A self-supervised item recommendation heterogeneous graph learning method based on unified spectral perspective

Abstract

This invention discloses a self-supervised heterogeneous graph learning method for item recommendation based on a unified spatial-spectral perspective. This method is a self-supervised heterogeneous graph learning framework that simultaneously considers spatial and spectral domain information, which is beneficial for graph representation learning. Specifically, a spatial encoder is used to aggregate information from local neighbor nodes, and a spectral encoder is used to extract global structural information, thereby suppressing noise from local aggregation. Subsequently, a homeomorphism mapping method is proposed, which can align dual-domain representations while avoiding the loss of domain-specific information. Furthermore, a spectral optimization module is proposed, which uses fused representations to generate cluster-level guidance for effectively optimizing the original representations. Extensive experimental results on various real-world heterogeneous graph datasets for item recommendation demonstrate the significant superiority of this invention.

Description

Technical Field

[0001] This invention belongs to the field of item recommendation technology, and more specifically, relates to a self-supervised heterogeneous graph learning method for item recommendation based on a unified spatial spectrum perspective. Background Technology

[0002] In e-commerce platforms, social media, or content distribution systems, items such as goods, videos, and news are accurately recommended to users. Traditional item recommendation methods often only utilize the binary interaction information between users and items, ignoring the rich variety of entities and their complex relationships in the recommendation system. The learned node representations are suboptimal, resulting in limited effectiveness in practical applications.

[0003] Self-supervised heterogeneous graph learning (SHGL) aims to uncover diverse interactions between nodes in heterogeneous graphs without labels, thereby generating node representations beneficial for downstream tasks. Therefore, this technique has been widely adopted in item recommendation systems.

[0004] However, existing self-supervised item recommendation methods for learning heterogeneous graphs can be broadly categorized into two types: meta-path-based methods and meta-path-free methods. The former typically utilizes multiple meta-paths to capture the homogeneity of heterogeneous graphs from different perspectives. For example, GTC generates tokens via meta-paths and inputs them into the Transformer encoder to capture multi-hop interaction information. However, the design of meta-paths relies on expert knowledge for manual construction. To address this issue, meta-path-free methods have been proposed to adaptively acquire graph structure and learn node representations without relying on predefined meta-paths. For example, HERO captures homogeneity through a self-representation matrix and models heterogeneity through information aggregation across node types.

[0005] While existing self-supervised item recommendation heterogeneous graph learning methods have demonstrated good performance, they still have significant limitations. Specifically, both methods based on metapaths and those without metapaths employ spatial message passing mechanisms. However, due to noise in the metapaths or graph structure, both methods may involve message passing between nodes of different categories. This phenomenon is detrimental to representation learning. Consequently, the node representations learned from heterogeneous graphs in self-supervised item recommendation exhibit weak generalization ability and low quality, resulting in insufficient improvement in the overall recommendation accuracy. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a self-supervised heterogeneous graph learning method for item recommendation based on a unified spatial spectrum perspective. This method reduces noise interference, achieves effective and efficient node representation extraction, improves the generalization ability of self-supervised item recommendation heterogeneous graph learning, enhances the generalization ability of node representation, and thus improves the overall recommendation accuracy of items.

[0007] To achieve the above-mentioned objectives, this invention provides a self-supervised item recommendation heterogeneous graph learning method based on a unified spatial spectrum perspective, characterized by the following steps:

[0008] (1) Constructing a heterogeneous graph for item recommendation based on a business dataset

[0009] Constructing a heterogeneous graph for item recommendation based on commercial datasets. , ,in, and Let them represent the set of nodes and the set of edges, respectively. Represents the node feature matrix, Indicates the first Each node's characteristics The total number of nodes. For node feature dimensions, and These represent the set of node types and the set of edge types, respectively. (Node type set) It includes three types of nodes: user nodes, item nodes, and attribute nodes, and a set of edge types. It includes three types of interaction or association: user-item interaction edge, user-user social edge, and item-attribute edge; the initial value of the node feature of the user node is obtained by embedding the user profile vector or the user's historical behavior sequence through the representation layer; the initial value of the node feature of the item node is obtained by embedding the item description text, image feature vector or category encoding; and the initial value of the node feature of the attribute node is obtained by one-hot encoding or pre-trained word vectors.

