Feature encoding method, device and computer readable storage medium

Abstract

The embodiment of the present application discloses a feature encoding method, device and computer-readable storage medium, which can be applied to various scenarios such as cloud technology, artificial intelligence, smart transportation, and assisted driving; by obtaining a weighted node relationship graph; identifying the connection attribute information of the node in the weighted node relationship graph, and determining the segmentation parameters of the weighted node relationship graph based on the connection attribute information; segmenting the weighted node relationship graph according to the connection weight and the segmentation parameter to obtain at least one unweighted node relationship subgraph; based on the segmentation parameter, screening out at least one target node edge corresponding to each node in the unweighted node relationship subgraph, and counting the edge distribution information of the node based on the target node edge; based on the edge distribution information, feature encoding the corresponding node in the weighted node relationship graph to obtain the node feature of the node. In this way, the accuracy of node feature encoding in the relationship graph is improved, and the encoding efficiency of node features is thereby improved.

Description

Technical Field

[0001] The present application relates to the field of Internet technology, and in particular to a feature encoding method, device and computer-readable storage medium. Background Art

[0002] In recent years, with the rapid development of Internet technology, graph computing technology has also been developing continuously. Among them, the graph node representation learning algorithm is an algorithm in graph computing technology that learns the low-dimensional dense representation of the nodes in the graph from the topological structure of the relationship graph and encodes the information of the nodes in the graph in a reduced dimension. In the existing feature encoding method, the similarity between nodes is generally calculated by using the existing graph node representation learning algorithm, and whether the nodes are connected is determined based on the similarity, so as to encode the node features according to the node connection results, so as to apply the encoded node features to downstream tasks such as user diversification management and risk warning.

[0003] In the process of research and practice of existing technologies, it was found that in the existing feature encoding method, determining the node connection relationship by calculating the similarity between nodes will result in connection similarity between each node and other nodes in the relationship graph, resulting in the node features encoded based on the feature encoding method being unable to accurately restore the node connection relationship of the nodes in the relationship graph, making the feature encoding accuracy of the nodes in the relationship graph low, which in turn leads to low efficiency of feature encoding. Summary of the invention

[0004] The embodiments of the present application provide a feature encoding method, device and computer-readable storage medium, which can improve the accuracy of node feature encoding in a relationship graph, thereby improving the encoding efficiency of node features.

[0005] The present application provides a feature encoding method, including:

[0006] Obtaining a weighted node relationship graph corresponding to a content recommendation application, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges, wherein objects in the content recommendation application correspond one-to-one to the nodes in the weighted node relationship graph, and the node edges and the connection weights corresponding to the node edges are determined according to association information of the objects in the content recommendation application, wherein the objects are users in the content recommendation application, and the association information includes interaction relationship information and / or interest similarity information between the users;

[0007] Identifying connection attribute information of the node in the weighted node relationship graph, and determining a segmentation parameter of the weighted node relationship graph based on the connection attribute information, wherein the segmentation parameter is a relevant parameter for segmenting the weighted node relationship graph;

[0008] According to the connection weights and the segmentation parameters, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, where the unweighted node relationship subgraph is a node relationship subgraph without connection weights;

[0009] Based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and edge distribution information of the node is counted based on the target node edge;

[0010] According to the edge distribution information, feature encoding is performed on the corresponding node in the weighted node relationship graph to obtain the node feature of the node;

[0011] Content recommendation is performed for each object in the content recommendation application based on the node features.

[0012] Accordingly, an embodiment of the present application provides a feature encoding device, including:

[0013] an acquisition unit, configured to acquire a weighted node relationship graph corresponding to a content recommendation application, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges, wherein objects in the content recommendation application correspond one-to-one to the nodes in the weighted node relationship graph, and the node edges and the connection weights corresponding to the node edges are determined according to association information of the objects in the content recommendation application, wherein the objects are users in the content recommendation application, and the association information includes information on interaction relationships between the users and / or information on the similarity of interests;

[0014] A determination unit, configured to identify connection attribute information of the node in the weighted node relationship graph, and determine a segmentation parameter of the weighted node relationship graph based on the connection attribute information, wherein the segmentation parameter is a relevant parameter for segmenting the weighted node relationship graph;

[0015] A segmentation unit, configured to segment the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph, wherein the unweighted node relationship subgraph is a node relationship subgraph without connection weights;

[0016] A screening unit, configured to screen out at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameter, and to calculate edge distribution information of the node based on the target node edge;

[0017] An encoding unit, configured to perform feature encoding on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes;

[0018] The recommendation unit is used to make content recommendations for each object in the content recommendation application based on the node features.

[0019] In one embodiment, the segmentation unit comprises:

[0020] A relationship graph type determination subunit, used to determine the relationship graph type of the weighted node relationship graph according to the connection weights corresponding to the node edges;

[0021] A multiple splitting sub-unit is used for splitting the multiple weighted node relationship graph to obtain at least two single weighted node relationship graphs when the relationship graph type is a multiple weighted node relationship graph, and splitting the single weighted node relationship graph according to the connection weight and the splitting parameter to obtain at least one unweighted node relationship sub-graph, wherein each node edge in the single weighted node relationship graph corresponds to a connection weight;

[0022] The single segmentation subunit is used to segment the single weighted node relationship graph according to the connection weight and segmentation parameter to obtain at least one unweighted node relationship subgraph when the relationship graph type is a single weighted node relationship graph.

[0023] In one embodiment, the relationship graph type determination subunit includes:

[0024] A weight quantity identification module, used to identify the quantity of connection weights corresponding to each node edge in the weighted node relationship graph;

[0025] A multiple weighted node relationship graph determination module, configured to determine the relationship graph type of the weighted node relationship graph as a multiple weighted node relationship graph when there are node edges corresponding to at least two connection weights in the node edges;

[0026] The single weighted node relationship graph determination module is used to determine the relationship graph type of the weighted node relationship graph as a single weighted node relationship graph when there is no node edge corresponding to at least two connection weights in the node edges.

[0027] In one embodiment, the multiple segmentation subunit and the single segmentation subunit include:

[0028] A pruning times determination module, used to determine the pruning times of the single weighted node relationship graph according to the segmentation parameters, and obtain a pruning threshold group corresponding to the single weighted node relationship graph, wherein the pruning threshold group includes at least one pruning threshold;

[0029] A segmentation module is used to segment the single weighted node relationship graph based on the pruning times, connection weights and pruning threshold group to obtain at least one unweighted node relationship subgraph.

[0030] In one embodiment, the segmentation module is used to:

[0031] A first pruning submodule is used for pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group when the pruning number is once, to obtain an unweighted node relationship subgraph corresponding to each pruning threshold;

[0032] The second pruning submodule is used to prune the single weighted node relationship graph based on the connection weight and the pruning threshold group when the pruning times are multiple times, to obtain an unweighted node relationship subgraph after pruning, and to use the pruned unweighted node relationship subgraph as the single weighted node relationship graph, and return to execute the step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group until the pruning times are reached to obtain at least one unweighted node relationship subgraph.

[0033] In one embodiment, the first pruning submodule is used to:

[0034] Comparing at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph;

[0035] Based on the comparison results corresponding to the node edges, pruning each node edge in the single weighted node relationship graph to obtain a pruned single weighted node relationship graph;

[0036] The connection weight corresponding to each node edge in the pruned single weighted node relationship graph is deleted to obtain an unweighted node relationship subgraph corresponding to each pruning threshold.

[0037] In one embodiment, the first pruning submodule may be specifically used for:

[0038] When the connection weight corresponding to the node edge is less than the pruning threshold, deleting the node edge in the single weighted node relationship graph to obtain a deleted node edge;

[0039] When the connection weight corresponding to the node edge is not less than the pruning threshold, the node edge is retained in the single weighted node relationship graph to obtain a retained node edge;

[0040] Based on the deleted node edges and the retained node edges, a pruned single-weighted node relationship graph corresponding to each pruning threshold is determined.

[0041] In one embodiment, the encoding unit includes:

[0042] A first feature encoding subunit is used for, when the weighted node relationship graph is a single weighted node relationship graph, to perform feature encoding on a corresponding node in the weighted node relationship graph according to the edge distribution information to obtain a node feature of the node;

[0043] The second feature encoding subunit is used for, when the weighted node relationship graph is a multiple weighted node relationship graph, to perform feature encoding on the corresponding nodes in at least two single weighted node relationship graphs of the weighted node relationship graph according to the edge distribution information, to obtain the single node features of the nodes in the corresponding single weighted node relationship graphs, and to concatenate the single node features to obtain the corresponding node features of the nodes in the weighted node relationship graphs.

[0044] In one embodiment, the first feature encoding subunit includes:

[0045] An edge distribution feature extraction module is used to extract edge distribution features corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information;

[0046] A combination module, used for combining the edge distribution features corresponding to the nodes to obtain the target edge distribution features of the nodes;

[0047] A node feature determination module is used to determine the node feature of the node according to the target edge distribution feature.

[0048] In one embodiment, the edge distribution feature extraction module includes:

[0049] An extraction submodule, configured to extract the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information;

[0050] The conversion submodule is used to use the edge information as an edge feature basis, and convert the distribution information into an edge distribution feature based on the edge feature basis.

[0051] In one embodiment, the screening unit comprises:

[0052] A node edge order determination subunit, used to determine the node edge order corresponding to each node based on the segmentation parameter;

[0053] A traversal subunit, used to traverse each node in the unweighted node relationship subgraph to obtain a connection sequence between node edges corresponding to each node;

[0054] The splicing subunit is used to screen out the node edges to be spliced ​​corresponding to the node edge order in each node in the unweighted node relationship subgraph, and splice the node edges to be spliced ​​according to the connection order to obtain at least one target node edge corresponding to each node.

[0055] In one embodiment, the screening unit comprises:

[0056] A counting subunit, used for counting the number of target node edges corresponding to each node in the unweighted node relationship subgraph;

[0057] The edge distribution information determination subunit is used to determine the edge distribution information of the node in the unweighted node relationship subgraph based on the number of edges of the target node.

[0058] In addition, an embodiment of the present application further provides a computer-readable storage medium, which stores a plurality of instructions, and the instructions are suitable for a processor to load to execute the steps in any one of the feature encoding methods provided in the embodiment of the present application.

[0059] In addition, an embodiment of the present application also provides a computer device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to implement the feature encoding method provided in the embodiment of the present application.

[0060] The embodiment of the present application also provides a computer program product or a computer program, wherein the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the steps in the feature encoding method provided in the embodiment of the present application.

