature k, arranges λ1(n, m, k, t1)=0.
Fig. 8 is step S4 of Fig. 3 and Fig. 5 method*Concrete methods of realizing.Comprise the steps.
Step S41*. read first element in set of relations R；
Step S42*. according to the described element read, obtain the mark of user therein and article；
Step S43*. read characteristic vector and the characteristic vector of described article of described user；
Step S44*. use the characteristic vector of described user to update the characteristic vector of described article；
Step S45*. read the next element in set of relations R；
Step S46*. judge whether next element exists, be then to return step S42*, otherwise terminate.
Described step S44 of said method*It is meant that the characteristic vector of described article is the spy of described user
Levy the function of vector.Assuming that described user is user m, described article are article n, the most described step S44*
A kind of concrete methods of realizing be
Example 2. uses the characteristic vector of equation below more new article n.
dwnk=dwnk+λ2(m, n, k, t) f2(uwmk) (for each k ∈ K)
Wherein, t represents iterations, λ2(m, n, k t) represent and user m, article n, feature k and iteration
What number of times t was relevant affects coefficient, and λ2(m, n, k, t) >=0.Described f2(uwmk) it is described uwmkIncreasing function,
And f2(x) >=0, f2(0)=0.Such as f2(uwmk)=σ3·uwmkOr f2(uwmk)=σ4·(uwmk)a, wherein σ3, σ4
With a for presetting normal number.It addition, access the different modes of article for user, as bought, putting into and buy
Car, comment on and browse, it is assumed that the λ of its correspondence2(m, n, k, numerical value t) diminishes successively.
In method described in Fig. 8, if setting t2During secondary iteration, certain feature k of article n2Can not be by
Update, then for each user m, λ is set2(n, m, k2, t2)=0.If setting iterations t2≥T2Time,
Certain feature k of article n2Can not be updated, then as iterations t2≥T2Time, for each user m,
λ is set2(m, n, k2, t2)=0, wherein T2For setting constant.
In method described in Fig. 4, if it is desired at t2Secondary iteration selects do not perform step S4*, then one
Planting implementation method is for each user m, article n and feature k, arranges λ2(m, n, k, t2)=0.
The several application examples of said method are described below.
Application example 1
Application example 1 is an application example of Fig. 4 method.This example by study user and article characteristics,
Recommend the article of its setting quantity that may like to user, i.e. TOP N recommends.This example is included in and connects
Enter in the server of the Internet, execution following steps:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；If iterations t=1；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is that the multiple article in described article collection I arrange characteristic vector initial value；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S4a. collects the characteristic vector of each user in U and implements normalized user；
Step S4b., to each k ∈ K, collects the kth user's column vector in U and implements normalized user；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S4a*. the characteristic vector of each article in article collection I is implemented normalized；
Step D4. judges that iterations t, whether more than T, is then to perform step S5, otherwise sets iterations
T=t+1, and return step S4；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
Application example 2
Application example 2 is an application example of Fig. 6 method.This example by study user and article characteristics,
Recommend the article of its setting quantity that may like to user, i.e. TOP N recommends.This example is included in and connects
Enter in the server of the Internet, execution following steps:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；If iterations t=1；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is multiple user setup characteristic vector initial values that described user collects in U；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S4a*. the characteristic vector of each article in article collection I is implemented normalized；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S4a. collects the characteristic vector of each user in U and implements normalized user；
Step S4b., to each k ∈ K, collects the kth user's column vector in U and implements normalized user；
Step D4. judges that iterations t, whether more than T, is then to perform step S5, otherwise sets iterations
T=t+1, and return step S4*；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
Application example 3
This application example introduces a kind of method recommending to set the article of quantity to user.
