Multi-behavior recommendation method for purchase preference perception based on intention

Abstract

The invention discloses an intention-based purchase preference perception multi-behavior recommendation method, which relates to the technical field of recommendation algorithms, and comprises the following steps of: obtaining a user set, a product set and a behavior set, and constructing a bigraph of each behavior; removing an interaction edge between the user and the purchased product in the bipartite graph of the non-target behavior; learning corresponding user characterization and product characterization from the bipartite graph of each behavior by using a depth map neural network; based on the user characterization and the product characterization under the target behavior, learning a user fusion weight and a product fusion weight under each non-target behavior through a deep neural network, and respectively fusing the user characterization and the product characterization under each non-target behavior to obtain a final user characterization and a final product characterization; calculating a loss function under each non-target behavior based on the final user representation and the final product representation; and carrying out multi-behavior intention learning in combination with the loss function under each non-target behavior so as to update model parameters to the convergence of an overall optimization target, thereby obtaining optimal parameters for realizing product recommendation meeting the intention of the user. According to the method, self-adaptive weight learning based on user purchase preference perception and behavior context perception is designed, multiple intention representation fusion weights meeting user personalization are obtained, and therefore products better meeting the future purchase intention of the user are recommended.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of recommendation algorithm, and in particular to a multi-behavior recommendation method based on intention-aware purchase preference. BACKGROUND

[0002] As a core component of modern information systems, recommendation systems have important application value in the field of e-commerce. E-commerce platforms based on personalized recommendation algorithms can optimize user experience and improve business efficiency by predicting user preferred items. Although traditional collaborative filtering-based recommendation systems perform well in single behavior type recommendation tasks, in actual e-commerce scenarios, due to limited user purchase records and large-scale product categories, the system faces serious data sparsity problems.

[0003] Notably, user-item interaction behaviors in e-commerce platforms have multi-dimensional characteristics, including browsing, adding to cart, and other interactive ways in addition to purchase behavior. These multi-behavior interaction data not only effectively alleviate the data sparsity problem caused by insufficient purchase records, but also fully depict user interest preferences. Existing technical solutions usually take purchase behavior as the target behavior and model other non-target behaviors as auxiliary features. This method does not fully consider the user purchase intention revealed by non-target behaviors, thus limiting the recommendation performance. Therefore, how to accurately identify potential purchase intention from user's multi-behavior records has become a key technical problem to improve the performance of multi-behavior recommendation systems. SUMMARY

[0004] To overcome the above-mentioned defects in the prior art, the present application provides a multi-behavior recommendation method based on intention-aware purchase preference, which encourages the model to focus on user intention information in non-target behaviors and user personalized purchase preferences to adaptively integrate user intention information under non-target behaviors.

[0005] To achieve the above-mentioned purposes, the present application adopts the following technical solutions, comprising:

[0006] A multi-behavior recommendation method based on intention-aware purchase preference, comprising:

[0007] S1, obtaining a user set, a product set, and a behavior set, constructing a bipartite graph for each behavior, in which the user and the product are nodes, and the interaction between the user and the product is an edge; the behavior includes non-target behavior and target behavior, and the target behavior refers to the behavior of purchasing a product;

[0008] S2, removing the interaction edge between the user and the purchased product in the bipartite graph of the non-target behavior;

[0009] S3, learning corresponding user representation and product representation from each behavior bipartite graph by using deep graph neural network;

[0010] S4, learning user fusion weight and product fusion weight under each non-target behavior by deep neural network based on user representation and product representation under target behavior, and respectively used for fusing user representation and product representation under each non-target behavior to obtain final user representation and final product representation;

[0011] S5, calculating loss function under each non-target behavior based on final user representation and final product representation;

[0012] S6, performing multi-behavior intention learning by combining loss function under each non-target behavior to update model parameters to overall optimization target convergence, thereby obtaining optimal parameters for realizing product recommendation meeting user intention.

