CN108334645A - One kind feeding back newer activity recommendation method based on graph model - Google Patents
One kind feeding back newer activity recommendation method based on graph model Download PDFInfo
- Publication number
- CN108334645A CN108334645A CN201810317533.0A CN201810317533A CN108334645A CN 108334645 A CN108334645 A CN 108334645A CN 201810317533 A CN201810317533 A CN 201810317533A CN 108334645 A CN108334645 A CN 108334645A
- Authority
- CN
- China
- Prior art keywords
- node
- activity
- graph model
- probability
- user
- 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
- 230000000694 effects Effects 0.000 title claims abstract description 141
- 238000000034 method Methods 0.000 title claims abstract description 36
- 239000011159 matrix material Substances 0.000 claims description 57
- 230000007704 transition Effects 0.000 claims description 42
- 238000005295 random walk Methods 0.000 claims description 18
- 239000013598 vector Substances 0.000 claims description 11
- 238000010276 construction Methods 0.000 claims description 8
- 238000013508 migration Methods 0.000 claims description 6
- 230000005012 migration Effects 0.000 claims description 6
- 238000012546 transfer Methods 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 description 22
- 238000010586 diagram Methods 0.000 description 8
- 238000011156 evaluation Methods 0.000 description 5
- 235000010627 Phaseolus vulgaris Nutrition 0.000 description 3
- 244000046052 Phaseolus vulgaris Species 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 238000001914 filtration Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000005303 weighing Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000013100 final test Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000008707 rearrangement Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
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
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Primary Health Care (AREA)
- Marketing (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Health & Medical Sciences (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses one kind feeding back newer activity recommendation method based on graph model, the method includes：On the basis of carving given M class data types at the beginning, each data is set as a node, N number of node is obtained, then this N number of node is divided into M parts according to data type, obtain the node set of M types, and according to the company side between the relevance structure node between each node, the graph model comprising N number of node and adjacent side is finally obtained, activity recommendation is carried out according to successive ignition and convergent probability on graph model.Feedback information is got in following instant, i.e., the frontier juncture system of company between node changes, and chooses changed graph model part and is ranked up again to activity to be recommended, and completes last recommendation in conjunction with first recommendation results.The present invention cannot can in time adjust after getting field feedback and improve the problem of recommending accuracy rate in the existing activity recommendation of effective solution.
Description
Technical field
The invention belongs to the proposed algorithm fields in commending system, are fed back more based on graph model more particularly, to one kind
New activity recommendation method is used for the activity recommendation of social networks.
Background technology
In recent years, major internet platform is widely applied to be based on movable society in efficient development Below-the-line
Hand over network.The participation how to allow user more convenient by movable social networks and oneself interested activity is collected,
Commending system comes into being in this case.It constructs user interest mould by analyzing a large number of users data being collected into
Type predicts movable scoring of each user to having neither part nor lot in, and finally recommends the highest activity of pre-judging score to user.Mainstream at present
Activity recommendation algorithm include traditional proposed algorithm and the proposed algorithm based on graph model.
Traditional proposed algorithm includes content-based recommendation algorithm and proposed algorithm based on collaborative filtering and by two
The algorithm that person combines.The core of content-based recommendation algorithm is that the activity description of participation is similar in the past to user for user's recommendation
Activity.The algorithm needs to extract content characteristic information from the text message of activity or user.Current text feature
Extractive technique is highly developed, and document subject matter generates model (Latent Dirichlet Allocation, LDA) and word frequency-
The calculating such as inverse document frequency (term frequency-inverse document frequency, TF-IDF) can be used for solving
Text character extraction problem.Proposed algorithm based on collaborative filtering then utilizes the history score information of user, calculates user's scoring
The similarity of vector, finds and the most like K of target user is a " seed user ", then utilizes this K user for candidate item
The weighted value of mesh scoring predicts scoring of the user to it.But above two algorithm cuts both ways, content-based recommendation is calculated
In the case of method has neglected the social networks of user itself, and the proposed algorithm based on collaborative filtering can not solve no historical data
Cold start-up recommend problem.To solve above-mentioned drawback, it is suggested in conjunction with the mixing proposed algorithm that above two is calculated.The algorithm from
Two dimensions of content and user social contact relationship consider that user to movable level of interest to be recommended, finally considers two again respectively
Kind in the case of interest and complete to recommend.
The core of proposed algorithm based on graph model is that structure one includes all related informations in movable social networks
Figure.In this way convert activity recommendation problem to the probabilistic forecasting problem of graph model interior joint.It is proposed at present based on
The proposed algorithm of graph model includes proposed algorithm based on path and the Random Walk Algorithm that band restarts.Band restart with
In machine migration algorithm, the probability distribution that each node is last in entire graph model is estimated using markovian convergence.
