CN103617289A - Micro-blog recommendation method based on user characteristics and network relations - Google Patents

Abstract

The invention relates to a micro-blog recommendation method based on user characteristics and network relations. The method includes the following steps of firstly, building and updating a network topological relation matrix, secondly, updating the network topological relation matrix, thirdly, determining the association degree among different users, and fourthly, finally determining an attention intensity matrix. According to the method, the network relations between the users and topics are built, the association degree between the users and the topics is acquired, behaviors of the users are analyzed according to the change of topic content at a period of time, the association relation is further corrected through an analysis result, topics to which the users pay attention are accurately acquired and recommended to the users, and compared with a traditional recommendation method, the method effectively improves recommendation accuracy of micro-blog topics.

Description

Microblogging recommend method based on user characteristics and cyberrelationship

Technical field

The present invention relates to the treatment technology of internet data, specifically a kind of microblogging recommend method based on user characteristics and cyberrelationship.

Background technology

Along with Internet technology, the especially development of Web2.0 network application, the social networks such as microblogging are being played the part of the role of ever more important in daily life.Compare with traditional portal website, forum, blog, in the social networks such as microblogging, information issue amount is larger, and user interactions is more frequent.The information such as daily record, microblogging, picture, state of paying close attention to good friend in social networks are all sent to associated user's homepage with the form pushing, this push mode strengthened user profile mutual in, also to user, brought the impact of information storm, so the personalized recommendation algorithm in social networks will be experienced the impact great with information interaction mass formation to improving user.

Traditional user-customized recommended algorithm is mainly by collaborative filtering and information filtering.Collaborative filtering comprises again collaborative filtering based on product and collaborative filtering based on user.The commercial product recommending algorithm of Amazon is the Typical Representative filtering based on product collaborative.It recommends thought to be, the joint purchase probability of two commodity is larger, and the relevance between corresponding goods is stronger, so utilize the incidence relation between commodity to realize recommendation.Collaborative filtering based on user is by similarity between user and item associations matrix computations user, thereby recommends targeted customer by the interested article of similar users are limited.Content-based proposed algorithm does not need the similarity of carrying out between user to calculate, and associated with the coupling of user profile according to the characteristic information of article itself, realizes the recommendation of article.

Recommend to compare with traditional article, the way of recommendation in social networks possesses 3 differences, take microblogging social networks as example:

(1) user force is for the impact of microblogging:

There is larger individual difference in the influence power of different user.In propagation relation specific to microblogging, the user that same microblogging is had Different Effects power issues, and its propagation relation is all different from Communication results.Even a microblogging from domestic consumer is had compared with the user of high-impact and forwards by one, also may greatly change relay path and the Communication results of this microblogging.

(2) user interactions relation is for the impact of microblogging:

In social networks, another important measure is the intimate degree between good friend.The phenomenon that user pays close attention to the whole microbloggings of some specific hail fellow is ubiquitous.This concern feature does not rely on any content similarity and text feature, and this is non-existent in traditional article proposed algorithm.

(3) repeat to recommend:

In traditional proposed algorithm, if user has bought certain article, just these article should not recommended targeted customer by recommended models more so.But in microblogging commending system, the main body of recommendation corresponds to a series of microblog topics after cluster, and a topic can comprise a large amount of micro-blog informations.When user is interested for one of them topic, microblogging proposed algorithm should more preferentially be recommended user by the microblogging of same topic, and this is the concept that repeats recommendation.

Therefore traditional proposed algorithm can not effectively be applied to the commending system of microblog topic, is unfavorable for the raising of microblogging recommendation topic accuracy.

Summary of the invention

Technical matters to be solved by this invention is, provides a kind of microblogging social network sites that can improve to user, to recommend to pay close attention to the microblogging recommend method based on user characteristics and cyberrelationship of the accuracy of the relevant microblogging of topic.

The step principle brief description of the microblogging recommend method based on user characteristics and cyberrelationship of the present invention is as follows:

(1) foundation of network topology matrix and renewal.

Network topology matrix is the basic foundation of carrying out user and topic correlation analysis.It is according to the simplest user of existing information foundation and the incidence relation between topic.

For example, user A issued 10 microbloggings, and wherein 9 belong to topic a, and other 1 belongs to topic b; Can find out, the microblog topic a of user A issue is 9:1 with the ratio of topic b, can illustrate that topic a has higher Preference to user A, and topic b does not possess enough representativenesses to user A.So when determining the degree of association of user and topic, will consider user's Preference, the degree of association of the topic that place ratio is large is high, just the coefficient of relationship of topic a and user A is set as to 9/10 here, and the coefficient of relationship of topic b and user A is set as 1/10.