[0013] Preferably, step S1 is specifically as follows:

[0014] User set U={u1,…u a ,…u M}, u a represents the a-th user, and M represents the total number of users;

[0015] Product set V={v1,…,v i ,…,v N}, v i represents the i-th product, and N represents the total number of products;

[0016] Behavior set B={b1,…,b j ,…,b K}, b j represents the j-th behavior, and K represents the total number of behaviors, wherein {b1,…,b j ,…,b K-1} represents non-target behavior, and b K represents target behavior;

[0017] Let represent the interaction data of the a-th user u a to the i-th product v i under the behavior b j , then the user-product interaction matrix under the behavior b j is If the a-th user u j has interaction record to the i-th product v a under the behavior b i , then Otherwise,

[0018] According to behavior b j The user-product interaction matrix R of behavior b j , with users and products as nodes and interaction records as edges, to construct a user-product bipartite graph G j of behavior b j = <U, V, R j >.

[0019] Preferably, step S2 is specifically as follows: for non-target behaviors {b1, …, b j , …, b K-1}, if the a-th user u a has interaction with the i-th product v i under the target behavior b K , then set the interaction matrix R j in the non-target behavior to

[0020] Preferably, step S3 is specifically as follows:

[0021] S31, initialize the user representation matrix P = {p1, …, p a , …, p M} shared by all behaviors, where p a represents the representation of the a-th user u a ; initialize the product representation matrix Q = {q1, …, q i , …, q N} shared by all behaviors, where q i represents the representation of the i-th product v i ;

[0022] S32, calculate the connection matrix A j of behavior b j :

[0023]

[0024] where R j is the user-product interaction matrix of behavior b j , if the a-th user u a has interaction record with the i-th product v i under behavior b j , then otherwise, M represents the total number of users, and N represents the total number of products;

[0025] S33, define that the graph neural network has L convolutional layers, the current convolutional layer is l, l = 0, 1, …, L; initialize the 0-th layer node representation matrix E of behavior b j .j,0 ={P,Q};

[0026] S34, will the behavior b j The representation matrix E of the l-th layer nodes j,l The input is fed into a graph neural network, and behavior b is calculated. j The node representation matrix E of the (l+1)th layer below j,l+1 :

[0027]

[0028] In the formula, D j Represents the connection matrix A j The degree matrix;

[0029] S35, aggregate the outputs of each convolutional layer and obtain behavior b. j The final representation matrix E j :

[0030]

[0031] S36, obtaining behavior b j The user representation matrix X learned below j and product characterization matrix Y j :

[0032] X j =E j [:M]

[0033] Y j =E j [M:]

[0034] In the formula, E j [:M] represents matrix E j Lines 1 to M, E j [M:] represents matrix E j The (M+1) to (M+N)th rows.

[0035] Preferably, step S4 is as follows:

[0036] S41, calculate the a-th user u a In non-target behavior b j Representation fusion weights under And the i-th product v i In non-target behavior b j Representation fusion weights under

[0037]

[0038] wherein j ∈ {1,..., K-1} represents a non-target behavior, K represents a target behavior; σ is a Sigmoid activation function; represents a non-target behavior b j the learned user representation matrix X j represents the representation vector of the a-th user in X represents a target behavior b K the learned user representation matrix X K represents the representation vector of the a-th user in X represents a non-target behavior b j the learned product representation matrix Y j represents the representation vector of the i-th product in Y represents a target behavior b K the learned product representation matrix Y K represents the representation vector of the i-th product in Y g and b g respectively represent a feature transformation matrix and a bias matrix in a gating network;

[0039] S42, the non-target behavior b j the learned user representation matrix X j and the product representation matrix Y j are fused by adaptive fusion weights, i.e. to obtain a final user representation matrix X and a final product representation matrix Y:

[0040]

[0041] wherein x a represents the final representation vector of the a-th user in the final user representation matrix X; y i represents the final representation vector of the i-th product in the final product representation matrix Y.

