CN115659277A - E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion - Google Patents

E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion Download PDF

Info

Publication number
CN115659277A
CN115659277A CN202211299136.8A CN202211299136A CN115659277A CN 115659277 A CN115659277 A CN 115659277A CN 202211299136 A CN202211299136 A CN 202211299136A CN 115659277 A CN115659277 A CN 115659277A
Authority
CN
China
Prior art keywords
commodity
behavior
vector
feature
feature vector
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202211299136.8A
Other languages
Chinese (zh)
Inventor
卢官明
丁佳伟
鲍秉坤
余鹏航
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202211299136.8A priority Critical patent/CN115659277A/en
Publication of CN115659277A publication Critical patent/CN115659277A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an e-commerce session recommendation method and system based on multi-behavior feature fusion. The method comprises the following steps: acquiring four kinds of session behavior data of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in an e-commerce database; an e-commerce conversation recommendation model based on multi-behavior feature fusion is constructed and comprises a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module; training the e-commerce session recommendation model by using four session behavior data in an e-commerce database; and recommending commodities to the users in the conversation by using the trained e-commerce conversation recommendation model, and outputting a recommendation result. The invention utilizes the multi-behavior characteristic in the electronic commerce conversation recommendation model and can effectively improve the recommendation performance.

Description

E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion
Technical Field
The invention relates to the field of data mining, in particular to an e-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion.
Background
The e-commerce session recommendation refers to predicting commodities which are about to interact with a user by utilizing anonymous session information of the user, namely commodities which are interacted with the user in a short-time session in an e-commerce application scene, and is an important branch and a sub-topic of a recommendation system.
Initial conversational recommendations widely used the markov chain approach, usually using the previous interactive item to predict the next possible interactive item. Rendle et al combine the Markov chain with matrix decomposition for predicting the next interaction term; le et al build a Markov payment model by adding context information to improve recommendation performance; however, this method is performed under the assumption of strong independence, which limits the accuracy of prediction. In recent years, with the development of deep learning, a recurrent neural network is widely used in conversational recommendation. Hidasi et al propose the GRU4Rec model, which uses a recurrent neural network for session recommendation for the first time; tan et al upgrades the GRU4Rec model, improves recommendation performance and alleviates the over-fitting problem; liu et al utilize multiple layers of perceptrons and attention networks to improve recommendation performance. But the recurrent neural network is based on a dependency assumption that any adjacent interaction must contain a sequential relationship. Under the assumption of dependency, this method can only capture point-by-point dependencies, but not collective dependencies. The graph neural network can capture complex dependency relationships among nodes by virtue of the strong graph structure advantages, and is widely applied to recent session recommendation research. The SR-GNN model proposed by Wu et al applies graph networks to conversational recommendations for the first time and enhances conversational representation by adding an attention mechanism; the GC-SAN proposed by Xu et al learns a more accurate representation of the session using a multi-layer self-attention network. The strong node learning capability of the graph neural network enables the performance of the session recommendation to be greatly improved.
Although the application of the neural network greatly improves the performance of the conversation recommendation, the characteristic learning based on the click behavior cannot fully reflect the commodity characteristics and the conversation characteristics. In fact, besides the clicking action on the commodity, the user may also take other actions such as collecting, purchasing, adding to a shopping cart, etc., which also affect the feature learning. The existing model rarely considers other behaviors except the click behavior, so that the characteristic learning of the commodity is incomplete, and the recommendation performance is reduced.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the defects of the prior art, the invention aims to provide an e-commerce conversation recommendation method and system based on multi-behavior feature fusion.
The technical scheme is as follows: in order to achieve the aim, the invention provides an e-commerce conversation recommendation method based on multi-behavior feature fusion, which comprises the following steps of:
s1: acquiring four kinds of session behavior data of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in an e-commerce database;
s2: an e-commerce conversation recommendation model based on multi-behavior feature fusion is constructed and comprises a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module;
the behavior feature extraction module is used for extracting feature vectors of four behaviors of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in a conversation; the click behavior is a primary behavior, and the three behaviors of collecting, purchasing and adding into the shopping cart are secondary behaviors occurring on the basis of the click behavior;
the commodity feature extraction module consists of a hierarchical gate control circulation unit network and is used for fusing the extracted four behavior feature vectors and outputting commodity feature vectors; the hierarchical gated cyclic unit network is composed of three levels of gated cyclic unit networks:
the first-level gating circulation unit network takes the collection behavior feature vector and the purchasing behavior feature vector as input, adaptively learns the importance weights of the collection behavior feature vector and the purchasing behavior feature vector, and performs weighted summation on the importance weights to obtain a fusion feature vector of the collection behavior feature and the purchasing behavior feature;
the second-level gate control circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weight of the two feature vectors, and performs weighted summation on the importance weight to obtain the fusion feature vector of the secondary behavior;
the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector;
the commodity high-order feature extraction module is composed of a gated graph neural network, and extracts commodity high-order feature vectors by taking an adjacency matrix constructed by the click sequence relation of users to commodities in a session and the commodity feature vectors as input;
the conversation feature extraction module consists of an attention module and a gate control circulation unit network, and is used for performing self-adaptive fusion on the current interest feature vector and the global interest feature vector and outputting a conversation feature vector;
the commodity recommendation module calculates the similarity between the conversation characteristic vector and the candidate commodity characteristic vector through vector dot product operation, sorts the conversation characteristic vector and the candidate commodity characteristic vector according to the similarity from high to low, and recommends K-bit commodities with the similarity ranking top to the user;
s3: training the e-commerce session recommendation model by using four session behavior data in an e-commerce database;