The renewal of network topology matrix adds up topic number and microblogging number again, upgrades the coefficient of relationship of topic, if update mechanism is set as to listen mode, just user delivers new microblogging, upgrades immediately matrix, can expend a lot of resources, and is not easy to realize.So set a time threshold t,, through the t time, just again do not add up topic number and microblogging number, set up new relational matrix.

The foundation of matrix and renewal are the bases of whole algorithm.

(2) analysis of network topology matrix update.

This network topology matrix update reflects the variation of user behavior feature, and this step is exactly according to this, to change to set up the concern intensity matrix of user session topic.

Pay close attention to the concern intensity that intensity matrix represents user session topic, if user i is a to the concern intensity of topic j _ij, the capable j column element of i of matrix is a _ij.Initial concern intensity matrix is identical with network topology matrix.

If a. the neutral element in network topology matrix becomes nonzero element, this just shows that user has delivered a new topic in a time period t.

For example, user A issued 10 microbloggings, and wherein 9 belong to topic a, and other 1 belongs to topic b; After a time period t, user A has delivered again two microbloggings about topic c, the coefficient of relationship of topic c and user A is just set as 1/6 by 0 so, this variation shows, user A is interested in topic c suddenly, and likely topic c is up-to-date hot ticket, and also likely user A starts to pay close attention to topic c, now, by a paying close attention in intensity matrix _acvalue be set as 100, and keep n time period t, at once topic c is added in the recommendation list of access customer A, and recommends constantly in nt time span user A.

This is according to the instantaneity feature of user behavior, to pass judgment on the attention rate of user session topic.This behavioural characteristic is very representative, and recommends accuracy rate very high.

If b. in the time period of nt, in network topology matrix, the value of the element of respective items is totally in downtrending, show that user diminishes relatively to the attention rate of this topic, even no longer pay close attention to this topic, so even this topic is large at network topology matrix relationship coefficients comparison, also no longer representative.

User A issued 100 microbloggings, and wherein 90 belong to topic a, and other 10 belong to topic b; After one month, user A has only delivered 10 microbloggings about topic c, even if now the coefficient of relationship of topic a is 9/11, but not representative in section at this moment.

Therefore be incorporated herein a factor alpha, a is proportional to the reduction of topic a coefficient of relationship.Pay close attention to element value corresponding in intensity matrix and be made as α a _ij.

If c., in the time period of nt, in network topology matrix, the value of the element of respective items, totally in ascendant trend, shows that user becomes large relatively to the attention rate of this topic.Just in time contrary with the situation in b., introduce the recruitment that a factor alpha a is proportional to topic coefficient of relationship.

(3) degree of association between different user.

Between different user, have certain degree of association, between the high user of the degree of association, have high similarity, the similar possibility of the topic paid close attention between user is also just larger.

For example, have three user: A, B, C in network, each is with having issued per family 5 microbloggings.5 microbloggings that party A-subscriber issues all belong to topic a; In the microblogging of party B-subscriber's issue, 4 belong to topic a, and 1 belongs to topic b; In the microblogging of C user's issue, 2 belong to topic a, remain 3 and belong to topic c.Objective definition user is A, and it is carried out to microblogging recommendation, in above-mentioned hypothesis, the microblogging of user A 100% belongs to topic a, in the microblogging of user B issue, have 80% microblogging to belong to topic a, and the microblogging that belongs to topic a in user C only accounts for 40%, so user A and user B should possess higher similarity.Other microbloggings that user B issues, the microblogging that belongs to topic b is compared user C and is issued the microblogging that belongs to topic c, for user A, should have higher recommendation relation.

The relation between different user can utilize the concern intensity matrix of generation to embody, and two users are added about the product of the concern intensity of all topics.Numerical value is larger, and explanation relation is stronger.

(4) paying close attention to the final of intensity matrix determines

Suppose to exist in network A, two users of B and a, b, tri-topics of c.Wherein all microbloggings of user A issue all belong to topic a, and user B has issued 10 microbloggings, and wherein 5 belong to topic a, and 4 belong to topic b, and 1 belongs to topic c.Obviously, for user A, the topic b that user B delivers has more attractive force than topic a.

The concern intensity of current definition user A topic unpublished with it is that user A and other users' the degree of association is multiplied by the coefficient of relationship of the corresponding topic of other users, adds up and just can obtain.