[0042] Preferably, step S5 is specifically as shown below:

[0043] S51, a heterogeneity standard loss function is used to calculate the loss of each non-target behavior b j , upper and lower limit values are introduced to indicate the selection criteria of the non-sampling strategy, and user heterogeneity matrix H and product heterogeneity matrix F are introduced to approximate the upper limit value and the lower limit value

[0044]

[0045] wherein, is the a-th user u j and the i-th product v a under the non-target behavior b iupper bound value of interaction probability of the a-th user u is the behavior b j the a-th user u a and the i-th product v i lower bound value of interaction probability; α is a hyper-parameter for controlling the proportion of upper and lower bound values, 0 < α < 1; the user heterogeneity matrix H corresponds to user traits, and the product heterogeneity matrix F corresponds to product attributes; is the user trait representation of the a-th user u j under the behavior b a ; is the product attribute representation of the i-th product v i under the behavior b j ;

[0046] S52, predicting the interaction probability of the a-th user u a to the i-th product v i ;

[0047]

[0048] wherein x a represents the final representation vector of the a-th user; y i represents the final representation vector of the i-th product;

[0049] S53, calculating the loss function L j under the non-target behavior b j :

[0050]

[0051] wherein D a is the total training data of the a-th user u a , represents the set of interaction products of the a-th user u a , represents the set of non-interaction products of the a-th user u a ; w is a hyper-parameter representing the loss weight of non-interaction data.

[0052] Preferably, step S6 is specifically as shown below:

[0053] S61, establishing an overall optimization objective L(θ):

[0054]

[0055] wherein θ represents the model parameters to be optimized; γ j is a hyper-parameter for adjusting the loss weight of each behavior; L j is the loss function under the non-target behavior b j ;

[0056] S62, solving the overall optimization target L(θ) by a gradient descent method to update the parameters θ until L(θ) converges to a minimum value, so as to obtain the optimal parameters θ * ;

[0057] S63, under the optimal parameters θ * , predicting the optimal interaction probability of the a th user u a to the i th product v i ; So as to obtain the optimal interaction matrix predicted by the user set U to the product set V for realizing product recommendation meeting the user intention, The greater the value is, the stronger the user intention is.

[0058] The optimal interaction probability The calculation method is as follows:

[0059]

[0060] In the formula, The optimal final feature vector of the a th user u a , The optimal final feature vector of the i th product v i .

[0061] The application further provides a readable storage medium, which has a computer program stored thereon, and the computer program is executed to realize the multi-behavior recommendation method based on intention-based purchase preference perception.

[0062] The application further provides an electronic device, which comprises a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to realize the multi-behavior recommendation method based on intention-based purchase preference perception.

[0063] The application further provides a computer program product, which comprises computer programs / instructions, and the computer programs / instructions are executed by a processor to realize the multi-behavior recommendation method based on intention-based purchase preference perception.

[0064] The application has the following advantages:

[0065] (1) The multi-behavior recommendation method based on intention-based purchase preference perception considers the user purchase intention semantics contained in the non-target behavior in multi-behavior data, designs adaptive weight learning based on user purchase preference perception and behavior context perception, obtains multi-intention representation fusion weights meeting user personalization, and thus recommends products more meeting the user future purchase intention.

[0066] (2) The application proposes a multi-behavior recommendation method based on intention-based purchase preference perception. The method extracts intention information in non-target behavior by focusing on user interaction but incomplete purchase conversion, learns user purchase preference information through user historical purchase records, and guides the adaptive fusion of multiple user intention semantics. The method can effectively utilize non-target behavior data and improve the performance of multi-behavior recommendation.

[0067] (3) The multi-behavior method based on intention-based purchase preference perception proposed by the application is very lightweight, has strong practical value, and performs superior performance on multiple data sets, and has strong application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 A flowchart of a multi-behavior recommendation method based on intention-based purchase preference perception of the application.

