CN109408734B - Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust - Google Patents
Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust Download PDFInfo
- Publication number
- CN109408734B CN109408734B CN201811138086.9A CN201811138086A CN109408734B CN 109408734 B CN109408734 B CN 109408734B CN 201811138086 A CN201811138086 A CN 201811138086A CN 109408734 B CN109408734 B CN 109408734B
- Authority
- CN
- China
- Prior art keywords
- trust
- user
- recommendation
- users
- 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
- 238000001914 filtration Methods 0.000 title claims abstract description 47
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000004364 calculation method Methods 0.000 claims abstract description 73
- 230000000694 effects Effects 0.000 claims description 13
- 238000011156 evaluation Methods 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000012546 transfer Methods 0.000 claims description 9
- 230000008859 change Effects 0.000 claims description 6
- 230000003993 interaction Effects 0.000 claims description 5
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 2
- 239000011541 reaction mixture Substances 0.000 claims description 2
- 230000005540 biological transmission Effects 0.000 claims 1
- 230000004927 fusion Effects 0.000 abstract description 6
- 230000002829 reductive effect Effects 0.000 abstract description 6
- 230000002401 inhibitory effect Effects 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 13
- 230000007246 mechanism Effects 0.000 description 12
- 238000012360 testing method Methods 0.000 description 12
- 238000012549 training Methods 0.000 description 8
- 230000006399 behavior Effects 0.000 description 7
- 238000011160 research Methods 0.000 description 7
- 230000003068 static effect Effects 0.000 description 6
- 238000010835 comparative analysis Methods 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 208000037004 Myoclonic-astatic epilepsy Diseases 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000000874 microwave-assisted extraction Methods 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 238000003012 network analysis Methods 0.000 description 2
- 238000013179 statistical model Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a collaborative filtering recommendation method fusing information entropy similarity and dynamic trust, which is based on two similarity calculation methods of information entropy similarity and trust recessive similarity of grading difference, constructs a comprehensive similarity calculation model and relieves the problem that the similarity of a cold start user is difficult to calculate; then, the scoring credibility and the recommendation reliability are measured comprehensively, a direct, indirect and global credibility calculation model is constructed, and the false recommendation of the unreliable user is reduced; then, constructing a fusion score prediction model of similarity and trust, and completing score prediction and personalized project recommendation of a target user; and finally, evaluating the scoring effectiveness of the recommended user, proposing a trust punishment strategy, dynamically updating a trust neighbor set for the target user, and inhibiting the negative influence of random false scoring of the user on the recommendation performance. Experimental results show that the method can improve the recommendation precision and recommendation reliability of the recommendation system and effectively relieve the problems of data reliability, data sparsity and cold start.
Description
Technical Field
The invention relates to the technical field of personalized recommendation, in particular to a collaborative filtering recommendation method integrating information entropy similarity and dynamic trust.
Background
With the rapid development of information technology and social networks, data resources are growing explosively, and the problem of information overload needs to be solved urgently. The recommendation system is developed according to the practical problem of how to assist the user in efficiently filtering and recommending useful information in a personalized manner from mass data. Currently, recommendation systems can be classified into 5 types: content-based recommendations, collaborative filtering recommendations, knowledge-based recommendations, social recommendations, and hybrid recommendations. As an effective strategy for solving the problem of information overload, recommendation systems have been greatly developed in a variety of application fields such as electronic commerce (amazon. com, Alibaba, etc.), social networks (Facebook, Twitter, microblog, etc.), information retrieval (iGoogle, GroupLens, hundreds, etc.), music, video, movie recommendation, and the like.
The 20 th 90 th generation Collaborative Filtering (CF) algorithm proposes to provide personalized recommendations for users by applying a collective intelligence viewpoint of "people gather, objects gather" which is widely applied due to the advantages of domain independence and mining of potential interests of users. Collaborative filtering recommendations are mainly classified into 2 types:
1) model-based collaborative filtering: a probabilistic statistical model or a machine learning method is adopted, a user characteristic model is constructed on a training set for recommendation, and the model comprises a linear programming model, a statistical model, a Bayesian model, a probabilistic correlation model, a decision tree model and the like; the algorithm is characterized by long model training time and high complexity;
2) memory-based collaborative filtering: mining similarity relations of users or items according to the score matrix to carry out score prediction and recommendation, and classifying the similarity relations into User-based recommendation (User-CF) and Item-based recommendation (Item-CF); such algorithms are characterized by being overly reliant on memory data.
Data sparseness and cold start problems always exist in a collaborative filtering recommendation system, researchers around the problems provide a plurality of improvement methods for core steps of a collaborative filtering recommendation algorithm, such as improvement of a similarity calculation method, introduction of auxiliary data such as demographic information (age, gender, occupation and the like), filling of a score matrix vacancy value and the like. Analysis finds that the method focuses on algorithm self optimization and can effectively solve the problem in the research environment, but with the popularization and scale expansion of the application of the Social Network, the existing method cannot effectively solve the problems of data sparsity and cold start, so that the academic community draws on the Social Network Analysis (SNA) research theory, introduces Social relations (trust relations, friend relations and the like) among individuals as important influence factors into the Recommendation generation process, and provides a Social Recommendation method (Social Recommendation).
Due to the prevalence of Social networks (Facebook, Twitter, microblog, wechat) based on the Web 2.0 technology, the public constructs a virtual Social network formed by connecting complex Social relationships in an ad hoc manner on line, so that the Social Recommendation System (Social Recommendation System) gradually develops into the research focus of scholars at home and abroad in the field. Nowadays, a collaborative filtering recommendation (CF) model has become an important basic model for most social network recommendation systems, and a social network recommendation system based on collaborative filtering technology generally comprises two parts: a collaborative filtering recommendation (CF) model, a social relationship model based on social network analysis, formally described above as: the social network recommended CF model is the basic collaborative filtering recommended CF model + social relationship model. Trust is used as a core foundation of social relations and has important characteristics of directly influencing user behaviors and decisions, so that researchers use the trust relations as important dimensions of a social recommendation method. By constructing a trust network among users, for a cold start user in the system, as long as a user has a direct or indirect trust relationship with the user in the network, and the user can be recommended in a personalized manner based on a trust transfer relationship and a user interest model. Relevant research results show that the social recommendation system becomes an effective strategy for information filtering and personalized recommendation, the recommendation precision of the recommendation system can be improved by introducing trust recommendation close to the real life, and the problems of data sparsity and cold start of a recommendation algorithm are relieved to a greater extent.
Research on a collaborative filtering recommendation algorithm based on a trust network mainly focuses on the following three aspects:
1) and (3) trust reasoning: the method comprises the steps of efficiently searching for single (multiple) optimal or approximately optimal trust inference paths among nodes in a complex trust network;
2) and (3) trust calculation: constructing a self-adaptive dynamic trust calculation model to measure the trust degree between nodes in the network;
3) and (3) trust fusion: and constructing a fusion model of similarity and trust relationship to realize collaborative filtering recommendation so as to improve algorithm precision.
Around the above points, researchers have proposed many classical trust recommendation algorithms: a trust recommendation algorithm based on width (depth) first search and shortest path search, a trust calculation model based on random probability path search, a trust recommendation algorithm based on matrix decomposition and the like. These classic recommendation algorithms all provide a key algorithm theory for the research in this field, but with the development of social networks and recommendation systems, the above algorithms cannot well meet the personalized recommendation requirements of users, and the reasons are as follows:
1) false scores of user randomness or maliciousness exist in the recommendation system, most of the existing methods are based on the assumption that user scores have authenticity, and the reliability and the recommendation effectiveness of data are not fully considered, so that the recommendation precision of the recommendation system is not high;
2) trust has dynamics, most of the existing trust calculation models describe static trust relationships, and the problem of dynamic evolution of the trust relationships among users is not fully considered, so that the recommendation reliability of a recommendation system is not high.
Disclosure of Invention
Aiming at the defects in the prior art, the invention mainly researches the problems of recommendation reliability and trust dynamic evolution of users and provides a collaborative filtering recommendation method integrating information entropy similarity and dynamic trust. On one hand, the scoring reliability and recommendation reliability of the user are evaluated, and the false recommendation of the unreliable user is reduced; on the other hand, a trust reward and punishment updating mechanism is introduced, and the negative influence of random and false scoring behaviors of the user on the recommendation performance is inhibited. The method integrates similarity and dynamic trust relations, can effectively relieve the problems of data reliability, data sparsity and cold start, and improves the recommendation precision and reliability of the recommendation system.
