WO2021179834A1 - 基于异构图进行业务处理的方法及装置 - Google Patents

基于异构图进行业务处理的方法及装置 Download PDF

Info

Publication number
WO2021179834A1
WO2021179834A1 PCT/CN2021/074248 CN2021074248W WO2021179834A1 WO 2021179834 A1 WO2021179834 A1 WO 2021179834A1 CN 2021074248 W CN2021074248 W CN 2021074248W WO 2021179834 A1 WO2021179834 A1 WO 2021179834A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
node
current
feature
entity
Prior art date
Application number
PCT/CN2021/074248
Other languages
English (en)
French (fr)
Inventor
胡斌斌
方精丽
贾全慧
张志强
周俊
方彦明
杨双红
余泉
Original Assignee
支付宝(杭州)信息技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 支付宝(杭州)信息技术有限公司 filed Critical 支付宝(杭州)信息技术有限公司
Publication of WO2021179834A1 publication Critical patent/WO2021179834A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/901Indexing; Data structures therefor; Storage structures
    • G06F16/9024Graphs; Linked lists
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Definitions

  • One or more embodiments of this specification relate to the field of computer technology, and in particular to a method and device for performing business processing based on a heterogeneous graph through a computer.
  • graph structures (relationship networks) usually have strong data description capabilities. For various entities and concepts that have association relationships in the real world, as well as their relationships, they can be described by a graph structure.
  • the original intention of the graph structure is to improve the capabilities of search engines and improve user search quality and search experience.
  • graph structures can be widely used in intelligent search, intelligent question and answer, personalized recommendation, content distribution and other fields.
  • the graph structure can be combined with the machine learning model to enhance the predictive ability of the model.
  • the intent recognition model can help the intent recognition model to perform semantic analysis and intent recognition on the questions posed by the user in natural language by introducing the graph structure related to the product and service, and then push the answer to the query according to the identified intent. user.
  • One or more embodiments of this specification describe a method and device for business processing based on heterogeneous graphs, which comprehensively characterize nodes through multiple relational networks with different structures to obtain more effective processing results.
  • a method for business processing based on a heterogeneous graph is provided, the heterogeneous graph is used to describe the association relationship between multiple entities under a plurality of predetermined connection relationship types, wherein the multiple The connection relationship types are independent of each other, the multiple connection relationship types include a first connection relationship type, and the relationship network describing the association relationship between the multiple entities under the first connection relationship type is the first relationship network,
  • the first relationship network each entity corresponds to each node one-to-one, and the two entities corresponding to the two nodes connected by the connecting edge have the association relationship of the first connection relationship type, and each node corresponds to each node.
  • the method includes: determining the current node corresponding to the current entity for the current service in the first relationship network; processing through a predetermined feature aggregation model
  • the first relationship network obtains a first service characterization vector for the current node in the first connection relationship type; and determines the first service according to the entity characteristics corresponding to the current entity in each connection relationship type.
  • a first importance coefficient corresponding to a characterization vector based at least on the first importance coefficient and the first service characterization vector, fuse each of the current entity in each relationship network corresponding to the multiple connection relationship types
  • the business characterization vector obtains a comprehensive evaluation result of the current entity, so as to use the comprehensive evaluation result to perform subsequent business processing on the current entity.
  • the neighbor nodes of the current node include a first node, the first node corresponds to a first neighbor weight, and the entity feature corresponding to the first node includes a first feature
  • the first feature corresponds to the first feature weight
  • the processing of the first relationship network through the predetermined feature aggregation model to obtain the first service characterization vector for the current node in the first connection relationship type includes: The product of the first feature weight and the first neighbor weight is determined as the first feature aggregation coefficient of the first node on the first feature; based on the first node on the first feature
  • the product of the feature expression vector of and the first feature aggregation coefficient determines the element value corresponding to the first feature vector in the first service feature vector.
  • the neighbor nodes of the current node include a second node, the second node corresponds to a second neighbor weight, and the predetermined feature aggregation model is a first graph neural network.
