A Diffusion Graph Recommendation Method and System Based on a Multi-Hop Mechanism

Abstract

The present invention provides a diffusion graph recommendation method and system based on a multi-hop mechanism, belonging to the technical fields of deep learning, knowledge graphs, and recommendation systems. The present invention consists of a knowledge graph construction, a feature extraction network, and a multi-hop mechanism diffusion model; different types of POI features are respectively input into different feature extraction hidden layers, and operations such as different weight assignments, embedding mappings, and non-linear transformations are performed on them to better capture the feature differences of different types of interest points, thereby improving the expression ability and generalization ability of the model; through the information transmission of the graph, the user portrait is enriched, potential associations are discovered, and personalized recommendations are made; the water wave diffusion algorithm enables the model to have the ability to find more entities to improve its generalization performance; the multi-hop model can gradually expand and enrich the user's set of interest points, mine more potential interest points related to the user's historical behavior, effectively utilize the relationship information in the user and the knowledge graph, and achieve personalized and diversified recommendations.

Description

Technical Field

[0001] The present invention belongs to the technical fields of deep learning, knowledge graph and recommendation system, and particularly relates to a diffusion graph recommendation method and system based on a multi-hop mechanism. Background Art

[0002] With the rapid development of information technology and the wide popularization of Internet applications, users are faced with a vast amount of information and choices in their daily lives. In this era of information explosion, in order to accurately obtain information that users are interested in, points of interest are often used to describe personal information such as user information and preferences. As an important branch of the recommendation system, point-of-interest recommendation aims to recommend content that users may be interested in according to information such as users' interests, preferences, and historical behaviors, so as to help users obtain useful information more efficiently, and thus recommendation algorithms came into being. With the rise of knowledge graph technology, point-of-interest recommendation algorithms based on knowledge graphs have gradually attracted attention. A knowledge graph is a semantic network that can describe concepts, entities in the real world, and the relationships between them, providing rich semantic information and context information for the recommendation system. By combining the knowledge graph with the recommendation algorithm, the interests and needs of users can be understood more accurately, and the accuracy and diversity of recommendations can be improved.

[0003] In recent years, personalized recommendation systems and knowledge graphs have made remarkable progress in both academia and industry. Traditional recommendation algorithms mainly make recommendations based on users' historical behavior data, such as collaborative filtering, content filtering, etc. In the prior art, a rating preference model and a time weight factor are introduced to reconstruct the user-item matrix, optimizing the traditional collaborative filtering algorithm, but there is still the problem of incomplete mining of user preferences. To address such problems, Li Xiangkun et al. proposed a collaborative filtering method (OCRIF) that fuses overlapping community regularization and implicit feedback. This method considers the community structure of users in the community network and the implicit feedback of users' rating information and social information, further improving the performance of collaborative filtering algorithms. However, these methods have limitations in dealing with complex semantic information and context information, and cannot fully understand the interests and needs of users, and there are common problems such as difficulty in high-order modeling and insufficient modeling of user characteristics. Summary of the Invention

[0004] To solve the above problems, the present invention proposes a diffusion graph recommendation method based on a multi-hop mechanism, which mines high-order semantic information in the knowledge graph in an end-to-end manner, covering three parts: knowledge graph construction, feature extraction network construction, and multi-hop mechanism diffusion model construction, specifically including:

[0005] Construct a knowledge graph containing basic information of various points of interest and integrating semantic information and social relationships;

[0006] Construct a POI-GCN feature extraction network based on the graph convolutional neural network to capture the deep semantic information in the knowledge graph;

[0007] Based on the features extracted by the POI-GCN feature extraction network, introduce the water wave diffusion and multi-hop mechanism to construct a Diffussion-POI recommendation model to generate a recommendation sequence.

[0008] Furthermore, the construction method of the knowledge graph is as follows:

[0009] Combine the crawled Yelp2018 dataset, Gowalla dataset, public commercial address data, and open platform merchant data obtained using the MeiTuan-POI interface to form a knowledge graph dataset. The knowledge graph dataset consists of structured point-of-interest data and unstructured and semi-structured data, providing high-quality data for subsequent model training;

[0010] Perform data preprocessing on the data in the knowledge graph dataset;

[0011] Construct the triples <entity, relationship, attribute> that make up the knowledge graph according to the preprocessed text results.