The technical scheme for realizing the purpose of the invention is as follows:
a collaborative filtering recommendation method integrating information entropy similarity and dynamic trust comprises the steps of firstly, calculating user similarity and trust based on user item scoring and explicit trust relation of a recommendation system, and selecting a Top-K similarity and trust recommendation user set of a target user from big to small; then, a scoring prediction model is constructed by fusing the similarity and the trust, and the scoring of the target user on the project is predicted; secondly, selecting Top-N projects according with the target user preference according to the prediction scores, and completing personalized project recommendation of the target user; finally, evaluating the scoring effectiveness of the recommended users, and giving trust rewards and punishments to the active and passive recommended users;
in the recommendation system, the user item scoring data is represented as an m × n-order matrix R and comprises m user sets U ═ U1,u2,…,umAnd n sets of items I ═ I1,i2,…,inR, each element R in the matrix Rij(1 ≦ i ≦ m,1 ≦ j ≦ n) represents user uiFor item ijIf user u scoresiFor item ijNo score is expressed asThe explicit trust relationship between users is represented as m x m order trust matrix T, Tij(i is not less than 1 and not more than m, j is not less than 1 and not more than m) represents user uiWith user ujTrust relationship of, TijE is {0,1}, 0 represents distrust, and 1 represents trust; (ii) a
The method specifically comprises the following steps:
1) constructing an information entropy similarity calculation method based on the grading difference among users, and calculating the grading similarity RatingSim of the users;
2) constructing a trust relationship implicit similarity calculation method based on the explicit trust relationship, and calculating the trust similarity TrustSim of the user;
3) fusing RatingSim and TrustSim to construct a comprehensive similarity calculation model and calculate the comprehensive similarity SIM of the user;
4) analyzing the Trust relationship among users, constructing a Trust level Trust calculation model, and calculating the direct Trust level DTrust, the indirect Trust level ITrust and the global Trust level Auth of the users;
4-1) if explicit trust exists among users or scoring interaction of a co-rating item set exists among the users, calculating scoring credibility Cre, scoring prediction capability Cap and a co-rating item set Jaccard coefficient of the users; integrating three influence factors of the user scoring credibility Cre, the scoring prediction capability Cap and the co-scoring project set Jaccard coefficient, and calculating the recommendation credibility ReTrust of the target user to the recommendation user; integrating explicit trust and recommended trust to calculate the direct trust DTrust of the target user to the recommended user;
4-2) if no direct trust relationship exists among users but other trust transfer paths meeting the conditions exist, calculating single-path indirect trust or multi-path indirect trust ITrust according to the number of the paths;
4-3) if no direct trust and indirect trust relationship exists between the users, calculating the global trust Auth of the users, and using the global trust of the users as the trust of the target users to the recommended users;
5) obtaining user comprehensive similarity SIM and Trust Trust according to the step 3) and the step 4), selecting a Top-K similarity and Trust recommendation user set of a target user according to descending order from big to small, constructing a scoring prediction model by fusing the similarity and the Trust, and predicting the scoring of the target user on the project;
6) obtaining the prediction result of the target user on the unscored items according to the step 5), arranging the prediction result in a descending order from big to small, constructing a Top-N item recommendation list according with the preference of the target user, and recommending specific items to the target user in a personalized way;
7) and evaluating the effectiveness of the recommended user score according to the difference between the user feedback score and the recommended user score, and giving trust rewards and punishments for the active and passive recommended users.
In the step 1), the method for calculating the scoring similarity RatingSim comprises the following steps:
calculating the scoring difference information entropy RDE (u, v) of the users u and v by adopting a formula (1), and calculating the scoring similarity RatingSim (u, v) of the users u and v by adopting a formula (2):
in the formulae (1) and (2), NfailRepresenting the number of items with the evaluation attitudes of the users u and v to the common evaluation item set being opposite; n represents the number of the co-evaluation items; the absolute value of the score difference of the user u and v common score items is expressed as a set c ═ c1,c2,c3,…ck},p(ci) Indicating the probability of occurrence of each of the absolute values of the score difference.
In the step 2), the trust similarity TrustSim calculation method comprises the following steps:
in formula (3), t (u) and t (v) represent the set of neighbor nodes trusted by nodes u and v, respectively; t (u, v) ═ T (u) andt (v) denotes a set of neighbor nodes that are trusted in common by two users; din(k) Indicating the degree of the k-entry of the node.
In step 3), the method for calculating the comprehensive similarity of the user SIM comprises the following steps:
calculating the comprehensive similarity SIM between the users u and v by adopting a formula (4), and calculating the self-adaptive weight factor alpha between the scoring similarity RatingSim and the trust similarity TrustSim by adopting a formula (5):
SIM(u,v)=αRatingSim(u,v)+(1-α)TrustSim(u,v) (4)
in the step 4), the calculation method of the Trust level Trust comprises the following steps:
4-1) if explicit trust exists among users or scoring interaction of a co-scoring item set exists among users, fusing the explicit trust and the recommendation trust, and calculating the direct trust of the users by adopting a formula (6):
DTrust(u,v)=ρReTrust(u,v)+(1-ρ)Tuv (6)
in formula (6), DTrust (u, v) represents the direct trust level of the target user u for the recommended user v; ReTrust (u, v) represents the recommendation confidence level; t isuvRepresenting explicit trust of users u and v, TuvE is {0,1}, 0 represents distrust, and 1 represents trust; rho is a weight adjustment factor value of 0.5.
Calculating the recommended trust ReTrust (u, v) by adopting the formula (7):
ReTrust(u,v)=Jacu,v[βCrev+(1-β)Capu,v] (7)
in formula (7), CrevRepresenting the scoring credibility of the recommendation user v; capu,vThe score prediction capability of the recommended user v is represented; jacu,vA Jaccard coefficient representing a co-rating item of users u and v; beta is a weight adjusting factor, and the value of beta is 0.2;
calculating the scoring credibility Cre of the user v by adopting a formula (8)v:
Crev=Actv×Errv (8)
Calculating the user activity Act by adopting a formula (9)v:
In formula (9), ImA set of scoring items representing the mth user; ivL represents the number of user v scoring items;
calculating the user score deviation Err by adopting the formula (10)v:
In the formula (10), rv,iRepresents the rating of item i by user v;a score mean representing item i;
calculating recommended user using formula (11)Score prediction capability of Capu,v:
Capu,v=1-MSEu,v
In the formula (11), the reaction mixture,representing the predictive scoring of the co-scoring item set by taking the user v as the only recommended user of the user u; r isu,iRepresents the scoring of item i by user u; iuvL represents the number of the co-rated items of the two users;
In the formula (12), rv,iRepresents the rating of the recommended user v for the item i;andrespectively representing the average values of the scores of the users u and v; SIM (u, v) represents the integrated similarity between users u and v as stated in equation (4);
and (3) calculating the co-rating item Jaccard coefficient of the users u and v by adopting the formula (13):
4-2) if the users have no direct trust relationship but other trust transfer paths meeting the conditions, respectively calculating the single-path indirect trust or the multi-path indirect trust ITrush according to the number of the paths:
single path indirect trust: if there is one and only one shortest path P (u, u) reachable between users u and v1,u2… v), path length (path ≦ dmaxThe single-path indirect trust calculation is shown in equation (14):
in the formula (14), DTrust (u)1,u2) Indicating u between adjacent users on the path1For u is paired2Direct trust of; dpIndicating the length of the path;expressing the attenuation factor based on the belief propagation distance, the calculation formula is shown in (15):
in the formula (15), dmaxMaximum depth of 3, d representing a trusted path searchuvRepresents the trust path length between users u and v;
multipath indirect trust: if there are multiple shortest Paths reachable between user u and user v, the path set is expressed as Paths (u, v) ═ P1(u,v),P2(u,v),…Pn(u, v) }, where the path length d of any reachable path P (u, v) ∈ Paths (u, v)PSatisfy dp≤dmaxThen the multipath indirect trust calculation is shown in equation (16):
ITrust(u,v)=maxPaths(u,v)(ITrust(u,v)) (16)
formula (16) shows that the optimal or near optimal trust transfer path is selected from the multiple single paths in a maximum value mode;
4-3) if the direct trust and the indirect trust relationship do not exist between the users, calculating the global trust Auth of the users by adopting a formula (17):
in formula (17), Trust (k, u) represents the direct/indirect Trust of user k to user u in the Trust network; c denotes the number of users having a direct/indirect trust relationship with user u.
In step 5), the method for calculating the project prediction score is as follows:
in the formula (18), SN and TB respectively represent a Top-K similar user set and a trusted user set; SIM (u, v) represents the integrated similarity between users calculated by formula (4); trust (u, t) represents the Trust degree of the target user u on the Trust recommendation user t; theta represents the value of the weight adjustment parameter of 0.5.
And 6), carrying out personalized recommendation on the Top-N item which accords with the target user preference to the target user.