[0069] Figure 2 A schematic diagram of a multi-behavior recommendation method based on intention-based purchase preference perception of the application. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0071] Embodiment 1

[0072] As shown in Figure 1 and Figure 2 , the multi-behavior recommendation method based on intention-based purchase preference perception of the application considers the user purchase intention semantics contained in the non-target behavior in multi-behavior data, designs adaptive weight learning based on user purchase preference perception and behavior context perception, obtains multi-intention representation fusion weight meeting user personalization, and thus recommends products more consistent with user future purchase intention. The method of the application includes the following steps:

[0073] S1, constructing original data, including: user interaction matrix of product.

[0074] Let U represent the user set, and U = {u1,…u a ,…u M}, u a represents the a-th user, M represents the total number of users, and 1≤a≤M.

[0075] Let V denote the product set, and V = {v1, ..., v}. i ,…,v N}, v i Let i represent the i-th product, and N represent the total number of products, where 1 ≤ i ≤ N.

[0076] Let B denote the set of rows, and B = {b1, ..., b}. j ,…,b K}, b j Let {b1, ..., bj} represent the j-th action, and K represent the total number of actions, where {b1, ..., bj} are all actions. j ,…,b K-1} represents non-target behavior, j∈{1,...,K}, b K Indicates the target behavior.

[0077] make Represents the a-th user u a For the i-th product v i In behavior b j The interaction data below will then lead to behavior b. j The following is a user interaction matrix for the product. If in behavior b j Next user a a For the i-th product v i If there is an interaction record, then otherwise,

[0078] Based on user behavior regarding the product (b) j The interaction matrix R under j Using users and products as nodes and their interaction records as edges, behavior b is constructed. j The user-product section of the diagram G j = <U,V,R j >; Use G = {G1, ..., G} j ,…,G K} represents a user-product bipartite diagram under multiple behaviors.

[0079] S2. Process the original data to remove the interaction edges with purchased products in the bipartite graph under non-target behavior, and obtain the intentional interaction bipartite graph G' under the user's non-target behavior.

[0080] Using the bipartite graph G under the target behavior K Data processing is performed on the bipartite graph under non-target behavior. For the bipartite graph G under non-target behavior... j ∈G, where j∈{1,...,K-1}, if the a-th user u a For the i-th product v i In target behavior b KIf there is interaction under the non-target behavior, then the bipartite graph G j under the non-target behavior is obtained j j j

[0081] j j j K

[0082] In the formula, R K represents the interaction matrix under the target behavior, and represents the matrix point multiplication.

[0083] In this embodiment, training and testing are performed on real data sets Beibei, Taobao and Tmall, and the data set is divided. Specifically, the latest purchase interaction of each user in the data set is used as the test, and all the remaining interaction data is used for training.

[0084] S3, using a deep graph neural network and a share-bottom framework, learning user and product representation matrices under multiple views, obtaining user representation matrices X j and product representation matrices Y j learned from behavior b j .

[0085] S31, randomly initialize all behavior-shared user representation matrices P={p1,...,p a ,...,p M} using a Gaussian distribution, wherein p a represents the representation of the a-th user u a ; randomly initialize all behavior-shared product representation matrices Q={q1,...,q i ,...,q N} using a Gaussian distribution, wherein q i represents the representation of the i-th product v i .

[0086] S32, calculate the connection matrix A j of behavior b j :

[0087]

[0088] S33, define that the graph neural network has L convolutional layers, the current convolutional layer is l, l=0,1,..,L. Initialize the 0th layer node representation matrix E j of behavior b j,0 ={P,Q}.​​​​​​​​

[0089] S34, will the behavior b j The representation matrix E of the l-th layer nodes j,l The input is fed into a graph neural network, and behavior b is calculated. j The node representation matrix E of the (l+1)th layer below j,l+1 :

[0090]

[0091] In the formula, D j Represents the connection matrix A j The degree matrix.