Step S1. is provided with M user and forms user and collect U, N number of article form article collection I, have p group
Feature composition characteristic collection K={K1, K2..., Kp, wherein Ki(i=1,2 ..., p) it is characterized the i-th subset of collection K；
Step S2. obtains and stores set of relations R, and each element representation user in set of relations R collects in U
One user accesses the record of article in article collection I；If iterations t=1；
Step S3. is that the multiple article in article collection I arrange characteristic vector initial value, and collects U for user
In multiple user setup characteristic vector initial values；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element, implements step as follows:
Step S41. reads first element in set of relations R；
Step S42., according to the element read, obtains the mark of user therein and article；
Step S43. reads characteristic vector and the characteristic vector of described article of described user；
Step S44. uses the characteristic vector of described article to update the characteristic vector of described user, it is assumed that described use
Family is user m, and described article are article n, then concrete update method is
uwmk=uwmk+λ1(n, m, k, t) f1(dwnk) (for each k ∈ K)
Wherein, t represents iterations, λ1(n, m, k t) are and user m, article n, feature k and iteration time
What number t was relevant affects coefficient, f1(dwnk) it is described dwnkIncreasing function；
Step S45. reads the next element in set of relations R；
Step S46. judges whether next element exists, and is then to return step S42, otherwise performs step S47；
Step S47. collects the characteristic vector of each user in U and implements normalized user；
Step S48., to each k ∈ K, collects the kth user's column vector in U and implements normalized user；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；Implement step as follows:
Step S41*. read first element in set of relations R；
Step S42*. according to the element read, obtain the mark of user therein and article；
Step S43*. read characteristic vector and the characteristic vector of described article of described user；
Step S44*. use the characteristic vector of described user to update the characteristic vector of described article, it is assumed that described
User is user m, and described article are article n, then concrete update method is
dwnk=dwnk+λ2(m, n, k, t) f2(uwmk) (for each k ∈ K)
Wherein, t represents iterations, λ2(m, n, k t) represent and user m, article n, feature k and iteration
What number of times t was relevant affects coefficient, f2(uwmk) it is described uwmkIncreasing function.
Step S45*. read the next element in set of relations R；
Step S46*. judge whether next element exists, be then to return step S42*, otherwise perform step S47*；
Step S47*. the characteristic vector of each article in article collection I is implemented normalization；
Step S48*. to each k ∈ K, the kth columns of items vector in article collection I is implemented normalized；
Step D4. judges that iterations t, whether more than T, is then to perform step S5, otherwise sets iterations
T=t+1, and return step S4；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.Assume to use
Sim (m, n, Ki) represent that user m and article n is at character subset KiIn similarity, with Sim (m, n) represent
User m and the similarity of article n, then have
Sim (m, n, Ki)=[∑(k∈Ki)(uwmk·dwnk)]/{[∑(k∈Ki)(uwmk)2]1/2·[∑(k∈Ki)(dwnk)2]1/2}
Sim (m, n)=∑I=1,2...., p{αiSim (m, n, Ki)}
Wherein, α1, α2..., αpIt it is one group of weight coefficient.
The above application example is only the preferably application example of the present invention, is not limited to the present invention's
Protection domain.
Claims (9)
1. a personalized recommendation method, it is characterised in that described method is included in the service accessing the Internet
Execution following steps in device:
Step S1. obtains and stores the user being made up of multiple ID and collects U, is made up of multiple article marks
Article collection I, and feature set K being made up of multiple signature identifications；
Step S2. obtains and stores set of relations R, user described in each element representation in described set of relations R
A user in collection U accesses the record of article in described article collection I；
Step S3. is that the multiple article in described article collection I arrange characteristic vector initial value, and is described use
Multiple user setup characteristic vector initial values in the collection U of family；
Step S4. reads each element in described set of relations R, often reads an element, just uses described unit
The characteristic vector of the article in element updates the characteristic vector of the user in described element；
Step S4*. read each element in described set of relations R, often read an element, just use described unit
The characteristic vector of the user in element updates the characteristic vector of the article in described element；
Step S5. calculates appointment user and each article in described article collection I that described user collects in U
Similarity, and recommend to set the article of quantity according to described similarity to the described user of appointment.
