CN114707427B - Personalized modeling method of graph neural network based on effective neighbor sampling maximization - Google Patents
Personalized modeling method of graph neural network based on effective neighbor sampling maximization Download PDFInfo
- Publication number
- CN114707427B CN114707427B CN202210571867.7A CN202210571867A CN114707427B CN 114707427 B CN114707427 B CN 114707427B CN 202210571867 A CN202210571867 A CN 202210571867A CN 114707427 B CN114707427 B CN 114707427B
- Authority
- CN
- China
- Prior art keywords
- user
- representation
- module
- fashion
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/12—Geometric CAD characterised by design entry means specially adapted for CAD, e.g. graphical user interfaces [GUI] specially adapted for CAD
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/12—Cloth
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Geometry (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Architecture (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Molecular Biology (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- User Interface Of Digital Computer (AREA)
- Image Analysis (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an effective neighbor sampling maximization-based personalized modeling method for a graph neural network, and belongs to the technical field of personalized fashion recommendation systems. The invention also provides an individualized model which comprises a composition module, a fashion single product representation module, a user representation module, an upper garment-lower garment compatibility module, a user individualized module and a recommendation module; the fashion single representation module is connected with the upper garment-lower garment compatibility module, the fashion single representation module and the user representation module are connected with the user personalization module, and the upper garment-lower garment compatibility module and the user personalization module are connected with the recommendation module. The invention provides the specific application of the graph neural network in the field of fashion recommendation, and the recommendation accuracy is improved; different sampling probabilities are adopted in different layers of the graph neural network, and the representation capability of the model on the target nodes and the robustness of the model are improved.
Description
Technical Field
The invention belongs to the technical field of personalized fashion recommendation systems, and particularly relates to a personalized and compatible modeling method of a graph neural network based on effective neighbor sampling maximization.
Background
In recent years, with the rapid development of fashion industry, the demand of consumers for fashion recommendations is increasing, thereby creating a strong demand for dress matches. In the aspect of garment compatibility modeling, current work can be divided into two broad categories: one is to model the compatibility of two fashion items (such as upper garment and lower garment) and the other is to model the compatibility of a complete set of clothing (multiple fashion items). In the aspect of fashion item modeling, most methods treat a pair of fashion items with compatible relationships as a metric learning problem by assuming that they are close to each other in a potential space; later, with the progress of related research and the sequential proposal of a data independent function and a data dependent function, researchers begin to use the methods to model the compatibility between fashion singles; in the aspect of suit modeling, most of the current research methods are to regard the suit as a sequence, model by using a Bi-LSTM or RNN method, and then start to apply the method to suit modeling along with the wide application of the graph neural network.
Based on the above, when garment compatibility modeling is performed in the fashion recommendation field, the current method mainly has the following three disadvantages: firstly, the association degree between different clothes is neglected, and only the visual features or text features of the clothes are used for feature representation, which is not enough to obtain accurate single-item feature representation; secondly, in a modeling scheme considering personalized factors, influence of a user relationship network on a user is ignored, so that accurate user characteristic representation cannot be obtained, and a recommendation system cannot make accurate clothing recommendation according to different users; thirdly, in the scheme of modeling by using the graph neural network, all neighbor information of the target node is generally aggregated, and the influence of the noise node on the representation of the target node is not considered.
Disclosure of Invention
The invention aims to provide an individualized modeling method of a graph neural network based on effective neighbor sampling maximization, so as to make up for the defects of the prior art.
In order to achieve the aim of the invention, the invention adopts the following specific technical scheme:
the personalized modeling method of the graph neural network based on the effective neighbor sampling maximization comprises the following steps:
s1: firstly, constructing a basic model: namely the personalization, compatibility model;
s2: according to an interaction diagram between the user and the fashion single product: constructing a user-fashion single picture, a user interaction picture, a fashion single-user picture and a fashion single-fashion single interaction picture;
s3: according to the fashion single item interaction diagram created in the step S2, creating a fashion single item feature representation matrix, constructing a diagram neural network model with a common L layer, sending the fashion single item feature representation matrix into the diagram neural network model, and obtaining the updated fashion single item feature representation matrix, wherein the updated fashion single item feature representation matrix comprises the feature representation of an upper garment and the feature representation of a lower garment;
s4: creating a characteristic representation matrix of the user according to the user interaction diagram created in the step S2, and sending the characteristic representation matrix of the user into the neural network model by using the neural network model created in the step S3 to obtain an updated user characteristic representation matrix, namely user characteristic representation;
s5: through the step of S3, acquiring the feature representation of the upper garment and the feature representation of the lower garment, sending the two feature representations into the upper garment-lower garment compatibility module, and calculating the compatibility scores of the upper garment and the lower garment;
s6: respectively obtaining the characteristic representation and the user characteristic representation of the lower garment through the steps of S3 and S4, and sending the two representations into the user personalization module to obtain the preference score of the user for the given lower garment;
s7: the compatibility scores of the upper garment and the lower garment obtained from the step of S5 and the preference score of the user for the given lower garment obtained from the step of S5 are fed to a recommending module to obtain the compatibility scores of the upper garment and the lower garment with the personalization factors fused.