[0092] S35, aggregate the outputs of each convolutional layer and obtain behavior b. j The final representation matrix E j :

[0093]

[0094] S36, obtaining behavior b j The user representation matrix X learned below j and product characterization matrix Y j :

[0095] X j =E j [:M]

[0096] Y j =E j [M:]

[0097] In the formula, E j [:M] represents matrix E j Lines 1 to M, E j [M:] represents matrix E j The (M+1) to (M+N)th rows.

[0098] S4, based on user purchase preference representation, learns adaptive fusion weights for multiple intent representations through a deep neural network, and uses the learned weights to fuse behavioral representation X. j and Y j And obtain the end-user representation matrix X and the end-product representation matrix Y.

[0099] S41, calculate the a-th user u a In behavior b j Representation fusion weights And the i-th product v i In behavior b j Representation fusion weights

[0100]

[0101] wherein, j ∈ {1,...,K-1} represents a non-target behavior, K represents a target behavior; σ is a Sigmoid activation function; represents a non-target behavior b j a learned user representation matrix X j , wherein the a-th row representation vector in the learned user representation matrix X represents a target behavior b K a learned user representation matrix X K , wherein the a-th row representation vector in the learned user representation matrix X represents a non-target behavior b j a learned product representation matrix Y j , wherein the i-th row representation vector in the learned product representation matrix Y represents a target behavior b K a learned product representation matrix Y K , wherein the i-th row representation vector in the learned product representation matrix Y g and b g respectively represent a feature conversion matrix and a bias matrix in a gating network.

[0102] S42, according to the following formula, the non-target behavior b j a learned intention representation X j , Y j is fused to obtain a final user representation matrix X and a final product representation matrix Y:

[0103]

[0104] wherein, x a represents the a-th row representation vector in the final user representation matrix X, i.e., the final representation vector of the a-th user; y i represents the i-th row representation vector in the final product representation matrix Y, i.e., the final representation vector of the i-th product.

[0105] S5, based on the final user representation matrix X and the final product representation matrix Y, a loss function L j under the non-target behavior b j is calculated.

[0106] S51, a heterogeneity standard loss function is used to calculate the loss under each behavior, for the non-target behavior b j , an upper limit and a lower limit domain value are introduced to indicate the selection standard of the non-sampling strategy, a user heterogeneity matrix H and a product heterogeneity matrix F are introduced to approximate the upper domain value and the lower domain value ​

[0107]

[0108] wherein j ∈ {1,..., K-1} represents non-target behavior; is the non-target behavior b j the a-th user u a and the i-th product v i interaction probability of the upper bound value, is the behavior b j the a-th user u a and the i-th product v i interaction probability of the lower bound value; a is a hyperparameter for controlling the upper and lower bound value ratio, 0 < a < 1; the user heterogeneity matrix H ∈ R |M|×D corresponds to the user trait, the product heterogeneity matrix F ∈ R |N|×D corresponds to the product attribute, D represents the dimension of the user trait and the product attribute, in the user heterogeneity matrix H, H a represents the user trait representation of the a-th user u a , in the product heterogeneity matrix F, F i represents the product attribute representation of the i-th product v i ; each behavior b j has a corresponding user heterogeneity matrix H and product heterogeneity matrix F, represents the user trait representation of the a-th user u j under the behavior b a ; represents the product attribute representation of the i-th product v j under the behavior b i .

[0109] S52, predicting the interaction probability of the a-th user u a to the i-th product v i

[0110]

[0111] wherein x a represents the final representation vector of the a-th user; y i represents the final representation vector of the i-th product.

[0112] S53, calculating the loss function L j under the non-target behavior b j :

[0113]

[0114] wherein j ∈ {1,..., K-1} represents non-target behavior; D​a It is the a-th user u a All training data, Represents the a-th user u a A collection of interactive products Represents the a-th user u a The set of non-interactive products; w is a hyperparameter representing the loss weight of non-interactive data.

