Embedded preference enhanced community search method, system and program and storage medium

Abstract

The invention provides an embedded preference enhanced community search method, system and program and a storage medium. The system comprises a social network node embedding device, a community searching device, a random walk sequence obtaining device and a system control device. According to the method, embedding learning and community search are dynamically combined, node embedding is learned by using a Check2Vec model, the modeling capability of sign-in information is improved, and node attributes and sign-in information are comprehensively considered in community search, so that a more accurate user-position bipartite community is obtained. The method has the advantages of multi-modal information integration, suitability for multiple data fields, flexible adjustability and the like, and provides a powerful means for improving the community search effect and the individuation degree.

Description

Technical Field

[0001] The present invention belongs to the technical field of social network, and in particular relates to a community search method, system, program and storage medium embedded with preference enhancement. Background Art

[0002] Generally speaking, the community search (CS) problem in large-scale graphs can be divided into community search based on graph structure and community search based on the combination of graph attributes and graph structure [Fang Y, Huang X, Qin L, et al. Survey of community search over big graphs [J]. VLDB J., 2020, 29(1): 353-392.].

[0003] In general, the goal of CS is to search for high-quality communities in an online manner based on query requests. Specifically, given a homogeneous graph G and a query node Q, the goal is to find a community or dense subgraph that contains Q and satisfies the following properties:

[0004] Connectivity, that is, the nodes in the community are connected;

[0005] Cohesion, that is, the nodes in the community are closely connected to each other, is a special goodness measure.

[0006] In the community search based on graph structure, the main focus is on the cohesion of the community structure. The measurement of community structure cohesion is usually defined using some classic subgraph cohesion indicators, mainly k-core, k-truss and k-clique metrics [Zhang C, Zhang Y, Zhang W, et al. Efficient Maximal Spatial CliqueEnumeration [C] / / 35th IEEE International Conference on Data Engineering, ICDE2019, Macao, China, April 8-11, 2019. IEEE, 2019: 878-889.], as shown below:

[0007] The k-core is the largest (maximal) subgraph of G, where every node has degree at least k.

[0008] k-truss[Akbas E, Zhao P.Truss-based Community Search:a Truss-equivalence Based Indexing Approach[J].Proc.VLDB Endow.,2017,10(11):1298-1309.] is the largest subgraph of G, in which every edge is contained in at least (k-2) triangles.

[0009] k-clique[Lin YK, Lin WC, Wang C CK-Closest Points and Maximum CliquePruning for Efficient and Effective 3-D Laser Scan Matching[J].IEEE Robotics Autom. Lett., 2022, 7(2): 1471-1477.] is a set of k nodes of G, and each pair of nodes has an edge.

[0010] The three structures have their own advantages and disadvantages. Currently, the most studied community search methods are based on k-core. Compared with the other two, the k-core structure consumes the least computing resources because it uses the node degree as the subgraph compactness control condition. k-clique uses the complete subgraph as the compactness control condition and is often used to identify overlapping communities. The k-truss-based community search method is often used for some problems that require high structural cohesion. The algorithm uses the number of triangles where the edges are located as the subgraph compactness control condition. The k-truss structure usually has higher structural cohesion than other methods, but its computing resource consumption is also large.

[0011] Classic community search algorithms are mainly based on some isomorphic graphs, such as the user social network parts of Facebook, Twitter, and Amazon. Some studies also extract the user social network of LBSN for community search. Sozio et al. [SozioM, Gionis A. The community-search problem and how to plan a successful cocktail party [C] / / RAO B, KRISHNAPURAM B, TOMKINS A, et al. Proceedings of the 16th ACMSIGKDDInternational Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, July 25-28, 2010. ACM, 2010: 939-948.] proposed the first global algorithm to find the (kd)-core containing q. Zhu et al. [Zhu Junchao, Wang Chaokun. Community search method under complex conditions [J]. Journal of Software, 2019 (03vo 30): 552-572.] proposed community search under complex conditions and conducted experiments on the LBSN dataset. However, these works assume non-spatial graphs and ignore the location information of nodes.