s4: and recommending commodities to the users in the conversation by using the trained e-commerce conversation recommendation model, and outputting a recommendation result.
Preferably, the behavior feature extraction module in step S2 extracts four behavior feature vectors of clicking, collecting, purchasing, and adding a shopping cart to a commodity by a user in a session, and includes the following specific steps:
s2.1.1: for n commodities { v in e-commerce database 1 ,v 2 ,…,v n One-hot coding is carried out, and the codes are respectively embedded into d 1 Dimension space, get d 1 Commodity characterization vector of dimensions
Figure BDA0003902123170000031
Wherein i =1,2, \8230;, n, d 1 Is a positive integer;
s2.1.2: one-hot coding is carried out on three secondary behaviors of collection, purchase and addition to a shopping cart, and the three secondary behaviors are respectively embedded into d 2 Dimension space, get d 2 Collection behavior feature vector of dimension
Figure BDA0003902123170000032
Purchasing behavior feature vector
Figure BDA0003902123170000033
Adding shopping cart behavior feature vector
Figure BDA0003902123170000034
Wherein d is 2 Is a positive integer;
s2.1.3: for a commodity v i Characterizing the merchandise by a vector c i Respectively spliced with the three secondary behavior feature vectors to obtain a commodity v i Collected behavior feature vector of
Figure BDA0003902123170000035
Purchasing behavior feature vector
Figure BDA0003902123170000036
Adding behavior feature vectors to a shopping cart
Figure BDA0003902123170000037
S2.1.4: for a commodity v i Characterizing the commodity as a vector c i Performing 0 complementing operation to obtain click behavior feature vector
Figure BDA0003902123170000038
Wherein 0 is d 2 Zero vector of dimensions.
Preferably, the commodity feature extraction module in step S2 fuses the four extracted behavior feature vectors to output a commodity feature vector, and the specific steps are as follows:
s2.2.1: first-level gated cyclic unit network with commodity v i The collection behavior feature vector and the purchasing behavior feature vector are used as input, the importance weights of the collection behavior feature vector and the purchasing behavior feature vector are learned in a self-adaptive mode, the importance weights are subjected to weighted summation, the fusion feature vector of the collection behavior feature and the purchasing behavior feature is obtained, and the expression is as follows:
g 1 ＝σ(W 1 f i +W 2 b i )
m i ＝g 1 ⊙f i +(1-g 1 )⊙b i
wherein,
Figure BDA0003902123170000039
and
Figure BDA00039021231700000310
is a matrix of weights that can be trained,
Figure BDA00039021231700000311
is a weight vector used for regulating and controlling the contribution degree of the collection behavior feature vector and the purchasing behavior feature vector, sigma (-) represents sigmoid nonlinear activation function,
Figure BDA00039021231700000312
is a full 1 vector, which indicates a hadamard product,
Figure BDA00039021231700000313
is a fusion feature vector of collection and purchase behavior features;
s2.2.2: the second-level gating circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weights of the two input feature vectors, and performs weighted summation on the importance weights to obtain the fusion feature vector of the secondary behavior, wherein the expression is as follows:
g 2 ＝σ(W 3 m i +W 4 a i )
n i ＝g 2 ⊙m i +(1-g 2 )⊙a i
wherein,
Figure BDA0003902123170000041
and
Figure BDA0003902123170000042
is a matrix of weights that can be trained,
Figure BDA0003902123170000043
is a weight vector used for regulating and controlling the contribution degree of the collection and purchase fusion feature vector and the feature vector added into the behavior of the shopping cart,
Figure BDA0003902123170000044
is a fused feature vector of the secondary behavior;
s2.2.3: the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two input characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector, wherein the expression is as follows:
g 3 ＝σ(W 5 n i +W 6 r i )
Figure BDA0003902123170000045
wherein,
Figure BDA0003902123170000046
and
Figure BDA0003902123170000047
is a matrix of weights that can be trained in the process,
Figure BDA0003902123170000048
is a weight vector for regulating and controlling the contribution degree of the fusion feature vector of the click behavior feature vector and the secondary behavior,
Figure BDA0003902123170000049
is a commodity feature vector.
Preferably, the commodity high-order feature extraction module in step S2 takes an adjacency matrix and a commodity feature vector constructed by a user' S click order relationship to the commodity in the session as input, and the specific steps of extracting the commodity high-order feature vector are as follows:
s2.3.1: for different goods v in a conversation s s,1 ,v s,2 ,...,v s,l The commodity feature vectors are respectively expressed as
Figure BDA00039021231700000410
According to the click sequence relation v of the user to the commodity in the session s,1 →v s,2 →…→v s,m Building a adjacency matrix
Figure BDA0003902123170000051
Element x in the adjacency matrix ow When o = w, the value is 1; when o ≠ w, if the merchandise v s,o And the commodity v s,w There is a sequential interaction relationship, i.e. clicking on the item v s,o Then click on the commodity v s,w Then x ow =1, otherwise x ow =0; where l represents the number of different items in the conversation s, l is a positive integer, o =1,2,. L, w =1,2,...l，m∈{1,2,...,l}；
s2.3.2: for item v in conversation s s,j Feature vectors of merchandise in conversation s and adjacency matrix A s Neutralization commodity v s,j Inputting the related column vectors into an L-layer gating graph neural network; after k-layer gating graph neural network learning, a commodity v is obtained s,j The expression of the aggregated feature vector of (2) is:
Figure BDA0003902123170000052
wherein,
Figure BDA0003902123170000053
and
Figure BDA0003902123170000054
is a learnable parameter that, when k =1,
Figure BDA0003902123170000055
is that
Figure BDA0003902123170000056
Is a commodity feature vector that, when k ∈ {2, 3.., L },
Figure BDA0003902123170000057
is a commodity feature vector output by the neural network of the gate map of the k-1 level,
Figure BDA0003902123170000058
represents the adjacency matrix A s Neutralization commodity v s,j The associated column vector is then used to determine,
Figure BDA0003902123170000059
commodity v representing k-th layer gate control graph neural network output s,j L is a positive integer, j belongs to {1, 2.. Multidot.l }, and k belongs to {1,2, 3.. Multidot.l };
s2.3.3: using the updated gate of the gating network, adaptively learning in the k-th layer gating graph neural network, and obtaining the commodity v s,j Commodity characteristic vector output by k-1 layer network
Figure BDA00039021231700000510
The information needing to be updated is expressed as follows:
Figure BDA00039021231700000511
wherein,
Figure BDA00039021231700000512
and
Figure BDA00039021231700000513
is a parameter that can be learned by the user,
Figure BDA00039021231700000514
the weight vector of the output of the update gate in the gate control network determines the commodity characteristic vector output by the k-1 layer network
Figure BDA00039021231700000515
How much information is updated;
s2.3.4: using reset gate of gate control network, self-adaptive learning in k-th layer gate control graph neural network s,j Commodity characteristic vector output in k-1 layer network
Figure BDA0003902123170000061
The information to be discarded is needed, and a candidate state vector of the k-th layer network is calculated, and the expression is as follows:
Figure BDA0003902123170000062
Figure BDA0003902123170000063
wherein,
Figure BDA0003902123170000064
and
Figure BDA0003902123170000065
is a parameter that can be learned by the user,
Figure BDA0003902123170000066
is the output vector of the reset gate in the gate control network, and determines the commodity characteristic vector output by the k-1 layer network