Further, the basic model comprises a composition module, a fashion single representation module, a user representation module, an upper garment-lower garment compatibility module, a user personalization module and a recommendation module; the fashion single representation module is connected with the upper garment-lower garment compatibility module, the fashion single representation module and the user representation module are connected with the user personalization module, and the upper garment-lower garment compatibility module and the user personalization module are connected with the recommendation module.
Further, the patterning module: according to the relation between the user and the fashion single, constructing a user-fashion single graph, a user-user relation graph, a fashion single-user graph and a fashion single-fashion single graph;
the fashion singles representation module: the system comprises a graph neural network model, a fashion single product-user graph, a fashion single product-fashion single product graph, a user and single product characteristic representation matrix, a user and a single product characteristic representation matrix, wherein the user and the single product characteristic representation matrix are initialized randomly;
the user representation module: the system comprises a neural network model, a user-fashion single-item graph, a user-user relation graph, a user and single-item feature representation matrix, a user-fashion single-item graph and a user and single-item feature representation matrix, wherein the user-fashion single-item graph and the user-user relation graph are initialized randomly;
the upper garment-lower garment compatibility module: the upper garment feature representation and the lower garment feature representation obtained from the fashion single representation module are used as input information and are sent into the module, and the compatibility degree of the upper garment and the lower garment is calculated;
the user personalization module: the lower clothes feature representation obtained from the fashion single representation module and the user feature representation obtained from the user representation module are used as input information and are sent into the module, and the like degree of the user for the given lower clothes is calculated;
the recommendation module: and taking the compatibility degree of the upper garment and the lower garment obtained by the upper garment-lower garment compatibility module and the user like degree of the lower garment obtained by the user personalization module as input, calculating the compatibility score of the upper garment and the lower garment integrated with the personalization factors, and generating recommendation for the user according to the score.
Furthermore, the fashion single representation module comprises a first representation submodule, a second representation submodule and a third representation submodule, wherein the first representation submodule and the second representation submodule are connected with the third representation submodule;
a first representation submodule: the method is used for obtaining single-item feature representation from a single-item level by means of a graph neural network model and combining an attention mechanism from a fashion single-item interaction space;
the second representation submodule: the system is used for obtaining feature representation of the single item from a user level by means of a graph neural network model and combining an attention mechanism from a fashion single item-user interaction space;
a third representation submodule: and taking the output contents of the first representation submodule and the second representation submodule as input for fusing the single-item feature representation obtained from the single-item level and the single-item feature representation obtained from the user level to obtain the final feature representation of the target fashion single item.
Furthermore, the user representation module comprises a first representation submodule, a second representation submodule and a third representation submodule, wherein the first representation submodule and the second representation submodule are connected with the third representation submodule;
a first representation submodule: the system is used for obtaining user feature representation from a single product level by means of a graph neural network and combining an attention mechanism from a user-fashion single product interaction space;
the second representation submodule: the system is used for obtaining a feature representation of a user from a user interaction layer by means of a graph neural network and combining an attention mechanism from a user-user interaction relation space;
a third representation submodule: and taking the output contents of the first expression submodule and the second expression submodule as input, and fusing the user characteristic expression obtained from the single-item level and the user characteristic expression obtained from the user interaction level to obtain the final characteristic expression of the target user.
Further, in step S2, the specific method for creating the interaction graph includes:
s21: first, two symbolic representations are definedThe method comprises the steps of setting a user set containing N users;a garment set containing M fashion singles;
s22: creating a dictionary-style fashion item-fashion item interaction diagramWherein, in the step (A),andrepresenting the destination node of the fashion singles,and withAre respectively nodes,The neighbor node set of (2);
s23: creating a dictionary-style fashion singleton-user interaction diagramWherein, in the step (A),andthe destination node of the fashion item is represented,andare respectively nodes,The neighbor node set of (2);
s24: creating a dictionary-form user-user interaction graphWherein, in the step (A),andwhich represents the destination node of the user,and withAre respectively nodes,The neighbor node set of (2);
s25: creating a dictionary-style user-fashion singleton interaction diagramWherein, in the step (A),andwhich represents the destination node of the user,andare respectively nodes,Set of neighbor nodes.
