A knowledge graph driven remote sensing image recommendation method

Abstract

The application discloses a kind of based on knowledge graph driving's remote sensing image recommendation method, including obtaining user business document, constructs implicit theme model based on document;According to the theme label in user and implicit theme model, industry knowledge graph is constructed;Knowledge graph network feature learning is carried out using the representation learning method based on deep learning, and theme fusion feature model is constructed;User-remote sensing image historical interaction data matrix is established, and user neighbor set is calculated;The recommendation result of theme fusion feature model and the recommendation result of user neighbor set are aggregated, and the final remote sensing image recommendation result is obtained.The application combines the result of image labeling modeling, converts knowledge graph from the graph structure that is not easy to calculate into mathematical vector representation, realizes to each kind of military and civilian user push potential interest remote sensing data, and is conducive to relieving the problem of data sparsity and cold start, improves recommendation effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image recommendation, and particularly relates to a remote sensing image recommendation method based on knowledge graph driving. BACKGROUND

[0002] The rapid development of mobile Internet, artificial intelligence and other technologies has brought a lot of convenience to people's work and life, but at the same time, it has also brought the problem of information overload. Search engines and recommendation systems are representative technologies to solve the problem of information overload. Compared with search engines, the superiority of recommendation systems lies in that they can take into account personalized needs and provide the most suitable options. Recommendation systems have experienced twenty years since their inception, including content-based recommendation algorithms, collaborative filtering-based recommendation algorithms and hybrid recommendation algorithms.

[0003] The idea of content-based recommendation algorithm is to use the historical behavior information of users to learn the hidden features of users or items from different information, calculate the preference degree of users by obtaining text information, user comments and the like, and then calculate the similarity, and recommend items to users according to the value.

[0004] The collaborative filtering-based recommendation algorithm uses the "people are divided into groups" idea to analyze the user group similar to the target user, mines their preference information and generates a recommendation list for the target user. Usually, it first finds the most similar users or the most similar item evaluations, and then analyzes the similarity between neighbors and users or items to push the optimal recommendation result to the user.

[0005] The hybrid recommendation algorithm is used when the recommendation accuracy obtained by using only one recommendation algorithm is not high. Different recommendation algorithms are integrated into one algorithm for hybrid recommendation to improve the recommendation accuracy.

[0006] With the advent of the big data era, data mining has become a key problem faced by recommendation systems. The emergence of knowledge graph provides an effective way for the design of recommendation systems in the big data environment, and also provides a new solution for the integration of data. Based on representation learning, the user and item structure and semantics are embedded into a low-dimensional semantic space, and the similarity between entities and relationships is calculated by using semantic distance (based on link number, Euclidean distance, cosine distance), correlation coefficient, implicit semantic feedback path, entropy weight coefficient, association rule and other similarity algorithms to recommend to alleviate the problem of data sparsity and cold start, improve the recommendation effect and storage efficiency.

[0007] Currently, remote sensing image data (RSID) faces problems of efficient storage, management, analysis and pushing, among which the pushing problem is related to the accurate service of the remote sensing data distribution system to the business of multiple industries. The existing acquisition of remote sensing image data needs professional field knowledge to identify the attribute information of the image to be acquired, but often does not utilize the object data related to the remote sensing image data, causing the fragmentation of the data, making the acquired remote sensing image data seriously single, and unable to obtain the rich object information behind the remote sensing image, so that the data utilization rate is not high. SUMMARY

[0008] In view of the above problems in the prior art, the present application provides a remote sensing image recommendation method based on a knowledge graph.

[0009] In order to achieve the above-mentioned purpose of the application, the technical scheme adopted by the present application is as follows:

[0010] A remote sensing image recommendation method based on a knowledge graph, comprising the following steps:

[0011] S1, acquiring user business documents and constructing an implicit topic model based on the documents;

[0012] S2, constructing an industry knowledge graph according to the user and the topic labels in the implicit topic model;

[0013] S3, using a representation learning method based on deep learning to learn the network features of the knowledge graph and constructing a topic fusion feature model;

[0014] S4, establishing a user-remote sensing image historical interaction data matrix and calculating a user neighbor set;

[0015] S5, aggregating the recommendation results of the topic fusion feature model and the user neighbor set to obtain the final remote sensing image recommendation results.

[0016] Optionally, step S1 specifically comprises the following steps:

[0017] S11, acquiring user business documents;

[0018] S12, using an implicit Dirichlet allocation model to model the topics of the user business documents and constructing an implicit topic model;

[0019] S13, using Gibbs sampling to sample the implicit topics of the implicit topic model to obtain the topic distribution of the documents and keywords.