Figure BDA0003902123170000067
How much information is discarded and how much information is discarded,
Figure BDA0003902123170000068
is a candidate state vector for the k-th layer network;
s2.3.5: calculating the goods v s,j And (3) outputting the commodity feature vector after passing through the k-th layer gate control diagram neural network, wherein the expression of the commodity feature vector is as follows:
Figure BDA0003902123170000069
wherein,
Figure BDA00039021231700000610
goods v s,j Commodity feature vectors output after passing through a k-layer gating graph neural network;
s2.3.6: when k = L, we obtain v s,j The expression of the commodity high-order feature vector is as follows:
Figure BDA00039021231700000611
wherein,
Figure BDA00039021231700000612
is v s,j The commodity high-order feature vector of (1); goods v in conversation s s,1 ,v s,2 ,...,v s,l The high-order feature vectors thereof all are from aboveThe steps are calculated and respectively expressed as
Figure BDA00039021231700000613
Preferably, the session feature extraction module in step S2 performs adaptive fusion on the current interest feature vector and the global interest feature vector to output a session feature vector, and the specific steps are as follows:
s2.4.1: for items v clicked in order in session s s,1 →v s,2 →…→v s,m The last clicked commodity v in the conversation s,m Higher order feature vector of
Figure BDA00039021231700000614
Considering as a current interest feature vector
Figure BDA00039021231700000615
S2.4.2: calculating a weight coefficient of a high-order feature vector of each commodity in the conversation and a current interest feature vector through a soft attention mechanism, and calculating to obtain a global interest feature vector according to the weight coefficient, wherein the expression is as follows:
Figure BDA00039021231700000616
Figure BDA0003902123170000071
wherein,
Figure BDA0003902123170000072
and
Figure BDA0003902123170000073
is a matrix of weights that can be trained,
Figure BDA0003902123170000074
is a parameter that can be learned by the user,
Figure BDA0003902123170000075
is the transpose of q, α j Is the weight coefficient obtained by the calculation,
Figure BDA0003902123170000076
is a global interest feature vector;
s2.4.3: using a gate control cycle unit network to perform self-adaptive fusion on the obtained current interest feature vector and the global interest feature vector to obtain a session feature vector, wherein the expression is as follows:
g 4 ＝σ(W 9 s l +W 10 s g )
s h ＝g 4 ⊙s l +(1-g 4 )⊙s g
wherein,
Figure BDA0003902123170000077
and
Figure BDA0003902123170000078
is a matrix of weights that can be trained,
Figure BDA0003902123170000079
is a weight vector for regulating and controlling the contribution degree of the current interest feature vector and the global interest feature vector,
Figure BDA00039021231700000710
is the session feature vector.
Preferably, the commodity recommending module in step S2 calculates a similarity score vector between the conversation feature vector and a candidate commodity feature vector through a vector dot product operation, where the candidate commodities are n commodities in an e-commerce database, and the expression of the similarity score vector is as follows:
Figure BDA00039021231700000711
wherein,
Figure BDA00039021231700000712
is s h Transposing; scoring a similarity vector z using a softmax function i And (4) carrying out normalization, wherein the expression is as follows:
y i ＝softmax(z i )
and sorting the commodities according to the normalized value of the similarity score vector from high to low, and recommending K-bit commodities before ranking to the user, wherein K is a positive integer.
In addition, the invention provides an e-commerce conversation recommendation system based on multi-behavior feature fusion, which comprises a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module;
the behavior feature extraction module is used for extracting feature vectors of four behaviors of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in a conversation; the click behavior is a primary behavior, and the three behaviors of collecting, purchasing and adding into the shopping cart are secondary behaviors occurring on the basis of the click behavior;
the commodity feature extraction module consists of a hierarchical gate control circulation unit network and is used for fusing the extracted four behavior feature vectors and outputting commodity feature vectors; the hierarchical gated cyclic unit network is composed of three levels of gated cyclic unit networks:
the first-stage gating circulation unit network takes the collection behavior characteristic vector and the purchasing behavior characteristic vector as input, adaptively learns the importance weights of the collection behavior characteristic vector and the purchasing behavior characteristic vector, and performs weighted summation on the importance weights to obtain a fusion characteristic vector of the collection behavior characteristic and the purchasing behavior characteristic;
the second-level gate control circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weight of the two feature vectors, and performs weighted summation on the importance weight to obtain the fusion feature vector of the secondary behavior;
the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector;
the commodity high-order feature extraction module is composed of a gated graph neural network, and takes an adjacency matrix constructed by a user clicking sequence relation of commodities and commodity feature vectors in a session as input to extract the commodity high-order feature vectors;
the session feature extraction module is composed of an attention module and a gate control circulation unit network, and is used for performing self-adaptive fusion on the current interest feature vector and the global interest feature vector and outputting a session feature vector;
the commodity recommendation module calculates the similarity between the conversation characteristic vector and the candidate commodity characteristic vector through vector dot product operation, sorts the conversation characteristic vector and the candidate commodity characteristic vector according to the similarity from high to low, and recommends K-bit commodities with the similarity ranking to a user;
furthermore, the present invention provides an e-commerce session recommendation apparatus based on multi-behavior feature fusion, including at least one computing device, the computing device including a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program, when loaded into the processor, implementing an e-commerce session recommendation method based on multi-behavior feature fusion according to any one of claims 1 to 7.
Furthermore, the present invention proposes a computer-readable storage medium having stored thereon a computer program which, when being executed by a processor, implements a multi-behavior feature fusion based e-commerce session recommendation method according to any one of claims 1 to 7.
Has the beneficial effects that: compared with the prior art, the technical scheme of the invention has the following beneficial technical effects:
1. by means of multi-behavior feature fusion, commodity feature representation is enriched, in the subsequent feature learning process, the conversation purpose of a user can be better grasped, and more targeted recommendation is implemented;
2. the relationship among various behaviors is learned in a self-adaptive mode by utilizing the multi-level gating cycle unit network, so that the model has good trainability, and the robustness and the generalization of the session recommendation system are improved;
3. and a global interest characteristic and a current interest characteristic are fused by using the gated cyclic unit network, a session characteristic learning mode is updated, and the performance of the session recommendation system is further improved.
Drawings