Further, the step S3 specifically includes:
s31: constructing a static feature matrix of the fashion single item according to the created interactive graph;
Wherein, the fashion single item feature matrixEach row of (A) represents a fashion itemIs expressed asThis is also a single itemAn initial vector of a first layer in the graph neural network;
s32: from the level of fashion single products, by means of the neural network of the graph, the states of the neighbor nodes are aggregated, and the single products are matchedThe characteristic vector is updated, and the single product is obtained through state updatingNode representation at the kth level;
S33: from the user level, aggregating the states of the neighbor nodes by means of the graph neural network, and aligning the singlesThe feature vectors are updated, and the list is obtained through state updatingArticle (A)Node representation at the kth level;
S34: the items to be obtained from the fashion item levelTo representAnd the singleton obtained from the user planeTo representSplicing to obtain the final single productRepresent。
Further, the step S4 specifically includes:
Wherein, the user feature matrixEach row of (a) represents a userIs expressed asThis is also the userAn initial vector of a first layer in the graph neural network;
s42: from the level of fashion singles, by means of a graph neural network, aggregating the states of neighbor nodes and aiming at usersThe characteristic vector is updated, and the user can be obtained through state updatingNode representation at the kth level;
S43: from the user interaction relation level, by means of the graph neural network, the states of the neighbor nodes are aggregated, and the users are provided with the informationThe characteristic vector is updated, and the user can be obtained through state updatingNode representation at the kth level;
S44: users to be obtained from the fashion singles levelTo representAnd users obtained from the user interaction layerTo representSplicing to obtain the final userTo represent。
Furthermore, when the graph neural network aggregates information of neighboring nodes, a strategy of sampling effective nodes to participate in information aggregation in different network layers according to different probabilities p is adopted, and the detailed steps are as follows:
a1: the relation between the sampling probability and the number of layers of the neural network of the graph is as follows:whereinjAs the first of a neural networkjA layer of a material selected from the group consisting of,pthe sampling probability of the layer network;
a2: for the target nodeoAccording to the sampling probabilitypAt the first of the networkjThe layers perform sampling nodes to participate in information aggregation of target nodes, wherein the number of the sampling nodesnumAnd probabilitypThe relationship of (1) is:
wherein, in the step (A),dis a target nodeoThe degree of (a) is greater than (b),ceilis a function rounded up;
a3: calculating target nodesoAll neighbors and nodes ofoBefore removal ofnumThe nodes with large similarity participate in the target nodeoThe feature of (1) is aggregated.
Further, the step S5 specifically includes:
s51: the features of the jacket obtained through S3 are expressedWith lower clothesRepresentation of featuresSending the data to an upper garment-lower garment compatibility modeling module;
In the formula (I), the compound is shown in the specification,representing a dot product operation.
Further, the step S6 specifically includes:
s61: the lower clothes characteristics obtained through the step of S3 are expressedAnd the user feature representation obtained through the step of S4Sending the data to a user personalized modeling module;
Further, the step S7 specifically includes:
s71: obtaining the upper garment and lower garment compatibility scores according to the step of S52And the user' S like score for a given lower garment obtained according to the step of S52Garment compatibility with user personalized preferences computed:
Where pi is a non-negative trade-off parameter between (0, 1).
in the formula (I), the compound is shown in the specification,,coat with personalized factors fused respectivelyWith lower clothesLower clothesThe compatibility of (c).
The method can be applied to the fields of fashion recommendation, clothes matching and the like.
The invention has the following advantages and beneficial effects:
(1) the invention provides a specific application of a graph neural network in the field of fashion recommendation, and by mining interaction relations between users and fashion singles and by means of a graph representation learning method, feature representations of the users and the fashion singles can be fully learned, so that the recommendation accuracy is improved.
(2) According to the method, different sampling probabilities are adopted in different layers of the graph neural network, effective nodes are sampled as much as possible, and noise nodes are abandoned, so that the representation capability of the model on target nodes is improved; meanwhile, the problem of over-smoothness caused by excessive layers in the neural network of the graph is effectively solved by the probability sampling strategy, and the robustness of the model is improved.
Drawings
FIG. 1 is a block diagram of the overall modeling process of the present invention;
FIG. 2 is a flow chart of a modeling method of use of the present invention.
Detailed Description
The invention will be further elucidated with reference to the drawing.
Example (b):
a user personalization and compatibility method of a graph neural network based on effective neighbor sampling maximization comprises the following steps:
firstly, establishing a basic model (as shown in figures 1-2), wherein the model mainly comprises a composition module, a fashion single-item representation module, a user representation module, an upper garment-lower garment compatibility module, a user personalization module and a recommendation module, the composition module is connected with the fashion single-item representation module and the user representation module, the fashion single-item representation module is connected with the upper garment-lower garment compatibility module, the user representation module is connected with the user personalization module, and the upper garment-lower garment compatibility module and the user personalization module are connected with the recommendation module;
a patterning module: constructing a user-fashion single picture, a user-user interaction relation picture, a fashion single-user picture and a fashion single-fashion single picture so as to provide an interaction picture for a fashion single representation module and a user representation module;
fashion singleton representation module: and the method is used for inputting two interaction graphs of the fashion single item-user graph and the fashion single item-fashion single item graph as well as a randomly initialized user and fashion single item feature representation matrix into the constructed graph neural network model so as to obtain the updated fashion single item feature representation. The fashion singles representation module comprises: the first representation submodule, the second representation submodule and the third representation submodule are connected with the third representation submodule;
a first representation submodule: the method is used for obtaining the fashion single-item feature representation from the single-item level by means of a graph neural network and combining an attention mechanism from a fashion single-item interaction space;
a second representation sub-module: the system is used for obtaining feature representation of the single item from a user level by means of a graph neural network and combining an attention mechanism from a fashion single item-user interaction space;
a third representation submodule: and taking the output contents of the first representation submodule and the second representation submodule as input for fusing the single-item feature representation obtained from the single-item level and the single-item feature representation obtained from the user level to obtain the final feature representation of the target fashion single item.