[0020] Optionally, step S3 specifically comprises the following steps:

[0021] S31, knowledge graph network feature learning is performed by using a representation learning method based on deep learning, and entities representing users, topics and remote sensing images in the knowledge graph are mapped to a K-dimensional space;

[0022] S32, a cosine similarity algorithm of vectors is used to calculate a correlation coefficient between a topic entity and a user;

[0023] S33, a user-topic pair is labeled, similarity under a single user context feature is calculated, and a topic fusion feature vector is constructed.

[0024] Optionally, step S4 specifically comprises the following steps:

[0025] S41, a user-remote sensing image historical interaction data matrix is established;

[0026] S42, a user-topic score matrix is calculated according to the user-remote sensing image historical interaction data matrix;

[0027] S43, the user-remote sensing image historical interaction data matrix and the user-topic score matrix are merged to obtain a merged user-topic score matrix;

[0028] S44, a Weighted Slope One algorithm is used to fill in empty items in the merged user-topic score matrix to obtain a filled user-topic score matrix;

[0029] S45, a K-means clustering algorithm based on density and distance is used to cluster users;

[0030] S46, distances between a target user and cluster centers in a clustering result are calculated, and the target user is added to a cluster closest to the target user to obtain a nearest neighbor set of the target user.

[0031] Optionally, the calculation method of the user-topic score in step S42 is:

[0032] According to a user set of the scored topic, a score deviation between a user scored topic and a user topic to be predicted is calculated;

[0033] According to the score deviation and a score of the user scored topic, a predicted score of the user topic to be predicted is calculated.

[0034] Optionally, the calculation formula of the score deviation is:

[0035]

[0036] dis i,j is a score deviation between a user scored topic i and a user topic to be predicted j, U i,jFor the user set scored on both topic i and topic j, u' is a user in the user set U i,j u′,i r is the score of user u' on topic i, r u′,j r is the score of user u' on topic j, num i,j num is the number of users in the user set U i,j .

[0037] Optionally, the calculation formula of the predicted score of the user to be predicted on the score topic is:

[0038]

[0039] Wherein, pre u,j r is the predicted score of user u on topic j, I u I is the set of all scored topics of user u, r u,i r is the score of user u on topic i.

[0040] Optionally, the method for recommending the result of the topic fusion feature model in step S5 is:

[0041] The similarity between the topic fusion feature vector and the entity vector of the remote sensing image is calculated to obtain the correlation score between the remote sensing image and the topic, and the Top-K recommendation list with the largest correlation is obtained according to the association between the user and the topic.

[0042] Optionally, the method for recommending the result of the user neighbor set in step S5 is:

[0043] The similarity between the target user and other users in the cluster is calculated based on the Pearson correlation coefficient, the predicted score of the target user on the un-scored project is calculated according to the score of the user in the nearest neighbor set of the target user on the project, and the first N projects with the highest predicted score are recommended to the target user.

[0044] The present application has the following beneficial effects:

[0045] The present application realizes the modeling of users or businesses by deeply analyzing the internal business documents and industry demand documents related to land, agriculture, ecology and other fields of industry users, and automatically extracting key element labels to form a user and business knowledge system and obtain application requirements of each business field. The knowledge graph is converted from a non-easy-to-calculate graph structure to a mathematical vector representation by combining the modeling results of image labeling, potential interest remote sensing data is pushed to various military and civilian users, the problems of data sparsity and cold start are alleviated, and the recommendation effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a flowchart of a remote sensing image recommendation method based on knowledge graph driving in an embodiment of the present application.​

[0047] Figure 2 Fig. 1 is a probability graph based on an implicit topic model LDA in an embodiment of the present application;

[0048] Figure 3 Fig. 2 is a schematic diagram of a hybrid recommendation generation process based on a knowledge graph in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0050] The embodiment provides a remote sensing image recommendation method based on knowledge graph driving. By deeply analyzing internal business documents and industry demand documents related to industry users in the fields of land, agriculture and ecology, key element labels are automatically extracted to form a user and business knowledge system, and application requirements of each business field are obtained to realize user or business modeling. The knowledge graph is converted from a non-computable graph structure to a mathematical vector representation by combining the modeling results of image labeling, potential interest remote sensing data is pushed to various military and civilian users, and the problems of data sparsity and cold start are alleviated, and the recommendation effect is improved.

[0051] As shown in Figure 1 The remote sensing image recommendation method based on knowledge graph driving provided by the embodiment includes the following steps S1 to S5:

[0052] S1, obtaining user business documents, and constructing an implicit topic model based on the documents;

[0053] In an optional embodiment of the present application, step S1 specifically includes the following sub-steps:

[0054] S11, obtaining user business documents;

[0055] S12, modeling the topics of the user business documents by using an implicit Dirichlet allocation model, and constructing an implicit topic model;

[0056] S13, sampling the implicit topics of the implicit topic model by using Gibbs sampling, and obtaining the topic distribution of the documents and keywords.

[0057] Specifically, the user business documents are industry user and business documents, and the documents are daily published documents and business texts in the industry, which lay a foundation for extracting entity library and relationship library in subsequent data.

[0058] In order to realize the analysis of user and business demand in the industry, and solve the problem of unstable performance of the method of extracting article theme based on clustering, the key technology is to model the theme of the business text. The existing method is mainly to construct the theme of the document by the method of word clustering. That is, the internal information of the document, that is, the aggregation degree of the word, is used to discover the theme of the document. But this method will exist the problem of limited information of single document, and is easily affected by the similarity degree of the word, so the present application selects to use the implicit theme model to construct the theme of the document.

[0059] The implicit theme model based on the document can learn the implicit theme by using large-scale documents, which avoids the problem of insufficient information of single document itself, and can obtain relatively meaningful and stable theme information, effectively avoiding the uncertainty factors generated by clustering on single document. Accordingly, the present application selects the theme modeling mode based on the latent Dirichlet allocation (LDA) to model the document, and the probability graph based on the implicit theme model LDA is as shown in Figure 2 The basic idea is that a single document can be mixed by multiple themes, and each theme has multiple theme words. First, a P (word / theme) is calculated according to the probability of the theme word appearing in the theme, then a P (theme / document) is calculated according to the probability of the theme appearing in the document, and according to the P (word / theme) and the P (theme / document), the probability of the word appearing is calculated, and the calculation formula is:

[0060]

[0061] Therefore, each document can be represented as a mixed distribution of K implicit themes, and each theme K is represented as a multinomial distribution on W words For any document d j , its theme distribution θ j is generated from a Dirichlet prior, and the prior parameter is α. For the i-th word x j in the document d ij , its theme z ij =k is generated from θ j , and the word x ij is generated from φ k . The LDA generation process can be represented as

[0062] θ j ~Dir (α)

[0063]

[0064] z ij =k~θ j

[0065]

[0066] Where Dir(*) represents the Dirichlet distribution.

[0067] Then, the LDA model is learned using Gibbs sampling. First, the integrals of θ and φ are removed using the duality property, and only the latent variable z is sampled. The topic z is then used to calculate... ij Given a training corpus excluding w ij Besides the other words and the topics obtained from their previous sampling, z can be obtained. ij The conditional probability of a word, and the formula for calculating its topic probability, is as follows:

[0068]

[0069] in, The corresponding word w ij The topics from the previous round of sampling were excluded.

[0070] w is obtained each time through sampling ij New theme z ij Then, the co-occurrence frequency matrix C of "word-topic" was analyzed. word (where each element C) wk The document-topic co-occurrence matrix C represents the number of times word w is assigned to topic k in the training corpus. doc (where each element C) kj Document d represents document d j The number of times each word is assigned to topic k), and a vector C recording the number of topics. topic (where each element C) k This represents the number of times topic k appears in the training corpus. After multiple sampling iterations, based on matrix C... word C doc Sum vector C topic The two parameters θ and φ of LDA are calculated as follows:

[0071]

[0072] Where, θ kj Document d represents document d j The probability φ of topic k in the middle wk This represents the probability of word w in topic k.

[0073] After learning a topic model from large-scale data, the topic distribution of documents and given words is obtained, and the higher distribution is selected as the topic of the document and given words.

[0074] The present application uses a text analysis tool to perform subject modeling, divides key information in a document and original business documents in a data source into words, divides different subject subgroups based on a clustering method, and analyzes original text in the data source to obtain demand subject information for different industry users.

[0075] S2, constructing an industry knowledge graph according to a user and a subject label in an implicit subject model;

[0076] In an optional embodiment of the present application, the knowledge graph often contains rich object entity information, and the present application releases data value by using the network structure, and realizes accurate recommendation of remote sensing images to various users by matching user interest labels and image labels through similarity calculation.

[0077] S3, using a representation learning method based on deep learning to learn network features of the knowledge graph, and constructing a subject fusion feature model;

[0078] In an optional embodiment of the present application, the present application uses a deep learning classification model to classify the label data of the remote sensing image according to the same subject, associates the user with the remote sensing image, and realizes accurate recommendation.

[0079] Step S3 specifically includes the following steps:

[0080] S31, using a representation learning method based on deep learning to learn network features of the knowledge graph, and mapping entities representing users, subjects and remote sensing images in the knowledge graph to K-dimensional space;

[0081] S32, using a cosine similarity algorithm of a vector to calculate a correlation coefficient between a subject entity and a user;

[0082] S33, labeling a user subject pair, calculating similarity under a single user context feature, and constructing a subject fusion feature vector.

[0083] Specifically, the present application uses a knowledge embedding algorithm to represent a triple represented by a department interest label knowledge graph with a vector. The knowledge embedding algorithm includes but is not limited to: TransE algorithm, TransH algorithm, TransR algorithm, TransG algorithm, Node2Vec algorithm and HoLE algorithm. The knowledge embedding algorithm is trained by using a training set, verified by using a verification set, and tested by using a test set. The knowledge graph includes several triples (h, r, t), h represents a head entity, r represents a relationship, and t represents a tail entity.

[0084] The application uses a deep learning-based representation learning method such as Node2Vec to learn the knowledge graph network features, and maps entities representing users, topics and remote sensing images in the knowledge graph to a K-dimensional space. In the K-dimensional vector space, the closer the entities are in geometry, the greater the relevance. The subject group plans to calculate the relevance Sim(e i ,e j ) between the topic entity e i and the user e j by the cosine similarity algorithm of the vector:

[0085]

[0086] By processing the training set, the user topic pair (U i ,I j ) is labeled y ij , and the similarity of (U i ,I j ) under the single user context feature is calculated on the basis of the constructed basic knowledge graph, and the fusion feature vector is constructed to comprehensively consider the structural information and semantic information between the user and the topic, which is represented as:

[0087]

[0088] S4, a user-remote sensing image historical interaction data matrix is established, and a user neighbor set is calculated;

[0089] In an optional embodiment of the application, step S4 specifically comprises the following steps:

[0090] S41, a user-remote sensing image historical interaction data matrix is established;

[0091] S42, a user-topic score matrix is calculated according to the user-remote sensing image historical interaction data matrix;

[0092] S43, the user-remote sensing image historical interaction data matrix and the user-topic score matrix are merged to obtain a merged user-topic score matrix;

[0093] S44, a Weighted Slope One algorithm is used to fill in the empty items in the merged user-topic score matrix to obtain a filled user-topic score matrix;

[0094] S45, a K-means clustering algorithm based on density and distance is used to cluster users;

[0095] S46, the distance between the target user and each cluster center in the clustering result is calculated, and the target user is added to the nearest cluster to obtain the nearest neighbor set of the target user.

[0096] Specifically, the present application is based on user / service-remote sensing image historical interaction data matrix R m,n The user-topic score matrix R1 is calculated, and two score matrices are combined to obtain score matrix R2, and then the Weighted Slope One algorithm is used to fill in the empty items in the score matrix, reduce the sparsity of the score matrix, and obtain score matrix R3. Specifically, assuming that there is a user u, the scored items of the user u are i, and I u The user u has scored all the item sets, i.e. i∈I u The user u to be predicted is j, and the score deviation dis i,j The calculation formula is:

[0097]

[0098] dis i,j is the score deviation between the scored topic i of the user and the topic j to be predicted by the user, U i,j is the user set that has scored both the topic i and the topic j, u' is a user in the user set U i,j r u′,i is the score of the user u' to the topic i, r u′,j is the score of the user u' to the topic j, and num i,j is the number of users in the user set U i,j .

[0099] Then, the predicted score pre i,j of the target user u to the item j is calculated according to the score deviation dis u,i and the score r u,j of the target user u to the item i, and the calculation formula is:

[0100]

[0101] pre u,j is the predicted score of the user u to the topic j, I u is the set of all scored topics of the user u, r u,i is the score of the user u to the topic i.

[0102] Then, R3 is taken as the clustering input data, the initial cluster centers of the K-means clustering combined with density and distance are found, and the users are clustered into K clusters through the initial cluster centers combined with the K-means algorithm. The Euclidean distance between the target user u and the K cluster centers is calculated, and the target user u is added to the nearest cluster. In the cluster where the target user u is located, the nearest neighbor set N u of the user u is obtained.

[0103] S5, the theme fusion feature model recommendation result and the user neighbor set recommendation result are recommended, and a final remote sensing image recommendation result is obtained.

[0104] In an optional embodiment of the application, the method for recommending the theme fusion feature model in step S5 is:

[0105] The theme fusion feature vector is subjected to similarity calculation with the entity vector of the remote sensing image, to obtain a correlation score between the remote sensing image and the theme, and a Top-K recommendation list with the largest correlation is obtained according to the association between the user and the theme.

[0106] The method for recommending the user neighbor set in step S5 is:

[0107] The similarity between the target user and other users in the cluster is calculated based on the Pearson correlation coefficient, the predicted score of the target user for an un-scored item is calculated according to the score of the item by the users in the nearest neighbor set of the target user, and the first N items with the highest predicted scores are recommended to the target user.

[0108] Specifically, the knowledge graph-based recommendation generation refers to jointly analyzing seemingly unrelated data by using a knowledge graph, and through the introduction of more semantic relationships, the potential interest of a user for remote sensing image data can be discovered in depth, which is beneficial to the divergence of the recommendation result and no longer limited to a single type. However, this method cannot effectively balance the semantic information and structural information of the user or the item to be recommended, resulting in that the recommendation accuracy cannot be effectively improved. In addition, the collaborative filtering-based recommendation generation refers to clustering target users or items to be recommended according to whether they have the same attribute label, and then pushing the item to be recommended that is interested in by a certain user to other users of the same type, or pushing the item to be recommended with the same attribute to a certain user at the same time. This method can effectively improve the accuracy of recommendation, but it does not consider the potential relationship between the user or the item to be recommended, and the interest drift problem that may exist in the user, resulting in that the diversity of the recommendation result is low. To solve the above problems, the application combines the advantages of the knowledge graph-based recommendation and the collaborative filtering-based recommendation, and proposes a hybrid recommendation technology based on the knowledge graph and the collaborative filtering to complete the intelligent pushing of remote sensing image information.

[0109] The hybrid recommendation generation technology route based on the knowledge graph and the collaborative filtering is as shown in Figure 3As shown, the basic idea is to start from the traditional recommendation method based on the idea of collaborative filtering and the deep model method based on the knowledge graph, wherein first, the user / business-remote sensing image historical interaction data matrix is established by using the idea of collaborative filtering, and the recommendation of remote sensing images that the user may be interested in is completed based on the matrix decomposition recommendation technology; meanwhile, in the industry knowledge graph in which the user / business and the related theme information are associated, the theme classification of the label data of the remote sensing image is completed by using the deep learning classification model based on the remote sensing image, that is, the user and the remote sensing image can be associated according to the same theme, and the accurate recommendation is realized.

[0110] After obtaining the fusion feature vector expression of the theme, the entity vector expression representing the remote sensing image is calculated for similarity, the correlation score between the remote sensing image and the theme is obtained, that is, the theme most related to the current remote sensing image is obtained, and the Top-K recommendation list with the maximum correlation is obtained according to the association between the theme and the user business.

[0111] The similarity sim(u,v) of the target user u and other users in the cluster based on the Pearson correlation coefficient is calculated according to the similarity calculation target, and is expressed as:

[0112]

[0113] Then, according to the ratings of the items by the users in the nearest neighbor set N u of the target user u, the predicted rating of the target user u to the un-rated item is calculated, and the first N items with the highest predicted rating are recommended to the user u.

[0114] Finally, the influence of the collaborative filtering recommendation result and the knowledge graph recommendation result on the final recommendation result is comprehensively considered, and the bootstrap algorithm is used to obtain the final result.

[0115] In summary, the related theme business label is extracted from the user business document, the theme modeling of the business text is performed, and the industry knowledge graph is constructed; the user / business-remote sensing image historical interaction data matrix is established by using the idea of collaborative filtering, and the recommendation of remote sensing images that the user may be interested in is completed based on the matrix decomposition recommendation technology; meanwhile, in the industry knowledge graph in which the user / business and the related theme information are associated, the theme classification of the label data of the remote sensing image is completed by using the deep learning classification model based on the remote sensing image, that is, the user and the remote sensing image can be associated according to the same theme, to realize the accurate recommendation and retrieval application of the remote sensing image information. The knowledge graph is converted from the graph structure which is not easy to calculate into a mathematical vector representation by using the knowledge embedding algorithm, and the collaborative information between the user and the interest label is fused, which is beneficial to alleviate the data sparsity problem and improve the recommendation effect.

[0116] The present application is described in reference to the flowchart and / or block diagram of the method, device (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate a device that implements the flow Figure 1 one or more flows and / or blocks Figure 1 one or more flows and / or blocks

[0117] These computer program instructions can also be stored in a computer readable memory that can direct the computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction devices that implement the flow Figure 1 one or more flows and / or blocks Figure 1 one or more flows and / or blocks

[0118] These computer program instructions can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are performed on the computer or other programmable device to produce a computer implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 one or more flows and / or blocks

[0119] The principles and implementation manners of the present application are described in the specific embodiments, and the above embodiment description is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application scope will be changed, and the above description should not be understood as the limitation of the present application.

[0120] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application, and should be understood as the protection scope of the present application is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the protection scope of the present application.

Claims

1. A knowledge graph driven remote sensing image recommendation method, characterized in that, The method comprises the following steps: S1, obtaining user business documents, and constructing an implicit topic model based on the documents; the step S1 specifically comprises the following sub-steps: S11, obtaining user business documents; S12, modeling the user business documents by using an implicit Dirichlet allocation model, and constructing an implicit topic model; S13, sampling the implicit topics of the implicit topic model by using Gibbs sampling, and obtaining topic distributions of the documents and keywords; S2, constructing an industry knowledge graph according to the users and the topic labels in the implicit topic model; S3, performing knowledge graph network feature learning by using a deep learning-based representation learning method, and constructing a topic fusion feature model; the step S3 specifically comprises the following sub-steps: S31, performing knowledge graph network feature learning by using a deep learning-based representation learning method, and mapping entities representing users, topics and remote sensing images in the knowledge graph to a K-dimensional space; S32, calculating a correlation coefficient between the topic entities and the users by using a cosine similarity algorithm of vectors; S33, labeling the user-topic pairs, calculating the similarity under a single user context feature, and constructing a topic fusion feature vector; S4, establishing a user-remote sensing image historical interaction data matrix, and calculating a user neighbor set; the step S4 specifically comprises the following sub-steps: S41, establishing a user-remote sensing image historical interaction data matrix; S42, calculating a user-topic score matrix according to the user-remote sensing image historical interaction data matrix; S43, merging the user-remote sensing image historical interaction data matrix and the user-topic score matrix to obtain a merged user-topic score matrix; S44, filling empty items in the merged user-topic score matrix by using a Weighted Slope One algorithm, and obtaining a filled user-topic score matrix; S45, clustering the users by using a K-means clustering algorithm based on density and distance; S46, calculating distances between a target user and cluster centers in the clustering results, and adding the target user to the nearest cluster to obtain a nearest neighbor set of the target user; S5, aggregating a topic fusion feature model recommendation result and a user neighbor set recommendation result to obtain a final remote sensing image recommendation result; in the step S5, the method of the topic fusion feature model recommendation result is as follows: calculating the similarity between the topic fusion feature vector and an entity vector of the remote sensing image to obtain a correlation score between the remote sensing image and the topic; and according to the association between the user and the topic, obtaining a Top-K recommendation list with the largest correlation. 2.The knowledge graph driven remote sensing image recommendation method of claim 1, wherein, The calculation method of the user-topic score in the step S42 is as follows: calculating a score deviation between a user-scored topic and a user-predicted score topic according to a user set of the scored topic; calculating a predicted score of the user-predicted score topic according to the score deviation and the score of the user-scored topic. 3.The knowledge graph driven remote sensing image recommendation method of claim 2, wherein, The calculation formula of the score deviation is as follows: where dis i,j is the rating deviation between a user rated topic i and a user to be predicted rated topic j, u i,j is the set of users who have rated both topic i and topic j, u′ is a user in the set of users U i,j , r u′,i is the rating of topic i by user u′, r u′,j is the rating of topic j by user u′, num i,j is the number of users in the set of users U i,j .

4. The knowledge graph driven remote sensing image recommendation method according to claim 3, characterized in that, The calculation formula of the predicted score of the user-predicted score topic is as follows: where pre u,j is the predicted rating of user u for topic j, I u is the set of all rated topics for user u, r u,i is the rating of user u for topic i. 5.The knowledge graph driven remote sensing image recommendation method of claim 1, wherein, The method for recommending the result of the user neighbor set in step S5 is: calculating the similarity between the target user and other users in the cluster based on the Pearson correlation coefficient; calculating the predicted score of the target user for the items which have not been scored by the target user according to the scores of the users in the nearest neighbor set of the target user for the items, and recommending the first N items with the highest predicted scores to the target user.

Images