Geographic information knowledge graph recommendation system and recommendation method thereof

Abstract

The application discloses a geographic information knowledge graph recommendation system and a recommendation method thereof, and relates to the technical field of recommendation systems.The geographic information knowledge graph recommendation system comprises the following steps: obtaining two kinds of knowledge graphs in the input layer, namely, a current user interest interaction project knowledge graph and a geographic knowledge graph of an environment where the current user interaction project is located.In order to learn different information from the two knowledge graphs, the information in the two knowledge graphs is mapped to two different spaces by using one-hot encoding.A graph convolution module learns high-dimensional features of the two data, and the information of the two databases is fused in an anti-fact learning module, so that the data recommendation work is realized.The application has the advantages that: the effective knowledge of the geographic knowledge graph and the user-project knowledge graph can be fully fused and utilized, and the user experience and the service quality are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electronic information technology, and in particular to a geographic information knowledge graph recommendation system based on a graph neural network and a recommendation method thereof. BACKGROUND

[0002] Nowadays, human beings are in the era of massive data. A large amount of information accumulation provides an important information source for human activities, but information overload is also a serious problem. People have difficulty in selecting the correct content from a large number of options. A recommendation system (RS) can recommend results that meet the user's preferences based on the user's historical behavior, and even can achieve the function of associated information recommendation through improved algorithms. The function of the recommendation system can solve the problem of information explosion and is widely used in music recommendation [1], movie recommendation [2] and online shopping [3] and other applications.

[0003] Traditional recommendation systems usually rely on specific indicators to model core functions. The recommended results are not only small in quantity, but also low in quality, and often cannot meet the user's needs. When the recommendation task changes, most cases require experts to manually modify the core function, which is not flexible and time-consuming and laborious. In some regionally restricted tasks, the recommended results are extremely prone to large deviations.

[0004] Traditional knowledge graph-based recommendation systems are mainly divided into collaborative filtering (CF) based recommendation systems and content-based recommendation systems. (1) The collaborative filtering method assumes that users may be interested in the items selected by people who share similar interaction records. The interaction mode can be explicit interaction such as user evaluation [4][5], or implicit interaction behavior such as clicking and viewing [6]. The collaborative filtering algorithm is mainly divided into content-based methods and model-based methods, where the memory-based method first learns the similarity of users from the user-item interaction data center, and then recommends similar data to the current user according to the selection of similar users. The model-based method [7] mainly solves the sparsity problem by establishing a reasoning model to extract the latent representation from the high-dimensional user and item interaction matrix, and then calculates the similarity between users and items using inner product or other methods. (2) The content-based filtering algorithm learns the representation of users and items from global user-item interaction data, assuming that users may be interested in similar items to the items they have interacted with in the past. Item representation is obtained by extracting attributes from the auxiliary information of the item (including text, image, etc.), while user representation is based on the characteristics of the personal interaction items. The process of comparing candidate items with user profiles is essentially matching them with the user's previous records, that is, recommending items similar to the items the user likes in the past to the user [8].

[0005] The limitations of such methods are strong, and the function of the function is weak, and the cold start (new users have not yet made screening so that they cannot recommend), data sparsity (the number of data in the user-item matrix is small), malicious attacks (user ratings are copied by other users, causing the similarity between users to soar) and gray sheep problems (users differ greatly between each preference, and the system cannot correctly generate recommendations) when encountering these difficult problems. Traditional recommendation methods cannot work.