[0010] (2) Constructing a self-supervised item recommendation heterogeneous graph learning model from a unified spatial spectrum perspective

[0011] 2.1) Spatial representation is generated sequentially using a spatial encoder and a multilayer perceptron. and its projected spatial representation

[0012] First, a spatial representation is generated using a spatial domain encoder based on spatial domain aggregation, where nodes... In the The spatial representation of a layer is as follows:

[0013] (1)

[0014] in, Represents a non-linear activation function. Indicates by edge type With nodes The set of connected neighbor nodes, Represents the set of neighboring nodes The number of nodes in For the set of neighbor nodes Middle node In the The spatial representation of the layer, and the first Layer space representation Initialize as node features , It is for edge type The learnable weight matrix;

[0015] All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. ;

[0016] Then, a multilayer perceptron is used to obtain the spatial representation of the projected domain. :

[0017] (2)

[0018] in, It is a non-linear activation function. It is a multilayer perceptron;

[0019] 2.2) A spectral domain representation is generated sequentially using a spectral domain encoder and a multilayer perceptron. and its projected spectral domain characterization

[0020] First, a spectral domain representation is generated using a positive noncommutative spectral encoder. :

[0021] (3)

[0022] in, This represents the set of all normalized adjacency matrices of two node types. Based on adjacency matrix set matrix polynomial, Construct a spectral domain filter, It is a spectrum encoder;

[0023] Then, a multilayer perceptron is used to obtain the projection spectral domain representation. :

[0024] (4)

[0025] in, It is a non-linear activation function. It is a multilayer perceptron;

[0026] 2.3) Representing the projected spatial domain and projection spectral domain characterization By splicing, a fused representation is obtained. ;

[0027] 2.4) Obtaining the spatial and spectral domain representations after homeomorphism mapping

[0028] Characterizing the projected spatial domain The Line is a node airspace representation and projection spectral domain characterization The Line is a node spectral domain representation Consider as the initial input of a homeomorphism and ;

[0029] Then in the In the layer, they are mapped from the source domain to another representation space to achieve an invertible transformation, which can be formally expressed as:

[0030] (5)

[0031] (6)

[0032] in, This indicates element-wise multiplication, while Represents the learnable function implemented by the multilayer perceptron;

[0033] All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them in rows sequentially to obtain the spectral domain representation. All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. ;

[0034] 2.5) Use a spectral neural network module to generate effective clustering level guidance information.

[0035] By calculating the affinity matrix And generate clustering assignment matrix Further pseudo-labels were obtained. Specifically, it includes the following steps:

[0036] 2.5.1) Obtain the affinity matrix

[0037] Affinity Matrix element value Through fusion characterization calculate:

[0038] (7)

[0039] in, Represents a node of Nearest neighbor set Set as a scaling parameter for the node. To its first The distance to the nearest neighbor, , Fusion characterization The line, number The nodes obtained through action learning and Node representation;

[0040] 2.5.2) Obtain the clustering assignment matrix and pseudo-labels

[0041] First, use a spectral neural network Projection fusion characterization Characterization :

[0042] (8)

[0043] Among them, characterization , It is the number of clusters. For activation functions;

[0044] Then, the characterization is orthogonalized through QR decomposition. This yields a clustering assignment matrix that provides information for soft clustering. ;

[0045] Finally, in the clustering assignment matrix The above uses the argmax operation to obtain pseudo-tags. ;

[0046] (3) Training a self-supervised item recommendation heterogeneous graph learning model from a unified spatial spectrum perspective

[0047] For the Iteration:

[0048] First, after obtaining the projection spatial domain representation and spectral domain characterization Projected spectral domain characterization and spatial representation Then, calculate the objective function. :

[0049] (9)

[0050] Among them, the consistency loss function , for:

[0051]

[0052] (10)

[0053] in, It is a non-negative parameter. , It is alignment loss. , It is a regularization term;

[0054] Then, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under a unified spatial spectrum perspective is optimized. After optimization, the objective function is calculated. :

[0055] (11)

[0056] in, To refine the loss function, It is a non-negative parameter. It is the spectral loss function. The loss function is guided by pseudo-labels;

[0057] Finally, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under the unified spatial spectrum perspective is optimized. After optimization, it returns to the next iteration, i.e. the next round of alternating optimization, until the set number of iterations is reached.

[0058] (4) Output the final trained model and generate fused representations. , among which, the OK For the nodes learned Node representation.

[0059] The objective of this invention is achieved as follows.

[0060] This invention proposes a self-supervised heterogeneous graph learning framework for item recommendation that simultaneously considers spatial and spectral domain information, which is beneficial for graph representation learning. Specifically, a spatial encoder is used to aggregate information from local neighbor nodes, and a spectral encoder is used to extract global structural information, thereby suppressing noise from local aggregation. Subsequently, a homeomorphism mapping method is proposed, which can align dual-domain representations while avoiding the loss of domain-specific information. In addition, we propose a spectral optimization module, which uses cluster-level guidance generated by the fused representation to effectively optimize the original representation. Extensive experimental results on various real-world heterogeneous graph datasets for item recommendation demonstrate the significant superiority of this invention.

[0061] Compared with the prior art, the present invention has the following advantages:

[0062] 1) This invention is the first to unify the spatial spectrum perspective for a self-supervised item recommendation heterogeneous graph learning method;

[0063] 2) This invention proposes an efficient and effective method for spatial and spectral representation encoding and optimization, and further introduces homeomorphism mapping technology to extract consensus and complementary information from both spatial and spectral perspectives;

[0064] 3) This invention has been comprehensively empirically studied on four public benchmark datasets of heterogeneous graphs. Compared with five comparative methods, it has verified the superior performance of this invention in node classification. Attached Figure Description

[0065] Figure 1 This is a flowchart of a specific implementation of the self-supervised item recommendation heterogeneous graph learning method based on a unified spatial spectrum perspective of the present invention;

[0066] Figure 2 This is a schematic diagram illustrating the principle of a specific implementation of the self-supervised item recommendation heterogeneous graph learning method based on a unified spatial spectrum perspective of the present invention.

[0067] Figure 3 This is a flowchart of the generation of cluster-level guidance information in this invention. Detailed Implementation

[0068] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0069] For heterogeneous graph learning in self-supervised item recommendation, a spectral domain perspective offers a potential solution from a graph signal processing standpoint. Unlike spatial domain-based methods that aggregate local messages, the spectral domain can encode global structure and absolute node position information. Therefore, a unified spatial-spectral perspective can complement information from both domains, reduce noise interference, and thus improve the generalization ability of self-supervised heterogeneous graph learning, enhance node representation quality, and ultimately improve the overall recommendation accuracy of items, thereby improving its effectiveness in practical applications.

[0070] Specifically, this invention provides a novel heterogeneous graph learning method for item recommendation, namely a self-supervised heterogeneous graph representation learning method based on a unified spatial-spectral perspective. This method effectively achieves heterogeneous graph representation learning for item recommendation, enabling efficient and effective node representation extraction. To achieve efficient encoding and representation optimization, this invention designs two specific encoders to obtain spatial and spectral domain representations respectively, and then designs a spectral optimization module to guide representation optimization. To effectively unify the dual-domain representations, this invention designs a homeomorphism mapping module to preserve specific information within the domains while aligning the spatial and spectral domain representations. Experiments on various real-world heterogeneous graph datasets and downstream tasks demonstrate that the proposed method significantly outperforms existing state-of-the-art self-supervised heterogeneous graph learning methods for item recommendation.

[0071] Figure 1 , 2 These are a flowchart and a schematic diagram of a specific implementation of the self-supervised item recommendation heterogeneous graph learning method based on a unified spatial spectrum perspective of the present invention.

[0072] In this embodiment, as Figure 1 As shown, the self-supervised item recommendation heterogeneous graph learning method based on a unified spatial spectrum perspective of the present invention includes the following steps:

[0073] Step S1: Construct a heterogeneous graph for item recommendation based on a business dataset

[0074] like Figure 2 As shown, a heterogeneous graph for item recommendation is constructed based on a business dataset. , ,in, and Let them represent the set of nodes and the set of edges, respectively. Represents the node feature matrix, Indicates the first Each node's characteristics The total number of nodes. For node feature dimensions, and These represent the set of node types and the set of edge types, respectively.

[0075] Node type collection It includes three types of nodes: user nodes, item nodes, and attribute nodes, a set of node types. This can be represented as {user, item, attribute}. User nodes represent individual users participating in the item recommendation system, item nodes represent entities such as products, videos, or articles to be recommended, and attribute nodes represent attribute tags related to users or items, such as product category, brand, or user age range, interest tags, etc.

[0076] Edge type set It includes three types of interaction or association: user-item interaction edges, user-user social edges, and item-attribute edges. The set of edge types can be represented as {interact, social, belong_to}. Among them, user-item interaction edges indicate that there are historical interaction behaviors between users and items, such as clicking, purchasing, watching, or rating, and can carry timestamps or rating values ​​as edge attributes. User-user social edges indicate social relationships such as friends and following to introduce social influence information. Item-attribute edges indicate the subordinate relationship between items and attributes, such as a product belonging to the category of "electronic products".

[0077] The initial values ​​of node features for user nodes are obtained by embedding representation layers using user profile vectors or user history behavior sequences. The initial values ​​of node features for item nodes are obtained by embedding vectors of item description text, image feature vectors, or category encoding. The initial values ​​of node features for attribute nodes are obtained by one-hot encoding or pre-trained word vectors.

[0078] like Figure 2 As shown, this invention first employs two specific encoders to acquire spatial and spectral domain representations respectively. Then, a multilayer perceptron (MLP) is used to obtain projected representations. Subsequently, a homeomorphism mapping module is used to map the projected representations to another representation space, thus avoiding direct alignment. Furthermore, a consensus loss is designed to capture the consensus of the dual-domain representations. The model further utilizes the dual-domain fused representations, designing a refined loss function to generate effective guidance signals while introducing specific pseudo-labels to guide the loss optimization of the original representations. Specifically, it includes the following steps:

[0079] Step S2: Construct a self-supervised heterogeneous graph learning model for item recommendation from a unified spatial spectrum perspective.

[0080] Step S2.1: Generate spatial representations sequentially using a spatial encoder and a multilayer perceptron. and its projected spatial representation

[0081] like Figure 2 As shown, firstly, a spatial domain representation is generated using a spatial domain encoder based on spatial domain aggregation, where nodes... In the The spatial representation of a layer is as follows:

[0082] (1)

[0083] in, Represents a non-linear activation function. Indicates by edge type With nodes The set of connected neighbor nodes, Represents the set of neighboring nodes The number of nodes in For the set of neighbor nodes Middle node In the The spatial representation of the layer, and the first Layer space representation Initialize as node features , It is for edge type The learnable weight matrix.

[0084] All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. .

[0085] Then, a multilayer perceptron is used to obtain the spatial representation of the projected domain. :

[0086] (2)

[0087] in, It is a non-linear activation function. It is a multilayer perceptron.

[0088] Step S2.2: Generate spectral domain representations sequentially using a spectral domain encoder and a multilayer perceptron. and its projected spectral domain characterization

[0089] like Figure 2 As shown, firstly, a spectral domain representation is generated using a spectral encoder based on positive noncommutative spectral encoding. :

[0090] (3)

[0091] in, This represents the set of all normalized adjacency matrices of two node types. Based on adjacency matrix set matrix polynomial, Construct a spectral domain filter, It is a spectrum encoder;

[0092] Then, a multilayer perceptron is used to obtain the projection spectral domain representation. :

[0093] (4)

[0094] in, It is a non-linear activation function. It is a multilayer perceptron.

[0095] Step S2.3: As Figure 2 As shown, the projected spatial domain is represented. and projection spectral domain characterization By splicing, a fused representation is obtained. .

[0096] Step S2.4: Obtain the spatial and spectral domain representations after homeomorphic mapping.

[0097] To avoid information loss caused by alignment in a common space, this invention employs a continuous bijective structure, namely a homeomorphism, to transform the original representation into another representation space before alignment. The homeomorphism consists of multiple invertible neural network blocks (INNs). Specifically, it includes the following steps:

[0098] like Figure 2 As shown, the projected spatial domain is represented. The Line is a node airspace representation and projection spectral domain characterization The Line is a node spectral domain representation Consider as the initial input of a homeomorphism and ;

[0099] Then in the In the layer, they are mapped from the source domain to another representation space to achieve an invertible transformation, which can be formally expressed as:

[0100] (5)

[0101] (6)

[0102] in, This indicates element-wise multiplication, while This represents the learnable function implemented by the multilayer perceptron.

[0103] All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them in rows sequentially to obtain the spectral domain representation. All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. .

[0104] This bijective nature ensures that the representation retains information from the source domain during the mapping process. Therefore, using homeomorphisms, we can obtain the mapped spectral domain representation. and spatial representation .

[0105] Step S2.5: Use the spectral neural network module to generate effective clustering level guidance information.

[0106] This invention uses a spectral neural network, mathematically equivalent to spectral clustering, to obtain node clustering assignments. For example... Figure 2 As shown, the affinity matrix is ​​calculated. And generate clustering assignment matrix Further pseudo-labels were obtained. , specifically Figure 3 As shown, it includes the following steps:

[0107] Step S2.5.1: Obtain the affinity matrix

[0108] Affinity Matrix element value Through fusion characterization calculate:

[0109] (7)

[0110] in, Represents a node of Nearest neighbor set Set as a scaling parameter for the node. To its first The distance to the nearest neighbor, , Fusion characterization The line, number The nodes obtained through action learning and Node representation.

[0111] The resulting affinity matrix element value Nodes were measured and The similarity between them.

[0112] Step S2.5.2: Obtain the clustering assignment matrix and pseudo-labels

[0113] First, in order to efficiently obtain the clustering assignment matrix... This invention uses a spectral neural network. Projection fusion characterization Characterization :

[0114] (8)

[0115] Among them, characterization , It is the number of clusters. This is the activation function.

[0116] Then, the characterization is orthogonalized through QR decomposition. This yields a clustering assignment matrix that provides information for soft clustering. .

[0117] Finally, in the clustering assignment matrix The above uses the argmax operation to obtain pseudo-tags. .

[0118] In summary, by utilizing affinity matrices and cluster assignment, this invention can efficiently obtain cluster-level guidance information to optimize the original representation.

[0119] Step S3: Train a self-supervised heterogeneous graph learning model for item recommendation from a unified spatial spectrum perspective.

[0120] This invention formalizes the training process as an alternating optimization, in which, in each iteration, the fused representation... and clustering assignment Consistency loss was designed in the previous steps. With refining loss This effectively achieves inter-domain alignment and representation optimization. Specifically, it includes the following steps:

[0121] For the Iteration:

[0122] First, after obtaining the projection spatial domain representation and spectral domain characterization Projected spectral domain characterization and spatial representation Then, calculate the objective function. :

[0123] (9)

[0124] Among them, the consistency loss function , for:

[0125]

[0126] (10)

[0127] in, It is a non-negative parameter. , It is alignment loss. , It is a regularization term.

