CN105740430B - A kind of personalized recommendation method of mosaic society's information - Google Patents
A kind of personalized recommendation method of mosaic society's information Download PDFInfo
- Publication number
- CN105740430B CN105740430B CN201610067099.6A CN201610067099A CN105740430B CN 105740430 B CN105740430 B CN 105740430B CN 201610067099 A CN201610067099 A CN 201610067099A CN 105740430 B CN105740430 B CN 105740430B
- Authority
- CN
- China
- Prior art keywords
- user
- project
- label
- information
- similarity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Abstract
A kind of personalized recommendation method of mosaic society's information, includes the following steps：S1, structure user's users to trust matrix；S2, structure item label similarity matrix；S3, structure and training pattern：S4, prediction user are for the unknown purpose preference.The present invention mainly has the following advantages：1) method for study of sorting in information retrieval field is applied in Top K recommendations, efficiently solves the sequencing problem in commending system, while overcoming traditional the shortcomings that Top K recommendations can not effectively be carried out based on score in predicting method；2) socialization information i.e. user social contact information and item label information have been merged in the model based on sequence study, has improved the accuracy rate of recommendation results.
Description
Technical field
The present invention relates to personalized recommendation, sequence study and community network field, especially a kind of mosaic society's information
Personalized recommendation method.
Background technology
With the rapid development of Internet technology especially e-commerce, in internet the growth rate of data considerably beyond
The reception speed of the mankind, problem of information overload seem increasingly severe.Us are helped to filter out useful number from mass data
According to Information Filtering Technology become more and more important, personalized recommendation technology be exactly it is a kind of according to user preference from large-scale data
In find the ideal method of user's data of interest.
Currently, the application of personalized recommendation is broadly divided into two major classes.The first kind is score in predicting problem, i.e., by giving one
The history scoring behavior prediction of a user scores to the unknown purpose, and score value is the fancy grade for indicating user to project.The
Two classes are that Top-K recommends problem, Top-K to recommend to be dedicated to K project for recommending its most probable to like for user.Since user is past
The project of front is come toward most concern, therefore is compared with score in predicting problem, Top-K more intuitively provides sequence to the user
Recommendation list, thus it is more practical, and this is also that current major e-commerce website is dedicated to solving the problems, such as.
The core of personalized recommendation technology is proposed algorithm, it is presently recommended that algorithm is broadly divided into two major classes, it is interior respectively
Hold filtering and collaborative filtering.Information filtering recommends method mainly by analyzing the content information of user and project, such as the people of user
Mouth statistical information, the description information etc. of project, to construct the series of features of user and project, eventually by matching user
Recommend to make with the similarity of project.In contrast to this, collaborative filtering method does not need the content of any user or project
Information is a kind of method completely unrelated with field.Collaborative filtering method efficiently utilizes group intelligence, it is based on such
It is assumed that user can like and oneself have the project that same interest user likes, meanwhile, the joint act between user is more at most
Interest between user is more similar.Collaborative filtering method is broadly divided into two major classes at present, is the collaborative filtering based on memory respectively
With the collaborative filtering based on model.Collaborative filtering method efficiently avoids the problem of needing expert's markup information, and
It has been widely used in various commending systems.
Although the above method can improve recommendation accuracy rate to a certain extent, it is but faced with " number in practical applications
According to sparse " and " cold start-up " problem." Sparse " problem refers to that the hollow element of user-project matrix is excessive, there is value element mistake
Less so as to cause the very few problem of availability data；" cold start-up " problem refer to the behavioral data of new user it is very few cause system without
Method analyzes the problem of its preference.In recent years, with the development of online social networks, the personalized recommendation side based on socialization information
Method is increasingly paid attention to by industrial quarters and researcher.Socialization information mainly include social network and socialized label,
Hao Ma in 2008 et al. propose SoRec methods, and user-project rating matrix is decomposed simultaneously using the method that probability matrix decomposes
Matrix is trusted with user-user to be recommended；Mohsen in 2010 et al. proposes SocMF methods, during matrix decomposition
The difference for constraining the feature vector between user and user friend simultaneously, based on social networks using the matrix decomposition for trusting conduction
Method；Le Wu in 2012 et al. propose NHPMF methods, and using the label information of user and project, mould is decomposed in probability matrix
The label constraint item of user and project is added in type to carry out model training, and then obtains the potential feature square of user and project
Battle array, predicts user-project preference value.
For above-mentioned recommendation method to optimize score in predicting accuracy rate as target, there have in score in predicting problem to be higher accurate
Rate, but the sequencing problem of recommendation list is not accounted for, when carrying out Top-K recommendations with certain limitation.Top-K lists
Recommendation is considered as a sequencing problem, and study of sorting is a kind of method optimizing document ordering in information retrieval field.
By user-project to analogizing to inquiry-document of information retrieval field to being inputted as data, the method for study of sorting is also gradual
It applies in personalized recommendation field, similar with tradition sequence learning method, the sequence learning method in personalized recommendation is also led
It is divided into three categories, is Point-wise methods, Pair-wise methods and List-wise methods respectively.Point-wise methods
It is intended that single project forecast goes out accurate hobby value, traditional collaborative filtering method based on score in predicting belongs to Point-
Wise methods；Nathan etc. proposed the Bradley-TerryMF methods based on Pair-wise in 2010, and Xin Liu et al. exists
Propose RankNet-MF methods within 2014, the method based on Pair-wise instructs the partial ordering relation that project is liked using user
Practice model, to optimize the sorted lists of project；Yue Shi in 2010 et al. propose ListRank-MF methods, use cross entropy
(Cross-Entropy) it is used as loss function, it is proposed that a kind of matrix disassembling method based on List-wise sequence study, purport
Entire recommendation list is farthest being fitted and is being optimized.