[0012] Community search research on LBSN is the latest research hotspot. For example, in the SAC algorithm [Fang Y, Wang Z, Cheng R, et al. On Spatial-Aware Community Search [J]. IEEE Trans. Knowl. Data Eng., 2019, 31 (4): 783-798.], it is assumed that each user is associated with a location and a space-aware community (SAC) is determined. This paper studies the problem of online community search on spatial graphs, and calls communities that meet connectivity, structural cohesion, and spatial cohesion space-aware communities. However, this paper associates each user with a location, ignoring the embedded information of LBSN and the attribute information of nodes. Guo et al. [Guo F, YuanY, Wang G, et al. Multi-attributed Community Search in Road-social Networks[C] / / 37th IEEE International Conference on Data Engineering, ICDE 2021, Chania, Greece, April 19-22, 2021. IEEE, 2021: 109-120.] In order to solve the problem of inaccurate results caused by inaccurate preference vectors input by users, a multi-attributed community search method (Multi-attributed Community Search, MAC) based on road networks was studied and proposed. However, this paper only considers the user's attribute information, and ignores the attribute information of the location and embedded heterogeneous information.Junghoon Kim et al. [Kim J, Guo T, Feng K, et al. Densely Connected User Community and Location Cluster Search in Location-Based Social Networks [C] / / MAIER D, POTTINGER R, DOAN A, et al. Proceedings of the 2020 International Conference on Management of Data, SIGMOD Conference 2020, [Portland, OR, USA], June 14-19, 2020. ACM, 2020: 2199-2209.] proposed the community search problem (GeoSocial Community Search, GCS) with integrated location information given a location-based social network and a set of query nodes. This paper studies the community search algorithm with the goal of maximizing the embedding density. By traversing the k-core structure containing all query nodes in the user network graph and the location subgraph, a pair of k-core structure combinations with the largest embedding density are selected as the result set. Summary of the invention

[0013] The object of the present invention is to provide a community search method, system, program and storage medium embedded with preference enhancement.

[0014] The purpose of the present invention is achieved through the following technical solutions:

[0015] A community search method with embedded preference enhancement, the specific steps are as follows:

[0016] Step 1: Construct a model Check2Vec that contains the embedded representation of LBSN user nodes and location nodes, and ensure node proximity by maximizing the cosine similarity between the four types of nodes in the hypergraph: location, time, semantics, and user domain, and their best fit line;

[0017] Step 2: In the improved community search algorithm AGSCS+, the check-in vector representation of the user and location nodes learned by Check2Vec and the corresponding best fit line are used to perform community search;

[0018] Step 3: The AGSCS+ algorithm combines node attribute similarity and check-in similarity when calculating the score, and adjusts the impact of attribute information and check-in information on the search result community by inputting parameter p;

[0019] Step 4: Through the random walk strategy, perform jump walk sequence sampling in the user-location bipartite graph to obtain the sequence of user and location nodes for learning the check-in hyperedge information;

[0020] Step 5: Check2Vec algorithm learns node embeddings and uses the idea of ​​the best fit line in linear regression to maximize the cosine similarity between node embeddings and the best fit line.

[0021] Step 6: Use negative sampling techniques to learn vector representations of user and location nodes to better capture the relationship between nodes;

[0022] Step 7: In the community search phase, the comprehensive score calculation formula in the AGSCS+ algorithm is used to comprehensively consider the attribute similarity and check-in similarity of the user and location nodes to obtain the optimal user-location bipartite community.