The concrete technical scheme of the inventive method comprises the following steps:

1) foundation of network topology matrix and renewal;

For m microblog users, the topic that these users deliver adds up to n, sets up the matrix of a n * m;

N wherein _jbe total microblogging number of j user's issue, t (i, j) represents the microblogging number that belongs to topic i of user j issue;

Wherein

when exactly user j is recommended, the initial value of topic i;

2) network topology matrix update;

Set up and pay close attention to intensity matrix, initial concern intensity matrix B is identical with network topology matrix; Be B=A;

If a in a.T moment network topology matrix _ij(T)=0, the network topology matrix after a period of time t for element a _ij(T+t) > 0, pays close attention to the element b of correspondence position in intensity matrix B _ij(T+t)=100;

If a in b.T moment network topology matrix _ij(T) ≠ 0, through the time period of nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) < a _ij(T);

α=k ₁(a _ij(T)-a _ij(T+nt)) (k ₁for constant)

Pay close attention to the element of correspondence position in intensity matrix B

b _ij(T+nt)＝(1-α)a _ij(T+nt)；

If a in c.T moment network topology matrix _ij(T) ≠ 0, through the time period of nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) > a _ij(T).

α=k ₂(a _ij(T+nt)-a _ij(T)) (k ₂for constant)

Pay close attention to the element of correspondence position in intensity matrix B

b _ij(T+nt)＝(1+α)a _ij(T+nt)；

3) the determining of the degree of association between different user;

For user A, be shown with the association table of other user j

$R_{\mathrm{Aj}}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{b}_{\mathrm{iA}}{b}_{\mathrm{ij}},(1\&NotEqual;A)$

B wherein _iArepresent the concern intensity between user A and topic i, b _ijfor the concern intensity between user j and i topic; By this formula, can calculate user A and every other user's the degree of association;

4) paying close attention to the final of intensity matrix determines

Suppose that user j has delivered topic i, but user A does not deliver.User A can be expressed as for the concern intensity of topic i:

$b_{\mathrm{iA}}=\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{R}_{\mathrm{Aj}}{k}_{\mathrm{ij}}.$

Described constant k ₁k ₂preparation method be:

Carry out following test:

Select the targeted customer of some as test source user, capture some (as 100) microbloggings of the up-to-date issue of source user and the review information of microblogging; The 1-layer-user user profile that crawl source user is paid close attention to and some (as 200) microbloggings and the microblogging review information of the up-to-date issue of 1-layer-user user; Capture the bean vermicelli user of all 1-layer-user, i.e. 2-layer-user information simultaneously; Above data have formed the raw data set of test;

In test, training set is used for training constant k ₁k ₂; The parameter k that test set utilizes training set to obtain ₁k ₂calculate realistic model Output rusults, with the recommendation effect of true online data comparative evaluation model in test set; Test utilizes all test source users and them to pay close attention to the comment relation of user institute issuing microblog, and with time-sequencing, the microblogging comment relation of choosing front certain proportion (as front 80%) enters test training set, in order to training parameter k ₁k ₂, the microblogging comment relation of all the other (as rear 20%) enters experimental test collection, is used for the recommendation effect of evaluation algorithms;

In training process, parameter k ₁k ₂in [0,1] span, change, for each k ₁k ₂value is calculated respectively concern intensity, selects the best k of effect ₁k ₂as model parameter training result.

The present invention is by having built the cyberrelationship between user and topic, obtain the correlation degree between user and topic, by the mutation analysis user behavior of topic content in a period of time, utilize the result of analyzing further to revise incidence relation, thereby obtain accurately the topic that user pays close attention to, user is recommended, compare with traditional recommend method, the method has effectively improved the accuracy that microblog topic is recommended.

Accompanying drawing explanation

Fig. 1 is the schematic diagram that the embodiment of the present invention is assessed and compared with traditional NBI proposed algorithm.

Embodiment

The embodiment of the inventive method comprises the following steps:

1) foundation of network topology matrix and renewal;

For m microblog users, the topic that these users deliver adds up to n, sets up the matrix of a n * m;

N wherein _jbe total microblogging number of j user's issue, t (i, j) represents the microblogging number that belongs to topic i of user j issue;

Wherein

when exactly user j is recommended, the initial value of topic i;

2) network topology matrix update;

Set up and pay close attention to intensity matrix, initial concern intensity matrix B is identical with network topology matrix; Be B=A;

If a in a.T moment network topology matrix _ij(T)=0, the network topology matrix after a period of time t for element a _ij(T+t) > 0, pays close attention to the element b of correspondence position in intensity matrix B _ij(T+t)=100;

If a in b.T moment network topology matrix _ij(T) ≠ 0, through the time period of nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) < a _ij(T);

α=k ₁(a _ij(T)-a _ij(T+nt)) (k ₁for constant)

Pay close attention to the element of correspondence position in intensity matrix B

b _ij(T+nt)＝(1-α)a _ij(T+nt)；

If a in c.T moment network topology matrix _ij(T) ≠ 0, through the time period of nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) > a _ij(T).

