Collaborative filtering recommendation method and system based on decoupled dual-channel hypergraph neural network

Abstract

The present application proposes a collaborative filtering recommendation method and system based on a decoupled dual-channel hypergraph neural network, which relates to the technical field of recommendation systems, wherein the method includes: obtaining a user-item interaction graph, and randomly initializing to obtain user features and item features; using a dual-channel hypergraph construction method to construct a user hypergraph structure and an item hypergraph structure; using an intention-aware feature extraction module to obtain user feature representation and item feature representation; using multiple superimposed hypergraph decoupling modules to obtain new user features and new item features; performing feature fusion on the new user features and item features to obtain a user final feature representation and an item final feature representation; using the dot product result of the user final feature representation and the item final feature representation to represent the user's preference for the item, and the larger the dot product result, the greater the user's preference for the item. The present invention models the various potential intentions that exist when users interact with items, and more accurately describes the relationship between users and items.

Description

Technical Field

[0001] The present application relates to the technical field of recommendation systems, and in particular to a collaborative filtering recommendation method and system based on a decoupled dual-channel hypergraph neural network. Background Art

[0002] As an effective way to solve the problem of information overload, recommendation systems have been widely used in fields such as e-commerce. In recommendation systems, collaborative filtering based on user historical behavior information is one of the widely used methods. For collaborative filtering, how to accurately model the relationship between users and items and mine collaborative information is a key issue. In recent years, due to the stronger expressiveness and flexibility of hypergraph neural networks in modeling complex data associations, some researchers have proposed collaborative filtering algorithms based on hypergraph neural networks. However, existing methods fail to fully consider the different intentions of their interactions when modeling the relationship between users and items, and it is difficult to accurately describe the relationship between users and items. Summary of the invention

[0003] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0004] To this end, the first purpose of the present application is to propose a collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network, which solves the technical problem that the existing methods fail to fully consider the different intentions that exist during their interaction when modeling the relationship between users and items, and it is difficult to accurately describe the relationship between users and items. It realizes the use of multiple probabilistic hypergraphs to model the various potential intentions that exist when users interact with items, and by proposing a hypergraph decoupling module, the hypergraph structures and node features of different intentions are made as different as possible, thereby focusing on information under different intentions and more accurately describing the relationship between users and items; at the same time, the recommended items are filtered to improve the accuracy of the recommended items.

[0005] The second objective of this application is to propose a collaborative filtering recommendation system based on a decoupled dual-channel hypergraph neural network.

[0006] To achieve the above-mentioned purpose, the first embodiment of the present application proposes a collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network, including: obtaining a user-item interaction graph, and randomly initializing to obtain user features and item features; according to the user-item interaction graph, using a dual-channel hypergraph construction method to construct a user hypergraph structure and an item hypergraph structure; according to the user features and item features, using an intent-aware feature extraction module to obtain a user feature representation and an item feature representation; according to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation, using multiple superimposed hypergraph decoupling modules to obtain new user features and new item features; performing feature fusion on the new user features and the new item features to obtain a user final feature representation and an item final feature representation; using the dot product result of the user final feature representation and the item final feature representation to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item.

[0007] Optionally, in one embodiment of the present application, a dual-channel hypergraph construction method is used to construct a user hypergraph structure, including the following steps:

[0008] The k-order reachability matrix of an item is defined as:

[0009]

[0010] in, The adjacency matrix H is the user-item interaction graph, and M is the number of items;

[0011] Based on the k-order reachability matrix of the item, the corresponding user's hyperedge group adjacency matrix is ​​expressed as:

[0012]

[0013] in, N is the number of users, M is the number of items;

[0014] The first-order matrix and the second-order matrix in the user's hyperedge group adjacency matrix are used to concatenate the hyperedge groups to obtain the final user hypergraph. The user hypergraph is expressed as:

[0015]

[0016] in, is the first-order matrix in the user's hyperedge group adjacency matrix, is the second-order matrix in the user's hyperedge group adjacency matrix;