  • the i-th layer of the first graph neural network processes the first relational network in the following way: splicing the current feature expression vector of the current node and the current feature expression vector of the second node to obtain the first Stitching vector; based on the product of the first weight matrix and the first stitching vector, determine the neighbor weight of the second node in the i-th layer, the first weight matrix is the first graph neural network in the i-th layer
  • the model parameters of is determined when the first graph neural network is trained; the i-th neighbor of the current node is determined according to the neighbor weight of the second node in the i-th layer and the current feature expression vector of the second node Aggregation vector; fusion of the neighbor aggregation vector and the current feature expression vector of the current node to obtain the characterization
  • the current feature expression vector of the current node and the current feature expression vector of the second node are determined by the current node and the second node in the first The entity feature in a relational network is determined; when i is a natural number greater than 1, the current feature expression vector of the current node and the current feature expression vector of the second node are the current node and the The second node is the representation vector processed by the i-1th layer graph neural network.
  • the fusing the neighbor aggregation vector with the current feature expression vector of the current node to obtain the representation vector of the current node after the i-th layer graph neural network processing includes: combining the neighbor aggregation vector Splicing with the current feature expression vector of the current node to obtain a second splicing vector; based on the product of the second weight matrix and the second splicing vector, the feature weight vector in the i-th layer graph neural network is determined, and the first The two-weight matrix is the model parameter of the first graph neural network in the i-th layer, which is determined when the first graph neural network is trained; the neighbor aggregation vector is corrected according to the feature weight vector to obtain the current The representation vector of the node processed by the i-th layer graph neural network.
  • the modified representation vector is the first service representation vector.
  • the correcting the neighbor aggregation item according to the feature weight vector includes taking the product of the k-th element in the feature weight vector and the k-th element in the neighbor aggregation vector as the State the k-th element of the representation vector after the current node is processed by the i-th layer graph neural network.
  • the determining the first importance coefficient corresponding to the first service characterization vector according to the entity characteristics corresponding to the current entity in each connection relationship type includes: according to the respective connection relationship types obtained through pre-training. Corresponding attention vectors, determine that the current entity corresponds to each attention value of each connection relationship type; compare the first attention value corresponding to the current entity under the first connection relationship type with the value of each connection relationship type The ratio of the sum of the attention values is determined as the first importance coefficient.
  • the first attention value is an exponential function whose independent variable is the following value: the product of the transposition vector of the first attention vector corresponding to the first connection relationship type and the splicing vector of each service characterization vector .
  • the fusion of each service characterization vector of the current entity in each relationship network corresponding to the multiple connection relationship types is based at least on the first importance coefficient and the first service characterization vector .
  • Obtaining a comprehensive evaluation result of the current entity includes: taking each importance coefficient as the weight of the corresponding characterization vector, and determining the weighted sum of each service characterization vector, where the first importance coefficient is the first service characterization The weight of the vector; the weighted sum is used as a comprehensive evaluation result of the current entity.
  • the comprehensive evaluation result includes one of the following: a prediction score in a prediction service, and a score in each category in a classification service.
  • the multiple entities include a first entity, and each node corresponding to each connection relationship type of the first entity is associated with at least one user identifier of the first entity.
  • an apparatus for performing business processing based on a heterogeneous graph the heterogeneous graph being used to describe association relationships between multiple entities under multiple predetermined connection relationship types, wherein the The multiple connection relationship types are independent of each other, the multiple connection relationship types include a first connection relationship type, and the relationship network describing the association relationship between the multiple entities under the first connection relationship type is the first relationship network
  • each entity has a one-to-one correspondence with each node, and the two entities corresponding to the two nodes connected by the connecting edge have an association relationship of the first connection relationship type, and each node respectively Corresponding to the entity characteristics of the corresponding entity under the first connection relationship type;
  • the device includes:
  • a node determining unit configured to determine the current node corresponding to the current entity targeted by the current service in the first relational network
  • the feature aggregation unit is configured to process the first relationship network through a predetermined feature aggregation model to obtain a first service characterization vector for the current node in the first connection relationship type;
  • the importance determination unit determines the first importance coefficient corresponding to the first service characterization vector according to the entity characteristics corresponding to each connection relationship type of the current entity respectively;
  • the fusion unit is configured to fuse, based on at least the first importance coefficient and the first service characterization vector, each service characterization vector of the current entity in each relationship network corresponding to the multiple connection relationship types, to obtain The comprehensive evaluation result of the current entity is used to perform subsequent business processing on the current entity by using the comprehensive evaluation result.