But based on the algorithm of graph model since the complexity of graph model leads to the effect for rebuilding graph model when system update information
Rate is low.
In recent years, with the development of internet, movable social networks (Event-Based Social are based on
Network, EBSNs), such as external meetup and the domestic same city of bean cotyledon, obtain extensive development.These websites are to use
Family and the activity side of holding provide a convenient line upper mounting plate and can participate in and share activity.Then, individual character in social networks
The key areas that the activity recommendation of change also just becomes industrial quarters and academia pays close attention to jointly, while but also traditional pushes away
The method of recommending is faced with more challenges.
Invention content
For the disadvantages described above or Improvement requirement of the prior art, it is newer based on graph model feedback that the present invention provides one kind
Thus activity recommendation method solves in the existing activity recommendation technology based on graph model, structure again completely is needed when updating the data
The technical issues of building graph model and re-starting random walk, decline so as to cause efficiency.
To achieve the above object, the present invention provides one kind feeding back newer activity recommendation method based on graph model, including：
The information of activity, user group and sponsor that the label concentrated according to preset data is marked, passes through non-weighting
Similar tags cluster is several themes by group average method, as theme node S；
It is that several are lived before each activity matching is most like according to the movable characteristic attribute that the preset data is concentrated
It is dynamic, and the oriented even side between the adjacent activity of construction activities；
It carves at the beginning, using U, E, H, G, initial graph model PG is constructed on the company side between S nodes and each node, and
The random walk restarted from the band of target user's node is carried out on the initial graph model PG, will obtain active node
Scoring of the convergent probability as the target user to movable first moment, wherein U, E, H, G indicate described pre- respectively
If the node that the user, activity, sponsor and user group in data set are constituted；
In subsequent time, according to the feedback information of the target user, structure feedback graph model FG, in the feedback artwork
The random walk restarted from the band of target user's node is carried out on type FG, will obtain the convergent probability of active node
Scoring as the target user to movable second moment, wherein the feedback graph model FG includes user and activity
Company side between node and user and activity；
The initial artwork is obtained by the comentropy of the initial graph model PG and the comentropy of the feedback graph model FG
The weight that type PG and the feedback graph model FG influence recommendation results, and by the weight, the scoring at first moment
And the score in predicting target user at second moment recommends to each movable level of interest, and to the target user
Preceding several activities.
Preferably, the information of the label concentrated according to preset data is marked activity, user group and sponsor,
It is several themes to be clustered similar tags by non-weighting group average method, including：
Select two most like Label Mergings for a target cluster from all labels；
Using the target cluster as new label, and calculate the similarity between the target cluster and other labels, wherein should
Similarity is equal to the average value of other labels and all label similarities in the target cluster, and executes described from all labels
Select two most like Label Mergings for a target cluster, until the quantity of the cluster finally clustered, which meets, presets number of clusters amount.
Preferably, the movable characteristic attribute concentrated according to the preset data is that each activity matching is most like
Several preceding activities, and the oriented even side between the adjacent activity of construction activities, including：
Adjacency matrix, activity and the activity venue of adjacency matrix, activity and activity time that activity is spent with activity
Adjacency matrix and activity are connected with the adjacency matrix of Activity Type, obtain movable eigenmatrix；
Similarity between each activity is calculated by the row vector corresponding to movable eigenmatrix；
Be several activities before each movable matching similarity is highest according to the similarity between each activity, and build from
Goal activities are directed toward the directed edge of similar active.
Preferably, the target user is to the scoring at movable first moment：Wherein,Indicate that target is used
Q-th of the activity in u pairs of family is in moment t0Scoring, M indicates movable number.