The method of Point-wise is still the model towards score in predicting, does not account for the characteristic of sequence；Pair-wise
Method need to consider the partial ordering relation between all items, the complexity of model training is excessively high；Although ListRank-MF methods
The sequence for optimizing entire recommendation list is considered, the sequencing problem of project can be solved to a certain extent, but due in model
The information of middle involvement is very little, does not account for the influence of user and project socialization information, is still difficult to avoid that " Sparse
Property " and " cold start-up " problem, there is significant limitation in practical applications.
Invention content
The object of the present invention is to provide one kind closer to people's thinking habit, and there is the mosaic societyization compared with high-accuracy to believe
The personalized recommendation method of breath.
The present invention solves technical solution used by prior art problem：A kind of personalized recommendation of mosaic society's information
Method includes the following steps：
S1, structure user-user trust matrix：
A1, oriented degree of belief between user is obtained：Known containing in the database of socialization information, according to the database
In include social networks in concern relation between user obtain oriented degree of belief between user, oriented degree of belief between the user
Acquisition methods are as follows：
Wherein, tukIndicate user u oriented degree of belief, d between the user of user k in database social networks-(vk) indicate
The quantity that user k is concerned, d+(vu) indicate that user u pays close attention to the quantity of user；
A2, normalized：Then the oriented degree of belief t between userukNormalized is done, t is obtainedukUser between have
To degree of belief initialization value Tuk, and matrix is trusted by oriented degree of belief initialization value structure user-user between the user；
S2, structure project-item label similarity matrix：
B1, the weight vectors for obtaining project：Item label information in acquisition database, it is each to use tf*idf weights
Each label in project marks weight, and is constituted with tf*idf weighted values the element of the weight vectors of each project, project
The dimension of label weight vectors is the quantity of label in database, and the label weight vectors of project represent the label spy of project
Sign, the specific method is as follows：
Wherein, N is the number of project in database, wjtThe weighted value of label t in expression project j, tf (j, t) indicate item
Mesh j is marked with the number of label t, and when no visible marking's number is denoted as 1, df (t) and indicates the labeled project numbers of label t,
Do not have markd label weight to be denoted as 0 automatically；
Label similarity between b2, calculating project：
Pass through the label similarity between the cosine similarity acquisition project for the weight vectors for calculating two projects, calculation formula
It is as follows：
Wherein, sim (j1,j2) it is project j1With project j2Between label similarity,For the item obtained in step b1
Mesh j1With project j2Weight vectors；
B3, it chooses k nearest neighbor and normalizes the label similarity between project：According to the label similarity options between project
Purpose k nearest neighbor, the label similarity between the project of k nearest neighbor are normalized, between the k nearest neighbor project after being normalized
Label similarity, and the label similarity between the project except k nearest neighbor is set as 0, the project initialized-item label phase
Like degree matrix；
S3, structure and training pattern：
C1, acquisition training dataset：It randomly selects the 80% of the known database containing socialization information and is used as training number
According to collection；
C2, structure user-project rating matrix：
The each user for including is concentrated to assign the score value of 1-5 for the scoring of each project and with these points training data
Value builds user-project rating matrix as matrix element；
C3, the first probability of the user for some score value is calculated：
It is R for the scoring of project j to be calculated by the following formula user iijThe first probability Wherein,All meet for increasing function and for all xParameter D is that user i was commented
The number of entry divided；The first place probabilityIt is user i for the scoring R of project jijIs discharged in sorted lists
One probability value；
The list ordering of c4, structure and training mosaic society information learn recommended models, including：
The loss function L (U, V) of d1, structure mosaic society information：
Wherein,RmaxFor the highest score of scoring；IijFor indicator function, NiIt is the trust user of user i
Set, Np is the item destination aggregation (mda) with label similarity with project j, and M is number of users, and N is the number of entry, if user
I has scoring record to project j, and then value is 1, and otherwise value is 0；UiAnd VjIt is that the potential eigenmatrix of user and project are potential respectively
The row vector of eigenmatrix U, V, and the dimension of U and V is less than the quantity of number of users and project；WithUser is indicated respectively
Social networks penalty coefficient and item label penalty coefficient；WithTwo norms of respectively U and V；TijIndicate user i and
The degree of belief of user j, s (j, k) indicate the label similarity of project j and project k；
D2, the list ordering for creating mosaic society's information learn recommended models：
The gradient of user potential eigenmatrix U and the potential eigenmatrix V of project, and profit are obtained from the loss function L (U, V)
Loss function is trained with gradient descent method；Wherein, the ladder of the potential eigenmatrix U of user and the potential eigenmatrix V of project
The acquisition methods of degree are as follows：
S4, prediction user are for the unknown purpose preference：
Preference of the prediction user to project as follows：
R'ij=Ui TVj
Wherein R'ijIt is the preference that user i predicts project j, UiAnd VjIt is by the row of mosaic society's information respectively
The vector for the potential eigenmatrix of user and the potential eigenmatrix of project that list sorting study recommended models are trained.
