Community structure discovery-oriented large-scale complex network visualization method

Abstract

The invention relates to a community structure discovery-oriented large-scale complex network visualization method. The method comprises the following steps: acquiring a large-scale complex network; extracting topological structure features of the large-scale complex network by using a representation learning technology to obtain a node representation vector set; performing dimension reduction processing on the node representation vector set by using PCA and t-SNE algorithms to obtain an output vector set; recursive clustering is carried out on the output vector set, each level in the hypergraph is divided, a community formed after clustering of each level is constructed as a super node, the relation between nodes in the community is combined as a hyperedge, and a hypergraph structure is constructed; and visualizing the hypergraph structure by using a diversified visualization algorithm. By adopting the method, the visualization effect of the complex social network can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of social network visualization, and in particular to a large-scale complex network visualization method for community structure discovery. Background Art

[0002] With the continuous expansion of social network scale and the deep complexity of structure, the interaction data and relationship links between users have exploded, which has led to a sharp increase in the amount of social network graph data. In this context, social network visualization research faces unprecedented challenges. Traditional social network visualization methods need to consume huge computing resources for data processing and graph drawing when dealing with massive user data and complex relationships, resulting in low visualization efficiency. When the scale of social networks expands significantly, the density of user nodes and relationship edges increases significantly, and the visualization results are prone to visual crowding and information confusion. Faced with the huge amount of user information and complex relationships in social networks, traditional visualization methods are difficult to accurately filter and effectively present key social relationships and important user information. Once information overload occurs, it is difficult for users to quickly gain insight into key social contexts. What is even more difficult is that in social networks, it is extremely difficult to abstract and filter relationship edges, and it is difficult to accurately identify and highlight core social structures, such as key social circles, influence propagation paths, etc. At the same time, existing social network visualization methods are difficult to accurately express network structure and characteristics while taking into account aesthetics and readability. They cannot clearly present important community structures and social clusters in large-scale social networks, and often fall into the dilemma of losing one thing while gaining another. Traditional clustering and visualization methods are not effective when dealing with complex social networks. Therefore, how to choose a better visualization presentation method is still an important problem that needs to be solved urgently. Summary of the invention

[0003] Based on this, it is necessary to provide a large-scale complex network visualization method for community structure discovery that can improve the visualization effect of complex social networks in response to the above technical problems.

[0004] A large-scale complex network visualization method for community structure discovery, the method comprising: Obtain a large-scale complex network; a large-scale complex network is a social network; a social network includes nodes and edges; a node represents a user; an edge represents a relationship between users; The topological structure features of large-scale complex networks are extracted using representation learning technology to obtain a node representation vector set; the node representation vector set is subjected to dimensionality reduction processing using PCA and t-SNE algorithms to obtain an output vector set; Perform recursive clustering on the output vector set to divide each level in the hypergraph. The communities formed after clustering each level are constructed as supernodes, and the connections between the internal nodes of the community are merged into hyperedges to construct the hypergraph structure. Use diverse visualization algorithms to visualize the hypergraph structure.

[0005] The above-mentioned visualization method for large-scale complex networks for community structure discovery first defines the social network as a graph structure of "nodes (users) + edges (relationships)", and uses representation learning technology to extract the topological structure features of the complex network. Since the original social network data has high dimensions and a lot of noise, the direct processing efficiency is low. Representation learning can map the high-dimensional topological structure into a low-dimensional representation vector through non-linear transformation, realizing information compression and key feature extraction while retaining the structural association between nodes, providing an accurate feature basis for subsequent clustering. After obtaining the representation vector set, use the PCA and t-SNE algorithms for dimensionality reduction. PCA retains the main variance through linear dimensionality reduction, reduces the amount of calculation, and avoids the "curse of dimensionality" during clustering; t-SNE performs non-linear dimensionality reduction and maps the vectors into 2D / 3D space to facilitate visual verification of the clustering effect, assist in adjusting the clustering parameters, and ensure reasonable community division. Then perform recursive clustering on the output vector set after dimensionality reduction to construct the hypergraph structure. Recursive clustering aggregates nodes into communities (supernodes) step by step through multi-layer aggregation, and aggregates the edges between communities into hyperedges to achieve multi-scale abstraction. This process not only compresses the visualization scale, reduces the number of node edges, and alleviates visual congestion, but also directly shows the community distribution, core communities, and key connections, highlighting the core structure of the social network, conforming to its natural hierarchical features, and meeting the user's cognitive needs from macro to micro. Finally, use diverse visualization algorithms to visualize the hypergraph structure. For the hierarchical structure of the hypergraph, use a hierarchical layout or a force-directed layout algorithm to improve the aesthetics; combine front-end technologies to achieve interactive visualization and support users to dynamically filter information. These measures effectively balance the structural accuracy and aesthetic readability, and solve the problem of one-sidedness in traditional methods. Brief Description of the Drawings