Preferably, the random walk that the band on the initial graph model PG restarts is expressed as：
h(j+1)=αEHe(j)PEH+(1-αEH)s(j)PSH
g(j+1)=αUGu(j)PUG+(1-αUG)s(j)PSG
s(j+1)=αGSg(j)PGS+αHSh(j)PHS+(1-αGS-αHS)e(j)PES
Wherein, u(j), e(j), h(j), g(j), s(j)It is probability vector of all kinds of nodes during iteration j,WithIndicate t0Moment user node is to active node and active node to the transition probability matrix of user node, αEUIt indicates from work
Dynamic node is transferred to the weight of itself probability of user node shared by the probability of user node, αGUIt indicates to be transferred to use from small group node
Family node probability accounts for the weight of user node itself probability, PGUIndicate transition probability matrix of the small group node to user node, αUE
It indicates to be transferred to the weight that active node probability accounts for active node itself probability, α from user nodeHEIt indicates to turn from sponsor's node
Move on to the weight that active node probability accounts for active node itself probability, PHEIndicate transition probability of sponsor's node to active node
Matrix, αSEIt indicates to be transferred to the weight that active node probability accounts for active node itself probability, P from theme nodeSEIndicate theme section
Put the transition probability matrix to active node, PEEIndicate transition probability matrix of the active node to active node, αEHIt indicates from work
Dynamic node is transferred to the weight that sponsor's node probability accounts for itself probability of sponsor's node, PEHIndicate that active node is saved to sponsor
The transition probability matrix of point, PSHIndicate theme node to the transition probability matrix of sponsor's node, αUGIt indicates to turn from user node
Move on to the weight that small group node probability accounts for itself probability of small group node, PUGIndicate transition probability square of the user node to small group node
Battle array, PSGIndicate theme node to the transition probability matrix of small group node, αGSIt indicates to be transferred to theme node probability from small group node
Account for the weight of itself probability of theme node, PGSIndicate small group node to the transition probability matrix of theme node, αHSIt indicates from sponsoring
Fang Jiedian is transferred to the weight that theme node probability accounts for itself probability of theme node, PHSIndicate sponsor's node to theme node
Transition probability matrix, PESIndicate transition probability matrix of the active node to theme node.
Preferably, the progress on the feedback graph model FG is restarted from the band of target user's node
Random walk will obtain scoring of the convergent probability of active node as the target user to movable second moment, packet
It includes：
SettingWithIt is t respectivelykThe transition probability matrix at moment, the two transition probability matrixs are respectively by abutting
MatrixWithIt is normalized into every trade, whereinWithTwo matrixes transposed matrix each other,In it is each
Row represents corresponding user with all movable participation relationships, and the representation of activity participated in is 1, and the representation of activity having neither part nor lot in is 0；
ByWithIt is carried out for FG graph models random
Migration, until to target user, the convergence in probability of active node obtains the target user to each movable second moment
Scoring Indicate target
U pairs of q-th of activity of user is in moment tkScoring.
Preferably, described that institute is obtained by the comentropy of the initial graph model PG and the comentropy of the feedback graph model FG
The weight that initial graph model PG and the feedback graph model FG influence recommendation results is stated, including：
ByThe comentropy of the initial graph model PG is obtained, byIt obtains described
Feed back the comentropy of graph model FG, wherein rijIndicate from
Node niTo node njThe probability of transfer, NPGIndicate the number of nodes that the initial graph model PG includes, NFGIndicate the feedback
The number of nodes that graph model FG includes；
ByThe initial graph model PG and the feedback graph model FG are obtained to recommending
As a result the weight influenced.
Preferably, it is described by the weight, the scoring at first moment and the scoring at second moment it is pre-
Target user is surveyed to each movable level of interest, including：
ByPredict target user to each movable interest
Degree.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain down and show
Beneficial effect：
(1) it improves and recommends accuracy rate：It is that each activity matches by the movable characteristic attribute concentrated according to preset data
Several most like preceding activities, and the oriented even side between the adjacent activity of construction activities, can build more practical
The graph model of expression activity social networks.Recessive even side between the activity built in the graph model not only avoids increase activity
Occur the problem of hanging node reduces graph model degree of communication caused by characteristic node, and effectively enhances content between activity
Similar weighing factor, to improve the accuracy of activity recommendation, wherein hanging node is the class being only connected with a kind of node
Node.By the scoring and the score in predicting target user at second moment by weight, first moment to each movable
Level of interest, and recommend preceding several activities to target user, predict that user is emerging respectively with feedback information in conjunction with initial information
Interest as a result, simultaneously dynamically adjusting different predictions according to initial graph model and the feedback graph model information content that includes in different moments
As a result weight, and then improve and recommend movable accuracy.
(2) reduce the workload for rebuilding graph model：It is restarted as a result of the band for feeding back graph model random
Migration, in the feedback information of system update user, since the most information in initial graph model is to retain constant, institute
To rebuild graph model just for the relationship occurred between the user that update changes and activity, and only in this section on graph model
Carry out random walk.Therefore reduce structure graph model and carry out the workload of random walk.
(3) influence of the feedback information to recommendation results can dynamically be adjusted：Pass through the comentropy by initial graph model PG
Initial graph model PG is obtained with the comentropy of feedback graph model FG and feeds back the weight that graph model FG influences recommendation results, system
Can by feed back the information content for including in different time in graph model number, feedback information pair when dynamic adjustment is recommended every time
The weighing factor of recommendation results.