[0061] The embodiment of the present application obtains a weighted node relationship graph, which includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges; identifies connection attribute information of the nodes in the weighted node relationship graph, and determines segmentation parameters of the weighted node relationship graph based on the connection attribute information; segments the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, screens out at least one target node edge corresponding to each node in the unweighted node relationship subgraph, and counts the edge distribution information of the node based on the target node edge; based on the edge distribution information, feature encodes the corresponding nodes in the weighted node relationship graph to obtain node features of the nodes. In this way, by dividing the weighted node relationship graph into at least one unweighted node relationship subgraph, and counting the edge distribution information of the nodes in the unweighted node relationship subgraph, and then feature encoding the nodes in the weighted node relationship graph according to the edge distribution information corresponding to each unweighted node relationship subgraph, the node features of the nodes are obtained, and the node connection relationship between the node and other nodes in the weighted node relationship graph is integrated into the node features, thereby improving the accuracy of node feature encoding in the relationship graph, and then improving the encoding efficiency of node features. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0063] Figure 1 This is a schematic diagram of an implementation scenario of a feature encoding method provided in an embodiment of the present application;

[0064] Figure 2 It is a flowchart of a feature encoding method provided in an embodiment of the present application;

[0065] Figure 3 It is a schematic diagram of a node relationship diagram of a feature encoding method provided in an embodiment of the present application;

[0066] Figure 4a It is a schematic diagram of a feature encoding process of a multiple weighted node relationship graph of a feature encoding method provided in an embodiment of the present application;

[0067] Figure 4b It is a single weighted node relationship of a feature encoding method provided in an embodiment of the present application Figure 1 Schematic diagram of the secondary pruning process;

[0068] Figure 4cIt is a single weighted node relationship of a feature encoding method provided in an embodiment of the present application Figure 2 Schematic diagram of the secondary pruning process;

[0069] Figure 5 is another flowchart of a feature encoding method provided in an embodiment of the present application;

[0070] Figure 6 is a schematic diagram of the structure of a feature encoding device provided in an embodiment of the present application;

[0071] Figure 7 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0073] The embodiment of the present application provides a feature coding method, device and computer readable storage medium. The feature coding device can be integrated in a computer device, which can be a server or a terminal.

[0074] Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, network acceleration services (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms. Terminals may include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. Terminals and servers can be directly or indirectly connected via wired or wireless communications, and this application does not limit this.

[0075] See also Figure 1 , taking the feature encoding device integrated into a computer device as an example, Figure 1A schematic diagram of an implementation scenario of the feature encoding method provided in an embodiment of the present application, wherein the computer device can be a server or a terminal, and the computer device can obtain a weighted node relationship graph, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and a connection weight corresponding to the node edge; the connection attribute information of the node is identified in the weighted node relationship graph, and based on the connection attribute information, the segmentation parameters of the weighted node relationship graph are determined; according to the connection weight and the segmentation parameter, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, and the unweighted node relationship subgraph is a node relationship subgraph without connection weight; based on the segmentation parameter, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and the edge distribution information of the node is counted based on the target node edge; according to the edge distribution information, the corresponding node in the weighted node relationship graph is feature encoded to obtain the node feature of the node.

[0076] It should be noted that the embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, assisted driving, etc. Figure 1 The implementation environment scenario diagram of the feature coding method shown is only an example. The implementation environment scenario of the feature coding method described in the embodiment of the present application is to more clearly illustrate the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided by the embodiment of the present application. It is known to those skilled in the art that with the evolution of feature coding and the emergence of new business scenarios, the technical solution provided by the present application is also applicable to similar technical problems.

[0077] The solutions provided in the embodiments of the present application involve technologies such as natural language processing of artificial intelligence, which are specifically described by the following embodiments. It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments.

[0078] This embodiment will be described from the perspective of a feature coding device. The feature coding device may be integrated into a computer device, which may be a server. This application does not limit this.

[0079] See also Figure 2 , Figure 2 : is a flow chart of a feature encoding method provided in an embodiment of the present application. The feature encoding method includes:

[0080] In step 101, a weighted node relationship graph is obtained.

[0081] The weighted node relationship graph may be a node relationship graph including connection weights, the node relationship graph may be a network graph including nodes and node edges connecting the nodes, which may be used to characterize the connection relationship between the nodes, and may be an undirected graph, the weighted node relationship graph may include at least two nodes, at least one node edge connecting the nodes, and the connection weights corresponding to the node edges. The connection weight may be used to characterize the degree of association between the nodes, for example, it may be used to characterize the meaning of the distance, cost, time, probability of event occurrence, etc. corresponding to the node edge between the two nodes. The probability of the event occurrence may be forwarding probability, click probability, risk probability, etc.

[0082] In the prior art, graph walk algorithms (Node2Vec, DeepWalk), graph neural networks (LINE) and other graph node representation learning algorithms are often used to encode node features in a relationship graph. This method often focuses on the similarity of neighboring nodes, calculates the similarity between nodes, and determines whether the nodes are connected according to the similarity, thereby encoding the node features according to the node connection results. However, determining the node connection relationship by calculating the similarity between nodes will result in a connection similarity between each node and other nodes in the relationship graph, that is, there is a connection relationship between each node, which causes the node features encoded based on the feature encoding method to be unable to accurately restore the node connection relationship of the node in the relationship graph, making the feature encoding accuracy of the node in the relationship graph low, thereby resulting in low feature encoding efficiency. To this end, an embodiment of the present application provides a feature encoding method, by vectorizing the local connection edge structure distribution of each node in a weighted node relationship graph to restore the actual connection features of the node in the node relationship graph as much as possible, so that the node connection relationship between the node and other nodes in the weighted node relationship graph can be integrated into the node feature, thereby improving the accuracy of node feature encoding in the relationship graph, thereby improving the encoding efficiency of node features. The feature encoding method provided by this application is described in detail below.

[0083] In step 102, connection attribute information of nodes in the weighted node relationship graph is identified, and segmentation parameters of the weighted node relationship graph are determined based on the connection attribute information.

[0084] Among them, the connection attribute information can be attribute information that characterizes the connection relationship between each node in the weighted node relationship graph and other nodes, the attribute information can be information including the properties of the node in the weighted node relationship graph and the connection relationship between other nodes, and the segmentation parameter can be relevant parameters for segmenting the weighted node relationship graph, for example, it can include parameters such as the number of segmentations and the segmentation method.

[0085] Among them, there can be multiple ways to identify the connection attribute information of nodes in the weighted node relationship graph. For example, the nodes and node edges in the weighted node relationship graph can be traversed, and the connection attribute information of each node in the weighted node relationship graph can be counted based on the traversal results.

[0086] After identifying the connection attribute information of the nodes in the weighted node relationship graph, the segmentation parameters of the weighted node relationship graph can be determined based on the connection attribute information. There are many ways to determine the segmentation parameters of the weighted node relationship graph based on the connection attribute information. For example, the node connection status of each node in the weighted node relationship graph can be determined based on the connection attribute information corresponding to the node, and the longest node edge of each node in the weighted node relationship graph can be obtained based on the node connection status of each node. Therefore, the segmentation parameters of the weighted node relationship graph can be determined based on the longest node edge of each node in the weighted node relationship graph. The longest node edge can be the longest path traversing the weighted node relationship graph with a certain node as the root node. For example, please refer to Figure 3 , Figure 3 It is a schematic diagram of a node relationship diagram of a feature encoding method provided by an embodiment of the present application, wherein the node relationship diagram includes node 1, node 2, node 3, node 4 and node 5. Taking the node relationship diagram as a weighted node relationship diagram as an example, assuming that the node 1 is the root node and the node relationship diagram is traversed, the longest path of node 1 can be obtained as two node edges, which can be two node edges from node 1 to node 4 or from node 1 to node 5, namely node edges ab and ad. Assuming that the node 3 is the root node and the node relationship diagram is traversed, the longest path of node 3 can be obtained as three node edges, which can be three node edges from node 3 to node 4 or from node 3 to node 5, namely node edges cab and cad. It can also be assumed that the node 2 is the root node and the node relationship diagram is traversed, then the longest path of node 2 can be obtained as two node edges, which can be two node edges from node 2 to node 3, namely ac, etc.

[0087] There are many ways to determine the segmentation parameters of the weighted node relationship graph based on the longest node edge of each node in the weighted node relationship graph. For example, the longest node edge of each node can be counted, and the longest node edge with the most occurrences can be determined based on the statistical results. Then, the number of node edges corresponding to the longest node edge with the most occurrences can be determined as the segmentation parameter. For example, please continue to refer to Figure 3 ,exist Figure 3In the weighted node relationship graph shown, the longest node edge of node 1 is two node edges, the longest node edge of node 2 is two node edges, the longest node edge of node 3 is three node edges, the longest node edge of node 4 is three node edges, and the longest node edge of node 5 is three node edges, so that the longest node edges of the five nodes can be counted, and according to the statistical results, it can be determined that the longest node edge with the most occurrences is 3 node edges. Therefore, the number of splits can be determined as 3 according to the longest node edge with the most occurrences, thereby determining the splitting parameters of the weighted node relationship graph.

[0088] In one embodiment, the segmentation parameters of the weighted node relationship graph can be further determined according to the computing power and computing accuracy of the feature coding device. For example, taking the segmentation parameter as the number of segmentation times as an example, when the computing power of the feature coding device is strong, a larger number of segmentation times can be selected, and when the computing power of the feature coding device is weak, a smaller number of segmentation times can be selected. When the computing accuracy of the feature coding is high, a larger number of segmentation times can be selected to obtain more node information based on high-order features, and when the computing accuracy of the feature coding is low, a smaller number of segmentation times can be selected.

[0089] In step 103, the weighted node relationship graph is segmented according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph.

[0090] The unweighted node relationship subgraph may be a node relationship subgraph without connection weights, that is, there are connected and unconnected relationships between all nodes in the unweighted node relationship subgraph, and each node edge is a node edge without corresponding connection weights.

[0091] Among them, there are multiple ways to split the weighted node relationship graph according to the connection weight and splitting parameters. For example, the relationship graph type of the weighted node relationship graph can be determined according to the connection weight corresponding to the node edge. When the relationship graph type is a multiple weighted node relationship graph, the multiple weighted node relationship graph is split to obtain at least two single weighted node relationship graphs, and the single weighted node relationship graph is split according to the connection weight and splitting parameters to obtain at least one unweighted node relationship subgraph. When the relationship graph type is a single weighted node relationship graph, the single weighted node relationship graph is split according to the connection weight and splitting parameters to obtain at least one unweighted node relationship subgraph.

[0092] The relationship graph type may be a type of weighted node relationship graph, which may include a multiple weighted node relationship graph and a single weighted node relationship graph. The multiple weighted node relationship graph may be a node edge relationship graph in which there are at least two connection weights corresponding to the node edge, the single weighted node relationship graph may be a relationship graph in which each node edge corresponds to one connection weight, and each node edge in the single weighted node relationship graph may correspond to one connection weight.

[0093] Among them, there can be multiple ways to determine the relationship graph type of the weighted node relationship graph based on the connection weights corresponding to the node edges. For example, the number of connection weights corresponding to each node edge can be identified in the weighted node relationship graph. When there are node edges corresponding to at least two connection weights in the node edges, the relationship graph type of the weighted node relationship graph is determined to be a multiple weighted node relationship graph. When there are no node edges corresponding to at least two connection weights in the node edges, the relationship graph type of the weighted node relationship graph is determined to be a single weighted node relationship graph.

[0094] There are many ways to identify the number of connection weights corresponding to each node edge in the weighted node relationship graph. For example, the weighted node relationship graph can be traversed to obtain the number of connection weights corresponding to each node edge based on the traversal result.