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of the first aspect.
  • a computing device including a memory and a processor, characterized in that executable code is stored in the memory, and when the processor executes the executable code, the method of the first aspect is implemented .
  • heterogeneous graphs composed of multiple relational networks with different structures can be used to directly perform business processing.
  • the multiple relationship networks under different connection relationship types are processed separately to obtain the respective business representation vectors of the current entity in each relationship network, and then, according to the current business In each relationship network relative to the respective importance coefficients of the current entity, these service representation vectors are merged to obtain a comprehensive evaluation result, which can be used to perform subsequent business processing for the current entity. Due to the use of multiple relationship networks of different connection relationship types, the characteristics of entities can be more comprehensively described.
  • each relationship network is processed separately to obtain the business characterization vector.
  • each relationship network there is no need to synthesize each relationship network, which can avoid cumbersome Manual feature merging and/or extraction.
  • the importance coefficient (weight) of each relationship network under the current business can be automatically determined, and information fusion in each relationship network can be realized, thus, the evaluation result of the current entity can be improved. precise.
  • Figure 1 shows a schematic diagram of an implementation scenario of an embodiment disclosed in this specification
  • Fig. 2 shows a flow chart of a method for business processing based on a heterogeneous graph according to an embodiment
  • FIG. 3 shows a schematic diagram of a specific example of processing the first relational network to perform feature aggregation through a predetermined feature aggregation model
  • FIG. 4 shows a schematic diagram of business processing for user risk prediction based on a heterogeneous graph in a specific example
  • Fig. 5 shows a schematic block diagram of an apparatus for performing service processing based on a heterogeneous graph according to an embodiment.
  • Figure 1 shows a schematic diagram of a scenario for predicting financial risk of users based on a heterogeneous graph describing the relationship between users.
  • the heterogeneous graph can be used to describe the association relationship between multiple entities under multiple predetermined connection relationship types.
  • the connection relationship types shown in FIG. 1 are, for example, the connection relationship types of shared terminal applications (APP), the connection relationship types of financial platform transfers, the connection relationship types saved by contacts in the address book, and so on.
  • Each type of connection relationship can form an independent relationship network.
  • the shared APP is connected to the network
  • the transfer relationship is connected to the network
  • the address book is connected to the network, and so on.
  • the entities may be users.
  • each node and each user can correspond one-to-one.
  • the same user can be related to each other through the user identification (entity identification) under the corresponding connection relationship.
  • Each relationship network can exist independently of each other (each connection relationship type is independent of each other).
  • these relationship networks can also be merged to obtain a comprehensive relationship network.
  • the integration mentioned here can be understood as the merger of nodes, but the connection relationship is still diversified (the types of connection relationships are independent of each other).
  • the fusion of these relational networks can be to represent the nodes corresponding to the same user in each relational network with the same node ID (or entity ID). Since the relationship networks under each connection relationship type are always independent of each other, these multiple relationship networks can be called heterogeneous graphs.
  • the business to be processed may be to predict the financial risk of user A (for example, the risk of repaying a loan item, etc.).
  • the corresponding node can be determined in the relationship network corresponding to each connection relationship type of the heterogeneous graph, through
  • fusion is performed to obtain the risk assessment result for the user, that is, the risk score.
  • the risk score can be further used for the risk prediction service of the user. For example, if the risk score exceeds a first threshold, the user is determined to be a high-risk user, and the user is prohibited from lending business on the current financial platform.
  • connection relationship type of financial platform transfer there is no node corresponding to user A (has not participated in any platform transfer behavior), and other connection relationship types Below, both include the node corresponding to user A.
  • the processing result for user A may be empty or zero.