[0006] Figure 1 It is a schematic flowchart of a visualization method for large-scale complex networks for community structure discovery in an embodiment; Figure 2 It is a diagram of the data structure relationship of clustered nodes in an embodiment; Figure 3 It is a schematic diagram of a hypergraph in an embodiment; Figure 4 It is a schematic diagram of a two-layer hypergraph in another embodiment; Figure 5 It is a schematic diagram of hypergraph construction in an embodiment; Figure 6Schematic diagram of force-directed layout in an embodiment; Figure 7 Schematic diagram of circular layout in an embodiment; Figure 8 Architectural diagram implemented by a web platform in an embodiment. Detailed implementation manners

[0007] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0008] In one embodiment, as Figure 1 shown, a large-scale complex network visualization method for community structure discovery is provided, including the following steps: Step 102, obtain a large-scale complex network; the large-scale complex network is a social network; the social network includes nodes and edges; the nodes represent users; the edges represent the relationships between users.

[0009] Step 104, use representation learning technology to extract the topological structure features of the large-scale complex network to obtain a node representation vector set; perform dimensionality reduction processing on the node representation vector set using PCA and t-SNE algorithms to obtain an output vector set.

[0010] In today's complex network analysis field, the discovery of community structure is a key task for understanding the network topology and function. As an important branch of it, large-scale network visualization intuitively displays the network structure through a graphical interface and reveals community characteristics.

[0011] Since large-scale networks usually contain tens of thousands of nodes and edges, directly visualizing these networks will lead to visual chaos and it is difficult to identify the community structure. In this application, through representation learning, first, by learning the low-dimensional representation of nodes, the complex edge set network is represented as 128-dimensional feature vectors; secondly, these representations can be further used for dimensionality reduction techniques, such as multi-dimensional scaling PCA or t-SNE, to achieve network visualization. In this way, even in the visual space, the important relationships between nodes and the community structure can be maintained, making community discovery intuitive and easy to interpret. This learned node representation can be used for various network analysis tasks, such as node classification, link prediction, and community detection, etc.

[0012] Step 106, perform recursive clustering on the output vector set to divide each layer in the hypergraph. Each community formed after clustering in each layer is constructed as a super node, and the connections between nodes within the community are merged into hyperedges to construct a hypergraph structure; use diverse visualization algorithms to visualize the hypergraph structure.

[0013] The recursive clustering to construct a hypergraph structure is the core innovation of this application. Traditional clustering algorithms mostly divide communities based on node degrees or edge weights, making it difficult to cope with the dynamic complexity of large-scale networks. This solution adopts a recursive strategy. Through algorithms such as hierarchical clustering, starting from the underlying nodes, it gradually aggregates to form multi-level structures such as communities and super-communities. The communities generated by each layer of clustering are abstracted as super-nodes, and the internal connections are merged into hyper-edges, realizing the "hierarchical compression" of data. Tens of thousands of user nodes are compressed into dozens of super-nodes, effectively solving the problem of visual congestion while retaining the connection relationships between communities (such as the hyper-edge weights reflecting the intensity of community interactions). The hypergraph structure can also clearly display key structures such as the core social circles and influence propagation paths, meeting the dual needs of users for macro and micro analysis of the network.

[0014] In the visualization section, aiming at the characteristics of the hypergraph structure, diverse algorithms are adopted to improve the display effect. In terms of layout algorithms, hierarchical layouts (such as DAG layout) can intuitively present the hierarchical relationships, and force-directed layouts (such as ForceAtlas2) can make the nodes in the same community naturally gather and different communities separate, enhancing the visual hierarchy. In terms of visual encoding, the size of the super-nodes represents the community scale, the color of the hyper-edges distinguishes the relationship types, and the edge width reflects the interaction frequency, etc., converting the abstract data into intuitive visual signals. Combining with front-end technologies such as Vue, interactive visualization is realized. Users can freely zoom in and out, expand the hierarchy, query the details of the communities, and dynamically filter information, further reducing the cognitive load.

[0015] The above-mentioned large-scale complex network visualization method for community structure discovery first defines the social network as a graph structure of "nodes (users) + edges (relationships)", and uses representation learning technology to extract the topological structure features of the complex network. Since the original social network data has high dimensions and a lot of noise, direct processing is inefficient. Representation learning can map the high-dimensional topological structure into a low-dimensional representation vector through non-linear transformation, realizing information compression and key feature extraction while retaining the structural associations between nodes, providing an accurate feature basis for subsequent clustering. After obtaining the set of representation vectors, PCA and t-SNE algorithms are used for dimensionality reduction. PCA retains the main variance through linear dimensionality reduction, reduces the amount of calculation, and avoids the "curse of dimensionality" during clustering; t-SNE performs non-linear dimensionality reduction, mapping the vectors into a 2D / 3D space to facilitate visual verification of the clustering effect, assisting in adjusting the clustering parameters, and ensuring reasonable community division. Then, the output vector set after dimensionality reduction is recursively clustered to construct a hypergraph structure. Recursive clustering aggregates nodes into communities (supernodes) step by step through multi-layer aggregation, and the edges between communities are aggregated into hyperedges, realizing multi-scale abstraction. This process not only compresses the visualization scale, reduces the number of node edges, and alleviates visual congestion, but also directly displays the community distribution, core communities, and key connections, highlighting the core structure of the social network, conforming to its natural hierarchical features, and meeting the user's cognitive needs from macro to micro. Finally, diverse visualization algorithms are used to visualize the hypergraph structure. For the hypergraph hierarchical structure, hierarchical layout or force-directed layout algorithms are adopted to improve the aesthetics; combined with front-end technology, interactive visualization is realized to support users to dynamically filter information. These measures effectively balance the structural accuracy and aesthetic readability, solving the problem of one-sidedness in traditional methods.

[0016] In one embodiment, the output vector set is recursively clustered to divide each level in the hypergraph. The communities formed after each level of clustering are constructed as supernodes, and the connections between the internal nodes in the community are merged into hyperedges to construct the hypergraph structure, including: The output vector set is recursively clustered to generate initial node structure relationship data, including num, nodes, and connect; where num represents the number of nodes in the next layer inside the supernode, nodes represents the numbers of the corresponding internal nodes, and connect represents the connection relationships between the internal nodes between the supernodes, which is used for hyperedge normalization generation.

[0017] In a specific embodiment, the clustering node data structure relationship is as Figure 2As shown, based on these clustering results, further node merging is performed, treating each cluster as a supernode. By converting multiple original nodes or edges into a single supernode, their common features or functions can be better represented. Subsequently, hyperedges are defined to connect these supernodes, thereby capturing the relationships between the entire set of nodes, revealing higher-order connection information, and then forming an initial top-level hypergraph, as Figure 3 shown.

[0018] In the figure, the highlighted part represents the supernode. After contact, the number of clusters subordinate to this node can be seen, showing the relevant information of the next-level hypergraph. The second-layer clustering situation is based on the clustering supernodes of the first layer. Secondary clustering is performed on each supernode of the first layer respectively, and then the construction of the second-layer hypergraph is formed, as Figure 4 . In the process of constructing the hypergraph, there will be a situation where the number of supernodes is too small to effectively display information. Here, the present application sets a threshold to limit the generation of hypergraph levels, so that the constructed hypergraph can better display the structural relationship information on the basis of simplicity. The schematic diagram of hypergraph construction is as Figure 5 shown, Figure 5 where k-means plus represents the improved k-means algorithm, agglomeration represents the agglomeration algorithm, cure represents the Cure algorithm, dbscan represents the DBSCAN algorithm, deepwalk represents the deep walk algorithm, node2vec represents the node vectorization algorithm, line represents the large-scale information network embedding algorithm, sdne represents the structured deep network embedding algorithm, and grarep represents the graph representation learning algorithm. By constructing the hypergraph structure, the present application can simplify the complexity of the original graph and at the same time display the internal connections of the data at a higher level of abstraction. Hypergraphs can more effectively capture and express high-dimensional data relationships, especially suitable for large-scale and highly complex data sets. This structure helps to discover hidden patterns and group dynamics that are difficult to identify in conventional graphs, providing a deeper perspective for data analysis and decision-making, and helping users better understand and utilize complex network data.

[0019] In one embodiment, each cluster is regarded as a supernode. By converting multiple original nodes into a single supernode and converting the original edges into hyperedges, an initial top-level hypergraph is formed based on the hyperedges connecting the supernodes; The second-layer clustering situation is based on the clustering supernodes of the initial top level. Secondary clustering is performed on each supernode of the initial top level respectively to form the second-layer hypergraph.

[0020] In one embodiment, representation learning technology is used to extract the topological structure features of a large-scale complex network, obtaining a set of representation vectors, including: Calculating the first similarity of nodes and the second similarity between nodes in the large-scale complex network; Using negative sampling technology, iteratively optimize the objective function of the first-order similarity and the objective function of the second-order similarity to obtain the embedding vector of each node, and construct a representation vector set from the embedding vectors of all nodes.

[0021] In one embodiment, calculating the first similarity of nodes in a large-scale complex network and the second similarity between nodes includes: Calculating the first similarity of nodes in a large-scale complex network as: ; where, is the low-dimensional vector representation of vertex , is the low-dimensional vector representation of vertex , , representing different vertices; Calculating the second similarity of nodes in a large-scale complex network as: ; where T represents the transpose operation, represents the total number of nodes in the node set V of the network, is the low-dimensional vector representation of vertex .

[0022] In one embodiment, the objective function of the first-order similarity is: ; where, represents the weight of the edge between node i and node j, and E represents the set of edges in the graph.

[0023] In one embodiment, the objective function of the second-order similarity is: ; where, represents the weight of the edge between node i and node j, and E represents the set of edges in the graph.

[0024] In one embodiment, the diversification visualization algorithm includes a force-directed layout algorithm and a circular layout algorithm.

[0025] In a specific embodiment, visualization is the process of converting data into visual graphics, which presents data in a graphical way to make it more intuitive and easier to understand. In the visualization process, the layout algorithm plays a very important role in beautifying and optimizing the layout effect of the graphics. In this implementation, the force-directed layout algorithm and the circular layout algorithm are used to beautify the layout of hypergraph visualization. The force-directed layout algorithm is a commonly used graph layout algorithm. Its core idea is to simulate the action of physical forces, enabling the nodes in the graph to be automatically arranged on a two-dimensional plane and maintaining an appropriate distance between the nodes, thereby achieving an aesthetic and easy-to-understand layout effect. The visualization page implemented based on the force-directed layout in this application is as shown in Figure 6 . In the force-directed layout algorithm, the node positions in the initial layout are randomly distributed. During the continuous iteration process, the node coordinates are updated at a certain speed, and under the continuous action of gravitational and repulsive forces, the final layout effect is formed. In the research on the force-directed layout algorithm, it is found that in the case of fewer nodes, a good layout effect has been achieved before the iteration stops; in the case of more nodes, a better layout effect has not been achieved when the iteration stops. Therefore, if the iteration times can be dynamically adjusted, it can optimize the layout effect. Based on following the aesthetic standards of the point-placement algorithm, this application uses the node deviation as the evaluation parameter for the layout effect, forms a relationship mapping between the node deviation and the iteration times, and integrates it into the force-directed layout algorithm. During the layout process, the node deviation value is used as the annealing parameter to determine the progress of the iteration. The optimization of the algorithm is based on the WSPD (well-separated pair decomposition) method, which calculates the force between the centroid of the subset after decomposition and other nodes instead of calculating the force between each pair of nodes, and can reduce the time complexity of the force-directed algorithm to O( ).

[0026] The circular layout is a common layout method, which is often used to highlight the cyclic structure in the network or display groups of nodes with similar characteristics. Through this layout method, users can quickly obtain a centralized graphical view, which helps to more intuitively understand the structure of the network and the relationships between nodes. When performing a circular layout, it is important to ensure that the positions of the nodes on the circumference can accurately reflect their properties and relationships. By reasonably arranging the positions and attribute values of the nodes, this application can effectively display the structure and organization of the network and help better understand the information in the network. In addition, the circular layout can also help this application more intuitively discover possible cyclic structures or groups of nodes with similar characteristics in the network, thereby enhancing the understanding and cognition of the network. The visualization page implemented based on the circular layout in this application is shown in Figure 7.

[0027] In one of the embodiments, based on Vue and Spring Boot, an integrated large-scale complex network visualization Web platform is built by running a Python script to visually display the network files uploaded by users.

[0028] In a specific embodiment, in the backend architecture of the platform, the Spring Boot framework and Maven build a robust service environment specifically designed to receive complex network data files uploaded by users. By cleverly integrating Python scripts, the system can parse this data, paving the way for subsequent advanced analysis. Using multiple network representation learning algorithms such as DeepWalk, Node2Vec, LINE, GraRep, and SDNE, the structural information of the network is transformed into high-dimensional feature vectors, which are then processed by dimensionality reduction methods such as PCA and t-SNE to adapt the data for clustering analysis. In the community detection step, with the help of advanced clustering algorithms such as K-Means++, Agglomerative Clustering, DBSCAN, CURE, and SOMSC, the hidden community structure in the network is effectively revealed. The outputs of these algorithms are then organized into a hypergraph structure and serialized into a JSON file for delivery to the front end for display. The front end uses modern technology stacks such as Vue3, Vite, TypeScript, and ElementPlus to build a beautiful and highly interactive user interface. Users can easily upload data files and select different clustering algorithms and view layouts through intuitive interface operations. The ECharts chart library dynamically loads the hypergraph JSON generated by the backend and supports diverse layout algorithms such as force-directed layout and circular layout, thus presenting the complex hypergraph structure to users in a clear and aesthetic way. The front-end application also integrates functions such as search, filtering, zooming, and dragging, greatly enhancing the user's exploration experience and the flexibility of data analysis. The overall architecture of the front and back ends is as Figure 8As shown, where "web" represents a web page, Vue3 is a popular JavaScript front-end framework for building user interfaces. Element - plus UI is a set of desktop component libraries based on Vue3. Echarts is a JavaScript-based data visualization library that can be used to generate various charts. Axios is a Promise-based HTTP client. POST request is an HTTP request method, and Get request is also an HTTP request method. SpringBoot is a rapid development framework in the Java field for building enterprise-level applications. Maven is a build tool and project management tool for Java projects, used to manage project dependencies, compile code, package projects, etc., and standardize the project build process. Java ProcessBuilder is a class in Java for creating and starting processes, which can be used to execute external programs, scripts, etc. In the backend, it can be used to call other programs to complete tasks. During user usage, the user uploads network data files through an intuitive web interface built with the Vue3 technology stack, ensuring high interactivity and responsiveness. The backend server, based on the SpringBoot framework and managed by Maven, securely receives these uploads. Subsequently, with the help of the Java ProcessBuilder command-line processing library, it calls a Python script for initial data cleaning and format standardization. The command-line format example is as follows: python generate.py --openne deepwalk --cluster kmeans. The options of the command-line parameters are shown in Table 1.

[0029]

[0030] The preliminarily processed data is fed into the feature learning process. Using algorithms such as DeepWalk and Node2Vec, complex network nodes are mapped into low-dimensional vectors, extracting the most crucial information of the data to reveal the hidden connections and network structure between nodes. Then, various clustering methods are adopted to further process the node data. Based on the clustering results, the system constructs a hypergraph model, with hyperedges representing the associations between node groups, more efficiently abstracting the network structure. Subsequently, this structure is encoded into JSON format for seamless transfer between the front and backend. The front end uses a high-performance chart library like ECharts to parse and convert the received JSON data into vivid network graphs, which are displayed through layout methods such as force-directed layout. The combination of Vue3 and ElementPlus enables a highly dynamic and interactive interface where users can easily zoom, drag the view, and even click to view specific nodes.

[0031] It should be understood that although Figure 1The steps in the flowchart are shown in sequence according to the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this application, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in Figure 1 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0032] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0033] The above-described embodiments merely represent several implementation manners of this application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application should be subject to the appended claims.

Claims

1. A large-scale complex network visualization method for community structure discovery, characterized in that, The method includes: Obtain a large-scale complex network; the large-scale complex network is a social network; the social network includes nodes and edges; the nodes represent users; the edges represent the relationships between users. Use representation learning technology to extract the topological structure features of the large-scale complex network to obtain a set of node representation vectors; perform dimensionality reduction processing on the set of node representation vectors using the PCA and t-SNE algorithms to obtain an output vector set. Perform recursive clustering on the output vector set to divide each level in the hypergraph. The communities formed after each level of clustering are constructed as supernodes, and the connections between the nodes within the communities are merged into hyperedges to construct a hypergraph structure; use a diverse visualization algorithm to visualize the hypergraph structure.

2. The method according to claim 1, characterized in that, Performing recursive clustering on the output vector set to divide each level in the hypergraph. The communities formed after each level of clustering are constructed as supernodes, and the connections between the nodes within the communities are merged into hyperedges to construct a hypergraph structure, including: Perform recursive clustering on the output vector set to generate initial node structure relationship data, including num, nodes, and connect; where num represents the number of nodes in the next level within the supernode, nodes represents the numbers of the corresponding internal nodes, and connect represents the connection relationships between the internal nodes of the supernodes, which is used for hyperedge normalization generation.

3. The method according to claim 2, wherein The method includes: Regard each clustering as a supernode. By converting the original multiple nodes into one supernode and the original edges into hyperedges, form an initial top-level hypergraph based on the hyperedges connecting the supernodes. Based on the clustering supernodes at the initial top level, perform secondary clustering on each supernode at the initial top level respectively to form a second-level hypergraph.

4. The method according to claim 1, wherein Use representation learning technology to extract the topological structure features of the large-scale complex network to obtain a set of representation vectors, including: Calculate the first similarity of the nodes in the large-scale complex network and the second similarity between the nodes. Use negative sampling technology to iteratively optimize the objective function of the first-order similarity and the objective function of the second-order similarity to obtain the embedding vector of each node, and the embedding vectors of all nodes construct a set of representation vectors.

5. The method according to claim 4, wherein Calculating the first similarity of the nodes in the large-scale complex network and the second similarity between the nodes, including: The calculation of the first similarity of the nodes in the large-scale complex network is: ； Among them, is the low-dimensional vector representation of vertex , is the low-dimensional vector representation of vertex , , represent different vertices; The calculation of the second similarity of the nodes in the large-scale complex network is: ； where T represents the transpose operation, represents the total number of nodes in the node set V of the network, represents a vertex with its low-dimensional vector representation.

6. The method according to claim 5, wherein The objective function of the first-order similarity is: ； Among them, represents the weight of the edge between node i and node j, and E represents the set of edges in the graph.

7. The method according to claim 5, characterized in that, The objective function of the second-order similarity is: ； Among them, represents the weight of the edge between node i and node j, and E represents the set of edges in the graph.

8. The method according to claim 1, characterized in that, The diverse visualization algorithm includes a force-directed layout algorithm and a circular layout algorithm.

9. The method according to claim 1, characterized in that, The method further includes: Based on Vue and Spring Boot, construct an integrated large-scale complex network visualization Web platform by running a Python script to visually display the network files uploaded by users.

Images