[0023] A community search system with embedded preference enhancement, comprising a social network node embedding device, a community search device, a random walk sequence acquisition device and a system control device;

[0024] The social network node embedding device learns vector representations of user and location nodes, optimizes cosine similarity through a node embedding learning algorithm, and combines random walk strategy to sample node sequences and embed hyperedge sequences to provide informative node embedding representations for community search;

[0025] The community search device uses the learned node vectors to construct a user-location bipartite graph, calculates the comprehensive score of the node, and obtains the optimal community through a community search algorithm;

[0026] The random walk sequence acquisition device is a sequence acquisition engine, which generates random walk sequences of users and location nodes, and samples random walk sequences embedded in hyperedges;

[0027] The system control device is a central control unit that coordinates the work processes of various devices and manages data transmission and information exchange.

[0028] Furthermore, the main task of the social network node embedding device is to learn the vector representation of user and location nodes, and optimize the cosine similarity in the form of a best fit line through a node embedding learning algorithm to capture the correlation between nodes; through a random walk strategy, a sequence of user and location nodes is generated, as well as a sequence of sampled embedded hyperedges, providing a rich and meaningful node embedding representation for community search.

[0029] Furthermore, the community search device uses the node vectors learned by Check2Vec to construct a user-location bipartite graph to implement the core algorithm of community search; by calculating the comprehensive score of the node and comprehensively considering the weights of attribute information and embedded information, the system can more accurately evaluate the similarity between nodes; finally, the optimal community is found through the community search algorithm, realizing community search based on embedded information.

[0030] Furthermore, the random walk sequence acquisition device is responsible for generating random walk sequences of users and position nodes; through the jumping walk strategy, the problem of non-existence of relationship between structural relationship and position nodes is solved, ensuring that the generated sequence is more diverse.

[0031] Furthermore, the system control device acts as the central control unit of the entire system, responsible for coordinating the workflow between the Check2Vec model learning device, the community search algorithm device and the random walk sequence acquisition device; through effective data transmission and information exchange, it ensures that each device operates in an orderly manner and works in coordination.

[0032] A computer-readable storage medium stores a computer program / instruction thereon, characterized in that when the computer program / instruction is executed by a processor, the steps of a community search method with embedded preference enhancement are implemented.

[0033] A computer program product includes a computer program / instruction, wherein the computer program / instruction implements the steps of a community search method with embedded preference enhancement when executed by a processor.

[0034] The beneficial effects of the present invention are:

[0035] The present invention combines embedding learning with community search, uses the Check2Vec model to learn node embedding, improves the modeling ability of check-in information, and comprehensively considers node attributes and check-in information in community search to obtain more accurate user-location binary communities. The present invention has the advantages of multimodal information integration, applicability to multiple data domains, and flexible adjustability, and provides a powerful means to improve the community search effect and personalization. . BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the LBSN of the present invention;

[0037] Figure 2 A structural block diagram of the system described in the present invention;

[0038] Figure 3 Schematic diagram of hyperedges for representation learning embedding;

[0039] Figure 4 Schematic diagram of random walk strategy;

[0040] Figure 5 Comparison chart of running time for changing parameter p;

[0041] Figure 6 Attribute similarity and check-in density map on Yelp;

[0042] Figure 7 Attribute similarity and check-in density graph on Gowalla;

[0043] Figure 8 Attribute similarity and check-in density map on Foursquare;

[0044] Fig. 9 This is a comparison chart of running time on Yelp;

[0045] Fig.10 Score comparison chart for attributes;

[0046] Fig.11 It is a structural density comparison chart;

[0047] Fig.12 This is a comparison chart of check-in density. DETAILED DESCRIPTION

[0048] The present invention is further described below in conjunction with the accompanying drawings.

[0049] Figure 1 This is a schematic diagram of the LBSN of the present invention, for example:

[0050] User u eats at breakfast shop p1 in the morning, starts working at company p2 in the morning, eats lunch at restaurant p3 at noon, goes to supermarket p4 after lunch, and then returns to company p2 to continue working. After get off work, he goes to the cinema p5 and returns to apartment p6 in the evening. In this case, user u checks in at multiple locations in one day, and user u may rate these places on social platforms. At the same time, these locations are always checked in by a large number of other people, such as colleagues and clients in the company, employees and customers in the mall, etc.

[0051] Figure 2 The structure block diagram of the system of the present invention is shown, comprising:

[0052] Social network node embedding device: This device is responsible for learning the vector representation of user and location nodes, optimizing cosine similarity through node embedding learning algorithm, combining random walk strategy to sample node sequence and embed hyperedge sequence, and providing informative node embedding representation for community search. This vector representation plays a key role in subsequent community search, helping the system better understand the relationship between users, locations and embedded information.

[0053] Community search device: Use the node vectors learned by Check2Vec to construct a user-location bipartite graph and implement the core algorithm of community search. By calculating the comprehensive score of the node and considering the weights of attribute information and embedded information, the system can more accurately evaluate the similarity between nodes. Finally, the optimal community is found through the community search algorithm, realizing community search based on embedded information. The role of this device is to provide more intelligent and personalized community search results to better meet user needs.

[0054] Random walk sequence acquisition device: As a sequence acquisition engine, this device generates random walk sequences of users and location nodes, and samples random walk sequences embedded in hyperedges. Through the jump walk strategy, the problem of non-existent relationships between structural relationships and location nodes is solved, ensuring that the generated sequences are more diverse. These sequences are not only used for learning node embedding, but also provide rich sequence information for community search, helping the system to more comprehensively understand the association between users and location nodes, and improving the robustness and generalization ability of the model.

[0055] System control device: As the central control unit, this device is responsible for coordinating the workflow between the Check2Vec model learning device, the community search algorithm device, and the random walk sequence acquisition device. Through effective data transmission and information exchange, it ensures that each device operates in an orderly manner and works together. The role of the overall system control device is to improve the overall efficiency of the system and ensure that each component can cooperate organically, so as to achieve the smooth operation of the community search system with embedded preference enhancement.

[0056] Figure 3 To represent the learning embedding hyperedge diagram, specifically:

[0057] For user nodes in the user-location bipartite graph, the random walk scheme of the present invention performs a classical random walk algorithm on user nodes and location nodes based on the path guided by the check-in edge. For example, for each user node encountered, a set of check-in hyperedges are sampled from the corresponding user by staying at the user node. When iterating on each user node in the random walk sequence, not only is it learned from the heterogeneous edges such as {user-location-user...}, but it is also learned from its check-in hyperedges {location, time, semantics, user} by staying at the corresponding user node. In other words, the node embedding learning process alternates between the user-location heterogeneous structure and the check-in hyperedges. In order to perform a classical random walk on a heterogeneous network, the present invention uses the same strategy as the existing work [13,23] to generate a sequence of length on each user node, and the order of walking is determined according to the relationship between the user and the location.

[0058] In addition, in order to balance the impact of attribute and check-in hyperedges on learned node embeddings, Check2Vec introduces an adjustable parameter p to control the emphasis of both. Specifically, the present invention uses parameter p to determine the impact of attributes and check-ins on node attribute scores. In the following, two types of node embedding learning processes are described in detail, which preserve the n-way node proximity of hyperedges. The present invention first introduces the AGSCS community search algorithm process combined with representation learning, and then introduces the random walk sequence acquisition algorithm and the method of learning on check-in hyperedges.

[0059] Figure 4 This is a schematic diagram of the random walk strategy, specifically:

[0060] In order to better learn the information of the check-in hyperedge, a jumping walk sequence is selected as the random walk sequence for learning the check-in hyperedge. When learning the check-in vector representation of the user node, for a given context window of length k, k contexts are iterated for each user node u to obtain k pairs of user nodes. Specifically, a window of length k is generated on the left side of the user node u, and another user node connected to this position node is found through a random position node connected to u. This is repeated twice to complete the window sequence sampling of u. For the user node u, k pairs of user nodes for learning are obtained. At the same time, for the user node u, k pairs of check-in hyperedges need to be learned, and a total of 2k edges are learned.

[0061] Because the community search algorithm AGSCS of the present invention needs to consider both user and location node types, it is also necessary to learn the reverse check-in representation of the location node. The learning of the location node also follows a similar process, except that the roles of user and location are swapped, resulting in 2k edges for learning. Similar to the learning process on the user graph, the learning of the location node is also carried out in the same way, such as Figure 4 It shows the two-way movement from user to location and from location to user.

[0062] In order to verify the effectiveness of the present invention, experiments are conducted on three real datasets: Yelp, Foursquare and Gowalla. The dataset attributes are shown in Table 1, and the dataset size and index construction time are shown in Table 2.

[0063] Table 1

[0064]

[0065] Table 2

[0066]

[0067] The evaluation indicators are running time, attribute similarity, structural density, and embedding density. Specifically, the total number of sampled hyperedges is determined based on the number of check-in data and PageRank value of the node. That is, for the user u and position v, the number of check-in data is n, and the PageRank value is pr, then the number of sampled hyperedges for this user is (Round down). Secondly, in order to sample more recent embedded hyperedges, set a sampling ratio. First, sort the check-in data from the user to the location in chronological order. The closer the date, the greater the sampling probability. For example, between user u and location v, the number of sampled hyperedges is Divide the n embedded data into 10 time periods t according to time i , i∈[1,10], the closer the time (the smaller i is), the more samples are taken. If there are not enough samples in this time period, the samples are taken in reverse order. In the above way, the valid check-in embedding information between user u and location v can be sampled, thereby obtaining a data set for representation learning.

[0068] For the method Check2Vec of the present invention, the same random walk parameters (l=80, r=10, k=10) as those of DeepWalk and Node2vec are used. If not specified, the node embedding dimension d of all methods in all experiments is uniformly set to 128 in tensor form, and the number of negative samples γ is set to 10.

[0069] First, we change the value of parameter p to conduct experiments, mainly to observe the changes in attribute similarity and embedding density after changing parameter p. First, the experimental results of running time are as follows Figure 5 As shown in the figure, it can be seen that the overall trend is fluctuating within a range, indicating that changing the parameter p does not affect the experimental time. This is because the main time consumption of the algorithm of the present invention comes from the construction of the position graph and the search of the k-TTC infrastructure. Whether the parameter p is introduced or not only causes some changes in the number of nodes within this k-TTC range, but the total time remains at a similar level.

[0070] Figure 6is the change of attribute similarity and check-in density of communities on Yelp with parameter p. The attribute similarity of communities decreases as parameter p increases, while the check-in density is just the opposite. It increases with the increase of parameter p, and the increase slows down after p = 0.4. Combining these two figures, we can see that, as shown in Definition 5, parameter p makes the attribute similarity and check-in density of communities show an inverse trend. It can be found that when p = 0.45, the attribute similarity has a higher value, and the check-in density has entered a stable stage of slow increase. Therefore, setting the parameter to p = 0.45 can obtain higher attribute similarity and check-in density, so that the searched community can satisfy both attribute similarity and check-in preference.

[0071] akin, Figure 7 This is an experiment conducted on the Gowalla dataset. It can be found that the attribute similarity on the Gowalla dataset also decreases as the parameter p increases. Unlike Yelp, the variation range of Gowalla is relatively large. This may be because Gowalla's attributes are self-generated, which is relatively more random and causes a larger variation range. The check-in density is the same as Yelp. It increases first and then stabilizes as the parameter p increases. It shows a steady growth after p = 0.35. Figure 8 The following are the experimental results on the Foursquare dataset. It can be found that the trends of attribute similarity and check-in density are the same as those of Yelp and Gowalla. Unlike these two, the changes in attribute similarity and check-in density on Foursquare are the smallest. This is because Foursquare has the smallest proportion of check-in data, and Foursquare is a sparse graph, so check-in has little impact on the results.

[0072] In summary, these experimental results can prove that parameter p has a relatively small impact on the running time. As the value of parameter p increases, the attribute similarity of the community decreases and the check-in density increases. This is just as shown in the scoring calculation formula. By changing the value of parameter p, the influence ratio of attributes and check-ins can be changed. On different data sets, the community can simultaneously satisfy both high check-in density and attribute similarity by selecting a suitable p value. For example, the parameters on Yelp, Gowalla, and Foursquare can be selected as 0.45, 0.35, and 0.5, respectively. Therefore, the experiment proves that the method of the present invention can make the community results meet the attribute similarity and check-in preference required by the AGSCS problem.

[0073] Next, the present invention continues to experiment on the effectiveness of representation learning methods based on the three algorithms of GCS, AGSCS, and AGSCS+. First, the comparison is made in terms of running time. Fig. 9As shown in the figure, similar to the previous results, the time consumed by GCS on query nodes of different sizes is much higher than that of AGSCS and AGSCS+, while the time consumption of AGSCS and AGSCS+ is not much different.

[0074] The following is a comparison of the three algorithms GCS, AGSCS, and AGSCS+ in attribute scoring. Fig.10 These are the results on the Yelp, Gowalla, and Foursquare datasets. It can be seen that no matter which dataset is used, the attribute scores show the same trend. The attribute scores of GCS in the three datasets are much lower than those of AGSCS and AGSCS+, which shows that whether or not to consider node attributes has a great impact on the community. If node attributes are not considered, the searched community may contain a large number of low-quality nodes that are irrelevant to the query node. The attribute score of AGSCS+ is slightly lower than that of AGSCS, but the difference range on the Yelp and Gowalla datasets is not large, which means that the attribute scores of the community obtained by all algorithms proposed in the present invention are related to the attribute characteristics of the nodes themselves in the dataset.

[0075] Similar, such as Fig.11 The figure shows the comparison of structural similarity of the three algorithms GCS, AGSCS, and AGSCS+ under the same settings. It can be seen that compared with GCS, AGSCS and AGSCS+ have greater structural similarity, which shows that the present invention uses the k-TTC structure as the structural constraint to effectively improve the cohesion of the community structure. The structural similarity of AGSCS and AGSCS+ is basically the same, because these two algorithms ensure that they are connected truss structures with the maximum support value when searching for the previous UL-Community, and regardless of the attribute optimization or the introduction of the influence of the signed-in hyperedge, the algorithm always keeps optimizing the community under higher structural constraints, so their structural similarity can be maintained at the same level.

[0076] Fig.12 The following is a comparison of the check-in density between AGSCS and AGSCS+. Fig.10 From the attribute scores, we can see that although the attribute score of AGSCS+ is slightly lower than that of AGSCS, the check-in density of AGSCS+ is much higher than that of AGSCS. This is because AGSCS+ takes the check-in information into account during the score calculation process before attribute optimization. In this way, during the Pareto optimization process, some points with high attribute scores but poor check-in conditions may be deleted, and some nodes with low attribute scores but increased community check-in density may be retained. The final optimization result is a comprehensive optimal solution that considers both attribute and check-in hyperedge information. Although the attribute score may decrease slightly, the check-in density can be significantly improved.