Method the most according to claim 1, it is characterised in that after performing described step S4, institute
Method of stating also includes judging that described user collects the characteristic vector of each user in U and whether is updated, and right
Each user characteristics vector being updated implements the step of normalized.
Method the most according to claim 1, it is characterised in that after performing described step S4, institute
Method of stating also includes, to each feature k ∈ K, described user collecting the kth user's column vector in U and implementing
The step of normalized.
Method the most according to claim 1, it is characterised in that performing described step S4*Afterwards,
Described method also includes whether the characteristic vector judging each article in described article collection I is updated, and
Each article characteristics vector being updated is implemented the step of normalized.
Method the most according to claim 1, it is characterised in that performing described step S4*Afterwards,
Described method also includes, to each feature k ∈ K, implementing the kth columns of items vector in described article collection I
The step of normalized.
Method the most according to claim 1, it is characterised in that performing the most described step S4 and institute
State step S4*It is defined as an iteration, then, performing described step S4*Afterwards, described method also includes
Perform the step (t >=2) of t-1 iteration again, and in each iterative process, it is also possible to select following iteration
One in mode: mode 1 is for performing described step S4 and described step S4*, mode 2 is not for performing institute
State step S4*Only carry out described step S4.
Method the most according to claim 1, it is characterised in that exchange described step S4 and described step
S4*Execution sequence, i.e. first carry out described step S4*, described step S4 of rear execution.
Method the most according to claim 7, it is characterised in that performing the most described step S4*With
Described step S4 is defined as an iteration, then, after performing described step S4, described method is also wrapped
Include the step (t >=2) performing t-1 iteration again, and in each iterative process, it is also possible to select following changing
One in mode: mode 1 is for performing described step S4*With described step S4, mode 2 is not for perform
Described step S4 only carries out described step S4*。
Method the most according to claim 1, it is characterised in that described feature set K is first divided in advance to
Few two character subsets, therefore, are calculating the described appointment article specified in user and described article collection I
During similarity, first calculate described appointment user and the described appointment article phase under each character subset
Like degree, then, further according to each sub-similarity, calculate described appointment user and the phase of described appointment article
Like degree.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410578957.4A CN105894310A (en) | 2014-10-15 | 2014-10-15 | Personalized recommendation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410578957.4A CN105894310A (en) | 2014-10-15 | 2014-10-15 | Personalized recommendation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN105894310A true CN105894310A (en) | 2016-08-24 |
Family
ID=57000471
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410578957.4A Pending CN105894310A (en) | 2014-10-15 | 2014-10-15 | Personalized recommendation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105894310A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107977865A (en) * | 2017-12-07 | 2018-05-01 | 畅捷通信息技术股份有限公司 | Advertisement sending method, device, computer equipment and readable storage medium storing program for executing |
| CN108090807A (en) * | 2017-12-13 | 2018-05-29 | 北京小度信息科技有限公司 | Information recommendation method and device |
| CN108182621A (en) * | 2017-12-07 | 2018-06-19 | 合肥美的智能科技有限公司 | The Method of Commodity Recommendation and device for recommending the commodity, equipment and storage medium |
| CN108509626A (en) * | 2018-04-08 | 2018-09-07 | 百度在线网络技术（北京）有限公司 | Method and apparatus for verify data |
| CN109934629A (en) * | 2019-03-12 | 2019-06-25 | 重庆金窝窝网络科技有限公司 | A kind of information-pushing method and device |