In step 7), the trust reward and punishment method is as follows:
7-1) using the calculation method of formula (19) to give trust rewards to actively recommending users:
in the formula (19), the first and second groups,representing the trust degree of the target user u to the recommended user v after the trust is updated;representing the trust degree of the target user u to the recommended user v before the trust updating; mSA control factor indicating a maximum value of the trust change amount, andESrepresenting the influence of the recommendation score error on the reward amplitudeWherein Q represents the maximum value of the scoring error in the recommendation system, R represents the maximum value of the item rating, and when the scoring error d is equal to | Ru,i-rv,iWhen | ≦ epsilon, the larger the error d, the smaller the reward value; pNExpressing the influence factor of the user recommendation times on the reward forceWherein ω CrevRepresenting the user scoring credibility; pSRepresenting the recommendation success rate of the recommendation node, and when the recommendation user v actively recommends the target user u, namely | ru,i-rv,iWhen | < epsilon, the recommendation is successful, PSThe calculation formula is shown as (20):
in the formula (20), N ═ TotalSet (u, v) | represents the total number of times of recommendation of the target user u by the recommending user v; n is a radical ofsRepresenting the successful recommendation times of the target user u by the recommending user v;
7-2) performing trust penalty on the negative recommended users by adopting a calculation method of a formula (21):
in the formula (21), MfA control factor indicating a maximum value of the trust change amount, andEfrepresenting the impact of the recommendation score error on the penalty magnitude, orderWhen inter-user scoring error d ═ ru,i-rv,i|>When epsilon is larger, the larger the error d is, the larger the penalty value is; pNFactor representing influence of user recommendation times on penalty degree, orderPfRepresenting the recommendation failure rate of the recommendation node, when the recommendation user v makes a negative recommendation to the target user u, namely | ru,i-rv,i|>When ε, the recommendation fails, PfThe calculation formula is shown as (22):
in the formula (22), Nf| False (u, v) | represents the number of failed recommendations of the target user u by the recommending user v.
Compared with the prior art, the method has the following advantages:
1) constructing an information entropy interest similarity calculation method based on score difference, introducing a trust relationship implicit similarity for avoiding a trust network hotspot effect, and combining the two in a self-adaptive dynamic mode to calculate the comprehensive similarity of users, so that the problem that the score similarity accuracy is low or the similarity of cold-start users is difficult to calculate due to data sparsity is solved;
2) according to the user rating data, calculating recommendation trust degrees from three dimensions of rating trust degrees, prediction rating capabilities and Jaccard coefficients of a common rating item set, fusing explicit trust and recommendation trust, accurately calculating direct trust degrees among users, and reducing false recommendation of unreliable users; analyzing and researching the single-path trust weak propagation and multi-path trust aggregation problem based on the reconstructed trust network, and constructing indirect and global trust degree calculation models;
3) a trust reward and punishment updating mechanism is introduced, so that the negative influence of random and false scoring behaviors of a user on the recommendation performance is inhibited; evaluating the scoring effectiveness of the recommended users, reasonably rewarding and punishing the trust degrees of the active recommenders and the passive recommenders respectively, representing the dynamic evolution of the trust relationship among the users, and dynamically updating the trust neighbor set for the target user.
Drawings
FIG. 1 is a schematic flow chart of a collaborative filtering recommendation method fusing information entropy similarity and dynamic trust;
FIG. 2 is a schematic diagram of Pearson similarity;
FIG. 3 is a schematic diagram of improved Pearson similarity;
FIG. 4 is a diagram illustrating cosine similarity;
FIG. 5 is a schematic view of integrated similarity;
FIG. 6 is a comparison of MAEs for different similarity algorithms;
FIG. 7 is a graph of the effect of θ on MAE;
FIG. 8 is a graph of the effect of θ on RMSE;
FIG. 9 is a comparison of MAEs for different recommendation algorithms;
FIG. 10 is a comparison of Recall (RL) for different recommendation algorithms;
fig. 11 is a comparison of MAEs for different recommendation algorithms.
Detailed Description
The present invention will be described in further detail with reference to the following drawings and examples, but the present invention is not limited thereto.
Example (b):
as shown in fig. 1, a collaborative recommendation method fusing information entropy similarity and dynamic trust specifically includes the following steps:
this example uses the FilmTrust movie scoring dataset: the scoring data comprises 35497 scoring of 1508 users on 2071 movies, the scoring value interval is 0.5-4, and the sparsity of the scoring data is 98.86%; the trust data includes 1853 explicit trust relationships between 1642 users, with the sparsity of the trust data being 99.93%.
1) Constructing an information entropy similarity calculation method based on the grading difference among users, and calculating the grading similarity RatingSim of the users:
wherein N isfailRepresenting the number of items with the evaluation attitudes of the users u and v to the common evaluation item set being opposite; n represents the number of the co-evaluation items; the absolute value of the score difference of the user u and v common score items is expressed as a set c ═ c1,c2,c3,…ck},p(ci) Indicating the probability of occurrence of each of the absolute values of the score difference.
2) Constructing a trust relationship implicit similarity calculation method based on an explicit trust relationship, and calculating the trust similarity TrustSim of the user:
wherein T (u) and T (v) represent the set of neighbor nodes trusted by nodes u and v, respectively; t (u, v) ═ T (u) andt (v) denotes a set of neighbor nodes that are trusted in common by two users; din(k) Indicating the degree of the k-entry of the node.
3) And (3) fusing the RatingSim and the TrustSim to construct a comprehensive similarity calculation model, and calculating the comprehensive similarity SIM of the user:
SIM(u,v)=αRatingSim(u,v)+(1-α)TrustSim(u,v) (4)
4) analyzing the Trust relationship among users, constructing a Trust level Trust calculation model, and calculating the direct Trust level DTrust, the indirect Trust level ITrust and the global Trust level Auth of the users:
4-1) if explicit trust exists among users or scoring interaction of a co-rating item set exists among users, fusing the explicit trust and the recommendation trust, and calculating the direct trust of the users:
DTrust(u,v)=ρReTrust(u,v)+(1-ρ)Tuv (6)
wherein DTrust (u, v) represents the direct trust degree of the target user u on the recommended user v; ReTrust (u, v) represents the recommendation confidence level; t isuvRepresenting explicit trust of users u and v, TuvE is {0,1}, 0 represents distrust, and 1 represents trust; rho is a weight adjustment factor value of 0.5.
Calculating the recommended trust ReTrust (u, v) by adopting the formula (7):
ReTrust(u,v)=Jacu,v[βCrev+(1-β)Capu,v] (7)
wherein CrevRepresenting the scoring credibility of the recommendation user v; capu,vThe score prediction capability of the recommended user v is represented; jacu,vA Jaccard coefficient representing a co-rating item of users u and v; beta is the value of the weight adjusting factor of 0.2.
Calculating the scoring credibility Cre of the user v by adopting a formula (8)v:
Crev=Actv×Errv (8)
Calculating the user activity Act by adopting a formula (9)v:
Wherein, ImA set of scoring items representing the mth user; ivL represents the number of user v scoring items;
calculating the user score deviation Err by adopting the formula (10)v:
calculating the score prediction capability Cap of the recommended user by adopting a formula (11)u,v:
Capu,v=1-MSEu,v
Wherein,representing the predictive scoring of the co-scoring item set by taking the user v as the only recommended user of the user u; r isu,iRepresents the scoring of item i by user u; iuvL represents the number of the co-rated items of the two users;
Wherein r isv,iRepresents the rating of the recommended user v for the item i;andrespectively representing the average values of the scores of the users u and v; SIM (u, v) represents the integrated similarity between users u and v as stated in equation (4);
and (3) calculating the co-rating item Jaccard coefficient of the users u and v by adopting the formula (13):
4-2) if the users have no direct trust relationship but other trust transfer paths meeting the conditions, respectively calculating the single-path indirect trust or the multi-path indirect trust ITrush according to the number of the paths:
single path indirect trust: if users u andv has one and only one shortest path P ═ to (u, u) that can be reached1,u2… v), path length (path ≦ dmaxThe single-path indirect trust calculation is shown in equation (14):
among them, DTrust (u)1,u2) Indicating u between adjacent users on the path1For u is paired2Direct trust of; dpIndicating the length of the path;expressing the attenuation factor based on the belief propagation distance, the calculation formula is shown in (15):
wherein d ismaxMaximum depth of 3, d representing a trusted path searchuvRepresents the trust path length between users u and v;
multipath indirect trust: if there are multiple shortest Paths reachable between user u and user v, the path set is expressed as Paths (u, v) ═ P1(u,v),P2(u,v),…Pn(u, v) }, where the path length d of any reachable path P (u, v) ∈ Paths (u, v)PSatisfy dp≤dmaxThen the multipath indirect trust calculation is shown in equation (16):
ITrust(u,v)=maxPaths(u,v)(ITrust(u,v)) (16)
formula (16) shows that the optimal or near optimal trust transfer path is selected from the multiple single paths in a maximum value mode;
4-3) if the direct trust and the indirect trust relationship do not exist between the users, calculating the global trust Auth of the users:
wherein Trust (k, u) represents the direct/indirect Trust of the user k to the user u in the Trust network; c denotes the number of users having a direct/indirect trust relationship with user u.