The method that oriented degree of belief does normalized between user in step a2 is：
Wherein, TukFor tukUser between oriented degree of belief initialization value, NuFor user u all users set of interest.
In step b3, it is as follows that method is normalized in the label similarity between the project of the k nearest neighbor：
Wherein NjExpression project j1K nearest neighbor set.
The method being trained to loss function using gradient descent method is as follows：
E1. arrange parameter d, λu,λv, λ, learning rate η and maximum iteration maxIter；
E2. it is uniformly distributed according to [0,1] and the potential eigenmatrix of user and project potential eigenmatrix U, V is carried out initially
Change, it is all infinity, i.e. currError=preError=∞, iterations to initialize error current and last error
IterCount=0；
E3. the potential eigenmatrix U and potential eigenmatrix V of project of user is updated as follows：
E4. iterations add 1, error current are calculated according to the loss function L (U, V) in step d1, if error current
It is more than maximum iteration more than last error or iterations, i.e., then meets the condition of convergence to step volume e5；Otherwise, in order
First-order error is error current, continues e3；
E5. the potential eigenmatrix U of user and the potential eigenmatrix V of project, training process that output has restrained terminate.
In step c3, function
K in step b3 is 10.
, the socialization information includes social network and socialized label.
In step d1, the highest score R of scoringmaxIt is 5.
The beneficial effects of the present invention are：The present invention has following obvious advantage compared with existing collaborative filtering method：
1) traditional probability matrix decomposition method is using minimization score in predicting error as target, and the method for the invention with
It is target to be fitted entire entry sorting list.
2) present invention simultaneously using the socialization of user and project information, that is, users to trust degree and item label similarity as
The bound term of matrix decomposition makes have the feature vector of the user of larger degree of belief as similar as possible, meanwhile, there is larger label
The feature vector of the project of similarity is as close possible to being more nearly the thinking habit of the mankind.
3, in stand-alone environment (CPU is double-core 3.0GHz, inside saves as 4G), in d=5, λ=0.001, λu=0.05, λv=
When 0.001, item nearest neighbor number k=10, by taking NDCG@1 as an example, method disclosed by the invention and Random, UserAvg,
RankNet-MF, Bradley-TerryMF, ListRank-MF are compared and are respectively increased 16.6%, 6.8%, 12.4%, 12.5%,
0.9%.
Description of the drawings
Fig. 1 is the overview flow chart of the present invention.
Fig. 2 is users to trust matrix calculation flow chart.
Fig. 3 is item label similarity matrix calculation flow chart.
Fig. 4 is the training flow for learning recommended models to the list ordering of mosaic society's information using gradient descent method
Figure.
Specific implementation mode
Below in conjunction with the drawings and the specific embodiments, the present invention will be described：
Fig. 1 is a kind of overview flow chart of the personalized recommendation method of mosaic society's information of the present invention.A kind of fusion society
The personalized recommendation method that information can be changed, includes the following steps：
S1, structure user-user trust matrix：
A1, oriented degree of belief between user is obtained：Known containing in the database of socialization information, according in the database
Including social networks in concern relation between user obtain oriented degree of belief between user.Wherein, socialization information includes society
Change network and socialized label, socialization information are mainly derived from online social networks, such as bean cotyledon net；Concern between user is closed
If owner obtains from social networks, and this concern with to be concerned the social networks of relationship be in social networks by user
Actively statement it is as shown in table 1 below for include in the social networks of the known database containing socialization information user between close
Relation table is noted, the value for being 1 in table represents row user has concern relation to row user；
Concern relation table between 1 user of table
| User | 1 | 2 | 3 | 4 |
| 1 | 1 | 1 | 1 | |
| 2 | 1 | |||
| 3 | 1 | 1 | ||
| 4 | 1 | 1 |
Assuming that M user is co-existed in social networks, if tukUser u is indicated to the degree of belief of user k, then tukIt is bigger
Indicate that user k is bigger to the influence power of user's u interest；Conversely, user k is smaller to the influence power of user u.Meanwhile if user u
Multi-user is got in concern, then tukIt should be with reduction；If user k is paid close attention to by more multi-user, tukIt should increase.More than being based on
It analyzes, the acquisition methods of oriented degree of belief are as follows between user of the invention：
Wherein, tukIndicate user u oriented degree of belief, d between the user of user k in database social networks-(vk) indicate
The quantity that user k is concerned, d+(vu) indicate that user u pays close attention to the quantity of user；
A2, normalized：Then the oriented degree of belief t between userukNormalized is done, t is obtainedukUser between have
To degree of belief initialization value Tuk, i.e.,：
Wherein, TukFor tukUser between oriented degree of belief initialization value, NuFor user u institutes of interest
There is user's set.
Oriented degree of belief initialization value T between the userukTrust matrix P ∈ R as matrix element structure user-userM ×M。
The method that table 1 is trusted to matrix according to step S1, structure user-user obtains believing with user-user shown in table 2
The user-user that the element of degree matrix table is constituted is appointed to trust matrix, the element in table 2 represents letter of the row user to row user
The relationship of appointing, 0 indicates do not have trusting relationship, and trusting relationship is oriented, and has made normalized.