FIG. 1 is an e-commerce session recommendation method based on multi-behavior feature fusion according to the present invention;
FIG. 2 is a block diagram of a multi-level gated loop cell network used in the present invention;
FIG. 3 is a schematic diagram of a neural network for a gated graph of high-order feature vectors of merchandise in a learning session according to the present invention;
FIG. 4 is a block diagram of a gated cyclic unit network incorporating global interest feature vectors and current interest feature vectors in accordance with the present invention.
Detailed Description
The following provides a more detailed description of embodiments of the present invention, as illustrated in the accompanying drawings.
As shown in fig. 1, an e-commerce session recommendation method based on multi-behavior feature fusion provided by an embodiment of the present invention mainly includes the following steps:
step S1, four kinds of conversation behavior data of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in an E-commerce database are obtained: the present embodiment uses the UserBehavior database, which is a Taobao user behavior database provided by Aliiba; the database comprises all behaviors (behaviors comprise clicking, collecting, purchasing and adding a shopping cart) of about one million random users between 2017-09-11 and 2017-12-03, each row of the data set represents one user behavior, and each row of the data set consists of a user ID, a commodity category ID, a behavior type and a timestamp which are separated by commas; preprocessing the session data in the database, regarding the session with the length of w, taking the commodity clicked in the first w-1 times of the session as a training sample in the embodiment, taking the commodity clicked in the last time of the session as a verification sample of the training sample, and processing all the sessions in the database according to the method.
S2, constructing an e-commerce conversation recommendation model based on multi-behavior feature fusion, wherein the model comprises a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module;
s2.1, a behavior feature extraction module: the specific steps of extracting four behavior feature vectors of clicking, collecting, purchasing and adding the shopping cart to the commodity in the conversation by the user are as follows:
s2.1.1: for n commodities { v in e-commerce database 1 ,v 2 ,…,v n Respectively embedding and coding to obtain d 1 Commodity characterization vector of dimensions
Figure BDA0003902123170000091
Wherein i =1,2, \ 8230, n, n =4162024 in this example 1 ＝100；
S2.1.2: three secondary behaviors of collection, purchase and addition of a shopping cart are respectively embedded and coded to obtain d 2 Collection behavior feature vector of dimension
Figure BDA0003902123170000092
Purchasing behavior feature vector
Figure BDA0003902123170000093
Adding shopping cart behavior feature vector
Figure BDA0003902123170000101
In this example d 2 ＝1；
S2.1.3: for commodity v i Characterizing the merchandise by a vector c i Respectively spliced with the three secondary behavior feature vectors to obtain a commodity v i Collected behavior feature vector of
Figure BDA0003902123170000102
Purchasing behavior feature vector
Figure BDA0003902123170000103
Behavior of joining shopping cartEigenvector
Figure BDA0003902123170000104
S2.1.4: for a commodity v i Characterizing the merchandise by a vector c i Performing 0 complementing operation to obtain the click behavior characteristic vector
Figure BDA0003902123170000105
In this embodiment, 0 is a 1-dimensional constant.
S2.2, commodity feature extraction: as shown in fig. 2, the specific steps of fusing the extracted four behavior feature vectors and outputting the commodity feature vector are as follows:
s2.2.1: first-level gated cyclic unit network with commodity v i The collection behavior feature vector and the purchasing behavior feature vector are used as input, the importance weights of the collection behavior feature vector and the purchasing behavior feature vector are learned in a self-adaptive mode, the importance weights are weighted and summed, the fusion feature vector of the collection behavior feature and the purchasing behavior feature is obtained, and the expression is as follows:
g 1 ＝σ(W 1 f i +W 2 b i )
m i ＝g 1 ⊙f i +(1-g 1 )⊙b i
wherein,
Figure BDA0003902123170000106
and
Figure BDA0003902123170000107
is a matrix of weights that can be trained in the process,
Figure BDA0003902123170000108
is a weight vector for regulating and controlling contribution degrees of a collection behavior characteristic vector and a purchase behavior characteristic vector, sigma (-) represents a sigmoid nonlinear activation function,
Figure BDA0003902123170000109
is a full 1 vector, which indicates a hadamard product,
Figure BDA00039021231700001010
is a fusion feature vector of collection and purchase behavior features;
s2.2.2: the second-level gating circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weights of the two input feature vectors, and performs weighted summation on the importance weights to obtain the fusion feature vector of the secondary behavior, wherein the expression is as follows:
g 2 ＝σ(W 3 m i +W 4 a i )
n i ＝g 2 ⊙m i +(1-g 2 )⊙a i
wherein,
Figure BDA00039021231700001011
and
Figure BDA00039021231700001012
is a matrix of weights that can be trained,
Figure BDA00039021231700001013
is a weight vector used for regulating and controlling the contribution degree of the collection and purchase fusion feature vector and the feature vector added into the behavior of the shopping cart,
Figure BDA00039021231700001014
is a fused feature vector of the secondary behavior;
s2.2.3: the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two input characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector, wherein the expression is as follows:
g 3 ＝σ(W 5 n i +W 6 r i )
Figure BDA0003902123170000111
wherein,
Figure BDA0003902123170000112
and
Figure BDA0003902123170000113
is a matrix of weights that can be trained,
Figure BDA0003902123170000114
is a weight vector for regulating and controlling the contribution degree of the fusion feature vector of the click behavior feature vector and the secondary behavior,
Figure BDA0003902123170000115
is a commodity feature vector.
And continuously and iteratively training the commodity feature extraction module through an error back propagation algorithm until the model parameters reach the optimal values. Then, the four behavior feature vectors can be input into a trained commodity feature extraction module to extract a commodity feature vector v i 。
S2.3, a commodity high-order feature extraction module: as shown in fig. 3, the specific steps of extracting the commodity high-order feature vector by using the adjacency matrix constructed by the click sequence relationship of the user to the commodity in the session and the commodity feature vector as input are as follows:
s2.3.1: take training sample session s in the database as an example, and verify the sample as v s,c (ii) a For different goods v in a training sample session s s,1 ,v s,2 ,...,v s,l In this embodiment, l =6, training the commodity v in the sample session s s,1 ,v s,2 ,v s,3 ,v s,4 ,v s,5 ,v s,6 The commodity feature vectors can be respectively expressed as
Figure BDA0003902123170000116
According to the click sequence relation of the user to the commodity in the training sample session s: v. of s,1 →v s,2 →v s,3 →v s,4 →v s,3 →v s,5 →v s,1 →v s,6 →v s,3 Building a adjacency matrix
Figure BDA0003902123170000117
In the present embodiment, the first and second electrodes are,
Figure BDA0003902123170000118
considering the influence of commodities, we add a self-loop to each commodity, so the adjacency matrix A s The main diagonal elements of (1);
s2.3.2: to train the commodity v in the sample session s s,3 For example, the commodity feature vector in the training sample session s and the adjacency matrix A s Neutralization commodity v s,3 Inputting the related column vectors into an L-layer gating graph neural network; after the neural network learning of the k-th layer gate control graph, a commodity v is obtained s,3 The expression of the aggregated feature vector of (2) is:
Figure BDA0003902123170000121
wherein,
Figure BDA0003902123170000122
and
Figure BDA0003902123170000123
is a learnable parameter that, when k =1,
Figure BDA0003902123170000124
is that
Figure BDA0003902123170000125
Is a commodity feature vector that, when k ∈ {2, 3.., L },
Figure BDA0003902123170000126
is a commodity feature vector output by the neural network of the gate map of the k-1 layer,
Figure BDA0003902123170000127
representing an adjacency matrix A s Neutralizing commodity v s,3 In connection withThe number of column vectors is such that,
Figure BDA0003902123170000128
commodity v representing k-th layer gate control graph neural network output s,3 L =3, j ∈ {1,2,..., L }, k ∈ {1,2,3,..., L }, in this example;
s2.3.3: using the updated gate of the gating network, adaptively learning in the k-th layer gating graph neural network, and obtaining the commodity v s,3 Commodity characteristic vector output by k-1 layer network
Figure BDA0003902123170000129
The information needing to be updated is expressed as follows:
Figure BDA00039021231700001210
wherein,
Figure BDA00039021231700001211
and
Figure BDA00039021231700001212
is a parameter that can be learned by the user,
Figure BDA00039021231700001213
the weight vector output by the update gate in the gate control network determines the commodity characteristic vector output by the k-1 layer network
Figure BDA00039021231700001214
How much information is updated;
s2.3.4: using reset gate of gate control network, self-adaptive learning in k-th layer gate control graph neural network s,3 Commodity characteristic vector output by k-1 layer network
Figure BDA00039021231700001215
The information to be discarded is calculated, and a candidate state vector of the k-th layer network is calculated, wherein the expression of the candidate state vector is as follows:
Figure BDA00039021231700001216
Figure BDA00039021231700001217
wherein,
Figure BDA0003902123170000131
and
Figure BDA0003902123170000132
is a parameter that can be learned by the user,
Figure BDA0003902123170000133
is the output vector of the reset gate in the gate control network, and determines the commodity characteristic vector output by the k-1 layer network
Figure BDA0003902123170000134
How much information is discarded and how much information is discarded,
Figure BDA0003902123170000135
is a candidate state vector for the k-th layer network;
s2.3.5: calculating the goods v s,3 The commodity feature vector output after passing through the k-th layer gate control graph neural network has the expression:
Figure BDA0003902123170000136
wherein,
Figure BDA0003902123170000137
commodity v s,3 Commodity feature vectors output after passing through a k-layer gating pattern neural network;
s2.3.6: when k = L, we obtain v s,3 The expression of the commodity high-order feature vector is as follows:
Figure BDA0003902123170000138
wherein,
Figure BDA0003902123170000139
is v s,3 The commodity high-order feature vector; other goods v in training sample sessions s s,1 ,v s,2 ,v s,4 ,v s,5 ,v s,6 The high-order feature vectors can be obtained by the above steps and are respectively expressed as
Figure BDA00039021231700001310
And continuously and iteratively training the commodity high-order feature extraction module through an error back propagation algorithm until the model parameters reach the optimal values. Then, the commodity feature vector can be input into a trained commodity high-order feature extraction module to extract the corresponding commodity high-order feature vector.
S2.4 session feature extraction module: as shown in fig. 4, the specific steps of adaptively fusing the current interest feature vector and the global interest feature vector to output the session feature vector are as follows:
s2.4.1: taking training sample session s in database as an example, aiming at commodity v clicked in sequence in training sample session s s,1 →v s,2 →v s,3 →v s,4 →v s,3 →v s,5 →v s,1 →v s,6 →v s,3 The last clicked commodity v in the training sample session s,3 Higher order feature vector of
Figure BDA00039021231700001311
Considering as a current interest feature vector
Figure BDA00039021231700001312
S2.4.2: calculating a weight coefficient of the high-order characteristic vector of each commodity in the conversation and the current interest characteristic vector through a soft attention mechanism, and calculating to obtain a global interest characteristic vector according to the weight coefficient, wherein the expression of the global interest characteristic vector is as follows:
Figure BDA00039021231700001313
Figure BDA00039021231700001314
wherein,
Figure BDA0003902123170000141
and
Figure BDA0003902123170000142
is a matrix of weights that can be trained in the process,
Figure BDA0003902123170000143
is a parameter that can be learned by the user,
Figure BDA0003902123170000144
is the transpose of q, α j Is the weight coefficient obtained by the calculation,
Figure BDA0003902123170000145
is a global interest feature vector, in this embodiment, j belongs to {1,2,3,4,5,6};
s2.4.3: using a gate control cycle unit network to perform self-adaptive fusion on the obtained current interest feature vector and the global interest feature vector to obtain a session feature vector, wherein the expression of the session feature vector is as follows:
g 4 ＝σ(W 9 s l +W 10 s g )
s h ＝g 4 ⊙s l +(1-g 4 )⊙s g
wherein,
Figure BDA0003902123170000146
and
Figure BDA0003902123170000147
is trainableThe weight matrix is a matrix of weights,
Figure BDA0003902123170000148
is a weight vector for regulating the contribution degree of the current interest feature vector and the global interest feature vector,
Figure BDA0003902123170000149
is the session feature vector.
And continuously iterating the training session characteristic extraction module through an error back propagation algorithm until the model parameters reach the optimal values. Then, the commodity high-order feature vector can be input into a trained conversation feature extraction module to extract the conversation feature vector.
S2.5, a commodity recommendation module: calculating a similarity score vector of the conversation characteristic vector and a candidate commodity characteristic vector through vector dot product operation, wherein the candidate commodities are n commodities in an e-commerce database, and the expression of the similarity score vector is as follows:
Figure BDA00039021231700001410
wherein,
Figure BDA00039021231700001411
is s h Transposing; scoring a similarity vector z using a softmax function i And (4) carrying out normalization, wherein the expression is as follows:
y i ＝softmax(z i )
and sorting according to the normalized value of the similarity score vector from high to low, and recommending K-bit commodities before ranking to the user, wherein K =5 in the embodiment.
Verification sample v of conversation s of recommendation result and training sample s,c And comparing, calculating two indexes of recall rate and average reciprocal ranking, and continuously performing iterative training through an error back propagation algorithm until the two indexes of recall rate and average reciprocal ranking reach the optimal.
Step S3, the E-commerce session recommendation model is trained by using four session behavior data in the E-commerce database: in this embodiment, the training samples in the UserBehavior database are used, a back propagation algorithm is adopted, and a cross entropy function is used as a loss function to perform iterative training on the e-commerce session recommendation model until all parameters in the model are optimal.
And S4, recommending commodities to the users in the conversation by using the trained e-commerce conversation recommendation model, and outputting a recommendation result.
Those skilled in the art will appreciate that the steps of the embodiments may be adaptively changed and arranged in one or more systems different from the embodiments. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components.