5) Obtaining the comprehensive similarity SIM and the Trust level Trust of the user according to the step 3) and the step 4), selecting a Top-K similarity and Trust recommendation user set of the target user according to descending order from big to small, building a scoring prediction model by fusing the similarity and the Trust level, and predicting the scoring of the target user on the project:
wherein SN and TB respectively represent a Top-K similar user set and a trusted user set; SIM (u, v) represents the integrated similarity between users calculated by formula (4); trust (u, t) represents the Trust degree of the target user u on the Trust recommendation user t; theta represents the value of the weight adjustment parameter of 0.5.
6) Obtaining the prediction result of the target user on the unscored items according to the step 5), arranging the prediction result in a descending order from big to small, constructing a Top-N item recommendation list according with the preference of the target user, and recommending specific items to the target user in a personalized way;
7) evaluating the effectiveness of the recommended user score according to the difference between the user feedback score and the recommended user score, and giving trust rewards and punishments to positive and negative recommended users:
7-1) using the calculation method of formula (19) to give trust rewards to actively recommending users:
wherein,representing the trust degree of the target user u to the recommended user v after the trust is updated;representing the trust degree of the target user u to the recommended user v before the trust updating; mSA control factor indicating a maximum value of the trust change amount, andESrepresenting the influence of the recommendation score error on the reward amplitudeWherein Q represents the maximum value of the scoring error in the recommendation system, R represents the maximum value of the item rating, and when the scoring error d is equal to | Ru,i-rv,iWhen | ≦ epsilon, the larger the error d, the smaller the reward value; in FilmTrust dataset, Q is 4, R is 3.5, and e is 0.5; pNExpressing the influence factor of the user recommendation times on the reward forceWherein ω CrevRepresenting the user scoring credibility; pSRepresenting the recommendation success rate of the recommendation node, and when the recommendation user v actively recommends the target user u, namely | ru,i-rv,iWhen | < epsilon, the recommendation is successful, PSThe calculation formula is shown as (20):
wherein N ═ TotalSet (u, v) | denotes the total number of times of recommendation of the target user u by the recommending user v; n is a radical ofsRepresenting the successful recommendation times of the target user u by the recommending user v;
7-2) performing trust penalty on the negative recommended users by adopting a calculation method of a formula (21):
wherein M isfControl factor representing maximum value of trust change amountLet us orderEfRepresenting the impact of the recommendation score error on the penalty magnitude, orderWhen inter-user scoring error d ═ ru,i-rv,i|>When epsilon is larger, the larger the error d is, the larger the penalty value is; pNFactor representing influence of user recommendation times on penalty degree, orderPfRepresenting the recommendation failure rate of the recommendation node, when the recommendation user v makes a negative recommendation to the target user u, namely | ru,i-rv,i|>When ε, the recommendation fails, PfThe calculation formula is shown as (22):
wherein N isf| False (u, v) | represents the number of failed recommendations of the target user u by the recommending user v.
The experimental results using the FilmTrust dataset for this example are as follows:
(1) performance evaluation indexes are as follows:
mean Absolute Error (MAE), where MAE represents the Mean Absolute Error between the predicted score value and the true score value, and a smaller MAE value indicates higher recommendation accuracy of the algorithm, and MAE calculation is shown in formula (23):
wherein, PkIndicates the value of the prediction score, RkRepresenting the true score value, and n represents the number of scoring predictions.
Root Mean Square Error (RMSE), where RMSE represents a Root of a ratio of a sum of squares of deviations of the predicted score values from the true score values to the number of times the score is predicted, and a smaller value indicates a higher recommendation accuracy of the algorithm, and RMSE is calculated as shown in equation (24):
recall (Recall, RL), also known as Recall, represents the ratio of the number of scores that can be predicted by the algorithm to all the scores to be measured, and is calculated as shown in equation (25):
wherein, M represents the number of the prediction scores available for the algorithm, and N represents the number of the scores to be tested in the test set.
(2) Comparison of similarity algorithms:
in order to verify the accuracy of the comprehensive similarity algorithm (SIM) proposed by the present invention, the SIM algorithm is compared with the following algorithms:
cosine similarity (Cosine): measuring the difference of the users according to the cosine value of the included angle of the scoring vectors of the two users in the vector space; the algorithm represents the difference of the two vectors in the direction, and does not represent the difference of the distance or the length;
pearson similarity (Pearson): calculating the quotient of covariance and standard deviation between the two variables, wherein the algorithm is used for measuring the strength of linear correlation of the two variables;
improved Pearson similarity (Adjusted _ Pearson): introducing a Logistic function of the number of common scoring items among users to optimize a calculation formula of the Pearson similarity;
experiment one: comparative analysis of similarity accuracy
From the FilmTrust score data, 80% were randomly selected as training set, and the remaining 20% were selected as test set. Selecting a 509 th user with a large number of scoring items as a target user, randomly selecting 1000 different users from the training set, and calculating the similarity between the target user and the 1000 users by respectively adopting Pearson, Adjusted _ Pearson, Cosine and SIM, wherein the distribution conditions of the four similarity calculation results are respectively shown in FIG. 2, FIG. 3, FIG. 4 and FIG. 5.
As can be seen from fig. 4, the similarity calculation result of the Cosine similarity (Cosine) algorithm is mainly distributed in the interval [0.8,1], and part of the similarity calculation value is exaggerated and inaccurate; moreover, because the distribution interval of the similarity is narrow, the difference degree shown between users is small, and the system is difficult to accurately screen the Top-K neighbor similar recommended users of the target user; this is because the cosine algorithm characterizes the difference in direction of the two vectors, and does not reflect the difference in distance or length; the experimental result shows that the similarity precision of the Cosine similarity (Cosine) algorithm is low.
As can be seen from fig. 2, the similarity calculation results of the Pearson correlation coefficient (Pearson) algorithm are mainly distributed in the interval [ -0.6,0.8], the similarity distribution interval is greatly expanded compared with the cosine similarity, but because the Pearson correlation coefficient has a negative number, a user negatively correlated with a target user cannot be used as a Top-K neighbor recommendation user of the target user, which is contrary to some practical situations; the reason is that the pearson correlation coefficient is only used for measuring the strength of linear correlation of two variables, and the similarity between users cannot be reflected completely and accurately; the experimental result shows that compared with cosine similarity, the accuracy of the Pearson similarity is improved to a certain extent, but the Top-K neighbor similarity recommendation user of the missed selection target user is inaccurate.
As can be seen from FIG. 3, the calculation results of the improved Pearson similarity (Adjusted _ Pearson) algorithm are mainly distributed in the interval [ -0.4,0.6], and compared with FIG. 2, the results of the improved Pearson similarity algorithm with a positive correlation similarity of 1 and a negative correlation similarity of-1 in FIG. 3 are obviously reduced, because the improved Pearson similarity algorithm introduces Logistic function influence factors of the common score items among users, so that the similarity calculation results are optimized to a certain extent. However, as can be seen from comparison between fig. 2 and fig. 3, the algorithm does not really solve the problem of negative correlation limitation existing in the pearson correlation coefficient itself, and there still exists an inaccuracy condition that Top-K of the target user is missed to select a similar recommending user.
As can be seen from fig. 5, the similarity calculation results of the integrated Similarity (SIM) algorithm proposed by the present invention are mainly distributed in the interval [0.4,0.9], and the similarity values in the interval are relatively uniformly distributed and have higher accuracy; compared with fig. 2, 3 and 4, the SIM algorithm avoids the imprecision situation that the similarity distribution density and the coincidence degree are too high; the comprehensive similarity calculation method SIM provided by the invention not only utilizes the information entropy of the grading difference between users to effectively quantify the difference between the users, but also considers the influence factors of the common-grade item sets between the users, and also fuses the trust implicit similarity between the users, so the comprehensive similarity calculation method SIM obtains the similarity with higher precision.
The experimental results are as follows: the comprehensive similarity precision of the user calculated by the two dimensions of the information entropy similarity and the trust relationship implicit similarity based on the grading difference is obviously improved, and the Top-K neighbor similar recommendation user of the target user can be screened out more accurately.