2 user-user of table trusts matrix element table
| User | 1 | 2 | 3 | 4 |
| 1 | 0 | 0.33 | 0.33 | 0.33 |
| 2 | 0 | 0 | 1 | 0 |
| 3 | 0.5 | 0 | 0 | 0.5 |
| 4 | 0 | 0.5 | 0.5 | 0 |
S2, structure project-item label similarity matrix：Detailed process is as shown in figure 3, include the following steps：
B1, the weight vectors for obtaining project：Assuming that a shared N number of project, L label, if label occurrence number is more,
The label is more important, while the project of label for labelling is more, then its discrimination is lower, thus in project i label t weight witIt adopts
With tf*idf weights weight is marked for each label of each project.
Item label information in acquisition database simultaneously uses tf*idf weights for each label label in each project
Weight, and the element of the weight vectors of each project is constituted with tf*idf weighted values, the dimension of the label weight vectors of project is
For the quantity of label in database, the label weight vectors of project represent the label characteristics of project, and the specific method is as follows：
Wherein, N is the number of project in database, wjtThe weighted value of label t in expression project j, tf (j, t) indicate item
Mesh j is marked with the number of label t, and when no visible marking's number is denoted as 1, df (t) and indicates the labeled project numbers of label t,
Do not have markd label weight to be denoted as 0 automatically；
So far, each project can be expressed as the weight vectors of L dimensions.
Label similarity between b2, calculating project：
Pass through the label similarity between the cosine similarity acquisition project for the weight vectors for calculating two projects, calculation formula
It is as follows：
Wherein, sim (j1,j2) it is project j1With project j2Between label similarity,For the item obtained in step b1
Mesh j1With project j2Weight vectors；
B3, it chooses k nearest neighbor and normalizes the label similarity between project：According to the label similarity options between project
Purpose k nearest neighbor, the value of K is preferably 10 herein.Label similarity between the project of k nearest neighbor is normalized, and obtains normalizing
The label similarity between k nearest neighbor project after change, and the label similarity between the project except k nearest neighbor is set as 0, it obtains just
The project of beginningization-item label similarity matrix Q ∈ RN×N；Label similarity wherein between the project of k nearest neighbor is normalized
Method is as follows：
Wherein NjExpression project j1K nearest neighbor set.
When neighbour's number k is set as 2, project shown in table 3-item label similarity matrix list of elements is obtained, member in table 3
Element represents the similarity of the row project and the list of items, and similarity has made normalized, and it is pair to have the matrix that the element is constituted
Claim matrix.
3 projects of the table-item label similarity matrix list of elements
S3, structure and training pattern：As shown in figure 4, including the following steps：
C1, acquisition training dataset：It randomly selects the 80% of the known database containing socialization information and is used as training number
According to collection；
C2, structure user-project rating matrix
Scoring R of each user for including by training data concentration for each projectijIt assigns 1 and arrives RmaxScore value, usually
It selects the score value of 1-5 and builds user-project rating matrix using these score values as matrix element；User-as shown in table 4 below
Project grade form is the partial data collection extracted from Epinions data sets.
4 users of table-project grade form
| 1 | 2 | 3 | 4 | 5 | 6 | |
| 1 | 2 | 4 | 4 | |||
| 2 | 3 | 5 | 5 | |||
| 3 | 3 | 1 | 4 | |||
| 4 | 3 | 5 | 2 |
C3, the first probability of the user for some score value is calculated
Assuming that a shared M user, N number of project, R are the matrix of a M × N, RijIndicate scorings of the user i to project j,
RijTypically one from 1 to Rmax, it is R for the scoring of project j to be calculated by the following formula user iijThe first probabilityWherein,All meet for increasing function and for all xSelect functionParameter D is that user i comments the excessive number of entry；
The first probabilityIt is i.e. the first general to illustrate that project is discharged to primary probability value in given sorted lists
Rate.Obviously, score value RijBigger, then user is bigger to the fancy grade of the project, and correspondingly the first probability value is higher, and more having can
It can be discharged to first in the ranking.
In information theory, usually a probability distribution and given probability point are weighed with cross entropy (Cross-Entropy)
The similarity degree of cloth, cross entropy is smaller, shows that two probability distribution are more similar, particularly, if two probability distribution complete one
It causes, then cross entropy is minimum.Similarly, the first probability distribution of prediction project sorted lists and known can be weighed with cross entropy
The similarity degree of the first probability distribution of entry sorting list.
The list ordering of c4, structure and training mosaic society information learn recommended models
Simultaneously in view of the interest that trusts each other between user is often more similar, while between the bigger user of degree of belief
The similarity of feature also tends to bigger, and the influence power between user also can be bigger；On the other hand, item label shown in table 5
As an important dimension of described project feature, have the characteristics that short and small refining, can largely reflect a project
Feature, therefore the feature vector between project between the more high then project of label similarity should be more similar.
5 item label table of table
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | |
| 1 | 1 | 1 | 1 | ||||
| 2 | 1 | 1 | 1 | ||||
| 3 | 1 | 1 | 1 | ||||
| 4 | 1 | 1 | 1 | ||||
| 5 | 1 | 1 | 1 | 1 | |||
| 6 | 1 | 1 | 1 | 1 |
Based on the above analysis, by the punishment for adding users to trust degree and label similarity in existing loss function
, obtain the list ordering study recommended models of mosaic society's information as follows.