[0012] Furthermore, the POI-GCN feature extraction network includes a word embedding layer, a network layer, a GCN layer, and a hidden layer;

[0013] For each point of interest, extract features including the name, location, category, and rating of the point of interest from the triples it belongs to, and then map these features into vectors of a fixed length and input them into the word embedding layer to obtain the embedding vector of the entity, which is used as the input of the network layer;

[0014] An attention mechanism is added in the network layer for weight assignment, and the calculation formula is as follows:

[0015]

[0016] where, a ij represents the attention weight; softplus is the activation function; e ij represents the similarity between user i and item j, and the formula is as follows:

[0017] e ij = ELU(a T |Wh r ||Wh i |) (2)

[0018]

[0019] where, ELU is the activation function; W is the linear transformation matrix, and its shape is W ∈ RF×F′ , where F is the dimension of the input features and F' is the dimension of the output features; Wh r represents the element in the h-th row and r-th column of the W matrix; Wh i represents the element in the h-th row and i-th column of the W matrix; a represents the parameter vector; f(x) represents the ELU activation function, which adjusts the similarity value e through a non-linear transformation ij , to better calculate the attention weights; exp represents the natural exponential function; γ is the scaling factor;

[0020] After weight assignment, the weighted entity embedding vector is output, and then message aggregation is performed to obtain the aggregated features;

[0021] The GCN layer performs feature fusion on the obtained aggregated features and outputs the entity feature vector of the fused features;

[0022] The hidden layer uses the ReLU activation function to perform non-linear transformation on the entity feature vector of the fused features, obtaining the feature vector that fuses all entity information.

[0023] Furthermore, a water wave diffusion and multi-hop mechanism is introduced to construct the Diffussion-POI recommendation model, and the process of generating the recommendation sequence is as follows:

[0024] S10: Extract the knowledge graph features according to the feature vector extracted by the POI-GCN feature extraction network

[0025] S101: Perform entity linking on the entities in the knowledge graph to obtain the matched entities;

[0026] S102: Extract a subgraph from the knowledge graph after entity linking;

[0027] S103: Introduce context text features for the entities in the subgraph:

[0028] context(n) = {n m |(n, r, n m ) ∈ G or (n m , r, n) ∈ G}

[0029] where context(n) is a function for describing the context features of entity n; G represents the graph; r is the relationship between nodes; n m represents the m-th entity; (n, r, n m ) represents a triple, entity n, relationship r and another entity n m ;

[0030] S20: Perform information transfer of the graph

[0031] S201: Set the initial value of all entity nodes after introducing context text features to 0, and set the initial value of the seed node to 1;

[0032] S202: Iteratively update the values of each entity node. In each iteration, for any entity node v, calculate the contribution value d(x, v) of its neighbor node x to it, and then update the value of v to the sum of d(x, v), and continue to spread in this way;

[0033] S203: Repeat the above iterative update process until the values of all entity nodes no longer change or reach the preset maximum number of iterations, and output the feature representations of all entities;

[0034] S30: Perform multi-hop recommendation

[0035] S301: Perform one-hop reasoning: Input the feature representation of the entity, the feature vector integrating all entity information, and the user's historical behavior and convert them into embedding vectors; Randomly select a node in the subgraph as the starting node, and calculate the transition probability of the relationship from the starting node to its direct neighbor nodes; Calculate the similarity between the user embedding vector and the entity embedding vector in the subgraph; Then use the transition probability to weight the similarity; Then perform diffusion through iteration to simulate the random walk of the user in the subgraph; After multiple rounds of iteration, find the entity set that is highly correlated with the user's historical behavior after one hop;

[0036] S302: Perform two-hop reasoning: Randomly select a new starting point from the obtained entity set; Then calculate the two-hop transition probability, combine the user embedding vector and the two-hop transition probability, and find the entity set that is more deeply associated with the user's historical behavior after two hops;

[0037] S303: Repeat the above process. After multiple iterations, obtain multiple entity sets with different degrees of association with the user's historical behavior, perform weighted summation on the multiple entity sets to obtain a weighted entity set, generate a recommendation sequence for the entities with the top-ranked weight values, and realize personalized recommendation.