[0006] Integrating additional information into the recommendation method can overcome the limitations of using only one method, which uses a variety of auxiliary information such as item attributes [9], brand, text description

[10] , image features

[11] , video features

[12] , etc. By merging the content information of users and items (also known as user-side information and item-side information) into the CF-based framework, better recommendation performance can be achieved

[13] . Among them, the structured auxiliary information of the knowledge graph (KG) often makes the recommendation results more in line with user requirements. The methods using the knowledge graph as auxiliary information are mainly divided into embedding-based methods, path-based methods and unified methods. (1) Embedding-based methods directly use the information in the KG to enrich the representation of items or users, and need to apply knowledge graph embedding (KGE) algorithm to encode the KG into low-rank embedding. Most embedding-based methods

[14] -

[15] use multiple types of item-side information to construct the KG to enrich the representation of the item, which can be used to model the user representation more accurately. Some papers use GAN

[16] or BEM

[17] to improve the embedding to obtain better recommendations. (2) Path-based algorithms

[18] generate recommendations according to user behavior knowledge graphs, which use path connectivity to standardize or enrich user and item representations. With the development of deep learning technology, people have proposed different models

[19] to explicitly encode path embeddings, and proposed methods that can be generated through path embeddings, or can achieve better recommendations by finding the most significant path connecting user-item pairs. (3) Embedding-based methods use semantic representations of users / items in KG for recommendations, while path-based methods use semantic connection information. Both methods only use one aspect of the information in the graph. In order to make full use of the information in the KG to obtain better recommendations, people have proposed unified methods that combine semantic representations of entities and relationships and connectivity information. Unified methods are based on the idea of embedding propagation. These methods refine entity representations under the guidance of KG connection structures. These algorithms are usually implemented based on GNN architecture, and since RippleNet

[20] was proposed in 2018, such methods have become a new research trend.

[0007] Although the recommendation algorithm with auxiliary information can solve the cold start, data sparsity, malicious attack and gray sheep problem, but some recommendations involving spatial information often cannot get good recommendation results, such as when choosing a restaurant, most people tend to choose a restaurant close to their location, and the recommendation result without considering geographic location information is not high in quality. In addition, the recommendation result accuracy, novelty, diversity and interpretability of the previous recommendation system model are not high. Users cannot get satisfactory diversified results from the recommendation system, and most algorithms cannot give users an explanation, and the probability of users choosing the recommendation algorithm result is also low.

[0008] REFERENCES

[0009] [1] Mittal S, Ranjan A, Roy B, et al. Mus-Emo: An Automated Facial Emotion-Based Music Recommendation System Using Convolutional Neural Network [M] / / Advances in Communication, Devices and Networking. Springer, Singapore, 2022: 267-276;

[0010] [2] Jena K K, Bhoi S K, Mallick C, et al. Neural model based collaborative filtering for movie recommendation system [J]. International Journal of Information Technology, 2022: 1-11;

[0011] [3] Xu X. Research on Recommendation System of Internet Shopping Platform Based on Cognitive Internet of Things [C] / / International Conference on Cognitive based Information Processing and Applications (CIPA 2021). Springer, Singapore, 2022: 796-801;

[0012] [4] Amatriain X, Pujol J M, Oliver N. I Like It. I Like It Not: Evaluating User Ratings Noise in Recommender Systems [C] / / International Conference on User Modeling. Springer-Verlag, 2009;

[0013] [5] Jawaheer G, Szomszor M, Kostkova P. Comparison of implicit and explicit feedback from an online music recommendation service [C] / / Proceedings of the 1st International Workshop on Information Heterogeneity and Fusion in Recommender Systems. 2010:47-51;

[0014] [6] Wang C, Zhu H, Zhu C, et al. Setrank: A setwise Bayesian approach for collaborative ranking from implicit feedback [C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2020, 34(04):6127-6136;

[0015] [7] Zhuang F, Zhang Z, Qian M, et al. Representation learning via dual-autoencoder for recommendation [J]. Neural Networks, 2017, 90:83-89;

[0016] [8] Lops P, Gemmis M, Semeraro G. Content-based recommender systems: State of the art and trends [J]. Recommender systems handbook, 2011:73-105;

[0017] [9] Zhen Y, Li W J, Yeung D Y. TagiCoFi: tag informed collaborative filtering [C] / / Proceedings of the third ACM conference on Recommendersystems. 2009: 69-76;