Description of the drawings
Fig. 1 is a kind of flow signal for feeding back newer activity recommendation method based on graph model provided in an embodiment of the present invention
Figure；
Fig. 2 is the schematic diagram for connecting side between a kind of construction activities provided in an embodiment of the present invention；
Fig. 3 is a kind of initial graph model and feedback diagram model structure schematic diagram that embodiment provides；
Fig. 4 is a kind of control for comentropy in initial graph model of each moment and feedback graph model that embodiment provides
Table；
Fig. 5 is the present invention and other comparison diagrams of 5 kinds of way of contrast in 1 evaluation indexes of P@；
Fig. 6 is the present invention and other comparison diagrams of 5 kinds of way of contrast in MAP evaluation indexes；
Fig. 7 is the present invention and other comparison diagrams of 5 kinds of way of contrast in Recall evaluation indexes；
Fig. 8 is the present invention and other comparison diagrams of 5 kinds of way of contrast in F1 evaluation indexes.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below
It does not constitute a conflict with each other and can be combined with each other.
Fed back the present invention is based on graph model newer activity recommendation method Integral Thought be this method first against
Data set carries out every pretreatment, is allowed to be suitable for structure graph model in next step.Then structure is carved at the beginning comprising all first
The initial graph model of beginning information, and random walk is carried out on initial graph model.After getting field feedback, update is used
Relationship and structure between family and activity only include a feedback graph model for User Activity relationship, are then carried out on feedback graph model
Random walk.The information content for including according to initial graph model and feedback graph model finally, adds random walk result twice
Weight average.
Include the following steps as shown in Figure 1, feeding back newer activity recommendation method the present invention is based on graph model：
(1) in given bean cotyledon data set, by comprising user, activity, sponsor, user group and tag types
Data setting is different node U, E, H, G, T, and the characteristic attribute that activity includes in addition is set as：Activity time Em, actively
Point Er, activity cost Ec, Activity Type Et；
(2) data prediction is carried out for movable characteristic attribute, such as：Activity time carries out according to the number of weeks held
It divides and increases and all hold this case daily, be set as 8 timing nodes, activity venue is corresponded to according to the GPS coordinate of mark
Practical administrative division divided, it is m section that activity, which is spent according to consuming capacity interval division, and corresponding m spends node,
Each type set is then corresponding node by Activity Type；
(3) label is clustered, the case where the activity marked according to label, user group and sponsor, by non-
It is k theme that group average method, which is weighted, by similar tags cluster, as theme node S；
The advantages of this step, is：By being clustered to label so that unrelated single label can on surface
Other labels that there is recessive relationship with it are found, the quantity of label node is reduced simultaneously also by cluster, reduces artwork
The complexity of type.
In an optional embodiment, specifically, step (3) includes：
(3.1) select the most similar two Label Mergings for a cluster from all labels；
(3.2) using the cluster newly obtained as new label, and the similarity between the cluster and other labels is calculated, the similarity
It is equal to, the average value of other labels and all label similarities in the cluster；
(3.3) repeat the above steps (3.1) and (3.2), until the quantity of the cluster finally clustered is met the requirements.
(4) as shown in Fig. 2, being that each activity matches the most similar preceding n activity according to movable 4 characteristic attributes,
And the oriented even side between the adjacent activity of construction activities, it is directed toward neighbouring activity from activity；
The advantages of this step, is：By the recessive even side between movable content information construction activities and activity, not only
The problem of avoiding and occur hanging node in graph model, reducing graph model connectivity, and enhance between activity and activity
Relevance.
In an optional embodiment, specifically, step (4) includes：
(4.1) it enablesExpression activity is spent with activity respectively, the activity time, activity venue
And the adjacency matrix of Activity Type node；
(4.2) the above-mentioned adjacency matrix of series connection, obtains movable eigenmatrix AE；
(4.3) calculating activity EiWith movable EjBetween similarityWherein,It is
AEThe row vector of i-th row,Indicate AEThe row vector of jth row..；
(4.4) be the highest preceding n activity of each movable matching similarity according to the similarity between activity, and structure from
Goal activities are directed toward the directed edge of similar active.