[0077] Then, we observe the check-in density comparison experimental results of AGSCS+ and GCS. It can be seen that the community check-in density obtained by AGSCS+ is slightly lower than that of GCS on different data sets. This is because the principle of the GCS method is to use check-in density as the only optimization goal, and the structural constraints on the community are not strong. There is no requirement for the attribute similarity between the community and the query node. Although the obtained community maximizes the check-in density, it contains many redundant nodes that are not related to the query node and cannot meet the user's personalized query expectations. In summary, AGSCS+ is a community search method that can simultaneously make the community meet structural cohesion, attribute similarity, and check-in preference. Experiments have proved that the method of the present invention is effective.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A community search method with embedded preference enhancement, characterized in that: The specific steps are as follows: Step 1: Construct a model Check2Vec that contains the embedded representation of LBSN user nodes and location nodes, and ensure node proximity by maximizing the cosine similarity between the four types of nodes in the hypergraph: location, time, semantics, and user domain, and their best fit line; Step 2: In the improved community search algorithm AGSCS+, the check-in vector representation of the user and location nodes learned by Check2Vec and the corresponding best fit line are used to perform community search; Step 3: The AGSCS+ algorithm combines node attribute similarity and check-in similarity when calculating the score, and adjusts the impact of attribute information and check-in information on the search result community by inputting parameter p; Step 4: Through the random walk strategy, perform jump walk sequence sampling in the user-location bipartite graph to obtain the sequence of user and location nodes for learning the check-in hyperedge information; Step 5: Check2Vec algorithm learns node embeddings and uses the idea of ​​the best fit line in linear regression to maximize the cosine similarity between node embeddings and the best fit line. Step 6: Use negative sampling techniques to learn vector representations of user and location nodes to better capture the relationship between nodes; Step 7: In the community search phase, the comprehensive score calculation formula in the AGSCS+ algorithm is used to comprehensively consider the attribute similarity and check-in similarity of the user and location nodes to obtain the optimal user-location bipartite community.

2. The community search system with embedded preference enhancement according to claim 1, characterized in that: It includes a social network node embedding device, a community search device, a random walk sequence acquisition device and a system control device; The social network node embedding device learns vector representations of user and location nodes, optimizes cosine similarity through a node embedding learning algorithm, and combines random walk strategy to sample node sequences and embed hyperedge sequences to provide informative node embedding representations for community search; The community search device uses the learned node vectors to construct a user-location bipartite graph, calculates the comprehensive score of the node, and obtains the optimal community through a community search algorithm; The random walk sequence acquisition device is a sequence acquisition engine, which generates random walk sequences of users and location nodes, and samples random walk sequences embedded in hyperedges; The system control device is a central control unit that coordinates the work processes of various devices and manages data transmission and information exchange.

3. The community search system with embedded preference enhancement according to claim 2, characterized in that: The main task of the social network node embedding device is to learn the vector representation of user and location nodes, and optimize the cosine similarity in the form of a best fit line through a node embedding learning algorithm to capture the correlation between nodes; through a random walk strategy, a sequence of user and location nodes is generated, as well as a sequence of sampled embedded hyperedges, providing a rich and meaningful node embedding representation for community search.

4. The community search system with embedded preference enhancement according to claim 2, characterized in that: The community search device uses the node vectors learned by Check2Vec to construct a user-location bipartite graph and implement the core algorithm of community search; by calculating the comprehensive score of the node and comprehensively considering the weights of attribute information and embedded information, the system can more accurately evaluate the similarity between nodes; finally, the optimal community is found through the community search algorithm, realizing community search based on embedded information.

5. The community search system with embedded preference enhancement according to claim 2, characterized in that: The random walk sequence acquisition device is responsible for generating random walk sequences of users and position nodes; through the jumping walk strategy, the problem of non-existence of relationship between structural relationship and position nodes is solved, ensuring that the generated sequence is more diverse.

6. The community search system with embedded preference enhancement according to claim 2, characterized in that: The system control device acts as the central control unit of the entire system, responsible for coordinating the workflow between the Check2Vec model learning device, the community search algorithm device and the random walk sequence acquisition device; through effective data transmission and information exchange, it ensures that each device operates in an orderly manner and works in coordination.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: The computer program / instructions implement the steps of the method as claimed in claim 1 when executed by a processor.

8. A computer program product comprising a computer program / instructions, characterized in that: The computer program / instructions implement the steps of the method as claimed in claim 1 when executed by a processor.

Images