[0018]

[10] Han J, Zheng L, Xu Y, et al. Adaptive deep modeling of users and items using side information for recommendation [J]. IEEE transactions on neural networks and learning systems, 2019, 31(3): 737-748;

[0019]

[11] Chu W T, Tsai Y L. A hybrid recommendation system considering visual information for predicting favorite restaurants [J]. World Wide Web, 2017, 20(6): 1313-1331;

[0020]

[12] Chen J, Zhang H, He X, et al. Attentive collaborative filtering: Multimedia recommendation with item-and component-level attention [C] / / Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval. 2017: 335-344;

[0021]

[13] Sun Z, Guo Q, Yang J, et al. Research commentary on recommendations with side information: A survey and research directions [J]. Electronic Commerce Research and Applications, 2019, 37: 100879;

[0022]

[14] Wang H, Zhang F, Zhao M, et al. Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation [J]. 2019;

[0023]

[15] Cao Y, Wang X, He X, et al. Unifying Knowledge Graph Learning and Recommendation: Towards a Better Understanding of User Preferences [J]. 2019;

[0024]

[16] Yang D, Guo Z, Wang Z, et al. A Knowledge-Enhanced Deep Recommendation Framework Incorporating GAN-Based Models [C] / / 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 2018;

[0025]

[17] Ye Y, Wang X, Yao J, et al. Bayes EMbedding (BEM): Refining Representation by Integrating Knowledge Graphs and Behavior-specific Networks [J]. 2019;

[0026]

[18] Shi C, Hu B, Zhao X, et al. Heterogeneous Information Network Embedding for Recommendation[J]. IEEE Transactions on Knowledge & Data Engineering, 2017: 1-1;

[0027]

[19] Wang X, Wang D, Xu C, et al. Explainable Reasoning over Knowledge Graphs for Recommendation[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2019, 33: 5329-5336;

[0028]

[20] Wang H, Zhang F, Wang J, et al. RippleNet: Propagating User Preferences on the Knowledge Graph for Recommender Systems[J]. ACM, 2018;

[0029]

[21] Wang S, Gong M, Li H, et al. Multi-objective optimization for longtail recommendation[J]. Knowledge-Based Systems, 2016, 104: 145-155;

[0030]

[22] Cui Z, Zhao P, Hu Z, et al. An improved matrix factorization based model for many-objective optimization recommendation[J]. Information Sciences, 2021, 579: 1-14;

[0031]

[23] Geng B, Li L, Jiao L, et al. NNIA-RS: A multi-objective optimization based recommender system[J]. Physica A: Statistical Mechanics and its Applications, 2015, 424: 383-397. SUMMARY

[0032] The present application provides a geographic information knowledge graph recommendation system and a recommendation method thereof to overcome the defects of the prior art. The geographic information knowledge graph is integrated into the recommendation system, effectively improving the recommendation quality of geographic location transactions. The powerful fitting ability of the graph neural network (GCN) is utilized to maximize four objective functions, while improving the accuracy, novelty, diversity and explainability of the recommendation results.

[0033] To achieve the above application purposes, the technical solutions adopted by the present application are as follows:

[0034] A geographic information knowledge graph recommendation system comprises an input layer, a graph convolution layer and a counterfactual learning layer.

[0035] The input layer is used to map the geographic knowledge graph and the user-item interaction knowledge graph into embedding vectors, unifying the knowledge storage forms of the two databases. This facilitates the subsequent network to fuse the information in the two databases for learning, and helps to obtain more complete, more structured and more fine-grained information.

[0036] The graph convolution layer is responsible for learning the two types of vectors obtained by the input layer, obtaining high-dimensional vector information of the two types of knowledge.

[0037] The counterfactual learning layer fuses the two completely different high-dimensional information obtained by the graph convolution layer, uses counterfactual reasoning operations, uses geographic information to constrain the recommendation process, and obtains more accurate recommendation results.