[0017] The user hypergraph structure is obtained according to the user hypergraph, and the user hypergraph structure is expressed as:

[0018]

[0019] Among them, Hu For users,

[0020] The dual-channel hypergraph construction method is used to construct the item hypergraph structure, including the following steps:

[0021] The k-order reachable matrix of a user is defined as:

[0022]

[0023] in, The adjacency matrix H is the user-item interaction graph, M is the number of items, and N is the number of users;

[0024] Based on the user's k-order reachability matrix, the corresponding hyperedge group adjacency matrix of the item is expressed as:

[0025]

[0026] in,

[0027] The first-order matrix and the second-order matrix in the hyperedge group adjacency matrix of the item are used to concatenate the hyperedge groups to obtain the final item hypergraph. The item hypergraph is expressed as:

[0028]

[0029] in, is the first-order matrix in the user's hyperedge group adjacency matrix, is the second-order matrix in the user's hyperedge group adjacency matrix;

[0030] The item hypergraph structure is obtained according to the item hypergraph, and the item hypergraph structure is expressed as:

[0031]

[0032] Among them, H i Super map for items.

[0033] Optionally, in one embodiment of the present application, an intention-aware feature extraction module is used to obtain user feature representation and item feature representation, which is specifically expressed as follows:

[0034]

[0035]

[0036]

[0037]

[0038] in, For user u j and its characteristics The potential features of the user in the kth intent, and is a trainable parameter, For item i t and its characteristics The potential features of the item in the kth intention, and is a trainable parameter, R is the real number domain, D is the feature dimension, K is the number of intents, and σ is the activation function.

[0039] Optionally, in one embodiment of the present application, multiple superimposed hypergraph decoupling modules are used to obtain new user features and item features, including the following steps:

[0040] According to the user feature representation and the item feature representation, the non-zero elements in the corresponding user hypergraph structure and the item hypergraph structure are updated, and the updated user hypergraph structure and the item hypergraph structure are normalized to obtain a new user hypergraph structure and a new item hypergraph structure;

[0041] Perform hypergraph convolution operations on the user feature representation and the item feature representation according to the updated user hypergraph structure and the updated item hypergraph structure to obtain updated user features and item features;

[0042] A distance-related loss function is used to calculate the correlation loss between different intent features in the updated user features and item features.

[0043] Optionally, in one embodiment of the present application, non-zero elements in the corresponding user hypergraph structure and item hypergraph structure are updated according to the user feature representation and the item feature representation, and the updated user hypergraph structure and item hypergraph structure are represented as follows:

[0044]

[0045]

[0046] Where j and m are the numbers of the user and hyperedge respectively, and c i，k (m) is to construct the user hyperedge H u，k (:,m) is the number of the item used, a(·) is the function for calculating the association, and dot multiplication can be used. For user u j and its characteristics The potential features of the user in the kth intent, is the feature of the corresponding item used to construct the user hyperedge, c u，k (m) is the construction of hyperedge H i，k (:, m) when the central user number, The features of the corresponding users used to construct the item hyperedge, For item i t and its characteristics The potential features of the item in the kth intention;

[0047] The Softmax function is used to normalize the updated user hypergraph structure and item hypergraph structure across intents, which can be expressed as:

[0048]

[0049]

[0050] Among them, K is the number of intent features, is the updated user hypergraph structure, is the hypergraph structure of the t-th user, is the updated item hypergraph structure, is the hypergraph structure of the t-th item, j is the user number, and m is the hyperedge number.

[0051] Optionally, in one embodiment of the present application, the updated user features and item features are expressed as:

[0052]

[0053]

[0054] in, is the node degree matrix of the user hypergraph, H u，k represents the updated user hypergraph structure, is a trainable parameter, R is the real number domain, D is the feature dimension, K is the number of intents, is the hyper-edge degree matrix of the user hypergraph, E u，k (l) is the user feature representation, is the hyperedge degree matrix of the item hypergraph, H i，k represents the updated item hypergraph structure, is the hyper-edge degree matrix of the item hypergraph, E i，k (l) is the characteristic representation of the item,

[0055] The correlation loss between different intent features in the updated user features and item features is expressed as:

[0056]

[0057] in, and e k′ (l)is the feature representation of the user or item about intention k and intention k′ at the lth layer, K is the number of intentions, d(·) is the distance association, and d(·) is expressed as:

[0058]

[0059] Here, dCov(·) and dVar(·) represent covariance and variance, respectively.