[0115] S6 combines various loss functions to perform multi-behavior intent learning, updating model parameters until the overall optimization objective converges, thereby obtaining the optimal parameters to achieve product recommendations that meet user intent.

[0116] S61, Establish the overall optimization objective L(θ):

[0117]

[0118] In the formula, θ=[P,Q,H,F,W g ,b g [] represents all parameters to be optimized, γ j These are hyperparameters used to adjust the loss weights for each behavior.

[0119] S62, the overall optimization objective L(θ) is solved using the gradient descent method to update the parameter θ until L(θ) converges to the minimum value, thus obtaining the optimal parameter θ. * ;

[0120] S63, Predicting the optimal parameter θ * Next user a a For the i-th product v i Optimal interaction probability This yields the optimal interaction matrix for predicting the product set V from the user set U. Used to implement product recommendations that meet user intent. The larger the value, the stronger the user's intent;

[0121] Optimal interaction probability The calculation method is as follows:

[0122]

[0123] In the formula, Represents the a-th user u a The optimal final representation vector, v represents the i-th product i The optimal final representation vector.

[0124] Example 2

[0125] The embodiment provides an electronic device, comprising a memory and a processor, the memory is used for storing a program supporting the processor to execute the method of the embodiment 1, and the processor is configured to execute the program stored in the memory.

[0126] Embodiment 3

[0127] The embodiment provides a computer readable storage medium, and a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to perform the steps of the method of the embodiment 1.

[0128] Embodiment 4

[0129] In order to verify the effectiveness of the method and evaluate the performance of the method, the application selects three data sets commonly used in the field of multi-behavior recommendation system, namely Beibei, Taobao and Tmall data sets, performs experiments, divides the latest purchase behavior of the user into a test set, and uses the remaining interaction data as a training set. In the test set, all the products not purchased by the user are selected for sorting when testing, and hit ratio (HR@K) and normalized discounted cumulative gain (NDCG@K) are used as evaluation indexes of multi-behavior recommendation performance, and the higher the index value is, the higher the recommendation accuracy is. Wherein, K represents that the products with the top K interaction probabilities are recommended.

[0130] The application selects five related multi-behavior recommendation methods for effect comparison, which are PKEF, GHCF, MBSSL, HPMR and CHCF. Specifically, tables 1-3 respectively show the experimental results on the above three data sets.

[0131] Table 1 comparison of recommendation results of the method and the comparison method of the application on the Beibei data set

[0132] Models HR@10 NDCG@10 HR@50 NDCG@50 PKEF 0.1130 0.0582 0.2173 0.0650 GHCF 0.1922 0.1012 0.3794 0.1426 MBSSL 0.2229 0.1277 0.3806 0.1626 HPMR 0.2322 0.1332 0.3814 0.1664 CHCF 0.2436 0.1340 0.4324 0.1755 PAIF (Invention) 0.4394 0.2260 0.5781 0.2599

[0133] Table 2 comparison of recommendation results of the method and the comparison method of the application on the Taobao data set

[0134]

[0135]

[0136] Table 3 comparison of recommendation results of the method and the comparison method of the application on the Tmall data set

[0137] Models HR@10 NDCG@10 HR@50 NDCG@50 PKEF 0.1385 0.0785 0.2797 0.1067 GHCF 0.0683 0.0414 0.1691 0.0553 MBSSL 0.0774 0.0464 0.1850 0.0649 HPMR 0.0790 0.0472 0.1874 0.0660 CHCF 0.1653 0.0892 0.3402 0.1245 PAIF (Invention) 0.1904 0.0988 0.4079 0.1469

[0138] From the above three tables, it can be seen that on the three data sets, the PAIF method proposed in the application is significantly better than the comparative method in the four indexes of HR@10, NDCG@10, HR@50 and NDCG@50.