[0095] After determining the relationship graph type of the weighted node relationship graph according to the connection weight corresponding to the node edge, the multiple weighted node relationship graph can be split to obtain at least two single weighted node relationship graphs when the relationship graph type is a multiple weighted node relationship graph. There are many ways to split the multiple weighted node relationship graph. For example, the multiple weighted node relationship graph can be split according to the behavior attribute information and connection weight of the node to obtain at least two single weighted node relationship graphs. The behavior attribute information can be the behavior of the node itself and the node attribute information, and the node attribute information can be information including the properties of the nodes in the weighted node relationship graph and the relationship between the nodes. Since there are node edges corresponding to multiple connection weights in the node edge in the multiple weighted node relationship graph, the actual meaning of different node edges representing the node relationship can be different, so as to accurately describe complex multiple relationship scenarios. For example, when performing payment behavior, two users can establish a payment relationship, and when chatting, two users can establish a chat relationship. Therefore, in order to accurately integrate the connection relationship between nodes, that is, the edge structure distribution information of the nodes into the feature encoding of the nodes, the node edges corresponding to at least two connection weights in the multiple weighted node relationship graph can be split according to the behavior attribute information of the nodes in the multiple weighted node relationship graph, so that each node edge in the multiple weighted node relationship graph corresponds to at most one connection weight. For example, please refer to Figure 4a , Figure 4aThe schematic diagram of the feature encoding process of a multiple weighted node relationship graph of a feature encoding method provided in an embodiment of the present application is as follows. The multiple weighted node relationship graph G includes nodes 1 to 9. The multiple weighted node relationship graph G can be split into a single weighted node relationship graph G according to the behavior attribute information of the nodes and the actual meaning of each node edge. (1) And the single weighted node relationship graph G (2) , single weighted node relationship graph G (1) And the single weighted node relationship graph G (2) Each node edge in corresponds to at most one connection weight.

[0096] For a single weighted node relationship graph, the single weighted node relationship graph can be segmented according to the connection weight and the segmentation parameter to obtain at least one unweighted node relationship subgraph, wherein there can be multiple ways to segment the single weighted node relationship graph according to the connection weight and the segmentation parameter. For example, the number of pruning times of the single weighted node relationship graph can be determined according to the segmentation parameter, and the pruning threshold group corresponding to the single weighted node relationship graph can be obtained. Based on the number of pruning times, the connection weight and the pruning threshold group, the single weighted node relationship graph can be segmented to obtain at least one unweighted node relationship subgraph.

[0097] Among them, the number of pruning times can be the number of times a single weighted node relationship graph is pruned, that is, the number of splits, and the pruning threshold group can be a threshold group for pruning a single weighted node relationship graph. The pruning threshold group can include at least one pruning threshold, and the pruning threshold can be a pre-set critical value. When the connection weight of the node edge between two nodes is greater than or equal to the critical value, it can be determined that a connection is established between the two nodes. When the connection weight of the node edge between the two nodes is less than the critical value, it can be determined that no connection is established between the two nodes.

[0098] Optionally, in the pruning threshold group, the value of each pruning threshold may be between 0 and 1. For example, it may be assumed that the pruning threshold group includes m pruning thresholds, namely T1 to Tm, where 0≤T1≤T2≤…≤Tm≤1.

[0099] There are many ways to determine the number of pruning times of the single weighted node relationship graph according to the segmentation parameter. For example, the number of segmentation times can be extracted from the segmentation parameter and used as the number of pruning times.

[0100] After the number of pruning times of the single weighted node relationship graph is determined according to the segmentation parameter, the single weighted node relationship graph can be segmented based on the number of pruning times, connection weights and pruning threshold groups to obtain at least one unweighted node relationship subgraph. There are multiple ways to segment the single weighted node relationship graph based on the number of pruning times, connection weights and pruning threshold groups. For example, when the number of pruning times is one, the single weighted node relationship graph can be pruned based on the connection weights and pruning threshold groups to obtain an unweighted node relationship subgraph corresponding to each pruning threshold; when the number of pruning times is multiple times, the single weighted node relationship graph can be pruned based on the connection weights and pruning threshold groups to obtain a pruned unweighted node relationship subgraph, and the pruned unweighted node relationship subgraph is used as a single weighted node relationship graph, and the step of pruning the single weighted node relationship graph based on the connection weights and pruning threshold groups is returned to execute until the number of pruning times is reached to obtain at least one unweighted node relationship subgraph.

[0101] Among them, the pruned unweighted node relationship subgraph can be a node relationship subgraph obtained by pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group, and the node relationship subgraph can be a node relationship graph obtained by segmenting the weighted node relationship graph.

[0102] Among them, based on the connection weight and the pruning threshold group, the single weighted node relationship graph is pruned, and there can be multiple ways to obtain the unweighted node relationship subgraph corresponding to each pruning threshold. For example, at least one pruning threshold in the pruning threshold group can be compared with the connection weight corresponding to each node edge in the single weighted node relationship graph, and based on the comparison result corresponding to the node edge, each node edge in the single weighted node relationship graph is pruned to obtain the pruned single weighted node relationship graph, and the connection weight corresponding to each node edge in the pruned single weighted node relationship graph is deleted to obtain the unweighted node relationship subgraph corresponding to each pruning threshold.

[0103] The pruned single weighted node relationship graph may be a node relationship graph obtained by pruning each node edge in the single weighted node relationship graph according to a comparison result between a connection weight corresponding to each node edge and a pruning threshold.

[0104] Among them, based on the comparison result corresponding to the node edge, there can be multiple ways to prune each node edge in the single weighted node relationship graph. For example, when the connection weight corresponding to the node edge is less than the pruning threshold, the node edge can be deleted in the single weighted node relationship graph to obtain the deleted node edge; when the connection weight corresponding to the node edge is not less than the pruning threshold, the node edge can be retained in the single weighted node relationship graph to obtain the retained node edge; based on the deleted node edge and the retained node edge, the pruned single weighted node relationship graph corresponding to each pruning threshold is determined.

[0105] Among them, the deleted node edge can be the node edge after deleting the node edge whose connection weight is less than the pruning threshold in the single weighted node relationship graph, that is, the node edge that is disconnected; the retained node edge can be the node edge after retaining the node edge whose connection weight is not less than the pruning threshold in the single weighted node relationship graph, that is, the node edge that keeps the connection established. After pruning each node edge in the single weighted node relationship graph according to the pruning threshold, the pruned single weighted node relationship graph can be obtained.

[0106] Optional, please refer to Figure 4b , Figure 4b It is a single weighted node relationship of a feature encoding method provided in an embodiment of the present application Figure 1 Schematic diagram of the pruning process. When the number of pruning is once, it can be assumed that the pruning threshold group includes m pruning thresholds, namely T1, T2, ..., Tm. The m pruning thresholds can be used to prune the single weighted node relationship graph G respectively, and m pruned single weighted node relationship graphs corresponding to the m pruning thresholds are obtained, which are G1, G2, ..., Gm respectively.

[0107] Optional, please refer to Figure 4c , Figure 4c It is a single weighted node relationship of a feature encoding method provided in an embodiment of the present application Figure 2Schematic diagram of the secondary pruning process. When the number of pruning is twice, the single weighted node relationship graph G can be pruned based on the connection weight and the pruning threshold group (T1, T2, ..., Tm) to obtain m pruned unweighted node relationship subgraphs. Then, the m pruned unweighted node relationship subgraphs can be pruned twice based on the connection weight and the pruning threshold group to obtain m×m unweighted node relationship subgraphs. Among them, it should be noted that when performing secondary pruning, the pruning threshold used can be the same as the pruning threshold used in the first pruning, or it can be different. It is only necessary to ensure that the pruning operation and dimension of each node in the weighted node relationship graph are the same. At the same time, the number of pruning thresholds for secondary pruning can also be different from the number of pruning thresholds for the first pruning. For example, in the first pruning, the pruning threshold group (T1, T2, ..., Tm) can be used for pruning, and in the second pruning, the pruning threshold group (T1, T2, ..., Tp) can be used for pruning, so that m×p unweighted node relationship subgraphs can be obtained. It should be noted that the specific value and number of the pruning threshold group can be selected according to the actual application, and are not limited here. In addition, for the case where the number of pruning is multiple times, it can be based on the case of single pruning and secondary pruning. Analogy.

[0108] In step 104, based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and the edge distribution information of the node is counted based on the target node edges.

[0109] The target node edge may be a node edge of a specific order selected from the unweighted node relationship subgraph based on the segmentation parameter, for example, a first-order node edge, a second-order node edge, or a third-order node edge. For example, please continue to refer to Figure 3 , taking the node relationship graph as an unweighted node relationship subgraph as an example, for node 1, node edge a and node edge c are first-order node edges of node 1, node edges ab and ad are second-order node edges of node 1, and for node 3, node edge cab is a third-order node edge of node 3. The edge distribution information may be information characterizing the distribution of target node edges corresponding to each node in the unweighted node relationship subgraph, and is used to characterize the distribution of target node edges of each node in the unweighted node relationship subgraph, for example, it may include the number of target node edges corresponding to each node in the unweighted node relationship subgraph.

[0110] Among them, based on the segmentation parameters, there can be multiple ways to filter out at least one target node edge corresponding to each node in the unweighted node relationship subgraph. For example, based on the segmentation parameters, the node edge order corresponding to each node can be determined, each node in the unweighted node relationship subgraph is traversed to obtain the connection order between the node edges corresponding to each node, and the node edges to be spliced ​​corresponding to the node edge order in each node in the unweighted node relationship subgraph are filtered out, and the node edges to be spliced ​​are spliced ​​according to the connection order to obtain at least one target node edge corresponding to each node.

[0111] The node edge order may be the order of the node edge, for example, the node edge order corresponding to the first-order node edge is 1, the node edge order corresponding to the second-order node edge is 2, the node edge order corresponding to the third-order node edge is 3, and so on. For example, please continue to refer to Figure 3 , taking the node relationship graph as an unweighted node relationship subgraph as an example, the node edges ab and ad are second-order node edges with a node edge order of 2 corresponding to node 1, and the node edge cab is a third-order node edge with a node edge order of 3 corresponding to node 3. Optionally, the number of splits (i.e., the number of pruning) can be extracted from the split parameters, and then the node edge order can be determined based on the number of splits. The node edge order can correspond to the number of splits. For example, when the number of splits is 1, the node edge order can be 1st order, that is, the target node edge can be a first-order node edge that still maintains a connection relationship after being split once. When the number of splits is 2, the node edge order can be 2nd order, and the target node edge can be a second-order node edge composed of two consecutive node edges that are still connected after being split. For example, please continue to refer to Figure 3 , assuming that the node relationship graph is a weighted node relationship graph, taking the second-order node edge of the target node edge as node 1 as an example, the node edges between node 1 and its first-order neighbor nodes (i.e., node 2 and node 3) (i.e., node edge a and node edge c) can be pruned for the first time. Assuming that the connection weights of node edge a and node edge c are not less than the pruning threshold of the first pruning, the retained node edges a and c can be obtained. Then, the node edges corresponding to the retained node edges a and c and their first-order neighbor nodes (i.e., node edges b and d) are pruned for the second time. Assuming that the connection weight corresponding to node edge b is not less than the pruning threshold of the second pruning, and the connection weight corresponding to node edge b is less than the pruning threshold of the second pruning, it can be determined that the target node edge is a second-order node edge ab. The connection order can be the order in which the node edges corresponding to each node are connected. For example, please continue to refer to Figure 3, taking the unweighted node relationship subgraph as an example, for node 1, node edge ab is connected to node edge b by node edge a, and node ad is connected to node edge d by node edge a. The node edge to be spliced ​​can be the node edge corresponding to the node edge order selected in the unweighted node relationship subgraph. For example, assuming that the node edge order is 1, for node 1, the node edges a and c to be spliced ​​corresponding to the node edge order can be selected in the unweighted node relationship subgraph. Assuming that the node edge order is 2, for node 1, the node edges a, b and c to be spliced ​​corresponding to the node edge order can be selected in the unweighted node relationship subgraph.

