CN114154902A - Recommendation method of hidden social relationship feedback technology fusing user social status - Google Patents

Recommendation method of hidden social relationship feedback technology fusing user social status Download PDF

Info

Publication number
CN114154902A
CN114154902A CN202111500966.8A CN202111500966A CN114154902A CN 114154902 A CN114154902 A CN 114154902A CN 202111500966 A CN202111500966 A CN 202111500966A CN 114154902 A CN114154902 A CN 114154902A
Authority
CN
China
Prior art keywords
user
social
label
relationship
item
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111500966.8A
Other languages
Chinese (zh)
Other versions
CN114154902B (en
Inventor
陈锐
张建伟
崔建涛
马军霞
蔡增玉
谷培培
马源
宁浩田
梁维德
彭嘉怡
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhengzhou University of Light Industry
Original Assignee
Zhengzhou University of Light Industry
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhengzhou University of Light Industry filed Critical Zhengzhou University of Light Industry
Priority to CN202111500966.8A priority Critical patent/CN114154902B/en
Publication of CN114154902A publication Critical patent/CN114154902A/en
Application granted granted Critical
Publication of CN114154902B publication Critical patent/CN114154902B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/01—Social networking

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Educational Administration (AREA)
  • Development Economics (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Computing Systems (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Algebra (AREA)
  • Software Systems (AREA)
  • Computational Mathematics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Physics (AREA)
  • Game Theory and Decision Science (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a recommendation method of a hidden social relationship feedback technology fusing user social statuses, which comprises the steps of constructing a TrustMF model according to the trust relationship of a user, obtaining user social status weights according to a PageRank algorithm, constructing a USSocialMF model according to user scores and the trust relationship, obtaining user weights and project weights by using social labels, constructing a TSoccialsMF model according to the user weights and the project weights, training the three models, obtaining implicit similarity of the user and the project by using a social matrix feedback technology, and constructing the EISocialMF model. The recommendation method integrates various social relationship influence factors, can effectively relieve the problems of data sparseness and cold start, has better recommendation quality, improves recommendation accuracy, and introduces the user social status relationship prediction acquired by the PageRank algorithm into a recommendation model based on social relationship feedback so as to distinguish the influence of different social status users.

Description

Recommendation method of hidden social relationship feedback technology fusing user social status
Technical Field
The invention relates to a recommendation method of a hidden social relationship feedback technology fusing the social status of a user.
Background
Complex social relations exist among users in the social network, and various social relations affect each other, so that it is difficult to accurately model real social relations among users by directly measuring the complex social relations, and recommendation performance is affected. In recent years, some scholars at home and abroad fuse influence factors such as social relations, social labels and personal interests of users into a recommendation model so as to improve recommendation quality. The user scoring information and the social relationship are mapped to the shared user feature space and the shared item feature space to improve the accuracy of prediction, so that although the problem of inaccurate recommendation caused by sparse scoring data is relieved to a certain extent, when the user features are measured through the neighbor relationship, the user social relationship is lack of further training to obtain the accurate similarity relationship, so that a preference model obtained through the neighbor user may have deviation from a real user preference model, and the amplitude of improving the accuracy of recommendation is limited. In recent years, some prediction research methods about user trust relationships are proposed successively, but few researches apply user social relationship prediction results to recommendation models, such as: by utilizing a matrix decomposition technology, integrating approximate scores of the users and similar friends on the unknown items through comprehensive weighting to complete recommendation; the user characteristics are considered to be composed of trusts and trusts, the trusting relationship is analyzed by utilizing a matrix decomposition technology, and the characteristics of the trusting person and the trusting person of the user are obtained, so that the unknown trusting relationship is predicted, and the scoring prediction is realized and the recommendation is carried out; user scores and social labels are combined, an improved matrix decomposition method based on neighbor similarity is provided, recommendation accuracy is improved, but the method does not consider the influence of explicit trust relationships and the inaccuracy problem of directly measuring social relationships; the method improves the prediction accuracy to a certain extent, but does not consider the negative influence of the extremely sparse user trust relationship on the user social relationship, the positive influence of the implicit evaluation (such as social labels) of the user-item and the approximate relationship estimation of the item social relationship on the recommendation quality, and ignores the influence of different actions and homogeneity of each user in the social network.
Therefore, most of the existing recommendation algorithms only concern trust relationships and friend relationships among user individuals, and consider the user individuals in the social network to be viewed equally and have the same authority. In fact, the authority degrees of the users in different fields are different, and the mutual influence degrees among the users are also different. The authoritativeness of users in a social network is referred to as social status. Although the influence of the social status of the user on the recommendation quality is considered in the establishment process of the recommendation model, the authority degree of the user is determined only by the number of linked-in and linked-out users in the social network, and the self social status of the linked user, namely the influence degree of the individual user is not considered. In addition, complex social relations exist among users in the social network, and various social relations affect each other, so that the real social relations among the users are difficult to be accurately modeled by directly measuring the complex social relations, the recommendation performance is affected, and the recommendation accuracy is reduced.
Disclosure of Invention
The invention provides a recommendation method of a hidden social relationship feedback technology fusing the social status of a user, which is used for solving the technical problem of low recommendation accuracy based on the conventional recommendation method.
A recommendation method of a hidden social relationship feedback technology fusing user social status comprises the following steps:
constructing a recommendation model TrustMF according to the trust relationship of the user;
calculating the social place value of the user according to a PageRank algorithm, and acquiring the social place weight of the user based on the social place value of the user;
constructing a recommendation model USSocialMF based on the social status of the user according to the user score and the trust relationship;
acquiring a user weight and a project weight by using the social label;
establishing a social recommendation model TSoccialaMF based on social label weight according to the user weight and the project weight;
training a TrustMF model, a USSocialMF model and a TSocialMF model;
obtaining implicit similarity of users and items by utilizing a social matrix feedback technology;
and constructing a social recommendation model EISocialMF based on user scores, trust relationships and social labels.
Further, the building of the recommendation model TrustMF according to the trust relationship of the user comprises the following steps:
setting two users uu and uvBy the trustperson feature vector B of the useruAnd trusted person feature vector EvNormal distribution expression;
obtaining posterior probability distribution of the characteristic vector of the trustperson and the characteristic vector of the trustperson according to a Bayesian formula:
Figure BDA0003402538950000031
wherein ,TuvRepresenting user uu and uvThe trust relationship of (2);
Figure BDA0003402538950000032
and
Figure BDA0003402538950000033
the variances of Gaussian distribution with the mean value of 0 obeyed by the characteristic vectors of the trustperson and the trustperson respectively;
Figure BDA0003402538950000034
is a trust relationship TuvThe variance of the gaussian distribution obeying a mean of 0; b isu and EvRespectively representing trustperson and trustperson feature vectors; n represents the number of users;
taking logarithm of the formula and maximizing the logarithm to obtain a TrustMF model; and decomposing T by minimizing the loss function l (T, B, E) as follows, resulting in B and E:
Figure BDA0003402538950000035
further, the initial access probability of each vertex is the same, and the initial probability of each vertex is set to be 1/N; the method for calculating the social place value of the user according to the PageRank algorithm and acquiring the social place weight of the user based on the social place value of the user comprises the following steps:
and calculating the social place value of the user according to the PageRank algorithm, wherein the calculation formula is as follows:
Figure BDA0003402538950000036
wherein ,PRuRepresenting user uuPageRank value of (C)uRepresenting user uuA set of trusted friends that are to be used,
Figure BDA0003402538950000037
is the probability value of jumping out of the current network, and the value range is [0, 1]]；
Adjusted user uu and uvWeight of relationship between WuvThe calculation formula is as follows:
Figure BDA0003402538950000041
wherein ,tuvRepresenting user uuFor user uvThe trust value of (c).
Further, the method for constructing the user social status-based recommendation model USSocialMF according to the user scores and the trust relationship comprises the following steps:
the USSocialMF model is as follows:
Figure BDA0003402538950000042
wherein R is user score, U and V respectively represent a user hidden factor feature matrix and a project hidden factor feature matrix, and a column vector Uu and ViRespectively corresponding user and project implicit feature vectors, ruiFor user uuFor item iiScore of, NuRepresenting user uuA set of neighboring users of (a); t (u) represents the user uuA set of trusted users; sim (u, t) denotes user uu and utSim (i, j) represents item ii and ijThe similarity between them is obtained by the following formula:
Figure BDA0003402538950000043
Figure BDA0003402538950000044
wherein ,rui and rtiRespectively represent users uu and utFor item iiScore of (I)u and ItRespectively represent users uu and utSet of scored items, IutRepresenting user uu and utThe set of items that are scored in common,
Figure BDA0003402538950000045
and
Figure BDA0003402538950000046
respectively represent users uu and utAverage score of (a); r isujRepresenting user uuFor item ijIs given to all items i togetheri and ijA set of scored users;
by minimizing
Figure BDA0003402538950000051
And decomposing R and W to obtain feature matrixes U and V.
Further, the obtaining of the user weight and the item weight by using the social tag includes:
using the TF-IDF algorithm to obtain:
Figure BDA0003402538950000052
Figure BDA0003402538950000053
wherein ,Guk and HikRespectively represent users uuAnd item iiFor label lkThe weight relationship of (c); c. CupRepresenting user uuSelection of tag lpThe number of times of the operation of the motor,
Figure BDA0003402538950000054
indicating usage label lpNumber of users, NtIndicates the total number of tags, user uuUse of the label lpThe more times, the weight value G corresponding to the labelukThe larger; c. CipPresentation label lpAppear in item iiThe number of times of the middle-time period,
Figure BDA0003402538950000055
indicating a tagged label lpNumber of items, taglpThe more times of occurrence in a certain item set, the corresponding weight value HikThe larger.
Further, the establishing of the social recommendation model TSocialMF based on the social label weight according to the user weight and the project weight includes:
according to the weight relationship between the user-label and the item-label, the user and item characteristics are obtained, and then the conditional probabilities of the user-label and item-label weight matrixes are as follows:
Figure BDA0003402538950000056
Figure BDA0003402538950000057
wherein G is mapped to a user feature U and a label feature space L, and H is mapped to an item feature V and a label feature space L;
Figure BDA0003402538950000058
and
Figure BDA0003402538950000059
the variance of the gaussian distribution with the mean value of 0 obeyed by the user-label and item-label weight relationships respectively;
according to the Bayesian formula, the TSocialmF model is as follows, and, the minimization
Figure BDA0003402538950000061
To decompose G and H, yielding U, V and L:
Figure BDA0003402538950000062
wherein ,
Figure BDA0003402538950000063
Figure BDA0003402538950000064
is the variance of the gaussian distribution with mean 0 obeyed by the label feature L.
Further, the training of the TrustMF model, the USSocialMF model and the TSocialMF model includes:
the training process of the TrustMF model comprises the following steps:
setting values of parameters lambda, T and eta, wherein eta is a descending rate; initializing matrixes B and E by random numbers uniformly distributed according to [0, 1 ]; iteratively updating matrices B and E according to the following equations:
Figure BDA0003402538950000065
Figure BDA0003402538950000066
the training process of the USSocialMF model comprises the following steps:
setting values of parameters λ, R and η, where η is a rate of decrease; initializing matrices R and V with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U and V according to the following formula:
Figure BDA0003402538950000067
Figure BDA0003402538950000068
the training process of the TSociialMF model comprises the following steps:
setting values of parameters λ, G, H and η, where η is a rate of descent; initializing matrices G and H with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U, V and L according to the following formula:
Figure BDA0003402538950000071
Figure BDA0003402538950000072
Figure BDA0003402538950000073
further, the obtaining of the implicit similarity between the user and the item by using the social matrix feedback technology includes:
implicit similarities between users are noted
Figure BDA0003402538950000074
Is formed by normal distribution of user characteristic similarity, implicit similarity of trust relationship and label similarity, then S(U)The conditional probability distribution of (a) is:
Figure BDA0003402538950000075
wherein ,
Figure BDA0003402538950000076
for user uu and uvThe degree of similarity of the scores of (a) and (b),
Figure BDA0003402538950000077
for preference similarity based on trust relationships,
Figure BDA0003402538950000078
for user uu and uvTag-based implicit preference similarity:
Figure BDA0003402538950000079
Figure BDA00034025389500000710
Figure BDA00034025389500000711
variance (variance)
Figure BDA00034025389500000712
Representing the noise situation of the estimated value, the function f (x) being an improved user uu and uvIs expressed as:
Figure BDA00034025389500000713
relation (u, v) represents user uu and uvA direct trust relationship exists between the two;
setting item similarity
Figure BDA0003402538950000081
The method is characterized by comprising the following steps of (1) normally distributing the item feature similarity and the social label relationship similarity:
Figure BDA0003402538950000082
according to Bayes reasoning, the following loss function is obtained:
Figure BDA0003402538950000083
wherein item ii and ijSimilarity of (2)
Figure BDA0003402538950000084
The project characteristics and social label influence factors are comprehensively considered, and the calculation formula is as follows:
Figure BDA0003402538950000085
Figure BDA0003402538950000086
representing social tag-based items ii and ijThe calculation formula is as follows:
Figure BDA0003402538950000087
Figure BDA0003402538950000088
representing i based on item featuresi and ijThe calculation formula is as follows:
Figure BDA0003402538950000089
further, the establishing of the social recommendation model EISocialMF based on the user score, the trust relationship and the social label comprises the following steps:
the EISocialMF model is as follows:
Figure BDA0003402538950000091
wherein ,
Figure BDA0003402538950000092
representing the similarity between trusted users of the trusted users;
Figure BDA0003402538950000093
representing the similarity of two users based on a trust relationship, by user uu and uvRespectively trust their common users ukObtaining the trust relationship;
Figure BDA0003402538950000094
indicating the similarity of two trusted people based on a trust relationship,by uu、uvWith co-trusted users ukThe trust relationship of (a) obtains:
Figure BDA0003402538950000095
Figure BDA0003402538950000096
the recommendation method of the hidden social relationship feedback technology fusing the social status of the user, provided by the invention, fuses a plurality of social relationship influence factors, can effectively relieve the problems of data sparseness and cold start, has better recommendation quality, and can simplify the recommendation inaccuracy caused by complex relationship measurement through the hidden social relationship feedback technology, thereby improving the recommendation accuracy; in addition, a social relationship prediction technology is introduced into the establishment process of the recommendation model to establish a social relationship feedback model, the social relationship of the user is introduced into the recommendation model, the social relationship is adjusted according to the PageRank algorithm, and the social relationship prediction of the user obtained by the PageRank algorithm is introduced into the recommendation model based on the social relationship feedback to distinguish the influence of the users with different social relationships.
Drawings
FIG. 1 is a flowchart of a recommendation method of a hidden social relationship feedback technology fusing user social status provided by the invention;
FIG. 2 is a schematic diagram of the ISocialmF recommendation framework.
Detailed Description
The embodiment provides a recommendation method of a hidden social relationship feedback technology fusing a user social status, and in the embodiment, the following settings can be made: the user preference can be influenced by the user, neighbor users and trust users, the social tags reflect the preference degree of the user to items from the side, and the scoring information, the trust relationship and the social tag weight are respectively modeled by analyzing the influence of the social tags on the user preference from the two aspects of explicit and implicit social relationships. Meanwhile, on the basis of the ISOICAlMF framework structure, user characteristics and project characteristics are mapped to a shared characteristic space from the angles of user-project scoring, user trust relationship, social labels and the like by utilizing a matrix decomposition hidden factor technology, approximate relationships between users and between projects are estimated according to the obtained user characteristics and project characteristics, and the user characteristics and the project characteristics are continuously trained so as to more accurately optimize user similarity and project similarity.
As shown in fig. 1, the recommendation method of the hidden social relationship feedback technology fusing the social status of the user provided by this embodiment includes the following steps:
step S1: and (3) constructing a recommendation model TrustMF according to the trust relationship of the user:
setting two users uu and uvBy the trustperson feature vector B of the useruAnd trusted person feature vector EvAnd (4) normal distribution representation. In this embodiment, the mean and variance of the feature vectors of the trusts and trusts are set to 0 and 0, respectively
Figure BDA0003402538950000111
And
Figure BDA0003402538950000112
a gaussian distribution of (a). According to a Bayes formula, the posterior probability distribution of the characteristic vector of the trustperson and the characteristic vector of the trustperson can be obtained:
Figure BDA0003402538950000113
wherein ,TuvRepresenting user uu and uvThe trust relationship of (2);
Figure BDA0003402538950000114
and
Figure BDA0003402538950000115
the variances of Gaussian distribution with the mean value of 0 obeyed by the characteristic vectors of the trustperson and the trustperson respectively;
Figure BDA0003402538950000116
is a trust relationship TuvThe variance of the gaussian distribution obeying a mean of 0; b isu and EvRespectively representing trustperson and trustperson feature vectors; n represents the number of users.
Taking logarithm of the formula and maximizing the logarithm to obtain TrustMF model, and minimizing the following loss function
Figure BDA0003402538950000117
By minimizing a loss function
Figure BDA0003402538950000118
T can be decomposed to obtain B and E at the same time.
Figure BDA0003402538950000119
Step S2: calculating the social place value of the user according to a PageRank algorithm, and acquiring the social place weight of the user based on the social place value of the user:
in this embodiment, it is intended for user u1The item of interest is recommended. From user u1At the beginning, with probability
Figure BDA00034025389500001110
From u1Randomly selects a path to reach the next vertex, e.g. i1Then with probability
Figure BDA00034025389500001111
From item i1Returning to continue from item i1Is given a probability
Figure BDA00034025389500001112
And (4) random walk. After many walks, the importance of each user converges and the probability of each user is the user's social locality value. The PageRank algorithm is used to calculate the social status of each user in the social network. Because the initial access probability of each vertex is the sameThus, the initial probability of each vertex is set to 1/N. Through the analysis, the social place value of the user is calculated according to the PageRank algorithm, and the calculation formula is as follows:
Figure BDA0003402538950000121
wherein ,PRuRepresenting user uuPageRank value of (C)uRepresenting user uuA set of trusted friends, N represents the number of users,
Figure BDA0003402538950000122
is the probability value of jumping out of the current network, and the value range is [0, 1]]。
User u with social statusu and uvThe trust relationship of (2). The higher the social status of a user in a certain area, the more influential it will be, and the more likely it will be that its advice will be accepted by others. Adjusted user u, taken into account this factoru and uvWeight of relationship between WuvThe calculation formula is as follows:
Figure BDA0003402538950000123
wherein ,tuvRepresenting user uuFor user uvThe trust value of (c).
In a social network, user uvHigher social status of (1), user uvThe higher the confidence of (c).
Step S3: and (3) constructing a recommendation model USSocialMF based on the user social status according to the user score and the trust relationship:
in the embodiment, the user u can be obtained according to the neighbor relation by introducing the trust network to the user characteristic vectoruThe condition distribution of (1):
Figure BDA0003402538950000124
the USSocialMF model is shown below, based on Bayesian inference. The logarithm of the posterior probability of the objective function can be considered as the objective function, keeping the parameters fixed, maximizing the two potential eigenvectors U and V can be considered as an unconstrained optimization problem, and the initial problem transforms into a problem that minimizes the following equation. For the user-item scoring matrix R, it can be mapped to the user and item hidden feature spaces, respectively, using a matrix decomposition technique. Wherein the scoring information R is provided by the user uuAnd the product of the characteristics of the neighboring users and the item characteristics. By minimizing
Figure BDA0003402538950000125
And decomposing R and W to obtain feature matrixes U and V.
Figure BDA0003402538950000131
Wherein, U and V respectively represent a user hidden factor feature matrix and a project hidden factor feature matrix, and a column vector Uu and ViRespectively corresponding user and project implicit feature vectors, ruiFor user uuFor item iiScoring of (4); n is a radical ofuRepresenting user uuRather than a set of users with a trust relationship; t (u) represents the user uuA set of trusted users; sim (u, t) denotes user uu and utSim (i, j) represents item ii and ijThe similarity between them.
sim (u, t) is obtained by the following formula:
Figure BDA0003402538950000132
wherein ,rui and rtiRespectively represent users uu and utFor item iiScore of (I)u and ItRespectively represent users uu and utSet of scored items, IutRepresenting user uu and utThe set of items that are scored in common,
Figure BDA0003402538950000133
and
Figure BDA0003402538950000134
respectively represent users uu and utAverage score of (3).
sim (i, j) is obtained by the following formula:
Figure BDA0003402538950000135
wherein ,rujRepresenting user uuFor item ijIs given to all items i togetheri and ijA set of scored users.
Step S4: acquiring user weight and project weight by using social labels:
because the labeling and comment information of the user on the item reflects the preference degree of the user on the item to a certain degree, the implicit social relationship between the user and the item can be extracted by utilizing the label weight information of the user and the item. Guk and HikRespectively represent users uuAnd item iiFor label lkThe weight relationship of (2) is obtained by using a TF-IDF algorithm:
Figure BDA0003402538950000141
Figure BDA0003402538950000142
wherein ,cupRepresenting user uuSelection of tag lpThe number of times of the operation of the motor,
Figure BDA0003402538950000143
indicating usage label lpNumber of users, NtIndicates the total number of tags, user uuUse of the label lpThe more times, the weight value G corresponding to the labelukThe larger. c. CipPresentation label lpAppear in item iiThe number of times of the middle-time period,
Figure BDA0003402538950000144
indicating a tagged label lpNumber of items, label lpThe more times of appearance in a certain item set, the more important the label is, and the corresponding weight value HikThe larger.
Step S5: establishing a social recommendation model TSoccialaMF based on social label weight according to the user weight and the project weight:
according to the weight relationship between the user-tag and the item-tag, user and item features are obtained (in this embodiment, the user and item features are indirectly obtained by using a latent semantic model), and then the conditional probabilities of the user-tag and item-tag weight matrices are as follows:
Figure BDA0003402538950000145
Figure BDA0003402538950000146
wherein G is mapped to a user feature U and a label feature space L, and H is mapped to an item feature V and a label feature space L;
Figure BDA0003402538950000147
and
Figure BDA0003402538950000148
the variance of the gaussian distribution with mean 0 to which the user-label and item-label weight relationships respectively obey is respectively.
According to the Bayesian formula, the TSoccialaMF model minimizes the loss function as follows
Figure BDA0003402538950000151
To decompose G andh, while yielding U, V and L:
Figure BDA0003402538950000152
wherein ,
Figure BDA0003402538950000153
Figure BDA0003402538950000154
is the variance of the gaussian distribution with mean 0 obeyed by the label feature L.
Step S6: training a TrustMF model, a USSocialMF model and a TSocialMF model:
the TrustMF model training process comprises the following steps:
setting values of parameters lambda, T and eta, wherein eta is a descending rate; initializing matrixes B and E by random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices B and E according to the following formula based on the objective function in equation (2):
Figure BDA0003402538950000155
Figure BDA0003402538950000156
the training process of the USSocialMF model comprises the following steps:
setting values of parameters λ, R and η, where η is a rate of decrease; initializing matrices R and V with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U and V according to the following equations based on the objective function in equation (6):
Figure BDA0003402538950000157
Figure BDA0003402538950000158
the training process of the TSociialMF model comprises the following steps:
setting values of parameters λ, G, H and η, where η is a rate of descent; initializing matrices G and H with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U, V and L according to the objective function in equation (11) as follows:
Figure BDA0003402538950000161
Figure BDA0003402538950000162
Figure BDA0003402538950000163
step S7: and obtaining the implicit similarity of the user and the item by utilizing a social matrix feedback technology:
implicit similarities between users are noted
Figure BDA0003402538950000164
Is formed by normal distribution of user characteristic similarity, implicit similarity of trust relationship and label similarity, then S(U)The conditional probability distribution of (a) is:
Figure BDA0003402538950000165
wherein ,
Figure BDA0003402538950000166
for user uu and uvThe degree of similarity of the scores of (a) and (b),
Figure BDA0003402538950000167
for preference similarity based on trust relationships,
Figure BDA0003402538950000168
for user uu and uvTag-based implicit preference similarity:
Figure BDA0003402538950000169
Figure BDA00034025389500001610
Figure BDA00034025389500001611
variance (variance)
Figure BDA00034025389500001612
Representing the noise situation of the estimated value, the function f (x) being an improved user uu and uvIs expressed as:
Figure BDA00034025389500001613
relation (u, v) represents user uu and uvThere is a direct trust relationship between them. Therefore, the problem of inaccurate user characteristic description caused by data sparseness and unbalance can be solved, the problem of deviation from real user characteristics caused by linear superposition of user similarity is solved, and user preference can be obtained without explicit user scoring and trust relationship data.
If the user likes an item, the user also tends to like the item with the similar characteristics of the item, so the recommendation quality can be improved by mining the similar item recommendation of the item that the user likes. Similarly, item similarity is set
Figure BDA0003402538950000171
From similarity of item features and similarity of social label relationshipsNormal distribution consists of:
Figure BDA0003402538950000172
according to Bayes reasoning, the following loss function is obtained:
Figure BDA0003402538950000173
wherein item ii and ijSimilarity of (2)
Figure BDA0003402538950000174
The project characteristics and social label influence factors are comprehensively considered, and the calculation formula is as follows:
Figure BDA0003402538950000175
Figure BDA0003402538950000176
representing social tag-based items ii and ijThe calculation formula is as follows:
Figure BDA0003402538950000177
Figure BDA0003402538950000178
representing i based on item featuresi and ijThe calculation formula is as follows:
Figure BDA0003402538950000179
step S8: constructing a social recommendation model EISocialMF based on user scores, trust relationships and social labels:
according to the analysis, the influence of the user trust relationship, the scoring information and the social label on the preference similarity and the project similarity of the user is comprehensively considered, the user and project feature regular term is introduced, and the social recommendation model EISocialMF is obtained, wherein the EISocialMF model is specifically as follows:
Figure BDA0003402538950000181
where sim (u, t) represents user uu and utSim (i, j) represents item ii and ijThe similarity between the two groups is similar to each other,
Figure BDA0003402538950000182
and
Figure BDA0003402538950000183
respectively represent quilt uuNumber of trusted users and trust uuThe number of users of (c);
Figure BDA0003402538950000184
representing the similarity between trusted users of the trusted users;
Figure BDA0003402538950000185
representing the similarity of two users based on a trust relationship, by user uu and uvRespectively trust their common users ukObtaining the trust relationship;
Figure BDA0003402538950000186
representing the similarity of two trusts based on trust relationship, by uu、uvWith co-trusted users ukThe trust relationship of (a) obtains:
Figure BDA0003402538950000191
Figure BDA0003402538950000192
therefore, recommendation can be made based on the obtained eisalicmf model.
And a similarity relation S formed by the user characteristics U, the trust relation characteristics B and E and the label weight relation G is used for restraining the user characteristic space, and a similarity relation S formed by the project characteristics V and the social label weight relation H is used for restraining the project characteristic space, so that the quality of predicting the interest preference of the user is improved.
Fig. 2 is a schematic diagram of a recommendation framework of the eisocial mf model, and a corresponding recommendation process is as follows: establishing a social recommendation model according to explicit interaction information such as user scores and direct trust relationships of users and implicit interaction information such as social labels: a trust relationship model, a CSIT model, and a social tag-based weight relationship model. The specific process is as follows: mapping user characteristics, trustor users and project characteristics to a shared space by using a matrix decomposition technology, and establishing a TrustMF (TrustMF) recommendation model; acquiring the social status of a user in a social network by using PageRank; in combination with a CSIT model, a recommendation model USSocialMF based on the user social status is constructed according to the user score and the trust relationship information; establishing user weight and project weight by using implicit interactive information such as social labels and the like; combining the hidden feature space and the social label weight information to respectively obtain the hidden similarity of the user and the item; combining the socialized recommendation model and the implicit similarity by utilizing a social IT framework structure to establish a recommendation model EISocialMF based on an implicit similarity social feedback technology with an explicit and implicit social relationship fused with a social status; learning and training the model to obtain a user and project hidden feature space; the scores are predicted and recommended.
The above-mentioned embodiments are merely illustrative of the technical solutions of the present invention in a specific embodiment, and any equivalent substitutions and modifications or partial substitutions of the present invention without departing from the spirit and scope of the present invention should be covered by the claims of the present invention.

Claims (9)

1. A recommendation method of a hidden social relationship feedback technology fusing the social status of a user is characterized by comprising the following steps:
constructing a recommendation model TrustMF according to the trust relationship of the user;
calculating the social place value of the user according to a PageRank algorithm, and acquiring the social place weight of the user based on the social place value of the user;
constructing a recommendation model USSocialMF based on the social status of the user according to the user score and the trust relationship;
acquiring a user weight and a project weight by using the social label;
establishing a social recommendation model TSoccialaMF based on social label weight according to the user weight and the project weight;
training a TrustMF model, a USSocialMF model and a TSocialMF model;
obtaining implicit similarity of users and items by utilizing a social matrix feedback technology;
and constructing a social recommendation model EISocialMF based on user scores, trust relationships and social labels.
2. The recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 1, wherein the building of the recommendation model TrustMF according to the trust relationship of the user comprises:
setting two users uu and uvBy the trustperson feature vector B of the useruAnd trusted person feature vector EvNormal distribution expression;
obtaining posterior probability distribution of the characteristic vector of the trustperson and the characteristic vector of the trustperson according to a Bayesian formula:
Figure FDA0003402538940000011
wherein ,TuvRepresenting user uu and uvThe trust relationship of (2);
Figure FDA0003402538940000012
and
Figure FDA0003402538940000013
the variances of Gaussian distribution with the mean value of 0 obeyed by the characteristic vectors of the trustperson and the trustperson respectively;
Figure FDA0003402538940000014
is a trust relationship TuvThe variance of the gaussian distribution obeying a mean of 0; b isu and EvRespectively representing trustperson and trustperson feature vectors; n represents the number of users;
taking logarithm of the formula and maximizing the logarithm to obtain a TrustMF model; and decomposing T by minimizing the loss function l (T, B, E) as follows, resulting in B and E:
Figure FDA0003402538940000021
3. the recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 2, wherein the calculating the social status value of the user according to the PageRank algorithm and obtaining the social status weight of the user based on the social status value of the user comprises:
and calculating the social place value of the user according to the PageRank algorithm, wherein the calculation formula is as follows:
Figure FDA0003402538940000022
wherein ,PRuRepresenting user uuPageRank value of (C)uRepresenting user uuA set of trusted friends that are to be used,
Figure FDA0003402538940000023
is the probability value of jumping out of the current network, and the value range is [0, 1]]；
Adjusted user uu and uvWeight of relationship between WuvThe calculation formula is as follows:
Figure FDA0003402538940000024
wherein ,tuvRepresenting user uuFor user uvThe trust value of (c).
4. The recommendation method of the hidden social relationship feedback technology fusing the user social status according to claim 3, wherein the constructing of the recommendation model USSocialMF based on the user social status according to the user score and the trust relationship comprises:
the USSocialMF model is as follows:
Figure FDA0003402538940000031
wherein R is user score, U and V respectively represent a user hidden factor feature matrix and a project hidden factor feature matrix, and a column vector Uu and ViRespectively corresponding user and project implicit feature vectors, ruiFor user uuFor item iiScore of, NuRepresenting user uuA set of neighboring users of (a); t (u) represents the user uuA set of trusted users; sim (u, t) denotes user uu and utSim (i, j) represents item ii and ijThe similarity between them is obtained by the following formula:
Figure FDA0003402538940000032
Figure FDA0003402538940000033
wherein ,rui and rtiRespectively representUser uu and utFor item iiScore of (I)u and ItRespectively represent users uu and utSet of scored items, IutRepresenting user uu and utThe set of items that are scored in common,
Figure FDA0003402538940000034
and
Figure FDA0003402538940000035
respectively represent users uu and utAverage score of (a); r isujRepresenting user uuFor item ijIs given to all items i togetheri and ijA set of scored users;
by minimizing l (R, U, V, W), R and W are decomposed to obtain feature matrices U and V.
5. The recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 4, wherein the obtaining of the user weight and the item weight by using the social tag comprises:
using the TF-IDF algorithm to obtain:
Figure FDA0003402538940000041
Figure FDA0003402538940000042
wherein ,Guk and HikRespectively represent users uuAnd item iiFor label lkThe weight relationship of (c); c. CupRepresenting user uuSelection of tag lpThe number of times of the operation of the motor,
Figure FDA0003402538940000043
indicating usage label lpBy usingNumber of houses, NtIndicates the total number of tags, user uuUse of the label lpThe more times, the weight value G corresponding to the labelukThe larger; c. CipPresentation label lpAppear in item iiThe number of times of the middle-time period,
Figure FDA0003402538940000044
indicating a tagged label lpNumber of items, label lpThe more times of occurrence in a certain item set, the corresponding weight value HikThe larger.
6. The recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 5, wherein the building of the social recommendation model TSocialmF based on the social label weight according to the user weight and the item weight comprises:
according to the weight relationship between the user-label and the item-label, the user and item characteristics are obtained, and then the conditional probabilities of the user-label and item-label weight matrixes are as follows:
Figure FDA0003402538940000045
Figure FDA0003402538940000046
wherein G is mapped to a user feature U and a label feature space L, and H is mapped to an item feature V and a label feature space L;
Figure FDA0003402538940000047
and
Figure FDA0003402538940000048
the variance of the gaussian distribution with the mean value of 0 obeyed by the user-label and item-label weight relationships respectively;
according to Bayesian formulation, the TSociialMF model is as follows, and L (U, V, L, G, H) is minimized to decompose G and H, resulting in U, V and L:
Figure FDA0003402538940000049
Figure FDA0003402538940000051
wherein ,
Figure FDA0003402538940000052
Figure FDA0003402538940000053
is the variance of the gaussian distribution with mean 0 obeyed by the label feature L.
7. The recommendation method of the implicit social relationship feedback technology fusing in user social status as claimed in claim 6, wherein the training of TrustMF model, USSocialMF model and TSocialMF model comprises:
the training process of the TrustMF model comprises the following steps:
setting values of parameters lambda, T and eta, wherein eta is a descending rate; initializing matrixes B and E by random numbers uniformly distributed according to [0, 1 ]; iteratively updating matrices B and E according to the following equations:
Figure FDA0003402538940000054
Figure FDA0003402538940000055
the training process of the USSocialMF model comprises the following steps:
setting values of parameters λ, R and η, where η is a rate of decrease; initializing matrices R and V with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U and V according to the following formula:
Figure FDA0003402538940000056
Figure FDA0003402538940000057
the training process of the TSociialMF model comprises the following steps:
setting values of parameters λ, G, H and η, where η is a rate of descent; initializing matrices G and H with random numbers uniformly distributed according to [0, 1 ]; iteratively updating the matrices U, V and L according to the following formula:
Figure FDA0003402538940000058
Figure FDA0003402538940000059
Figure FDA00034025389400000510
8. the recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 6, wherein the obtaining the implicit similarity between the user and the item by using the social matrix feedback technology comprises:
implicit similarities between users are noted
Figure FDA0003402538940000061
Is formed by normal distribution of user characteristic similarity, implicit similarity of trust relationship and label similarity, then S(U)The conditional probability distribution of (a) is:
Figure FDA0003402538940000062
wherein ,
Figure FDA0003402538940000063
for user uu and uvThe degree of similarity of the scores of (a) and (b),
Figure FDA0003402538940000064
for preference similarity based on trust relationships,
Figure FDA0003402538940000065
for user uu and uvTag-based implicit preference similarity:
Figure FDA0003402538940000066
Figure FDA0003402538940000067
Figure FDA0003402538940000068
variance (variance)
Figure FDA0003402538940000069
Representing the noise situation of the estimated value, the function f (x) being an improved user uu and uvIs expressed as:
Figure FDA00034025389400000610
relation (u, v) represents user uu and uvA direct trust relationship exists between the two;
setting item similarity
Figure FDA00034025389400000611
The method is characterized by comprising the following steps of (1) normally distributing the item feature similarity and the social label relationship similarity:
Figure FDA0003402538940000071
according to Bayes reasoning, the following loss function is obtained:
Figure FDA0003402538940000072
wherein item ii and ijSimilarity of (2)
Figure FDA0003402538940000073
The project characteristics and social label influence factors are comprehensively considered, and the calculation formula is as follows:
Figure FDA0003402538940000074
Figure FDA0003402538940000075
representing social tag-based items ii and ijThe calculation formula is as follows:
Figure FDA0003402538940000076
Figure FDA0003402538940000077
representing i based on item featuresi and ijThe calculation formula is as follows:
Figure FDA0003402538940000078
9. the recommendation method of the hidden social relationship feedback technology fusing the social status of the user according to claim 8, wherein the constructing of the social recommendation model eisociallmf based on the user score, the trust relationship and the social label comprises:
the EISocialMF model is as follows:
Figure FDA0003402538940000081
wherein ,
Figure FDA0003402538940000082
representing the similarity between trusted users of the trusted users;
Figure FDA0003402538940000083
representing the similarity of two users based on a trust relationship, by user uu and uvRespectively trust their common users ukObtaining the trust relationship;
Figure FDA0003402538940000084
representing the similarity of two trusts based on trust relationship, by uu、uvWith co-trusted users ukThe trust relationship of (a) obtains:
Figure FDA0003402538940000085
Figure FDA0003402538940000086
CN202111500966.8A 2021-12-09 2021-12-09 Recommendation method of hidden social relation feedback technology integrating user social status Active CN114154902B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111500966.8A CN114154902B (en) 2021-12-09 2021-12-09 Recommendation method of hidden social relation feedback technology integrating user social status

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111500966.8A CN114154902B (en) 2021-12-09 2021-12-09 Recommendation method of hidden social relation feedback technology integrating user social status

Publications (2)

Publication Number Publication Date
CN114154902A true CN114154902A (en) 2022-03-08
CN114154902B CN114154902B (en) 2023-04-28

Family

ID=80454192

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111500966.8A Active CN114154902B (en) 2021-12-09 2021-12-09 Recommendation method of hidden social relation feedback technology integrating user social status

Country Status (1)

Country Link
CN (1) CN114154902B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103399858A (en) * 2013-07-01 2013-11-20 吉林大学 Socialization collaborative filtering recommendation method based on trust
US20150187024A1 (en) * 2013-12-27 2015-07-02 Telefonica Digital España, S.L.U. System and Method for Socially Aware Recommendations Based on Implicit User Feedback
CN111400612A (en) * 2020-03-11 2020-07-10 重庆邮电大学 Personalized recommendation method integrating social influence and project association
CN111506830A (en) * 2020-04-09 2020-08-07 中南大学 Collaborative filtering recommendation algorithm fusing social trust influence

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103399858A (en) * 2013-07-01 2013-11-20 吉林大学 Socialization collaborative filtering recommendation method based on trust
US20150187024A1 (en) * 2013-12-27 2015-07-02 Telefonica Digital España, S.L.U. System and Method for Socially Aware Recommendations Based on Implicit User Feedback
CN111400612A (en) * 2020-03-11 2020-07-10 重庆邮电大学 Personalized recommendation method integrating social influence and project association
CN111506830A (en) * 2020-04-09 2020-08-07 中南大学 Collaborative filtering recommendation algorithm fusing social trust influence

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
陈锐: "《社交网络中多因子融合的个性化推荐方法研究》", 《优秀博士论文》 *

Also Published As

Publication number Publication date
CN114154902B (en) 2023-04-28

Similar Documents

Publication Publication Date Title
Hindman Building better models: Prediction, replication, and machine learning in the social sciences
Cheng et al. Evaluation methods and measures for causal learning algorithms
Pujahari et al. Pair-wise preference relation based probabilistic matrix factorization for collaborative filtering in recommender system
Bashar et al. Performance of machine learning algorithms in predicting the pavement international roughness index
Deng et al. An evidential game theory framework in multi-criteria decision making process
US11030265B2 (en) Cross-platform data matching method and apparatus, computer device and storage medium
Navgaran et al. Evolutionary based matrix factorization method for collaborative filtering systems
Zhang et al. Trust from the past: Bayesian personalized ranking based link prediction in knowledge graphs
CN111143704B (en) Online community friend recommendation method and system integrating user influence relationship
CN108268464A (en) A kind of personalized recommendation method and device returned based on collaborative filtering and logistic
Veeraiah et al. Machine Learning Frameworks in Carpooling
Mele A structural model of homophily and clustering in social networks
Otani et al. Quality control of crowdsourced classification using hierarchical class structures
Deng et al. Probabilistic matrix factorization recommendation approach for integrating multiple information sources
CN110795640B (en) Self-adaptive group recommendation method for compensating group member difference
Ramezani et al. A new generalized collaborative filtering approach on sparse data by extracting high confidence relations between users
Ifada et al. Do-rank: DCG optimization for learning-to-rank in tag-based item recommendation systems
CN114154902B (en) Recommendation method of hidden social relation feedback technology integrating user social status
CN108320218B (en) Personalized commodity recommendation method based on trust-score time evolution two-way effect
CN108960954B (en) Content recommendation method and system based on user group behavior feedback
Xie et al. A unified framework for long range and cold start forecasting of seasonal profiles in time series
Wang et al. Multi‐feedback Pairwise Ranking via Adversarial Training for Recommender
Chen et al. From tie strength to function: Home location estimation in social network
CN113095084A (en) Semantic service matching method and device in Internet of things and storage medium
Koskinen et al. Bayesian analysis of multilevel probit models for data with friendship dependencies

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