[0139] The above merely describes preferred embodiments of the present application, and is not used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-behavior recommendation method based on intent-based purchase preference perception, characterized in that, Comprise: S1, obtain a user set, a product set, and a behavior set, construct a bipartite graph of each behavior, in which users and products are nodes, and interactions between users and products are edges; the behaviors include non-target behaviors and target behaviors, and the target behaviors refer to behaviors of purchasing products; S2, remove the interaction edges between users and purchased products in the bipartite graph of non-target behaviors; S3, learn corresponding user representations and product representations from the bipartite graph of each behavior by using a deep graph neural network; S4, learn user fusion weights and product fusion weights under each non-target behavior based on the user representations and product representations under the target behavior by using a deep neural network, and use the user fusion weights and product fusion weights to fuse the user representations and product representations under each non-target behavior respectively to obtain final user representations and final product representations; S5, calculate a loss function under each non-target behavior based on the final user representations and final product representations; S6, perform multi-behavior intention learning by combining the loss functions under each non-target behavior to update model parameters to convergence of an overall optimization target, thereby obtaining optimal parameters for realizing product recommendation meeting user intentions.

2. The method of claim 1, wherein the method is based on an intent-based purchase preference perception of a plurality of behaviors. Step S1 is specifically as follows: User set U = {u1,... u a M M}, u a denotes the athuser, and M denotes the total number of users. V = {v1,..., v i N N}, v i denotes the i-th product, and N denotes the total number of products; Behavior set B = {b1,..., b j K K}, b j represents the jth behavior, K represents the total number of behaviors, where {b1,..., b j K K-1} represents non-target behaviors, b K represents a target behavior; Let denote the a-th user u a for the i-th product v i in the behavior b j , the interaction data of the user in the behavior b j , the interaction matrix of the user in the behavior b if the a-th user u j has an interaction record with the i-th product v a in the behavior b i , then otherwise, According to behavior b j The user-product interaction matrix R j , with users and products as nodes and their interaction records as edges, construct the user-product bipartite graph G j under behavior b j = <U, V, R j >.

3. The method of claim 2, wherein the method is based on an intent-based purchase preference perception of a plurality of behaviors. Step S2 is specified as follows: for non-target behaviors {b1,..., b j}, if the a-th user u K-1 has an interaction with the i-th product v a in the target behavior b i , then the interaction matrix R K in the non-target behavior is set to 1. j ​ 4. The method of claim 1, wherein the method is based on an intent-based purchase preference perception of a plurality of behaviors. Step S3 is specifically as follows: S31, Initialize the user representation matrix P = {p1,...,p...} shared by all behaviors. a ,...,p M }, where p a Represents the a-th user u a Representation; Initialize the product representation matrix Q = {q1,...,q} shared by all behaviors. i ,...,q N }, where q i v represents the i-th product i The representation; S32, calculate behavior b j Connection matrix A of j : In the formula, R j For behavior b j Below is the user interaction matrix for the product. If in behavior b j Next user a a For the i-th product v i If there is an interaction record, then otherwise, M represents the total number of users, and N represents the total number of products; S33, define that the graph neural network has L convolution layers, the current convolution layer is l, l = 0, 1,.., L; initialize the behavior b j The 0th layer node representation matrix E j,0 = {P, Q}; S34, the behavior b j under the first layer node representation matrix E j,l is input into the graph neural network, and the behavior b j under the first +1 layer node representation matrix E j,l+1 : where D j denotes the degree matrix of the connection matrix A j ; S35, aggregate the outputs of each convolutional layer and obtain a behavior b j under the final representation matrix E j : S36, obtaining behavior b j The learned user representation matrix X j and the product representation matrix Y j : X j = E j [:M] Y j = E j [M:] wherein E j [M] represents the first to M rows of the matrix E j E j [M:] represents the (M+1) to (M+N) rows of the matrix E j ​ 5. The method as claimed in claim 1, wherein, Step S4 is specifically as follows: S41, calculating the a-th user u a under the non-target behavior b j under the non-target behavior b and the i-th product v i under the non-target behavior b j under the non-target behavior b where j ∈ {1,..., K-1} denotes non-target behaviors, K denotes target behaviors; σ is a Sigmoid activation function; denotes non-target behaviors b j the learned user representation matrix X j the representation vector of the a-th user in X denotes target behaviors b K the learned user representation matrix X K the representation vector of the a-th user in X denotes non-target behaviors b j the learned product representation matrix Y j the representation vector of the i-th product in Y denotes target behaviors b K the learned product representation matrix Y K the representation vector of the i-th product in Y g and b g denote the feature transformation matrix and the bias matrix in the gating network, respectively S42, non-target behavior b j The learned user representation matrix X j And product representation matrix Y j Through adaptive fusion weight, that is, Fusion, get the final user representation matrix X and the final product representation matrix Y: where x a represents the final representation vector of the a-th user in the final user representation matrix X; y i represents the final representation vector of the i-th product in the final product representation matrix Y.

