A graph neural network micro-map recommendation method considering social information

Abstract

The application discloses a kind of graph neural network micro map recommendation method considering social information, comprising: first, the social connection between user, the interaction between user and project and the relationship between project are constructed as a collaborative heterogeneous graph containing multiple types of relationships;Then, through memory enhancement and information aggregation module to refine and strengthen these relationship data, so that the characteristics of each user and project are more distinctive and distinguished, and the embedding features are extracted by multi-layer neural network and attention mechanism;Further, using embedding features to decouple specific types of information, and decoupled information propagation;Finally, by activation function, layer normalization, residual connection, connection operation etc., integrate all information, use contrast learning to further optimize and refine these features, to generate personalized micro map recommendation result.

Description

Technical Field

[0001] This invention relates to the field of micromap dissemination in the era of social media, and more specifically to a decoupled graph neural network micromap recommendation method that takes into account social information. Background Technology

[0002] In traditional map distribution and dissemination mechanisms, the process is primarily unidirectional or broadcast-based, meaning that professional cartographers or organizations control the generation and publication of maps, forming a model that starts with map publications and spreads to end-user groups. This model is significantly limited in terms of dissemination efficiency and interactivity. Although adaptive electronic map technology and crowdsourcing-based "crowd-helping" map projects attempt to break through this traditional framework and achieve innovation in map production and dissemination, they have not fundamentally changed the "broadcast-based" dissemination mode of map information. Furthermore, while there have been pioneering inventions that introduced Web 2.0 concepts into map production, proposing ideas for greater user participation, these attempts have not yet explored the potential of using social media technology for map dissemination. With the rapid development of social media technology, the democratization of map dissemination has become a realistic possibility. The distribution and dissemination mechanism of electronic maps urgently needs to consider end-users and their social network environment, not only achieving information broadcasting from a single point to a wide area, but also supporting point-to-point and inter-group information interaction.

[0003] The We-map platform, as a comprehensive platform integrating public mapping and social network frameworks, embodies the characteristics of a new paradigm of heterogeneous information networks. Its structure comprises diverse components, involving multiple dimensions such as individual users, map information representation, and online social activities. The We-map platform encapsulates the dynamics of interaction between users and map entities, the complex social network structure between individuals, the interaction between users and specific map tags, the establishment of links based on personal preferences, content dissemination mechanisms, and the correlation between different map elements. The introduction of recommendation algorithms into the micro-map product recommendation mechanism aims to quickly and effectively capture the preference patterns of specific user groups based on social network information among users, thereby improving the dissemination efficiency of micro-map content. Through a precise recommendation system, it can quickly filter out the information that best meets user needs from a large amount of micro-map content, achieving widespread dissemination of user-generated content, meeting the requirements of rapid, accurate, and personalized dissemination, and effectively addressing the challenges of map dissemination in the era of self-media.

[0004] To address the often overlooked issue of social dissemination in micromap recommendation processes, this invention proposes a graph neural network-based micromap recommendation strategy that considers social information. This strategy integrates a decoupling method based on extended memory networks with heterogeneous graph neural network technology, aiming to comprehensively model user social relationships, item (map) relationships, and user behavior data. By extending collaborative filtering methods, a wider range of data types are introduced to achieve more accurate collaborative filtering results. Simultaneously, a decoupling representation learning method for heterogeneous information is employed to perform detailed analysis and learning on diverse data. A comprehensive objective function is constructed by combining decoupling neural networks and contrastive learning techniques. Furthermore, a target attention mechanism is introduced to assess the user's emphasis on specific items, and features obtained from contrastive learning and the attention mechanism are integrated through a multilayer perceptron to achieve the final recommendation output, thereby realizing the social dissemination of micromaps.

[0005] In summary, this invention aims to solve the core problems in the dissemination of micromaps in the era of social media by introducing a novel graph neural network micromap recommendation method that takes social information into account. Figure 1 The overall process of this recommendation method is demonstrated, involving two key parts: neural network model training and social recommendation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The graph neural network micromap recommendation method that takes social information into account has the following model training steps:

[0008] S1: Initialize all model parameters Θ to prepare for the subsequent training process;

[0009] S2: Construct a collaborative heterogeneous graph G (G={V,E}) containing multiple relationships using the given user-item interaction matrix X, user social matrix S, and relationship matrix A.

[0010] S3: Begin training the model of this invention, and set an iteration cycle number E for this purpose. Each iteration cycle includes the following steps:

[0011] For each layer l in graph G, the iteration is performed as follows:

[0012] S4: For each edge in the graph, i.e. (p, q ∈ E), the memory-enhanced encoder computes the propagation information, i.e.

[0013] S5: For each user node u∈U, update the user's embedded features D based on its social and project relationship aggregation information. (l+1) [u i ];

[0014] S6: For each project node z∈Z, update the micromap embedding D based on the relationships between project nodes. (l+1) [z j ];

[0015] S7: For each type of relation r∈R, update the relation embedding feature D. (l+1) [r];

[0016] S8: After the computation of each layer (l) is completed, the embedded features H are further refined through layer normalization and self-attention mechanism. (l+1)

[0017] S9: Aggregate cross-layer embedded features to obtain the user's comprehensive feature representation D. * ;

[0018] S10: Using aggregated features D via Bayesian personalized ranking loss * Calculate the loss function L for the training batch;

[0019] S11: Update the model parameters Θ according to the loss function L;

[0020] S12: After the set training period, the algorithm returns the trained model parameters Θ;

[0021] Using the final user and item embedding features, a recommendation list is generated for each user. This typically involves the following sub-steps:

[0022] S13: For each user, the model will consider their embedded features H * [u i ] and the embedding features D of all possible items * [z j ];

[0023] S14: Calculate the affinity score between the user and each item based on the scoring function (similarity metric). The higher the score, the greater the user's preference for that item.

[0024] S15: Based on the calculated affinity score, sort all items for each user and select the items with the highest scores as recommendations;

[0025] S16: Finally, display the sorted list of micromap recommendations to the user.

[0026] This invention presents an advanced and scientific method that ensures improved dissemination efficiency of micromaps, meeting the requirements of rapid, accurate, and personalized dissemination. Experiments demonstrate that this method utilizes user social information and micromap-related information to optimize micromap recommendation efficiency, providing an effective approach to personalized dissemination of micromap products. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only schematic diagrams of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0028] Figure 1 A flowchart of a decoupled graph neural network micromap recommendation method that takes social information into account is provided by the present invention;

[0029] Figure 2 This is a schematic diagram of information transmission in the heterogeneous graph encoder of the present invention;

[0030] Figure 3 This is a schematic diagram of information aggregation in the heterogeneous graph neural network of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] The following is the symbol definition section:

[0033] The micromap recommendation method proposed in this invention uses users' previous social information, interaction data, and item information as training data. Specifically, let U and Z represent the user set and item set, respectively. Assume there exists a set of users U = {u1, u2, ..., u...}. i}, and the set of items Z = {z1, z2, ..., z} j The structure of the matrix is ​​as follows: The interaction between users and items is recorded in a matrix. ZhongX m,n =1 indicates that user u i and item z j If observed behavior (such as browsing or liking) exists, then X... m,n =0. Besides the interaction matrix X, there also exists a matrix representing the social connections between users. Each element S i,i' =1 indicates that user u i with u i' There are social connections between them; if no such connections exist, then S i,i'=0. This invention enhances social recommendation effectiveness by introducing relationships between items. Therefore, for a pair of items (z... j ,z j' This invention defines the relationship between them as a triple (z). j ,r,z j' ), where r∈R represents the intermediary node between items, representing the meta-relationship between items (such as the association relationship between item categories), and z j With z j' All belong to set Z. Items and relations (z) j The relationships between (r) constitute an item relationship matrix. in The total number of relationships.

[0034] The following steps constitute the model training process:

[0035] 1. The preparation steps are as follows:

[0036] S1: Initialize all model parameters Θ to prepare for the subsequent training process;

[0037] S2: Construct a collaborative heterogeneous graph G (G={V,E}) containing multiple relationships using the given user-item interaction matrix X, user social matrix S, and relationship matrix A.