Experiment two: comparative analysis of similarity algorithm recommendation performance
Randomly selecting 80% from the FilmTrust scoring data as a training set, and taking the rest 20% as a test set; and selecting 10 different values of the neighbor user number K of the target user, wherein the number K is {5,10, …,60 }. The mean absolute error between the predicted score and the true score is calculated on the test set, and the comparison result of the recommendation accuracy MAE of the four similarity algorithms is shown in fig. 6.
As can be seen from fig. 6, as the number K of the neighboring users increases, the MAE values of the four similarity recommendation algorithms keep a generally consistent decreasing trend, which indicates that the recommendation accuracy all increases as the K value increases, and the recommendation accuracy tends to be more stable as the K value is larger; on different K values, the MAE values of the SIM of the comprehensive similarity calculation method are smaller than those of other similarity calculation methods, so that the recommendation performance of the SIM of the comprehensive similarity calculation method is superior to that of other similarity calculation methods, and the experiment proves the higher accuracy of the comprehensive similarity of the SIM of the comprehensive similarity calculation method; the comprehensive similarity calculation method SIM provided by the invention not only utilizes the information entropy of the grading difference between users to effectively quantify the difference between the users, but also considers the influence factor of the co-rating item set between the users, and introduces the trust relationship to imply the similarity, thereby further relieving the problem that the similarity of the cold-start users is difficult to calculate, and the precision of the similarity obtained by combining the two dimension calculations is obviously improved, thereby selecting more accurate Top-K neighbor similarity recommendation users.
The experimental results are as follows: the comprehensive similarity calculation method provided by the invention has higher SIM similarity precision, can more accurately screen out Top-K neighbor similar recommended users of target users, and improves the recommendation reliability of a recommendation system. (3) Influence of weight adjustment factor
Experiment three: analysis of influence of weight adjustment parameter theta on recommendation performance of CF-IESDT algorithm
The experiment is mainly used for analyzing the influence of the theta weight adjusting parameter in the score prediction formula (20) for fusing similarity and trust, which is provided by the invention, on the recommendation performance of the CF-IESDT algorithm; and theta is a weight adjusting parameter of the similar recommended user set and the trust recommended user set in the score prediction, and the value range of theta is [0,1 ].
Randomly selecting 90% from the FilmTrust scoring data as a training set, and taking the rest 10% as a test set; and selecting 5 different values of the neighbor user number K of the target user, wherein the number K is {5,10, …,25 }. Taking θ ═ 0,0.1,0.2, …,1 on the test set, the mean absolute error MAE and the root mean square error RMSE between the predicted score and the true score were calculated, the results are shown in tables 1 and 2, and the effect of different θ values on the recommended precision MAE and RMSE of the CF-IESDT algorithm is shown in fig. 7 and 8.
As can be seen from fig. 7 and 8, the recommendation accuracy MAE and RMSE of the CF-IESDT algorithm both vary with different θ values, and as the number K of neighboring users increases, both the MAE value and the RMSE value decrease; when theta is 1, the scoring prediction completely depends on a Top-K similar neighbor recommended user set screened by the comprehensive similarity provided by the invention, and the recommendation precision MAE value and the RMSE value are both high; when theta is 0, the scoring prediction completely depends on the Top-K trust neighbor recommended user set screened by the trust degree, and the recommendation precision MAE value and the RMSE value are also higher; when the value of theta is changed from 0-0.5, the MAE value and the RMSE value are reduced, but when the value of theta is more than 0.5 and less than or equal to 1, the MAE value and the RMSE value are increased along with the increase of theta.
The experimental results are as follows: when the value of theta approaches to 0.5, the values of the recommended precision MAE and the RMSE of the CF-IESDT algorithm reach approximately the optimal value (namely the minimum value). Therefore, in the following recommended performance comparison experiment, the weight adjustment parameter θ takes a value of 0.5.
(4) Recommended performance comparison
Jia Dongyan proposes a collaborative filtering recommendation algorithm (CF-DNC) based on a double neighbor selection strategy, and has similarity with the collaborative filtering recommendation algorithm (CF-IESDT) which fuses information entropy similarity and dynamic trust and is proposed by the invention. The CF-DNC algorithm selects an interest similar TOP-K neighbor user set of a target user based on the improved Pearson similarity, then calculates the score prediction capability of the interest similar TOP-K neighbor user by using a leave-one-out method so as to represent the trust of the target user on the similar user, and finally takes the trust as the basis for selecting the credible recommendation neighbor user. Compared with the algorithm CF-IESDT provided by the invention, the CF-DNC considers whether the recommended user and the target user are similar in interest and the rating prediction capability of the recommended user, but the CF-DNC does not expand the credibility implicit similarity and does not consider a plurality of credibility influence factors such as the credibility of the user, and the algorithm does not really fuse the similarity relation and the credibility relation.
Experiment four: comparative analysis of recommended performance of CF-IESDT algorithm
In order to verify the recommendation precision and recommendation reliability of the CF-IESDT algorithm, the following classical collaborative filtering recommendation algorithm is compared and analyzed:
collaborative filtering recommendation algorithm (CF);
a collaborative filtering recommendation algorithm (CF-DNC) based on a dual neighbor selection strategy;
meanwhile, in order to verify the rationality of the CF-IESDT algorithm fusion similarity relation and the trust relation and the effectiveness of a trust reward and punishment mechanism, the following recommendation performances of the collaborative filtering recommendation algorithm provided by the invention are respectively observed through experiments:
a collaborative filtering recommendation algorithm (RatingSim) based on score similarity;
collaborative filtering recommendation algorithm (TrustSim) based on trust similarity;
a collaborative filtering recommendation algorithm (SIM) based on integrated similarity;
a Dynamic Trust collaborative filtering recommendation algorithm (Dynamic _ Trust) based on Trust and punishment;
a static Trust collaborative filtering recommendation algorithm (Trust) based on no Trust and punishment;
a collaborative filtering recommendation algorithm (CF-IEST) that integrates similarity with static trust;
from the FilmTrust score data, 90% were randomly selected as training set, and the remaining 10% were selected as test set. Under the same experiment environment, the number K of the neighbor users is 5 different values {5,10, …,25}, and the score prediction weight adjusting parameter theta is 0.5. The average absolute error MAE between the prediction score and the true score is calculated on the test set by using the above recommendation algorithms, the result is shown in table 3, and the recommendation accuracy MAE comparison of the algorithms is shown in fig. 9.
As can be seen from FIG. 9, as the number K of the neighboring users increases, the recommendation accuracy MAE of the CF-IESDT algorithm is higher than that of the CF and CF-DNC algorithms, which indicates that the CF-IESDT algorithm has relatively optimal recommendation performance. The main performance is as follows:
1) the recommendation precision MAE of the RatingSim, SIM and Dynamic-Trust algorithms provided by the invention is obviously higher than that of the CF and CF-DNC algorithms: for example, when the number K of the neighbors is 15, the recommendation accuracy MAE of the three algorithms is improved by about 20% and 4% compared with that of CF and CF-DNC; when the number K of the neighbors is 25, the recommendation accuracy MAE of the three algorithms is improved by about 22% and 8% compared with that of CF and CF-DNC respectively. Therefore, the invention respectively obtains more accurate similarity and Trust degree through the SIM and the Dynamic _ Trust algorithm. The SIM algorithm respectively provides an information entropy interest similarity calculation method based on score difference and a trust relationship implicit similarity calculation method for avoiding the hotspot effect of the trust network, and more accurate comprehensive similarity is obtained by fusing interest similarity and trust similarity. Similarly, in order to integrate interest similarity and Trust relations, the Dynamic _ Trust algorithm firstly measures the recommendation Trust of the user according to the user scoring data from three dimensions of user scoring credibility, scoring prediction capability and Jaccard coefficient of a common scoring item set, and the essence of the recommendation Trust is also the representation of the similarity; then the direct trust, the indirect trust and the global trust of the user are accurately calculated by fusing the explicit trust with the direct trust; the influence of unreliable data on the calculation of the credibility is reduced by comprehensively measuring the credibility and the reliability of the rating recommendation of the user; and the Dynamic _ Trust algorithm introduces a Trust punishment updating mechanism, dynamically adjusts the Trust degree of the target user to the recommended user according to the recommendation result between users every time, effectively reduces the false recommendation of the unreliable user along with the system operation and the user feedback, and obtains more accurate and objective user Trust degree.