[0060] Optionally, in one embodiment of the present application, feature fusion is performed on new user features and new item features, and the fused features are expressed as:

[0061]

[0062]

[0063] in, represents the characteristics of users in layer l, Represents the characteristics of items in layer l.

[0064] Optionally, in one embodiment of the present application, the dot product result is:

[0065]

[0066] in, represents the final feature representation of the user, Represents the final feature representation of the item.

[0067] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a collaborative filtering recommendation system based on a decoupled dual-channel hypergraph neural network, including: a first acquisition module, used to acquire user and item information, and randomly initialize to obtain user features, item features and a user-item interaction graph; a construction module, used to construct a user hypergraph structure and an item hypergraph structure according to the user-item interaction graph using a dual-channel hypergraph construction method; a second acquisition module, used to obtain a user feature representation and an item feature representation according to the user features and the item features using an intention-aware feature extraction module; a third acquisition module, used to obtain new user features and new item features according to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation using multiple superimposed hypergraph decoupling modules; a fusion module, used to perform feature fusion on the new user features and the new item features to obtain a user final feature representation and an item final feature representation; a determination module, used to use the dot product result of the user final feature representation and the item final feature representation to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item.

[0068] Optionally, in one embodiment of the present application, the system is trained, and the specific process is to adopt Bayesian personalized sorting and learn through the following function:

[0069]

[0070] Among them, μ is the weight coefficient of distance association loss, L is the distance correlation loss function, σ is the Sigmoid function, Θ is the model trainable parameter, λ is the regularization hyperparameter, o is the training set triplet set used for training, are the interactions that appeared in the training set, is an interaction that has not appeared in the training set, u is the user, i is the item clicked by user u in the training set, and j is the item not clicked by the user in the training set.

[0071] The collaborative filtering recommendation method and system based on a decoupled dual-channel hypergraph neural network in the embodiment of the present application use multiple probabilistic hypergraphs to model the various potential intentions that exist when users interact with items, and propose a hypergraph decoupling module to make the hypergraph structures and node features of different intentions as different as possible, thereby focusing on information under different intentions and achieving a more accurate description of the relationship between users and items; at the same time, the recommended items are filtered, and according to the user's preference for the items, the recommended items whose dot product results are less than a preset threshold are eliminated, thereby improving the accuracy of the recommended items and reflecting the superior performance of the collaborative filtering recommendation method and system based on a decoupled dual-channel hypergraph neural network.

[0072] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0074] Figure 1 A flowchart of a collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network provided in Example 1 of the present application;

[0075] Figure 2 This is another flowchart of the collaborative filtering recommendation method based on the decoupled dual-channel hypergraph neural network according to an embodiment of the present application. DETAILED DESCRIPTION

[0076] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0077] The following describes a collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network according to an embodiment of the present application with reference to the accompanying drawings.

[0078] Figure 1 This is a flowchart of a collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network provided in Example 1 of the present application.

[0079] like Figure 1 As shown, the collaborative filtering recommendation method based on the decoupled dual-channel hypergraph neural network includes the following steps:

[0080] Step 101, obtaining a user-item interaction graph, and randomly initializing to obtain user features and item features;

[0081] Step 102, constructing a user hypergraph structure and an item hypergraph structure using a dual-channel hypergraph construction method according to the user-item interaction graph;

[0082] Step 103, according to the user features and the item features, an intention-aware feature extraction module is used to obtain a user feature representation and an item feature representation;

[0083] Step 104, according to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation, a plurality of superimposed hypergraph decoupling modules are used to obtain new user features and new item features;

[0084] Step 105, performing feature fusion on the new user features and the new item features to obtain a final user feature representation and a final item feature representation;

[0085] Step 106, using the dot product result of the user's final feature representation and the item's final feature representation to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item.