[0038] Further, the data preprocessing includes:

[0039] Perform data cleaning to remove irrelevant information;

[0040] Perform Chinese word segmentation on the cleaned data;

[0041] Perform entity recognition on the data after Chinese word segmentation. The recognized place names are the points of interest, and obtain the point-of-interest entity data: Among them, for structured data, extract place name entities based on the Bi-LSTM-CRF algorithm; For semi-structured data, use a rule-based method to classify and label specific address information;

[0042] Based on spatial geographic information, entity disambiguation, entity extraction, and entity alignment are performed on the identified place name entities to obtain point-of-interest (POI) corpus;

[0043] Based on dependency syntax analysis, relation extraction is performed on the POI corpus, and the POI relations are classified to obtain the final corpus text data.

[0044] Furthermore, an AC automaton based on a double-array trie is used for Chinese word segmentation.

[0045] Furthermore, a Bert-BiLSTM-CRF neural network algorithm is used for entity extraction and alignment, and the KMP algorithm is used to improve the efficiency of entity alignment.

[0046] Furthermore, the POI relations are divided into: subject-predicate-object, verb-complement structure, adverbial-verb structure, adverbial-verb-complement structure, postpositive attributive relation, prepositional object relation, and object-preposed structure.

[0047] The present invention also provides a diffusion graph recommendation system based on a multi-hop mechanism, including a knowledge graph construction module, a POI-GCN feature extraction network, and a Diffussion-POI recommendation model;

[0048] The knowledge graph construction module is used to construct a knowledge graph containing basic information of various POIs and integrating semantic information and social relations;

[0049] The POI-GCN feature extraction network is used to capture deep semantic information in the knowledge graph;

[0050] The Diffussion-POI recommendation model is used to perform personalized recommendation according to the features extracted by the POI-GCN feature extraction network, and combine water wave diffusion and a multi-hop mechanism to generate a recommendation sequence.

[0051] The beneficial effects of the present invention: The present invention provides a diffusion graph recommendation method and system based on a multi-hop mechanism, mainly composed of knowledge graph construction, feature extraction network construction, and multi-hop mechanism diffusion model construction; different types of POI features are respectively input into different feature extraction hidden layers, and different weight assignments, embedding mappings, and non-linear transformations are performed on them, which can better capture the feature differences of different types of POIs, thereby improving the expression ability and generalization ability of the model; through the information transmission of the graph, the user profile can be enriched, potential associations can be discovered, and personalized recommendations can be made; the water wave diffusion algorithm can enable the model to have the ability to find more entities to improve its generalization performance; the multi-hop model can gradually expand and enrich the user's POI set, mine more potential POIs related to the user's historical behavior, effectively utilize the relationship information in the user and the knowledge graph, and achieve personalized and diversified recommendations. Description of the Drawings

[0052] Figure 1 Schematic flowchart of the diffusion graph recommendation method based on the multi-hop mechanism provided by the embodiments of the present invention;

[0053] Figure 2 Overall architecture diagram of the diffusion graph recommendation model based on the multi-hop mechanism provided by the embodiments of the present invention;

[0054] Figure 3 Schematic flowchart of the knowledge graph construction provided by the embodiments of the present invention;

[0055] Figure 4 Overall architecture diagram of the feature extraction network model provided by the embodiments of the present invention;

[0056] Figure 5 Computational structure diagram of the Weighted Sum provided by the embodiments of the present invention;

[0057] Figure 6 Schematic diagram of the multi-hop model provided by the embodiments of the present invention; Detailed implementation manners

[0058] The following combines the accompanying drawings and embodiments to further describe in detail the specific implementation manners of the present invention. The following embodiments are used to illustrate the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0059] As Figure 1 and Figure 2 shown, the embodiments of the present invention provide a diffusion graph recommendation method based on the multi-hop mechanism, including three parts: knowledge graph construction, feature submission network construction, and multi-hop mechanism diffusion model construction. The specific steps are as follows:

[0060] Step 1: Construct a knowledge graph, see Figure 3 ;

[0061] Step 1.1: Combine the crawled public commercial address data such as the Yelp2018 dataset, Gowalla dataset, and Amap POI, as well as the open platform merchant data obtained using the MeiTuan-POI interface to form a knowledge graph dataset. The knowledge graph dataset consists of structured point-of-interest data and unstructured and semi-structured data, providing high-quality data for subsequent model training.