[0112] Among them, there are multiple ways to splice the node edges to be spliced ​​according to the connection order to obtain at least one target node edge corresponding to each node. For example, assuming that the node edge order is 1, for node 1, the node edges a and c to be spliced ​​corresponding to the node edge order can be screened out in the unweighted node relationship subgraph, so that the target node edges corresponding to node 1 can be determined as node edges a and c. Assuming that the node edge order is 2, for node 1, the node edges to be spliced ​​can be a, b and c according to the connection order, and the target node edges corresponding to node 1 include second-order node edges ab and second-order node edges ac.

[0113] After at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph based on the segmentation parameters, the edge distribution information of the node can be counted based on the target node edge. There are multiple ways to count the edge distribution information of the node based on the target node edge. For example, the number of target node edges corresponding to each node in the unweighted node relationship subgraph can be counted, and the edge distribution information of the node in the unweighted node relationship subgraph can be determined based on the number of target node edges.

[0114] The number of target node edges may be the number of corresponding target node edges for each node in the unweighted node relationship subgraph.

[0115] There are many ways to count the number of target node edges corresponding to each node in the unweighted node relationship subgraph. For example, please continue to refer to Figure 4b , for the weighted node relationship graph G, W ij represents the connection weight between nodes i and j in G, Edge(W ij ) indicates that the connection weight is W ij For the case where the node edge order is 1, that is, when the target node edge is a first-order node edge, the first-order node edge structure distribution of node i in G can be expressed as

[0116]

[0117] in, is the first-order node edge structure distribution of node i, which can represent the edge distribution information of node i in G for the first-order target node edge, and is used to characterize the distribution of the first-order node edges of node i in the weighted node relationship graph G. Then, the weighted node relationship graph can be pruned using the pruning threshold group (T1, T2, ..., Tm) to obtain m pruned unweighted node relationship subgraphs G 1 ,G 2 ,…,G m If node j and node i are any two nodes in G, then the kth unweighted node relationship subgraph G k The first-order connection between nodes i and j can be expressed as

[0118]

[0119] in, Indicates the connection status of the node edge between node i and node j after pruning, Equal to 1 means that node j is in the unweighted graph G k is connected to node i, that is, node j is the first-order neighbor node of node i.

[0120] The unweighted node relationship subgraph G k All first-order node edges in Edge(T k ) indicates that the connection between nodes i and j in G is Edge(W ij ) can be used G 1 ,G 2 ,…,G m The node edges in can be represented by

[0121]

[0122] Among them, T k It can be an unweighted node relationship subgraph G k The corresponding pruning threshold. Therefore, based on the number of edges of the target node, the node i in the unweighted node relationship subgraph G can be determined. k The edge distribution information in can be expressed as

[0123]

[0124] Among them, node i is in the unweighted node relationship subgraph G k The number of target node edges corresponding to the first-order target node edges in can be expressed as

[0125]

[0126] Correspondingly, the first-order node edge structure distribution of node i in the relationship graph G can be approximately expressed as

[0127]

[0128] For the case where the node edge order is 2, that is, the target node edge is a second-order node edge, please continue to refer to Figure 4c , the second-order node edge of the node in the weighted node relationship graph G can be composed of the first-order node edge and the second-order node edge. Assume that one of the second-order node edges of node i is composed of a connection weight of W io The first-order node edge and connection weight is W oj , where node o is the first-order neighbor node of node i, and node j is the first-order neighbor node of node o, that is, node j is the second-order neighbor node of node i. Then the second-order node edge can be represented as Edge(W io ,W oj ), then Edge(W io ,W oj ) is regarded as the first-order node edge Edge(W io ) and the first-order edge Edge(W oj ) are concatenated, the second-order node edge of node i can be defined as

[0129]

[0130] Therefore, the second-order node edge structure distribution of node i can be expressed as

[0131]

[0132] Substituting into formula (1), the second-order node edge structure distribution of node i in the relationship graph G is It can be expressed as

[0133]

[0134] Among them, T s Indicates that the connection weight is W io The pruning threshold corresponding to the first-order node edge, T v The connection weight can be expressed as W oj The pruning threshold corresponding to the second-order node edge of . Correspondingly, the number of target node edges of the second-order target node edge corresponding to node i in the unweighted node relationship subgraph can be expressed as

[0135]

[0136] Therefore, based on the number of edges of the target node, it can be determined that the node i in the unweighted node relationship subgraph is connected by a weight W io The first-order node edge and connection weight is W ojThe edge distribution information corresponding to the target node edge composed of the second-order node edges can be expressed as

[0137]

[0138] It should be noted that, in the process of deriving the above formula, the edge distribution information of the node is represented by the edge distribution information of the node in the weighted node relationship graph G, and the edge distribution information of each node in each unweighted node relationship subgraph can be determined according to the edge distribution information of the node in the weighted node relationship graph G. In addition, the above formula is derived by taking the first-order node edge and the second-order node edge as examples, and accordingly, the n-order case can be derived by analogy.

[0139] In step 105, feature encoding is performed on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes.

[0140] The node feature may be feature information characterizing a node in a weighted relationship graph, and may include attribute information of the node itself and edge distribution information corresponding to the node connection relationship of the node in the weighted relationship graph.

[0141] Among them, there can be multiple ways to feature encode the corresponding nodes in the weighted node relationship graph according to the edge distribution information. For example, when the weighted node relationship graph is a single weighted node relationship graph, the corresponding nodes in the weighted node relationship graph can be feature encoded according to the edge distribution information to obtain the node features of the nodes; when the weighted node relationship graph is a multiple weighted node relationship graph, the corresponding nodes in at least two single weighted node relationship graphs of the weighted node relationship graph can be feature encoded according to the edge distribution information to obtain the single node features of the node in the corresponding single weighted node relationship graph, and the single node features can be spliced ​​to obtain the corresponding node features of the node in the weighted node relationship graph.

[0142] Among them, the single node feature can be a node feature corresponding to the node in the multiple weighted node relationship graph in at least two single weighted node relationship graphs corresponding to the multiple node relationship graphs.

[0143] Among them, according to the edge distribution information, the corresponding node in the weighted node relationship graph is feature encoded, and there can be multiple ways to obtain the node characteristics of the node. For example, the edge distribution characteristics corresponding to the node in each unweighted node relationship subgraph can be extracted from the edge distribution information, and the edge distribution characteristics corresponding to the node can be combined to obtain the target edge distribution characteristics of the node. According to the target edge distribution characteristics, the node characteristics of the node are determined.

[0144] Among them, the edge distribution feature can be feature information characterizing the node connection relationship of the node in the unweighted node relationship subgraph, that is, it can be feature information characterizing the edge structure distribution of the node in the unweighted node relationship subgraph, and the target edge distribution feature can be feature information characterizing the node connection relationship of the node in the weighted node relationship graph, which can include the distribution of node edges corresponding to the nodes in the weighted node relationship graph and the distribution of connection weights corresponding to the node edges.

[0145] There may be multiple ways to extract the edge distribution features corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information. For example, the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs may be extracted from the edge distribution information, the edge information may be used as an edge feature basis, and the distribution information may be converted into an edge distribution feature based on the edge feature basis.

[0146] The edge information may be information representing a target node edge corresponding to a node in an unweighted node relationship subgraph, and the distribution information may be information representing the distribution of the target node edge corresponding to the node in the unweighted node relationship subgraph.

[0147] In one embodiment, when the node edge order is first order, that is, when the target node edge is a first order node edge, formula (2) can be converted into

[0148]

[0149] Thus, the edge information Edge(T) corresponding to the node in each unweighted node relationship subgraph can be extracted from the edge distribution information. k ) and the distribution information is

[0150]

[0151] Then, the edge information can be used as the edge feature basis, and based on the edge feature basis, the distribution information can be converted into edge distribution features. The edge distribution feature of the first-order node edge of node i in the kth unweighted node relationship subgraph can be expressed as

[0152]

[0153] in, The characteristic information of the distribution of the first-order node edges corresponding to node i in each unweighted node relationship subgraph, that is, the edge distribution feature, can be expressed as:

[0154]

[0155] For the weighted node relationship graph G, we can use the edge information Edge(T 1),…,Edge(T m ) as the edge feature basis, so that the edge distribution features corresponding to the node i can be combined to obtain the target edge distribution features of the first-order node edges of node i in G. The target edge distribution features can be expressed as

[0156]

[0157] Please continue to refer to Figure 4b , with Edge(T 1 ),…,Edge(T m ) as the edge feature basis, the target edge distribution feature of the first-order node edge of node 1 in the weighted node relationship graph G can finally be encoded into a vector (4,3,…,2).

[0158] In one embodiment, when the node edge order is second order, that is, the target node edge is a second order node edge, the pruning thresholds used in the two prunings are the same, which are T1, T2, ..., Tm. Similarly, based on formula (3), it can be obtained that the node i in the unweighted node relationship subgraph consists of a connection weight of W io The first-order node edge and connection weight is W oj The edge distribution characteristics of the target node edge composed of the second-order node edge can be expressed as

[0159]

[0160] Correspondingly, the second-order node edge structure distribution of node i in the weighted node relationship graph G can be expressed as

[0161]

[0162] Edge(T 1 ,T 1 ),…,Edge(T m ,T m ) as the edge feature basis, and combining the edge distribution features corresponding to the node, we can get the target edge distribution features of the second-order node edges of the node i in G. The target edge distribution features can be encoded as

[0163]

[0164] Please continue to refer to Figure 4c , the target edge distribution characteristics of the second-order node edge of node 1 in G can be encoded as

[0165]

[0166] After combining the edge distribution features corresponding to the node to obtain the target edge distribution features of the node, the node features of the node can be determined based on the target edge distribution features. There are many ways to determine the node features of the node based on the target edge distribution features, for example, extracting the features of the node's own behavior and attributes, fusing the features with the target edge distribution features, and obtaining the node features of the node.

[0167] Similarly, when the node edge order is n, that is, when the target node edge is an n-order node edge, the second-order structure is expanded to the n-order structure, and the n pruning uses the same pruning threshold (T1, T2, ..., Tm), so Edge (T 1 ,…,T 1 ),…,Edge(T m ,…,T m )Total m n The n-order node edge features of the node are encoded by a basis, and the obtained feature encoding is an m×m×…×m-dimensional tensor. Then, the target edge distribution feature of the n-order node edge of node i at the coordinate (k, s,…, v) can be expressed as

[0168]

[0169] Among them, node o 1 is the first-order neighbor node of node i, o 2 Yes 1 The first-order neighbor node of o n Yes n-1 The first-order neighbor nodes of .