  • other relationship networks may correspond to higher weights, that is, rely more on others. The relationship conducts current business processing. In this way, through the comprehensiveness of heterogeneous graphs, situations such as the inability to conduct business evaluations for new users caused by a single connection relationship type are avoided.
  • Fig. 2 shows a flow of business processing based on a heterogeneous graph according to an embodiment.
  • the execution subject of the method can be any system, equipment, device, platform or server with computing and processing capabilities.
  • This method is suitable for heterogeneous graphs that describe the relationship between entities through various connection relationships. Specifically, a connection relationship can be used as a dimension, and a corresponding meta-path can be established in each dimension, corresponding to a connection relationship type.
  • connection relationship can be described through the following meta-paths: (a) user-(save)-user: user address book path, such as A’s address book contains B, It constitutes a meta-path A-save-B; (b) user-(saved)-user: the user is stored path, if A is stored in B’s address book, it constitutes a meta-path A-saved-B; (c) user-(use)-app-(used)-user: terminal application shared path, if user A and user B both use terminal application C, a meta path A-use-APP C-used-B; (d )user-(connect)-Wi-Fi-(connected)-user: The shared path of the network.
  • both user A and user B connect to the Internet through the wireless network WiFi D, it constitutes a meta-path A-connect-Wi-Fi D -connected-B; (e)user-(friend)-user: interactive path, if there is an interactive relationship between user A and user B, it constitutes a meta path A-friend-B; and so on.
  • the interactive relationship in (e) can be a relationship generated by interactive behaviors such as chatting with each other, transferring money, and sending red envelopes.
  • each path describes a single and independent connection relationship between users.
  • the acquisition of this relationship is relatively simple.
  • the user address book path and the user stored path can be determined by obtaining the address book of each user
  • the terminal application shared path can be determined by detecting the application installed on the user terminal or the user group of each terminal application
  • the network shared path can be determined by the IP address of the access network when the user exchanges information with the server
  • the interactive path can be determined by the user information request received by the server, the recorded interaction record, and other information.
  • the number of meta-paths is large, for example, terminal applications share paths. Any two users who use the same terminal application can establish a connection relationship. When the user group of the terminal application is large, the amount of data increases sharply. Therefore, according to an embodiment, meta-paths can also be sampled. For example, for terminal application C, users associated with user A can be selected to establish meta-paths in a predetermined manner, and other users do not consider the relevance to user A.
  • the predetermined method here is, for example, randomly selecting a predetermined number (for example, 5) users, or selecting a predetermined number (for example, 5) users that are geographically associated with user A, and so on.
  • users can also correspond to corresponding user characteristics.
  • the user characteristics may include the number of users in the user address book;
  • the user stored path may include the number of times the user is stored, the storage (tag) relationship type, and other characteristics;
  • the terminal application shared path may include the user used The number of terminal applications, the number of users of shared terminal applications, and other characteristics;
  • the network shared path can include the frequency of user connection to the network, the number of shared networks between two users, and the frequency of changes in the user’s connection to the network; in the interactive path, It can correspond to features such as the frequency of interaction between users and the number of interactive users.
  • the various meta-paths above correspond to various connection relationship types.
  • the association relationships of these users under various connection types together constitute a heterogeneous graph.
  • the same user may have the same user ID, for example, the unique ID of the terminal device, the registered user ID of the user in the current platform, and so on.
  • the heterogeneous graph contains the relationship between users described by multiple meta-paths, the corresponding relationship between users in various meta-paths can still be clarified because they are described by consistent user identifiers.
  • the corresponding relationship of the same user in different meta-paths can also be recorded through a table or the like.
  • connection relations described by each meta-path form a heterogeneous graph
  • the meta-paths of various connection relations can be integrated together or stored separately, which is not limited here.
  • the heterogeneous graph is exemplarily described with the user as the entity, but in practice, the heterogeneous graph can also be other entities, such as documents, keywords, web pages, etc.
  • the meta path can also be It is the corresponding various reasonable meta-paths, which describe the connection relationship of the corresponding entities. Further, the corresponding entity features in the meta-path may also be other features, which will not be repeated here.