(5) as shown in figure 3, carving t at the beginning0, utilize above-mentioned U, E, H, G, company's side structure between S nodes and each node
Initial graph model PG (primary graph) is built out, A is enabledMNIndicate the adjacency matrix of M type nodes and N type nodes, M, N ∈
{ U, E, H, G, S }, wherein if AMN(m, n)=1 shows there are dominant or recessive relationship between node n and node m, no
Then AMN(m, n)=0；
(6) as shown in figure 3, through being located at moment t after a period of timekWhen, the target user's that is got according to system is anti-
The feedback that feedforward information, i.e. target user make activity, such as new activity, structure feedback graph model FG (feedback
Graph), this graph model includes only the company side between user and active node and user and activity.It enablesIndicate tkMoment
The adjacency matrix of user node and active node, as user u1In t0~tkMovable e has been participated in period1, then
It can be updated to 1 from 0, finally obtain new adjacency matrix
(7) in t0After moment has built PG, the random trip restarted from the band of target user's node is carried out on PG
It walks, will finally obtain scoring of the convergent probability of active node as target user to movable first moment
In an optional embodiment, specifically, step (7) includes：
(7.1) adjacency matrix between all kinds of nodes is normalized into every trade, by AMNIt is converted into transition probability matrix PMN,
M, N ∈ { U, E, H, G, S }, whereinWithIndicate t0Moment user node is saved to active node and active node to user
Point transition probability matrix；
(7.2) it defines user and restarts moving vector qu, to target user ujIf quThe serial number i of middle element is equal to user uj's
Serial number j, then qu(i)=1 it is otherwise, 0；
(7.3) the probability vector of definition user, activity, user group, sponsor and theme is u(0),e(0),g(0),h(0),s(0), and random initializtion is carried out to these vectors；
(7.4) the random walk formula that band restarts is as follows：
h(j+1)=αEHe(j)PEH+(1-αEH)s(j)PSH (3)
g(j+1)=αUGu(j)PUG+(1-αUG)s(j)PSG (4)
s(j+1)=αGSg(j)PGS+αHSh(j)PHS+(1-αGS-αHS)e(j)PES (5)
U in above-mentioned formula(j), e(j), h(j), g(j), s(j)It is probability vector of all kinds of nodes during iteration j, (j
=0,1 ...), after successive ignition, to target user, the convergence in probability of active node obtains target user to each activity
First moment scoringWherein,Indicate mesh
U pairs of q-th of activity of user is marked in moment t0Scoring, M indicates movable number, αEUIt indicates to be transferred to user's section from active node
The weight of user node itself probability, α shared by the probability of pointGUIndicate that being transferred to user node probability from small group node occupies family section
The weight of itself probability of point, PGUIndicate transition probability matrix of the small group node to user node, αUEIt indicates to shift from user node
The weight of active node itself probability, α are accounted for active node probabilityHEExpression is transferred to active node probability from sponsor's node and accounts for
The weight of active node itself probability, PHEIndicate transition probability matrix of sponsor's node to active node, αSEIt indicates from theme
Node is transferred to the weight that active node probability accounts for active node itself probability, PSEIndicate transfer of the theme node to active node
Probability matrix, PEEIndicate transition probability matrix of the active node to active node, αEHIt indicates to be transferred to sponsor from active node
Node probability accounts for the weight of itself probability of sponsor's node, PEHIndicate transition probability matrix of the active node to sponsor's node,
PSHIndicate theme node to the transition probability matrix of sponsor's node, αUGIt indicates to be transferred to small group node probability from user node
Account for the weight of itself probability of small group node, PUGIndicate user node to the transition probability matrix of small group node, PSGIndicate theme section
Put the transition probability matrix to small group node, αGSIndicate that being transferred to theme node probability from small group node accounts for theme node itself generally
The weight of rate, PGSIndicate small group node to the transition probability matrix of theme node, αHSIt indicates to be transferred to theme from sponsor's node
Node probability accounts for the weight of itself probability of theme node, PHSIndicate sponsor's node to the transition probability matrix of theme node, PES
Indicate transition probability matrix of the active node to theme node.
(8) in tkMoment, according to updated target user's feedback data build feedback graph model FG, on FG carry out from
The random walk that the band that target user's node sets out restarts will finally obtain the convergent probability of active node as target user
Scoring to movable second moment
In an optional embodiment, specifically, step (8) includes：
(8.1) it setsWithIt is t respectivelykThe transition probability matrix at moment, the two transition probability matrixs difference
By adjacency matrixWithIt is normalized into every trade, whereinWithTwo matrixes transposed matrix each other,
In corresponding user and all movable participation relationships are represented per a line, the representation of activity participated in is 1, the movable table having neither part nor lot in
It is shown as 0；
(8.2) the random walk formula for being directed to FG graph models is as follows：
After carrying out successive ignition, each active node is obtained to the convergent probability of target user, is denoted as Indicate u couples of target user
Q-th of activity is in moment tkScoring.