[0170] In this way, the weighted node relationship graph can be pruned according to multiple pruning thresholds in the pruning threshold group to obtain multiple unweighted node relationship subgraphs, so that the distribution of the target node edges of each node in the unweighted node relationship subgraph can be counted, and then the node connection relationship and edge distribution information of each node in the weighted node relationship graph can be accurately simulated according to the distribution of the target node edges of each node in the unweighted node relationship subgraph, so that the node connection relationship, node edges and corresponding connection weight distribution information of the node in the weighted node relationship graph can be integrated into the feature encoding of the node based on the target edge distribution characteristics, so as to determine the node characteristics of each node, improve the accuracy of the node feature encoding of each node in the weighted node relationship graph, and can be applied to application scenarios such as financial risk control, recommendation systems, social networks, transportation networks, and e-shopping. By encoding the target edge distribution characteristics of the user in the node relationship graph of the corresponding scenario, combined with the characteristics of the user's own behavior and attribute information, the importance of the user in the relationship graph is accurately layered, which is convenient for diversified management of users. In addition, the feature encoding method provided in the embodiment of the present application can be easily extended to n-order node edges for node feature encoding, reducing the information loss of nodes during the feature encoding process, and the computational complexity is low. At the same time, it can be extended to the feature encoding of edge distribution characteristics of nodes in multiple weighted node relationship graphs (Multigraph), further improving the encoding efficiency of node features.

[0171] In one embodiment, the feature encoding method provided in the embodiment of the present application can be applied to a recommendation system scenario. Specifically, a weighted node relationship graph corresponding to a content recommendation application can be obtained. The content recommendation application includes at least one object and association information between objects. The weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges. The objects in the content recommendation application correspond one-to-one to the nodes in the weighted node relationship graph. The node edges and the connection weights corresponding to the node edges can be determined based on the association information of the objects in the content recommendation application. The connection attributes of the nodes are identified in the weighted node relationship graph. According to the connection attribute information, the segmentation parameters of the weighted node relationship graph are determined; according to the connection weights and the segmentation parameters, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and the edge distribution information of the node is counted based on the target node edge; according to the edge distribution information, the corresponding nodes in the weighted node relationship graph are feature encoded to obtain node features of the nodes; based on the node features, content recommendation is performed for each object in the content recommendation application.

[0172] Among them, the content recommendation application can be an application that recommends content based on a recommendation system, the object can be a user in the content recommendation application, the content can be a carrier of information, for example, it can be content in the form of video, audio, article, picture, etc., and the related information can be information including the association relationship between users in the content recommendation application, for example, it can include information such as the interactive relationship between users, the similarity of interests, etc. The interactive relationship can include the relationship between users in interactive behaviors such as likes, attention, comments, sharing, browsing, etc.

[0173] Among them, based on the node characteristics, there can be multiple ways to recommend content to each object in the content recommendation application. For example, based on the node characteristics of each object in the content recommendation application, the interest preference of the object can be predicted, and the content to be recommended can be obtained and pushed to the object according to the predicted interest preference, thereby improving the accuracy of content recommendation. For example, a content recommendation model can be established based on the node characteristics of each object node in the content recommendation application, and the interest preference of the object in the content recommendation application can be predicted through the content recommendation model, so that the content that the user may be interested in can be obtained according to the interest preference of the object, and the content can be pushed to the user, thereby improving the accuracy of content recommendation and thus improving the efficiency of content recommendation.

[0174] In one embodiment, the feature encoding method provided in the embodiment of the present application can be applied to a financial risk control scenario. Specifically, a weighted node relationship graph corresponding to a target payment network can be obtained. The target payment network may include at least one object. The weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and a connection weight corresponding to the node edge. The objects in the target payment network correspond one-to-one to the nodes in the weighted node relationship graph. The node edge and the connection weight corresponding to the node edge can be determined based on the payment relationship of the objects in the target payment network. The connection attribute information of the nodes is identified in the weighted node relationship graph, and based on The method comprises the following steps: connecting attribute information, determining the segmentation parameters of the weighted node relationship graph; segmenting the weighted node relationship graph according to the connection weight and the segmentation parameters to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, screening out at least one target node edge corresponding to each node in the unweighted node relationship subgraph, and calculating the edge distribution information of the node based on the target node edge; according to the edge distribution information, feature encoding the corresponding nodes in the weighted node relationship graph to obtain the node features of the nodes; based on the node features, identifying abnormal objects in the objects corresponding to the target payment network.

[0175] Among them, the target payment network can be a payment network in a financial payment scenario, and the payment network is used to characterize the payment relationship between each object in a certain financial payment scenario. Each node in the weighted node relationship diagram can correspond one-to-one to each object in the target payment network. The object can be a user in the target payment network, and the abnormal object can be a user with abnormal behavior among the objects of the target payment network, for example, it can be a user with abnormal payment behavior, for example, it can be a user with illegal behavior, or it can be a user who has been deceived.

[0176] Among them, there can be multiple ways to perform feature encoding on the corresponding nodes in the weighted node relationship graph based on the edge distribution information. For example, the target edge distribution features corresponding to each node in the weighted node relationship graph can be determined based on the edge distribution information, and the behavioral attribute features corresponding to each node can be obtained, so that the behavioral attribute features and the target edge distribution features can be fused to obtain the node features of the node.

[0177] Among them, the behavior attribute feature may be a feature of the payment behavior and attributes of the user corresponding to the node, for example, it may be information including the user's payment habits, payment ability, payment preferences, etc. Among them, it can be understood that in the specific implementation of this application, when the above embodiments of this application are applied to specific products or technologies, it is necessary to obtain user permission or consent, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0178] After encoding the features of the corresponding nodes in the weighted node relationship graph according to the edge distribution information, the abnormal objects can be identified in the objects corresponding to the target payment network based on the node features. There are many ways to identify the abnormal objects in the objects corresponding to the target payment network based on the node features. For example, the payment behavior of the objects in the target payment network can be predicted based on the node features of each object node in the target payment network, and the abnormal objects in the target payment network can be identified based on the prediction results, so as to timely handle the abnormal objects in the target payment network and ensure the payment security of the target payment network. For example, an abnormal object recognition model can be established based on the node features of each object node in the target payment network, and potential abnormal objects in the target payment network can be identified through the abnormal object recognition model, and risk warnings for the abnormal users can be issued to avoid problems such as abnormal payment behaviors in the target payment network causing economic losses to users, thereby improving the recognition efficiency of abnormal objects and ensuring the normal operation of payment behaviors in the target payment network.

[0179] In one embodiment, when the weighted node relationship graph is a multiple weighted node relationship graph, after obtaining the single node features of the node in the corresponding single weighted node relationship graph, the single node features can be spliced ​​to obtain the node features corresponding to the node in the weighted node relationship graph, wherein there can be multiple ways to splice the single node features. For example, since the number of pruning thresholds corresponding to each single weighted node relationship graph is the same, the encoding dimension of each single node feature is the same, and then the single node features corresponding to the multiple weighted node relationship graphs can be spliced ​​in a form similar to the color system (RGB) value in the picture, so as to obtain the node features corresponding to the node in the weighted node relationship graph.

[0180] From the above, it can be seen that the embodiment of the present application obtains a weighted node relationship graph, which includes at least two nodes, at least one node edge connecting the nodes, and the connection weights corresponding to the node edges; identifies the connection attribute information of the nodes in the weighted node relationship graph, and determines the segmentation parameters of the weighted node relationship graph based on the connection attribute information; segments the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, screens out at least one target node edge corresponding to each node in the unweighted node relationship subgraph, and counts the edge distribution information of the node based on the target node edge; based on the edge distribution information, features the corresponding nodes in the weighted node relationship graph to obtain node features of the nodes. In this way, by dividing the weighted node relationship graph into at least one unweighted node relationship subgraph, and counting the edge distribution information of the nodes in the unweighted node relationship subgraph, and then feature encoding the nodes in the weighted node relationship graph according to the edge distribution information corresponding to each unweighted node relationship subgraph, the node features of the nodes are obtained, and the node connection relationship between the node and other nodes in the weighted node relationship graph is integrated into the node features, thereby improving the accuracy of node feature encoding in the relationship graph, and then improving the encoding efficiency of node features.

[0181] According to the method described in the above embodiment, the following is further described in detail with examples.

[0182] In this embodiment, the feature encoding device is specifically integrated in a computer device as an example for description. The feature encoding method is specifically described by taking a server as the execution subject and the weighted node relationship graph as a single weighted node relationship graph as an example.

[0183] For a better description of the embodiments of this application, please refer to Figure 5 , Figure 5 Another schematic diagram of the feature encoding method provided in the embodiment of the present application. The specific process is as follows:

[0184] In step 201, the server obtains a weighted node relationship graph, identifies connection attribute information of the node in the weighted node relationship graph, and determines a segmentation parameter of the weighted node relationship graph based on the connection attribute information.

[0185] Among them, there are many ways for the server to identify the connection attribute information of nodes in the weighted node relationship graph. For example, the server can traverse the nodes and node edges in the weighted node relationship graph, and based on the traversal results, the connection attribute information of each node in the weighted node relationship graph can be counted.

[0186] After the server identifies the connection attribute information of the node in the weighted node relationship graph, it can determine the segmentation parameters of the weighted node relationship graph based on the connection attribute information. There are many ways for the server to determine the segmentation parameters of the weighted node relationship graph based on the connection attribute information. For example, the server can determine the node connection status of each node in the weighted node relationship graph based on the connection attribute information corresponding to the node, and can obtain the longest node edge of each node in the weighted node relationship graph based on the node connection status of each node, so as to determine the segmentation parameters of the weighted node relationship graph based on the longest node edge of each node in the weighted node relationship graph. The longest node edge can be the longest path traversing the weighted node relationship graph with a certain node as the root node. For example, please refer to Figure 3 The node relationship graph includes node 1, node 2, node 3, node 4 and node 5. Taking the node relationship graph as a weighted node relationship graph as an example, assuming that node 1 is the root node and traverses the node relationship graph, the longest path of node 1 can be obtained as two node edges, which can be two node edges from node 1 to node 4 or from node 1 to node 5, namely node edges ab and ad. Assuming that node 3 is the root node and traverses the node relationship graph, the longest path of node 3 can be obtained as three node edges, which can be three node edges from node 3 to node 4 or from node 3 to node 5, namely node edges cab and cad. It can also be assumed that node 2 is the root node and traverses the node relationship graph, then the longest path of node 2 can be obtained as two node edges, which can be two node edges from node 2 to node 3, namely ac, etc.

[0187] There are many ways for the server to determine the segmentation parameters of the weighted node relationship graph based on the longest node edge of each node in the weighted node relationship graph. For example, the server can count the longest node edge of each node and determine the longest node edge with the most occurrences based on the statistical results, so as to determine the segmentation parameter based on the number of node edges corresponding to the longest node edge with the most occurrences. For example, please continue to refer to Figure 3 ,exist Figure 3In the weighted node relationship graph shown, the longest node edge of node 1 is two node edges, the longest node edge of node 2 is two node edges, the longest node edge of node 3 is three node edges, the longest node edge of node 4 is three node edges, and the longest node edge of node 5 is three node edges, so that the longest node edges of the five nodes can be counted, and according to the statistical results, it can be determined that the longest node edge with the most occurrences is 3 node edges. Therefore, the server can determine the number of splits to be 3 according to the longest node edge with the most occurrences, thereby determining the splitting parameters of the weighted node relationship graph.

[0188] In one embodiment, the server may further select the segmentation parameters of the weighted node relationship graph according to the computing power and computing accuracy of the feature coding device. For example, taking the segmentation parameter as the number of segmentation times as an example, when the computing power of the server is strong, a larger number of segmentation times may be selected, and when the computing power of the server is weak, a smaller number of segmentation times may be selected. When the computing accuracy of the feature coding is high, a larger number of segmentation times may be selected to obtain more node information based on high-order features, and when the computing accuracy of the feature coding is low, a smaller number of segmentation times may be selected.

[0189] In step 202, the server identifies the number of connection weights corresponding to each node edge in the weighted node relationship graph, and when there are no node edges corresponding to at least two connection weights in the node edges, the relationship graph type of the weighted node relationship graph is determined to be a single weighted node relationship graph.

[0190] Among them, there can be multiple ways for the server to identify the number of connection weights corresponding to each node edge in the weighted node relationship graph. For example, the server can traverse the weighted node relationship graph, and thus obtain the number of connection weights corresponding to each node edge based on the traversal results.

[0191] Optionally, when there are node edges corresponding to at least two connection weights in the node edges, the server may determine the relationship graph type of the weighted node relationship graph as a multiple weighted node relationship graph.

[0192] In step 203, the server determines the number of pruning times of the single weighted node relationship graph according to the segmentation parameter, and obtains a pruning threshold group corresponding to the single weighted node relationship graph.

[0193] Optionally, in the pruning threshold group, the value of each pruning threshold may be between 0 and 1. For example, it may be assumed that the pruning threshold group includes m pruning thresholds, namely T1 to Tm, where 0≤T1≤T2≤…≤Tm≤1.

[0194] There are many ways for the server to determine the number of pruning times of the single weighted node relationship graph according to the segmentation parameter. For example, the server can extract the number of segmentation times from the segmentation parameter and use the number of segmentation times as the number of pruning times.

[0195] In step 204, when the number of pruning times is one, the server compares at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph, and prunes each node edge in the single weighted node relationship graph based on the comparison result corresponding to the node edge to obtain the pruned single weighted node relationship graph.

[0196] Optional, please refer to Figure 4b , when the number of pruning is one, it can be assumed that the pruning threshold group includes m pruning thresholds, namely T1, T2, ..., Tm. The server can use m pruning thresholds to prune the single weighted node relationship graph G respectively, and obtain m pruned single weighted node relationship graphs corresponding to m pruning thresholds, namely G1, G2, ..., Gm. Assuming that node j and node i are any two nodes in G, the kth unweighted node relationship subgraph G k The first-order connection between nodes i and j can be expressed as

[0197]

[0198] Among them, T k Represents the kth unweighted node relationship subgraph G k The corresponding pruning threshold, W ij represents the connection weight between nodes i and j. If W ij ≥T k , the node edge of the connection weight can be retained. If W ij <

[0199] T k , then the node edge of the connection weight can be pruned. Equal to 1 means that node j is in the unweighted graph G k is connected to node i, that is, j is the first-order neighbor node of node i.

[0200] In step 205, the server deletes the connection weight corresponding to each node edge in the pruned single weighted node relationship graph to obtain an unweighted node relationship subgraph corresponding to each pruning threshold.

[0201] Among them, the server prunes the node edges corresponding to at least two connection weights in the weighted node relationship graph through the pruning threshold group, so as to obtain a pruned single weighted node relationship graph in which multiple node edges have no corresponding connection weights. At the same time, the original connection weight corresponding to each node edge in the pruned single weighted node relationship graph can be deleted to obtain an unweighted node relationship subgraph corresponding to each pruning threshold.

[0202] In step 206, when the number of pruning times is multiple, the single weighted node relationship graph is pruned based on the connection weight and the pruning threshold group to obtain a pruned unweighted node relationship subgraph.

[0203] When the number of pruning times is multiple, the server can compare at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the weighted node relationship graph, and perform multiple pruning on the weighted node relationship graph according to the comparison result.

[0204] In step 207, the server uses the pruned unweighted node relationship subgraph as a single weighted node relationship graph, and returns to execute the step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group until the number of pruning times is reached to obtain at least one unweighted node relationship subgraph.

[0205] Optional, please continue to refer to Figure 4c , when the number of pruning is two times, the server can prune the single weighted node relationship graph G based on the connection weight and the pruning threshold group (T1, T2, ..., Tm) to obtain m pruned unweighted node relationship subgraphs, and then, the m pruned unweighted node relationship subgraphs can be pruned again based on the connection weight and the pruning threshold group to obtain m×m unweighted node relationship subgraphs. It should be noted that when performing the second pruning, the pruning threshold used can be the same as or different from the pruning threshold used in the first pruning. At the same time, the number of pruning thresholds for the second pruning can also be different from the number of pruning thresholds for the first pruning. For example, the pruning threshold group (T1, T2, ..., Tm) can be used for pruning in the first pruning, and the pruning threshold group (T1, T2, ..., Tp) can be used for pruning in the second pruning, so that m×p unweighted node relationship subgraphs can be obtained. It should be noted that the specific value and number of the pruning threshold group can be selected according to actual application and are not limited here. In addition, for the case where the number of pruning is multiple times, it can be inferred based on the case of single pruning and double pruning.

[0206] In step 208, the server screens out at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameters, and counts the number of target node edges corresponding to each node in the unweighted node relationship subgraph, and determines the edge distribution information of the node in the unweighted node relationship subgraph based on the number of target node edges.

[0207] There are many ways for the server to count the number of target node edges corresponding to each node in the unweighted node relationship subgraph. For example, please continue to refer to Figure 4b , for the weighted node relationship graph G, W ij represents the connection weight between nodes i and j in G, Edge(W ij ) indicates that the connection weight is W ij For the first-order node edge of , when the node edge order is 1, that is, when the target node edge is a first-order node edge, the first-order node edge structure distribution of node i in G can be expressed as

[0208]

[0209] in, is the first-order node edge structure distribution of node i, which can represent the edge distribution information of node i in G for the first-order target node edge, and is used to characterize the distribution of the first-order node edges of node i in the weighted node relationship graph G. Then, the weighted node relationship graph can be pruned using the pruning threshold group (T1, T2, ..., Tm) to obtain m pruned unweighted node relationship subgraphs G 1 ,G 2 ,…,G m If node j and node i are any two nodes in G, then the kth unweighted node relationship subgraph G k The first-order connection between nodes i and j can be expressed as

[0210]

[0211] The unweighted node relationship subgraph G k All first-order node edges in Edge(T k ) indicates that the connection between nodes i and j in G is Edge(W ij ) Available G 1 ,G 2 ,…,G m The node edges in are represented as

[0212]

[0213] in, It can represent the number of target node edges of the first-order target node edge corresponding to node i, Tk It can be an unweighted node relationship subgraph G k The corresponding pruning threshold. Therefore, based on the number of edges of the target node, the node i in the unweighted node relationship subgraph G can be determined. k The edge distribution information in can be expressed as

[0214]

[0215] Among them, node i is in the unweighted node relationship subgraph G k The number of target node edges corresponding to the first-order target node edges in can be expressed as

[0216]

[0217] Correspondingly, the first-order node edge structure distribution of node i in the relationship graph G can be approximately expressed as formula (2).

[0218] For the case where the node edge order is second order, that is, when the target node edge is a second order node edge, please continue to refer to Figure 4c , the second-order node edge of the node in the weighted node relationship graph G can be composed of the first-order node edge and the second-order node edge. Assume that one of the second-order node edges of node i is composed of a connection weight of W io The first-order node edge and connection weight is W oj , where node o is the first-order neighbor node of node i, and node j is the first-order neighbor node of node o, that is, node j is the second-order neighbor node of node i. Then the second-order node edge can be represented as Edge(W io ,W oj ), then Edge(W io ,W oj ) is regarded as the first-order node edge Edge(W io ) and the first-order edge Edge(W oj ) are concatenated, the second-order node edge of node i can be defined as

[0219]

[0220] The second-order node edge structure distribution of node i can be expressed as

[0221]

[0222] Substituting into formula (1), we can get the second-order node edge structure distribution of node i in the relationship graph G: It can be expressed as

[0223]

[0224] It should be noted that, in the process of deriving the above formula, the edge distribution information of the node is represented by the edge distribution information of the node in the weighted node relationship graph G, and the edge distribution information of each node in each unweighted node relationship subgraph can be determined according to the edge distribution information of the node in the weighted node relationship graph G. In addition, the above formula is derived by taking the first-order node edge and the second-order node edge as examples, and accordingly, the n-order case can be derived by analogy.

[0225] In step 209, the server extracts the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information, uses the edge information as an edge feature basis, and converts the distribution information into an edge distribution feature based on the edge feature basis.

[0226] In one embodiment, when the node edge order is first order, that is, when the target node edge is a first order node edge, formula (2) can be converted into

[0227]

[0228] Thus, the edge information Edge(T) corresponding to the node in each unweighted node relationship subgraph can be extracted from the edge distribution information. k ) and the distribution information is

[0229]

[0230] Then the server can send the edge information Edge(T k ) as the edge feature basis, and based on the edge feature basis, the distribution information is converted into edge distribution features. The edge distribution feature of the first-order node edge of node i in the kth unweighted node relationship subgraph can be expressed as

[0231]

[0232] in, It can be the characteristic information of the distribution of the first-order node edges corresponding to node i in each of the unweighted node relationship subgraphs, that is, the edge distribution characteristics.

[0233] In step 210, the server combines the edge distribution features corresponding to the node to obtain the target edge distribution feature of the node, and determines the node feature of the node according to the target edge distribution feature.

[0234] Among them, the first-order node edge structure distribution of node i in the weighted node relationship graph G can be expressed as

[0235]

[0236] For the weighted node relationship graph G, we can use the edge information Edge(T1 ),…,Edge(T m ) as the edge feature basis, so that the server can combine the edge distribution features corresponding to the node i to obtain the target edge distribution features of the first-order node edges of node i in G. The target edge distribution features corresponding to node i in the weighted node relationship graph G can be expressed as

[0237]

[0238] Please continue to refer to Figure 4b , with Edge(T 1 ),…,Edge(T m ) as the edge feature basis, the target edge distribution feature of the first-order node edge of node 1 in the weighted node relationship graph G can finally be encoded into a vector (4,3,…,2).

[0239] After combining the edge distribution features corresponding to the node to obtain the target edge distribution features of the node, the node features of the node can be determined based on the target edge distribution features. There are many ways to determine the node features of the node based on the target edge distribution features, for example, extracting the features of the node's own behavior and attributes, combining the features with the target edge distribution features, and obtaining the node features of the node.