2) The recommendation precision MAE of the CF-IESDT algorithm which is similar and trusted in fusion provided by the invention is higher than that of the SIM and Dynamic-Trust algorithms before fusion, and is also obviously higher than that of the CF and CF-DNC algorithms: for example, when the number K of the neighbors is 15, the recommendation accuracy MAE of the CF-IESDT algorithm is improved by about 21% and 6% compared with CF and CF-DNC respectively, and the recommendation accuracy MAE is improved by 1% and 2% again compared with the SIM and the Dynamic _ Trust algorithm; when the number K of the neighbors is 25, the recommendation precision MAE of the CF-IESDT algorithm is respectively improved by about 23% and 9% compared with CF and CF-DNC, and compared with the SIM and the Dynamic _ Trust algorithm, the recommendation precision MAE of the CF-IESDT algorithm is respectively improved by 1% and 1% again. Therefore, the score prediction method provided by the CF-IESDT algorithm and integrating similarity and trust is reasonable and effective, and has the effect of improving the algorithm recommendation precision. The CF-IESDT algorithm obtains more accurate similarity and Trust degree by fusing and integrating the similar SIM and the Dynamic Trust Dynamic _ Trust algorithm, so that users which are both similar and credible are more accurately screened out to be used as Top-K neighbor recommendation users, and the optimal recommendation performance is finally obtained.
The experimental results are as follows: the comprehensive similarity algorithm SIM and the Dynamic Trust algorithm Dynamic _ Trust can obtain more accurate similarity and Trust; compared with the un-fused comprehensive similarity algorithm SIM and the Trust recommendation algorithm Dynamic _ Trust, the CF-IESDT algorithm fused with the similarity and Trust relationship has the function of improving the recommendation performance again, so that the similarity and Trust fused score prediction method provided by the invention is reasonable and effective; compared with the CF and CF-DNC algorithms, the CF-IESDT algorithm can effectively reduce the false recommendation of unreliable users, and has relatively better recommendation precision and recommendation reliability.
Experiment five: comparative analysis of CF-IESDT algorithm recall
From the fourth experiment, the recommendation precision MAE of the collaborative filtering recommendation algorithm TrustSim based on the Trust similarity is lower than those of the RatingSim, SIM and Dynamic _ Trust algorithms provided by the invention, the reason for this is that the sparsity of the Trust data in the experimental data set FilmTrust is as high as 99.93%, the similarity between the users is calculated only by the common Trust user of two users in the Trust network, and the implicit similarity of the Trust relationship is obviously insufficient and inaccurate as the selection basis of the Top-K neighbor recommendation user. As shown in table 4, the predictable fraction of TrustSim algorithm is only 883, and the recall rate RL is about 25%, which is much lower than other recommended algorithms, so the recommended performance of TrustSim algorithm is reasonably interpretable as being lower than other algorithms.
To further verify the recall rate RL of the CF-IESDT, the recall rates of the CF and CF-DNC algorithms, respectively, were compared. And randomly selecting 90% from the FilmTrust scoring data as a training set, taking the rest 10% as a test set, wherein the number N of scores to be tested in the test set is 3545, and RL comparison results are shown in a table 4 and a graph 10.
As can be seen from FIG. 10, the recall rate of the CF-IESDT algorithm is improved by about 13% and 3% compared with the CF and CF-DNC algorithms, respectively, thereby illustrating that in the case of extremely sparse data set, the CF-IESDT algorithm also obtains higher recall rate while improving the recommendation precision. Compared with CF and CF-DNC algorithms, the recall rate of the RatingSim, SIM and Dynamic _ Trust algorithms provided by the invention is also obviously improved. The CF-IESDT algorithm relieves the problem that the similarity of cold-start users is difficult to calculate by fusing interest similarity and trust similarity on one hand, and relieves the problems of sparsity of trust data and cold start to a certain extent by fusing recommendation trust and explicit trust and researching the trust transfer problem of single-path trust weak propagation and multi-path trust aggregation based on a reconstructed trust network on the other hand.
The experimental results are as follows: in the case of extremely sparse data set, the recommendation accuracy MAE of the CF-IESDT algorithm does not decrease with the increase of the recall rate RL, and on the contrary, the CF-IESDT algorithm has relatively optimal recommendation performance as shown in fig. 9; the CF-IESDT algorithm keeps high recommendation precision while obtaining high recall rate, and effectively relieves the problems of data sparsity and cold start of the recommendation algorithm.
Experiment six: verification analysis of validity of trust reward and punishment mechanism
In order to verify the effectiveness of the trust punishment mechanism provided by the CF-IESDT algorithm, experiments respectively observe the recommended performances of the following algorithms provided by the invention: (1) a static Trust collaborative filtering recommendation algorithm (Trust) based on the untrusty reward and punishment; (2) a Dynamic Trust collaborative filtering recommendation algorithm (Dynamic _ Trust) based on Trust and punishment; (3) a collaborative filtering recommendation algorithm (CF-IEST) which integrates the comprehensive similarity and the static trust; the recommended accuracy MAE comparison results of the four algorithms are shown in fig. 11.
As can be seen from fig. 11, compared with the recommendation performance of the static Trust recommendation algorithm Trust without the Trust punishment mechanism, the recommendation performance of the Dynamic Trust recommendation algorithm Dynamic _ Trust after the Trust punishment mechanism is introduced is improved, because the Trust punishment mechanism evaluates the effectiveness of the recommendation user score according to the difference between the user feedback score and the recommendation user score, and performs controlled reward and penalty on the Trust degree of the recommendation user score, thereby inhibiting the negative influence of random and false scoring behavior of the user on the recommendation performance of the system. With the continuous adjustment and update of the system operation and the Trust value, users with higher success rate of target user recommendation obtain higher Trust degrees, and the users become Top-K neighbor Trust recommendation users of the target users stably, so that the Dynamic _ Trust algorithm screens out more accurate Top-K neighbor Trust recommendation users, thereby improving the prediction precision of unscored items and ensuring that the recommendation result is more reliable; similarly, compared with the recommendation performance of the CF-IEST recommendation algorithm without the trust punishment mechanism, the recommendation performance of the CF-IESDT algorithm introduced with the trust punishment mechanism is also improved, because the trust and the similarity between users are reasonably adjusted according to each recommendation result of the user along with the operation of the system, the false recommendation of the unreliable user is effectively reduced, and the recommendation result always comes from the recommendation user which is trusted with the target user and has higher similarity.
The experimental results are as follows: the trust reward and punishment updating mechanism provided by the invention can effectively inhibit the negative influence of random and false scoring behaviors of the user on the recommendation performance of the system, and improves the recommendation precision of the recommendation system; the algorithm carries out trust reward and punishment on the active and passive recommenders respectively according to the recommendation behaviors of the users, can dynamically update the Top-K neighbor trust user set for the target user, and reduces the false recommendation of the unreliable user, thereby improving the prediction accuracy of the unscored items and finally providing reliable personalized accurate recommendation for the target user.
To summarize:
the invention provides a collaborative filtering recommendation method CF-IESDT integrating information entropy similarity and dynamic trust, and the algorithm can effectively reduce false recommendation of unreliable users, inhibit negative effects of random and false scoring behaviors of the users on recommendation performance, and relieve the problems of data reliability, data sparsity and cold start of a recommendation system. Compared with the CF and CF-DNC recommendation algorithms, the CF-IESDT algorithm has the advantages that the recommendation performance is obviously improved, and meanwhile, a higher recall rate is obtained. Experimental results show that the trust punishment strategy provided by the invention, the constructed similarity calculation model, the trust calculation model and the score prediction model integrating similarity and trust are reasonable and effective, and the method has the effects of improving the algorithm recommendation precision and the recommendation reliability.