(9) initial graph model PG and t is calculated separatelykMoment feeds back the comentropy H of graph model FGPG(t0) and HFG(tk), figure
The calculation process of the comentropy of model is as follows：
The comentropy of whole graph model is the sum for the nodal information entropy for including, the i.e. comentropy of PG graph modelsSimilarly,Wherein NPGAnd NFGThe number of nodes that two graph models include is indicated respectively
Amount,rijIt indicates from node niTo node njThat shifts is general
Rate, i.e.,ωijIt is node niWith node njThe even weight on side, NPGIndicate that the initial graph model PG includes
Number of nodes, NFGIndicate the number of nodes that the feedback graph model FG includes.
(10) weight that PG graph models and FG graph models influence recommendation results is calculated
(11) finally predict mesh user to each movable level of interest
And n activities before to target user recommending level of interest ranking；
Application example
As shown in figure 3, the present invention can be divided into multiple feedback information update points according to the time span used.Each
Feedback information renewal time point, the feedback information that user makes between the update current point in time of batch and a upper time point.
According to the bean cotyledon got with city data, data set is divided into three parts：Initial training collection, feedback data collection
With final test collection.Feedback data collection stochastic averagina is divided into 10 equal portions simultaneously to simulate the user feedback letter of 10 periods
Breath.
Fig. 4 shows the comentropy that 11 moment initial graph model and feedback diagram model include.It can be seen that just from the table
The comentropy of beginning graph model is greater than the comentropy of feedback graph model always, but with the increase of field feedback, initially
The prediction result proportion of graph model constantly declines.This adjustment mode also complies with the subjective consciousness of user, i.e. the time is closer
Information more can to user generate large effect.
Fig. 5, Fig. 6, Fig. 7 and Fig. 8 show proposed by the present invention based on the newer activity recommendation method of graph model feedback and base
In the recommendation method of content, the recommendation method of full figure migration is reused, full figure migration is reused and combines based on content rearrangement
Recommendation method, only consider the figure of the recommendations method and the consideration simultaneously of feedback information and activity similarity of the graph model of feedback information
The recommendation effect of model method compares.
It can be seen that by the comparing result of Fig. 5, Fig. 6, Fig. 7 and Fig. 8, it is evident that proposed by the present invention to be based on graph model
Newer activity recommendation method is fed back, all more traditional recommendation calculation either on P@1, MAP, Recall or F1 evaluation indexes
The proposed algorithm of method and other graph model structures is significantly improved.This is because not only considering when building graph model
Whole Given information also emphasis consider user interest preference change with the time, inclined to enrich description user interest
Good dimension, improves the accuracy rate of recommendation.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, all within the spirits and principles of the present invention made by all any modification, equivalent and improvement etc., should all include
Within protection scope of the present invention.
Claims (8)
1. one kind feeding back newer activity recommendation method based on graph model, which is characterized in that including：
The information of activity, user group and sponsor that the label concentrated according to preset data is marked, it is flat by non-set of weights
Similar tags cluster is several themes by equal method, as theme node S；
It is several activities before each activity matching is most like according to the movable characteristic attribute that the preset data is concentrated, and
Oriented even side between the adjacent activity of construction activities；
It carves at the beginning, using U, E, H, G, initial graph model PG is constructed on the company side between S nodes and each node, and in institute
The random walk that progress is restarted from the band of target user's node on initial graph model PG is stated, the receipts of active node will be obtained
Hold back scoring of the probability as the target user to movable first moment, wherein U, E, H, G indicate the present count respectively
The node constituted according to the user of concentration, activity, sponsor and user group；
In subsequent time, according to the feedback information of the target user, structure feedback graph model FG, in the feedback graph model FG
It is upper to carry out the random walk that is restarted from the band of target user's node, will obtain the convergent probability of active node as
Scoring of the target user to movable second moment, wherein the feedback graph model FG includes user and active node,
And the company side between user and activity；
The initial graph model PG is obtained by the comentropy of the initial graph model PG and the comentropy of the feedback graph model FG
The weight that recommendation results are influenced with the feedback graph model FG, and by the weight, the scoring at first moment and
The score in predicting target user at second moment is to each movable level of interest, and if recommending to the target user preceding
The activity of dry position.
2. according to the method described in claim 1, it is characterized in that, the work that the label concentrated according to preset data is marked
Dynamic, user group and sponsor information, it is several themes to be clustered similar tags by non-weighting group average method, including：
Select two most like Label Mergings for a target cluster from all labels；
Using the target cluster as new label, and calculate the similarity between the target cluster and other labels, wherein this is similar
Degree is equal to the average value of other labels and all label similarities in the target cluster, and is selected from all labels described in execution
Two most like Label Mergings are a target cluster, until the quantity of the cluster finally clustered, which meets, presets number of clusters amount.