α=k ₂(a _ij(T+nt)-a _ij(T)) (k ₂for constant)

Pay close attention to the element of correspondence position in intensity matrix B

b _ij(T+nt)＝(1+α)a _ij(T+nt)；

3) the determining of the degree of association between different user;

For user A, be shown with the association table of other user j

$R_{\mathrm{Aj}}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{b}_{\mathrm{iA}}{b}_{\mathrm{ij}},(1\&NotEqual;A)$

B wherein _iArepresent the concern intensity between user A and topic i, b _ijfor the concern intensity between user j and i topic; By this formula, can calculate user A and every other user's the degree of association;

4) paying close attention to the final of intensity matrix determines

Suppose that user j has delivered topic i, but user A does not deliver.User A can be expressed as for the concern intensity of topic i:

$b_{\mathrm{iA}}=\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{R}_{\mathrm{Aj}}{k}_{\mathrm{ij}}.$

Described constant k ₁k ₂preparation method be:

Carry out following test:

Test utilizes many app key and accept token to authorize Sina's microblogging data is gathered.Data set as test source user, captures the review information of 200 microbloggings and the microblogging of the up-to-date issue of source user by 68 targeted customers; Capture 200 microbloggings and the microblogging review information of 1-layer-user user profile that source user pays close attention to and the up-to-date issue of 1-layer-user user; In order to calculate the user force index of 1-layer-user, need to capture the bean vermicelli user of all 1-layer-user, i.e. 2-layer-user information simultaneously.Above data have formed the raw data set of this test;

In test, training set is used for training constant k ₁k ₂; The parameter k that test set utilizes training set to obtain ₁k ₂calculate realistic model Output rusults, with the recommendation effect of true online data comparative evaluation model in test set; Test utilizes 68 potential source users and them to pay close attention to the comment relation of user institute issuing microblog, and with time-sequencing, front 80% the microblogging comment relation chosen enters test training set, in order to training parameter k ₁k ₂, rear 20% microblogging comment relation enters experimental test collection, is used for the recommendation effect of evaluation algorithms, so training set and test set ratio data are 8:2 grouping.

In training process, parameter k ₁k ₂in [0,1] span, change, for each k ₁k ₂value is calculated respectively concern intensity, selects the best k of effect ₁k ₂as model parameter training result.

Adopt the algorithm of RS to assess above-described embodiment:

In proposed algorithm evaluation criteria, recommendation list length L has important impact to recommendation effect.After recommended models is paid close attention to all microbloggings that are present in test set intensity calculating, select the microblogging of the front L of rank to recommend targeted customer.If the value of L is larger, what microblogging was recommended is comprehensive, and recall ratio is also just higher, growth along with L, the interested most microbloggings of user can both be recommended targeted customer by model, but the garbage wherein comprising is also inevitable, increase thereupon, have reduced whole recommendation effect.If the value of L is less, in recommendation list, each microblogging has higher possibility to be paid close attention to by targeted customer, and microblogging recommends accuracy higher, but recall ratio can correspondingly reduce.No matter accuracy or recall ratio, all there is close contacting with the length L of recommendation list, the simultaneously selection of L needs again to consider the corresponding microblogging comment matrix size of user and density, so the recommendation effect that accuracy and recall ratio all cannot a kind of proposed algorithms of independent assessment.

RS passes judgment on the calculating that each micro-blog information that first may touch targeted customer by recommended models is paid close attention to intensity, by final concern matrix, result of calculation is sorted by height by concern intensity herein, pay close attention to the microblogging that intensity is identical and sort in no particular order.Test captures 200 microbloggings of every up-to-date issue of 1-layer-user, and RS standard is given a mark and sorted each test microblogging, and the RS scoring of every microblogging is as shown in formula 4.28, and the RS of system marks as shown in formula 4.29:

RS(t)＝P(t)/L (4.28)

$\mathrm{RS}=\frac{1}{n}\underset{t=1}{\overset{n}{\mathrm{\&Sigma;}}}\mathrm{RS}\left(t\right)---\left(4.29\right)$

In formula 4.28, t represents a micro-blog information in test set, and P (t) is this microblogging for specific user's recommendation score sequence, and L is relevant microblogging total number.For example: in test set, have 1000 microbloggings relevant to targeted customer A, first utilize recommended models that these 1000 microbloggings are marked and sorted.In reality, have 3 microbloggings finally by user A, to be commented on, and the sequence of these 3 microbloggings after 1000 microblogging marking is respectively in the 10th, the 50th and the 100th, utilizes so formula 4.28 calculating wall scroll microblogging RS scorings to be: 10/1000=0.01; 50/1000=0.05; 100/1000=0.1.So the total RS of system calculates mark: (0.01+0.05+0.1)/3=0.0533.As can be seen here, the microblogging comment existing in True Data is closed and is tied up in model recommendation score ranking results, and it is more forward that rank is positioned at recommendation list, and the RS evaluation of system is little, and the recommendation effect of corresponding model is also just better.