[0062] Step 1.2: Perform data preprocessing on the data in the knowledge graph dataset

[0063] Step 1.2.1: Use methods such as regular expressions and string matching for data cleaning to remove irrelevant information such as noise data, non-text characters, and stop words;

[0064] Step 1.2.2: Use the AC automaton based on the double-array trie for Chinese word segmentation on the cleaned data to achieve efficient multi-pattern matching, which can minimize the number of comparisons as much as possible and improve the matching efficiency;

[0065] Step 1.2.3: Perform entity recognition on the data after Chinese word segmentation. The recognized place names are the points of interest, and obtain the point-of-interest entity data: Among them, for structured data, extract place name entities based on the Bi-LSTM-CRF algorithm; for semi-structured data, use a rule-based method to classify and label specific address information;

[0066] Step 1.2.4: Based on spatial geographic information, perform entity disambiguation, entity extraction, and entity alignment on the recognized place name entities to obtain the point-of-interest corpus; among them, use the Bert-BiLSTM-CRF neural network algorithm for entity extraction and alignment, and use the KMP algorithm to improve the efficiency of entity alignment;

[0067] Step 1.2.5: Perform relationship extraction on the point-of-interest corpus based on dependency syntax analysis. Classify the point-of-interest relationships into 7 types: subject-verb-object, verb-complement structure, adverbial-verb structure, adverbial-verb-complement structure, postpositive attributive relationship, prepositional object relationship, object-preposing structure, for the convenience of relationship extraction, and obtain the final corpus text data.

[0068] Step 1.3: Construct the triples <entity, relationship, attribute> that make up the knowledge graph according to the preprocessed text results, store the knowledge graph in the repository neo4j and visualize it. The knowledge graph not only contains the basic information of various points of interest, but also incorporates rich semantic information and social relationships, providing users with more comprehensive and in-depth knowledge services.

[0069] Step 2: Build a feature extraction network based on POI-GCN with the graph convolutional neural network (GCN) as the basis, capture the deep semantic information in the knowledge graph, and fuse it into the recommendation system to enhance the recommendation effect. Take the triples in the knowledge graph as the semantic network, where the nodes represent entities and the edges represent various semantic relationships between entities and entities or entities and attributes.

[0070] Figure 4 is the overall architecture diagram of the feature extraction network, including a word embedding layer, a network layer, a GCN layer, and a hidden layer;

[0071] For each point of interest, extract features including the name, location, category, score, etc. of the point of interest from the triples it belongs to, and then map these features into fixed-length vectors and input them into the word embedding layer to obtain the embedding vector of the entity, which is used as the input of the network layer;

[0072] An attention mechanism is added to the network layer for weight allocation, and the calculation formula is as follows:

[0073]

[0074] where a ij represents the attention weight; softplus is the activation function; e ij represents the similarity between user i and item j, and the formula is as follows:

[0075] e ij = ELU(a T |Wh r ||Wh i |) (2)

[0076]

[0077] where ELU is the activation function; W is the linear transformation matrix, with the shape W ∈ R F×F′ , F is the dimension of the input feature, and F' is the dimension of the output feature; Wh r represents the element in the h-th row and r-th column of the W matrix; Wh i represents the element in the h-th row and i-th column of the W matrix; a represents the parameter vector; f(x) represents the ELU activation function, which adjusts the similarity value e ij through non-linear transformation to better calculate the attention weight; exp represents the natural exponential function; γ is the scaling factor;

[0078] After weight allocation, the weighted entity embedding vector is output, and then message aggregation is performed to obtain the aggregated feature;

[0079] The GCN layer performs feature fusion on the obtained aggregated feature and outputs the entity feature vector of the fused feature;

[0080] The hidden layer uses the ReLU activation function to perform non-linear transformation on the entity feature vector of the fused feature, solve the gradient disappearance and neuron death problems, and capture the feature vector that integrates all entity information.

[0081] Step 3: Based on the features extracted by the POI-GCN feature extraction network, introduce the water wave diffusion and multi-hop mechanism in an end-to-end manner, and construct a Diffussion-POI recommendation model for recommendation.

[0082] Step 3.1: Extract the knowledge graph features according to the feature vectors extracted by the POI-GCN feature extraction network

[0083] Step 3.1.1: Perform entity linking on the entities in the knowledge graph to obtain the matched entities;

[0084] Step 3.1.2: Extract a sub-graph from the knowledge graph after entity linking;

[0085] Step 3.1.3: Introduce context text features for the entities in the sub-graph to more accurately characterize the features of the entities. The formula is as follows:

[0086] context(n) = {n m | (n, r, n m ) ∈ G or (n m , r, n) ∈ G}

[0087] where context(n) is a function for describing the context features of entity n; G represents the graph; r is the relationship between nodes; m represents the m-th entity; (n, r, n m ) represents a triple, entity n, relationship r, and another entity n m .

[0088] Step 3.2: Perform information transfer of the graph

[0089] Step 3.2.1: Set the initial value of all entity nodes after introducing context text features to 0, and set the initial value of the seed node to 1;

[0090] Step 3.2.2: Iteratively update the value of each entity node. In each iteration, for any entity node v, calculate the contribution value d(x, v) of its neighbor node x to it, and then update the value of v to the sum of d(x, v), and continuously spread in this way;

[0091] Step 3.2.3: Repeat the above iterative update process until the values of all entity nodes no longer change or reach the preset maximum number of iterations; output the feature representations of all entities.

[0092] Step 3.3: Perform multi-hop recommendation, as Figure 6 shown, the specific process is as follows:

[0093] Step 3.3.1: Perform one-hop reasoning. The specific process is as follows: Input the feature representation of the entity, the feature vector integrating all entity information, and the user's historical behavior and convert them into embedding vectors; randomly select a node in the sub-graph as the starting node, calculate the transition probability of the relationship from the starting node to its direct neighbor nodes; calculate the similarity between the user embedding vector and the entity embedding vectors in the sub-graph; then weight the similarity using the transition probability; then perform diffusion through iteration to simulate the random walk of the user in the sub-graph; after multiple rounds of iteration, find the entity set that is highly relevant to the user's historical behavior after one hop;

[0094] Step 3.3.2: Conduct two-hop reasoning. The specific process is as follows: After completing one-hop, randomly select a new starting point from the obtained set of interest points; then calculate the two-hop transition probability, and combine the user embedding vector and the two-hop transition probability to find the set of entities that are more deeply associated with the user's historical behavior after two hops;

[0095] Step 3.3.3: Repeat the above process. After multiple iterations, obtain multiple sets of entities with different degrees of association with the user's historical behavior, and perform weighted summation on the multiple sets of entities using Weight Sum. The calculation process is shown in Figure 5 , obtain the weighted set of entities, generate a recommendation sequence for the entities with the top-ranked weight values, and achieve personalized recommendation.

[0096] Based on the above method, test some hyperparameters respectively to verify the robustness of the model.

[0097] 1) Set the linear transformation layer to linear and sigmid, the masking ratio to 0.2, the learning rate to 0.01, the hidden layer dimension to 64, the number of propagation layers to 2, and the other model settings remain constant. Adjust the comparison dimension setting of the adjacency matrix propagation layer. The experimental results are shown in Table 1:

[0098] Table 1 Adjacency matrix layer dimension setting

[0099]

[0100] The experimental results show that the dimension setting of the model is a key parameter, and the feature representation effect is the best with 512 dimensions.

[0101] 2) Set the linear transformation layer to linear and sigmid, the masking ratio to 0.2, the learning rate to 0.01, the hidden layer dimension to 64, and the other model settings remain constant. Adjust the comparison dimension setting of the adjacency matrix propagation layer. The experimental results are shown in Table 2:

[0102] Table 2 Propagation layer depth setting

[0103]

[0104] In this example, according to the characteristics of the specific task and dataset, adjust the number of propagation layers to 6 layers to achieve the best performance.

[0105] Name the recommendation model of the present invention as the MultiHop-GDN model and conduct ablation experiments. The experimental results are shown in Table 3.

[0106] Table 3 Performance comparison of ablation experiments

[0107]

[0108] Performance comparison experiments were conducted, considering precision, recall, and F1 value on three datasets, Gowalla, Yelp2018, and MeiTuan-POI, as evaluation metrics.

[0109] To ensure the quality of the Gowalla dataset, 10 core settings were used in this example, that is, users and items were retained through at least ten interactions. The Yelp2018 dataset was taken from the 2018 version of the Yelp Challenge. Among them, local businesses such as restaurants and bars were regarded as items, and the same 10 core settings were used to ensure data quality. For the MeiTuan-POI dataset, open platform merchant categories of life services, transportation hubs, and medical and health were selected from the collections. Similarly, 10 core settings were used to ensure that each user and item had at least 10 interactions. This example was compared with four models, Pinsage, NGCF, KM-CF, and BC-CF, and the experimental results are shown in Table 4.

[0110] Table 4 Overall comparison of performance (RQ1)

[0111]

[0112] The experimental results on the public datasets verified that the MultiHop-GDN model of the present invention had better recommendation effects compared with other baseline methods, showing the powerful ability of the model in capturing user and item features. The present invention provided a useful reference for further integrating knowledge graphs and deepening recommendation mechanisms in the field of recommendation systems in the future. Future research could explore the application of multi-hop mechanisms in more extensive types of recommendation systems based on this, as well as more efficient methods for integrating knowledge graphs and recommendation systems to further improve the overall performance and user satisfaction of recommendation systems.

[0113] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any form of limitation to the present invention. Those skilled in the art should fully understand that it is completely feasible to modify the technical solutions described in the foregoing embodiments or perform equivalent replacements on any part or all of the technical features therein. As long as these modifications or replacements do not deviate from the protection scope determined by the claims of the present invention, they should be regarded as reasonable extensions of the present invention.

Claims

1. A diffusion graph recommendation method based on a multi-hop mechanism, characterized in that: include: Construct a knowledge graph that contains basic information about various points of interest and incorporates semantic information and social relationships; A POI-GCN feature extraction network is constructed based on a graph convolutional neural network to capture deep semantic information in the knowledge graph; Based on the features extracted by the POI-GCN feature extraction network, the water wave diffusion and multi-hop mechanism are introduced to construct the Diffussion-POI recommendation model and generate the recommendation sequence. The specific process is as follows: S10: Extracting knowledge graph features based on the feature vector extracted by the POI-GCN feature extraction network S101: Perform entity linking on entities in the knowledge graph to obtain matched entities; S102: extracting subgraphs from the knowledge graph after entity linking; S103: Introducing contextual text features into the entities in the subgraph: context(n)={n m |(n,r,n m )∈G or(n m ,r,n)∈G} Where context(n) is a function used to describe the contextual features of entity n; G represents the graph; r represents the relationship between nodes; n m represents the mth entity; (n,r,n m ) represents a triple, entity n, relation r and another entity n m ; S20: Transmitting graph information S201: setting the initial value of all entity nodes after the contextual text feature is introduced to 0, and setting the initial value of the seed node to 1; S202: Iteratively update the value of each entity node. In each iteration, for any entity node v, calculate the contribution value d(x, v) of its neighbor node x to it, and then update the value of v to the sum of d(x, v), so as to continuously spread; S203: Repeat the above iterative update process until the values ​​of all entity nodes no longer change or the preset maximum number of iterations is reached, and output feature representations of all entities; S30: Perform multi-hop recommendation S301: Perform one-hop reasoning: input the feature representation of the entity, the feature vector integrating all entity information and the user's historical behavior and convert them into an embedding vector; randomly select a node in the subgraph as the starting node, and calculate the transition probability of the relationship from the starting node to its direct neighbor node; calculate the similarity between the user embedding vector and the entity embedding vector in the subgraph; then use the transition probability to weight the similarity; then diffuse through iteration to simulate the random walk of the user in the subgraph; after multiple rounds of iterations, find the entity set that is highly correlated with the user's historical behavior after one hop; S302: Perform two-hop reasoning: randomly select a new starting point from the obtained entity set; then calculate the two-hop transfer probability, combine the user embedding vector and the two-hop transfer probability, and find the entity set that is more deeply associated with the user's historical behavior after two hops; S303: Repeat the above process for multiple iterations to obtain multiple entity sets with different degrees of association with the user's historical behavior, perform weighted summation on the multiple entity sets to obtain a weighted entity set, generate a recommendation sequence for the entities ranked first in weight value, and implement personalized recommendation.