[0086] Step 107, recommending the recommended items according to the user's preference for the items, wherein the recommended items are filtered according to the user's preference for the items, and the recommended items whose dot product results are less than a preset threshold are removed from the recommended items.

[0087] The collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network in the embodiment of the present application obtains a user-item interaction graph and randomly initializes it to obtain user features and item features; based on the user-item interaction graph, a dual-channel hypergraph construction method is used to construct a user hypergraph structure and an item hypergraph structure; based on the user features and item features, an intention-aware feature extraction module is used to obtain user feature representation and item feature representation; based on the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation, multiple superimposed hypergraph decoupling modules are used to obtain new user features and new item features; feature fusion is performed on the new user features and the new item features to obtain a user final feature representation and an item final feature representation; the dot product result of the user final feature representation and the item final feature representation is used to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item. Therefore, the technical problem that the existing methods fail to fully consider the different intentions existing in their interaction when modeling the relationship between users and items, and it is difficult to accurately describe the relationship between users and items can be solved. It realizes the use of multiple probabilistic hypergraphs to model the various potential intentions existing in the interaction between users and items, and by proposing a hypergraph decoupling module, the hypergraph structures and node features of different intentions are made as different as possible, so as to focus on the information under different intentions and more accurately describe the relationship between users and items; at the same time, the recommended items are filtered, and according to the user's preference for the items, the recommended items whose dot product results are less than the preset threshold are eliminated, which improves the accuracy of the recommended items and reflects the superior performance of the collaborative filtering recommendation method and system based on the decoupled dual-channel hypergraph neural network.

[0088] Furthermore, in the embodiment of the present application, given a set of N users and a set of M items Use and Represents the initialization features of users and items, where D is the feature dimension of users and items. Use A∈{0,1} N×M Represents the user-item interaction graph of the model input, where the non-zero element A i，j Indicates that there is an interaction between user i and item j. In addition, assume that the number of intentions is K. The interaction graph A is regarded as the adjacency matrix H.

[0089] The dual-channel hypergraph construction method is used to construct the user hypergraph structure, including the following steps:

[0090] The k-order reachability matrix of an item is defined as:

[0091]

[0092] in, The adjacency matrix H is the user-item interaction graph, and M is the number of items;

[0093] Based on the k-order reachability matrix of the item, the corresponding user's hyperedge group adjacency matrix is ​​expressed as:

[0094]

[0095] in, N is the number of users, M is the number of items;

[0096] The first-order matrix and the second-order matrix in the user's hyperedge group adjacency matrix are used to concatenate the hyperedge groups to obtain the final user hypergraph. The user hypergraph is expressed as:

[0097]

[0098] in, is the first-order matrix in the user's hyperedge group adjacency matrix, is the second-order matrix in the user's hyperedge group adjacency matrix;

[0099] The user hypergraph structure is obtained according to the user hypergraph, and the user hypergraph structure is expressed as:

[0100]

[0101] Among them, H u Hypergraph for users.

[0102] Similarly, for items, the user's k-order reachability matrix is ​​defined as:

[0103]

[0104] in,

[0105] Based on the user's k-order reachability matrix, the corresponding hyperedge group adjacency matrix of the item is:

[0106]

[0107] in,

[0108] The first-order and second-order reachable item information based on the user is used to construct the item hyperedge group, and the final item hypergraph H is obtained by splicing the hyperedge groups. i ,Right now:

[0109]

[0110] For intent k, we initialize it as:

[0111]

[0112] Furthermore, in the embodiment of the present application, an intention-aware feature extraction module is used to obtain user feature representation, which is specifically expressed as follows:

[0113]

[0114]

[0115] Among them, e uj，k For user u j and its characteristics The potential features of the user in the kth intent, and is a trainable parameter, R is the real number domain, D is the feature dimension, and K is the number of intents.

[0116] For item i t and its characteristics The potential features of the item in the kth intent are extracted in the following way:

[0117]

[0118]

[0119] Where σ is the activation function, and is a trainable parameter.

[0120] Similarly, for user i t and its characteristics We extract the potential features of the user in the kth intent in the following way:

[0121]

[0122]

[0123] Where σ is the activation function, and is a trainable parameter, R is the real number domain, D is the feature dimension, and K is the number of intents.

[0124] Furthermore, in an embodiment of the present application, multiple superimposed hypergraph decoupling modules are used to obtain new user features and item features, including the following steps:

[0125] According to the user feature representation and the item feature representation, the non-zero elements in the corresponding user hypergraph structure and the item hypergraph structure are updated, and the updated user hypergraph structure and the item hypergraph structure are normalized to obtain a new user hypergraph structure and a new item hypergraph structure;

[0126] Perform hypergraph convolution operations on the user feature representation and the item feature representation according to the updated user hypergraph structure and the updated item hypergraph structure to obtain updated user features and item features;

[0127] A distance-related loss function is used to calculate the correlation loss between different intent features in the updated user features and item features.

[0128] Furthermore, in the embodiment of the present application, the corresponding user hypergraph structure is updated according to the user feature representation, and the intention-aware hypergraph structure is updated: the corresponding hypergraph structure is updated based on the node features under different intentions, so that different hypergraph structures pay attention to the associated information under their corresponding intentions. For a user, given the feature representation of the user under different intentions {E u，1 , ..., E u，K}, the feature representation of items under different intentions {E i，1 , ..., E i，K}, the hypergraph association matrix {H u，1 , ..., H u，K} and a vector {c i，1 , ..., c i，K}, we first update the non-zero elements in the hypergraph incidence matrix using the following method:

[0129]

[0130] Where j and m are the numbers of the user and hyperedge respectively, and c i，k (m) is the construction of hyperedge H u，k (:,m) is the number of the item used, a(·) is the function for calculating the association, and dot multiplication can be used. For user u j and its characteristics The potential features of the user in the kth intent, The features of the corresponding items used to construct user hyperedges.

[0131] After updating the user hypergraph association matrix under each intent, we use the Softmax function to normalize the hypergraph structure across intents:

[0132]

[0133] Similarly, for items, given the feature representation of items and users {E i，1 , ..., E i，K} and {E u，1 , ..., E u，K}, the association matrix of the item hypergraph {Hi，1 , ..., H i，K} and a vector {c u，1 , ..., c u，K}, we update the non-zero elements in the hypergraph incidence matrix in the following way:

[0134]

[0135] Where j and m are the numbers of the item and hyperedge respectively, and c u，k (m) is the construction of hyperedge H i，k (:, m) when the central user number, For item i t and its characteristics The features of the corresponding users used to construct the item hyperedge,

[0136] After the latent features of the item in the kth intent are updated, we also perform cross-intent hypergraph structure normalization

[0137]

[0138] Further, in the embodiment of the present application, based on the updated hypergraph structure information, hypergraph convolution is performed on the hypergraph structures of different intentions, and the hypergraph convolution operation is respectively applied to the item hypergraphs under different intentions to update the corresponding item node features. The updated user features and item features are expressed as:

[0139]

[0140]

[0141] in, is the node degree matrix of the user hypergraph, H u，k represents the updated user hypergraph structure, is a trainable parameter, R is the real number domain, D is the feature dimension, K is the number of intents, is the hyper-edge degree matrix of the user hypergraph, E u，k (l) is the user feature representation, is the hyperedge degree matrix of the item hypergraph, H i，k represents the updated item hypergraph structure, is the hyperedge degree matrix of the item hypergraph, is the characteristic representation of the item,

[0142] The correlation loss between different intent features in the updated user features and item features is expressed as:

[0143]

[0144] in, and e k′ (l) is the feature representation of the user or item about intention k and intention k′ at the lth layer, K is the number of intentions, d(·) is the distance association, and d(·) is expressed as:

[0145]

[0146] Here, dCov(·) and dVar(·) represent covariance and variance, respectively.

[0147] Furthermore, in an embodiment of the present application, feature fusion is performed on new user features and new item features, and features of different intents in each layer of the network are fused, and then the features at different levels are fused to obtain the final feature representation of the user and the item.

[0148] For user j and item t at layer l, their features at layer l can be obtained as follows:

[0149]

[0150]

[0151] The final features of users and items can be obtained as follows:

[0152]

[0153]

[0154] Furthermore, in the embodiment of the present application, the dot product result is:

[0155]

[0156] in, represents the final feature representation of the user, Represents the final feature representation of the item.

[0157] The dot product of user features and item features is used to represent the user's preference for the item. The larger the value, the greater the user's preference for the item.

[0158] Figure 2 This is another flowchart of the collaborative filtering recommendation method based on the decoupled dual-channel hypergraph neural network according to an embodiment of the present application.

[0159] like Figure 2As shown, the collaborative filtering recommendation method based on the decoupled dual-channel hypergraph neural network includes: using the commonly used dual-channel hypergraph construction method to construct a hypergraph of users and items under different intentions; obtaining the feature representation of users and items under different intentions by using the intention-aware feature extraction module; based on the hypergraph structure and node feature information of users and items under different intentions, superimposing multiple hypergraph decoupling modules to capture the correlation information of users and items at different levels; fusing the features of different intentions in each layer of the network, and then fusing the features of different levels to obtain the final feature representation of users and items; for the training of the model, Bayesian personalized sorting is mainly used for learning.

[0160] In order to implement the above-mentioned embodiment, the present application also proposes a collaborative filtering recommendation system based on a decoupled dual-channel hypergraph neural network, including: a first acquisition module, used to acquire user and item information, and obtain user features, item features and a user-item interaction graph according to the user and item information; a construction module, used to construct a user hypergraph structure and an item hypergraph structure according to the user-item interaction graph using a dual-channel hypergraph construction method; a second acquisition module, used to obtain a user feature representation and an item feature representation according to the user features and the item features using an intention-aware feature extraction module; a third acquisition module, used to obtain new user features and new item features according to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation using multiple superimposed hypergraph decoupling modules; a fusion module, used to perform feature fusion on the new user features and the new item features to obtain a user final feature representation and an item final feature representation; a determination module, used to use the dot product result of the user final feature representation and the item final feature representation to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item.

[0161] Furthermore, in the embodiment of the present application, the decoupled dual-channel hypergraph neural network is trained, Bayesian personalized sorting is adopted, and learning is performed through the following function:

[0162]

[0163] in, σ is the Sigmoid function, Θ is the model trainable parameter, λ is the regularization hyperparameter, o is the training set triplet set used for training, are the interactions that appeared in the training set, is an interaction that has not appeared in the training set, μ is the weight coefficient of the distance association loss, L is the association loss function, u is the user, i is the item clicked by user u in the training set, and j is the item not clicked by the user in the training set.