3. method according to claim 1 or 2, which is characterized in that described according to the movable of preset data concentration
Characteristic attribute is that each activity matches several most like preceding activities, and the oriented company between the adjacent activity of construction activities
Side, including：
Adjacency matrix, activity and the adjacency matrix of activity time, the adjoining of activity and activity venue that activity is spent with activity
Matrix and activity are connected with the adjacency matrix of Activity Type, obtain movable eigenmatrix；
Similarity between each activity is calculated by the row vector corresponding to movable eigenmatrix；
It is several activities before each movable matching similarity is highest according to the similarity between each activity, and builds from target
The directed edge of similar active is directed toward in activity.
4. according to the method described in claim 3, it is characterized in that, scoring of the target user to movable first moment
For：Wherein,Indicate mesh
U pairs of q-th of activity of user is marked in moment t0Scoring, M indicates movable number.
5. according to the method described in claim 4, it is characterized in that, the band on the initial graph model PG restart it is random
Migration is expressed as：
h(j+1)=αEHe(j)PEH+(1-αEH)s(j)PSH
g(j+1)=αUGu(j)PUG+(1-αUG)s(j)PSG
s(j+1)=αGSg(j)PGS+αHSh(j)PHS+(1-αGS-αHS)e(j)PES
Wherein, u(j), e(j), h(j), g(j), s(j)It is probability vector of all kinds of nodes during iteration j,With
Indicate t0Moment user node is to active node and active node to the transition probability matrix of user node, αEUIt indicates from active section
Point is transferred to the weight of itself probability of user node shared by the probability of user node, αGUIt indicates to be transferred to user's section from small group node
Point probability accounts for the weight of user node itself probability, PGUIndicate transition probability matrix of the small group node to user node, αUEIt indicates
It is transferred to the weight that active node probability accounts for active node itself probability, α from user nodeHEExpression is transferred to from sponsor's node
Active node probability accounts for the weight of active node itself probability, PHEIndicate transition probability square of sponsor's node to active node
Battle array, αSEIt indicates to be transferred to the weight that active node probability accounts for active node itself probability, P from theme nodeSEIndicate theme node
To the transition probability matrix of active node, PEEIndicate transition probability matrix of the active node to active node, αEHIt indicates from activity
Node is transferred to the weight that sponsor's node probability accounts for itself probability of sponsor's node, PEHIndicate active node to sponsor's node
Transition probability matrix, PSHIndicate theme node to the transition probability matrix of sponsor's node, αUGIt indicates to shift from user node
The weight of itself probability of small group node, P are accounted for small group node probabilityUGIndicate transition probability square of the user node to small group node
Battle array, PSGIndicate theme node to the transition probability matrix of small group node, αGSIt indicates to be transferred to theme node probability from small group node
Account for the weight of itself probability of theme node, PGSIndicate small group node to the transition probability matrix of theme node, αHSIt indicates from sponsoring
Fang Jiedian is transferred to the weight that theme node probability accounts for itself probability of theme node, PHSIndicate sponsor's node to theme node
Transition probability matrix, PESIndicate transition probability matrix of the active node to theme node.
6. according to the method described in claim 4, it is characterized in that, described carry out on the feedback graph model FG from the mesh
The random walk that the band that mark user node sets out restarts, will obtain the convergent probability of active node as the target user couple
The scoring at movable second moment, including：
SettingWithIt is t respectivelykThe transition probability matrix at moment, the two transition probability matrixs are respectively by adjacency matrixWithIt is normalized into every trade, whereinWithTwo matrixes transposed matrix each other,In per a line generation
Table corresponds to user and all movable participation relationships, and the representation of activity participated in is 1, and the representation of activity having neither part nor lot in is 0；
ByWithRandom walk is carried out for FG graph models,
Until to target user, the convergence in probability of active node obtains the target user and comments each movable second moment
PointIndicate that target is used
Q-th of the activity in u pairs of family is in moment tkScoring.
7. according to the method described in claim 6, it is characterized in that, the comentropy by the initial graph model PG and described
The comentropy of feedback graph model FG obtains the power that the initial graph model PG and the feedback graph model FG influence recommendation results
Weight, including：
By HThe comentropy of the initial graph model PG is obtained, byObtain the feedback
The comentropy of graph model FG, wherein rijIt indicates from node
niTo node njThe probability of transfer, NPGIndicate the number of nodes that the initial graph model PG includes, NFGIndicate the feedback artwork
The number of nodes that type FG includes；
ByThe initial graph model PG and the feedback graph model FG are obtained to recommendation results
The weight of influence.
