CN116645174B - Personalized recommendation method based on decoupling multi-behavior characterization learning - Google Patents
Personalized recommendation method based on decoupling multi-behavior characterization learning Download PDFInfo
- Publication number
- CN116645174B CN116645174B CN202310926715.9A CN202310926715A CN116645174B CN 116645174 B CN116645174 B CN 116645174B CN 202310926715 A CN202310926715 A CN 202310926715A CN 116645174 B CN116645174 B CN 116645174B
- Authority
- CN
- China
- Prior art keywords
- user
- behavior
- embedding
- representing
- embedded
- 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
- 238000000034 method Methods 0.000 title claims abstract description 29
- 238000012512 characterization method Methods 0.000 title claims abstract description 19
- 230000006399 behavior Effects 0.000 claims abstract description 158
- 238000006243 chemical reaction Methods 0.000 claims abstract description 12
- 239000013598 vector Substances 0.000 claims description 75
- 230000009471 action Effects 0.000 claims description 50
- 230000006870 function Effects 0.000 claims description 30
- 230000003993 interaction Effects 0.000 claims description 29
- 239000011159 matrix material Substances 0.000 claims description 22
- 238000013528 artificial neural network Methods 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 9
- 230000009466 transformation Effects 0.000 claims description 8
- 230000002776 aggregation Effects 0.000 claims description 6
- 238000004220 aggregation Methods 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 230000003542 behavioural effect Effects 0.000 claims description 3
- 230000002452 interceptive effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 2
- 238000004378 air conditioning Methods 0.000 claims 1
- 238000005192 partition Methods 0.000 claims 1
- 230000004931 aggregating effect Effects 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 abstract description 2
- 239000000284 extract Substances 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 3
- FGUUSXIOTUKUDN-IBGZPJMESA-N C1(=CC=CC=C1)N1C2=C(NC([C@H](C1)NC=1OC(=NN=1)C1=CC=CC=C1)=O)C=CC=C2 Chemical compound C1(=CC=CC=C1)N1C2=C(NC([C@H](C1)NC=1OC(=NN=1)C1=CC=CC=C1)=O)C=CC=C2 FGUUSXIOTUKUDN-IBGZPJMESA-N 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention relates to the technical field of recommendation systems and deep learning, in particular to a personalized recommendation method based on decoupling multi-behavior characterization learning. And learning user embedding and article embedding in each behavior by using the cascade behavior relation, wherein the characteristic information of the user and the article in each behavior is used as the meta-knowledge, and the meta-network model extracts the meta-knowledge to be the personalized characteristic conversion between the modeling behaviors of the user and the article. On this basis, the user embedding and the article embedding in each behavior are decoupled by using decoupling characterization learning. For decoupled user and item embedding, feature preference attention network modeling users have been designed to have different preferences for different features in different behaviors. And finally, aggregating the predictive scores of the users on the articles in all behaviors. Therefore, the recommendation system models more finely-grained user embedding and article embedding on the one hand, and more individuates the recommendation model on the other hand, so that the performance of the recommendation system is improved.
Description
Technical Field
The invention relates to the technical field of recommendation systems and deep learning, in particular to a personalized recommendation method based on decoupling multi-behavior characterization learning.
Background
The recommendation has made tremendous progress in the last decade as an effective tool for dealing with information overload problems in the information age. Model-based collaborative filtering learns representations of users and items to recommend based on their interactions, which has become a mainstream recommendation technique since it was successful in Netflix recommendation algorithm contests. Collaborative filtering models tend to build on a single behavior that is directly related to platform profit, such as purchasing behavior of an e-commerce platform or downloading behavior of an App platform. Because this behavior brings real economic or time costs to the user, the single behavior recommendation model can encounter serious data sparseness problems in practical applications. Fortunately, users typically take other types of actions (i.e., clicking and shopping carts) to obtain more information about the item to assist them in making the final decision. Other types of behavior have richer interactions than purchasing behavior. These types of behavior or auxiliary behavior may be used to help capture user preferences, alleviating data sparseness problems in recommendations.
In recent years, many behavioral recommendation models have made tremendous progress in utilizing cascading relationships of behaviors, such as the multi-behavioral recommendation model cascading graph convolutional network (MB-CGCN), which employs cascading graph convolutional neural network (GCN) modules to explicitly model cascading relationships between behaviors. In this model, the behavior features learned on the previous behavior by the structural model in GCN (LightGCN) are passed on to the next behavior in the chain after the feature transformation operation, and the embeddings learned from all behaviors are clustered together for final prediction.
Despite the tremendous advances made by cascading multi-behavior recommendation model techniques, two major limitations remain. On the one hand, user embedding and item embedding all features of each behavior are entangled together, creating sub-optimal problems. The user's preferences are different for different features in different behaviors, all of which are entangled together and cannot model the user's preferences. On the other hand, feature transitions between behaviors are shared. The feature transformation between each behavior should model the personalized feature transformation based on the user or item information, resulting in a more personalized recommendation model.
Disclosure of Invention
Aiming at the defects of the prior art, the personalized recommendation method based on decoupling multi-behavior characterization learning is developed, the preference degree of a user on an article in each behavior can be displayed, and personalized feature conversion by utilizing meta-learning network modeling is provided.
The technical scheme for solving the technical problems is as follows:
a personalized recommendation method based on decoupling multi-behavior characterization learning comprises the following steps:
a) Preprocessing in an electronic commerce data set to obtain a training set and a testing set;
b) The data set comprises M users, N articles and B interaction behaviors;
c) Constructing a personalized cascade GCN module to obtain user embedding and article embedding of B interactive behaviors, wherein the module consists of B GCNs, and personalized information is transmitted between the GCNs by a meta-learning network;
d) Decoupling the user embedding and the article embedding of the B interaction behaviors, and modeling the user with different feature preferences in different behaviors by using a feature preference attention network;
e) Calculating the overall preference of the user for the article under the B behaviors;
f) Model parameters are optimized by an Adam optimizer using model loss.
The method for personalized recommendation based on decoupling multi-behavior characterization learning comprises the following steps: for an e-commerce data set, the articles interacted by the users in the B behaviors are respectively divided according to a leave-one-out method, the last interaction of each user in the B behaviors is reserved as a test set, the rest data are used as a training set, and the sequence of the B interaction behaviors follows the sequence of the data set or the real behaviors.
The optimized scheme of the personalized recommendation method based on decoupling multi-behavior characterization learning comprises the following steps:
b-1) the data set comprises M users forming a user set U, N articles forming an article set I, B interaction behaviors forming B interaction matrixes
b-2) mapping user ID and item ID into one-hot vector matrix ID, respectively U And IDI, embedding matrix for user by using Xavier methodAnd an article embedding matrix->Initializing parameters, wherein d is the dimension of the embedded vector;
b-3) by the formulaAnd formula->Calculating initial embedding vector of user ID and article ID respectively>And->Wherein->For matrix ID U Is the u-th user's one-hot vector,/->For matrix ID I One-hot vector of the i-th item,/->d is the embedded vector dimension.
The optimized scheme of the personalized recommendation method based on decoupling multi-behavior characterization learning comprises the following steps:
c-1) by the formulaUser embedding layer 1 of graph convolution result of computing behavior b +.>Where l= {1,2,.,. The term..l-1 }, L is the number of layers of the graph convolution,representing the result of a convolution of an item representing the b-th action embedded in the first layer, +.>Representing a collection of items interacting with user u +.>Representing a set of users interacting with item i, |·| representing the size of the set;
c-2) passing through the formulaUser embedding +.>Where L is the number of layers of the graph convolution, +.>A user embedded layer 1 graph convolution result representing a b-th behavior;
c-3) passing through the formulaUser-embedded meta-knowledge of the b-th behavior calculated>Wherein Concat (&) is a splicing function, & lt/L + & gt>User-embedded +.>Item insert representing aggregation interacting with user u, +.>
c-4) passing through the formulaUser-embedded personalized feature transformation parameters for computing behavior b>In-> And->First-layer and second-layer neural network weight matrices respectively representing users in a b-th behavior,/->And->Respectively representing first-layer and second-layer neural network bias vectors of a user in the b-th action, wherein Relu (·) represents an activation function;
c-5) passing through the formulaCalculating to obtain the personalized feature conversion from the b-th behavior to the b+1-th behavior>Wherein->User-initiated embedding for the b+1th behavior;
c-6) passing through the formulaCalculating the result of the picture convolution of the item of action b embedded in layer 1 +.>Where l= {1,2,.,. The term..and L-1}, L is the number of layers of the graph convolution,/-)>User-embedded layer 1 graph convolution results representing behavior b,/>Representing a set of users interacting with item i, +.>Representing a collection of items interacted with by user u, |·| representing the size of the collection;
c-7) passing through the formulaItem embedding for calculating b-th behavior>Where L is the number of layers of the graph convolution, +.>A graph convolution result representing the embedding of the item of the b-th behavior into the first layer;
c-8) passing through the formulaItem-embedded meta-knowledge of action b is calculated>Wherein Concat (&) is a splicing function, & lt/L + & gt>Item insert representing action b, +.>User-embedded +.f representing aggregation interactions with item i>
c-9) passing through the formulaCalculating the article embedded personalized feature conversion parameter P of the b-th behavior i (b) In the formula->Anda first layer and a second layer neural network weight matrix representing the item in the b-th act respectively,and->Representing the first and second layer neural network bias vectors, relu (·) of the item in action b, respectively, represents the activation function.
c-10) passing through the formulaCalculating the personalized feature conversion from the b-th behavior to the b+1-th behavior object embedding +.>Wherein->Is the item initialization embedding of the b+1th action.
The optimized scheme of the personalized recommendation method based on decoupling multi-behavior characterization learning comprises the following steps:
d-1) is represented by the formulaDivision->User embedding +.>Wherein K represents the number of feature vector divisions, +.>User embedding representing behavior b>The corresponding kth feature vector, d representing the feature vector dimension;
d-2) by the formulaDivision->Item insert for obtaining action b>Wherein K represents the number of feature vector divisions, +.>Item insert representing action b +.>The corresponding kth feature vector, d representing the feature vector dimension;
d-3) by the formulaCalculating preference weight of user embedded kth block feature vector to article embedded kth block feature vector of the b-th action> Attention weight matrix representing the user in action b,/>Represents a bias vector, [ ·; carrying out]Representing a splicing operation, and Tanh (·) represents an activation function;
d-4) passing through the formulaSplicing to obtain preference weight of user embedded K block feature vector of b-th behavior to object embedded corresponding block feature vector>Wherein Concat (·) represents a splicing function;
d-5) is represented by the formulaCalculating the attention score of the characteristic preference of the user embedded K block characteristic vector of the b-th action on the article embedded corresponding block characteristic vector>Wherein Softmax (·) represents the normalization function;
d-6) is calculated by the formulaUser embedding of the b-th behavior calculated +.>New user embedding after fusion of feature preference attention scores>In the middle ofX represents->And->Is multiplied by each dimension of (a).
The optimized scheme of the personalized recommendation method based on decoupling multi-behavior characterization learning comprises the following steps:
e-1) by the formulaCalculating user embedding +.>Embedding +.>In the formula>And->Respectively representing the attention score of the characteristic preference of the user-embedded kth block characteristic vector in the b-th action on the article-embedded corresponding block characteristic vector, the user-embedded kth block characteristic vector in the b-th action and the article-embedded kth block characteristic vector in the b-th action, wherein T is a transposition operation;
e-2) passing through the formulaCalculating the total score of the user to the article in the B behaviors, whereinRepresenting user embedding +.>Embedding +.>Is a preference score of (c).
The personalized recommendation method based on decoupling multi-behavior characterization learning preferably comprises the following steps:
f-1) according to the formulaCalculating BPR loss, wherein ∈R is calculated>For training set, -> To observe user u and item i + Interaction set existing between->Sampling unobserved item i in the interaction set for user u - ，/>Representing a Sigmoid activation function;
f-2) according to the formulaUser embedding +.>Is>In the middle of
dCov (·,) represents the distance covariance between the two matrices, dVar (·) represents the distance variance of the matrices;
f-3) according to the formulaCalculating to obtain user embedded decoupling total loss L in B behaviors u ；
f-4) according to the formulaArticle inlay for calculating b-th behaviorEnter->Is>In the middle of
dCov (·,) represents the distance covariance between the two matrices, dVar (·) represents the distance variance of the matrices;
f-5) according to the formulaCalculating to obtain the total loss L of the embedded decoupling of the article in the B behaviors i′ ；
f-6) according to the formulaCalculating to obtain a model total loss function>Wherein alpha and beta are weights of different super-parameter control modules, and the super-parameter control modules comprise user embedding and article embedding decoupling total loss L u +L i L2 regularization L of model parameter θ θ ，/>L2 regularization of the model parameters θ.
The effects provided in the summary of the invention are merely effects of embodiments, not all effects of the invention, and the above technical solution has the following advantages or beneficial effects:
the invention provides a personalized recommendation method based on decoupling multi-behavior characterization learning, which consists of a decoupling characterization learning module and a meta learning module. Firstly, learning user embedding and article embedding in each behavior by utilizing a cascade behavior relation, taking characteristic information of the user and the article in each behavior as meta-knowledge, and extracting the meta-knowledge through a meta-network to model personalized characteristic conversion between the behaviors of the user and the article. On this basis, the user embedding and the article embedding in each behavior are decoupled by using decoupling characterization learning. For decoupled user and item embedding, feature preference attention network modeling users have been designed to have different preferences for different features in different behaviors. And finally, aggregating the predictive scores of the users on the articles in all behaviors. Therefore, the recommendation system models more finely-grained user embedding and article embedding on the one hand, and more individuates the recommendation model on the other hand, so that the performance of the recommendation system is improved.
Drawings
The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate the invention and together with the embodiments of the invention, serve to explain the invention.
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram of a decoupled multi-behavioral representation study of the present invention.
Detailed Description
In order to clearly illustrate the technical features of the present solution, the present invention will be described in detail below with reference to fig. 1 and 2 of the accompanying drawings.
As shown in fig. 1 to 2, a personalized recommendation method based on decoupling behavior characterization learning is characterized by comprising the following steps:
a) Preprocessing in an electronic commerce data set to obtain a training set and a testing set;
b) The data set comprises M users, N articles and B interaction behaviors, wherein the interaction behaviors comprise browsing, shopping cart adding, purchasing and the like;
c) Constructing a personalized cascade GCN module to obtain user embedding and article embedding of B interactive behaviors, wherein the module consists of B GCNs, and personalized information is transmitted between the GCNs by a meta-learning network;
d) Decoupling the user embedding and the article embedding of the B interaction behaviors, and modeling the user with different feature preferences in different behaviors by using a feature preference attention network;
e) Calculating the overall preference of the user for the article under the B behaviors;
f) Model parameters are optimized by an Adam optimizer using model loss.
And learning user embedding and article embedding in each behavior by using the cascade behavior relation, wherein the characteristic information of the user and the article in each behavior is used as the meta-knowledge, and the meta-network model extracts the meta-knowledge to be the personalized characteristic conversion between the modeling behaviors of the user and the article. On this basis, the user embedding and the article embedding in each behavior are decoupled by using decoupling characterization learning. For decoupled user and item embedding, feature preference attention network modeling users have been designed to have different preferences for different features in different behaviors. And finally, aggregating the predictive scores of the users on the articles in all behaviors. Therefore, the recommendation system models more finely-grained user embedding and article embedding on the one hand, and more individuates the recommendation model on the other hand, so that the performance of the recommendation system is improved.
In this embodiment, step a) includes the steps of: for an e-commerce data set, the articles interacted by the users in the B behaviors are respectively divided according to a leave-one-out method, the last interaction of each user in the B behaviors is reserved as a test set, the rest data are used as a training set, and the sequence of the B interaction behaviors follows the sequence of the data set or the real behaviors.
In this embodiment, step b) includes the steps of:
b-1) the data set comprises M users forming a user set U, N articles forming an article set I, B interaction behaviors forming B interaction matrixes
b-2) mapping user ID and item ID into one-hot vector matrix ID, respectively U And IDI, embedding matrix for user by using Xavier methodAnd an article embedding matrix->Carrying out parameter initialization, wherein d is an embedded vector dimension, one-hot is the simplest text feature representation method which is quite common, and on word feature representation, the essence of the text feature representation directly takes the subscript of a word in a word set as the representation of word change, and the Xavier method is a neural network initialization method;
b-3) by the formulaAnd formula->Calculating initial embedding vector of user ID and article ID respectively>And->Wherein->For matrix ID U Is the u-th user's one-hot vector,/->For matrix ID I One-hot vector of the i-th item,/->d is the embedding vector dimension, P is the user embedding matrix, and Q is the item embedding matrix.
In this embodiment, step c) includes the steps of:
c-1) by the formulaUser embedding layer 1 of graph convolution result of computing behavior b +.>Where l= {1,2,.,. The term..l-1 }, L is the number of layers of the graph convolution,representing the result of a convolution of an item representing the b-th action embedded in the first layer, +.>Representing a collection of items interacting with user u +.>Representing a set of users interacting with item i, |·| representing the size of the set;
c-2) passing through the formulaUser embedding +.>Where L is the number of layers of the graph convolution, +.>A user embedded layer 1 graph convolution result representing a b-th behavior;
c-3) passing through the formulaUser-embedded meta-knowledge of the b-th behavior calculated>Wherein Concat (&) is a splicing function, & lt/L + & gt>User-embedded +.>Item insert representing aggregation interacting with user u, +.>
c-4) passing through the formulaUser-embedded personalized feature transformation parameters for computing behavior b>In-> And->First-layer and second-layer neural network weight matrices respectively representing users in a b-th behavior,/->And->Respectively representing first-layer and second-layer neural network bias vectors of a user in the b-th action, wherein Relu (·) represents an activation function;
c-5) passing through the formulaCalculating to obtain the personalized feature conversion from the b-th behavior to the b+1-th behavior>Wherein->Is the bUser-initiated embedding of +1 behaviors;
c-6) passing through the formulaCalculating the result of the picture convolution of the item of action b embedded in layer 1 +.>Where l= {1,2,.,. The term..and L-1}, L is the number of layers of the graph convolution,/-)>User-embedded layer 1 graph convolution results representing behavior b,/>Representing a set of users interacting with item i, +.>Representing a collection of items interacted with by user u, |·| representing the size of the collection;
c-7) passing through the formulaItem embedding for calculating b-th behavior>Where L is the number of layers of the graph convolution, +.>A graph convolution result representing the embedding of the item of the b-th behavior into the first layer;
c-8) passing through the formulaItem-embedded meta-knowledge of action b is calculated>In the middle ofConcat (&) is a splicing function, & lt + & gt>Item insert representing action b, +.>User-embedded +.f representing aggregation interactions with item i>
c-9) passing through the formulaCalculating the article embedded personalized feature conversion parameter P of the b-th behavior i (b) In the formula->Andrepresenting a first and second layer neural network weight matrix representing an item in action b,and->Representing the first and second layer neural network bias vectors, relu (·) of the item in action b, respectively, represents the activation function.
c-10) passing through the formulaCalculating the personalized feature conversion from the b-th behavior to the b+1-th behavior object embedding +.>Wherein->Item initialization embedding, P, being the b+1th behavior i (b) Item-embedded personalized feature transformation parameter of action b,/->The item representing the b-th action is embedded.
In this embodiment, step d) includes the steps of:
d-1) is represented by the formulaDivision->User embedding +.>Wherein K represents the number of feature vector divisions, +.>User embedding representing behavior b>The corresponding kth feature vector, d representing the feature vector dimension;
d-2) by the formulaDivision->Item insert for obtaining action b>Wherein K represents the number of feature vector divisions, +.>Item insert representing action b +.>The corresponding kth feature vector, d representing the feature vector dimension;
d-3) by the formulaCalculating preference weight of user embedded kth block feature vector to article embedded kth block feature vector of the b-th action>In the middle ofAttention weight matrix representing the user in action b,/>Representing the attention offset vector of the user in action b,>representing the output weight matrix of the user in the b-th action, [; carrying out]Representing a splicing operation, and Tanh (·) represents an activation function;
d-4) passing through the formulaSplicing to obtain preference weight of user embedded K block feature vector of b-th behavior to object embedded corresponding block feature vector>Wherein Concat (·) represents a splicing function;
d-5) is represented by the formulaUser embedded K block characteristic direction for calculating b-th behaviorQuantity +.A characteristic preference attention score for embedding an item into a corresponding block feature vector>Wherein Softmax (·) represents the normalization function;
d-6) is calculated by the formulaUser embedding of the b-th behavior calculated +.>New user embedding after fusion of feature preference attention scores>In the middle ofX represents->And->Is multiplied by each dimension of (a).
In this embodiment, step e) includes the steps of:
e-1) by the formulaCalculating user embedding +.>Embedding +.>Preference score of (2), wherein->And->Respectively representing the attention score of the characteristic preference of the user-embedded kth block characteristic vector in the b-th action on the article-embedded corresponding block characteristic vector, the user-embedded kth block characteristic vector in the b-th action and the article-embedded kth block characteristic vector in the b-th action, wherein T is a transposition operation;
e-2) passing through the formulaCalculating the total score of the user to the article in the B behaviors, whereinRepresenting user embedding +.>Embedding +.>Is a preference score of (c).
In this embodiment, step f) includes the steps of:
f-1) according to the formulaCalculating BPR loss, wherein ∈R is calculated>For training set, -> To observe user u and item i + Interaction set existing between->For user u to meetSample unobserved items i in the mutual collection - ，/>Representing a Sigmoid activation function;
f-2) according to the formulaUser embedding +.>Is>In the middle of
dCov (·,) represents the distance covariance between the two matrices, dVar (·) represents the distance variance of the matrices;
f-3) according to the formulaCalculating to obtain user embedded decoupling total loss L in B behaviors u ；
f-4) according to the formulaItem embedding for calculating b-th behavior>Is>
In the middle ofdCov (·,) represents the distance covariance between the two matrices, dVar (·) represents the momentThe distance variance of the array;
f-5) according to the formulaCalculating to obtain the total loss L of the embedded decoupling of the article in the B behaviors i′ ；
f-6) according to the formulaCalculating to obtain a model total loss function>Wherein alpha and beta are weights of different super-parameter control modules, and the super-parameter control modules comprise user embedding and article embedding decoupling total loss L u +L i L2 regularization L of model parameter θ θ ，/>L2 regularization of the model parameters θ.
In order to more intuitively explain how the method is applied to practical problems, the scheme is more thoroughly and specifically described below in combination with practical application of the user to the use behavior of the article in a certain field.
The following is a further description of an electronic commerce platform, a skyhook mall:
data set description: the kitten data set includes 15449 users and 11953 items, and the behavior information includes 873954 browsing behaviors, 195476 shopping cart adding behaviors, and 104329 purchasing behaviors. The cascading order of behaviors is browse- > join shopping cart- > purchase.
And c) constructing a personalized cascade GCN module according to the step c) to obtain user embedding and article embedding of the data set browsing behavior, the shopping cart adding behavior and the purchasing behavior of the skyhook mall. Decoupling user embedding and article embedding of the heaven mall data set browsing behaviors, the shopping cart adding behaviors and the purchasing behaviors according to the step d), and modeling that the user has different characteristic preferences in different behaviors by utilizing a characteristic preference attention network; calculating the overall preference of the user for the article under the actions of browsing, joining shopping cart and purchasing according to the step e); finally, calculating the total loss of the model according to the step f) and optimizing the model parameters through an Adam optimizer.
While the foregoing description of the embodiments of the present invention has been presented with reference to the drawings, it is not intended to limit the scope of the invention, but rather, it is apparent that various modifications or variations can be made by those skilled in the art without the need for inventive work on the basis of the technical solutions of the present invention.
Claims (3)
1. A personalized recommendation method based on decoupling multi-behavior characterization learning is characterized by comprising the following steps:
(a) Preprocessing in an electronic commerce data set to obtain a training set and a testing set;
(b) The data set comprises M users, N articles and B interaction behaviors;
(c) Constructing a personalized cascade GCN module to obtain user embedding and article embedding of B interactive behaviors, wherein the module consists of B GCNs, and personalized information is transmitted between the GCNs by a meta-learning network;
the step (c) is specifically as follows:
(c-1) passing through the formulaUser embedding layer 1 of graph convolution result of computing behavior b +.>Where l= {1,2,.,. The term..and L-1}, L is the number of layers of the graph convolution,/-)>Graph convolution result of item embedded in layer 1 representing behavior b, +.>Representing a collection of items interacting with user u +.>Representing a set of users interacting with item i, |·| representing the size of the set;
(c-2) passing through the formulaUser embedding +.>Where L is the number of layers of the graph convolution, +.>A user embedded layer 1 graph convolution result representing a b-th behavior;
(c-3) passing through the formulaUser-embedded meta-knowledge of the b-th behavior calculated>Wherein Concat (&) is a splicing function, & lt/L + & gt>User embedding showing behavior b, +.>Item insert representing action b, +.>Representing the aggregation of the item embeddings interacting with user u,
(c-4) passing through a male memberA kind of electronic device with high-pressure air-conditioning systemUser-embedded personalized feature transformation parameters for computing behavior b>In->And->The first and second layer neural network weight matrices representing the user in the b-th behavior respectively,and->Respectively representing first-layer and second-layer neural network bias vectors of a user in the b-th action, wherein Relu (·) represents an activation function; (c-5) by the formula>Calculating to obtain the personalized feature conversion from the b-th behavior to the b+1-th behavior>Wherein->User-initiated embedding for the b+1th behavior;
(c-6) passing through the formulaCalculating the result of the picture convolution of the item of action b embedded in layer 1 +.>Where l= {1,2,.,. The term..and L-1}, L is the number of layers of the graph convolution,/-)>User-embedded layer 1 graph convolution results representing behavior b,/>Representing a set of users interacting with item i, +.>Representing a collection of items interacted with by user u, |·| representing the size of the collection;
(c-7) passing through the formulaItem embedding for calculating b-th behavior>Where L is the number of layers of the graph convolution, +.>A graph convolution result representing the embedding of the item of the b-th behavior into the first layer;
(c-8) passing through the formulaItem-embedded meta-knowledge of action b is calculated>Wherein Concat (&) is a splicing function, & lt/L + & gt>The article representing the b-th action is embedded,user-embedded +.f representing aggregation interactions with item i>
(c-9) passing through the formulaCalculating the item-embedded personalized feature transformation parameter of the b-th behavior>In->Anda first layer and a second layer neural network weight matrix representing the item in the b-th act respectively,and->The first layer neural network bias vector and the second layer neural network bias vector respectively representing the articles in the b-th action, wherein Relu (·) represents an activation function;
(c-10) passing through the formulaCalculating the personalized feature conversion from the b-th behavior to the b+1-th behavior object embedding +.>Wherein->Article initialization embedding for the (b+1) th behavior;
(d) Decoupling the user embedding and the article embedding of the B interaction behaviors, and modeling the user with different feature preferences in different behaviors by using a feature preference attention network;
the step (d) is specifically as follows:
(d-1) passing through the formulaDivision->User embedding +.>Wherein K represents the number of feature vector divisions, +.>User embedding representing behavior b>The corresponding kth feature vector, d representing the feature vector dimension;
(d-2) passing through the formulaDivision->Item insert for obtaining action b>Where K represents the number of feature vector partitions,/>item insert representing action b +.>The corresponding kth feature vector, d representing the feature vector dimension;
(d-3) passing through the formulaCalculating the preference weight of the user embedded kth block feature vector of the b-th action to the article embedded kth block feature vectorIn->Representing the attention weight matrix of the user in the b-th behavior,representing the attention offset vector of the user in action b,>representing the output weight matrix of the user in the b-th action, [; carrying out]Representing a splicing operation, and Tanh (·) represents an activation function;
(d-4) passing through the formulaSplicing to obtain preference weight of user embedded K block feature vector of b-th behavior to object embedded corresponding block feature vector>Wherein Comcat (·) represents a stitching function;
(d-5) passing through the formulaCalculating the attention score of the characteristic preference of the user embedded K block characteristic vector of the b-th action on the article embedded corresponding block characteristic vector>Wherein Softmax (·) represents the normalized exponential function;
(d-6) passing through the formula
User embedding of the b-th behavior calculated +.>New user embedding after feature preference attention score fusionIn->X represents->And->Performing a multiplication operation on each dimension of (1);
(e) Calculating the overall preference of the user for the article under the B behaviors;
firstly, obtaining preference scores of user embedding on articles in the B-th behavior, and then obtaining total scores of the users on the articles in the B-th behavior;
the step (e) is specifically as follows:
(e-1) passing through the formulaCalculating user embedding +.>Embedding +.>Preference score of (2), wherein->And->Respectively representing the attention score of the characteristic preference of the user-embedded kth block characteristic vector in the b-th action on the article-embedded corresponding block characteristic vector, the user-embedded kth block characteristic vector in the b-th action and the article-embedded kth block characteristic vector in the b-th action, wherein T is a transposition operation;
(e-2) passing through the formulaCalculating the total score of the user to the article in the B behaviors, whereinRepresenting user embedding +.>Embedding +.>Is a preference score of (2);
(f) Optimizing model parameters by using model loss through an Adam optimizer;
firstly, BPR losses are obtained, then a user embedded decoupling loss function of the B-th behavior is obtained, further a user embedded decoupling total loss in the B-th behavior, a decoupling loss function of the B-th behavior in which articles are embedded and a decoupling total loss of the B-th behavior in which articles are embedded are obtained, and finally a model total loss function is obtained according to the loss result;
the step (f) is specifically as follows:
(f-1) according to the formulaCalculating BPR loss, wherein ∈R is calculated>For training set, -> To observe user u and item i + Interaction set existing between->Sampling unobserved item i in the interaction set for user u - ，/>Representing a Sigmoid activation function;
(f-2) according to the formulaUser embedding +.>Is>In the middle of
dCov (·, ·) represents the distance covariance between the two matrices, dVar (·, ·) represents the distance variance of the matrices;
(f-3) according to the formulaCalculating to obtain user embedded decoupling total loss L in B behaviors u ；
(f-4) according to the formulaItem embedding for calculating b-th behavior>Is>In the middle ofdCov (·,) represents the distance covariance between the two matrices, dVar (·) represents the distance variance of the matrices;
(f-5) according to the formulaCalculating to obtain the total loss L of the embedded decoupling of the article in the B behaviors i ；
(f-6) according to formula l=l BPR +α(L u +L i )+βL θ Calculating to obtain a model total loss function L, wherein alpha and beta are weights of different super-parameter control modules, and the super-parameter control modules comprise user embedding and article embedding decoupling total loss L u +L i L2 regularization L of model parameter θ θ ，L2 regularization of the model parameters θ.
2. The personalized recommendation method based on decoupled behavioral characteristics learning of claim 1, wherein step (a) comprises the steps of: for an e-commerce data set, the articles interacted by the users in the B behaviors are respectively divided according to a leave-one-out method, the last interaction of each user in the B behaviors is reserved as a test set, the rest data are used as a training set, and the sequence of the B interaction behaviors follows the sequence of the data set or the real behaviors.
3. The personalized recommendation method based on decoupled behavioral characteristics learning of claim 1, wherein step (b) comprises the steps of:
(B-1) the dataset comprising M user-formed user sets U, N items-formed item sets I, B interaction behaviors-formed B interaction matrices { Y ] 1 ，Y 2 ，…，Y B }；
(b-2) mapping the user ID and the item ID into one-hot vector matrix IDs, respectively U And ID I Embedding matrix to user by using Xavier methodAnd an article embedding matrix->Initializing parameters, wherein d is the dimension of the embedded vector;
(b-3) passing through the formulaAnd formula->Calculating initial embedding vector of user ID and article ID respectively>And->Wherein->For matrix ID U Is the u-th user's one-hot vector,/->For matrix ID I One-hot vector of the i-th item,/->d is the embedded vector dimension.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310926715.9A CN116645174B (en) | 2023-07-27 | 2023-07-27 | Personalized recommendation method based on decoupling multi-behavior characterization learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310926715.9A CN116645174B (en) | 2023-07-27 | 2023-07-27 | Personalized recommendation method based on decoupling multi-behavior characterization learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116645174A CN116645174A (en) | 2023-08-25 |
| CN116645174B true CN116645174B (en) | 2023-10-17 |
Family
ID=87643757
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310926715.9A Active CN116645174B (en) | 2023-07-27 | 2023-07-27 | Personalized recommendation method based on decoupling multi-behavior characterization learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116645174B (en) |
Citations (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20220042598A (en) * | 2020-09-28 | 2022-04-05 | (주)브레인콜라 | Personalized recommendation method and system based on future interaction prediction |
| CN114385804A (en) * | 2021-11-12 | 2022-04-22 | 中山大学 | Comment recommendation method of heterogeneous graph attention neural network based on meta-learning |
| CN114519147A (en) * | 2022-02-17 | 2022-05-20 | 云南大学 | Social recommendation method based on GCN enhanced influence diffusion |
| CN114693397A (en) * | 2022-03-16 | 2022-07-01 | 电子科技大学 | Multi-view multi-modal commodity recommendation method based on attention neural network |
| CN114861065A (en) * | 2022-05-31 | 2022-08-05 | 山东省人工智能研究院 | Multi-behavior based personalized recommendation method for cascaded residual error graph convolution network |
| CN115114526A (en) * | 2022-06-30 | 2022-09-27 | 南京邮电大学 | Weighted graph convolution network score prediction recommendation method for multi-behavior enhancement information |
| CN115187343A (en) * | 2022-07-20 | 2022-10-14 | 山东省人工智能研究院 | Multi-behavior recommendation method based on attention map convolution neural network |
| WO2022222037A1 (en) * | 2021-04-20 | 2022-10-27 | 中国科学院深圳先进技术研究院 | Interpretable recommendation method based on graph neural network inference |
| CN115357804A (en) * | 2022-07-19 | 2022-11-18 | 大连民族大学 | Two-stage collaborative filtering multi-behavior recommendation method based on graph convolution network |
| CN115687772A (en) * | 2022-11-09 | 2023-02-03 | 安徽大学 | Sequence recommendation method based on sequence dependence enhanced self-attention network |
| CN115878904A (en) * | 2023-02-22 | 2023-03-31 | 深圳昊通技术有限公司 | Intellectual property personalized recommendation method, system and medium based on deep learning |
| CN115905687A (en) * | 2022-11-01 | 2023-04-04 | 山东建筑大学 | Cold start-oriented recommendation system and method based on meta-learning graph neural network |
| CN116010718A (en) * | 2023-01-10 | 2023-04-25 | 合肥工业大学 | Fair personalized recommendation method, equipment and storage medium based on mutual information decoupling |
| CN116167812A (en) * | 2022-12-02 | 2023-05-26 | 重庆邮电大学 | Heterogeneous collaborative filtering method for multi-behavior recommendation |
| CN116340648A (en) * | 2023-03-27 | 2023-06-27 | 广东工业大学 | Knowledge graph attention network recommendation method based on graph collaborative filtering |
| CN116401542A (en) * | 2023-03-20 | 2023-07-07 | 华南理工大学 | Multi-intention multi-behavior decoupling recommendation method and device |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20230055699A1 (en) * | 2021-08-06 | 2023-02-23 | Rakuten Group, Inc. | Recommender system and method using shared neural item representations for cold-start recommendations |
-
2023
- 2023-07-27 CN CN202310926715.9A patent/CN116645174B/en active Active
Patent Citations (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR20220042598A (en) * | 2020-09-28 | 2022-04-05 | (주)브레인콜라 | Personalized recommendation method and system based on future interaction prediction |