[0164] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0165] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0166] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0167] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0168] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0169] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0170] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0171] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A collaborative filtering recommendation method based on a decoupled dual-channel hypergraph neural network, characterized in that: The following steps are involved: Obtain the user-item interaction graph and randomly initialize it to obtain user features and item features; According to the user-item interaction graph, a dual-channel hypergraph construction method is used to construct a user hypergraph structure and an item hypergraph structure; According to the user features and the item features, an intention-aware feature extraction module is used to obtain a user feature representation and an item feature representation; According to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation, a plurality of superimposed hypergraph decoupling modules are used to obtain new user features and new item features; Performing feature fusion on the new user feature and the new item feature to obtain a final user feature representation and a final item feature representation; Using the dot product result of the final feature representation of the user and the final feature representation of the item to represent the user's preference for the item, wherein the larger the dot product result, the greater the user's preference for the item; Wherein, the user hypergraph structure is constructed by using a dual-channel hypergraph construction method, comprising the following steps: The k-order reachability matrix of an item is defined as: in, The adjacency matrix H is the user-item interaction graph, and M is the number of items; Based on the k-order reachability matrix of the item, the corresponding user's hyperedge group adjacency matrix is ​​expressed as: in, N is the number of users, M is the number of items; The first-order matrix and the second-order matrix in the hyperedge group adjacency matrix of the user are used to perform hyperedge group concatenation to obtain the final user hypergraph, and the user hypergraph is expressed as: in, is the first-order matrix in the hyperedge group adjacency matrix of the user, is the second-order matrix in the hyperedge group adjacency matrix of the user; A user hypergraph structure is obtained according to the user hypergraph, and the user hypergraph structure is expressed as: Among them, H u is the user hypergraph, The item hypergraph structure is constructed using a dual-channel hypergraph construction method, including the following steps: The k-order reachable matrix of a user is defined as: in, The adjacency matrix H is the user-item interaction graph, M is the number of items, and N is the number of users; Based on the k-order reachability matrix of the user, the hyperedge group adjacency matrix of the corresponding item is expressed as: in, The first-order matrix and the second-order matrix in the hyperedge group adjacency matrix of the item are used to perform hyperedge group concatenation to obtain the final item hypergraph, and the item hypergraph is expressed as: in, is the first-order matrix in the hyperedge group adjacency matrix of the user, is the second-order matrix in the hyperedge group adjacency matrix of the user; The item hypergraph structure is obtained according to the item hypergraph, and the item hypergraph structure is expressed as: Among them, H i A super picture of the item.

2. The method according to claim 1, characterized in that The intention-aware feature extraction module is used to obtain user feature representation and item feature representation, which are specifically expressed as follows: in, For user u j and its characteristics The potential features of the user in the kth intent, and is a trainable parameter, For item i t and its characteristics The potential features of the item in the kth intention, and is a trainable parameter, R is the real number domain, D is the feature dimension, K is the number of intents, and σ is the activation function.

3. The method according to claim 1, characterized in that Using the multiple superimposed hypergraph decoupling modules to obtain new user features and item features includes the following steps: According to the user feature representation and the item feature representation, the non-zero elements in the corresponding user hypergraph structure and the item hypergraph structure are updated, and the updated user hypergraph structure and the item hypergraph structure are normalized to obtain a new user hypergraph structure and a new item hypergraph structure; Performing a hypergraph convolution operation on the user feature representation and the item feature representation according to the updated user hypergraph structure and the updated item hypergraph structure to obtain updated user features and item features; A distance-related loss function is used to calculate the correlation loss between different intent features in the updated user features and item features.

4. The method according to claim 3, characterized in that According to the user feature representation and the item feature representation, the non-zero elements in the corresponding user hypergraph structure and item hypergraph structure are updated. The updated user hypergraph structure and item hypergraph structure are expressed as: Where j and m are the numbers of the user and hyperedge respectively, and c i,k (m) is to construct the user hyperedge H u,k (:,m) is the number of the item used, a(·) is the function for calculating the association, and the dot product can be used. For user u j and its characteristics The potential features of the user in the kth intent, is the feature of the corresponding item used to construct the user hyperedge, c u,k (m) is the construction of hyperedge H i,k (:,m) when the central user number, The features of the corresponding users used to construct the item hyperedge, For item i t and its characteristics The potential features of the item in the kth intention; The Softmax function is used to normalize the updated user hypergraph structure and item hypergraph structure across intents, which can be expressed as: Among them, K is the number of intent features, is the updated user hypergraph structure, is the hypergraph structure of the tth user, is the updated item hypergraph structure, is the hypergraph structure of the t-th item, j is the user number, and m is the hyperedge number.

