CN109543109A - A kind of proposed algorithm of time of fusion window setting technique and score in predicting model - Google Patents
A kind of proposed algorithm of time of fusion window setting technique and score in predicting model Download PDFInfo
- Publication number
- CN109543109A CN109543109A CN201811425529.2A CN201811425529A CN109543109A CN 109543109 A CN109543109 A CN 109543109A CN 201811425529 A CN201811425529 A CN 201811425529A CN 109543109 A CN109543109 A CN 109543109A
- Authority
- CN
- China
- Prior art keywords
- user
- project
- score
- time
- scoring
- 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.)
- Granted
Links
Classifications
-
- 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
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
Landscapes
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses the proposed algorithms of a kind of time of fusion window setting technique and score in predicting model, belong to technical field of electronic commerce, the problem of data sparsity problem present in the technical problem to be solved in the present invention Collaborative Filtering Recommendation Algorithm and user interest change over time, the technical solution of use are as follows: be based on Collaborative Filtering Recommendation Algorithm and time window setting technique, using a kind of new score in predicting model come completion rating matrix, the score in predicting model can overcome the problems, such as that time complexity caused by traditional Non-negative Matrix Factorization score in predicting model is excessively high and score in predicting result is not unique to a certain extent, mitigate influence of the Sparse to collaborative filtering while improving score in predicting precision, it recycles recommended models to predict the project of the possible preference of user according to the interest of user and chooses the higher project of user in predicting scoring and utilizeTopNRecommended method generates combined recommendation list, the recommendation effect of boosting algorithm.
Description
Technical field
The present invention relates to a kind of technical field of electronic commerce, specifically a kind of time of fusion window setting technique and score in predicting
The proposed algorithm of model.
Background technique
Internet develops rapidly so that human information total amount explosive growth, information overload occurs, people are by magnanimity
Information is flooded, such as has millions of books on Amazon, and 1,000,000,000 web page storage, people are had more than above Del.icio.us
It was found that be to oneself interested content it is very difficult, the appearance of personalized recommendation system lacks and changes this status, benefit
The behavior record of user is recorded with recommender system, and combines the relationship between user, commodity, and user is any even without making
Form search just can be appreciated that the commodity interested to oneself.Therefore recommender system either academia or industry all by
Great concern and research.
Recommender system method mainly includes based on content algorithms and collaborative filtering two major classes.Content-based recommendation algorithm master
User interest to be indicated by text-processing technology with a multi-C vector, while also carry out feature extraction to project and establishing spy
Levy vector.Recommended by calculating the similitude between user interest vector sum item feature vector.
Collaborative filtering then includes based on Memory algorithm and based on model algorithm and various blending algorithms.Association based on memory
With filter algorithm behavior record previous according to user first, calculate between user, the similitude between project.Right rear line
Recommend and the big user of its similarity buy, the project that score or recommendation and the former bought item similarity of the user greatly
Project.However actual e-commerce system data volume is very huge, user's rating matrix is very sparse, using it is traditional based on
It is lower to remember proposed algorithm accuracy rate.
The proposed algorithm of mainstream has collaborative filtering recommending and context to recommend, and Collaborative Filtering Recommendation Algorithm surely belongs to classical push away
Algorithm is recommended, current collaborative filtering recommending mainly there are three classes, is based on memory, based on model and mixed recommendation, difference respectively
Classification cope with different demand and application, than more prominent the advantages of collaborative filtering, model passability is strong, implements ratio
The problem of relatively simple, effect is also fine, and disadvantage is also obvious, changes over time such as cold start-up problem and user interest.
Collaborative Recommendation algorithm based on model calculates user's row according to the pervious scoring of user and various implicit preferences
For model.Then the scoring behavior of user is predicted according to model.Matrix decomposition proposed algorithm is by for n user and m
A project establishes k dimensional feature vector, the rating matrix of n × m size is converted to two matrixes of n × k and k × m, then
The dot product of user characteristics vector sum item feature vector is calculated to score project.Many proposed algorithms all pass through calculating and use
The global similarity at family is recommended to find neighbour, but the interest of user may be only similar in one aspect, therefore base
Scoring is classified in the recommended models of Bayes's classification, different similar users are respectively adopted and are recommended.In recommendation field
There are many model, and wherein matrix decomposition model is with good performance, thus many scholars be devoted to using matrix decomposition come into
Row personalized recommendation, such as Non-negative Matrix Factorization, major defect are that time complexity is relatively high, and obtained result is not unique
To be difficult to obtain global minima point.
Probability enigmatic language meaning (PLSA) method is also a kind of based on model algorithm, and the implicit variable of algorithm extraction comes inclined to user
It is modeled well, relatively high accuracy rate can be obtained.These widely used proposed algorithms are mostly static models, only merely
Integration user's history data, not consider user interest situation of change.
Recommender system and Personalized service are as overcoming the important method of data overload to have been widely used for electronics
Commercial field, but the height of time complexity present in traditional nonnegative matrix how is coped with, obtained structure is not unique and assists
Changing over time with user interest in filtering recommendation algorithms is current technical problem urgently to be solved.
Summary of the invention
Technical assignment of the invention is to provide the proposed algorithm of a kind of time of fusion window setting technique and score in predicting model, to solve
Time complexity present in certainly traditional nonnegative matrix is high, obtained structure is not unique and Collaborative Filtering Recommendation Algorithm in user
The problem of interest changes over time.