| WO2022222037A1 (en) * | 2021-04-20 | 2022-10-27 | 中国科学院深圳先进技术研究院 | Interpretable recommendation method based on graph neural network inference |
| CN114385804A (en) * | 2021-11-12 | 2022-04-22 | 中山大学 | Comment recommendation method of heterogeneous graph attention neural network based on meta-learning |
| CN114519147A (en) * | 2022-02-17 | 2022-05-20 | 云南大学 | Social recommendation method based on GCN enhanced influence diffusion |
| CN114693397A (en) * | 2022-03-16 | 2022-07-01 | 电子科技大学 | Multi-view multi-modal commodity recommendation method based on attention neural network |
| CN114861065A (en) * | 2022-05-31 | 2022-08-05 | 山东省人工智能研究院 | Multi-behavior based personalized recommendation method for cascaded residual error graph convolution network |
| CN115114526A (en) * | 2022-06-30 | 2022-09-27 | 南京邮电大学 | Weighted graph convolution network score prediction recommendation method for multi-behavior enhancement information |
| CN115357804A (en) * | 2022-07-19 | 2022-11-18 | 大连民族大学 | Two-stage collaborative filtering multi-behavior recommendation method based on graph convolution network |
| CN115187343A (en) * | 2022-07-20 | 2022-10-14 | 山东省人工智能研究院 | Multi-behavior recommendation method based on attention map convolution neural network |
| CN115905687A (en) * | 2022-11-01 | 2023-04-04 | 山东建筑大学 | Cold start-oriented recommendation system and method based on meta-learning graph neural network |
| CN115687772A (en) * | 2022-11-09 | 2023-02-03 | 安徽大学 | Sequence recommendation method based on sequence dependence enhanced self-attention network |
| CN116167812A (en) * | 2022-12-02 | 2023-05-26 | 重庆邮电大学 | Heterogeneous collaborative filtering method for multi-behavior recommendation |
| CN116010718A (en) * | 2023-01-10 | 2023-04-25 | 合肥工业大学 | Fair personalized recommendation method, equipment and storage medium based on mutual information decoupling |
| CN115878904A (en) * | 2023-02-22 | 2023-03-31 | 深圳昊通技术有限公司 | Intellectual property personalized recommendation method, system and medium based on deep learning |
| CN116401542A (en) * | 2023-03-20 | 2023-07-07 | 华南理工大学 | Multi-intention multi-behavior decoupling recommendation method and device |
| CN116340648A (en) * | 2023-03-27 | 2023-06-27 | 广东工业大学 | Knowledge graph attention network recommendation method based on graph collaborative filtering |
Non-Patent Citations (3)
| Title |
|---|
| "基于轻量图卷积和注意力增强的多行为推荐模型";高钰澜等;《计算机应用研究》;第第39卷卷(第第6期期);第1753-1759页 * |
| 基于属性偏好自学习的推荐方法;刘志;林振涛;鄢致雯;陈波;;浙江工业大学学报(02);全文 * |
| 李琳 ; 唐守廉 ; .基于多层注意力表示的音乐推荐模型.电子学报.2020,(09),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116645174A (en) | 2023-08-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dong et al. | A hybrid collaborative filtering model with deep structure for recommender systems | |
| CN111611472B (en) | Binding recommendation method and system based on graph convolution neural network | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN111881342A (en) | Recommendation method based on graph twin network | |
| CN110674407B (en) | Hybrid recommendation method based on graph convolution neural network | |
| CN111222332A (en) | Commodity recommendation method combining attention network and user emotion | |
| CN111310063A (en) | Neural network-based article recommendation method for memory perception gated factorization machine | |
| CN112328900A (en) | Deep learning recommendation method integrating scoring matrix and comment text | |
| Yu et al. | Cross-domain recommendation based on latent factor alignment | |
| CN110910218A (en) | Multi-behavior migration recommendation method based on deep learning | |
| CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
| JP7427717B2 (en) | Multimodal transformer-based item classification system, data processing system, data processing method, and computer-implemented method | |
| Wang et al. | SDDRS: stacked discriminative denoising auto-encoder based recommender system | |
| CN113918832A (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
| CN113918833A (en) | Product recommendation method realized through graph convolution collaborative filtering of social network relationship | |
| Zhang et al. | A deep dual adversarial network for cross-domain recommendation | |
| CN115630153A (en) | Research student literature resource recommendation method based on big data technology | |
| CN116401542A (en) | Multi-intention multi-behavior decoupling recommendation method and device | |
| CN115858919A (en) | Learning resource recommendation method and system based on project field knowledge and user comments | |
| CN111079011A (en) | Deep learning-based information recommendation method | |
| CN112765482A (en) | Product delivery method, device, equipment and computer readable medium | |
| CN116645174B (en) | Personalized recommendation method based on decoupling multi-behavior characterization learning | |
| Yao et al. | Convolutional geometric matrix completion | |
| CN114861065B (en) | Personalized recommendation method of cascade residual error graph convolution network based on multiple behaviors | |
| CN116910357A (en) | Data processing method and related device |
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 |