Its good friend's quantity of the targeted customer that this test is chosen is all greater than 200, selects L=1000.The result of comparing with traditional NBI proposed algorithm, as shown in Figure 1.Compare with traditional recommend method, this present invention has obvious advantage.

Claims (2)

1. the microblogging recommend method based on user characteristics and cyberrelationship, is characterized in that: comprises the following steps,

1) foundation of network topology matrix and renewal;

For m microblog users, the topic that these users deliver adds up to n, sets up the matrix of a n * m;

N wherein _fbe total microblogging number of j user's issue, t (i, j) represents the microblogging number that belongs to topic i of user j issue;

Wherein

when exactly user j is recommended, the initial value of topic i;

2) network topology matrix update;

Set up and pay close attention to intensity matrix, initial concern intensity matrix B is identical with network topology matrix; Be B=A;

If a in a.T moment network topology matrix _ij(T)=0, the network topology matrix after a period of time t for element a _ij(T+t) > 0, pays close attention to the element b of correspondence position in intensity matrix B _ij(T+t)=100;

If a in b.T moment network topology matrix _ij(T) ≠ 0, through the time period of nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) < a _ij(T);

α=k ₁(a _ij(T)-a _ij(T+nt)) (k ₁for constant)

Pay close attention to the element of correspondence position in intensity matrix B

$b_{\mathrm{ij}}(T+\mathrm{nt})=(1-a)a_{\mathrm{ij}}(T+\mathrm{nt});$

If a in c.T moment network topology matrix _ij(T) ≠ 0 time period through nt, in network topology matrix, the value of the element of respective items becomes a _ij(T+nt), if a _ij(T+nt) > a _ij(T).

α=k ₂(a _ij(T+nt)-a _ij(T)) (k ₂for constant)

Pay close attention to the element of correspondence position in intensity matrix B

b _ij(T+nt)＝(1+α)a _ij(T-nt)；

3) the determining of the degree of association between different user;

For user A, be shown with the association table of other user j $R_{\mathrm{Aj}}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{b}_{\mathrm{iA}}{b}_{\mathrm{ij}},(1\&NotEqual;A)$

B wherein _iArepresent the concern intensity between user A and topic i, b _ijfor the concern intensity between user j and i topic; By this formula, can calculate user A and every other user's the degree of association;

4) paying close attention to the final of intensity matrix determines

Suppose that user j has delivered topic i, but user A does not deliver.User A can be expressed as for the concern intensity of topic i:

$b_{\mathrm{iA}}=\underset{j=1}{\overset{m}{\mathrm{\&Sigma;}}}{R}_{\mathrm{Aj}}{k}_{\mathrm{ij}}.$

2. the microblogging recommend method based on user characteristics and cyberrelationship according to claim 1, is characterized in that: described constant k ₁k ₂preparation method be,

Carry out following test:

Select the targeted customer of some as test source user, capture some microbloggings of the up-to-date issue of source user and the review information of microblogging; The 1-layer-user user profile that crawl source user is paid close attention to and some microbloggings and the microblogging review information of the up-to-date issue of 1-layer-user user; Capture the bean vermicelli user of all 1-layer-user, i.e. 2-layer-user information simultaneously; Above data have formed the raw data set of test;

In test, training set is used for training constant k ₁k ₂; The parameter k that test set utilizes training set to obtain ₁k ₂calculate realistic model Output rusults, with the recommendation effect of true online data comparative evaluation model in test set; Test utilizes all test source users and them to pay close attention to the comment relation of user institute issuing microblog, and with time-sequencing, the microblogging comment relation of choosing front certain proportion (as front 80%) enters test training set, in order to training parameter k ₁k ₂, the microblogging comment relation of all the other (as rear 20%) enters experimental test collection, is used for the recommendation effect of evaluation algorithms;

In training process, parameter k ₁k ₂in [0,1] span, change, for each k ₁k ₂value is calculated respectively concern intensity, selects the best k of effect ₁k ₂as model parameter training result.