| CN111797319A (en) * | 2020-07-01 | 2020-10-20 | 喜大（上海）网络科技有限公司 | Recommendation method, device, equipment and storage medium |
| CN111814032A (en) * | 2019-04-11 | 2020-10-23 | 阿里巴巴集团控股有限公司 | Cold start recommendation method and device and electronic equipment |
| US11822447B2 (en) | 2020-10-06 | 2023-11-21 | Direct Cursus Technology L.L.C | Methods and servers for storing data associated with users and digital items of a recommendation system |
-
2014
- 2014-10-15 CN CN201410578957.4A patent/CN105894310A/en active Pending
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107977865A (en) * | 2017-12-07 | 2018-05-01 | 畅捷通信息技术股份有限公司 | Advertisement sending method, device, computer equipment and readable storage medium storing program for executing |
| CN108182621A (en) * | 2017-12-07 | 2018-06-19 | 合肥美的智能科技有限公司 | The Method of Commodity Recommendation and device for recommending the commodity, equipment and storage medium |
| CN108090807A (en) * | 2017-12-13 | 2018-05-29 | 北京小度信息科技有限公司 | Information recommendation method and device |
| CN108509626A (en) * | 2018-04-08 | 2018-09-07 | 百度在线网络技术（北京）有限公司 | Method and apparatus for verify data |
| CN109934629A (en) * | 2019-03-12 | 2019-06-25 | 重庆金窝窝网络科技有限公司 | A kind of information-pushing method and device |
| CN111814032A (en) * | 2019-04-11 | 2020-10-23 | 阿里巴巴集团控股有限公司 | Cold start recommendation method and device and electronic equipment |
| CN111814032B (en) * | 2019-04-11 | 2024-05-28 | 阿里巴巴集团控股有限公司 | Cold start recommendation method and device and electronic equipment |
| CN111797319A (en) * | 2020-07-01 | 2020-10-20 | 喜大（上海）网络科技有限公司 | Recommendation method, device, equipment and storage medium |
| CN111797319B (en) * | 2020-07-01 | 2023-10-27 | 喜大（上海）网络科技有限公司 | Recommendation method, recommendation device, recommendation equipment and storage medium |
| US11822447B2 (en) | 2020-10-06 | 2023-11-21 | Direct Cursus Technology L.L.C | Methods and servers for storing data associated with users and digital items of a recommendation system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105894310A (en) | Personalized recommendation method | |
| CN110162693B (en) | Information recommendation method and server | |
| CN102789462B (en) | A kind of item recommendation method and system | |
| CN103377250B (en) | Top k based on neighborhood recommend method | |
| CN102982042B (en) | A kind of personalization content recommendation method, platform and system | |
| CN102142033B (en) | Method and device for providing relative sub-link information in search result | |
| US20130030950A1 (en) | Providing social product recommendations | |
| CN102214169B (en) | The offer method and device of key word information and target information | |
| CN105488233A (en) | Reading information recommendation method and system | |
| EP2126767A1 (en) | Intentionality matching | |
| CN107644036A (en) | A kind of method, apparatus and system of data object push | |
| CN113191838B (en) | Shopping recommendation method and system based on heterogeneous graph neural network | |
| CA3062119A1 (en) | Method and device for setting sample weight, and electronic apparatus | |
| CN103294692A (en) | Information recommendation method and system | |
| CN110197404A (en) | The personalized long-tail Method of Commodity Recommendation and system of popularity deviation can be reduced | |
| CN103744904B (en) | A kind of method and device that information is provided | |
| CN107180078A (en) | A kind of method for vertical search based on user profile learning | |
| CN109410001A (en) | A kind of Method of Commodity Recommendation, system, electronic equipment and storage medium | |
| CN112907334A (en) | Object recommendation method and device | |
| CN106708871A (en) | Method and device for identifying social service characteristics user | |
| CN110866191A (en) | Recommendation recall method, apparatus and storage medium | |
| CN104992352A (en) | Individualized resource retrieval method | |
| CN111797319A (en) | Recommendation method, device, equipment and storage medium | |
| CN111104606A (en) | Weight-based conditional wandering chart recommendation method | |
| CN110413880A (en) | Based on user's personality single classification collaborative filtering method layered |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20160824 |