The list ordering study of the loss function L (U, V) of d1, structure mosaic society information, mosaic society's information push away
Model is recommended to be built based on this loss function
Wherein,RmaxFor the highest score of scoring, generally 5；IijFor indicator function NiIt is user i
Trust the set of user, Np is the item destination aggregation (mda) with label similarity with project j, and M is number of users, and N is item number
Amount, value is 1 if user i has project j scoring to record, and otherwise value is 0；Ui、UjAnd VjIt is the potential feature square of user respectively
The row vector of battle array and the potential eigenmatrix U and V of project, and the dimension of U and V will be far smaller than the quantity of number of users and project；WithUser social contact relationship penalty coefficient and item label penalty coefficient are indicated respectively；WithThe two of respectively U and V
Norm；TijIndicate that the degree of belief of user i and user j, s (j, k) indicate the label similarity of project j and project k；
D2, the list ordering for creating mosaic society's information learn recommended models
The gradient of user potential eigenmatrix U and the potential eigenmatrix V of project, and profit are obtained from the loss function L (U, V)
Loss function is trained with gradient descent method；Wherein, the ladder of the potential eigenmatrix U of user and the potential eigenmatrix V of project
The acquisition methods of degree are as follows：
The method being trained to loss function using gradient descent method is as follows：
User-user is trusted into matrix, project-item label similarity matrix and user-project rating matrix as defeated
Enter,
E1. arrange parameter d, λu,λv, λ, learning rate η and maximum iteration maxIter, i.e. initiation parameter；
E2. it is uniformly distributed according to [0,1] and the potential eigenmatrix of user and project potential eigenmatrix U, V is carried out initially
Change, it is all infinity, i.e. currError=preError=∞, iterations to initialize error current and last error
IterCount=0；
E3. gradient is calculated according to formula (2) and (3), as follows by UiAnd VjUpdate realize it is latent to user
In the update of eigenmatrix U and the potential eigenmatrix V of project：
E4. iterations add 1, and error current currError is calculated according to formula (1).If judging, error current is more than
Last error, that is, currError>PreError or iterations are more than maximum iteration, that is, iterCount>MaxIter is then
Meet the condition of convergence to step 5；Otherwise, it is currError to enable preError, continues e3；
E5. the potential eigenmatrix U of user and the potential eigenmatrix V of project, training process that output has restrained terminate.
The potential eigenmatrix of user and the potential eigenmatrix list of elements difference of project are obtained as shown in table 6 and table 7, wherein joining
Number is set as d=5, λ=0.001, λu=0.05, λv=0.001, η=0.01, maximum iteration 500.It is each in table 6-7
Row indicates the feature vector U of row user (project)i(Vj)。
The potential eigenmatrix list of elements of 6 user of table
| 1 | 2 | 3 | 4 | 5 | |
| 1 | 0.557 | 0.097 | 0.55 | 0.516 | -0.074 |
| 2 | 0.58 | -0.402 | 0.55 | 0.364 | 0.186 |
| 3 | -0.051 | -0.561 | 0.456 | 0.585 | 0.036 |
| 4 | -0.235 | -0.489 | 0.403 | 0.68 | -0.048 |
The potential eigenmatrix list of elements of 7 project of table
| 1 | 2 | 3 | 4 | 5 | |
| 1 | -0.88 | -1.26 | -0.32 | -0.19 | 0.43 |
| 2 | -1.04 | 0.74 | -0.6 | 0.025 | -0.586 |
| 3 | 1.295 | 0.903 | -0.013 | -0.421 | 0.151 |
| 4 | -0.262 | -1.16 | 0.691 | 1.076 | 0.074 |
| 5 | 0.286 | 0.455 | 0.315 | 0.362 | -0.228 |
| 6 | 1.04 | 0.322 | -0.025 | -0.705 | 0.223 |
S4, prediction user are for the unknown purpose preference
Preference of the prediction user to project as follows：
R'ij=Ui TVj
Wherein R'ijIt is the preference that user i predicts project j, UiAnd ViIt is by the row of mosaic society's information respectively
The row vector for the potential eigenmatrix U of user and the potential eigenmatrix V of project that list sorting study recommended models are trained.
By taking user 1 as an example, can be obtained project 2, project 4, the preference value of project 6 that user 1 do not score to it be respectively-
0.782,0.67,0.22, thus according to preference value descending arrange user 1 recommended project list be project 4, project 6, project 2,
Similarly, the bulleted list that similar method can recommend the interested project of other users.So far, this hair
The method of bright proposition completes a recommended flowsheet.
Contrast on effect：
Apply the present invention to real data set Baidu film and recommends contest data set, abbreviation BaiduMovie data sets.
BaiduMovie data sets are disclosed by baidu company in the film commending system algorithm creative contest that in May, 2013 holds,
The data set owner will have following information：User-film scoring record, user's concern relation, film label information.It is wrapped in data set
1256998 scoring records of 7889 projects containing 9722 users couple, the density of score data is 1.64%, while these are used
Have 7898 concern relations between family, the density of concern relation is 0.0083%, there is 1121 labels, average each project have by
10 labels are marked.
It applies the method for the invention in BaiduMovie data sets, to keep experimental result more convincing, uses
5- rolls over cross validation method, data set is random and be averagely divided into 5 parts, chooses wherein 1 part successively and is used as test set, residue 4
Part data simultaneously obtain 5 experimental results as training set, 5 models of training, and the average value for choosing 5 results is tested as final
As a result.Unlike score in predicting problem, the present invention selects NDCG values as the index of evaluation sequence quality, user's set Q
In preceding K project ranking results NDCG calculation formula it is as follows：
Wherein ZuNormalization factor so that optimal sequence NDCG values be 1, R (u,p) represent in user's u optimal sequencings
P score values.