A user representation module: and the method is used for inputting a user-fashion single-item diagram, two interaction diagrams of the user-user interaction diagram and a randomly initialized user and single-item feature representation matrix into the constructed diagram neural network model to obtain the updated user feature representation. The user representation module comprises: the first representation submodule, the second representation submodule and the third representation submodule are connected with the third representation submodule;
a first representation submodule: the system is used for obtaining user feature representation from a single-item level by means of a graph neural network and combining an attention mechanism from a user-fashion single-item interaction space;
the second representation submodule: the system comprises a neural network module, a user interaction space and an attention mechanism module, wherein the neural network module is used for acquiring a feature representation of a user from a user interaction layer by means of a graph neural network and fusing the attention mechanism;
a third representation submodule: and taking the output contents of the first representation submodule and the second representation submodule as input, and fusing the user feature representation obtained from the single-product level and the user feature representation obtained from the user interaction level to obtain the final feature representation of the target user.
Upper garment-lower garment compatibility module: the upper garment feature representation and the lower garment feature representation obtained from the fashion single representation module are used as input information and are sent into the module, and the compatibility degree of the upper garment and the lower garment is calculated;
a user personalization module: the lower clothing feature representation obtained from the fashion item representation module and the user feature representation obtained from the user representation module are used as input information and are sent into the module, and the user's liking degree of a given lower clothing is calculated;
a recommendation module: and taking the compatibility degree of the upper garment and the lower garment obtained by the upper garment-lower garment compatibility module and the user like degree of the lower garment obtained by the user personalization module as input, calculating the compatibility score of the upper garment and the lower garment integrated with the personalization factors, and generating recommendation for the user according to the score.
The concrete modeling method based on the model comprises the following steps:
s1: constructing a user-fashion single picture, a user-user interaction relation picture, a fashion single-user picture and a fashion single-fashion single picture, wherein the four constructed interaction pictures are undirected pictures, and the detailed interaction picture constructing mode is as follows:
s11: first, two symbolic representations are definedTo compriseNA user set of individual users;to compriseMA set of garments for each fashion item;
s12: creating a dictionary-style fashion item-fashion item interaction diagramWherein, in the step (A),andrepresenting the destination node of the fashion singles,andare respectively nodes,The neighbor node set of (2);
s13: creating a dictionary-style fashion singleton-user interaction diagramWherein, in the process,andrepresenting the destination node of the fashion singles,andare respectively nodes,The neighbor node set of (2);
s14: creating a dictionary-form user-user interaction graphWherein, in the process,andwhich represents the destination node of the user,andare respectively nodes,The neighbor node set of (2);
s15: creating a dictionary-style user-fashion singleton interaction diagramWherein, in the step (A),andwhich represents the destination node of the user,andare respectively nodes,The neighbor node set of (2);
in the four interactive graphs constructed above, two points need to be noted, wherein self-loops are not allowed to appear in one graph and are reflected to a dictionary formIn the interaction graph ofvalueNone of the sets withkeyThe same elements; two of itvalueLength of collectionI.e. the degree of the target node.
The embodiment adopts a method of different probabilities in the neural networkpThe strategy of sampling effective nodes in different network layers to participate in information aggregation comprises the following detailed steps:
a1: the relation between the sampling probability and the number of layers of the neural network of the graph is as follows:in whichjAs the first of a neural networkjA layer of a material selected from the group consisting of,pthe sampling probability of the network of the layer;
a2: for the target nodeoAccording to the sampling probabilitypAt the first of the networkjThe layers perform sampling nodes to participate in information aggregation of target nodes, wherein the number of the sampling nodesnumAnd probabilitypThe relationship of (1) is:
wherein, in the step (A),dis a target nodeoThe degree of (a) is greater than (b),ceilis a function of the rounding-up,ceilfunction and probabilitypEnsures the sampling node number of the target node;
A3: calculating target nodesoAll neighbors and nodes ofoBefore removal ofnumThe nodes with large similarity participate in the target nodeoThe operation of calculating the similarity specifically comprises the following steps:
wherein the content of the first and second substances,is a target nodeoIs determined by the feature vector of (a),Eis a target nodeoOf the neighbor of (a), each row of the matrix representing itThe feature vector of the one neighbor in the set,the function has the effect ofAnd feature matrixEEach line of (1) is dot-product, and finally, the dot product is calculated according to the calculatedMMatrix, before selection thereofnumThe index with large similarity is used for finding the corresponding node according to the index to participate in the target nodeoThe feature of (1) is aggregated.
S2: through the step S1, the four interaction maps are created, and the static interaction matrix of the fashion singleton is constructedUsing a Pythrch frameWhereinnum_embIs the total number of fashion items,emb_dimis the dimension of the feature vector. Wherein, the matrixEach row of (A) represents a fashion itemIs expressed asThis is also a single itemInitial vectors at the first layer in the neural network of the graph.
The information gathering process in the neural network of the graph is described in detail by taking the fashion single product as an example:
s21: gathering information at the fashion singles level:
wherein the content of the first and second substances,coat capable of showing clothes-clothes spaceIn the neural network of the figureThe representation vector of the layer is represented by,is a weighted aggregator that performs aggregation operations in the garment-to-garment space.Show coatAll of the neighbors of (a) are,coat capable of showing clothes spaceIs in the neural network of the graphThe representation vector of the layer is represented by,andrespectively representing weights and bias which can be learned in the neural network, wherein sigma is a sigmoid nonlinear activation function, and ^ indicates a splicing operation.
The detailed neighbor aggregation operation at the fashion singleton level is as follows:
an attention mechanism is used in information aggregation, in which,indicating the attention coefficients of the different network layers. The detailed attention coefficient is obtained as follows:
wherein the content of the first and second substances,and withGarment with separate indicationUser ofIn the neural network of the figureA representation of the layer.
S22: and (3) aggregating information at a user plane:
wherein the content of the first and second substances,show clothing-user figure middle upper garmentIn a neural network of the figureThe representation vector of the layer is represented by,is a weighted aggregator that performs aggregation operations in the garment-user space.Show coatAll of the neighbors of (a) are,clothing in presentation clothing-user spaceOf (2)In the first placeThe representation vector of the layer.Andrespectively, representing weights and biases that can be learned among the neural network of the garment-user space.
The detailed neighbor aggregation operation at the user plane is as follows:
in the formula, an attention mechanism and an attention coefficient are also adopted in the polymerization processThe obtaining manner is the same as S21.
S23: jacket to be obtained from fashion singlesRepresentAnd a jacket obtained from the user levelTo representSplicing to obtain the final coatTo representThe vector splicing method is as follows:
S3: via S1, the above-created four are utilizedAn interaction graph for constructing a static interaction matrix of usersUsing a Pythrch frameWhereinnum_embIs the total number of users and is,emb_dimis the dimension of the feature vector. Wherein, the matrixEach row of (a) represents a userIs expressed asThis is also the userInitial vectors at the first layer in the neural network of the graph. The detailed information gathering process of the neural network of the graph is as follows:
s31: gathering information from the fashion singles level:
wherein, the first and the second end of the pipe are connected with each other,representing users in a user-clothing spaceIn the networkThe representation vector of the layer is represented by,is a weighted aggregator employed in the present invention.Representing a userAll of the neighbors of (a) are,representing users in a user-clothing spaceOf (2)In the first placeThe representation vector of the layer is represented by,andrespectively representing weights and biases that can be learned in the neural network.
The detailed neighbor aggregation operation at the fashion singleton level is as follows:
in the formula, an attention mechanism and an attention coefficient are also adopted in the polymerization processThe obtaining manner is the same as S21.
S32: aggregating information from the user interaction level:
wherein the content of the first and second substances,representing that user u is in the neural network of the graph in the user interaction spaceThe representation vector of the layer is represented by,is the weighted aggregator employed in the present invention,representing a userAll of the neighbors of (a) are,representing users in a user-interaction spaceOf (2)In the neural network of the figureThe representation vector of the layer is represented by,andweights and biases among the neural networks representing the user interaction space, respectively, can be learned.
The detailed neighbor aggregation operation of the user interaction level is as follows:
in the formula, an attention mechanism is also adopted in the polymerization process, and the attention coefficient η is obtained in the same manner as in S21.
S33: user representation to be obtained from fashion singles levelWith user representation obtained from the user interaction layerSplicing to obtain the final user representation:
S4: the jacket obtained by the step of S2Representation of featuresAnd a lower garmentIs characterized byIs sent to
Clothes-lower clothes compatibility module and calculation upper garmentWith lower clothesThe specific calculation process of the compatibility score is as follows:
the simplest dot product method is used in the above equation to calculate the compatibility score of the upper garment and the lower garment.
S5: the lower clothes obtained through the step of S2Is characteristic ofAnd the user characteristic representation obtained by the step of S3Sending the user preference score to a user personalization module, and calculating the preference score of the user for the given lower garment, wherein the specific calculation process is as follows:
the simplest dot product method is used in the above equation to calculate the user's like score for a given under-garment.
S6: the fashion recommendation of the lower garment is mainly completed in the recommendation module, and the detailed steps are as follows:
s61: the upper garment is obtained through the steps of S4 and S5 respectivelyWith lower clothesIs given a compatibility scoreAnd the userFor clothesDegree of preference ofGarment compatibility with user personalized preferences fusedCan be expressed as follows:
s62: four tuples are constructed from the existing dataset:
wherein, in quadrupletOShowing a set of upper and lower garments. QuadrupletShown in a given coatIn the case of (2), the userMore like getting off the clothesRather than to。
S63: by definition of the BPR loss function, the objective function is as follows:
wherein, λ is a non-negative hyperparameter,representing the parameters that occur throughout the model.
S64: according to the calculated loss value, parameters in the neural network of the graph are updated by adopting a random gradient descent method until the model converges, and then the training work of the model is completed.
S65: meanwhile, in the whole experiment process, the data set is divided into a training set, a verification set and a test set, the proportion of the three data sets is 7:2:1, wherein the training set is used for training model parameters, the verification set is used for adjusting hyper-parameters in the model, and the test set is used for evaluating the performance of the model.
The personalized modeling method of the graph neural network based on the effective neighbor sampling maximization fully utilizes the user, the fashion single product and the interaction information between the user and the fashion single product, constructs the user, the fashion single product and the interaction graph between the user and the fashion single product, and accurately represents the user and the fashion single product by means of the graph neural network and by adopting a maximization neighbor sampling strategy, so that high-quality recommendation conforming to the actual situation is realized. The invention realizes more accurate entity feature representation, thereby completing the personalized fashion recommendation task.
Finally, although the present description refers to embodiments, not every embodiment contains only a single technical solution, and such description of the present description is for clarity reasons only, and those skilled in the art should make the description as a whole, and the technical solutions in the embodiments can be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims (9)
1. The personalized modeling method of the graph neural network based on the effective neighbor sampling maximization is characterized by comprising the following steps of:
s1: firstly, constructing a basic model; the basic model comprises a composition module, a fashion single item representation module, a user representation module, an upper garment-lower garment compatibility module, a user personalization module and a recommendation module; the fashion single-item representation module and the user representation module are connected with the user personalization module, and the upper garment-lower garment compatibility module and the user personalization module are connected with the recommendation module;
s2: according to an interaction diagram between the user and the fashion single product: constructing a user-fashion single picture, a user interaction picture, a fashion single-user picture and a fashion single-fashion single interaction picture;
s3: according to the fashion single item interaction diagram created in the step S2, creating a fashion single item feature representation matrix, constructing a diagram neural network model with a common L layer, sending the fashion single item feature representation matrix into the diagram neural network model, and obtaining the updated fashion single item feature representation matrix, wherein the updated fashion single item feature representation matrix comprises the feature representation of an upper garment and the feature representation of a lower garment;
s4: creating a characteristic representation matrix of the user according to the user interaction diagram created in the step S2, and sending the characteristic representation matrix of the user into the neural network model by using the neural network model created in the step S3 to obtain an updated user characteristic representation matrix, namely user characteristic representation;
in S3 and S4, the neural network model adopts a strategy of performing sampling on effective nodes in different network layers according to different probabilities p to participate in information aggregation;
s5: through the step of S3, acquiring the feature representation of the upper garment and the feature representation of the lower garment, sending the two feature representations into the upper garment-lower garment compatibility module, and calculating the compatibility scores of the upper garment and the lower garment;
s6: respectively obtaining the characteristic representation and the user characteristic representation of the lower garment through the steps of S3 and S4, and sending the two representations into the user personalization module to obtain the preference score of the user for the given lower garment;
s7: the compatibility scores of the upper garment and the lower garment obtained from the step of S5 and the preference score of the user for the given lower garment obtained from the step of S5 are fed to a recommending module to obtain the compatibility scores of the upper garment and the lower garment with the personalized factors fused.
2. The personalized modeling method of claim 1, wherein the composition module: according to the relation between the user and the fashion single, constructing a user-fashion single graph, a user-user relation graph, a fashion single-user graph and a fashion single-fashion single graph;
the fashion singles representation module: the system comprises a graph neural network model, a fashion single product-user graph, a fashion single product-fashion single product graph, a user and single product characteristic representation matrix, a user and a single product characteristic representation matrix, wherein the user and the single product characteristic representation matrix are initialized randomly;
the user representation module: the system comprises a neural network model, a user-fashion single-item graph, a user-user relation graph, a user and single-item feature representation matrix, a user-fashion single-item graph and a user and single-item feature representation matrix, wherein the user-fashion single-item graph and the user-user relation graph are initialized randomly;
the upper garment-lower garment compatibility module: the upper garment feature representation and the lower garment feature representation obtained from the fashion single representation module are used as input information and are sent into the module, and the compatibility degree of the upper garment and the lower garment is calculated;
the user personalization module: the lower clothes feature representation obtained from the fashion single representation module and the user feature representation obtained from the user representation module are used as input information and are sent into the module, and the like degree of the user for the given lower clothes is calculated;
the recommendation module: and taking the compatibility degree of the upper garment and the lower garment obtained by the upper garment-lower garment compatibility module and the user like degree of the lower garment obtained by the user personalization module as input, calculating the compatibility score of the upper garment and the lower garment integrated with the personalization factors, and generating recommendation for the user according to the score.
3. The personalized modeling method of claim 1, wherein the fashion singles representation module comprises a first representation submodule, a second representation submodule, and a third representation submodule, wherein the first representation submodule and the second representation submodule are connected to the third representation submodule;
a first representation submodule: the method is used for obtaining single-item feature representation from a single-item level by means of a graph neural network model and combining an attention mechanism from a fashion single-item interaction space;
the second representation submodule: the system is used for obtaining the feature representation of the single product from the user level by means of a graph neural network model and combining an attention mechanism from a fashion single product-user interaction space;
a third representation submodule: and taking the output contents of the first representation submodule and the second representation submodule as input for fusing the single-item feature representation obtained from the single-item level and the single-item feature representation obtained from the user level to obtain the final feature representation of the target fashion single item.
4. The personalized modeling method of claim 1, wherein the user representation module comprises a first representation sub-module, a second representation sub-module, and a third representation sub-module, wherein the first representation sub-module and the second representation sub-module are connected to the third representation sub-module;
a first representation submodule: the system is used for obtaining user feature representation from a single-item level by means of a graph neural network and combining an attention mechanism from a user-fashion single-item interaction space;
the second representation submodule: the system comprises a neural network and a user interaction interface, wherein the neural network is used for acquiring a characteristic representation of a user from a user interaction layer by means of a graph neural network and fusing an attention mechanism;
a third representation submodule: and taking the output contents of the first representation submodule and the second representation submodule as input, and fusing the user feature representation obtained from the single-product level and the user feature representation obtained from the user interaction level to obtain the final feature representation of the target user.
5. The personalized modeling method according to claim 1, wherein in the step S2, the specific method for creating the interaction graph is as follows:
s21: first, two symbolic representations are definedThe method comprises the steps of (1) setting a user set containing N users;a garment set containing M fashion singles;
s22: creating a dictionary-style fashion item-fashion item interaction diagramWhich, it does,andthe node of the target is represented by,andare respectively nodes,The neighbor node set of (2);
s23: creating a dictionary-style fashion item-user interaction diagramWherein, in the step (A),andthe node of the target is represented by,andare respectively nodes,The neighbor node set of (2);
s24: creating a dictionary-form user-user interaction graphWherein, in the step (A),andthe node of the target is represented by,and withAre respectively nodes,The neighbor node set of (2);
6. The personalized modeling method according to claim 1, wherein the step S3 specifically comprises:
s31: constructing a static feature matrix of the fashion single item according to the created interactive graph;
Wherein, the fashion single item feature matrixEach row of (1) representsFashion single articleIs expressed asThis is also a single itemAn initial vector of a first layer in the graph neural network;
s32: from the level of fashion single products, by means of the neural network of the graph, the states of the neighbor nodes are aggregated, and the single products are matchedThe characteristic vectors are updated, and the singles are obtained through state updatingNode representation at the kth level;
S33: from the user level, aggregating the states of the neighbor nodes by means of the graph neural network, and aligning the singlesThe characteristic vectors are updated, and the singles are obtained through state updatingNode representation at the kth level;
7. The personalized modeling method according to claim 1, wherein the step S4 specifically comprises:
Wherein, the user feature matrixEach row of (a) represents a userIs expressed asThis is also the userAn initial vector of a first layer in the graph neural network;
s42: from the fashion single-item level, the neighbors are aggregated by means of a graph neural networkThe state of the node, to the userThe characteristic vector is updated, and the user can be obtained through state updatingNode representation at the kth level;
S43: from the user interaction relation level, by means of the graph neural network, the states of the neighbor nodes are aggregated, and the users are provided with the informationThe characteristic vector is updated, and the user can be obtained through state updatingNode representation at the kth level;
8. The personalized modeling method according to claim 1, wherein the step S5 specifically comprises:
s51: the jacket characteristics obtained by S3 are expressedWith lower garment characterizationSending the data to an upper garment-lower garment compatibility modeling module;
In the formula (I), the compound is shown in the specification,representing a dot product operation;
the step S6 specifically includes:
s61: the lower clothes characteristics obtained by the step of S3 are expressedAnd the user feature representation obtained through the step of S4Sending the data to a user personalized modeling module;
The step S7 specifically includes:
s71: obtaining the upper garment and lower garment compatibility scores according to the step of S52And the user' S like score for a given lower garment obtained according to the step of S52Garment compatibility with user personalized preferences integrated in calculation:
Where π is a non-negative trade-off parameter between (0, 1);