8. the method according to the description of claim 7 is characterized in that it is described by the weight, the scoring at first moment
And the score in predicting target user at second moment is to each movable level of interest, including：
ByPredict target user to each movable level of interest.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810317533.0A CN108334645A (en) | 2018-04-10 | 2018-04-10 | One kind feeding back newer activity recommendation method based on graph model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810317533.0A CN108334645A (en) | 2018-04-10 | 2018-04-10 | One kind feeding back newer activity recommendation method based on graph model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108334645A true CN108334645A (en) | 2018-07-27 |
Family
ID=62933031
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810317533.0A Pending CN108334645A (en) | 2018-04-10 | 2018-04-10 | One kind feeding back newer activity recommendation method based on graph model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108334645A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110287420A (en) * | 2019-06-28 | 2019-09-27 | 深圳前海微众银行股份有限公司 | A kind of training method and device of information recommendation system |
| WO2020029412A1 (en) * | 2018-08-09 | 2020-02-13 | 平安科技（深圳）有限公司 | Tag recommendation method and apparatus, computer device, and computer-readable storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106980659A (en) * | 2017-03-20 | 2017-07-25 | 华中科技大学鄂州工业技术研究院 | A kind of doings based on isomery graph model recommend method |
-
2018
- 2018-04-10 CN CN201810317533.0A patent/CN108334645A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106980659A (en) * | 2017-03-20 | 2017-07-25 | 华中科技大学鄂州工业技术研究院 | A kind of doings based on isomery graph model recommend method |
Non-Patent Citations (1)
| Title |
|---|
| SHENGHAO LIU等: "SERGE: Successive Event Recommendation Based on Graph Entropy for Event-Based Social Networks", 《IEEE ACCESS》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020029412A1 (en) * | 2018-08-09 | 2020-02-13 | 平安科技（深圳）有限公司 | Tag recommendation method and apparatus, computer device, and computer-readable storage medium |
| CN110287420A (en) * | 2019-06-28 | 2019-09-27 | 深圳前海微众银行股份有限公司 | A kind of training method and device of information recommendation system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yang et al. | Promoting cooperation by the social incentive mechanism in mobile crowdsensing | |
| Siegel | Asymmetric all-pay auctions with interdependent valuations | |
| CN106980659A (en) | A kind of doings based on isomery graph model recommend method | |
| CN107038184B (en) | A kind of news recommended method based on layering latent variable model | |
| Shao et al. | Estimation of distribution algorithm with path relinking for the blocking flow-shop scheduling problem | |
| CN106991133B (en) | It is a kind of based on any active ues group recommending method for restarting random walk model | |
| Wang et al. | Multi-objective fuzzy flexible job shop scheduling using memetic algorithm | |
| Zhang et al. | Balancing energy consumption and reputation gain of UAV scheduling in edge computing | |
| CN106372101B (en) | A kind of video recommendation method and device | |
| CN111723292B (en) | Recommendation method, system, electronic equipment and storage medium based on graph neural network | |
| US9489638B2 (en) | Method and apparatus for propagating user preference information in a communications network | |
| Wang et al. | Logistics-aware manufacturing service collaboration optimisation towards industrial internet platform | |
| CN107657043B (en) | Content-based mixed graph model image recommendation method | |
| Wu et al. | Two layered approaches integrating harmony search with genetic algorithm for the integrated process planning and scheduling problem | |
| CN108334645A (en) | One kind feeding back newer activity recommendation method based on graph model | |
| CN109918574A (en) | Item recommendation method, device, equipment and storage medium | |
| Hafezi et al. | Identification of representative patterns of time use activity through fuzzy C-means clustering | |
| Fang et al. | Large-scale personalized delivery for guaranteed display advertising with real-time pacing | |
| Ahmed et al. | Solving the vehicle routing problem with time windows using modified football game algorithm | |
| Kang et al. | Honeybee mating optimization algorithm for task assignment in heterogeneous computing systems | |
| CN109195186A (en) | A kind of network slice access intensified learning method based on evolutionary Game | |
| CN112669084B (en) | Policy determination method, device and computer readable storage medium | |
| Xie et al. | We know your preferences in new cities: Mining and modeling the behavior of travelers | |
| CN110196949A (en) | Information-pushing method, device, electronic equipment and computer-readable medium | |
| CN114492849B (en) | Model updating method and device based on federal learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180727 |