[0128] In this embodiment, the alignment loss is designed to be represented in the spatial domain. With the mapped spectral domain representation Align between them, and vice versa. To learn inter-domain invariance:

[0129]

[0130]

[0131] in, The cosine similarity function is used. For temperature parameters, and , and , and as well as and Representing the spatial characterization of projection respectively Spectral domain representation after mapping Projected spectral domain characterization Spatial representation after mapping No. row and number OK.

[0132] By forcing the corresponding spatial and spectral domain node representations, such as and They must be consistent with each other in order to capture consensus information.

[0133] Nevertheless, the learned representation may suffer from dimensionality collapse, i.e., the representation collapses into a subspace with reduced dimensionality. To address this issue, this invention introduces a regularization term to avoid the risk of dimensionality collapse:

[0134]

[0135]

[0136] in, They are respectively in and , and The correlation matrix is ​​calculated between different dimensions (columns). , These are the correlation matrices The diagonal element values, , These are the correlation matrices No. Line number The element values ​​of the column, These are the weight parameters.

[0137] The first term of the regularization term preserves the correlation between corresponding dimensions, while the second term penalizes the correlation between different dimensions, thereby achieving decorrelation of the representation dimensions. Therefore, a representation that integrates the spatial and spectral domains is achieved. The method effectively captures consensus information by applying consistency loss, while retaining domain-specific information by using homeomorphism mapping.

[0138] Then, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under a unified spatial spectrum perspective is optimized. After optimization, the objective function is calculated. :

[0139] (11)

[0140] in, To refine the loss function, It is a non-negative parameter. It is the spectral loss function. It is a pseudo-label-guided loss function.

[0141] In this invention, a spectral loss function is designed to ensure that the learned clustering assignments can approximate the eigenvalues ​​of the graph Laplacian matrix, thereby obtaining effective clustering information.

[0142]

[0143] in, These are nodes , Spectral allocation, i.e., clustering allocation matrix No. line, number The row represents the soft assignment probability. It is the entropy of the distribution probability. , It is a clustering assignment matrix No. List, It is a non-negative balance parameter.

[0144] Spectral loss function The first term aims to minimize the distance between nodes with high affinity, encouraging them to obtain similar cluster assignments. The second term maximizes the entropy of the assignment distribution, preventing trivial solutions that assign all nodes to the same cluster. Thus, the cluster assignment matrix is ​​obtained. It provides soft clustering information learned from the fused representations. Furthermore, this invention uses the argmax operation to obtain pseudo-labels. Using these pseudo-labels, we designed a pseudo-label-guided loss function. To optimize the original representation:

[0145]

[0146] in, Representing nodes respectively Projection-based spatial and spectral domain representations belong to corresponding pseudo-labels. The probability of the category, and It is a non-negative parameter.

[0147] This invention employs a multilayer sensor as the projection head. and ,Will and Projected as logical values ​​respectively and Right now , Then, the softmax function is applied to the logical values ​​to obtain the probability vector. and , These represent probability vectors respectively. and Corresponding pseudo-tags The probability of a category is a node. Projection-based spatial and spectral domain representations belong to corresponding pseudo-labels. The probability of the category.

[0148] Pseudo-label guidance loss function This invention encourages nodes that may belong to the same category to be closer together in the representation space. Therefore, the present invention uses an objective function... Cluster assignments were effectively extracted, and the original representation was optimized.

[0149] Finally, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under the unified spatial spectrum perspective is optimized. After optimization, it returns to the next iteration, i.e., the next round of alternating optimization, until the set number of iterations is reached.

[0150] Through alternating optimization, the objective function First, capture the consensus between spatial and spectral domain representations, then the objective function... Effective cluster-level guidance is generated to alternately optimize the original representation, thereby enabling efficient model training.