5. The method according to claim 3 or 4, characterized in that The updated user features and item features are expressed as: in, is the node degree matrix of the user hypergraph, H u,k represents the updated user hypergraph structure, is a trainable parameter, R is the real number domain, D is the feature dimension, K is the number of intents, is the hyper-edge degree matrix of the user hypergraph, E u,k (l) is the user feature representation, is the hyper-edge degree matrix of the item hypergraph, H i,k represents the updated item hypergraph structure, is the hyper-edge degree matrix of the item hypergraph, E i,k (l) is the characteristic representation of the item, The correlation loss between different intent features in the updated user features and item features is expressed as: in, and e k′ (l) is the feature representation of the user or item about intention k and intention k′ at the lth layer, K is the number of intentions, dCor(·) is the distance correlation, and dCor(·) is expressed as: Here, dCov(·) and dVar(·) represent covariance and variance, respectively.

6. The method according to claim 1, characterized in that The new user features and the new item features are fused, and the fused features are expressed as: in, represents the characteristics of users in layer l, Represents the characteristics of items in layer l.

7. The method according to claim 1, characterized in that The dot product result is: in, represents the final feature representation of the user, Represents the final feature representation of the item.

8. A collaborative filtering recommendation system based on a decoupled dual-channel hypergraph neural network, characterized in that: include: The first acquisition module is used to acquire the user-item interaction graph and randomly initialize it to obtain user features and item features; A construction module, used to construct a user hypergraph structure and an item hypergraph structure according to the user-item interaction graph by using a dual-channel hypergraph construction method; A second acquisition module is used to obtain user feature representation and item feature representation based on the user feature and item feature by using an intention-aware feature extraction module; A third acquisition module is used to acquire new user features and new item features by using multiple superimposed hypergraph decoupling modules according to the user hypergraph structure, the item hypergraph structure, the user feature representation, and the item feature representation; A fusion module, used to perform feature fusion on the new user feature and the new item feature to obtain a final user feature representation and a final item feature representation; a determination module, configured to use a dot product result of the final feature representation of the user and the final feature representation of the item to represent the user's preference for the item, wherein a larger dot product result represents a greater preference of the user for the item; Wherein, the user hypergraph structure is constructed by using a dual-channel hypergraph construction method, comprising the following steps: The k-order reachability matrix of an item is defined as: in, The adjacency matrix H is the user-item interaction graph, and M is the number of items; Based on the k-order reachability matrix of the item, the corresponding user's hyperedge group adjacency matrix is ​​expressed as: in, N is the number of users, M is the number of items; The first-order matrix and the second-order matrix in the hyperedge group adjacency matrix of the user are used to perform hyperedge group concatenation to obtain the final user hypergraph, and the user hypergraph is expressed as: in, is the first-order matrix in the hyperedge group adjacency matrix of the user, is the second-order matrix in the hyperedge group adjacency matrix of the user; A user hypergraph structure is obtained according to the user hypergraph, and the user hypergraph structure is expressed as: Among them, H u is the user hypergraph, The item hypergraph structure is constructed using a dual-channel hypergraph construction method, including the following steps: The k-order reachable matrix of a user is defined as: in, The adjacency matrix H is the user-item interaction graph, M is the number of items, and N is the number of users; Based on the k-order reachability matrix of the user, the hyperedge group adjacency matrix of the corresponding item is expressed as: in, The first-order matrix and the second-order matrix in the hyperedge group adjacency matrix of the item are used to perform hyperedge group concatenation to obtain the final item hypergraph, and the item hypergraph is expressed as: in, is the first-order matrix in the hyperedge group adjacency matrix of the user, is the second-order matrix in the hyperedge group adjacency matrix of the user; The item hypergraph structure is obtained according to the item hypergraph, and the item hypergraph structure is expressed as: Among them, H i A super picture of the item.

9. The system according to claim 8, characterized in that The system is trained by using Bayesian personalized sorting and learning through the following function: Among them, μ is the weight coefficient of distance association loss, L is the distance correlation loss function, σ is the Sigmoid function, Θ is the model trainable parameter, λ is the regularization hyperparameter, o is the training set triplet set used for training, are the interactions that appeared in the training set, is an interaction that has not appeared in the training set, u is the user, i is the item clicked by user u in the training set, and j is the item not clicked by the user in the training set.

Images