2. The diffusion graph recommendation method based on a multi-hop mechanism according to claim 1, characterized in that: The method for constructing the knowledge graph is as follows: The crawled Yelp2018 dataset, Gowalla dataset, public business address data, and open platform merchant data obtained using the MeiTuan-POI interface are used to form a knowledge graph dataset, which consists of structured point of interest data and unstructured and semi-structured data; Performing data preprocessing on the data in the knowledge graph dataset; The triples <entity, relationship, attribute> that make up the knowledge graph are constructed based on the preprocessed text results.

3. The method for recommending diffusion graphs based on a multi-hop mechanism according to claim 2, characterized in that: The POI-GCN feature extraction network includes a word embedding layer, a network layer, a GCN layer and a hidden layer; For each point of interest, extract features including the name, location, category, and score of the point of interest from its triplet, and then map these features into vectors of fixed length and input them into the word embedding layer to obtain an embedding vector of the entity as the input of the network layer; The attention mechanism is added to the network layer to distribute weights. The calculation formula is as follows: Among them, a ij represents the attention weight; softplus is the activation function; e ij Represents the similarity between user i and item j. The formula is as follows: have been ij =TWO(a T |Wh r ||Wh i |) (2) Among them, ELU is the activation function; W is the linear transformation matrix, the shape is W∈R F×F′ , F is the dimension of the input feature, F' is the dimension of the output feature; Wh r Represents the element in the hth row and rth column of the W matrix; Wh i represents the element in the hth row and ith column of the W matrix; a represents the parameter vector; f(x) represents the ELU activation function, which adjusts the similarity value e through nonlinear transformation ij ; exp represents the natural exponential function; γ is the scaling factor; After weight assignment, the weighted entity embedding vector is output, and then message aggregation is performed to obtain aggregate features; The GCN layer performs feature fusion on the obtained aggregated features and outputs an entity feature vector of the fused features; The hidden layer uses the ReLU activation function to perform nonlinear transformation processing on the entity feature vector of the fused feature to obtain a feature vector that integrates all entity information.

4. The method for recommending diffusion graphs based on a multi-hop mechanism according to claim 2, characterized in that: The data preprocessing includes: Perform data cleaning to remove irrelevant information; Perform Chinese word segmentation on the cleaned data; Perform entity recognition on the data after Chinese word segmentation, and the recognized place names are the points of interest, and obtain the entity data of the points of interest: for structured data, the place name entities are extracted based on the Bi-LSTM-CRF algorithm; for semi-structured data, a rule-based method is used to classify and label the specific address information; Based on spatial geographic information, the identified place name entities are disambiguated, extracted and aligned to obtain the point of interest corpus; Based on dependency syntactic analysis, the relationship of the interest point corpus is extracted, and the relationship of the interest points is divided to obtain the final corpus text data.

5. The method for recommending diffusion graphs based on a multi-hop mechanism according to claim 4, characterized in that: The AC automaton based on double-array dictionary tree is used for Chinese word segmentation.

6. The method for recommending diffusion graphs based on a multi-hop mechanism according to claim 4, characterized in that: The Bert-BiLSTM-CRF neural network algorithm is used for entity extraction and alignment, and the KMP algorithm is used to improve the efficiency of entity alignment.

7. The method for recommending diffusion graphs based on a multi-hop mechanism according to claim 4, characterized in that: The interest point relationship is divided into: subject-predicate-object, verb-complement structure, adverbial-verb structure, adverbial-verb-complement structure, attributive postposition relationship, preposition-object relationship, and object preposition structure.

8. A diffusion graph recommendation system based on a multi-hop mechanism, used to execute the diffusion graph recommendation method according to any one of claims 1 to 7, characterized in that: Including knowledge graph construction module, POI-GCN feature extraction network and Diffussion-POI recommendation model; The knowledge graph construction module is used to construct a knowledge graph containing basic information of various points of interest and integrating semantic information and social relationships; The POI-GCN feature extraction network is used to capture deep semantic information in the knowledge graph; The Diffussion-POI recommendation model is used to generate a recommendation sequence by performing personalized recommendation based on the features extracted by the POI-GCN feature extraction network and combining water wave diffusion with a multi-hop mechanism.

Images