Experiment have chosen 5 existing recommendation methods, including recommend at random method (Random), UserAvg and two
Method (RankNet-MF and Bradley-TerryMF) based on Pair-wise, a method based on List-wise
(ListRank-MF), the experimental result comparison of 6 methods is as shown in table 8.
8 accuracy rate contrast table of table
The experimental results showed that method of the invention recommends accuracy rate more other, the recommendation method based on sequence study has obviously
Raising.
The above content is combine specific optimal technical scheme it is made for the present invention be further described, and it cannot be said that
The specific implementation of the present invention is confined to these explanations.For those of ordinary skill in the art to which the present invention belongs, exist
Under the premise of not departing from present inventive concept, a number of simple deductions or replacements can also be made, all shall be regarded as belonging to the present invention's
Protection domain.
Claims (8)
1. a kind of personalized recommendation method of mosaic society's information, which is characterized in that include the following steps：
S1, structure user-user trust matrix：
A1, oriented degree of belief between user is obtained：Known containing in the database of socialization information, wrapped according in the database
Concern relation between user obtains oriented degree of belief between user in the social networks contained, the acquisition of oriented degree of belief between the user
Method is as follows：
Wherein, tukIndicate user u oriented degree of belief, d between the user of user k in database social networks-(vk) indicate user k
The quantity being concerned, d+(vu) indicate that user u pays close attention to the quantity of user；
A2, normalized：Then the oriented degree of belief t between userukNormalized is done, t is obtainedukUser between oriented trust
Spend initialization value Tuk, and matrix is trusted by oriented degree of belief initialization value structure user-user between the user；
S2, structure project-item label similarity matrix：
B1, the weight vectors for obtaining project：Item label information in acquisition database, uses tf*idf weights for each project
In each label mark weight, and constituted with tf*idf weighted values the element of the weight vectors of each project, the label of project
The dimension of weight vectors is the quantity of label in database, and the label weight vectors of project represent the label characteristics of project,
The specific method is as follows：
Wherein, N is the number of project in database, wjtThe weighted value of label t in expression project j, tf (j, t) indicate project j
It is marked with the number of label t, when no visible marking's number is denoted as 1, df (t) and indicates the labeled project numbers of label t, does not have
The label weight of label is denoted as 0 automatically；
Label similarity between b2, calculating project：
By the label similarity between the cosine similarity acquisition project for the weight vectors for calculating two projects, calculation formula is such as
Under：
Wherein, sim (j1,j2) it is project j1With project j2Between label similarity,For the project j obtained in step b11
With project j2Weight vectors；
B3, it chooses k nearest neighbor and normalizes the label similarity between project：According to the label similarity options purpose between project
K nearest neighbor, the label similarity between the project of k nearest neighbor are normalized, the label between k nearest neighbor project after being normalized
Similarity, and the label similarity between the project except k nearest neighbor is set as 0, the project initialized-item label similarity
Matrix；
S3, structure and training pattern：
C1, acquisition training dataset：Randomly select the known database containing socialization information 80% is used as training dataset；
C2, structure user-project rating matrix：
Training data is concentrated each user for including assign the score value of 1-5 for the scoring of each project and is made with these score values
User-project rating matrix is built for matrix element；
C3, the first probability of the user for some score value is calculated：
It is R for the scoring of project j to be calculated by the following formula user iijThe first probability Wherein,All meet for increasing function and for all xParameter D is that user i was commented
The number of entry divided；The first place probabilityIt is user i for the scoring R of project jijIt is discharged to primary probability value；
The list ordering of c4, structure and training mosaic society information learn recommended models, including：
The loss function L (U, V) of d1, structure mosaic society information：
Wherein,RmaxFor the highest score of scoring；IijFor indicator function, M is number of users, and N is item number
Amount, the I if user i has project j scoring to recordijValue is 1, otherwise IijValue is 0；NiIt is the trust user set of user i,
Np is the project set most like with project j labels；UiAnd VjIt is the potential eigenmatrix of user and the potential eigenmatrix of project respectively
The row vector of U, V, and the dimension of U and V is less than the quantity of number of users and project；WithUser social contact relationship is indicated respectively
Penalty coefficient and item label penalty coefficient,Indicate regularization coefficient；WithTwo norms of respectively U and V；TijTable
Show that the degree of belief of user i and user j, s (j, k) indicate the label similarity of project j and project k；
D2, the list ordering for creating mosaic society's information learn recommended models：
From the gradient of acquisition user potential eigenmatrix U and the potential eigenmatrix V of project in loss function L (U, V), and utilize ladder
Degree descent method is trained loss function；Wherein, the gradient of the potential eigenmatrix U of user and the potential eigenmatrix V of project
Acquisition methods are as follows：
Wherein NlIt is the project set most like with project l labels, NkIt is the trust user set of user k；
S4, prediction user are for the unknown purpose preference：
Preference of the prediction user to project as follows：
Ri'j=Ui TVj
Wherein Ri'jIt is the preference that user i predicts project j, UiAnd VjIt is by the list ordering of mosaic society's information respectively
The vector for the potential eigenmatrix of user and the potential eigenmatrix of project that study recommended models are trained.
2. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that step a2
In oriented degree of belief does normalized between user method be：Wherein, TukFor tukUser between oriented letter
Appoint degree initialization value, NuFor user u all users set of interest.
3. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that step b3
In, it is as follows that method is normalized in the label similarity between the project of the k nearest neighbor：
Wherein NjExpression project j1K nearest neighbor set.
4. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that utilize ladder
The method that degree descent method is trained loss function is as follows：
E1. arrange parameter d, λu,λv, λ, learning rate η and maximum iteration maxIter；
E2. initial to the potential eigenmatrix of user and potential eigenmatrix U, V progress of project according to Z~U (0,1) is uniformly distributed
Change, it is all infinity, i.e. currError=preError=∞, iterations to initialize error current and last error
IterCount=0；
E3. the potential eigenmatrix U and potential eigenmatrix V of project of user is updated as follows：
E4. iterations add 1, error current are calculated according to the loss function L (U, V) in step d1, if error current is more than
Last error currError ＞ preError or iterations are more than maximum iteration iterCount ＞ maxIter, i.e.,
Then meet the condition of convergence to step e5；Otherwise, preError=currError continues e3；
E5. the potential eigenmatrix U of user and the potential eigenmatrix V of project, training process that output has restrained terminate.
5. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that step c3
In, function
6. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that step b3
In K be 10.
7. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that the society
It includes social networks and socialized label that information, which can be changed,.
8. a kind of personalized recommendation method of mosaic society's information according to claim 1, which is characterized in that step d1
In, the highest score R of scoringmaxIt is 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610067099.6A CN105740430B (en) | 2016-01-29 | 2016-01-29 | A kind of personalized recommendation method of mosaic society's information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610067099.6A CN105740430B (en) | 2016-01-29 | 2016-01-29 | A kind of personalized recommendation method of mosaic society's information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105740430A CN105740430A (en) | 2016-07-06 |
| CN105740430B true CN105740430B (en) | 2018-08-14 |
Family
ID=56247215
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610067099.6A Active CN105740430B (en) | 2016-01-29 | 2016-01-29 | A kind of personalized recommendation method of mosaic society's information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105740430B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110362738A (en) * | 2018-11-06 | 2019-10-22 | 杭州电子科技大学 | A kind of personalized recommendation method of combination trust and influence power based on deep learning |
Families Citing this family (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106682114B (en) * | 2016-12-07 | 2020-10-27 | 广东工业大学 | Personalized recommendation method integrating user trust relationship and comment information |
| CN106886559A (en) * | 2016-12-28 | 2017-06-23 | 浙江浙大网新集团有限公司 | The collaborative filtering method of good friend's feature and similar users feature is incorporated simultaneously |
| CN106980648B (en) * | 2017-02-27 | 2021-07-27 | 南京邮电大学 | Personalized recommendation method based on probability matrix decomposition and combined with similarity |
| CN107026755B (en) * | 2017-03-13 | 2019-07-16 | 南京邮电大学 | A kind of service recommendation method based on sequence study |
| CN107169830B (en) * | 2017-05-15 | 2020-11-03 | 南京大学 | Personalized recommendation method based on clustering PU matrix decomposition |
| CN107025311A (en) * | 2017-05-18 | 2017-08-08 | 北京大学 | A kind of Bayes's personalized recommendation method and device based on k nearest neighbor |
| CN107239993B (en) * | 2017-05-24 | 2020-11-24 | 海南大学 | Matrix decomposition recommendation method and system based on expansion label |
| CN107770365A (en) * | 2017-08-30 | 2018-03-06 | 天津大学 | A kind of prediction of multiple features APP streams and recommendation method based on Android |
| CN107507073A (en) * | 2017-09-14 | 2017-12-22 | 中国人民解放军信息工程大学 | Based on the service recommendation method for trusting extension and the sequence study of list level |
| CN107679242B (en) * | 2017-10-30 | 2018-07-27 | 河海大学 | Merge the label recommendation method that multiple information sources Coupling Tensor is decomposed |
| CN108132963A (en) * | 2017-11-23 | 2018-06-08 | 广州优视网络科技有限公司 | Resource recommendation method and device, computing device and storage medium |
| CN108230104A (en) * | 2017-12-29 | 2018-06-29 | 努比亚技术有限公司 | Using category feature generation method, mobile terminal and readable storage medium storing program for executing |
| CN108376152A (en) * | 2018-02-02 | 2018-08-07 | 西北工业大学 | A kind of worker's recommendation method based on matrix decomposition |
| CN108491469B (en) * | 2018-03-07 | 2021-03-30 | 浙江大学 | Neural collaborative filtering concept descriptor recommendation method introducing concept label |
| CN108681913A (en) * | 2018-04-04 | 2018-10-19 | 淮阴工学院 | A kind of digraph recommendation method based on AUC optimizations |
| CN108470075A (en) * | 2018-04-12 | 2018-08-31 | 重庆邮电大学 | A kind of socialization recommendation method of sequencing-oriented prediction |
| CN110209922B (en) * | 2018-06-12 | 2023-11-10 | 中国科学院自动化研究所 | Object recommendation method and device, storage medium and computer equipment |
| CN110717116B (en) * | 2018-06-27 | 2023-12-05 | 北京京东尚科信息技术有限公司 | Link prediction method and system of relational network, equipment and storage medium |
| CN108959655B (en) * | 2018-08-07 | 2020-04-03 | 南京大学 | Self-adaptive online recommendation method for dynamic environment |
| CN109597899B (en) * | 2018-09-26 | 2022-12-13 | 中国传媒大学 | Optimization method of media personalized recommendation system |