Technical assignment of the invention realizes in the following manner, a kind of time of fusion window setting technique and score in predicting model
Proposed algorithm, this method are to calculate user's using score in predicting model based on Collaborative Filtering Recommendation Algorithm and time window setting technique
Interest Similarity recycles recommended models to predict the project of the possible preference of user according to the interest of user and choose user in predicting to comment
Higher project is divided to produce recommendation list using TopN recommended method；
The specific method is as follows:
S1, user-project rating matrix is constructed to the scoring of project by user, and scoring is normalized
To normalization user-project rating matrix；
S2, according to score in predicting model, calculated by recommended models and obtain the eigenmatrix of user and the feature of project
Matrix, the two, which is multiplied and then obtains user, scores to the prediction of non-scoring item, and obtains TopN2 using TopN recommended method and push away
Recommend list；
S3, the prediction of acquisition, which is scored, reverts to original scoring according to the principle of normalized, thus obtain one it is thick
Close rating matrix, and multiple and different time windows is divided using time window setting technique, it is item tax of having scored according to time window
Time scale is given, the overall similarity of project set in each user in predicting scoring item and each time window is calculated, it will most
Time scale of any one time scale as prediction article scoring behavior in the time window of high similitude；
S4, the Interest Similarity between user is calculated using collaborative filtering and constructs user interest similarity matrix,
Target user is calculated again to the interest preference of resource, and the highest top n commodity of user interest are produced using TopN recommended method
TopN1 recommendation list, and it is merged with TopN2 recommendation list and generates TopN recommendation list.
Preferably, in the step S1 to the method that is normalized of scoring specific formula is as follows:
Wherein, u indicates user；I indicates project；M indicates the most higher assessment that the score data range supported according to system obtains
Score value；ru,iIndicate user to the true score value of project；After indicating that user does normalized to the true score value of project
Score value,
Preferably, the score in predicting model is for predicting scoring of the user to non-scoring item；It is specific as follows:
It can be calculated by the true scoring of user uWhereinρu,iFor calculating normalizing
The appraisal result of change, and then determine that known variables are used to be determined the feature vector of user and project accordingly
Scoring of the user to non-scoring item can be obtained after the feature vector of family and project by matrix operation.,
In Fig. 1, the parameter of α and β expression model；α represents user characteristics overlapping degree, value range [0,1], when α is close to 0
When, it represents user and tends to same feature, be in brief exactly that user characteristics are relatively simple, α is larger, illustrates that corresponding user tends to
The aspect ratio of different user characteristics, user is more；β is greater than 1, needed for the bigger representative of value proves that some feature of certain user is prominent
The information wanted is more, and k indicates the dimension of vector；U indicates that user, i indicate project, and V indicates item feature vector, item characteristic square
The initialization of battle array is distributed V using betai,k~Beta (β, β), value range are [0,1]；UnIndicate user feature to
Amount is obeyed the distribution of Di Li Cray specific to the feature vector of user u, is denoted asVector (Uu,1,…,Uu,k) table
Show component of the user u in each dimension, due to having used Di Li Cray distributed model,There is no real number value or shaping values；
Z in modelu,iAnd ρi,uIt is the stochastic variable being arranged for each user u and project i；Zu,iIt is the stochastic variable for obeying classification distribution,Zu,iIn value indicate scoring of the user u to project i；ρu,iIt is the stochastic variable for obeying bi-distribution,User is represent to some preferences purpose confidence level.
More preferably, the prediction model is used to predict the project of the possible preference of user according to the interest preference of user；It pushes away
The concrete methods of realizing for calculating model is as follows:
(1), α, β and rating matrix R are inputted；Wherein, α and β indicates the parameter of prediction model；α represents user characteristics overlapping
Degree, 0≤α≤1；When α is close to 0, represents user and tend to same feature, is i.e. user characteristics are relatively simple；α is larger, then illustrates
Tend to different user characteristics, i.e. the aspect ratio of user is more；β indicate the prominent information needed of some feature of user number, β >
1, value is bigger represent some feature of certain user it is prominent required for information it is more；
(2), normalized generator matrix R ' is done to rating matrix R；
(3), random initializtion free parameter γu,k、Withγu,kFor u × k dimension matrix,WithRespectively v × k
The matrix of dimension；Wherein, u indicates user, and k representing matrix is dimension, and v indicates project；
(4), it is recorded according to the true scoring of user uRelative users u is calculated to the λ of respective item iu,i,k:
Wherein, λu,i,kIndicate correlated variables Zu,iThe parameter in classification distribution obeyed, λ 'u,i,kIt indicates to calculate λu,i,k's
Intermediate quantity；
Ψ is defined as the logarithmic derivative of gamma function,
Wherein, Γ (x) indicates gamma function；The derivative of Γ ' (x) expression gamma function；Y indicates the highest supported with system
It scores related, Y is constant, is fixed as 4；ρu,iIndicate the conditional probability distribution of user u preference project i；X is indicated；
(5), it is recorded according to the true scoring of user uCalculate and update the γ of relative users uu,k:
Wherein, γu,kIndicate the relevant parameter in the Di Li Cray distribution of user u obedience；
(6), it is recorded according to the true scoring of user uIt calculates and updates respective item i's
(7)、γu,k、WithParameter value whether change obviously:
If 1., it is obvious, repeat step (4) to (6)；
If 2., it is unobvious, then follow the steps (8)；
(8), the eigenmatrix U of user u is calculatedu,k:Calculating project eigenmatrix Vi,k:
(9), user u is calculated to predict the preference of project i:
(10), score in predicting completion matrix R ' generates TopN2 recommendation list using TopN recommended method；
(11), time window is divided according to time window setting technique；
(12), item similarity is calculated:
Wherein, sim (i, j) indicates the similarity of project i and project j；W (u, i) and W (u, j) indicates time weighting sum number
The weight combined according to weight；WithThe average score of project i and project j are respectively indicated,Indicate being averaged for user u marking
Point；
(13), the comprehensive similarity of project in the non-scoring item of each user and time window is calculated:
Wherein, Iu,jIndicate the project set in j-th of time window of user u, size (Iu,j) indicate set Iu,jIt is big
It is small；
(14), (the selection highest preceding k ' of similarity is a not for the highest project imparting time scale of a similarity of k ' before choosing
Scoring item assigns time scale, and the optimal value of k ' is different according to the difference of data set, needs to obtain by experiment optimal
Value)；
(15), the dense matrix with time scale is converted into user-project-time three-dimensional rating matrix；
(16), the similarity between user is obtained:
Wherein,tiIt indicates the time weight that corresponding user scores to respective item, passes through use
Family-project-time three-dimensional matrice obtains；
(17), user is obtained to the preference value of non-scoring item:
Wherein, S (u, k) indicates that preceding k user similar with user interest, N (i) indicate the use for having scoring to project i
Family set, sim (u, v) indicate the Interest Similarity between user u and user v；
(18), it on the basis of score in predicting, chooses the higher project of user in predicting scoring and generates TopN2 recommendation list,
And it is aggregated into new TopN recommendation list with the weighting of TopN1 recommendation list, formula is as follows:
TopN=ε TopN1+ (1- ε) TopN2
Wherein, for ε value between 0 and 1, different data sets has different ε optimal values.
More preferably, the calculation method for dividing time window according to time window setting technique in the step (11) is as follows:
Tuk(k)=Tu0-θ(k-1)-ka1
Wherein, α1The size for indicating first time window represents the length of user interest, and value is bigger, and window is bigger；θ
It indicates time window button interval increasing degree, is followed successively by Tu1, Tu2..., Tuk, user's Long-term Interest variation speed degree is represented, value is got over
It is big to indicate that user interest variation is faster, conversely, variation is slower.
More preferably, the time window difference weight that correspondence is different, using Chinese mugwort this great forgetting curve of guest as the time
Function:
F (u, i)=0.318 × (T0-Tuk)-0.125
Wherein, TukThe time window where time when indicating user u access i；The range of f (u, i) value is [0,1], is presented
Forgetting law first quick and back slow；
Interest-degree of the combined data weight definition target user to project:
W (u, i)=f (u, i) × β1+w(u,i)×(1-β1),β1∈[0,1]
Wherein, w (u, i) indicates the interest-degree in the user u nearest period to project i；I indicates user's nearest period
(Tu1~Tu0) in the project set that accessed；The similarity of the expression project i and project j of sim (i, j)；
More preferably, the calculating of item similarity sim (i, j) uses modified cosine similarity in w (u, i):
Wherein,WithThe average score of project i and project j are respectively indicated,Indicate the average mark of user u marking.
The proposed algorithm of time of fusion window setting technique and score in predicting model of the invention has the advantage that
(1), the present invention includes mainly two parts, is score in predicting model and collaborative filtering recommending model respectively, for when
Between processing using time window setting technique, influence of the Deta sparseness to collaborative filtering can make up for it by this method,
And due to the introducing of time window setting technique, recommendation results is enable preferably to agree with the interests change of user；
(2), the present invention uses new score in predicting model, and compared to traditional matrix decomposition algorithm, score in predicting is obtained
It greatly improves；
(3), present invention employs time window setting techniques to divide time window, it is contemplated that time factor is to user interest
It influences, to keep recommendation results more reasonable；
(4), score in predicting model is combined with time window setting technique with collaborative filtering, collaboration can not only be made up
It is cold-started problem present in filter algorithm, while can also improve influence of the Deta sparseness to collaborative filtering, especially
The addition of time window setting technique makes collaborative filtering can adapt to the interests change of user, to improve the effect of recommendation；
(5), the present invention merges new Algorithms of Non-Negative Matrix Factorization and time window setting technique in collaborative filtering, comprehensive
Both the advantages of to cope in defect present in traditional nonnegative matrix and Collaborative Filtering Recommendation Algorithm user interest at any time
Between the problem of changing, so that collaborative filtering is can adapt to the interests change of user, to improve the effect of recommendation.
Detailed description of the invention
The following further describes the present invention with reference to the drawings.
Attached drawing 1 is flow diagram of the invention；
Attached drawing 2 is the relation schematic diagram of score in predicting model；
Attached drawing 3 is that time window divides schematic diagram；
Attached drawing 4 is the curve graph of accuracy rate；
Attached drawing 5 is the curve graph of recall rate.
Specific embodiment
Referring to Figure of description and specific embodiment to a kind of time of fusion window setting technique of the invention and score in predicting model
Proposed algorithm be described in detail below.
Embodiment one:
The proposed algorithm of time of fusion window setting technique and score in predicting model of the invention, this method is pushed away based on collaborative filtering
Recommend algorithm and time window setting technique, using score in predicting model calculate user Interest Similarity, recycle recommended models according to
The interest prediction user at family may preference project and choose score higher project of user in predicting and given birth to using TopN recommended method
Produce TopN recommendation list；
The specific method is as follows:
S1, user-project rating matrix is constructed to the scoring of project by user, and scoring is normalized
To normalization user-project rating matrix；
S2, according to score in predicting model, calculated by recommended models and obtain the eigenmatrix of user and the feature of project
Matrix, the two, which is multiplied and then obtains user, scores to the prediction of non-scoring item, and obtains TopN2 using TopN recommended method and push away
Recommend list；
To the method that is normalized of scoring specific formula is as follows:
Wherein, u indicates user；I indicates project；M indicates the most higher assessment that the score data range supported according to system obtains
Score value；ru,iIndicate user to the true score value of project；After indicating that user does normalized to the true score value of project
Score value,
S3, the prediction of acquisition, which is scored, reverts to original scoring according to the principle of normalized, thus obtain one it is thick
Close rating matrix, and multiple and different time windows is divided using time window setting technique, it is item tax of having scored according to time window
Time scale is given, the overall similarity of project set in each user in predicting scoring item and each time window is calculated, it will most
Time scale of any one time scale as prediction article scoring behavior in the time window of high similitude；It is pre- by scoring
It surveys model and completion is carried out to rating matrix, and then obtain dense rating matrix, be the scoring in rating matrix on this basis
Time scale is assigned, can learn that the interest of user has timeliness according to correlative study, can change with the variation of time,
But short-term user interest is basically unchanged, according to the anthropomorphic forgetting law time graph proposed in correlative study, user is earliest
The time of scoring item be set as 0 and user last scoring item time be set as Tu0, by 0 to Tu0This is divided into for a period of time
Multiple time slices, as shown in Figure 3.
S4, the Interest Similarity between user is calculated using collaborative filtering and constructs user interest similarity matrix,
Target user is calculated again to the interest preference of resource, and the highest top n commodity of user interest are produced using TopN recommended method
TopN1 recommendation list, and it is merged with TopN2 recommendation list and generates TopN recommendation list.
Wherein, score in predicting model is for predicting scoring of the user to non-scoring item；It is specific as follows:
It can be calculated by the true scoring of user uWherein,ρu,iFor calculating normalizing
The appraisal result of change, and then determine that known variables are used to be determined the feature vector of user and project accordingly
Scoring of the user to non-scoring item can be obtained after the feature vector of family and project by matrix operation.
As shown in Fig. 2, α and β indicates the parameter of model；α represents user characteristics overlapping degree, value range [0,1], when
It when α is close to 0, represents user and tends to same feature, be in brief exactly that user characteristics are relatively simple, the larger then explanation of α is to application
Family tends to different user characteristics, and the aspect ratio of user is more；β is greater than 1, and the bigger representative of value proves that some feature of certain user is prominent
Required information is more out, and k indicates the dimension of vector；
U indicates that user, i indicate project, and V indicates item feature vector, and the initialization of item characteristic matrix is using shellfish
Tower is distributed Vi,k~Beta (β, β), value range are [0,1]；UnIndicate user feature vector, specific to user u feature to
Amount obeys the distribution of Di Li Cray, is denoted asVector (Uu,1,…,Uu,k) indicate point of the user u in each dimension
Amount, due to having used Di Li Cray distributed model,There is no real number value or shaping values；Z in modelu,iAnd ρi,uIt is for each use
The stochastic variable of family u and project i setting；Zu,iIt is the stochastic variable for obeying classification distribution,Zu,iIn value
Indicate scoring of the user u to project i；ρi,uIt is the stochastic variable for obeying bi-distribution,Represent user
To some preferences purpose confidence level.
As shown in Fig. 1, prediction model is used to predict the project of the possible preference of user according to the interest preference of user；It pushes away
The concrete methods of realizing for calculating model is as follows:
(1), α, β and rating matrix R are inputted；Wherein, α and β indicates the parameter of prediction model；α represents user characteristics overlapping
Degree, 0≤α≤1；When α is close to 0, represents user and tend to same feature, is i.e. user characteristics are relatively simple；α is larger, then illustrates
Tend to different user characteristics, i.e. the aspect ratio of user is more；β indicate the prominent information needed of some feature of user number, β >
1, value is bigger represent some feature of certain user it is prominent required for information it is more；
(2), normalized generator matrix R ' is done to rating matrix R；
(3), random initializtion free parameter γu,k、Withγu,kFor u × k dimension matrix,WithRespectively v × k
The matrix of dimension；Wherein, u indicates user, and k representing matrix is dimension, and v indicates project；
(4), it is recorded according to the true scoring of user uRelative users u is calculated to the λ of respective item iu,i,k:
Wherein, λu,i,kIndicate correlated variables Zu,iThe parameter in classification distribution obeyed, λ 'u,i,kIt indicates to calculate λu,i,k's
Intermediate quantity；
Ψ is defined as the logarithmic derivative of gamma function in above formula,
Wherein, Γ (x) indicates gamma function；The derivative of Γ ' (x) expression gamma function；Y indicates the highest supported with system
It scores related, Y is constant, is fixed as 4；ρu,iIndicate the conditional probability distribution of user u preference project i；X is indicated；
(5), it is recorded according to the true scoring of user uCalculate and update the γ of relative users uu,k:
Wherein, γu,kIndicate the relevant parameter in the Di Li Cray distribution of user u obedience；
(6), it is recorded according to the true scoring of user uIt calculates and updates respective item i's
(7)、γu,k、WithParameter value whether change obviously:
If 1., it is obvious, repeat step (4) to (6)；
If 2., it is unobvious, then follow the steps (8)；
(8), the eigenmatrix U of user u is calculatedu,k:Calculating project eigenmatrix Vi,k:
(9), user u is calculated to predict the preference of project i:
(10), score in predicting completion matrix R ' generates TopN2 recommendation list using TopN recommended method；
(11), time window is divided according to time window setting technique；The calculation method of time window is divided according to time window setting technique
It is as follows:
Tuk(k)=Tu0-θ(k-1)-ka1
Wherein, α1The size for indicating first time window represents the length of user interest, and value is bigger, and window is bigger；θ
It indicates time window button interval increasing degree, is followed successively by Tu1, Tu2..., Tuk, user's Long-term Interest variation speed degree is represented, value is got over
It is big to indicate that user interest variation is faster, conversely, variation is slower.In the present embodiment, time window is carried out impartial division by θ=0.
The time window difference weight that correspondence is different, using Chinese mugwort this great forgetting curve of guest as the function of time:
F (u, i)=0.318 × (T0-Tuk)-0.125
Wherein, TukThe time window where time when indicating user u access i；The range of f (u, i) value is [0,1], is presented
Forgetting law first quick and back slow；
Interest-degree of the combined data weight definition target user to project:
W (u, i)=f (u, i) × β1+w(u,i)×(1-β1),β1∈[0,1]
Wherein, w (u, i) indicates the interest-degree in the user u nearest period to project i；I indicates user's nearest period
(Tu1~Tu0) in the project set that accessed；The similarity of the expression project i and project j of sim (i, j)；The calculating of sim (i, j)
Using modified cosine similarity:
Wherein,WithThe average score of project i and project j are respectively indicated,Indicate the average mark of user u marking.
(12), item similarity is calculated:
Wherein, sim (i, j) indicates the similarity of project i and project j；W (u, i) and W (u, j) indicates time weighting sum number
The weight combined according to weight；WithThe average score of project i and project j are respectively indicated,Indicate being averaged for user u marking
Point；
(13), the comprehensive similarity of project in the non-scoring item of each user and time window is calculated:
Wherein, Iu,jIndicate the project set in j-th of time window of user u, size (Iu,j) indicate set Iu,jIt is big
It is small；
(14), (the selection highest preceding k ' of similarity is a not for the highest project imparting time scale of a similarity of k ' before choosing
Scoring item assigns time scale, and the optimal value of k ' is different according to the difference of data set, needs to obtain by experiment optimal
Value)；
(15), the dense matrix with time scale is converted into user-project-time three-dimensional rating matrix；
(16), the similarity between user is obtained, calculates user using based on time-weighted Pearson came associated similarity
Interest Similarity:
Wherein, weighting function is done using Logistic function, time windows is assigned with different weights, same time window
In scoring to assign identical weight equation as follows:tiIndicate that corresponding user scores to respective item
Time weight, pass through user-project-time three-dimensional matrice and obtain；
(17), user is obtained to the preference value of non-scoring item:
Wherein, S (u, k) indicates that preceding k user similar with user interest, N (i) indicate the use for having scoring to project i
Family set, sim (u, v) indicate the Interest Similarity between user u and user v；
(18), it on the basis of score in predicting, chooses the higher project of user in predicting scoring and generates TopN2 recommendation list,
And it is aggregated into new TopN recommendation list with the weighting of TopN1 recommendation list, formula is as follows:
TopN=ε TopN1+ (1- ε) TopN2
Wherein, for ε value between 0 and 1, different data sets has different ε optimal values.
Embodiment two: specific implementation
A, using data set Netflix, Movielens 20M, Movielens 10M, Movielens1M and Epinion；
User-project rating matrix is constructed to the scoring of project according to user, and scoring is normalized to obtain to normalize and is commented
Sub-matrix；
B, user is obtained by score in predicting model to score to the prediction of project, and generate TopN2 recommendation list, scoring
Accuracy uses MAE and CMAE, and calculation formula is as follows:
As a result as shown in the table:
C, rating matrix can be carried out by completion by score in predicting, dense user's rating matrix is obtained, using the time
Window setting technique (sets 6 for the number of time window according to related research result, equalization divides), and by between calculating project
Similarity sim (i, j), and then obtain the overall similarity of project set in each user in predicting scoring item and each time window
sim(i,Iu,j), select the highest preceding a non-scoring item of k ' of similarity to assign time scale.
D, the dense matrix with temporal information is finally converted into user-project-time three-dimensional rating matrix, in this base
User interest similarity sim (u, v) is calculated on plinth, and then calculates user to the interested degree of project, and generates TopN1 recommendation
List ultimately produces TopN recommendation list and feeds back to user, and the effect of recommendation list is measured using accuracy rate and recall rate,
As shown in figs. 4 and 5.
Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention., rather than its limitations；To the greatest extent
Pipe present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: its according to
So be possible to modify the technical solutions described in the foregoing embodiments, or to some or all of the technical features into
Row equivalent replacement；And these are modified or replaceed, various embodiments of the present invention technology that it does not separate the essence of the corresponding technical solution
The range of scheme.
Claims (7)
1. a kind of proposed algorithm of time of fusion window setting technique and score in predicting model, which is characterized in that this method is based on collaboration
Filtering recommendation algorithms and time window setting technique calculate the Interest Similarity of user using score in predicting model, recycle recommended models
The project of the possible preference of user is predicted according to the interest of user and chooses the higher project of user in predicting scoring, and TopN being utilized to recommend
Method produces recommendation list；
The specific method is as follows:
S1, user-project rating matrix is constructed to the scoring of project by user, and scoring is normalized and is returned
One changes user-project rating matrix；
S2, according to score in predicting model, the eigenmatrix of user and the eigenmatrix of project are calculated and obtained by recommended models,
The two, which is multiplied and then obtains user, scores to the prediction of non-scoring item, and obtains TopN2 using TopN recommended method and recommend column
Table；
S3, the prediction of acquisition, which is scored, reverts to original scoring according to the principle of normalized, thus obtain one it is dense
Rating matrix, and multiple and different time windows is divided using time window setting technique, it is when having scored item imparting according to time window
Between scale, the overall similarity of project set in each user in predicting scoring item and each time window is calculated, by highest phase
Like property time window in any one time scale as prediction article scoring behavior time scale；
S4, the Interest Similarity between user is calculated using collaborative filtering and constructs user interest similarity matrix, then counted
Target user is calculated to the interest preference of resource, the highest top n commodity of user interest are produced into TopN1 using TopN recommended method
Recommendation list, and it is merged with TopN2 recommendation list and generates TopN recommendation list.
2. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 1, which is characterized in that institute
State in step S1 to the method that is normalized of scoring specific formula is as follows:
Wherein, u indicates user；I indicates project；M indicates the highest score value that the score data range supported according to system obtains；
ru,iIndicate user to the true score value of project；Indicate that user does commenting after normalized to the true score value of project
Score value,
3. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 1 or 2, feature exist
In the score in predicting model is for predicting scoring of the user to non-scoring item；It is specific as follows:
It can be calculated by the true scoring of user uWherein,ρu,iIt is normalized for calculating
Appraisal result, and then determine that known variables to be determined the feature vector of user and project are obtaining corresponding user
With scoring of the user to non-scoring item can be obtained after the feature vector of project by matrix operation.
4. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 3, which is characterized in that institute
Prediction model is stated for predicting the project of the possible preference of user according to the interest preference of user；The specific implementation side of prediction model
Method is as follows:
(1), α, β and rating matrix R are inputted；Wherein, α and β indicates the parameter of prediction model；α represents user characteristics overlapping degree,
0≤α≤1；When α is close to 0, represents user and tend to same feature, is i.e. user characteristics are relatively simple；α is larger, then explanation tends to not
Same user characteristics, the i.e. aspect ratio of user are more；β indicate the prominent information needed of some feature of user number, β > 1, value
It is bigger represent some feature of certain user it is prominent required for information it is more；
(2), normalized generator matrix R ' is done to rating matrix R；
(3), random initializtion free parameter γu,k、Withγu,kFor u × k dimension matrix,WithRespectively v × k dimension
Matrix；Wherein, u indicates user, and k representing matrix is dimension, and v indicates project；
(4), it is recorded according to the true scoring of user uRelative users u is calculated to the λ of respective item iu,i,k:
Wherein, λu,i,kIndicate correlated variables Zu,iThe parameter in classification distribution obeyed, λ "u,i,kIt indicates to calculate λu,i,kCentre
Amount；
Ψ is defined as the logarithmic derivative of gamma function in above formula,
Wherein, Γ (x) indicates gamma function；The derivative of Γ ' (x) expression gamma function；Y indicates to score with the highest that system is supported
Related, Y is constant, is fixed as 4；ρu,iIndicate the conditional probability distribution of user u preference project i；
(5), it is recorded according to the true scoring of user uCalculate and update the γ of relative users uu,k:
Wherein, γu,kIndicate the relevant parameter in the Di Li Cray distribution of user u obedience；
(6), it is recorded according to the true scoring of user uIt calculates and updates respective item i's
(5)、γu,k、WithParameter value whether change obviously:
If 1., it is obvious, repeat step (4) to (6)；
If 2., it is unobvious, then follow the steps (8)；
(8), the eigenmatrix U of user u is calculatedu,k:
Calculating project eigenmatrix Vi,k:
(9), user u is calculated to predict the preference of project i:
(10), score in predicting completion matrix R ' generates TopN2 recommendation list using TopN recommended method；
(11), time window is divided according to time window setting technique；
(12), item similarity is calculated:
Wherein, sim (i, j) indicates the similarity of project i and project j；W (u, i) and W (u, j) indicates time weighting and data power
The weight that heavy phase combines；WithThe average score of project i and project j are respectively indicated,Indicate the average mark of user u marking；
(13), the comprehensive similarity of project in the non-scoring item of each user and time window is calculated:
Wherein, Iu,jIndicate the project set in j-th of time window of user u, size (Iu,j) indicate set Iu,jSize；
(14), the highest project of a similarity of k ' assigns time scale before choosing；
(15), the dense matrix with time scale is converted into user-project-time three-dimensional rating matrix；
(16), the similarity between user is obtained:
Wherein,tiIt indicates the time weight that corresponding user scores to respective item, passes through user-item
Mesh-time three-dimensional matrice obtains；
(17), user is obtained to the preference value of non-scoring item:
Wherein, S (u, k) indicates that preceding k user similar with user interest, N (i) indicate to have project i the user of scoring to collect
It closes, sim (u, v) indicates the Interest Similarity between user u and user v；
(18), on the basis of score in predicting, the higher project generation TopN2 recommendation list of selection user in predicting scoring, and with
The weighting of TopN1 recommendation list is aggregated into new TopN recommendation list, and formula is as follows:
TopN=ε TopN1+ (1- ε) TopN2
Wherein, for ε value between 0 and 1, different data sets has different ε optimal values.
5. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 4, which is characterized in that institute
It is as follows to state the calculation method that time window is divided according to time window setting technique in step (11):
Tuk(k)=Tu0-θ(k-1)-ka1
Wherein, α1The size for indicating first time window represents the length of user interest, and value is bigger, and window is bigger；When θ is indicated
Between window button interval increasing degree, be followed successively by Tu1, Tu2..., Tuk, user's Long-term Interest variation speed degree is represented, bigger expression is worth
User interest variation is faster, conversely, variation is slower.
6. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 5, which is characterized in that institute
The time window difference weight that correspondence is different is stated, using Chinese mugwort this great forgetting curve of guest as the function of time:
F (u, i)=0.318 × (T0-Tuk)-0.125
Wherein, TukThe time window where time when indicating user u access i；The range of f (u, i) value is [0,1], is presented first fast
Slow forgetting law afterwards；
Interest-degree of the combined data weight definition target user to project:
W (u, i)=f (u, i) × β1+w(u,i)×(1-β1),β1∈[0,1]
Wherein, w (u, i) indicates the interest-degree in the user u nearest period to project i；I indicates the nearest period (T of useru1~
Tu0) in the project set that accessed；The similarity of the expression project i and project j of sim (i, j).
7. the proposed algorithm of time of fusion window setting technique and score in predicting model according to claim 6, which is characterized in that institute
The calculating of item similarity sim (i, j) in claim 6 is stated using modified cosine similarity:
Wherein,WithThe average score of project i and project j are respectively indicated,Indicate the average mark of user u marking.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811425529.2A CN109543109B (en) | 2018-11-27 | 2018-11-27 | Recommendation algorithm integrating time window technology and scoring prediction model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811425529.2A CN109543109B (en) | 2018-11-27 | 2018-11-27 | Recommendation algorithm integrating time window technology and scoring prediction model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109543109A true CN109543109A (en) | 2019-03-29 |
| CN109543109B CN109543109B (en) | 2021-06-22 |
Family
ID=65851074
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811425529.2A Active CN109543109B (en) | 2018-11-27 | 2018-11-27 | Recommendation algorithm integrating time window technology and scoring prediction model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109543109B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110544129A (en) * | 2019-09-05 | 2019-12-06 | 创新奇智(青岛)科技有限公司 | Personalized recommendation method for social e-commerce users |
| CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111310033A (en) * | 2020-01-23 | 2020-06-19 | 山西大学 | Recommendation method and recommendation device based on user interest drift |
| CN111339435A (en) * | 2020-02-10 | 2020-06-26 | 南京邮电大学 | Matrix decomposition completion hybrid recommendation method based on potential factors |
| CN111382361A (en) * | 2020-03-12 | 2020-07-07 | 腾讯科技（深圳）有限公司 | Information pushing method and device, storage medium and computer equipment |
| CN111475744A (en) * | 2020-04-03 | 2020-07-31 | 南京理工大学紫金学院 | Personalized position recommendation method based on ensemble learning |
| CN112069417A (en) * | 2020-08-24 | 2020-12-11 | 北京神舟航天软件技术有限公司 | Work breakdown structure WBS template recommendation method |
| CN112530520A (en) * | 2019-09-17 | 2021-03-19 | 中山大学 | CircRNA function prediction method based on scoring mechanism and LightGBM |
| CN112860984A (en) * | 2019-11-27 | 2021-05-28 | 中移(苏州)软件技术有限公司 | Recommendation method, recommendation device and storage medium |
| CN113011950A (en) * | 2021-03-30 | 2021-06-22 | 吉林亿联银行股份有限公司 | Product recommendation method and device |
| CN113360759A (en) * | 2021-06-09 | 2021-09-07 | 南京大学 | Crowd-sourcing task recommendation method based on dual timing sequence correlation of user and project |
| CN116028727A (en) * | 2023-03-30 | 2023-04-28 | 南京邮电大学 | Video recommendation method based on image data processing |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106339502A (en) * | 2016-09-18 | 2017-01-18 | 电子科技大学 | Modeling recommendation method based on user behavior data fragmentation cluster |
| CN107729542A (en) * | 2017-10-31 | 2018-02-23 | 咪咕音乐有限公司 | A kind of information methods of marking and device and storage medium |
-
2018
- 2018-11-27 CN CN201811425529.2A patent/CN109543109B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106339502A (en) * | 2016-09-18 | 2017-01-18 | 电子科技大学 | Modeling recommendation method based on user behavior data fragmentation cluster |
| CN107729542A (en) * | 2017-10-31 | 2018-02-23 | 咪咕音乐有限公司 | A kind of information methods of marking and device and storage medium |
Non-Patent Citations (2)
| Title |
|---|
| 刘乔 等: ""基于时间加权的协同过滤推荐算法的改进"", 《计算机工程与设计》 * |