6. The method of claim 1, wherein the method is based on an intent-based purchase preference perception of a multi-behavior recommendation. Step S5 is specifically as follows: S51, calculate the loss of each non-target behavior b using the heterogeneity standard loss function j , introduce upper and lower limit domain values to indicate the selection criteria of the non-sampling strategy, introduce the user heterogeneity matrix H and the product heterogeneity matrix F to approximate the upper domain value and the lower domain value wherein, is the non-target behavior b j the a-th user u a and the i-th product v i the upper bound value of the interaction probability, is the lower bound value of the interaction probability of the a-th user u j and the i-th product v a under behavior b i ; a is a hyper-parameter for controlling the proportion of upper and lower bound values, 0 < a < 1; the user heterogeneity matrix H corresponds to user characteristics, and the product heterogeneity matrix F corresponds to product attributes; represents the user characteristic representation of the a-th user u j under behavior b a ; represents the product attribute representation of the i-th product v j under behavior b i ; S52, predict the a-th user u a the interaction probability of the i-th product v i ​ where x a represents the final representation vector of the a-th user; y i represents the final representation vector of the i-th product; S53, calculate non-target behavior b j loss function L j : where D a is the total training data of the a-th user u a , is the interacted product set of the a-th user u a , is the non-interacted product set of the a-th user u a ; w is a hyper-parameter, representing the loss weight of non-interacted data.

7. The multi-behavior recommendation method of intent-based purchase preference perception according to claim 1, characterized in that, Step S6 is specifically as follows: S61, establish an overall optimization target L(θ): In the formula, θ represents a model parameter to be optimized; γ j is a hyperparameter, used to regulate the loss weight of each behavior; L j is a loss function of a non-target behavior b j . S62, solve the overall optimization objective L(0) by gradient descent method to update the parameter 0 until L(0) converges to a minimum value, thereby obtaining the optimal parameter 0 * ; S63, in the optimal parameter θ * Next, the optimal interaction probability of the a-th user u a for the i-th product v i Thus, the optimal interaction matrix predicted by the user set U for the product set V is obtained for realizing product recommendation meeting user intention, The greater the value, the stronger the user intention.​ optimal interaction probability is calculated as follows: wherein, represents the optimal final representation vector of the a-th user u a represents the optimal final representation vector of the i-th product v i .​ 8. A readable storage medium, characterized by, The computer program stored thereon, when executed, implements the multi-behavior recommendation method based on intention-based purchase preference perception according to any one of claims 1-7.

9. An electronic device, comprising: It includes a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor implements the multi-behavior recommendation method based on intention-based purchase preference perception according to any one of claims 1-7 when executing the computer program.

10. A computer program product, characterised in that, It includes computer programs / instructions, which, when executed by a processor, implement the multi-behavior recommendation method based on intention-based purchase preference perception according to any one of claims 1-7.

Images