Claims (9)

1. An e-commerce conversation recommendation method based on multi-behavior feature fusion is characterized by comprising the following steps of:
s1: acquiring four kinds of session behavior data of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in an e-commerce database;
s2: an e-commerce conversation recommendation model based on multi-behavior feature fusion is constructed and comprises a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module;
the behavior feature extraction module is used for extracting feature vectors of four behaviors of clicking, collecting, purchasing and adding a shopping cart in a conversation by a user; the click behavior is a primary behavior, and the three behaviors of collecting, purchasing and adding into the shopping cart are secondary behaviors occurring on the basis of the click behavior;
the commodity feature extraction module consists of a hierarchical gate control circulation unit network and is used for fusing the extracted four behavior feature vectors and outputting commodity feature vectors; the hierarchical gated cyclic unit network is composed of three levels of gated cyclic unit networks:
the first-stage gating circulation unit network takes the collection behavior characteristic vector and the purchasing behavior characteristic vector as input, adaptively learns the importance weights of the collection behavior characteristic vector and the purchasing behavior characteristic vector, and performs weighted summation on the importance weights to obtain a fusion characteristic vector of the collection behavior characteristic and the purchasing behavior characteristic;
the second-level gate control circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weight of the two feature vectors, and performs weighted summation on the importance weight to obtain the fusion feature vector of the secondary behavior;
the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector;
the commodity high-order feature extraction module is composed of a gated graph neural network, and extracts commodity high-order feature vectors by taking an adjacency matrix constructed by the click sequence relation of users to commodities in a session and the commodity feature vectors as input;
the conversation feature extraction module consists of an attention module and a gate control circulation unit network, and is used for performing self-adaptive fusion on the current interest feature vector and the global interest feature vector and outputting a conversation feature vector;
the commodity recommendation module calculates the similarity between the conversation characteristic vector and the candidate commodity characteristic vector through vector dot product operation, sorts the conversation characteristic vector and the candidate commodity characteristic vector according to the similarity from high to low, and recommends K-bit commodities with the similarity ranking to a user;
s3: training the e-commerce session recommendation model by using four session behavior data in an e-commerce database;
s4: and recommending commodities to the users in the conversation by using the trained e-commerce conversation recommendation model, and outputting a recommendation result.
2. The e-commerce conversation recommendation method based on multi-behavior feature fusion of claim 1, wherein the behavior feature extraction module in the step S2 extracts four behavior feature vectors of clicking, collecting, purchasing and adding a shopping cart to a commodity of a user in a conversation as follows:
s2.1.1: for n commodities { v in e-commerce database 1 ,v 2 ,…,v n One-hot coding is carried out, and the codes are respectively embedded into d 1 Dimension space, get d 1 Commodity characterization vector of dimensions
Figure FDA0003902123160000021
Wherein i =1,2, \8230;, n, d 1 Is a positive integer;
s2.1.2: one-hot coding is carried out on three secondary behaviors of collection, purchase and addition to a shopping cart, and the three secondary behaviors are respectively embedded into d 2 Dimension space, get d 2 Collection behavior feature vector of dimension
Figure FDA0003902123160000022
Purchasing behavior feature vector
Figure FDA0003902123160000023
Adding behavior feature vectors to a shopping cart
Figure FDA0003902123160000024
Wherein, d 2 Is a positive integer;
s2.1.3: for commodity v i Characterizing the merchandise by a vector c i Respectively spliced with the three secondary behavior feature vectors to obtain a commodity v i Collected behavior feature vector of
Figure FDA0003902123160000025
Purchasing behavior feature vector
Figure FDA0003902123160000026
Adding shopping cart behavior feature vector
Figure FDA0003902123160000027
S2.1.4: for a commodity v i To show the commodityEigenvector c i Performing 0 complementing operation to obtain the characteristic vector of the click behavior
Figure FDA0003902123160000028
Wherein 0 is d 2 Zero vector of dimensions.
3. The e-commerce conversation recommendation method based on multi-behavior feature fusion as claimed in claim 2, wherein the commodity feature extraction module in step S2 fuses the extracted four behavior feature vectors to output a commodity feature vector, and the specific steps of:
s2.2.1: first-level gated cyclic unit network with commodity v i The collection behavior feature vector and the purchasing behavior feature vector are used as input, the importance weights of the collection behavior feature vector and the purchasing behavior feature vector are learned in a self-adaptive mode, the importance weights are subjected to weighted summation, the fusion feature vector of the collection behavior feature and the purchasing behavior feature is obtained, and the expression is as follows:
g 1 ＝σ(W 1 f i +W 2 b i )
m i ＝g 1 ⊙f i +(1-g 1 )⊙b i
wherein,
Figure FDA0003902123160000031
and
Figure FDA0003902123160000032
is a matrix of weights that can be trained in the process,
Figure FDA0003902123160000033
is a weight vector used for regulating and controlling the contribution degree of the collection behavior feature vector and the purchasing behavior feature vector, sigma (-) represents sigmoid nonlinear activation function,
Figure FDA0003902123160000034
is a full 1 vector, which indicates a hadamard product,
Figure FDA0003902123160000035
is a fusion feature vector of collection and purchase behavior features;
s2.2.2: the second-level gating circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weights of the two input feature vectors, and performs weighted summation on the importance weights to obtain the fusion feature vector of the secondary behavior, wherein the expression is as follows:
g 2 ＝σ(W 3 m i +W 4 a i )
n i ＝g 2 ⊙m i +(1-g 2 )⊙a i
wherein,
Figure FDA0003902123160000036
and
Figure FDA0003902123160000037
is a matrix of weights that can be trained in the process,
Figure FDA0003902123160000038
is a weight vector used for regulating and controlling the contribution degree of the collection and purchase fusion feature vector and the feature vector added into the behavior of the shopping cart,
Figure FDA0003902123160000039
is a fused feature vector of the secondary behavior;
s2.2.3: the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two input characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector, wherein the expression is as follows:
g 3 ＝σ(W 5 n i +W 6 r i )
Figure FDA00039021231600000310
wherein,
Figure FDA00039021231600000311
and
Figure FDA00039021231600000312
is a matrix of weights that can be trained in the process,
Figure FDA00039021231600000313
is a weight vector for regulating and controlling the contribution degree of the fusion feature vector of the click behavior feature vector and the secondary behavior,
Figure FDA00039021231600000314
is a commodity feature vector.
4. The e-commerce conversation recommendation method based on multi-behavior feature fusion as claimed in claim 3, wherein the commodity high-order feature extraction module in step S2 takes an adjacency matrix constructed by a user' S click sequence relation to a commodity in a conversation and a commodity feature vector as input, and the specific steps of extracting the commodity high-order feature vector are as follows:
s2.3.1: for different goods v in a conversation s s,1 ,v s,2 ,…,v s,l The commodity feature vectors are respectively represented as
Figure FDA0003902123160000041
According to the click sequence relation v of users to commodities in the session s,1 →v s,2 →…→v s,m Building an adjacency matrix
Figure FDA0003902123160000042
Element x in the adjacency matrix ow When o = w, the value is 1; when o ≠ w, if the merchandise v s,o And the commodity v s,w There is a sequential interaction relationship, i.e. clicking on the item v s,o Then, immediatelyClick on commodity v s,w Then x ow =1, otherwise x ow =0; wherein l represents the number of different commodities in the session s, l is a positive integer, o =1,2, \8230l, w =1,2, \8230l, m belongs to {1,2, \8230;, l };
s2.3.2: for item v in conversation s s,j The commodity feature vector in the session s and the adjacency matrix A s Neutralization commodity v s,j Inputting the related column vectors into an L-layer gating graph neural network; after k-layer gating graph neural network learning, a commodity v is obtained s,j The expression of the aggregated feature vector of (2) is:
Figure FDA0003902123160000043
wherein,
Figure FDA0003902123160000044
and
Figure FDA0003902123160000045
is a learnable parameter that, when k =1,
Figure FDA0003902123160000046
is that
Figure FDA0003902123160000047
Is the commodity feature vector, when k ∈ {2,3, \8230;, L },
Figure FDA0003902123160000048
is a commodity feature vector output by the neural network of the gate map of the k-1 layer,
Figure FDA0003902123160000049