| 张鹏飞 等: ""融合标签和多元信息的个性化推荐算法研究"", 《计算机工程与应用》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110544129A (en) * | 2019-09-05 | 2019-12-06 | 创新奇智(青岛)科技有限公司 | Personalized recommendation method for social e-commerce users |
| CN112530520A (en) * | 2019-09-17 | 2021-03-19 | 中山大学 | CircRNA function prediction method based on scoring mechanism and LightGBM |
| CN112860984A (en) * | 2019-11-27 | 2021-05-28 | 中移(苏州)软件技术有限公司 | Recommendation method, recommendation device and storage medium |
| CN111310033A (en) * | 2020-01-23 | 2020-06-19 | 山西大学 | Recommendation method and recommendation device based on user interest drift |
| CN111339435B (en) * | 2020-02-10 | 2022-09-23 | 南京邮电大学 | Matrix decomposition completion hybrid recommendation method based on potential factors |
| CN111339435A (en) * | 2020-02-10 | 2020-06-26 | 南京邮电大学 | Matrix decomposition completion hybrid recommendation method based on potential factors |
| CN111311324B (en) * | 2020-02-18 | 2022-05-20 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111311324A (en) * | 2020-02-18 | 2020-06-19 | 电子科技大学 | User-commodity preference prediction system and method based on stable neural collaborative filtering |
| CN111382361A (en) * | 2020-03-12 | 2020-07-07 | 腾讯科技（深圳）有限公司 | Information pushing method and device, storage medium and computer equipment |
| CN111382361B (en) * | 2020-03-12 | 2023-05-02 | 腾讯科技（深圳）有限公司 | Information pushing method, device, storage medium and computer equipment |
| CN111475744A (en) * | 2020-04-03 | 2020-07-31 | 南京理工大学紫金学院 | Personalized position recommendation method based on ensemble learning |
| CN112069417A (en) * | 2020-08-24 | 2020-12-11 | 北京神舟航天软件技术有限公司 | Work breakdown structure WBS template recommendation method |
| CN113011950A (en) * | 2021-03-30 | 2021-06-22 | 吉林亿联银行股份有限公司 | Product recommendation method and device |
| CN113360759A (en) * | 2021-06-09 | 2021-09-07 | 南京大学 | Crowd-sourcing task recommendation method based on dual timing sequence correlation of user and project |
| CN113360759B (en) * | 2021-06-09 | 2023-08-25 | 南京大学 | Crowd measurement task recommendation method based on user and project dual time sequence correlation |
| CN116028727A (en) * | 2023-03-30 | 2023-04-28 | 南京邮电大学 | Video recommendation method based on image data processing |
| CN116028727B (en) * | 2023-03-30 | 2023-08-18 | 南京邮电大学 | Video recommendation method based on image data processing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109543109B (en) | 2021-06-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109543109A (en) | A kind of proposed algorithm of time of fusion window setting technique and score in predicting model | |
| Reddy et al. | Content-based movie recommendation system using genre correlation | |
| CN108256093B (en) | Collaborative filtering recommendation algorithm based on multiple interests and interest changes of users | |
| CN110704739B (en) | Resource recommendation method and device and computer storage medium | |
| Kumar et al. | A movie recommender system: Movrec | |
| CN109299994B (en) | Recommendation method, device, equipment and readable storage medium | |
| CN105574003B (en) | A kind of information recommendation method based on comment text and scoring analysis | |
| CN104063481A (en) | Film individuation recommendation method based on user real-time interest vectors | |
| CN103744966A (en) | Item recommendation method and device | |
| CN105005589A (en) | Text classification method and text classification device | |
| CN106777051A (en) | A kind of many feedback collaborative filtering recommending methods based on user's group | |
| US8903822B2 (en) | Apparatus and method for measuring contents similarity based on feedback information of ranked user and computer readable recording medium storing program thereof | |
| Eliyas et al. | Recommendation systems: Content-based filtering vs collaborative filtering | |
| CN105426528A (en) | Retrieving and ordering method and system for commodity data | |
| CN110175895B (en) | Article recommendation method and device | |
| CN104090919A (en) | Advertisement recommending method and advertisement recommending server | |
| CN103824213A (en) | Individualized recommendation method based on user preferences and commodity properties | |
| CN111177559B (en) | Text travel service recommendation method and device, electronic equipment and storage medium | |
| CN109710835B (en) | Heterogeneous information network recommendation method with time weight | |
| CN103559622A (en) | Characteristic-based collaborative filtering recommendation method | |
| CN101266620A (en) | Method and apparatus for providing target information to user | |
| CN109840833A (en) | Bayes's collaborative filtering recommending method | |
| Pujahari et al. | Preference relation based collaborative filtering with graph aggregation for group recommender system | |
| Ji et al. | Making recommendations from top-N user-item subgroups | |
| CN108920647A (en) | Low-rank matrix based on spectral clustering fills TOP-N recommended method |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |