Property transaction recommendation system and recommendation method based on graph neural network

Abstract

The invention relates to the technical field of property right transaction recommendation, in particular to a property right transaction recommendation system and method based on a graph neural network. The method comprises the steps that a data collection unit collects property transaction data; the graph construction unit constructs a graph structure according to the property transaction data; the graph neural network unit performs feature learning on the constructed graph structure by using a graph neural network GCN, extracts a high-order relationship and generates feature embedding; and the recommendation generation unit performs matching calculation through cosine similarity based on feature embedding, and generates a property right transaction recommendation result. According to the method, multiple relations (such as a transaction relation, a click behavior relation and an asset category information relation) between the user and the asset are integrated, and the influence of the different relations on the node characteristics is optimized, so that the interest and preference of the user can be more accurately captured, and in addition, the user experience is improved. Factors such as geographic positions and user behavior sequences are also considered, so that recommendation results are closer to actual demands of users.

Description

Technical Field

[0001] The present invention relates to the technical field of property rights transaction recommendation, and in particular to a property rights transaction recommendation system and recommendation method based on graph neural network. Background Art

[0002] Property transactions involve complex interactions between users and assets, which include not only direct transaction records, but also a large number of indirect interactions, such as users' browsing history, collection preferences, etc. Traditional recommendation algorithms are difficult to effectively model such multi-type and multi-level interaction patterns; property transaction data usually has high dimensionality and sparsity, that is, the characteristics of each user or asset may be very rich but the amount of data is relatively small, which makes it difficult to learn meaningful patterns from the data. Traditional machine learning methods often perform poorly in this case because they cannot fully explore the potential associations in the data; for users who have just joined the system or newly listed assets, due to the lack of sufficient historical data, it is difficult for traditional recommendation systems to give accurate recommendations. Therefore, a property transaction recommendation system and recommendation method based on graph neural networks are provided. Summary of the invention

[0003] The purpose of the present invention is to provide a property rights transaction recommendation system and recommendation method based on graph neural network, so as to solve the problem that the traditional recommendation system proposed in the above background technology is difficult to effectively model multi-type and multi-level interaction patterns and the association mining of high-dimensional sparse data.

[0004] To achieve the above objectives, on the one hand, the present invention aims to provide a property rights transaction recommendation system based on graph neural network, comprising: A data collection unit, the data collection unit is used to collect property rights transaction data; A graph construction unit, wherein the graph construction unit constructs a graph structure according to the property rights transaction data; A graph neural network unit, which uses a graph neural network GCN to perform feature learning on the constructed graph structure, extract high-order relationships and generate feature embeddings, and optimizes the graph structure feature learning process according to the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on the node features in the graph structure; The recommendation generation unit generates a property rights transaction recommendation result based on feature embedding and matching calculation through cosine similarity.

[0005] As a further improvement of the present technical solution, the property rights transaction data includes property rights transaction history data, user behavior data and asset information.

[0006] As a further improvement of the technical solution, the graph construction unit constructs a graph structure according to the property rights transaction data, including the following steps: S1.1. Generate user nodes based on transaction history and behavior data gather: ; Among them, user node Contains a list of assets purchased in the past and browsing history property; Generate asset nodes based on transaction history and asset information gather: ; Among them, asset nodes Includes asset classes ; S1.2. Extract the transaction relationship between users and assets from the historical data of property rights transactions: ; in, Represents the set of all transaction records, represents the user node, Represents an asset node, Indicates the transaction amount. Indicates the transaction time. Represents a user node The index in Represents an asset node The index in Represents the attribute index on the edge; Extract the click behavior relationship between users and assets from user behavior data: ; in, Represents the collection of all click records, Represents a user node For asset nodes Number of clicks; Extract asset category features from asset information: ; in, Represents the collection of all assets, Represents an asset node Category of; S1.3. Generate transaction edges based on transaction history data : ; Generate behavior edges based on user behavior data : ; S1.4. Use the tool NetworkX to load nodes and edges and build a graph structure: ; in, A representation of the graph structure, Represents a collection of user nodes and a collection of asset nodes A collection of Indicates the transaction side and behavior edge A collection of .

[0007] As a further improvement of the technical solution, the graph neural network unit includes a graph convolution module and a high-order relationship extraction module; The graph convolution module receives the graph structure data generated by the graph construction unit, and learns the features of the user nodes and the asset nodes in the graph structure data through information transmission; The high-order relationship extraction module extracts high-order relationships between user nodes and asset nodes in graph structure data based on a multi-order neighbor aggregation method.

[0008] As a further improvement of the technical solution, the graph convolution module receives and processes the graph structure data generated by the graph construction unit, and learns the features of the nodes in the graph structure data through information transfer, including the following steps: S2.1. Vectorize the attributes of each node. For each user node , the eigenvector is , for each asset node , the eigenvector is ; S2.2. The receiving module parses the attributes of the edge and the attributes of the transaction edge It is expressed as: , the attributes of the behavior edge It is expressed as: ; S2.3. Embed the data of user nodes, asset nodes, transaction edges and behavior edges into vector representations. The initial node embedding vector for: , the edge embedding vector is: ; S2.4. The structure of the graph Represented as an adjacency matrix ; S2.5. Adjacency Matrix Normalize the embedding vectors; S2.6. The graph convolution formula is used to propagate and update feature vectors. The graph convolution formula is optimized based on the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features. The graph convolution formula is further optimized by considering the synergy and conflict between different types of relationships. S2.7. Graph convolution module outputs enhanced node feature embedding : .

[0009] As a further improvement of this technical solution, in S2.6, the graph convolution formula is: ; in, represents the node embedding feature, represents the normalized adjacency matrix, Indicated in The node feature matrix of the layer, Indicated in The node feature matrix of the layer, Indicates The weight matrix of the layer, ReLU activation function. The graph convolution formula is optimized according to the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features: ; in, Indicates that after optimization The node embedding features of the layer, Representing relationships The corresponding normalized adjacency matrix is, Indicates the relationship type The weight coefficient of An index representing the type of relationship, Represents the total number of relationship types, Representing relationships The corresponding weight matrix; Considering the collaboration and conflict between different types of relationships, the graph convolution formula is further optimized: ; in, Indicates that after further optimization The node embedding features of the layer, Represents the user node and asset node index, Representing relationships and The interaction matrix between Represents the weight coefficient.

[0010] As a further improvement of the technical solution, the high-order relationship extraction module extracts high-order relationships between nodes in the graph structure data based on a multi-order neighbor aggregation method, including the following steps: S2.8. Receive the node embedding features output by the graph convolution module and the adjacency matrix ; S2.9. Define the propagation matrix of g-order neighbors : ,in, Represents the adjacency matrix Power, represents the order of neighbors; S2.10. Aggregate the features of the g-order neighbors of each user node and asset node into the embedding of the current user node and asset node: ,in, represents the feature representation obtained by propagation through g-order neighbors, Represents the linear transformation weight of the g-order feature; S2.11. Generate a path sequence of user nodes and asset nodes through random walks, learn the potential relationship between user nodes and asset nodes, and extract global structural information to obtain the global structural feature embedding matrix ; S2.12. Combining high-order neighbor aggregation and global structural information: ; in, represents the enhanced node embedding matrix, represents the node embedding features output by the graph convolution module, represents the feature representation obtained by propagation through g-order neighbors, represents the global structural feature embedding matrix, represents the weight coefficient of the node's high-order neighbor g-order information, The weight coefficient representing the global structural information; Get the enhanced node embedding matrix: .

[0011] As a further improvement of the technical solution, the recommendation generation unit generates property rights transaction recommendation results based on feature embedding and matching calculation by cosine similarity, including the following steps: S3.1. Receive user node embedding matrix , asset node embedding matrix and graph structure ; S3.2, calculating the similarity score between the user embedding matrix and the asset embedding matrix through the cosine similarity of the user embedding matrix and the asset embedding matrix; S3.3, find the user's candidate assets by embedding the k-nearest neighbor method; S3.4, sorting the candidate assets according to the similarity scores and generating a final recommendation list; S3.5. Output a recommended list of property rights transactions sorted by similarity scores.

[0012] As a further improvement of the technical solution, in S3.2, the cosine similarity between the user embedding matrix and the asset embedding matrix is: ; in, Represents a user node The embedding matrix of Represents an asset node The embedding matrix of represents the similarity score; Optimize cosine similarity based on the user's geographic location and behavior sequence: ; in, represents the optimized similarity score, Indicates based on user node The feature vector is constructed based on the geographic location and behavior sequence.

[0013] On the other hand, the present invention provides a method for recommending property rights transactions based on a graph neural network, which is used in any one of the above-mentioned property rights transaction recommendation systems based on a graph neural network, and comprises the following steps: S4.1. Collect property transaction data; S4.2. Construct a graph structure based on property rights transaction data; S4.3. Use the graph neural network GCN to learn the features of the constructed graph structure, extract high-order relationships and generate feature embeddings. Optimize the graph structure feature learning process based on the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on the node features in the graph structure; S4.4. Based on feature embedding, matching calculation is performed to generate property rights transaction recommendation results.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. In the property transaction recommendation system and recommendation method based on graph neural network, by integrating multiple relationships between users and assets (such as transaction relationships, click behavior relationships, and asset category information relationships), and optimizing the impact of these different relationships on node features, it is possible to more accurately capture user interests and preferences. In addition, it also takes into account factors such as geographic location and user behavior sequence, making the recommendation results closer to the actual needs of users and improving the personalization of recommendations. This meticulous relationship processing and feature learning ensures that the output of the recommendation system and method not only meets the long-term interests of users but also reflects their short-term intentions, greatly improving the accuracy and relevance of recommendations.

[0015] 2. In the property transaction recommendation system and recommendation method based on graph neural network, by introducing high-order neighbor aggregation and global structural information extraction mechanism, it not only focuses on the information propagation of directly adjacent nodes, but also effectively utilizes multi-hop neighbors and global structural features, which helps to capture the potential but important connection between users and assets. Such a design enhances the model's ability to learn complex relationship patterns and avoids recommendation bias caused by local view limitations. At the same time, by optimizing the recommendation algorithm to reduce search costs and improve user satisfaction, it provides users with a smoother and more efficient property transaction exploration platform, significantly improving the overall user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall flow chart of the present invention; The meaning of each number in the figure is: 1. Data collection unit; 2. Graph construction unit; 3. Graph neural network unit; 31. Graph convolution module; 32. High-order relationship extraction module; 4. Recommendation generation unit. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Example 1: Please refer to Figure 1 As shown, a property rights transaction recommendation system based on graph neural network is provided, including: The data collection unit 1 is used to collect property rights transaction data; In this embodiment, the property rights transaction data includes property rights transaction history data, user behavior data and asset information.

[0019] The graph construction unit 2 constructs a graph structure according to the property rights transaction data; In this embodiment, the graph construction unit 2 constructs a graph structure according to the property rights transaction data, including the following steps: S1.1. Generate user node set based on transaction history and behavior data (Each user node contains a list of assets purchased in the past and browsing history): ; Among them, user node Contains a list of assets purchased in the past and browsing history property; Generate asset node set based on transaction history and asset information (Asset nodes contain asset classes): ; Among them, asset nodes Includes asset classes ; S1.2. Extract the transaction relationship between users and assets from the historical data of property rights transactions: ; in, Represents the set of all transaction records, represents the user node, Represents an asset node, Indicates the transaction amount. Indicates the transaction time. Represents a user node The index in Represents an asset node The index in Represents the attribute index on the edge (transaction edge or behavior edge); Extract the click behavior relationship between users and assets from user behavior data: ; in, Represents the collection of all click records, Represents a user node For asset nodes Number of clicks; Extract asset category features from asset information: ; in, Represents the collection of all assets, Represents an asset node Category of; S1.3. Generate transaction edges based on transaction history data , the transaction edge contains the transaction amount and transaction time attributes: ; Generate behavior edges based on user behavior data , if the user has only browsed or performed other non-transactional interactions on an asset, a behavior edge is created. Such edges can help capture the user’s potential interest, even if they have not yet converted into actual transactions: ; S1.4. Use the tool NetworkX to load nodes and edges and build a graph structure: ; in, A representation of the graph structure, Represents a collection of user nodes and a collection of asset nodes A collection of Indicates the transaction side and behavior edge A collection of .

[0020] Graph neural network unit 3 uses graph neural network GCN to learn features of the constructed graph structure, extract high-order relationships and generate feature embeddings; In this embodiment, the graph neural network unit 3 includes a graph convolution module 31 and a high-order relationship extraction module 32; The graph convolution module 31 receives the graph structure data generated by the graph construction unit 2, and learns the features of the user nodes and asset nodes in the graph structure data through information transmission. The graph convolution module learns the features by aggregating the neighbor node information of each node, so that the feature embedding of each node not only contains its own attribute information, but also integrates the features of its neighboring nodes in the graph structure, so as to capture the complex relationship between users and assets. The high-order relationship extraction module 32 extracts the high-order relationship between the user nodes and the asset nodes in the graph structure data based on the multi-order neighbor aggregation method. The graph neural network gradually integrates the information of the distant nodes through multi-layer propagation. These distant relationships are “high-order relationships”. Furthermore, learning the features of nodes in graph-structured data through information transfer can effectively capture the complex nonlinear relationships and dependencies between nodes, which is crucial for applications such as understanding user behavior, recommendation systems, and risk assessment. It makes full use of the structural information in graph data, not just the attributes of nodes, which helps to reveal the relationship patterns hidden in the data, such as the community structure in social networks or the abnormal patterns in financial transaction networks. The graph convolution module 31 receives and processes the graph structure data generated by the graph construction unit 2, and learns the features of the nodes in the graph structure data through information transfer, including the following steps: S2.1. Vectorize the attributes of each node. For each user node , the eigenvector is , for each asset node , the eigenvector is ; S2.2. The receiving module parses the attributes of the edge and the attributes of the transaction edge It is expressed as: , the attributes of the behavior edge It is expressed as: ; S2.3. Embed the data of user nodes, asset nodes, transaction edges and behavior edges into vector representations. The initial node embedding vector for: , the edge embedding vector is: ; S2.4. The structure of the graph Represented as an adjacency matrix ,in Representation Node and The connection relationship (transaction or behavior edge) between The adjacency matrix of a graph with nodes is A matrix where the value of each element is determined by the presence or absence of an edge: Adjacency Matrix : ; in, , represents the set of all edges; Graph structure data output format: Node feature matrix : ; Adjacency Matrix : ; S2.5. Adjacency Matrix Normalize the embedding vectors; S2.6. Propagation and feature vector update are performed through the graph convolution formula. The graph convolution formula uses the topological structure and node feature information of the graph. Through neighborhood aggregation and feature transformation, the node features are gradually integrated with the information of its neighboring nodes, thereby learning feature embedding that reflects the global graph structure. The graph convolution formula is optimized for the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features. The graph convolution formula is further optimized by considering the synergy and conflict between different types of relationships. The graph convolution operation allows each node to aggregate information from its neighbors. In this way, the node can "sense" the structure of its local neighborhood and learn feature representations based on that structure; since the graph convolution operation is based on the relationship between nodes rather than the absolute position, it is invariant to the rearrangement of nodes in the graph, which means that no matter how the nodes are arranged, as long as the connection relationship between them remains unchanged, the learned features will be the same; it can naturally process graphs of different sizes and structures without pre-defining a fixed input size or shape, which facilitates solving diverse problems in practical applications; the graph convolution formula can combine the characteristics of the node itself (such as node attributes) with its position in the graph (i.e., the topological structure), thereby more comprehensively characterizing the characteristics of the node; Furthermore, the graph convolution formula is: ; in, represents the node embedding feature, represents the normalized adjacency matrix, ,in, is the degree matrix, Indicated in The node feature matrix of the layer, Indicated in The node feature matrix of the layer (where each row represents the feature vector of a node), the feature of each node is obtained by information transmission and aggregation of its neighboring nodes, Indicates The weight matrix of the layer, which is used to linearly transform the features of each node, ReLU activation function. The interaction between users and assets in the property transaction recommendation system is very complex. The transaction relationship directly reflects the user's purchase decision, while the click behavior more represents the user's interest exploration process. The category information of the asset provides static features about the asset itself, which is crucial for understanding the user's preferences. By distinguishing and processing these different types of relationships, the diversity of user behaviors can be captured more accurately; each relationship (such as transaction relationship, click behavior relationship, category information relationship) may have a different effect on the feature update of the node in the recommendation system. For example, the transaction relationship may better reflect the user's actual preference for the asset, while the click behavior relationship may reflect the user's interest level; in the graph neural network, the information propagation between nodes is carried out according to the adjacency matrix and the feature matrix. Each relationship (transaction, click, category information) may affect the characteristics of the node in different ways. Therefore, simply merging these relationships into a unified propagation method may lead to over- or under-learning of certain relationships. In order to make the model better adapt to the impact of different relationships on the features, different weight coefficients need to be given to different types of relationships to reflect the importance of these relationships; The graph convolution formula is optimized according to the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features: ; in, Indicates that after optimization The node embedding features of the layer, Representing relationships The corresponding normalized adjacency matrix is, Indicates the relationship type The weight coefficient of , which indicates the influence of each relationship on the node feature update, An index representing the type of relationship, including the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets. Represents the total number of relationship types, Representing relationships The corresponding weight matrix; Different types of relationships may produce synergistic effects in some cases, or they may have conflicting effects. For example, if a user clicks on a certain asset category multiple times and eventually trades, the transaction relationship and the click behavior relationship may jointly strengthen the interest in the asset in this case. Therefore, there may be a synergistic effect between the click behavior relationship and the transaction relationship to enhance the accuracy of the recommendation; in some cases, the click behavior may conflict with the transaction behavior. For example, a user clicks on an asset but does not ultimately purchase it, which may indicate that the user is not really interested in the asset. In this case, there may be a certain conflict between click behavior and transaction behavior, and their influence should be balanced by appropriately adjusting the weight coefficient to prevent the system from mistakenly taking click behavior as a strong signal of purchase behavior; different relationships do not exist independently, and there may be complex interactions between them. By introducing the relationship interaction matrix, the interaction between these relationship types can be captured. For example, the interaction between the transaction relationship and the click behavior relationship may indicate the connection between the user's actual purchase behavior and his potential interest. Therefore, when optimizing the graph convolution formula, it is necessary to take into account the interactive information of different relationships to prevent a certain relationship type from excessively affecting the node feature update or causing conflicts between different relationships; Considering the synergy and conflict between different types of relationships (transaction relationship between users and assets, click behavior relationship between users and assets, and category information relationship between assets), the graph convolution formula is further optimized: ; in, Indicates that after further optimization The node embedding features of the layer, Represents the user node and asset node index, Representing relationships and The interaction matrix between them can capture the synergies and conflicts between different relationship types. Represents the weight coefficient, represents the node For relationships the degree of dependence; Defining a relationship interaction matrix , used to describe two relationships and Interaction strength between: and Represents two relationship types, for example, It can be a transactional relationship. It can be a click behavior relationship For each relationship type (such as transactions, clicks, category information, etc.), first construct the corresponding adjacency matrix , wait; The dot product method is used to calculate the interaction information between two or more relationships to obtain a relationship interaction matrix ,The matrix is ​​used to adjust the influence of different relationships in the information dissemination process and prevent the interference of conflicting relationships; S2.7. Graph convolution module outputs enhanced node feature embedding : .

[0021] Furthermore, the multi-order neighbor aggregation method recursively aggregates the information of each node's multi-hop neighbors to capture indirect relationships between nodes that go beyond direct connections, thereby revealing deeper structures and patterns in the graph; by aggregating the information of multi-hop (i.e., multiple edge connections) neighbors, the more distant indirect relationships between nodes can be effectively captured, which is crucial for understanding the global structure in complex networks; the introduction of high-order neighbor information enables the model to learn richer feature representations, thereby enhancing the understanding of node and edge attributes and improving the performance of tasks such as classification, clustering, or link prediction; in many practical applications, important relationships may not be directly reflected in the adjacency matrix, and multi-order neighbor aggregation can help discover those implicit but important association patterns, such as community structures in social networks or functional modules in biological networks; for recommendation systems, the potential connections between users and items often exist in high-order neighbors. By considering these high-order relationships, recommendation algorithms can provide more accurate and personalized suggestions; The high-order relationship extraction module 32 extracts high-order relationships between nodes in the graph structure data based on a multi-order neighbor aggregation method, including the following steps: S2.8, receiving the node embedding features output by the graph convolution module 31 and the adjacency matrix ; S2.9, in order to capture the high-order relationship between nodes, the module aggregates the information of multi-hop neighbors and defines the propagation matrix of g-order neighbors: , g-order neighbors refer to the set of nodes connected to the current node through g-hop paths, where Represents the adjacency matrix Power, Indicates the order of neighbors; S2.10. Aggregate the features of the g-order neighbors of each user node and asset node into the embedding of the current user node and asset node: ,in, represents the feature representation obtained by propagation through g-order neighbors, Represents the linear transformation weight of the g-order feature; S2.11. Generate a path sequence of user nodes and asset nodes through random walks, learn the potential relationship between user nodes and asset nodes, and extract global structural information to obtain the global structural feature embedding matrix (First, the Perform random walks to generate node path sequences , then, through the path sequence The co-occurrence probability of a node with other nodes is calculated, and the Skip-Gram method is used for embedding learning, and finally the global embedding of the node is learned. The path sequence generated by the random walk reflects the potential relationship between the nodes. These relationships may not be directly reflected in the adjacency matrix, but through multiple random walks, the global structure of the nodes in the graph can be revealed. The starting point can be a user node or an asset node. According to the edge weight and transition probability of the graph, jumps are made between user nodes and asset nodes. For example: from a user node Start and jump to the asset node it interacted with , from the asset node Start and jump to other user nodes that interact with it , if the user Purchased assets , while users Also viewed assets , then the random walk may generate the following paths: , where the path contains the user node , , and asset nodes ; S2.12. Combining high-order neighbor aggregation and global structural information: ; in, represents the enhanced node embedding matrix, represents the node embedding features output by the graph convolution module 31, represents the feature representation obtained by propagation through g-order neighbors, represents the global structural feature embedding matrix, The weight coefficient of the g-order information of the node's high-order neighbors is used to control the influence of different-order neighbor characteristics on the current node. The weight coefficient representing the global structural information indicates the contribution of the global feature to the node embedding; Get the enhanced node embedding matrix: .

[0022] The recommendation generating unit 4 performs matching calculation based on feature embedding and generates a property rights transaction recommendation result; In this embodiment, the similarity between property transaction items is quantified by calculating the cosine similarity between feature embedding vectors, so as to recommend the transaction opportunities that best match the user's interests or historical behaviors; the feature embedding generated by the graph neural network can capture the complex nonlinear relationship between users and assets, including direct and indirect interaction modes, which makes the recommendation more accurate; cosine similarity measures the angular difference between two vectors rather than their absolute size, which means that it focuses more on the consistency of feature directions rather than numerical distances, thereby better reflecting the semantic similarity between data points; using feature embedding combined with cosine similarity can effectively identify the assets most relevant to the user's interests, and provide highly relevant recommendations even if these assets have not been directly interacted with by the user before; The recommendation generation unit 4 generates a property rights transaction recommendation result by performing matching calculation based on feature embedding and cosine similarity, including the following steps: S3.1. Receive user node embedding matrix (represents the feature embedding of all user nodes, the shape is ,in is the number of user nodes, is the embedding dimension), asset node embedding matrix (represents the feature embedding of all asset nodes, the shape is ,in is the number of asset nodes, is the embedding dimension) and the graph structure ; S3.2, calculating the similarity score between the user embedding matrix and the asset embedding matrix through the cosine similarity of the user embedding matrix and the asset embedding matrix; Furthermore, the cosine similarity between the user embedding matrix and the asset embedding matrix is: ; in, Represents a user node The embedding matrix of Represents an asset node The embedding matrix of represents the similarity score; In many cases, property transactions (such as real estate, vehicles, etc.) are closely related to geographic location. Users are usually more interested in assets in their area or nearby areas. By taking geographic location information into consideration, it can ensure that the recommendation results are closer to the actual needs of users, thereby improving the relevance and practicality of recommendations; users' past behavior patterns can well reflect their interests and preferences. For example, if a user often views a certain type of property or a specific brand of car, the system should prioritize these factors to provide personalized recommendations. Utilizing behavioral sequence information can help capture users' long-term interests and short-term intentions, thereby generating more relevant recommendations; one of the goals of the recommendation system is to provide unique and valuable suggestions for each user. Relying solely on general feature embeddings may cause recommendations to be too generalized and fail to fully meet the needs of individual users. By adding information about geographic location and behavioral sequence, the recommendation results can be more personalized and better adapted to the unique preferences of different users; through optimized recommendations, users can more easily find assets they are interested in, reducing search costs and improving satisfaction; Optimize cosine similarity based on the user's geographic location and behavior sequence: ; in, represents the optimized similarity score, Indicates based on user node The feature vector constructed by the geographical location and behavior sequence; Furthermore, the geographic space is divided into grid cells, and use one-hot encoding to represent the set of areas visited by each user. , and obtain a length of The binary vector , among which A bit of 1 indicates that the user has visited grid cells; definition For user nodes In the The number of grid cells that appear, the position frequency distribution vector is written as: ; Introducing the time decay factor , so that newer access records have higher weights, if Is a user node Last visit The timestamp of each grid, then the time-aware position feature is written as: ,in, is the current timestamp, is the time decay coefficient, and the final position eigenvector for: , is the weight of the position frequency distribution, is the time-aware position feature weight; If it is a set of all possible behavior types, for the user Define the behavior frequency vector , , Represents a user node Execute The number of such behaviors; Each behavior Mapped to an embedding vector of fixed dimension , the behavior sequence becomes an embedding matrix , is the embedding dimension, and the long short-term memory network LSTM model is used to process this embedding matrix: ,in, Indicates that at time step The hidden state of the previous time step depends on the hidden state of and the current behavior embed ; Finally, take the hidden state of the last time step As a user node Behavior sequence feature representation ; Combining the above location features and behavior sequence features, construct a The feature vector constructed by the geographic location and behavior sequence , and the eigenvector Perform normalization; S3.3, finding candidate assets for the user by embedding a k-nearest neighbor method, where the embedded k-nearest neighbor method determines the assets most similar to the user based on a distance metric in the feature space; S3.4, sorting the candidate assets according to the similarity scores and generating a final recommendation list; S3.5. Output a recommended list of property rights transactions sorted by similarity score. The recommendation result may be a list of user-asset pairs, for example: .

[0023] Example 2: The difference between Example 2 of the present invention and Example 1 is that this example introduces a graph neural network property rights transaction recommendation method used in a property rights transaction recommendation system based on a graph neural network.

[0024] A method for recommending property rights transactions based on a graph neural network, used in any of the above-mentioned property rights transaction recommendation systems based on a graph neural network, comprises the following steps: S4.1. Collect property transaction data; S4.2. Construct a graph structure based on property rights transaction data; S4.3. Use the graph neural network GCN to learn the features of the constructed graph structure, extract high-order relationships and generate feature embeddings. Optimize the graph structure feature learning process based on the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on the node features in the graph structure; S4.4. Based on feature embedding, matching calculation is performed to generate property rights transaction recommendation results.

[0025] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention, and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A property rights transaction recommendation system based on graph neural network, characterized in that: include: A data collection unit (1), the data collection unit (1) being used to collect property rights transaction data; A graph construction unit (2), wherein the graph construction unit (2) constructs a graph structure according to the property rights transaction data; A graph neural network unit (3), wherein the graph neural network unit (3) uses a graph neural network GCN to perform feature learning on the constructed graph structure, extracts high-order relationships and generates feature embeddings, and optimizes the graph structure feature learning process according to the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on the node features in the graph structure; A recommendation generation unit (4), wherein the recommendation generation unit (4) performs matching calculation based on feature embedding through cosine similarity to generate a property rights transaction recommendation result.

2. The property rights transaction recommendation system based on graph neural network according to claim 1 is characterized by: The property rights transaction data includes property rights transaction history data, user behavior data and asset information.

3. The property rights transaction recommendation system based on graph neural network according to claim 2 is characterized in that: The graph construction unit (2) constructs a graph structure according to the property rights transaction data, including the following steps: S1.

1. Generate user node set based on transaction history and behavior data : ； Among them, user node Contains a list of assets purchased in the past and browsing history property; Generate asset node set based on transaction history and asset information : ； Among them, asset nodes Includes asset classes ; S1.

2. Extract the transaction relationship between users and assets from the historical data of property rights transactions: ； in, Represents the set of all transaction records, represents the user node, Represents an asset node, Indicates the transaction amount. Indicates the transaction time. Represents a user node The index in Represents an asset node The index in Represents the attribute index on the edge; Extract the click behavior relationship between users and assets from user behavior data: ； in, Represents the collection of all click records, Represents a user node For asset nodes Number of clicks; Extract asset category features from asset information: ； in, Represents the collection of all assets, Represents an asset node Category of; S1.

3. Generate transaction edges based on transaction history data : ； Generate behavior edges based on user behavior data : ； S1.

4. Use the tool NetworkX to load nodes and edges and build a graph structure: ； in, A representation of the graph structure, Represents a collection of user nodes and a collection of asset nodes A collection of Indicates the transaction side and behavior edge A collection of .

4. The property rights transaction recommendation system based on graph neural network according to claim 3 is characterized by: The graph neural network unit (3) includes a graph convolution module (31) and a high-order relationship extraction module (32); The graph convolution module (31) receives the graph structure data generated by the graph construction unit (2), and learns the features of the user nodes and asset nodes in the graph structure data through information transmission; The high-order relationship extraction module (32) extracts high-order relationships between user nodes and asset nodes in graph structure data based on a multi-order neighbor aggregation method.

5. The property rights transaction recommendation system based on graph neural network according to claim 4 is characterized in that: The graph convolution module (31) receives and processes the graph structure data generated by the graph construction unit (2), and learns the features of the nodes in the graph structure data through information transfer, including the following steps: S2.

1. Vectorize the attributes of each node. For each user node , the eigenvector is , for each asset node , the eigenvector is ; S2.

2. The receiving module parses the attributes of the edge and the attributes of the transaction edge It is expressed as: , the attributes of the behavior edge It is expressed as: ; S2.

3. Embed the data of user nodes, asset nodes, transaction edges and behavior edges into vector representations. The initial node embedding vector for: , the edge embedding vector is: ; S2.