[0038] The present application also discloses a geographic information knowledge graph recommendation method, comprising the following steps:

[0039] Step 1, in the input layer, obtain the current user-item interaction knowledge graph and the geographic information knowledge graph of the environment where the current user interacts with the project.

[0040] The embedding matrix of the geographic information knowledge graph and the user-item interaction knowledge graph is shown in the following formula:

[0041] E g =[u 1,g ;…;u N,g ;p1,g ;... ; p M,g ; v 1,g ;... ; v W,g ]

[0042] E f = [u 1,f ;... ; u N,f ; p 1,f ;... ; p M,f ; v 1,f ;... ; v Q,f ]

[0043] User-geographical entity vector embedding and user-item interaction vector embedding are represented as where u represents embedding vector embedding, i is used to identify user identity, g represents geographical entity number in knowledge graph, u i,g represents vector embedding of user-geographical information interaction; f is function number, u i,f represents vector embedding of user-item interaction, and d is the dimension of embedding.

[0044] In which p represents a location-based service item, p i,g represents the embedding vector embedding of the i-th geographical vector in the geographical knowledge graph, p i,f represents the embedding vector embedding of the i-th function vector in the geographical knowledge graph, and p serves to connect v and u.

[0045] The embedding vectors of function entities and geographical entities are respectively represented in user-item knowledge graph and geographical information knowledge graph as where v represents embedding vector embedding in knowledge graph; h is user identity, f is function number, and k is geographical entity number.

[0046] N and M represent the number of users in geographical information knowledge graph and user-item interaction knowledge graph respectively, and W and Q represent the number of entities in geographical information knowledge graph and user-item interaction knowledge graph respectively.

[0047] Step 2, in the graph convolution layer, the information in the two knowledge graphs needs to be aggregated to pass to the user, so the propagation rule in the graph convolution layer is defined as the following formula:

[0048]

[0049] where and respectively represent the k-th geographical entity and the i-th user knowledge graph embedding matrix, and is Hadamard inner product. represents a geographical entity sample related to a user, the implementation of the function of the recommendation system based on the geographical knowledge graph needs multiple intention driving, including geographical intention I g and functional intention I f , e j,g is the interest score of user u i , and β(i,j) is the attention score of interest score e j,g .

[0050] The calculation method of β(i,j) is as follows:

[0051]

[0052] Wherein is the one-hot encoded vector of the i-th user, is the one-hot encoded embedding vector of the j-th intention about geographical entity g in the geographical information knowledge graph, is all individuals belonging to geographical intention I g , and exp is the exponential function with base e.

[0053] The entire graph convolution network updates the weight of each node according to the following formula, and the specific propagation rule needs to be used to adjust the weight coefficients between multi-layer neural networks, in order to obtain the information in high-order graph neural network, the information propagation between low layer and high layer follows the following rules:

[0054]

[0055]

[0056]

[0057]

[0058] The information learned by the above two different partitions of graph convolution layer needs to be fused to form the embedding vector of the user and the embedding vector of the geographical entity, so as to calculate the matching degree of the two.(l) and (l-1) represent the l-th layer and the l-1-th layer in the graph convolution network, I g represents the geographical intention, I f represents the functional intention, r j,g is the embedding vector of the i-th relationship in the geographical knowledge graph, r t,f is the embedding vector of the t-th relationship in the functional knowledge graph. e j,g is the embedding vector of the j-th intention in the geographical knowledge graph, e t,fis the embedding vector of the t-th intent in the functional knowledge graph. is the relation of interest to the geographic entity in the geographic knowledge graph, is the user u i is interested in. β(i,j) is the attention score of the i-th user and the j-th geographic entity. i is the interest score of the geographic entity e j,g is the attention score of the i-th user and the j-th functional. β(i,t) is the interest score of the functional e i is the interest score of the geographic entity e t,f . The information fusion calculation method of the i-th user and the i-th geographic entity is as follows:

[0059]

[0060] Step 3, the counterfactual learning layer uses the counterfactual learning method to add the information of the geographic information knowledge graph to the final recommendation result, and assists the generation of the recommendation result.