represents the adjacency matrix A s Neutralization commodity v s,j The associated column vector is then used to determine,
Figure FDA00039021231600000410
to representCommodity v output by neural network of k-th layer gate control diagram s,j L is a positive integer, j belongs to {1,2, \8230;, L }, k belongs to {1,2,3, \8230;, L };
s2.3.3: using the updated gate of the gating network, adaptively learning in the k-th layer gating graph neural network, and obtaining the commodity v s,j Commodity characteristic vector output by k-1 layer network
Figure FDA00039021231600000411
The information needing to be updated is expressed as follows:
Figure FDA00039021231600000412
wherein,
Figure FDA0003902123160000051
and
Figure FDA0003902123160000052
is a parameter that can be learned by the user,
Figure FDA0003902123160000053
the weight vector output by the update gate in the gate control network determines the commodity characteristic vector output by the k-1 layer network
Figure FDA0003902123160000054
How much information is updated;
s2.3.4: using reset gate of gate control network, self-adaptive learning in k-th layer gate control graph neural network s,j Commodity characteristic vector output in k-1 layer network
Figure FDA0003902123160000055
The information to be discarded is needed, and a candidate state vector of the k-th layer network is calculated, and the expression is as follows:
Figure FDA0003902123160000056
Figure FDA0003902123160000057
wherein,
Figure FDA0003902123160000058
and
Figure FDA0003902123160000059
is a parameter that can be learned by the user,
Figure FDA00039021231600000510
is the output vector of the reset gate in the gate control network, and determines the commodity characteristic vector output by the k-1 layer network
Figure FDA00039021231600000511
How much information is discarded and how much information is discarded,
Figure FDA00039021231600000512
is a candidate state vector for the k-th layer network;
s2.3.5: calculating a commodity v s,j The commodity feature vector output after passing through the k-th layer gate control graph neural network has the expression:
Figure FDA00039021231600000513
wherein,
Figure FDA00039021231600000514
goods v s,j Commodity feature vectors output after passing through a k-layer gating graph neural network;
s2.3.6: when k = L, we obtain v s,j The expression of the commodity high-order feature vector is as follows:
Figure FDA00039021231600000515
wherein,
Figure FDA00039021231600000516
is v s,j The commodity high-order feature vector of (1); goods v in conversation s s,1 ,v s,2 ,…,v s,l The high-order feature vectors are calculated by the above steps and are respectively expressed as
Figure FDA00039021231600000517
5. The e-commerce conversation recommendation method based on multi-behavior feature fusion of claim 4, wherein the conversation feature extraction module in the step S2 performs adaptive fusion on the current interest feature vector and the global interest feature vector to output a conversation feature vector by the following specific steps:
s2.4.1: for items v clicked in order in session s s,1 →v s,2 →…→v s,m The last clicked commodity v in the conversation s,m Higher order feature vector of
Figure FDA0003902123160000061
Feature vector regarded as current interest
Figure FDA0003902123160000062
S2.4.2: calculating a weight coefficient of a high-order feature vector of each commodity in the conversation and a current interest feature vector through a soft attention mechanism, and calculating to obtain a global interest feature vector according to the weight coefficient, wherein the expression is as follows:
Figure FDA0003902123160000063
Figure FDA0003902123160000064
wherein,
Figure FDA0003902123160000065
and
Figure FDA0003902123160000066
is a matrix of weights that can be trained,
Figure FDA0003902123160000067
is a parameter that can be learned by the user,
Figure FDA00039021231600000615
is the transpose of q, α j Is the weight coefficient obtained by the calculation,
Figure FDA0003902123160000068
is a global interest feature vector;
s2.4.3: using a gate control cycle unit network to perform self-adaptive fusion on the obtained current interest feature vector and the global interest feature vector to obtain a session feature vector, wherein the expression of the session feature vector is as follows:
g 4 ＝σ(W 9 s l +W 10 s g )
s h ＝g 4 ⊙s l +(1-g 4 )⊙s g
wherein,
Figure FDA0003902123160000069
and
Figure FDA00039021231600000610
is a matrix of weights that can be trained in the process,
Figure FDA00039021231600000611
is a weight vector for regulating the contribution degree of the current interest feature vector and the global interest feature vector,
Figure FDA00039021231600000612
is the session feature vector.
6. The e-commerce conversation recommendation method based on multi-behavior feature fusion as claimed in claim 5, wherein the commodity recommendation module in step S2 calculates similarity score vectors between the conversation feature vector and candidate commodity feature vectors through vector dot product operation, the candidate commodities are n commodities in an e-commerce database, and the expression is as follows:
Figure FDA00039021231600000613
wherein,
Figure FDA00039021231600000614
is s h Transposing; similarity score vector z using softmax function i And (4) carrying out normalization, wherein the expression is as follows:
y i ＝softmax(z i )
and sorting the commodities according to the normalized value of the similarity score vector from high to low, and recommending K-bit commodities before ranking to the user, wherein K is a positive integer.
7. An e-commerce conversation recommendation system based on multi-behavior feature fusion is characterized by comprising a behavior feature extraction module, a commodity high-order feature extraction module, a conversation feature extraction module and a commodity recommendation module;
the behavior feature extraction module is used for extracting feature vectors of four behaviors of clicking, collecting, purchasing and adding a shopping cart to a commodity by a user in a conversation; the click behavior is a primary behavior, and the three behaviors of collection, purchase and addition to the shopping cart are secondary behaviors generated based on the click behavior;
the commodity feature extraction module consists of a hierarchical gate control circulation unit network and is used for fusing the extracted four behavior feature vectors and outputting commodity feature vectors; the hierarchical gated cyclic unit network is composed of three levels of gated cyclic unit networks:
the first-level gating circulation unit network takes the collection behavior feature vector and the purchasing behavior feature vector as input, adaptively learns the importance weights of the collection behavior feature vector and the purchasing behavior feature vector, and performs weighted summation on the importance weights to obtain a fusion feature vector of the collection behavior feature and the purchasing behavior feature;
the second-level gating circulation unit network takes the fusion feature vector of the collection and purchase behavior features and the behavior feature vector of the shopping cart as input, adaptively learns the importance weights of the two feature vectors, and performs weighted summation on the importance weights to obtain the fusion feature vector of the secondary behavior;
the third-level gating circulation unit network takes the fusion characteristic vector and the click behavior characteristic vector of the secondary behavior as input, adaptively learns the importance weight of the two characteristic vectors, and performs weighted summation on the importance weight to obtain a commodity characteristic vector;
the commodity high-order feature extraction module is composed of a gated graph neural network, and extracts commodity high-order feature vectors by taking an adjacency matrix constructed by the click sequence relation of users to commodities in a session and the commodity feature vectors as input;
the session feature extraction module is composed of an attention module and a gate control circulation unit network, and is used for performing self-adaptive fusion on the current interest feature vector and the global interest feature vector and outputting a session feature vector;
the commodity recommending module calculates the similarity between the conversation characteristic vector and the candidate commodity characteristic vector through vector dot product operation, sorts the conversation characteristic vector and the candidate commodity characteristic vector according to the similarity from high to low, and recommends K-bit commodities with the similarity ranking top to the user.
8. An e-commerce session recommendation apparatus based on multi-behavior feature fusion, comprising at least one computing device, wherein the computing device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and when the computer program is loaded into the processor, the e-commerce session recommendation apparatus based on multi-behavior feature fusion according to any one of claims 1 to 7 is implemented.
9. A computer-readable storage medium, wherein the computer-readable storage medium has a computer program stored thereon, and when the computer program is executed by a processor, the computer program implements a multi-behavior feature fusion based e-commerce session recommendation method according to any one of claims 1-7.
CN202211299136.8A 2022-10-21 2022-10-21 E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion Pending CN115659277A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202211299136.8A CN115659277A (en) 2022-10-21 2022-10-21 E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202211299136.8A CN115659277A (en) 2022-10-21 2022-10-21 E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion

Publications (1)

Publication Number Publication Date
CN115659277A true CN115659277A (en) 2023-01-31

Family

ID=84988606

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202211299136.8A Pending CN115659277A (en) 2022-10-21 2022-10-21 E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion

Country Status (1)

Country Link
CN (1) CN115659277A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117994008A (en) * 2024-04-03 2024-05-07 深圳市灵智数字科技有限公司 Mobile commodity recommendation method and device, electronic equipment and storage medium

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117994008A (en) * 2024-04-03 2024-05-07 深圳市灵智数字科技有限公司 Mobile commodity recommendation method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN110956497B (en) Method for predicting repeated purchasing behavior of user of electronic commerce platform
Wu et al. Session-based recommendation with graph neural networks
Xiao et al. Quantifying uncertainties in natural language processing tasks
CN109299396B (en) Convolutional neural network collaborative filtering recommendation method and system fusing attention model
CN110503531B (en) Dynamic social scene recommendation method based on time sequence perception
CN110717098B (en) Meta-path-based context-aware user modeling method and sequence recommendation method
CN111581520A (en) Item recommendation method and system based on item importance in session
Liu et al. Collaborative filtering with a deep adversarial and attention network for cross-domain recommendation
CN110516160A (en) User modeling method, the sequence of recommendation method of knowledge based map
CN115082147B (en) Sequence recommendation method and device based on hypergraph neural network
CN108665311B (en) Electric commercial user time-varying feature similarity calculation recommendation method based on deep neural network
Wang et al. Data mining of mass storage based on cloud computing
Li et al. A CTR prediction model based on user interest via attention mechanism
CN112100439B (en) Recommendation method based on dependency embedding and neural attention network
CN111178986A (en) User-commodity preference prediction method and system
CN114117229A (en) Project recommendation method of graph neural network based on directed and undirected structural information
CN114925270B (en) Session recommendation method and model
Gicić et al. Intelligent credit scoring using deep learning methods
CN115659277A (en) E-commerce session recommendation method, system, device and medium based on multi-behavior feature fusion
Cao et al. Implicit user relationships across sessions enhanced graph for session-based recommendation
Ren et al. A co-attention based multi-modal fusion network for review helpfulness prediction
Qin et al. A hybrid deep learning model for short‐term load forecasting of distribution networks integrating the channel attention mechanism
CN116452269A (en) Dynamic graph neural network recommendation method based on multiple behaviors
CN116757747A (en) Click rate prediction method based on behavior sequence and feature importance
Chen et al. An Optimized BP Neural Network Model and Its Application in the Credit Evaluation of Venture Loans

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