TABLE 1 Effect of θ on MAE Experimental results
TABLE 2 Effect of θ on RMSE test results
TABLE 3 MAE experimental results of different recommendation algorithms
TABLE 4 Recall (RL) test results for different recommendation algorithms
Claims (8)
1. A collaborative filtering recommendation method integrating information entropy similarity and dynamic trust is characterized in that firstly, based on a user item score and an explicit trust relationship of a recommendation system, user similarity and trust are calculated, and a Top-K similarity and trust recommendation user set of a target user is selected from big to small; then, a scoring prediction model is constructed by fusing the similarity and the trust, and the scoring of the target user on the project is predicted; secondly, selecting Top-N projects according with the target user preference according to the prediction scores, and completing personalized project recommendation of the target user; finally, evaluating the scoring effectiveness of the recommended users, and giving trust rewards and punishments to the active and passive recommended users;
in the recommendation system, the user item scoring data is represented as an m × n-order matrix R and comprises m user sets U ═ U1,u2,…,umAnd n sets of items I ═ I1,i2,…,inR, each element R in the matrix Rij(1 ≦ i ≦ m,1 ≦ j ≦ n) represents user uiFor item ijIf user u scoresiFor item ijNo score is expressed asThe explicit trust relationship between users is represented as m x m order trust matrix T, Tij(i is not less than 1 and not more than m, j is not less than 1 and not more than m) represents user uiWith user ujTrust relationship of, TijE is {0,1}, 0 represents distrust, and 1 represents trust;
the method specifically comprises the following steps:
1) constructing an information entropy similarity calculation method based on the grading difference among users, and calculating the grading similarity RatingSim of the users;
2) constructing a trust relationship implicit similarity calculation method based on the explicit trust relationship, and calculating the trust similarity TrustSim of the user;
3) fusing RatingSim and TrustSim to construct a comprehensive similarity calculation model and calculate the comprehensive similarity SIM of the user;
4) analyzing the Trust relationship among users, constructing a Trust level Trust calculation model, and calculating the direct Trust level DTrust, the indirect Trust level ITrust and the global Trust level Auth of the users;
a) if explicit trust exists among users or scoring interaction of a co-scoring item set exists among the users, calculating scoring credibility Cre, scoring prediction capability Cap and a co-scoring item set Jaccard coefficient of the users; integrating three influence factors of the user scoring credibility Cre, the scoring prediction capability Cap and the co-scoring project set Jaccard coefficient, and calculating the recommendation credibility ReTrust of the target user to the recommendation user; integrating explicit trust and recommended trust to calculate the direct trust DTrust of the target user to the recommended user;
b) if no direct trust relationship exists among users but other trust transmission paths meeting the conditions exist, calculating single-path indirect trust or multi-path indirect trust ITrust according to the number of the paths;
c) if no direct trust and indirect trust relationship exists between the users, calculating the global trust Auth of the users, and using the global trust of the users as the trust of the target users to the recommended users;
5) obtaining user comprehensive similarity SIM and Trust Trust according to the step 3) and the step 4), selecting a Top-K similarity and Trust recommendation user set of a target user according to descending order from big to small, constructing a scoring prediction model by fusing the similarity and the Trust, and predicting the scoring of the target user on the project;
6) obtaining the prediction result of the target user on the unscored items according to the step 5), arranging the prediction result in a descending order from big to small, constructing a Top-N item recommendation list according with the preference of the target user, and recommending specific items to the target user in a personalized way;
7) and evaluating the effectiveness of the recommended user score according to the difference between the user feedback score and the recommended user score, and giving trust rewards and punishments for the active and passive recommended users.
2. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in step 1), the scoring similarity RatingSim calculation method is as follows:
calculating the scoring difference information entropy RDE (u, v) of the users u and v by adopting a formula (1), and calculating the scoring similarity RatingSim (u, v) of the users u and v by adopting a formula (2):
in the formulae (1) and (2), NfailRepresenting the number of items with the evaluation attitudes of the users u and v to the common evaluation item set being opposite; n represents the number of the co-evaluation items; the absolute value of the score difference of the user u and v common score items is expressed as a set c ═ c1,c2,c3,…ck},p(ci) Indicating the probability of occurrence of each of the absolute values of the score difference.
3. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in the step 2), the trust similarity TrustSim calculation method is as follows:
in formula (3), t (u) and t (v) represent the set of neighbor nodes trusted by nodes u and v, respectively; t (u, v) ═ T (u) # T (v) denotes that two users trust in commonThe neighbor node set of (2); din(k) Indicating the degree of the k-entry of the node.
4. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in step 3), the user comprehensive similarity SIM calculation method comprises the following steps:
calculating the comprehensive similarity SIM between the users u and v by adopting a formula (4), and calculating the self-adaptive weight factor alpha between the scoring similarity RatingSim and the trust similarity TrustSim by adopting a formula (5):
SIM(u,v)=αRatingSim(u,v)+(1-α)TrustSim(u,v) (4)
5. the collaborative filtering recommendation method integrating information entropy similarity and dynamic Trust according to claim 1, wherein in the step 4), the calculation method of the Trust Trust is as follows:
4-1) if explicit trust exists among users or scoring interaction of a co-scoring item set exists among users, fusing the explicit trust and the recommendation trust, and calculating the direct trust of the users by adopting a formula (6):
DTrust(u,v)=ρReTrust(u,v)+(1-ρ)Tuv (6)
in formula (6), DTrust (u, v) represents the direct trust level of the target user u for the recommended user v; ReTrust (u, v) represents the recommendation confidence level; t isuvRepresenting explicit trust of users u and v, TuvE is {0,1}, 0 represents distrust, and 1 represents trust; rho is a weight adjustment factor value of 0.5;
calculating the recommended trust ReTrust (u, v) by adopting the formula (7):
ReTrust(u,v)=Jacu,v[βCrev+(1-β)Capu,v] (7)
in formula (7), CrevRepresenting the scoring credibility of the recommendation user v; capu,vThe score prediction capability of the recommended user v is represented; jacu,vA Jaccard coefficient representing a co-rating item of users u and v; beta is a weight adjusting factor, and the value of beta is 0.2;
calculating the scoring credibility Cre of the user v by adopting a formula (8)v:
Crev=Actv×Errv (8)
Calculating the user activity Act by adopting a formula (9)v:
In formula (9), ImA set of scoring items representing the mth user; ivL represents the number of user v scoring items;
calculating the user score deviation Err by adopting the formula (10)v:
In the formula (10), rv,iRepresents the rating of item i by user v;a score mean representing item i;
calculating the score prediction capability Cap of the recommended user by adopting a formula (11)u,v:
Capu,v=1-MSEu,v
In the formula (11), the reaction mixture,representing the predictive scoring of the co-scoring item set by taking the user v as the only recommended user of the user u; r isu,iRepresents the scoring of item i by user u; iuvL represents the number of the co-rated items of the two users;
In the formula (12), rv,iRepresents the rating of the recommended user v for the item i;andrespectively representing the average values of the scores of the users u and v; SIM (u, v) represents the integrated similarity between users u and v as stated in equation (4);
and (3) calculating the co-rating item Jaccard coefficient of the users u and v by adopting the formula (13):
4-2) if the users have no direct trust relationship but other trust transfer paths meeting the conditions, respectively calculating the single-path indirect trust or the multi-path indirect trust ITrush according to the number of the paths:
single path indirect trust: if there is one and only one shortest path P (u, u) reachable between users u and v1,u2… v), path length (path ≦ dmaxThe single-path indirect trust calculation is shown in equation (14):
in the formula (14), DTrust (u)1,u2) Indicating u between adjacent users on the path1For u is paired2Direct trust of; dpIndicating the length of the path;expressing the attenuation factor based on the belief propagation distance, the calculation formula is shown in (15):
in the formula (15), dmaxMaximum depth of 3, d representing a trusted path searchuvRepresents the trust path length between users u and v;
multipath indirect trust: if there are multiple shortest Paths reachable between user u and user v, the path set is expressed as Paths (u, v) ═ P1(u,v),P2(u,v),…Pn(u, v) }, where the path length d of any reachable path P (u, v) ∈ Paths (u, v)PSatisfy dp≤dmaxThen the multipath indirect trust calculation is shown in equation (16):
ITrust(u,v)=maxPaths(u,v)(ITrust(u,v)) (16)
formula (16) shows that the optimal or near optimal trust transfer path is selected from the multiple single paths in a maximum value mode;
4-3) if the direct trust and the indirect trust relationship do not exist between the users, calculating the global trust Auth of the users by adopting a formula (17):
in formula (17), Trust (k, u) represents the direct/indirect Trust of user k to user u in the Trust network; c denotes the number of users having a direct/indirect trust relationship with user u.
6. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in step 5), the item prediction score is calculated as follows:
in the formula (18), SN and TB respectively represent a Top-K similar user set and a trusted user set; SIM (u, v) represents the integrated similarity between users calculated by formula (4); trust (u, t) represents the Trust degree of the target user u on the Trust recommendation user t; theta represents the value of the weight adjustment parameter of 0.5.
7. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in step 6), the Top-N item meeting the target user preference is recommended to the target user in a personalized manner.