[0061] Further, in order to obtain a better recommendation result, the optimization process of the model is constrained by accuracy, diversity, novelty and explainability. The calculation method of the four indexes is shown in the following formula:

[0062]

[0063]

[0064]

[0065]

[0066] Among them, max Accuracy index is used to constrain the accuracy of the recommendation result, max Novelty index is used to constrain the novelty of the recommendation result, max Diversity index is used to constrain the diversity of the recommendation result, and max Explainability index is used to constrain the explainability of the recommendation result. R u represents the user-item recommendation list, T u represents the user's favorite items, |R u | represents the length of the recommendation list, |R u ∩T u | represents the length of the items in the recommendation list that the user likes. num is the number of users, r i is the number of ratings of the user to the item i. s(i,j) represents the similarity of different items in the recommendation result. r i,j-m is the rating of the user to the item j-m . S j-m,j′-m is the similarity between two items, sum riis the sum of the ratings of the user to all items.

[0067] Compared with the prior art, the application has the advantages that:

[0068] 1. The information of two knowledge graphs is fused to generate a recommendation result, high-order information of the two knowledge graphs is extracted using a graph convolution layer, and the high-order information from two different sources is fused in a counterfactual learning layer to obtain a recommendation result limited by a geographic factor. The user can easily find the most suitable item for himself, and the recommendation result can also provide detailed geographic information for the user, effectively improving the user experience and the quality of the recommendation.

[0069] 2. Multiple optimization objectives are designed when optimizing the graph neural network, and the maximum values of multiple optimization functions are obtained, which comprehensively improves the model performance and provides better recommendation services for the user. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a workflow diagram of the model EGKG-RS of the embodiment of the application;

[0071] Figure 2 is a structural schematic diagram of the model EGKG-RS of the embodiment of the application;

[0072] Figure 3 is a counterfactual learning method flowchart of the model EGKG-RS of the embodiment of the application. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical scheme and advantages of the application clearer, the following will further describe the application according to the drawings and examples.

[0074] In this embodiment, the model EGKG-RS is implemented in the PyTorch framework, and the model is deployed on the GPU of Tesla V100, and all operations are completed using the python programming language based on the Linux operating system.

[0075] In this embodiment, a recommendation system EGKG-RS (Explicable Geographic Knowledge Graph Recommendation System) using geographic information knowledge graph as auxiliary information is designed. Figure 1As shown, two kinds of knowledge graphs need to be obtained in the input layer: the current user interest interaction project interaction knowledge graph and the geographical information knowledge graph of the environment where the current user interacts with the project. In order to learn different information from the two knowledge graphs, One-Hot Encoding needs to be used to map the information in the two knowledge graphs to two different spaces. The graph convolution module learns the high-dimensional features of the two data, and the information of the two databases is fused in the counterfactual learning module, realizing the data recommendation work.

[0076] As shown in FIG. 1, the recommendation system EGKG-RS is composed of three functionally different layers, including an input layer, a graph convolution layer, and a counterfactual learning layer. Figure 2 As shown, the recommendation system EGKG-RS is composed of three functionally different layers, including an input layer, a graph convolution layer, and a counterfactual learning layer.

[0077] First, in the input layer, the geographical information knowledge graph vector embedding and the user-project interaction embedding are represented as where u i,g represents the geographical information knowledge graph vector embedding, u i,f represents the user-project interaction embedding, and d is the dimension of the embedding. The geographical entity v k is represented in the geographical information knowledge graph as The geographical entity v k is represented in the geographical information knowledge graph as The embedding matrix of the geographical information knowledge graph and the project interaction knowledge graph can be represented as shown in the following formula:

[0078] E g = [u 1,g ;... ; u N,g ; p 1,g ;... ; p M,g ; v 1,g ;... ; v W,g ]

[0079] E f = [u 1,f ;... ; u N,f ; p 1,f ;... ; p M,f ; v 1,f ;... ; v Q,f ]

[0080] where N and M represent the number of users in the geographical information knowledge graph and the user-project interaction knowledge graph, respectively, and W and Q represent the number of entities in the geographical information knowledge graph and the user-project interaction knowledge graph, respectively.

[0081] Second, in the graph convolution layer, the information in the two knowledge graphs needs to be aggregated and passed to the user, so the propagation rule in the graph convolution layer is defined as follows:

[0082]

[0083] wherein, and respectively represent the knowledge graph embedding matrix of the kth geographical entity and the ith user, and is the Hadamard inner product. represents the geographical entity sample related to the user, e j,g is the interest score of the user u i , and β(i,j) is the attention score of e j,g . The calculation method of β(i,j) is as shown in the following formula:

[0084]

[0085] wherein is the one-hot encoded vector of the ith user, and e j,g is the one-hot encoded embedding vector of the jth intent in the geographical information knowledge graph.

[0086] The entire graph convolution network updates the weight of each node according to the following formula, and the specific propagation rule needs to be used to adjust the weight coefficients between the multi-layer neural networks. In order to obtain the information in the high-order graph neural network, the information propagation between the low layer and the high layer follows the following rules:

[0087]

[0088]

[0089]

[0090]

[0091] The information learned by the above two different partitions of the graph convolution layer needs to be fused to form the embedding vector of the user and the embedding vector of the geographical entity, so as to calculate the matching degree of the two. The information fusion calculation method of the ith user and the ith geographical entity is as shown below:

[0092]

[0093] Thirdly, since the information of the geographical information knowledge graph is auxiliary information, the counterfactual learning method needs to be used to add the information of the geographical information knowledge graph into the final recommendation result, and assist in the generation of the recommendation result. The specific process and principle of the counterfactual learning method are as shown in Figure 3 Figure 3 ​(a) describes a standard form of a causal diagram, uppercase letters are random variables, lowercase letters are the corresponding values of these variables, the head node of the directed edge has an impact on the tail node, P represents the geographical location, F represents the function of the entity, G represents the specific geographical entity, Y is the probability of the user choosing to interact with this geographical location, and is also the final result needed in the entire counterfactual learning process. Figure 3 (b) is the relationship between entities and users in the real world, in the counterfactual world Figure 3 (c) the geographical location p fails to affect the user's interaction probability V, through the addition of geographical information such as g*, p* and f*, the geographical entity P has an impact on the user's choice, reaching the reference state Figure 3 (d), the state (d) obtained through deductive learning is closer to the real world (b), greatly improving the accuracy of the recommendation results.

[0094] Fourth, in order to obtain better recommendation results, the model constrains the optimization process of the model in four aspects of accuracy, diversity, novelty and explainability. Among them, the accuracy is the most intuitive indicator to measure whether the user likes the items in the recommendation results, and higher precision means that the recommended performance of the proposed model is much better. Novelty represents the popularity of the recommended result item, and the calculation method is the inverse of the number of times a user evaluates an item, and a higher novelty indicator means that the popularity of the recommended result list is lower. In order to provide users with diversified goods, the richness of the recommended result items needs to be calculated, and the higher the richness index, the more diversified the recommended results. Explainability is to explain each recommended result through path explanation, helping the target user to understand the reason for the recommended item, and the larger the explainability index value means the easier the recommended result is to be understood. The calculation methods of the four indicators are shown in the following formulas:

[0095]

[0096]

[0097]

[0098]

[0099] Where R u represents the user-item recommendation list, T u represents the items that the user likes, |R u | represents the length of the recommendation list, |R u ∩T u represents the length of the items in the recommendation list that the user likes. num is the number of users, r iis the number of ratings of item i by users. s(i, j) represents the similarity of different items in the recommendation result. i,j-m is the rating of item j-m by users, S j-m,j′-m is the similarity between two items, sum ri is the sum of ratings of all items by users. During the training process, a multi-objective optimization scheme is adopted to maximize the four indicators as the final goal, and the performance of the EGKG-RS model is comprehensively improved.

[0100] Finally, the embodiment is trained on the geographic information knowledge graph dataset and the collected user restaurant selection dataset, and the results of the MORS

[21] , MaORA

[22] , NNIA-RS

[23] and the EGKG-RS model proposed in the embodiment are obtained. The performance of the test dataset in accuracy (Acc), novelty (Nov), diversity (Div) and explainability (Exp) is calculated, and the optimal value (Best) and average value (Avg) of different models in each evaluation indicator are recorded. The specific results are shown in Table 1.

[0101] Table 1: Experimental results

[0102]

[0103] From the experimental results, it can be seen that since the EGKG-RS model proposed in the embodiment uses the geographic knowledge information of the geographic location where the user is located, fully considers the influence of regional factors on the recommendation result, and is obviously better than the other three models in the accuracy indicator (Acc), and is about 0.2 better than the other three models on average. In the novelty indicator (Nov), it is 0.1 better than the other models on average, and the optimal result is only 0.002 lower than the MaORA model, which still shows that the model proposed in the embodiment has better recommendation result quality and is more popular with users. Since EGKG-RS uses a graph neural network model, it has strong information representation and learning ability, can extract more rich content from a large knowledge graph and recommend it to users, and is therefore 0.5 better than the other indicators on average in the diversity indicator (Div). In the explainability indicator, the EGKG-RS model also leads the MaORA model by 14 in the average value, and can generate a path according to the recommendation result, and explain the recommendation reason by calculating the embedded path, so that users are more willing to accept the recommended result of the system.

[0104] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the methods of practicing the present application and are not intended to limit the scope of the present application in any way. Various modifications and alterations of this application can be made by those skilled in the art without departing from the scope and spirit of this application, which is measured by the appended claims and their equivalents.

Claims

1. A geographic information knowledge graph recommendation method, characterized in that, The geographic information knowledge graph recommendation method includes the following steps: Step 1: In the input layer, obtain the current user-project interaction knowledge graph and the geographic information knowledge graph of the environment in which the current user interacts with the project. The embedding matrices of the geographic information knowledge graph and the user-item interaction knowledge graph are shown in the following formula: E g =[u 1,G ;…;u N,g ;p 1,g ;…;p M,g ;v 1,g ;…;v W,g ] E f =[u 1,f ;…;u N,f ;p 1,f ;…;p M,f ;v 1,f ;…;v Q,f ] User-geographic entity vector embedding and user-item interaction vector embedding are represented as follows: Where u represents the embedding vector, i is used to identify the user, g represents the geographic entity number in the knowledge graph, and u i,g Represents the vector embedding of user-geographic information interaction; f is the function number, u i,f This represents the vector embedding of user-project interactions, where d is the dimension of the embedding. In this context, p represents a location-based service item. i,g p represents the embedding vector of the i-th geographic vector in the geographic knowledge graph. i,f This represents the embedding vector of the i-th functional vector in the geographic knowledge graph, where p serves to connect v and u. The embedding vectors of functional entities and geographic entities are represented in the user-item knowledge graph and the geographic information knowledge graph, respectively. Where v represents the embedding vector in the knowledge graph; h is the user identity identifier, f is the function number, and k is the geographic entity number; N and M represent the number of users in the geographic information knowledge graph and the user-item interaction knowledge graph, respectively; W and Q represent the number of entities in the geographic information knowledge graph and the user-item interaction knowledge graph, respectively. Step 2: In the graph convolutional layer, information from the two knowledge graphs needs to be aggregated and delivered to the user. Therefore, the propagation rule in the graph convolutional layer is defined as follows: in, This represents the knowledge graph of the i-th user. This represents the k-th geographic entity, and ⊙ is the Hadama inner product; This represents geographic entity samples related to the user. The implementation of this function in a geographic knowledge graph-based recommendation system requires multiple intent-driven mechanisms, including geographic intent I. g and functional intent I f e j,g User u i Interest score, β(i,j) is the interest score e j,g Attention score; The calculation method for β(i,j) is shown in the following formula: in It is the vector of the i-th user after one-hot encoding. It is the embedding vector of the j-th intent about geographic entity g in the geographic information knowledge graph, after one-hot encoding. It belongs to geographical intent I g For all individuals, exp is an exponential function with base e; The entire graph convolutional network updates the weights of each node according to the following formula. Specific propagation rules are needed between multi-layered neural networks to adjust the weight coefficients. To obtain information from higher-order graph neural networks, information propagation between lower and higher layers follows these rules: The information learned from the two different partitions by the convolutional layers in the above figure needs to be fused together to form the final user embedding vector and geographic entity embedding vector, so as to calculate the matching degree between the two; (l) and (l-1) represent the l-th and l-1-th layers in the graph convolutional network, respectively. g Indicating geographical intent, I f Indicates functional intent, r j,g It is the embedding vector of the i-th relation in the geographical knowledge graph, r t,f It is the embedding vector of the t-th relation in the functional knowledge graph; e j,g The embedding vector of the j-th intent in the geographic knowledge graph, e t,f It is the embedding vector of the t-th intent in the functional knowledge graph; It refers to the relationships of interest to geographical entities in a geographical knowledge graph. User u i The geographic entity of interest; β(i,j) is the user u i Interest score for geographic entities e j,g The attention score, β(i,t) is the user u i Interest score on function e t,f The attention score; the information fusion calculation method for the i-th user and the i-th geographic entity is as follows: Step 3: The counterfactual learning layer uses counterfactual learning methods to add information from the geographic information knowledge graph to the final recommendation results, assisting in the generation of the recommendation results.

2. The geographic information knowledge graph recommendation method according to claim 1, characterized in that: To obtain better recommendation results, the optimization process of the model was constrained by four indicators: accuracy, diversity, novelty, and interpretability. The calculation methods for the four indicators are shown in the following formulas: Among them, the max Accuracy metric is used to constrain the accuracy of the recommendation results, the max Novelty metric is used to constrain the novelty of the recommendation results, the max Diversity metric is used to constrain the diversity of the recommendation results, and the max Explainability metric is used to constrain the interpretability of the recommendation results. R u T represents a user-item recommendation list. u Indicates the items that users like, |R u | represents the length of the recommendation list, |R u ∩T u | indicates the length of the recommended items in the user's preferred list; num is the number of users, r i s is the number of times a user rates item i; s(i,j) represents the similarity between different items in the recommendation results; r i,j-m It is the user's rating of the project JM, S j-m,j′-m It's the similarity between the two projects. It is the sum of user ratings for all items.

3. A geographic information knowledge graph recommendation system, characterized in that, The method for implementing the geographic information knowledge graph recommendation method of claim 1 or 2 includes: an input layer, a graph convolutional layer, and a counterfactual learning layer; The input layer is used to map the geographic knowledge graph and the user-project interaction knowledge graph into embedding vectors, which unifies the knowledge storage format of the two databases. This facilitates the subsequent fusion of information from the two databases for learning and helps to obtain more complete, structured and fine-grained information. The graph convolutional layer is responsible for learning the two types of vectors obtained from the input layer, respectively, to obtain high-dimensional vector information of the two types of knowledge; The counterfactual learning layer fuses two completely different high-dimensional information obtained from the graph convolutional layer, uses counterfactual reasoning operations, and uses geographic information to constrain the recommendation process to obtain more accurate recommendation results.

Images