| CN110020883A (en) * | 2018-12-12 | 2019-07-16 | 阿里巴巴集团控股有限公司 | The method and device that unknown scoring in a kind of pair of rating matrix is predicted |
| CN109726747B (en) * | 2018-12-20 | 2021-09-28 | 西安电子科技大学 | Data fusion ordering method based on social network recommendation platform |
| CN109857928B (en) * | 2018-12-25 | 2020-12-29 | 北京交通大学 | User preference prediction method based on multivariate credit evaluation |
| CN109886779A (en) * | 2019-01-30 | 2019-06-14 | 上海大学 | A kind of intelligence commercial product recommending system |
| CN110008408B (en) * | 2019-04-12 | 2021-04-06 | 山东大学 | Session recommendation method, system, device and medium |
| CN110851705A (en) * | 2019-10-09 | 2020-02-28 | 天津大学 | Project-based collaborative storage recommendation method and recommendation device thereof |
| CN111177581A (en) * | 2019-12-25 | 2020-05-19 | 清华大学 | Multi-platform-based social e-commerce website commodity recommendation method and device |
| CN111292171B (en) * | 2020-02-28 | 2023-06-27 | 中国工商银行股份有限公司 | Financial product pushing method and device |
| CN111553657A (en) * | 2020-04-28 | 2020-08-18 | 贝壳技术有限公司 | Matching method and device based on behavior analysis, electronic equipment and storage medium |
| CN111723290B (en) * | 2020-06-09 | 2023-04-18 | 清华大学深圳国际研究生院 | User personalized preference prediction method based on multi-angle non-transmission preference relationship |
| CN111783980B (en) * | 2020-06-28 | 2023-04-07 | 大连理工大学 | Ranking learning method based on dual cooperation generation type countermeasure network |
| CN112069419B (en) * | 2020-09-08 | 2021-03-30 | 杭州电子科技大学 | Cross-correlation collaborative filtering method integrating user weak trace behavior preference |
| CN112883289B (en) * | 2021-04-16 | 2022-05-06 | 河北工程大学 | PMF recommendation method based on social trust and tag semantic similarity |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103092911B (en) * | 2012-11-20 | 2016-02-03 | 北京航空航天大学 | A kind of mosaic society label similarity is based on the Collaborative Filtering Recommendation System of k nearest neighbor |
| US9116894B2 (en) * | 2013-03-14 | 2015-08-25 | Xerox Corporation | Method and system for tagging objects comprising tag recommendation based on query-based ranking and annotation relationships between objects and tags |
| CN104268171B (en) * | 2014-09-11 | 2017-09-19 | 东北大学 | The social network friend recommendation system and method trusted based on the social activity of activity phase Sihe |
-
2016
- 2016-01-29 CN CN201610067099.6A patent/CN105740430B/en active Active
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110362738A (en) * | 2018-11-06 | 2019-10-22 | 杭州电子科技大学 | A kind of personalized recommendation method of combination trust and influence power based on deep learning |
| CN110362738B (en) * | 2018-11-06 | 2021-12-21 | 杭州电子科技大学 | Deep learning-based individual recommendation method combining trust and influence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105740430A (en) | 2016-07-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105740430B (en) | A kind of personalized recommendation method of mosaic society's information | |
| CN108763362A (en) | Method is recommended to the partial model Weighted Fusion Top-N films of selection based on random anchor point | |
| He et al. | SocoTraveler: Travel-package recommendations leveraging social influence of different relationship types | |
| CN105893609A (en) | Mobile APP recommendation method based on weighted mixing | |
| CN110210933B (en) | Latent semantic recommendation method based on generation of confrontation network | |
| CN112966091B (en) | Knowledge map recommendation system fusing entity information and heat | |
| CN108228867A (en) | A kind of theme collaborative filtering recommending method based on viewpoint enhancing | |
| CN105808786B (en) | A kind of collaboration theme recurrence label recommendation method based on content mutual trust | |
| CN111177538A (en) | Unsupervised weight calculation-based user interest tag construction method | |
| CN106897419A (en) | The study recommendation method that sorted to level of fusion social information | |
| CN108053050A (en) | Clicking rate predictor method, device, computing device and storage medium | |
| Zhao et al. | Improving recommendation accuracy using networks of substitutable and complementary products | |
| CN106384259A (en) | Recommend system solution method for fusing social information | |
| CN107016122A (en) | Knowledge recommendation method based on time-shift | |
| CN109213951A (en) | A kind of proposed algorithm calculated based on trust with matrix decomposition | |
| CN110334278A (en) | A kind of web services recommended method based on improvement deep learning | |
| CN104572915B (en) | One kind is based on the enhanced customer incident relatedness computation method of content environment | |
| Wang et al. | Research on hybrid collaborative filtering recommendation algorithm based on the time effect and sentiment analysis | |
| CN106021428A (en) | KNN and three-way decision-based movie recommendation method | |
| Duan et al. | A hybrid intelligent service recommendation by latent semantics and explicit ratings | |
| CN109857928A (en) | User preference prediction technique based on polynary credit evaluation | |
| CN109033241A (en) | News recommended method, device and electronic equipment | |
| CN103942278B (en) | Method for conducting friend recommendation through analysis of user active friends making will | |
| CN109711651B (en) | Hidden factor model wecker task recommendation method with correction vector | |
| CN111859113A (en) | Personalized content pushing method and device based on online taxi appointment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |