CN117216376A - Fair perception recommendation system and recommendation method based on depth map neural network - Google Patents
Fair perception recommendation system and recommendation method based on depth map neural network Download PDFInfo
- Publication number
- CN117216376A CN117216376A CN202310595117.8A CN202310595117A CN117216376A CN 117216376 A CN117216376 A CN 117216376A CN 202310595117 A CN202310595117 A CN 202310595117A CN 117216376 A CN117216376 A CN 117216376A
- Authority
- CN
- China
- Prior art keywords
- recommendation
- sensitive
- tower
- user
- attribute
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000008447 perception Effects 0.000 title claims abstract description 36
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 33
- 230000035945 sensitivity Effects 0.000 claims abstract description 38
- 230000013016 learning Effects 0.000 claims abstract description 19
- 230000028161 membrane depolarization Effects 0.000 claims abstract description 11
- 230000008485 antagonism Effects 0.000 claims abstract description 4
- 239000013598 vector Substances 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 25
- 230000006870 function Effects 0.000 claims description 23
- 238000013527 convolutional neural network Methods 0.000 claims description 18
- 230000003993 interaction Effects 0.000 claims description 15
- 230000002776 aggregation Effects 0.000 claims description 12
- 238000004220 aggregation Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 9
- 238000003062 neural network model Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 239000004973 liquid crystal related substance Substances 0.000 claims description 4
- 238000009825 accumulation Methods 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000003042 antagnostic effect Effects 0.000 claims description 2
- 238000005516 engineering process Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000000052 comparative effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000000547 structure data Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 206010070834 Sensitisation Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000008313 sensitization Effects 0.000 description 1
- 230000001568 sexual effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to a fair perception recommendation system and a recommendation method based on a depth map neural network, which pass through a recommendation tower f R Predicting the score of the user on the project and inputting the score into the sensitive attribute countermeasure f A The method comprises the steps of carrying out a first treatment on the surface of the Sensitive attribute estimation tower f E Predicting sensitive properties and learning user-embedded representations to limit sensitive properties against f A The method comprises the steps of carrying out a first treatment on the surface of the Through recommendation tower f R Input and sensitivity attribute estimation tower f E After the restriction of (a), sensitive attribute countermeasure (f) A Inferring sensitive attributes, and estimating tower f of sensitive attributes E Minimum constitutionMaximum game ensures fairness of user representation; the mutual information constraint module is used for relieving the problem of discrimination caused by failure of antagonism depolarization and convergence, and fairness of recommendation results of different sensitive attribute groups is ensured through fairness perception constraint; and the influence of sensitive attributes on the recommendation result is eliminated from the whole, so that the continuous and healthy operation of the platform is realized.
Description
Technical Field
The invention relates to the technical field of recommendation methods, in particular to a fair perception recommendation system and a recommendation method based on a depth map neural network.
Background
In the information age of the prosperous development of the internet, the recommendation method has become an important technology in human life. Because of the burst of big data, the overload of the information resources on the network can be quickly and conveniently screened out interested items for users without clear requirements by using the recommendation method.
Over the past few years, recommendation methods have played a vital role in finding items of interest to users and in alleviating the challenges of information overload. As a highly data driven application, the outcome of the recommendation method is susceptible to data bias and algorithm bias, resulting in an unfair recommendation. In addition, unfair recommendation methods recommend biased results based on implicit sensitivity attributes, which can lead to discrimination of the recommended results to the user and thus serious social impact. Therefore, ensuring fairness of the recommended results is particularly important.
As one of the mainstream tools of the recommendation method, the graph neural network has been remarkably successful in many recommendation scenarios. The aggregated neighbor mechanism and the higher order messaging mechanism of the graph neural network may then deepen the bias of the recommendation result against the user sensitive attribute, resulting in unfairness of the recommendation result.
To address this problem, fairGNN enhances fairness of node representations by alternately training encoders and discriminators against sexual learning; fairGo converts the original user and project embedding into a filtered embedding space based on a sensitive feature set by learning a combination of filters. The model then applies resistance learning on the user-centric graph to achieve fair recommendation.
The method adopts an opposite learning framework to learn fair representation, only ensures that the user embedding is irrelevant to sensitive information, but ignores the fact that the item embedding can also imply sensitive attribute information in the propagation process. In addition, the more classes of sensitive attributes, the harder it is to guarantee fairness.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a fair perception recommendation system and a recommendation method based on a depth map neural network, which eliminate the influence of sensitive attributes on recommendation results from a global angle, remarkably improve the fairness of recommendation and realize continuous and healthy operation of a platform.
In order to achieve the above objective, in one aspect, the present invention provides a fair perception recommendation system based on a depth map neural network, including a recommendation tower f R Sensitive attribute estimation tower f E Sensitive attribute countermeasure f A Mutual information constraint module and fairness perception constraint module, the recommendation tower f R And a sensitivity attribute estimation tower f E All comprise an embedded layer, a graph convolution neural network model and a multi-layer perceptron model, wherein the sensitive attribute countermeasure f A Only the multi-layer perceptron model;
the recommendation tower f R For predicting user scoring of items: converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, and then applying a graph convolutional neural network model to encode neighbor features to generate item embedding and user embedding representations after information aggregation; finally, calculating dot products of project embedding and user embedding, and inputting calculation results into the multi-layer perceptron model to predict the scores of the users on the projects;
the sensitive attribute estimation tower f E User sensitive properties for predictive property unknowns: the high-dimensional sparse feature vector is converted into a low-dimensional dense feature vector through an embedding layer, and then a graph convolutional neural network model is applied to information aggregationCombining to obtain node representation on the user-project interaction diagram, and finally predicting sensitive attributes of the unknown user through the multi-layer perceptron model;
The sensitive attribute challenge f A For recommending tower modules f R Deducing sensitive attributes from the generated embedded content;
the mutual information constraint module is used for relieving the problem of discrimination of the recommender caused by failure of antagonism depolarization convergence;
the fairness perception constraint module is used for improving recommendation accuracy and ensuring fairness of recommendation results of different sensitive attribute groups when different sensitive groups expect equal recommendation quality.
Specifically, the sensitive attribute is against f A Estimating tower f using sensitive properties E The predicted sensitive attribute is used as a label training model and is based on the recommendation tower f R The generated user embedded content deduces the user sensitive attribute; recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair.
Specifically, the mutual information constraint module is used for recommending the tower f through minimization R The generated estimation tower f for predicting the user scoring and sensitive attribute of the project E And the generated dependency relationship between the sensitive attributes is used for relieving the problem of discrimination of the recommender caused by failure of convergence of resistive depolarization.
Specifically, when the recommendation quality is expected to be equal by different sensitive groups, the fairness perception constraint module improves the recommendation accuracy by reducing the difference between the prediction scores and the recommendation quality among users of different sensitive attribute groups, so as to ensure the fairness of the recommendation results of the different sensitive attribute groups.
On the other hand, the invention also provides a recommendation method of the fairness-aware recommendation system, which comprises the following steps:
step S1, dividing the heterogeneous graph G into a plurality of subgraphs { G1, G2, & gt. GR }, wherein R represents the number of subgraphs and is taken as a recommendation tower f according to the user-project interaction scores R And a sensitivity attribute estimation tower f E Is input X of (a);
step S2, inputting the initial data X obtained in the step S1 into a recommendation tower f R Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to encode neighbor features, generating item embedding and user embedding representations after information aggregation, finally calculating dot products of the item embedding and the user embedding, and inputting calculation results into a multi-layer perceptron model to predict the scoring of the item by a user;
S3, inputting the initial data X obtained in the step S1 into a sensitive attribute estimation tower f E Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to perform information aggregation to obtain node representation on a user-project interaction graph, and finally predicting sensitive attributes of users with unknown attributes through a multi-layer perceptron model;
step S4, sensitive attribute countermeasure f A Estimating tower f by using sensitive attribute in step S3 E The predicted sensitive attribute is used as a label training model, and a tower f is recommended according to the step S2 R The generated embedded content deduces sensitive attribute to make recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair;
s5, reducing the difference between the prediction scores and the recommendation quality among users of different sensitive attribute groups through a fair perception constraint module, and minimizing a recommendation tower f through a mutual information constraint module R The generated predictive user's score and sensitivity attribute for the itemEstimation tower f E And when the trained parameters tend to be stable, the model training process is finished.
Specifically, the recommendation tower f in step S2 R The method comprises the steps of predicting scoring of projects by a user, converting high-dimensional sparse feature vectors into low-dimensional dense feature vectors through an embedding layer, connecting the embedded vectors with the same length into initial node features, and inputting the initial node features into a graph neural network model:
X ′ ＝E(X) (1)
in the above, X ′ Is an embedding matrix of nodes, E ()' represents an embedding operator;
since the nodes and edges in heterogeneous graphs are of different types, their features are typically located in different spaces, thus in subgraph G r The graph convolutional neural network is applied to encode the neighbor features:
in the above-mentioned method, the step of,a user representation representing a 1-hop neighborhood after propagation of the l-layer graph convolutional neural network model,representing user u in subgraph G r Item set of upper interaction, d u And d i Representing the degree, σ (), non-linear activation function, W, of user u and item i, respectively (l) Is the parameter matrix of layer l, < >>Is an initial representation of user u, aggregating representations obtained from different subgraphs using a pooling operation:
In the above description, accsum ()' represents accumulation operation, the propagation process of item nodes is similar to that of user nodes, and user embedding h is calculated u And item embedding h i And input the results to a multi-layer perceptron model to predict final scoresThe prediction function may be written as:
where w and b are trainable parameters of the multi-layer perceptron model,is a recommendation tower f R Loss value of->Is a softmax function, +.>Is the calculation of dot product, recommending tower f R Loss value of +.>Expressed as:
specifically, in step S3, the sensitivity attribute estimation tower f E Since the graph neural network is more likely to expose sensitive information, it is used as an estimator to predict unknown sensitive attributes, and recommendation tower f R Similarly, information propagation using a graph convolution neural network to obtain node representations on a user-project interaction graph, and then using a multi-layer perceptron to obtain predicted user-sensitive attributes
In the above, N u Representing a set of neighbor nodes for user u, W (l) Is the parameter matrix of layer l, W and b are trainable parameters, whereby the sensitivity attribute estimation tower f E The loss function of (2) can be written as:
in the above equation, s is the true value of the attribute, Is the predicted attribute value, V u Is a training user set, < >>Is a sensitive attribute estimation tower f E Is a loss of (2).
Specifically, the sensitivity attribute countermeasure f in step S4 A According to recommendation tower f R The generated embedded content infers sensitive attributes against f A Will recommend tower f R The generated user node representation h u As input and attempt to estimate the user's sensitive propertiesI.e. < ->Firstly, updating the parameters of the countermeasure to maximally improve the sensitivity attribute countermeasure f A The possibility of distinguishing sensitive information in the embedding and then fixing the sensitive attribute countermeasure f A Parameters to minimize differences between the generated representations of different sensitive property groups, which causes sensitive property countermeasure f A It is not possible to distinguish which sensitive group the user belongs to, forming a min-max game, expressed as:
in the above, θ ε 、θ R And theta A Respectively represent sensitive attribute estimation towers f E Recommendation tower f R And sensitive attribute countermeasure f A Is used for the control of the temperature of the liquid crystal display device,is a sensitive attribute countermeasure f A Loss value of->Representing user embedding h u When the parameter is theta ε 、θ R And theta A When using predicted sensitivity attribute k i Probability distribution represented by the user; k is the number of unique sensitivity attribute values, representing sensitivity attribute countermeasure f A The probability of the predicted outcome being.
Specifically, the difference between the prediction scores and the recommended quality among the users of the different sensitive attribute groups is reduced by a fairness perception constraint module in step S5, the fairness perception constraint module minimizes the statistical parity delta SP and the opportunity equality delta EO by adding constraint conditions in the loss function,and->Is a fair perceived constraint loss, < >>Narrowing the predictive score gap between different sensitive property groups and +.>Then focus is placed on the recommended quality gap between users focusing on different sets of sensitive attributes and their preference items (top scores);
in the above, if and only ifWhen (I)>Reaching a global minimum; in general, the above conditions are not the same as the actual situation, thus achieving +.>The global minimum of (2) would impair the accuracy of the recommendation for +.>The condition of the global minimum is sum +.>Equal to->In particular, when both items are equal to 0 at the same time, the accuracy and +.>Are all optimal; therefore, it is desirable to increase the recommendation accuracy as much as possible while ensuring that different sensitive groups expect the same recommendation quality; notably, constraints can be effectively applied to multiple classes of attributes, so the final objective function is expressed as:
in the above formula, α, β, γ, and λ control weights of mutual information constraint, antagonistic depolarization, statistical parity fairness constraint, and opportunity equality fairness constraint, respectively.
Specifically, in step S5, the recommendation tower f is minimized by the mutual information constraint module R The generated estimation tower f for predicting the user's scoring and sensitive attribute of the project E The generated dependency relationship between sensitive attributes is defined as follows:
in the above, D KL Indicating Kullback-Leibler divergence,representing sensitivity attribute estimation tower f E Predicted sensitivity attribute,/->Is a recommendation tower f R Predicted score,/->Is->And->Is a combination of (a) and (b) of (b)>Representing marginal distribution +.>Andproduct of>Is the mutual information constraint loss, is a non-negative number, and the global minimum value 0 means +.>And->Is independent.
Compared with the prior art, the invention has the beneficial effects that:
(1) The book is provided withInventive recommendation tower f R Predicting user scoring of items and inputting the generated embedded content into a sensitive attribute countermeasure f A The method comprises the steps of carrying out a first treatment on the surface of the Sensitive attribute estimation tower f E Sensitive attributes for users whose predicted attributes are unknown, learning user-embedded representations to limit subtended sensitivity attributes f A The sensitive group to which the user node belongs cannot be distinguished; through recommendation tower f R After input of (2) and limitation of sensitive attribute estimation tower fE, sensitive attribute countermeasure f A For inferring sensitive properties, with sensitive properties estimation tower f E Forming a minimum-maximum game, and ensuring fairness of user representation; mutual information constraint is used for relieving the problem of discrimination of a recommender caused by failure of opposite depolarization convergence, fairness perception constraint is used for ensuring that recommendation results of different sensitive attribute groups are fairer, and dependence between the recommendation results and sensitive information is limited from a global angle.
(2) The graphic neural network and the countermeasure training are introduced to innovate the unfairness of the traditional recommendation method, and a mutual information constraint and two fairness perception constraints are introduced to solve the problem that the fairness of the final recommendation result cannot be guaranteed in the countermeasure training, and meanwhile, the fairness performance is realized while the accuracy of the recommendation is kept almost the same, the transformation of the recommendation method is completed, the users and the information are contacted, the users are helped to find valuable information of the users, the information can be displayed in front of the users interested in the recommendation method, the personalized requirements of recommendation are met, so that win-win of information consumers and information producers are realized, the utilization efficiency of resources is improved, and the continuous, stable and healthy operation of a platform is guaranteed to the greatest extent.
Drawings
Fig. 1 is a schematic structural diagram of a fair perception recommendation system based on a depth map neural network;
fig. 2 is an algorithm flow diagram of a fair perception recommendation method based on a depth map neural network.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The graphic neural network and the countermeasure training technology, the graphic neural network is an algorithm generic term for learning graphic structure data by using the neural network, extracting and exploring features and modes in the graphic structure data and meeting the requirements of graphic learning tasks such as clustering, classifying, predicting, segmenting, generating and the like, and the idea is how to iteratively aggregate feature information from neighbors and combine the aggregated information with the current central node representation. The graph neural network technology can capture the dependency relationship of the graph through a neighbor aggregation and message transfer mechanism among graph nodes so as to solve the limitation of the traditional deep learning on the processing of topological structure relationship data; deep learning is generally divided into two categories: one is a generation countermeasure network (Generative Adversarial Networks, GAN), belonging to a generation model; the other is challenge attack, challenge defense, etc., using a min-max training approach similar to generating a challenge network, but with the difference that they are concerned with the robustness of the model under disturbance. In the embodiment, a first type of method is adopted, sensitive attributes are predicted from the representation of the encoder by using the countermeasure, the purpose of the encoder is to learn the representation, the encoder is used for deceptively predicting the countermeasure, the task is ensured to be accurately predicted, and the countermeasure depolarization is realized and is mainly applied to the following scenes:
In the intelligent development maintenance and evolution process of an open-source community group, users or other developers can fairly and accurately recommend proper question respondents to questioners in the open-source community in time and accurately when new function requests, poor experience of the existing functions, encountered errors and other problems are presented after the open-source project is used, and intelligent recommendation of electronic commerce, music and movie platforms is achieved, so that the accurate and personalized requirements of recommendation are met.
As shown in FIG. 1, in this embodiment, a fair perception recommendation system based on a depth map neural network is providedThe system comprises a recommendation tower f R Sensitive attribute estimation tower f E Sensitive attribute countermeasure f A Mutual information constraint module and fairness perception constraint module, the recommendation tower f R And a sensitivity attribute estimation tower f E All comprise an embedded layer, a graph convolution neural network model and a multi-layer perceptron model, wherein the sensitive attribute countermeasure f A Only the multi-layer perceptron model;
the recommendation tower f R For predicting user scoring of items: converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, and then applying a graph convolutional neural network model to encode neighbor features to generate item embedding and user embedding representations after information aggregation; finally, calculating dot products of project embedding and user embedding, and inputting calculation results into the multi-layer perceptron model to predict the scores of the users on the projects;
The sensitive attribute estimation tower f E User sensitive properties for predictive property unknowns: converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to perform information aggregation to obtain node representation on a user-project interaction graph, and finally predicting sensitive attributes of users with unknown attributes through a multi-layer perceptron model;
the sensitive attribute challenge f A For recommending tower modules f R The generated embedded content deduces sensitive attribute as a label training model and according to the recommendation tower f R The generated user embedded content deduces the user sensitive attribute; recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair
The mutual information constraint module is used for recommending the tower f through the minimum R Generated predictive user scoring and sensitive attribute estimation for itemsTower f E The generated dependency relationship between sensitive attributes is used for relieving the problem of discrimination of the recommender caused by failure of reactive depolarization convergence;
The fairness perception constraint module is used for improving the recommendation accuracy by reducing the difference between the prediction scores and the recommendation quality of the users of different sensitive attribute groups when the recommendation quality of different sensitive groups is expected to be equal, so that fairness of recommendation results of the different sensitive attribute groups is ensured.
According to the above, the present embodiment also provides a recommendation method of the fairness-aware recommendation system, including the following steps:
step S1, dividing the heterogeneous graph G into a plurality of subgraphs { G1, G2, & gt. GR }, wherein R represents the number of subgraphs and is taken as a recommendation tower f according to the user-project interaction scores R And a sensitivity attribute estimation tower f E Is input X of (a);
step S2, inputting the initial data X obtained in the step S1 into a recommendation tower f R Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to encode neighbor features, generating item embedding and user embedding representations after information aggregation, finally calculating dot products of the item embedding and the user embedding, and inputting calculation results into a multi-layer perceptron model to predict the scoring of the item by a user;
s3, inputting the initial data X obtained in the step S1 into a sensitive attribute estimation tower f E Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to perform information aggregation to obtain node representation on a user-project interaction graph, and finally predicting sensitive attributes of users with unknown attributes through a multi-layer perceptron model;
step S4, sensitive attribute countermeasure f A Estimating tower f by using sensitive attribute in step S3 E The predicted sensitive attribute is used as a label training model, and a tower f is recommended according to the step S2 R The generated embedded content deduces sensitive attribute to make recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair;
s5, reducing the difference between the prediction scores and the recommendation quality among users of different sensitive attribute groups through a fair perception constraint module, and minimizing a recommendation tower f through a mutual information constraint module R The generated estimation tower f for predicting the user's scoring and sensitive attribute of the project E And when the trained parameters tend to be stable, the model training process is finished.
Specifically, the recommendation tower f in step S2 R The method comprises the steps of predicting scoring of projects by a user, converting high-dimensional sparse feature vectors into low-dimensional dense feature vectors through an embedding layer, connecting the embedded vectors with the same length into initial node features, and inputting the initial node features into a graph neural network model:
X ′ ＝E(X) (1)
in the above, X ′ Is an embedding matrix of nodes, E ()' represents an embedding operator;
since the nodes and edges in heterogeneous graphs are of different types, their features are typically located in different spaces, thus in subgraph G r The graph convolutional neural network is applied to encode the neighbor features:
in the above-mentioned method, the step of,a user representation representing a 1-hop neighborhood after propagation of the l-layer graph convolutional neural network model,representing user u in subgraph G r Go up and crossMutual item set, d u And d i Representing the degree, σ (), non-linear activation function, W, of user u and item i, respectively (l) Is the parameter matrix of layer l, < >>Is an initial representation of user u, aggregating representations obtained from different subgraphs using a pooling operation:
in the above description, accsum ()' represents accumulation operation, the propagation process of item nodes is similar to that of user nodes, and user embedding h is calculated u And item embedding h i And input the results to a multi-layer perceptron model to predict final scoresThe prediction function may be written as:
where w and b are trainable parameters of the multi-layer perceptron model,is a recommendation tower f R Loss value of->Is a softmax function, +.>Is the calculation of dot product, recommending tower f R Loss value of +.>Expressed as:
in particular, the method comprises the steps of,the sensitive attribute estimation tower f in step S3 E Since the graph neural network is more likely to expose sensitive information, it is used as an estimator to predict unknown sensitive attributes, and recommendation tower f R Similarly, information propagation using a graph convolution neural network to obtain node representations on a user-project interaction graph, and then using a multi-layer perceptron to obtain predicted user-sensitive attributes
In the above, N u Representing a set of neighbor nodes for user u, W (l) Is the parameter matrix of layer l, W and b are trainable parameters, whereby the sensitivity attribute estimation tower f E The loss function of (2) can be written as:
in the above equation, s is the true value of the attribute,is the predicted attribute value, V u Is a training user set, < >>Is a sensitive attribute estimation tower f E Is a loss of (2).
Specifically, the sensitivity attribute countermeasure f in step S4 A According to recommendation tower f R The generated embedded content infers sensitive attributes against f A Will recommend tower f R The generated user node representation h u As input and attempt to estimate the user's sensitive propertiesI.e. < ->Firstly, updating the parameters of the countermeasure to maximally improve the sensitivity attribute countermeasure f A The possibility of distinguishing sensitive information in the embedding and then fixing the sensitive attribute countermeasure f A Parameters to minimize differences between the generated representations of different sensitive property groups, which causes sensitive property countermeasure f A It is not possible to distinguish which sensitive group the user belongs to, forming a min-max game, expressed as:
in the above, θ ε 、θ R And theta A Respectively represent sensitive attribute estimation towers f E Recommendation tower f R And sensitive attribute countermeasure f A Is used for the control of the temperature of the liquid crystal display device,is a sensitive attribute countermeasure f A Loss value of->Representing user embedding h u When the parameter is theta ε 、θ R And theta A When using predicted sensitivity attribute k i Probability distribution represented by the user; k is the number of unique sensitivity attribute values, representing sensitivity attribute countermeasure f A The probability of the predicted outcome being.
Min-max gaming generates a fair feasibility analysis of the user representation (applicable to multi-class sensitive attributes):
proposition 1: if equation (9) sensitive attribute countermeasure f A With sufficient capacity, its optimal solutionIs that
And (3) proving: constructing Lagrange dual function of formula (9) and constrainingIs solved under the constraint condition of (2).
Proposition 2: the goal of the min-max game formula is equivalent to maximizing the sensitivity attributeAnd user embedding h u Conditional entropy between; if and only if->The global maximum of conditional entropy is reached.
And (3) proving: challenge f by replacing the optimal sensitive attribute in proposition 1 A The target in equation (9) is equal to:
according to the maximum entropy model,satisfaction, it can be demonstrated that
Although the contribution of user embedding to fairness can be achieved through such min-max gaming, sensitive information may still be included in item embedding and recommendation scores during propagation of the graph neural networkThe fairness is still not guaranteed due to the common decision of user embedding and item embedding, since it is difficult to directly remove sensitive properties in the item embedded representation.
Therefore, in this embodiment, a mutual information constraint module and a fairness perception constraint module are provided, by limitingAnd sensitization toFeel attribute->The dependency relationship between the user and the item eliminates the influence of the sensitive attribute on the recommendation result from the global angle so as to ensure that the recommendation result determined by the user and the item embedding is fair for different sensitive attribute groups.
Specifically, the difference between the prediction scores and the recommended quality among the users of the different sensitive attribute groups is reduced by a fairness perception constraint module in step S5, the fairness perception constraint module minimizes the statistical parity delta SP and the opportunity equality delta EO by adding constraint conditions in the loss function,and->Is a fair perceived constraint loss, < >>Narrowing the predictive score gap between different sensitive property groups and +.>Then focus is placed on the recommended quality gap between users focusing on different sets of sensitive attributes and their preference items (top scores);
in the above, if and only ifWhen (I)>Reaching a global minimum; in general, the above conditions are not the same as the actual situation, thus achieving +.>The global minimum of (2) would impair the accuracy of the recommendation for +.>The condition of the global minimum is sum +.>Equal to->In particular, when both items are equal to 0 at the same time, the accuracy and +.>Are all optimal; therefore, it is desirable to increase the recommendation accuracy as much as possible while ensuring that different sensitive groups expect the same recommendation quality; notably, constraints can be effectively applied to multiple classes of attributes, so the final objective function is expressed as:
in the above, alpha, beta, gamma and lambda control mutual information constraint, antagonism depolarization, statistical parity constraint and opportunity equalization respectively
The weight of the fairness constraint.
Specifically, in step S5, the recommendation tower f is minimized by the mutual information constraint module R The generated estimation tower f for predicting the user's scoring and sensitive attribute of the project E The generated dependency relationship between sensitive attributes is defined as follows:
in the above, D KL Indicating Kullback-Leibler divergence,representing sensitive attribute estimatesJi Da f E Predicted sensitivity attribute,/->Is a recommendation tower f R Predicted score,/->Is->And->Is a combination of (a) and (b) of (b)>Representing marginal distribution +.>Andproduct of>Is the mutual information constraint loss, is a non-negative number, and the global minimum value 0 means +.>And->Is independent.
As shown in fig. 2, in order to make the calculation process of the method of the present invention clearer, it is convenient to understand that the following embodiment further provides an algorithm flow of the fair perception recommendation method based on the depth map neural network:
step one, randomly initializing a parameter θ ε 、θ R And theta A The following steps are executed and repeated;
step two, obtaining user embedded representation and recommendation scores through a recommendation tower fR;
step three, acquiring sensitive attributes through a sensitive attribute estimation tower fE;
step four, obtaining a predicted sensitive attribute embedded by a user through a sensitive attribute countermeasure fA;
Step five, defining a loss function
Step six, updating the parameter theta ε ：
Step seven, updating the parameter theta R ：
Step eight, updating the parameter theta A ：Up to theta ε 、θ R And theta A Reaching convergence;
step nine, returning sensitive attribute estimation tower f E Recommendation tower f R And challenge f by sensitive properties A Is used for training the parameters of the system.
In the above flow, G represents the heterogram, V represents the node set, E represents the edge set, X represents the node feature matrix, R represents the user rating set, S represents the user sensitive attribute set, α, β, γ and λ represent the weight parameters, η is the learning rate, θ ε 、θ R And theta A Respectively represent sensitive attribute estimation towers f E Recommendation tower f R And sensitive attribute countermeasure f A Is used for the control of the temperature of the liquid crystal display device,representing a recommendation tower f R Loss of->Is a sensitive attribute estimation tower f E Loss of->Representing sensitive attribute challenge f A Loss of->Representing mutual information constraint loss-> Representing fairness perception constraint loss, f E 、f R And f A Respectively represent sensitive attribute estimation towers f E Recommendation tower f R And challenge f by sensitive properties A Is used for training the parameters of the system.
The proposed method was validated on two reference data sets, the statistics of which are shown in table 1 below. Namely MovieLens-1 and Lastfm-360K. MovieLens-1M is a benchmark dataset for the recommended approach in which users are associated with three sensitive attributes, including gender (2 categories), age (7 categories), and occupation (21 categories). The historical scores of the user items were divided into training and test sets at a ratio of 8:2. Lastfm-360K is a music recommendation dataset that contains user scores for artists that Lastfm collects from a music website. The play time is regarded as the scoring value. Because of the large scoring range, it is first preprocessed using logarithmic transformation and then normalized to an integer range of 0 to 5. Attributes associated with the user profile, including gender, age, and the like. Here, these two properties are considered as sensitive properties. For the age attribute, it is normalized to 7 categories. Similarly, the historical interaction scores were divided into training and test sets at a ratio of 8:2.
Table 1 model training dataset table
| Data set | MovieLens-1M | Lastfm-360K |
| # node | 9940 | 359347 |
| # edge | 1,000,209 | 17,559,530 |
| Sensitive attribute # | 3 | 2 |
| Binary sensitive attribute | √ | √ |
| Multi-class sensitive attributes | √ | √ |
Example 1
Introduction of comparative model:
FairGNN is used as the latest model of node classification tasks on the graph, and additional sensitive feature estimators are deployed to increase the amount of sensitive information. Fairness in node representation is achieved by alternately training the encoder and discriminator for resistance learning; unlike FairGNN, the classification task is replaced with a recommendation task and the research objective is extended to multi-classification sensitive properties.
FairGo as the latest model of fair recommendation tasks, through a combination of learning filters, transforms the original user and project embedding into a filtered embedding space based on a set of sensitive features. The model then applies countermeasure learning on the user-centric graph to achieve fair recommendations.
In the present embodiment, a drawing roll is usedNeural network as recommendation tower f R And a sensitivity attribute estimation tower f E Is a backbone of the (c). Selecting a simple linear classifier as the opponent-sensitive attribute countermeasure f A . The dimension of node embedding is set to 64 and the leaklylrelu is used as the activation function. The hidden layer dimension is 128 using a 2-layer graph convolution neural network. The Dropout rate is set to 0.5, and the weight parameters α, β, γ and λ are set to 0.01, 1, 0.005 on the equation of MovieLens-1M, and 0.01, 1, 0.003, 0.005 on Lastfm-360K. Using Adam optimizer, the initial learning rate is 0.001 and the weight decay is set to 1e-5. For FairGNN1, ICML 20192, and FairGo3 (models designed by other researchers), the original code in the GitHub repository was used to provide a strict and fair comparison between different models by adjusting the hyper-parameters of all models individually.
Introduction of evaluation index: the mean error (Root Mean Square Error, RMSE) root mean square error measures the deviation between the observed value and the true value, and is used to measure the error between the predicted score and the true score, and is calculated by the following formula (14):
the F1-score was calculated using the precision and recall, and the weighted harmonic mean between them was calculated by the following equation (15):
the population equality Δsp is measured by the absolute value error expected from the prediction scores among users of different sensitive attribute groups, and can be obtained by calculation of a formula (16):
the equality chance deltaeo measures the expected absolute value error of the recommended quality between users of different sensitive property groups, in this embodiment the recommended quality uses the absolute values of the predicted score and the true score:
population equality and opportunity equality in recommendations indicate that the prediction score and recommendation quality, respectively, are not affected by sensitive attributes, which are often multi-classified in real scenes, and the fairness metrics that extend the definition to multi-class sensitive attributes are: Δsp: measuring the expected absolute value error of the prediction scores among users of different sensitive attribute groups, wherein the smaller DeltaSP indicates that the prediction scores are less influenced by the sensitive attributes, and the expression is shown as a formula (18):
Δeo: the expected absolute value error of the recommended quality among users of different sensitive attribute groups is measured, the smaller the delta EO is, the smaller the recommended quality is influenced by the sensitive attribute, and the calculation method is as follows:
as shown in Table 2 below, the experimental results for the graph roll-up neural network as a backbone network are compared to a comparative performance table for the model and the Movielens-1M upper baseline method.
In table 2 above, RMSE was used to measure recommended performance and F1-score for measuring exposure of sensitive information in user representation, Δsp and Δeo were used to represent fairness performance of model, smaller values mean more fair recommenders, on movieens-1M dataset, fair perception recommendation method based on depth map neural network was always superior to other models in fairness, when sensitive attributes were age, gender and occupation, respectively, recommendation accuracy of 0.85%, 0.07% and 1.51% was sacrificed, and as for F1-score, model performed best in most cases, and at the same time, it was found that the more categories of sensitive attribute values were, the more difficult to guarantee fairness, and compared with other models, the method improved fairness of multi-classification sensitive attribute more significantly.
In summary, the invention introduces a graphic neural network and an countermeasure training technology to innovate pain points of a task recommended by a traditional platform developer, eliminates the influence of sensitive attributes on a recommendation result from a global angle through the graphic neural network technology, solves the problem that the fairness of a final recommendation result cannot be ensured by countermeasure training through one mutual information constraint and two fairness perception constraints, and effectively expands to multi-classification sensitive attributes. The invention has strong expansibility on the model, and can greatly ensure the recommendation fairness under the condition of almost not losing the recommendation accuracy.
While the preferred embodiments of the present application have been illustrated and described, the present application is not limited to the embodiments, and various equivalent modifications and substitutions can be made by one skilled in the art without departing from the spirit of the present application, and these modifications and substitutions are intended to be included in the scope of the present application as defined in the appended claims.
Claims (10)
1. A fair perception recommendation system based on a depth map neural network is characterized by comprising a recommendation tower f R Sensitive attribute estimation tower f E Sensitive attribute countermeasure f A Mutual information constraint module and fairness perception constraint module, the recommendation tower f R And a sensitivity attribute estimation tower f E All comprise an embedded layer, a graph convolution neural network model and a multi-layer perceptron model, wherein the sensitive attribute countermeasure f A Only the multi-layer perceptron model;
the recommendation tower f R For predicting user scoring of items: converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, and then applying a graph convolutional neural network model to encode neighbor features to generate item embedding and user embedding representations after information aggregation; finally, calculating the dot product of the project embedding and the user embedding Inputting the results into a multi-layer perceptron model to predict the scores of the user on the items;
the sensitive attribute estimation tower f E User sensitive properties for predictive property unknowns: converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to perform information aggregation to obtain node representation on a user-project interaction graph, and finally predicting sensitive attributes of users with unknown attributes through a multi-layer perceptron model;
the sensitive attribute challenge f A For recommending tower modules f R Deducing sensitive attributes from the generated embedded content;
the mutual information constraint module is used for relieving the problem of discrimination of the recommender caused by failure of antagonism depolarization convergence;
the fairness perception constraint module is used for improving recommendation accuracy and ensuring fairness of recommendation results of different sensitive attribute groups when different sensitive groups expect equal recommendation quality.
2. The depth map neural network-based fair perception recommendation system according to claim 1, wherein the sensitivity attribute countermeasure f A Estimating tower f using sensitive properties E The predicted sensitive attribute is used as a label training model and is based on the recommendation tower f R The generated user embedded content deduces the user sensitive attribute; recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair.
3. The depth map neural network-based fair perception recommendation system according to claim 1, wherein the mutual information constraint module is configured to reduce the number of recommendation towers f by minimizing the number of recommendation towers f R The generated estimation tower f for predicting the user scoring and sensitive attribute of the project E And the generated dependency relationship between the sensitive attributes is used for relieving the problem of discrimination of the recommender caused by failure of convergence of resistive depolarization.
4. The fair perception recommendation system based on the depth map neural network according to claim 1, wherein the fair perception constraint module improves recommendation accuracy by reducing a difference between a prediction score and recommendation quality between users of different sensitive attribute groups when different sensitive groups expect equal recommendation quality, so as to ensure fairness of recommendation results of different sensitive attribute groups.
5. A recommendation method of a fairness-aware recommendation system according to claims 1-4, comprising the steps of:
step S1, dividing the heterogeneous graph G into a plurality of subgraphs { G1, G2, & gt. GR }, wherein R represents the number of subgraphs and is taken as a recommendation tower f according to the user-project interaction scores R And a sensitivity attribute estimation tower f E Is input X of (a);
step S2, inputting the initial data X obtained in the step S1 into a recommendation tower f R Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to encode neighbor features, generating item embedding and user embedding representations after information aggregation, finally calculating dot products of the item embedding and the user embedding, and inputting calculation results into a multi-layer perceptron model to predict the scoring of the item by a user;
s3, inputting the initial data X obtained in the step S1 into a sensitive attribute estimation tower f E Converting the high-dimensional sparse feature vector into a low-dimensional dense feature vector through an embedding layer, then applying a graph convolutional neural network model to perform information aggregation to obtain node representation on a user-project interaction graph, and finally predicting sensitive attributes of users with unknown attributes through a multi-layer perceptron model;
Step S4, sensitive attribute countermeasure f A Estimating tower f by using sensitive attribute in step S3 E PredictionIs used as a label training model and is based on the recommendation tower f in step S2 R The generated embedded content deduces sensitive attribute to make recommendation tower f R And sensitive attribute countermeasure f A Constitutes a min-max game, sensitive attribute countermeasure f A According to recommendation tower f R The generated user-embedded content infers user-sensitive attributes, and the recommendation tower f R Learning user-embedded representation to make sensitive attribute challenge f A The sensitive attribute groups of the users cannot be distinguished so as to ensure the recommendation tower f R The generated user-embedded representation is fair;
s5, reducing the difference between the prediction scores and the recommendation quality among users of different sensitive attribute groups through a fair perception constraint module, and minimizing a recommendation tower f through a mutual information constraint module R The generated estimation tower f for predicting the user's scoring and sensitive attribute of the project E And when the trained parameters tend to be stable, the model training process is finished.
6. The recommendation method of a fair-aware recommendation system according to claim 5, wherein the recommendation tower f in step S2 R The method comprises the steps of predicting scoring of projects by a user, converting high-dimensional sparse feature vectors into low-dimensional dense feature vectors through an embedding layer, connecting the embedded vectors with the same length into initial node features, and inputting the initial node features into a graph neural network model:
X ′ ＝E(X) (1)
In the above, X ′ Is an embedding matrix of nodes, E ()' represents an embedding operator;
since the nodes and edges in heterogeneous graphs are of different types, their features are typically located in different spaces, thus in subgraph G r The graph convolutional neural network is applied to encode the neighbor features:
in the above-mentioned method, the step of,user representation representing a 1-hop neighborhood after propagation of the l-layer graph convolutional neural network model,/>Representing user u in subgraph G r Item set of upper interaction, d u And d i Representing the degree, σ (), non-linear activation function, W, of user u and item i, respectively (l) Is the parameter matrix of layer l, < >>Is an initial representation of user u, aggregating representations obtained from different subgraphs using a pooling operation:
in the above description, accsum ()' represents accumulation operation, the propagation process of item nodes is similar to that of user nodes, and user embedding h is calculated u And item embedding h i And input the results to a multi-layer perceptron model to predict final scoresThe prediction function may be written as:
where w and b are trainable parameters of the multi-layer perceptron model,is a recommendation tower f R Loss value of->Is a softmax function, +.>Is the calculation of dot product, recommending tower f R Loss value of +.>Expressed as:
7. the recommendation method of a fair sensing recommendation system according to claim 5, wherein the sensitivity attribute estimation tower f in step S3 E Since the graph neural network is more likely to expose sensitive information, it is used as an estimator to predict unknown sensitive attributes, and recommendation tower f R Similarly, information propagation using a graph convolution neural network to obtain node representations on a user-project interaction graph, and then using a multi-layer perceptron to obtain predicted user-sensitive attributes
In the above, N u Representing a set of neighbor nodes for user u, W (l) Is the parameter matrix of layer l, W and b are trainable parameters, whereby the sensitivity attribute estimation tower f E The loss function of (2) can be written as:
in the above equation, s is the true value of the attribute,is the predicted attribute value, V u Is a training user set, < >>Is a sensitive attribute estimation tower f E Is a loss of (2).
8. The recommendation method of a fair sense recommendation system according to claim 5, wherein said sensitivity attribute countermeasure f in step S4 A According to recommendation tower f R The generated embedded content infers sensitive attributes against f A Will recommend tower f R The generated user node representation h u As input and attempt to estimate the user's sensitive propertiesI.e. < ->Firstly, updating the parameters of the countermeasure to maximally improve the sensitivity attribute countermeasure f A The possibility of distinguishing sensitive information in the embedding and then fixing the sensitive attribute countermeasure f A Parameters to minimize differences between the generated representations of different sensitive property groups, which causes sensitive property countermeasure f A It is not possible to distinguish which sensitive group the user belongs to, forming a min-max game, expressed as:
in the above, θ ε 、θ R And theta A Respectively represent sensitive attribute estimation towers f E Recommendation tower f R And sensitive attribute antagonismDevice f A Is used for the control of the temperature of the liquid crystal display device,is a sensitive attribute countermeasure f A Loss value of->Representing user embedding h u When the parameter is theta ε 、θ R And theta A When using predicted sensitivity attribute k i Probability distribution represented by the user; k is the number of unique sensitivity attribute values, representing sensitivity attribute countermeasure f A The probability of the predicted outcome being.
9. The recommendation method of a fair sense recommendation system according to claim 5, wherein the difference between the prediction scores and the recommendation qualities among users of different sensitive attribute groups is reduced by a fair sense constraint module in step S5, the fair sense constraint module minimizes statistical parity ΔSP and opportunity equality ΔEO by adding constraint conditions in a loss function,and->Is a fair perceived constraint loss, < > >Narrowing the predictive score gap between different sensitive property groups and +.>Then focus is placed on the recommended quality gap between users focusing on different sets of sensitive attributes and their preference items (top scores);
in the above, if and only ifWhen (I)>Reaching a global minimum; in general, the above conditions are not the same as the actual situation, thus achieving +.>The global minimum of (2) would impair the accuracy of the recommendation for +.>The condition of the global minimum is sum +.>Equal to->In particular, when both items are equal to 0 at the same time, the accuracy and +.>Are all optimal; therefore, it is desirable to increase the recommendation accuracy as much as possible while ensuring that different sensitive groups expect the same recommendation quality; notably, constraints can be effectively applied to multiple classes of attributes, so the final objective function is expressed as:
in the above formula, α, β, γ, and λ control weights of mutual information constraint, antagonistic depolarization, statistical parity fairness constraint, and opportunity equality fairness constraint, respectively.
10. The recommendation method of a fair sense recommendation system according to claim 5, wherein the minimizing of the recommendation tower f by the mutual information constraint module in step S5 R The generated estimation tower f for predicting the user's scoring and sensitive attribute of the project E The generated dependency relationship between sensitive attributes is defined as follows:
in the above, D KL Indicating Kullback-Leibler divergence,representing sensitivity attribute estimation tower f E Predicted sensitivity attribute,/->Is a recommendation tower f R Predicted score,/->Is->And->Is a combination of (a) and (b) of (b)>Representing marginal distribution +.>And->Product of>Is the mutual information constraint loss, is a non-negative number, and the global minimum value 0 means +.>And->Is independent.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310595117.8A CN117216376A (en) | 2023-05-25 | 2023-05-25 | Fair perception recommendation system and recommendation method based on depth map neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310595117.8A CN117216376A (en) | 2023-05-25 | 2023-05-25 | Fair perception recommendation system and recommendation method based on depth map neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117216376A true CN117216376A (en) | 2023-12-12 |
Family
ID=89035861
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310595117.8A Pending CN117216376A (en) | 2023-05-25 | 2023-05-25 | Fair perception recommendation system and recommendation method based on depth map neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117216376A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117633478A (en) * | 2024-01-08 | 2024-03-01 | 暨南大学 | GNNs-based social media mining fairness backdoor assessment method and system |
-
2023
- 2023-05-25 CN CN202310595117.8A patent/CN117216376A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117633478A (en) * | 2024-01-08 | 2024-03-01 | 暨南大学 | GNNs-based social media mining fairness backdoor assessment method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
| Yang et al. | Friend or frenemy? Predicting signed ties in social networks | |
| JP6414363B2 (en) | Prediction system, method and program | |
| Gao et al. | Evaluating user reputation in online rating systems via an iterative group-based ranking method | |
| US20170032398A1 (en) | Method and apparatus for judging age brackets of users | |
| CN112084427A (en) | Interest point recommendation method based on graph neural network | |
| CN111582538B (en) | Community value prediction method and system based on graph neural network | |
| US20190012746A1 (en) | System and method for evaluating the true reach of social media influencers | |
| CN109471982B (en) | Web service recommendation method based on QoS (quality of service) perception of user and service clustering | |
| JP6311851B2 (en) | Co-clustering system, method and program | |
| CN110209946B (en) | Social and community-based product recommendation method, system and storage medium | |
| CN108053050A (en) | Clicking rate predictor method, device, computing device and storage medium | |
| CN107329887B (en) | Data processing method and device based on recommendation system | |
| CN117216376A (en) | Fair perception recommendation system and recommendation method based on depth map neural network | |
| CN113239266B (en) | Personalized recommendation method and system based on local matrix decomposition | |
| CN109190040A (en) | Personalized recommendation method and device based on coevolution | |
| Shi et al. | Long-term effects of user preference-oriented recommendation method on the evolution of online system | |
| CN115840853A (en) | Course recommendation system based on knowledge graph and attention network | |
| CN113360772A (en) | Interpretable recommendation model training method and device | |
| CN113673773A (en) | Learning path recommendation method fusing knowledge background and learning time prediction | |
| CN112463964A (en) | Text classification and model training method, device, equipment and storage medium | |
| CN117633371B (en) | Recommendation method, device and readable storage medium based on multi-attention mechanism | |
| Sridhar et al. | Extending Deep Neural Categorisation Models for Recommendations by Applying Gradient Based Learning | |
| WO2024012219A1 (en) | Related apparatus of model training method | |
| Shen | Social network mining and its applications |
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 |