[0240] From the above, it can be seen that the embodiment of the present application obtains a weighted node relationship graph through a server, identifies the connection attribute information of the node in the weighted node relationship graph, and determines the segmentation parameters of the weighted node relationship graph based on the connection attribute information; the server identifies the number of connection weights corresponding to each node edge in the weighted node relationship graph, and when there is no node edge corresponding to at least two connection weights in the node edge, the relationship graph type of the weighted node relationship graph is determined to be a single weighted node relationship graph; the server determines the number of pruning times of the single weighted node relationship graph according to the segmentation parameters, and obtains The pruning threshold group corresponding to the single weighted node relationship graph; when the pruning times is once, the server compares at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph, and based on the comparison result corresponding to the node edge, prunes each node edge in the single weighted node relationship graph to obtain the pruned single weighted node relationship graph; the server deletes the connection weight corresponding to each node edge in the pruned single weighted node relationship graph to obtain an unweighted node relationship subgraph corresponding to each pruning threshold; when the pruning times is once, the server compares at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph to obtain an unweighted node relationship subgraph corresponding to each pruning threshold; When the number of pruning times is multiple, the single weighted node relationship graph is pruned based on the connection weight and the pruning threshold group to obtain a pruned unweighted node relationship subgraph; the server uses the pruned unweighted node relationship subgraph as a single weighted node relationship graph, and returns to execute the step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group until the number of pruning times is reached to obtain at least one unweighted node relationship subgraph; the server selects at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameter, and counts the number of target node edges of the target node edge corresponding to each node in the unweighted node relationship subgraph, and determines the edge distribution information of the node in the unweighted node relationship subgraph based on the target node edge number; the server extracts the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information, uses the edge information as an edge feature basis, and converts the distribution information into an edge distribution feature based on the edge feature basis; the server combines the edge distribution features corresponding to the node to obtain the target edge distribution feature of the node, and determines the node feature of the node according to the target edge distribution feature.In this way, the weighted node relationship graph can be pruned according to multiple pruning thresholds in the pruning threshold group to obtain multiple unweighted node relationship subgraphs, so that the distribution of target node edges of each node in the unweighted node relationship subgraph can be counted, and then the node connection relationship and edge distribution information of each node in the weighted node relationship graph can be accurately simulated according to the distribution of target node edges of each node in the unweighted node relationship subgraph, so that the node connection relationship, node edges and corresponding connection weight distribution information of the node in the weighted node relationship graph can be integrated into the feature coding of the node based on the target edge distribution characteristics, so as to determine the node feature of each node, thereby improving the accuracy of the node feature coding of each node in the weighted node relationship graph. In addition, the feature coding method provided in the embodiment of the present application is easy to expand to node feature coding based on multi-order node edges, reduces the information loss of nodes in the feature coding process, and has low computational complexity, further improving the coding efficiency of node features.

[0241] In order to better implement the above method, an embodiment of the present invention further provides a feature encoding device, which can be integrated in a computer device, and the computer device can be a server.

[0242] For example, Figure 6 As shown, it is a schematic diagram of the structure of the feature encoding device provided in an embodiment of the present application. The feature encoding device may include a sample acquisition unit 301, a determination unit 302, a segmentation unit 303, a screening unit 304 and an encoding unit 305, as follows:

[0243] An acquisition unit 301 is used to acquire a weighted node relationship graph, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and a connection weight corresponding to the node edge;

[0244] A determination unit 302, configured to identify connection attribute information of the node in the weighted node relationship graph, and determine a segmentation parameter of the weighted node relationship graph based on the connection attribute information;

[0245] A segmentation unit 303 is used to segment the weighted node relationship graph according to the connection weight and the segmentation parameter to obtain at least one unweighted node relationship subgraph, where the unweighted node relationship subgraph is a node relationship subgraph without connection weights;

[0246] A screening unit 304 is used to screen out at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameter, and to calculate edge distribution information of the node based on the target node edge;

[0247] The encoding unit 305 is used to perform feature encoding on the corresponding node in the weighted node relationship graph according to the edge distribution information to obtain the node feature of the node.

[0248] In one embodiment, the segmentation unit 303 includes:

[0249] A relationship graph type determination subunit, used to determine the relationship graph type of the weighted node relationship graph according to the connection weight corresponding to the node edge;

[0250] A multiple splitting sub-unit is used for splitting the multiple weighted node relationship graph to obtain at least two single weighted node relationship graphs when the relationship graph type is a multiple weighted node relationship graph, and splitting the single weighted node relationship graph according to the connection weight and the splitting parameter to obtain at least one unweighted node relationship sub-graph, wherein each node edge in the single weighted node relationship graph corresponds to a connection weight;

[0251] The single splitting sub-unit is used to split the single weighted node relationship graph according to the connection weight and the splitting parameter to obtain at least one unweighted node relationship sub-graph when the relationship graph type is a single weighted node relationship graph.

[0252] In one embodiment, the relationship graph type determination subunit includes:

[0253] A weight quantity identification module is used to identify the quantity of connection weights corresponding to each node edge in the weighted node relationship graph;

[0254] A multiple weighted node relationship graph determination module, used for determining the relationship graph type of the weighted node relationship graph as a multiple weighted node relationship graph when there are node edges corresponding to at least two connection weights in the node edges;

[0255] The single weighted node relationship graph determination module is used to determine the relationship graph type of the weighted node relationship graph as a single weighted node relationship graph when there is no node edge corresponding to at least two connection weights in the node edge.

[0256] In one embodiment, the multiple segmentation subunit and the single segmentation subunit include:

[0257] A pruning times determination module, used to determine the pruning times of the single weighted node relationship graph according to the segmentation parameter, and obtain a pruning threshold group corresponding to the single weighted node relationship graph, wherein the pruning threshold group includes at least one pruning threshold;

[0258] The segmentation module is used to segment the single weighted node relationship graph based on the pruning times, the connection weights and the pruning threshold group to obtain at least one unweighted node relationship subgraph.

[0259] In one embodiment, the segmentation module is used to:

[0260] A first pruning submodule is used for pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group when the pruning number is once, to obtain an unweighted node relationship subgraph corresponding to each pruning threshold;

[0261] The second pruning submodule is used to prune the single weighted node relationship graph based on the connection weight and the pruning threshold group when the pruning times are multiple times, to obtain a pruned unweighted node relationship subgraph, and to use the pruned unweighted node relationship subgraph as the single weighted node relationship graph, and return to execute the step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group until the pruning times are reached to obtain at least one unweighted node relationship subgraph.

[0262] In one embodiment, the first pruning submodule is used to:

[0263] Comparing at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph;

[0264] Based on the comparison result corresponding to the node edge, prune each node edge in the single weighted node relationship graph to obtain a pruned single weighted node relationship graph;

[0265] The connection weight corresponding to each node edge in the pruned single weighted node relationship graph is deleted to obtain an unweighted node relationship subgraph corresponding to each pruning threshold.

[0266] In one embodiment, the first pruning submodule may be specifically used for:

[0267] When the connection weight corresponding to the node edge is less than the pruning threshold, the node edge is deleted in the single weighted node relationship graph to obtain the deleted node edge;

[0268] When the connection weight corresponding to the node edge is not less than the pruning threshold, the node edge is retained in the single weighted node relationship graph to obtain a retained node edge;

[0269] Based on the deleted node edges and the retained node edges, a pruned single-weighted node relationship graph corresponding to each pruning threshold is determined.

[0270] In one embodiment, the encoding unit 305 includes:

[0271] A first feature encoding subunit is used for, when the weighted node relationship graph is a single weighted node relationship graph, to perform feature encoding on a corresponding node in the weighted node relationship graph according to the edge distribution information to obtain a node feature of the node;

[0272] The second feature encoding subunit is used for, when the weighted node relationship graph is a multiple weighted node relationship graph, to perform feature encoding on the corresponding nodes in at least two single weighted node relationship graphs of the weighted node relationship graph according to the edge distribution information, to obtain the single node feature of the node in the corresponding single weighted node relationship graph, and to concatenate the single node features to obtain the corresponding node feature of the node in the weighted node relationship graph.

[0273] In one embodiment, the first feature encoding subunit includes:

[0274] An edge distribution feature extraction module is used to extract the edge distribution feature corresponding to the node in each unweighted node relationship subgraph from the edge distribution information;

[0275] A combination module, used for combining the edge distribution features corresponding to the node to obtain the target edge distribution features of the node;

[0276] The node feature determination module is used to determine the node feature of the node according to the target edge distribution feature.

[0277] In one embodiment, the edge distribution feature extraction module includes:

[0278] An extraction submodule, used for extracting the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information;

[0279] The conversion submodule is used to use the edge information as an edge feature basis and convert the distribution information into an edge distribution feature based on the edge feature basis.

[0280] In one embodiment, the screening unit 304 includes:

[0281] A node edge order determination subunit, used to determine the node edge order corresponding to each node based on the segmentation parameter;

[0282] A traversal subunit is used to traverse each node in the unweighted node relationship subgraph to obtain the connection sequence between the node edges corresponding to each node;

[0283] The splicing subunit is used to screen out the node edges to be spliced ​​corresponding to the edge order of the node in each node in the unweighted node relationship subgraph, and splice the node edges to be spliced ​​according to the connection order to obtain at least one target node edge corresponding to each node.

[0284] In one embodiment, the screening unit 304 includes:

[0285] A statistical subunit, used for counting the number of target node edges corresponding to each node in the unweighted node relationship subgraph;

[0286] The edge distribution information determination subunit is used to determine the edge distribution information of the node in the unweighted node relationship subgraph based on the number of edges of the target node.

[0287] In specific implementation, the above units can be implemented as independent entities, or can be arbitrarily combined to be implemented as the same or several entities. The specific implementation of the above units can refer to the previous method embodiments, which will not be repeated here.

[0288] From the above, it can be seen that the embodiment of the present application obtains a weighted node relationship graph through an acquisition unit 301, and the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and the connection weights corresponding to the node edges; the determination unit 302 identifies the connection attribute information of the node in the weighted node relationship graph, and determines the segmentation parameters of the weighted node relationship graph based on the connection attribute information; the segmentation unit 303 segments the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph, and the unweighted node relationship subgraph is a node relationship subgraph without connection weights; the screening unit 304 screens out at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameters, and counts the edge distribution information of the node based on the target node edge; the encoding unit 305 performs feature encoding on the corresponding node in the weighted node relationship graph according to the edge distribution information to obtain the node features of the node. In this way, by dividing the weighted node relationship graph into at least one unweighted node relationship subgraph, and counting the edge distribution information of the nodes in the unweighted node relationship subgraph, and then feature encoding the nodes in the weighted node relationship graph according to the edge distribution information corresponding to each unweighted node relationship subgraph, the node features of the nodes are obtained, and the node connection relationship between the node and other nodes in the weighted node relationship graph is integrated into the node features, thereby improving the accuracy of node feature encoding in the relationship graph, and then improving the encoding efficiency of node features.

[0289] The present application also provides a computer device, such as Figure 7 As shown, it shows a schematic diagram of the structure of a computer device involved in an embodiment of the present application, and the computer device may be a server, specifically:

[0290] The computer device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will appreciate that Figure 7 The computer device structure shown in the figure does not constitute a limitation on the computer device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0291] The processor 401 is the control center of the computer device, and uses various interfaces and lines to connect various parts of the entire computer device. By running or executing software programs and / or modules stored in the memory 402, and calling data stored in the memory 402, the processor 401 performs various functions of the computer device and processes data. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 401.

[0292] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and feature coding by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 402 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0293] The computer device also includes a power supply 403 for supplying power to various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to manage charging, discharging, and power consumption through the power management system. The power supply 403 can also include any components such as one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, and power status indicators.

[0294] The computer device may further include an input unit 404, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0295] Although not shown, the computer device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 401 in the computer device will load the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402, thereby realizing various functions, as follows:

[0296] A weighted node relationship graph is obtained, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges; connection attribute information of the nodes is identified in the weighted node relationship graph, and based on the connection attribute information, a segmentation parameter of the weighted node relationship graph is determined; according to the connection weights and the segmentation parameters, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and edge distribution information of the node is counted based on the target node edge; according to the edge distribution information, feature encoding is performed on the corresponding nodes in the weighted node relationship graph to obtain node features of the nodes.

[0297] The specific implementation of each of the above operations can be found in the previous embodiments, which will not be described in detail here. It should be noted that the computer device provided in the embodiment of the present application and the method applicable to feature encoding in the above embodiment belong to the same concept, and its specific implementation process is detailed in the above method embodiment, which will not be described in detail here.

[0298] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0299] To this end, an embodiment of the present application provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any feature encoding method provided in the embodiment of the present application. For example, the instructions can execute the following steps:

[0300] A weighted node relationship graph is obtained, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges; connection attribute information of the nodes is identified in the weighted node relationship graph, and based on the connection attribute information, a segmentation parameter of the weighted node relationship graph is determined; according to the connection weights and the segmentation parameters, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, which is a node relationship subgraph without connection weights; based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and edge distribution information of the node is counted based on the target node edge; according to the edge distribution information, feature encoding is performed on the corresponding nodes in the weighted node relationship graph to obtain node features of the nodes.

[0301] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0302] Since the instructions stored in the computer-readable storage medium can execute the steps in any one of the feature encoding methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any one of the feature encoding methods provided in the embodiments of the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0303] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the various optional implementations provided in the above embodiments.

[0304] The above is a detailed introduction to a feature encoding method, device and computer-readable storage medium provided in an embodiment of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A feature encoding method, It is characterized in that include: Obtaining a weighted node relationship graph corresponding to a content recommendation application, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges, wherein objects in the content recommendation application correspond one-to-one to the nodes in the weighted node relationship graph, and the node edges and the connection weights corresponding to the node edges are determined according to association information of the objects in the content recommendation application, wherein the objects are users in the content recommendation application, and the association information includes interaction relationship information and / or interest similarity information between the users; Identifying connection attribute information of the node in the weighted node relationship graph, and determining a segmentation parameter of the weighted node relationship graph based on the connection attribute information, wherein the segmentation parameter is a relevant parameter for segmenting the weighted node relationship graph; According to the connection weights and the segmentation parameters, the weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph, where the unweighted node relationship subgraph is a node relationship subgraph without connection weights; Based on the segmentation parameters, at least one target node edge corresponding to each node is screened out in the unweighted node relationship subgraph, and edge distribution information of the node is counted based on the target node edge; According to the edge distribution information, feature encoding is performed on the corresponding node in the weighted node relationship graph to obtain the node feature of the node; Content recommendation is performed for each object in the content recommendation application based on the node features.

2. The feature encoding method according to claim 1, It is characterized in that The step of segmenting the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph includes: Determining a relationship graph type of the weighted node relationship graph according to the connection weights corresponding to the node edges; When the relationship graph type is a multiple weighted node relationship graph, the multiple weighted node relationship graph is split to obtain at least two single weighted node relationship graphs, and the single weighted node relationship graph is split according to the connection weight and the splitting parameter to obtain at least one unweighted node relationship subgraph, wherein each node edge in the single weighted node relationship graph corresponds to a connection weight; When the relationship graph type is a single weighted node relationship graph, the single weighted node relationship graph is segmented according to the connection weight and segmentation parameter to obtain at least one unweighted node relationship subgraph.

3. The feature encoding method according to claim 2, It is characterized in that The determining the relationship graph type of the weighted node relationship graph according to the connection weights corresponding to the node edges includes: Identify the number of connection weights corresponding to each node edge in the weighted node relationship graph; When there are node edges corresponding to at least two connection weights in the node edges, determining the relationship graph type of the weighted node relationship graph as a multiple weighted node relationship graph; When there is no node edge corresponding to at least two connection weights in the node edges, the relationship graph type of the weighted node relationship graph is determined to be a single weighted node relationship graph.

4. The feature encoding method according to claim 2, It is characterized in that The step of segmenting the single weighted node relationship graph according to the connection weight and the segmentation parameter to obtain at least one unweighted node relationship subgraph includes: Determine the number of pruning times of the single weighted node relationship graph according to the segmentation parameter, and obtain a pruning threshold group corresponding to the single weighted node relationship graph, wherein the pruning threshold group includes at least one pruning threshold; Based on the pruning times, the connection weights and the pruning threshold group, the single weighted node relationship graph is segmented to obtain at least one unweighted node relationship subgraph.

5. The feature encoding method according to claim 4, It is characterized in that The method of segmenting the single weighted node relationship graph based on the pruning times, the connection weights and the pruning threshold group to obtain at least one unweighted node relationship subgraph includes: When the number of pruning times is one, the single weighted node relationship graph is pruned based on the connection weight and the pruning threshold group to obtain an unweighted node relationship subgraph corresponding to each pruning threshold; When the number of pruning times is multiple, the single weighted node relationship graph is pruned based on the connection weight and the pruning threshold group to obtain a pruned unweighted node relationship subgraph, and the pruned unweighted node relationship subgraph is used as the single weighted node relationship graph, and the step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group is returned to execute until the number of pruning times is reached to obtain at least one unweighted node relationship subgraph.

6. The feature encoding method according to claim 5, It is characterized in that The step of pruning the single weighted node relationship graph based on the connection weight and the pruning threshold group to obtain an unweighted node relationship subgraph corresponding to each pruning threshold includes: Comparing at least one pruning threshold in the pruning threshold group with the connection weight corresponding to each node edge in the single weighted node relationship graph; Based on the comparison results corresponding to the node edges, pruning each node edge in the single weighted node relationship graph to obtain a pruned single weighted node relationship graph; The connection weight corresponding to each node edge in the pruned single weighted node relationship graph is deleted to obtain an unweighted node relationship subgraph corresponding to each pruning threshold.

7. The feature encoding method according to claim 6, It is characterized in that The step of pruning each node edge in the single weighted node relationship graph based on the comparison result corresponding to the node edge to obtain the pruned single weighted node relationship graph includes: When the connection weight corresponding to the node edge is less than the pruning threshold, deleting the node edge in the single weighted node relationship graph to obtain a deleted node edge; When the connection weight corresponding to the node edge is not less than the pruning threshold, the node edge is retained in the single weighted node relationship graph to obtain a retained node edge; Based on the deleted node edges and the retained node edges, a pruned single-weighted node relationship graph corresponding to each pruning threshold is determined.

8. The feature encoding method according to claim 2, It is characterized in that The step of performing feature encoding on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes includes: When the weighted node relationship graph is a single weighted node relationship graph, feature encoding is performed on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes; When the weighted node relationship graph is a multiple weighted node relationship graph, feature encoding is performed on the corresponding nodes in at least two single weighted node relationship graphs of the weighted node relationship graph according to the edge distribution information to obtain the single node features of the nodes in the corresponding single weighted node relationship graphs, and the single node features are concatenated to obtain the corresponding node features of the nodes in the weighted node relationship graphs.

9. The feature encoding method according to claim 8, It is characterized in that The step of performing feature encoding on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes includes: Extracting edge distribution features corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information; Combining the edge distribution features corresponding to the nodes to obtain the target edge distribution features of the nodes; The node feature of the node is determined according to the target edge distribution feature.

10. The feature encoding method according to claim 9, It is characterized in that Extracting the edge distribution features corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information includes: Extracting the edge information and distribution information corresponding to the node in each of the unweighted node relationship subgraphs from the edge distribution information; The edge information is used as an edge feature basis, and the distribution information is converted into an edge distribution feature based on the edge feature basis.

11. The feature encoding method according to claim 1, It is characterized in that The step of selecting at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameter includes: Based on the segmentation parameters, determining the node edge order corresponding to each node; Traversing each node in the unweighted node relationship subgraph to obtain a connection sequence between node edges corresponding to each node; Filter out the node edges to be spliced ​​corresponding to the node edge order in each node in the unweighted node relationship subgraph, and splice the node edges to be spliced ​​according to the connection order to obtain at least one target node edge corresponding to each node.

12. The feature encoding method according to claim 1, It is characterized in that The step of calculating edge distribution information of the node based on the edge statistics of the target node includes: Counting the number of target node edges corresponding to each node in the unweighted node relationship subgraph; Based on the number of edges of the target node, edge distribution information of the node in the unweighted node relationship subgraph is determined.

13. A feature encoding device, It is characterized in that include: an acquisition unit, configured to acquire a weighted node relationship graph corresponding to a content recommendation application, wherein the weighted node relationship graph includes at least two nodes, at least one node edge connecting the nodes, and connection weights corresponding to the node edges, wherein objects in the content recommendation application correspond one-to-one to the nodes in the weighted node relationship graph, and the node edges and the connection weights corresponding to the node edges are determined according to association information of the objects in the content recommendation application, wherein the objects are users in the content recommendation application, and the association information includes information on interaction relationships between the users and / or information on the similarity of interests; A determination unit, configured to identify connection attribute information of the node in the weighted node relationship graph, and determine a segmentation parameter of the weighted node relationship graph based on the connection attribute information, wherein the segmentation parameter is a relevant parameter for segmenting the weighted node relationship graph; A segmentation unit, configured to segment the weighted node relationship graph according to the connection weights and the segmentation parameters to obtain at least one unweighted node relationship subgraph, wherein the unweighted node relationship subgraph is a node relationship subgraph without connection weights; A screening unit, configured to screen out at least one target node edge corresponding to each node in the unweighted node relationship subgraph based on the segmentation parameter, and to calculate edge distribution information of the node based on the target node edge; An encoding unit, configured to perform feature encoding on corresponding nodes in the weighted node relationship graph according to the edge distribution information to obtain node features of the nodes; The recommendation unit is used to make content recommendations for each object in the content recommendation application based on the node features.

14. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the feature encoding method according to any one of claims 1 to 12.

15. A computer device, It is characterized in that The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the feature encoding method according to any one of claims 1 to 12 when executing the computer program.

16. A computer program product, It is characterized in that The computer program product comprises a computer program / instruction, and when the computer program / instruction is executed by a processor, the steps in the feature encoding method according to any one of claims 1 to 12 are implemented.

Images