9. The personalized modeling method according to claim 1 can be applied to the technical fields of fashion recommendation and clothes matching.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210571867.7A CN114707427B (en) | 2022-05-25 | 2022-05-25 | Personalized modeling method of graph neural network based on effective neighbor sampling maximization |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210571867.7A CN114707427B (en) | 2022-05-25 | 2022-05-25 | Personalized modeling method of graph neural network based on effective neighbor sampling maximization |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114707427A CN114707427A (en) | 2022-07-05 |
CN114707427B true CN114707427B (en) | 2022-09-06 |
Family
ID=82176860
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210571867.7A Active CN114707427B (en) | 2022-05-25 | 2022-05-25 | Personalized modeling method of graph neural network based on effective neighbor sampling maximization |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114707427B (en) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113378048A (en) * | 2021-06-10 | 2021-09-10 | 浙江工业大学 | Personalized recommendation method based on multi-view knowledge graph attention network |
CN113763300A (en) * | 2021-09-08 | 2021-12-07 | 湖北工业大学 | Multi-focus image fusion method combining depth context and convolution condition random field |
Family Cites Families (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875910A (en) * | 2018-05-23 | 2018-11-23 | 山东大学 | Garment coordination method, system and the storage medium extracted based on attention knowledge |
CN108960959B (en) * | 2018-05-23 | 2020-05-12 | 山东大学 | Multi-mode complementary clothing matching method, system and medium based on neural network |
US10824949B2 (en) * | 2018-09-27 | 2020-11-03 | Babylon Partners Limited | Method and system for extracting information from graphs |
CN110825963B (en) * | 2019-10-18 | 2022-03-25 | 山东大学 | Generation-based auxiliary template enhanced clothing matching scheme generation method and system |
CN110807477B (en) * | 2019-10-18 | 2022-06-07 | 山东大学 | Attention mechanism-based neural network garment matching scheme generation method and system |
CN111382309B (en) * | 2020-03-10 | 2023-04-18 | 深圳大学 | Short video recommendation method based on graph model, intelligent terminal and storage medium |
KR20210127464A (en) * | 2020-04-14 | 2021-10-22 | 주식회사 제이어스 | Coodinating and styling methods and systems through deep learning |
CN111881342A (en) * | 2020-06-23 | 2020-11-03 | 北京工业大学 | Recommendation method based on graph twin network |
CN113158088A (en) * | 2021-04-16 | 2021-07-23 | 桂林电子科技大学 | Position recommendation method based on graph neural network |
CN113592609B (en) * | 2021-08-17 | 2024-06-04 | 中山大学 | Personalized clothing collocation recommendation method and system utilizing time factors |
CN114444369A (en) * | 2021-09-29 | 2022-05-06 | 中南大学 | Fashion suit compatibility modeling method based on heterogeneous graph neural network |
CN114090902B (en) * | 2021-11-22 | 2022-09-09 | 中国人民解放军国防科技大学 | Social network influence prediction method and device based on heterogeneous network |
-
2022
- 2022-05-25 CN CN202210571867.7A patent/CN114707427B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113378048A (en) * | 2021-06-10 | 2021-09-10 | 浙江工业大学 | Personalized recommendation method based on multi-view knowledge graph attention network |
CN113763300A (en) * | 2021-09-08 | 2021-12-07 | 湖北工业大学 | Multi-focus image fusion method combining depth context and convolution condition random field |
Also Published As
Publication number | Publication date |
---|---|
CN114707427A (en) | 2022-07-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109299396A (en) | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model | |
CN112364976B (en) | User preference prediction method based on session recommendation system | |
CN111932336A (en) | Commodity list recommendation method based on long-term and short-term interest preference | |
CN107563841A (en) | A kind of commending system decomposed that scored based on user | |
CN110263257B (en) | Deep learning based recommendation method for processing multi-source heterogeneous data | |
CN111310063A (en) | Neural network-based article recommendation method for memory perception gated factorization machine | |
CN113590900A (en) | Sequence recommendation method fusing dynamic knowledge maps | |
CN110008377B (en) | Method for recommending movies by using user attributes | |
CN112115377A (en) | Graph neural network link prediction recommendation method based on social relationship | |
CN115438732A (en) | Cross-domain recommendation method for cold start user based on classification preference migration | |
CN115082147A (en) | Sequence recommendation method and device based on hypergraph neural network | |
CN113592609A (en) | Personalized clothing matching recommendation method and system using time factors | |
CN106897776A (en) | A kind of continuous type latent structure method based on nominal attribute | |
CN115221413B (en) | Sequence recommendation method and system based on interactive graph attention network | |
CN107391577A (en) | A kind of works label recommendation method and system based on expression vector | |
CN115269977A (en) | Recommendation method for fusion knowledge and collaborative information based on graph neural network | |
CN111414555A (en) | Personalized recommendation method based on collaborative filtering | |
Ahamed et al. | A recommender system based on deep neural network and matrix factorization for collaborative filtering | |
CN115374288A (en) | Recommendation method based on knowledge graph and attention mechanism | |
CN114997476A (en) | Commodity prediction method fusing commodity incidence relation | |
CN114742564A (en) | False reviewer group detection method fusing complex relationships | |
CN114707427B (en) | Personalized modeling method of graph neural network based on effective neighbor sampling maximization | |
CN116342228B (en) | Related recommendation method based on directed graph neural network | |
CN111429175B (en) | Method for predicting click conversion under sparse characteristic scene | |
CN110570226A (en) | scoring prediction method combining topic model and heterogeneous information network |
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 |