[0151] Step S4: Output the final trained model and generate fused representations. , among which, the OK For the nodes learned Node representation.

[0152] To better illustrate the technical effects of this invention, specific examples are used to experimentally verify it. The node representations generated by this invention can be applied to downstream tasks such as click-through rate prediction in item recommendation systems. The quality of the representations is evaluated through node classification experiments. This experiment uses the commercial dataset Yelp for verification, comparing one unsupervised learning algorithm for homogeneous graphs (GCN) and four self-supervised learning algorithms for heterogeneous graphs (HERO, GTC, D2CMG, and LatGRL).

[0153]

[0154] Table 1

[0155] Table 1 presents the classification performance of various methods on real heterogeneous graph datasets, namely the Macro-F1 and Micro-F1 metrics. From a practical application perspective, the Macro-F1 score effectively measures the model's ability to represent cold-start items or long-tail categories with sparse interactions in the recommendation system, while the Micro-F1 score reflects the model's overall recommendation accuracy for mainstream items. The results show that this invention outperforms all compared methods in classification performance. Compared to other self-supervised methods, this invention achieves an average performance improvement of 13.2% over the homogeneous graph method GCN and 1.2% over the current best heterogeneous graph method LatGRL. Therefore, for node classification tasks, this invention outperforms other compared methods.

[0156] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A self-supervised heterogeneous graph learning method for item recommendation based on a unified spatial spectrum perspective, characterized in that, Includes the following steps: (1) Constructing a heterogeneous graph for item recommendation based on a business dataset , Constructing a heterogeneous graph for item recommendation based on commercial datasets. , ,in, and Let them represent the set of nodes and the set of edges, respectively. Represents the node feature matrix, Indicates the first Each node's characteristics The total number of nodes. For node feature dimensions, and These represent the set of node types and the set of edge types, respectively. (Node type set) It includes three types of nodes: user nodes, item nodes, and attribute nodes, and a set of edge types. It includes three types of interaction or association: user-item interaction edge, user-user social edge, and item-attribute edge; the initial value of the node feature of the user node is obtained by embedding the user profile vector or the user's historical behavior sequence through the representation layer; the initial value of the node feature of the item node is obtained by embedding the item description text, image feature vector or category encoding; and the initial value of the node feature of the attribute node is obtained by one-hot encoding or pre-trained word vectors. (2) Construct a self-supervised heterogeneous graph learning model for item recommendation from a unified spatial spectrum perspective. 2.1) Spatial representation is generated sequentially using a spatial encoder and a multilayer perceptron. and its projected spatial representation , First, a spatial representation is generated using a spatial domain encoder based on spatial domain aggregation, where nodes... In the The spatial representation of a layer is as follows: (1), in, Represents a non-linear activation function. Indicates by edge type With nodes The set of connected neighbor nodes, Represents the set of neighboring nodes The number of nodes in For the set of neighbor nodes Middle node In the The spatial representation of the layer, and the first Layer space representation Initialize as node features , It is for edge type The learnable weight matrix; All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. ; Then, a multilayer perceptron is used to obtain the spatial representation of the projected domain. : (2), in, It is a non-linear activation function. It is a multilayer perceptron; 2.2) A spectral domain representation is generated sequentially using a spectral domain encoder and a multilayer perceptron. and its projected spectral domain characterization , First, a spectral domain representation is generated using a positive noncommutative spectral encoder. : (3), in, This represents the set of all normalized adjacency matrices of two node types. Based on adjacency matrix set matrix polynomial, Construct a spectral domain filter, It is a spectrum encoder; Then, a multilayer perceptron is used to obtain the projection spectral domain representation. : (4), in, It is a non-linear activation function. It is a multilayer perceptron; 2.3) Representing the projected spatial domain and projection spectral domain characterization By splicing, a fused representation is obtained. ; 2.4) Obtain the spatial and spectral domain representations after homeomorphic mapping. Characterizing the projected spatial domain The Line is a node airspace representation and projection spectral domain characterization The Line is a node spectral domain representation Consider as the initial input of a homeomorphism and ; Then in the In the layer, they are mapped from the source domain to another representation space to achieve an invertible transformation, which can be formally expressed as: (5), (6), in, This indicates element-wise multiplication, while Represents the learnable function implemented by the multilayer perceptron; All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them in rows sequentially to obtain the spectral domain representation. All nodes The last layer, i.e., the first Spatial representation obtained from layers Arrange them row by row to obtain the spatial representation. ; 2.5) A spectral neural network module is used to generate effective clustering-level guidance information. By calculating the affinity matrix And generate clustering assignment matrix Further pseudo-labels were obtained. Specifically, it includes the following steps: 2.5.1) Obtain the affinity matrix , Affinity Matrix element value Through fusion characterization calculate: (7), in, Represents a node of Nearest neighbor set Set as a scaling parameter for the node. To its first The distance to the nearest neighbor, , Fusion characterization The line, number The nodes obtained through action learning and Node representation; 2.5.2) Obtain the clustering assignment matrix and pseudo-labels , First, use a spectral neural network Projection fusion characterization Characterization : (8), Among them, characterization , It is the number of clusters. For activation functions; Then, the characterization is orthogonalized through QR decomposition. This yields a clustering assignment matrix that provides information for soft clustering. ; Finally, in the clustering assignment matrix The above uses the argmax operation to obtain pseudo-tags. ; (3) Training a self-supervised heterogeneous graph learning model for item recommendation from a unified spatial spectrum perspective. For the Iteration: First, after obtaining the projection spatial domain representation and spectral domain characterization Projected spectral domain characterization and spatial representation Then, calculate the objective function. : (9), Among them, the consistency loss function , for: , (10), in, It is a non-negative parameter. , It is alignment loss. , It is a regularization term; Then, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under a unified spatial spectrum perspective is optimized. After optimization, the objective function is calculated. : (11), in, To refine the loss function, It is a non-negative parameter. It is the spectral loss function. The loss function is guided by pseudo-labels; Finally, based on the objective function The self-supervised item recommendation heterogeneous graph learning model under the unified spatial spectrum perspective is optimized. After optimization, it returns to the next iteration, i.e. the next round of alternating optimization, until the set number of iterations is reached. (4) Output the final trained model and generate fused representations. , among which, the OK For the nodes learned Node representation.

2. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 1, characterized in that, The alignment loss , They are respectively: , , in, The cosine similarity function is used. For temperature parameters, and , and , and as well as and Representing the spatial characterization of projection respectively Spectral domain representation after mapping Projected spectral domain characterization Spatial representation after mapping No. row and number OK.

3. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 1, characterized in that, The regularization term , for: , , in, They are respectively in and , and The correlation matrix is ​​calculated between different dimensions. , These are the correlation matrices The diagonal element values, , These are the correlation matrices No. Line number The element values ​​of the column, These are the weight parameters.

4. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 1, characterized in that, The spectral loss function for: , in, These are nodes , Spectral allocation, i.e., clustering allocation matrix No. line, number The row represents the soft assignment probability. It is the entropy of the distribution probability. , It is a clustering assignment matrix No. List, It is a non-negative balance parameter.

5. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 1, characterized in that, The pseudo-label-guided loss function for: , in, Representing nodes respectively Projection-based spatial and spectral domain representations belong to corresponding pseudo-labels. The probability of the category, and It is a non-negative parameter.

6. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 5, characterized in that, The probability Obtained through the following steps: Each uses a multilayer sensor as the projection head. and ,Will and Projected as logical values ​​respectively and Right now , Then, the softmax function is applied to the logical values ​​to obtain the probability vector. and , These represent probability vectors respectively. and Corresponding pseudo-tags The probability of a category is a node. Projection-based spatial and spectral domain representations belong to corresponding pseudo-labels. The probability of the category.

Images