[0038] 2. The model training steps are as follows:

[0039] S3: Begin training the model of this invention, and set an iteration cycle number E for this purpose. Each iteration cycle includes the following steps:

[0040] For each layer l in graph G, the iteration is performed as follows:

[0041] S4: For each edge in the graph, i.e. (p, q ∈ E), the memory-enhanced encoder computes the propagation information, i.e. Defined as:

[0042]

[0043] Among them, W m D is the weight matrix associated with the m-th memory unit. (l) [q] and D[p] are the embeddings of the target node and the source node, respectively, b m For bias terms;

[0044] S5: For each user node u∈U, update the user's embedded features D based on its social and project relationship aggregation information. (l+1) [u i ];

[0045] S6: For each project node z∈Z, update the micromap embedding D based on the relationships between project nodes. (l+1) [z j ];

[0046] S7: For each type of relation r∈R, update the relation embedding feature D. (l+1) [r];

[0047] S8: After the computation of each layer (l) is completed, the embedded features D are further refined through layer normalization and self-attention mechanism. (l+1)

[0048] S9: Aggregate cross-layer embedded features to obtain the user's comprehensive feature representation D. * ;

[0049] S10: Using aggregated features D through contrastive learning loss. * The loss function L for the training batch is defined as:

[0050]

[0051] Where D(q) is the embedding after neural network encoding, v + It is the embedding of positive samples associated with v. and It is the selected set of positive and negative samples, and τ is the temperature parameter;

[0052] S11: Update the model parameters Θ according to the loss function L;

[0053] S12: After the set training period, the algorithm returns the trained model parameters Θ;

[0054] 3. The recommended steps for using the micro-map are as follows:

[0055] S13: For each user, the model will consider their embedded features D * [u i ] and the embedding features D of all possible items * [z j ];

[0056] S14: Calculate the affinity score between the user and each item based on the scoring function (similarity metric). The higher the score, the greater the user's preference for that item, as defined below:

[0057] score(u,i)=D[u j ] T ·D[z i (3)

[0058] S15: Based on the calculated affinity score, sort all items for each user and select the items with the highest scores as recommendations;

[0059] S16: Finally, display the sorted list of micromap recommendations to the user.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in this scheme may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown in this scheme, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A graph neural network-based micromap recommendation method that takes social information into account, comprising three parts: preparation, model training, and micromap recommendation. The preparation steps are as follows: S1: Put all model parameters Initialization prepares the system for the subsequent training process; S2: Utilizing the given user-item interaction matrix User social matrix and relation matrix Construct a collaborative heterogeneous graph G containing multiple relationships. ; The model training steps are as follows: S3: Start training the model of this invention, and set an iteration period T for this purpose; For each layer in graph G The iterations are as follows: S4: For each edge in the graph, the memory-enhanced encoder computes the propagation information. Defined as: (1) in, Is with the first The weight matrix associated with each memory unit. , These are the embeddings of the target node and the source node, respectively. It is a bias term, and ; S5: For each user node It aggregates information based on users' social and project relationships to update users' embedded features. ; S6: For each project node Based on the relationships between project nodes, update the micromap embedding. ; S7: For each type of relation Update relation embedding features ; S8: In each layer After the computation is complete, the embedded features are further refined through layer normalization and self-attention mechanisms. ; S9: Aggregate cross-layer embedded features to obtain a comprehensive feature representation of the user. ; S10: Using aggregated features through contrastive learning loss Calculate the loss function for the training batch Defined as: （2） in, It is an embedding encoded by a neural network. It is the embedding of positive samples that are positively correlated with v. and It is the selected set of positive and negative samples. It is a temperature parameter; S11: According to the loss function Update model parameters ; S12: After the set training period, the algorithm returns the trained model parameters. ; The recommended steps for using the micro-map are as follows: S13: For each user, the model will consider their embedded features. and the embedding features of all possible items ; S14: Calculate the affinity score between the user and each item based on the scoring function. The higher the score, the greater the user's preference for that item. S15: Based on the calculated affinity score, sort all items for each user and select the items with the highest scores as recommendations; S16: Introduce a target attention mechanism to assess the user's level of attention to specific items, and integrate features obtained from contrastive learning and attention mechanisms through a multilayer perceptron to achieve the final recommendation output.

2. The graph neural network micromap recommendation method considering social information according to claim 1, characterized in that, By iterating through steps S8 to S10, memory enhancement and feature aggregation of node information in heterogeneous graphs are achieved.

3. A graph neural network micromap recommendation method that takes social information into account, as described in claim 1 or claim 2, characterized in that... In the micromap recommendation step S16, the target attention mechanism is used to assess the user's level of importance to a specific item, and the features obtained from contrastive learning and attention mechanisms are integrated through a multilayer perceptron to achieve the final recommendation output.

Images