4. The structure of the graph Represented as an adjacency matrix ; S2.

5. Adjacency Matrix Normalize the embedding vectors; S2.

6. The graph convolution formula is used to propagate and update feature vectors. The graph convolution formula is optimized based on the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features. The graph convolution formula is further optimized by considering the synergy and conflict between different types of relationships. S2.

7. Graph convolution module outputs enhanced node feature embedding : 。 6. The property rights transaction recommendation system based on graph neural network according to claim 5 is characterized by: In S2.6, the graph convolution formula is: ； in, represents the node embedding feature, represents the normalized adjacency matrix, Indicated in The node feature matrix of the layer, Indicated in The node feature matrix of the layer, Indicates The weight matrix of the layer, ReLU activation function. The graph convolution formula is optimized according to the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on node features: ； in, Indicates that after optimization The node embedding features of the layer, Representing relationships The corresponding normalized adjacency matrix is, Indicates the relationship type The weight coefficient of An index representing the type of relationship, Represents the total number of relationship types, Representing relationships The corresponding weight matrix; Considering the collaboration and conflict between different types of relationships, the graph convolution formula is further optimized: ； in, Indicates that after further optimization The node embedding features of the layer, Represents the user node and asset node index, Representing relationships and The interaction matrix between Represents the weight coefficient.

7. The property rights transaction recommendation system based on graph neural network according to claim 6 is characterized by: The high-order relationship extraction module (32) extracts high-order relationships between nodes in the graph structure data based on a multi-order neighbor aggregation method, comprising the following steps: S2.

8. Receive the node embedding features output by the graph convolution module (31) and the adjacency matrix ; S2.

9. Define the propagation matrix of g-order neighbors : ,in, Represents the adjacency matrix Power, represents the order of neighbors; S2.

10. Aggregate the features of the g-order neighbors of each user node and asset node into the embedding of the current user node and asset node: ,in, represents the feature representation obtained by propagation through g-order neighbors, Represents the linear transformation weight of the g-order feature; S2.

11. Generate a path sequence of user nodes and asset nodes through random walks, learn the potential relationship between user nodes and asset nodes, and extract global structural information to obtain the global structural feature embedding matrix ; S2.

12. Combining high-order neighbor aggregation and global structural information: ； in, represents the enhanced node embedding matrix, represents the node embedding features output by the graph convolution module (31), represents the feature representation obtained by propagating through g-order neighbors, represents the global structural feature embedding matrix, represents the weight coefficient of the node's high-order neighbor g-order information, The weight coefficient representing the global structural information; Get the enhanced node embedding matrix: 。 8. The property rights transaction recommendation system based on graph neural network according to claim 7 is characterized by: The recommendation generation unit (4) generates a property rights transaction recommendation result by performing matching calculation based on feature embedding and cosine similarity, including the following steps: S3.

1. Receive user node embedding matrix , asset node embedding matrix and graph structure ; S3.2, calculating the similarity score between the user embedding matrix and the asset embedding matrix through the cosine similarity of the user embedding matrix and the asset embedding matrix; S3.3, find the user's candidate assets by embedding the k-nearest neighbor method; S3.4, sorting the candidate assets according to the similarity scores and generating a final recommendation list; S3.

5. Output a recommended list of property rights transactions sorted by similarity scores.

9. The property rights transaction recommendation system based on graph neural network according to claim 8 is characterized by: In S3.2, the cosine similarity between the user embedding matrix and the asset embedding matrix is: ； in, Represents a user node The embedding matrix of Represents an asset node The embedding matrix of represents the similarity score; Optimize cosine similarity based on the user's geographic location and behavior sequence: ； in, represents the optimized similarity score, Indicates based on user node The feature vector is constructed based on the geographic location and behavior sequence.

10. A method for recommending property rights transactions based on a graph neural network, used in a property rights transaction recommendation system based on a graph neural network as claimed in any one of claims 1 to 9, characterized in that: The steps include: S4.

1. Collect property transaction data; S4.

2. Construct a graph structure based on property rights transaction data; S4.

3. Use the graph neural network GCN to learn the features of the constructed graph structure, extract high-order relationships and generate feature embeddings. Optimize the graph structure feature learning process based on the different effects of the transaction relationship between users and assets, the click behavior relationship between users and assets, and the category information relationship of assets on the node features in the graph structure; S4.

4. Based on feature embedding, matching calculation is performed to generate property rights transaction recommendation results.

Images