8. The collaborative filtering recommendation method integrating information entropy similarity and dynamic trust according to claim 1, wherein in step 7), the trust punishment method is as follows:
7-1) using the calculation method of formula (19) to give trust rewards to actively recommending users:
in the formula (19), the first and second groups,representing the trust degree of the target user u to the recommended user v after the trust is updated;representing the trust degree of the target user u to the recommended user v before the trust updating; mSA control factor indicating a maximum value of the trust change amount, andESrepresenting the influence of the recommendation score error on the reward amplitudeWherein Q represents the maximum value of the scoring error in the recommendation system, R represents the maximum value of the item rating, and when the scoring error d is equal to | Ru,i-rv,iWhen | ≦ epsilon, the larger the error d, the smaller the reward value; pNExpressing the influence factor of the user recommendation times on the reward forceWherein ω CrevRepresenting the user scoring credibility; pSRepresenting the recommendation success rate of the recommendation node, and when the recommendation user v actively recommends the target user u, namely | ru,i-rv,iWhen | < epsilon, the recommendation is successful, PSThe calculation formula is shown as (20):
in the formula (20), N ═ TotalSet (u, v) | represents the total number of times of recommendation of the target user u by the recommending user v; n is a radical ofsRepresenting the successful recommendation times of the target user u by the recommending user v;
7-2) performing trust penalty on the negative recommended users by adopting a calculation method of a formula (21):
in the formula (21), MfA control factor indicating a maximum value of the trust change amount, andEfrepresents a recommendation score error versus penaltyInfluence factor of penalty amplitude, orderWhen inter-user scoring error d ═ ru,i-rv,i|>When epsilon is larger, the larger the error d is, the larger the penalty value is; pNFactor representing influence of user recommendation times on penalty degree, orderPfRepresenting the recommendation failure rate of the recommendation node, when the recommendation user v makes a negative recommendation to the target user u, namely | ru,i-rv,i|>When ε, the recommendation fails, PfThe calculation formula is shown as (22):
in the formula (22), Nf| False (u, v) | represents the number of failed recommendations of the target user u by the recommending user v.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811138086.9A CN109408734B (en) | 2018-09-28 | 2018-09-28 | Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811138086.9A CN109408734B (en) | 2018-09-28 | 2018-09-28 | Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109408734A CN109408734A (en) | 2019-03-01 |
CN109408734B true CN109408734B (en) | 2021-07-27 |
Family
ID=65465549
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811138086.9A Active CN109408734B (en) | 2018-09-28 | 2018-09-28 | Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109408734B (en) |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110377841B (en) * | 2019-06-04 | 2022-01-07 | 深思考人工智能机器人科技(北京)有限公司 | Similarity calculation method and system applied to collaborative filtering method |
CN110390058A (en) * | 2019-06-28 | 2019-10-29 | 哈尔滨理工大学 | Consider the credible mixed recommendation method of Web service of timeliness |
CN110334286A (en) * | 2019-07-10 | 2019-10-15 | 南京工业大学 | Trust relationship based personalized recommendation method |
CN110457592A (en) * | 2019-07-11 | 2019-11-15 | 天津大学 | A kind of social networks recommended method based on figure entropy |
CN110442750A (en) * | 2019-08-12 | 2019-11-12 | 深圳清华大学研究院 | Sea area network video recommended method based on time and space sequence information |
CN110704496B (en) * | 2019-09-02 | 2023-01-24 | 山东科技大学 | Proposed aggregation method based on asymmetric response |
CN110889045B (en) * | 2019-10-12 | 2024-04-23 | 平安科技(深圳)有限公司 | Label analysis method, device and computer readable storage medium |
CN110990713B (en) * | 2019-10-24 | 2023-06-02 | 广东技术师范大学 | Collaborative filtering recommendation method based on optimal trust path |
CN110750689A (en) * | 2019-10-30 | 2020-02-04 | 北京大学 | Multi-graph fusion method |
CN111314348B (en) * | 2020-02-19 | 2022-07-12 | 中国联合网络通信集团有限公司 | Method and device for establishing trust degree model, trust evaluation and equipment authentication |
CN111506830B (en) * | 2020-04-09 | 2022-08-09 | 中南大学 | Collaborative filtering recommendation algorithm fusing social trust influence |
CN111641637B (en) * | 2020-05-28 | 2021-05-11 | 重庆邮电大学 | Edge node calculation result credibility judgment method based on trust evaluation |
CN112765465B (en) * | 2021-01-15 | 2023-04-14 | 电子科技大学 | User map-based recommendation method |
CN112765488A (en) * | 2021-01-25 | 2021-05-07 | 广东工业大学 | Recommendation method, system and equipment fusing social network and knowledge graph |
CN113051487B (en) * | 2021-04-20 | 2022-08-02 | 安徽师范大学 | Travel group recommendation method based on user trust and social influence |
CN113269609A (en) * | 2021-05-25 | 2021-08-17 | 中国联合网络通信集团有限公司 | User similarity calculation method, calculation system, device and storage medium |
CN113344671B (en) * | 2021-06-23 | 2023-04-07 | 昆明理工大学 | Trust factor fused personalized recommendation model and construction method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106484876A (en) * | 2016-10-13 | 2017-03-08 | 中山大学 | A kind of based on typical degree and the collaborative filtering recommending method of trust network |
CN106682114A (en) * | 2016-12-07 | 2017-05-17 | 广东工业大学 | Personalized recommending method fused with user trust relationships and comment information |
CN107943897A (en) * | 2017-11-17 | 2018-04-20 | 东北师范大学 | A kind of user recommends method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170323019A1 (en) * | 2016-05-04 | 2017-11-09 | Agingo Corporation | Ranking information providers |
-
2018
- 2018-09-28 CN CN201811138086.9A patent/CN109408734B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106484876A (en) * | 2016-10-13 | 2017-03-08 | 中山大学 | A kind of based on typical degree and the collaborative filtering recommending method of trust network |
CN106682114A (en) * | 2016-12-07 | 2017-05-17 | 广东工业大学 | Personalized recommending method fused with user trust relationships and comment information |
CN107943897A (en) * | 2017-11-17 | 2018-04-20 | 东北师范大学 | A kind of user recommends method |
Non-Patent Citations (7)
Title |
---|
A matrix factorization technique with trust propagation for recommendation in social networks;Mohsen Jamali 等;《RecSys "10: Proceedings of the fourth ACM conference on Recommender systems》;20100926;135-142 * |
A new confidence-based recommendation approach: Combining trust and certainty;Faezeh SadatGohari 等;《Information Sciences》;20180131;第422卷;21-50 * |
基于上下文因素的P2P动态信任模型;刘义春 等;《通信学报》;20160831;第37卷(第8期);34-45 * |
基于加权信息熵相似度的协同过滤算法研究;刘文龙;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140415(第04期);I138-1210 * |
复杂网络环境下基于信任传递的推荐模型研究;李慧 等;《自动化学报》;20180228;第44卷(第2期);363-376 * |
融合信息熵与信任机制的防攻击推荐算法研究;郝志峰 等;《计算机应用与软件》;20150331;第32卷(第3期);284-288 * |
融合信息熵与兴趣度的协同过滤算法;石坤 等;《计算机与数字工程》;20170220;第45卷(第2期);338-342 * |
Also Published As
Publication number | Publication date |
---|---|
CN109408734A (en) | 2019-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109408734B (en) | Collaborative filtering recommendation method integrating information entropy similarity and dynamic trust | |
Steck | Calibrated recommendations | |
Ahmadian et al. | A social recommender system based on reliable implicit relationships | |
Javari et al. | A probabilistic model to resolve diversity–accuracy challenge of recommendation systems | |
CN109190030B (en) | Implicit feedback recommendation method fusing node2vec and deep neural network | |
Eliyas et al. | Recommendation systems: Content-based filtering vs collaborative filtering | |
Yu et al. | Joint user knowledge and matrix factorization for recommender systems | |
CN109635206B (en) | Personalized recommendation method and system integrating implicit feedback and user social status | |
US11636394B2 (en) | Differentiable user-item co-clustering | |
Jiao et al. | A novel learning rate function and its application on the SVD++ recommendation algorithm | |
Sinha et al. | Evolution of recommender paradigm optimization over time | |
Feng et al. | Improving group recommendations via detecting comprehensive correlative information | |
Wang et al. | Group recommendation based on a bidirectional tensor factorization model | |
Zarei et al. | A memory-based collaborative filtering recommender system using social ties | |
Bahrani et al. | User and item profile expansion for dealing with cold start problem | |
Liu et al. | Towards context-aware collaborative filtering by learning context-aware latent representations | |
Symeonidis et al. | Counteracting the filter bubble in recommender systems: Novelty-aware matrix factorization | |
Li et al. | From reputation perspective: a hybrid matrix factorization for qos prediction in location‐aware mobile service recommendation system | |
Hong et al. | Context-aware recommendation using role-based trust network | |
Lee | Fuzzy clustering with optimization for collaborative filtering-based recommender systems | |
Wang et al. | A hybrid collaborative filtering recommendation algorithm: integrating content information and matrix factorisation | |
Zhu et al. | Gtrm: A top-n recommendation model for smartphone applications | |
Shambour et al. | Restaurant Recommendations Based on Multi-Criteria Recommendation Algorithm | |
Chen et al. | Trust-based collaborative filtering algorithm in social network | |
Roy | A comparative analysis of different trust metrics in user-user trust-based recommender system |
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 |