  • the method for business processing based on a heterogeneous graph may include the following steps: step 201, determining the current node corresponding to the current entity for the current business in each relational network; step 202, passing predetermined The feature aggregation model processes each relationship network separately, and obtains each service characterization vector of the current node corresponding to each relationship network; step 203, according to the entity characteristics corresponding to each connection relationship type of the current entity, determine the corresponding service characterization vector Each importance coefficient; Step 204, based on each importance coefficient, fuse each business characterization vector of the current entity in each relationship network corresponding to multiple connection relationship types to obtain a comprehensive evaluation result of the current entity to use the comprehensive evaluation result Perform follow-up business processing for the current entity.
  • each relationship network may have a node corresponding to the current entity. These nodes may correspond to the entity identifier of the current entity, or establish a corresponding relationship with the current entity through a table.
  • the corresponding node can be determined in each relationship network.
  • the heterogeneous graph formed by the integration of multiple relational networks is a relational network of merged nodes (such as the aforementioned unified node identification in the triplet representation)
  • the current entity can only have the corresponding node under each connection relationship type one.
  • connection relationship type Taking any connection relationship type (referred to as the first connection relationship type) among various connection relationship types as an example, the any connection relationship type is called the first connection relationship type, and the corresponding relationship network is called the first relationship network.
  • the node corresponding to the current entity in the first relationship network may be referred to as the current node in the first relationship network.
  • step 202 the corresponding relationship network is processed through each pre-trained graph neural network, and each service characterization vector corresponding to each connection relationship type of the current node is obtained.
  • the current entity not only corresponds to the corresponding node, but also corresponds to the entity characteristics under the corresponding connection relationship.
  • these entity characteristics can be represented by symbols, such as the value corresponding to the frequency of transfers, the value corresponding to the frequency of jumps between pages, and so on.
  • these multiple features can also be represented by entity vectors.
  • each node may correspond to a corresponding characterization vector.
  • Each dimension of this representation vector may be an expression of a feature that determines the meaning, or it may be a vector expression with no definite meaning in each dimension.
  • This kind of characterization vector can also be referred to as the node's feature expression vector.
  • Each node can have an initial feature expression vector (or called an initial characterization vector).
  • the initial feature expression vector can be directly determined according to the entity feature.
  • the entity feature corresponding to a certain dimension is the frequency of transfers between users, and a number positively related to the corresponding actual transfer frequency can be used as the value of this dimension in the initial feature expression vector.
  • the initial feature expression vector of each node can be determined according to the training sample training graph neural network, and the graph nerve Other model parameters of the network will not be repeated here.
  • a business representation vector can be a representation vector used to reflect business characteristics in a specific business. For example, in a classification service that relies on a single relational network, a certain node's service characterization vector can be mapped to the probability of the node in each classification category through an incentive function or the like.
  • the feature aggregation method of neighbor nodes is usually used to aggregate the feature expression vector of the previous layer of the current node and the feature expression vector of the neighbor node to obtain the feature expression vector of the current layer as the output of the current layer .
  • this feature aggregation method is described by a feature aggregation model.
  • the feature aggregation model may be a model with a preset aggregation method (for example, a feature weighting method, etc.), or a graph neural network model.
  • a node and its neighboring nodes usually have different degrees of association.
  • the degree of association is different, the impact on the current node is also different.
  • high-order nodes have less influence on the current node than low-order nodes, and neighbor nodes with higher transfer frequency have less influence on the current node than neighbor nodes with lower transfer frequency. Therefore, according to a possible design, each neighbor node can correspond to neighbor importance (neighbor weight), which is used to describe the importance of each neighbor node relative to the current node.
  • these neighbor weights can be used as parameters of the feature aggregation model, which can be adjusted and determined according to the sample features corresponding to the sample entity and the pre-labeled sample business results.
  • different neighbor weights can also be determined for each neighbor node. For example, in a first-order neighbor node, the neighbor weight is positively correlated with the frequency of mutual transfer between the neighbor node and the current node.
  • the feature aggregation model is a graph convolutional neural network (hereinafter also referred to as a graph neural network), and each neighbor node corresponds to a different neighbor importance coefficient.
  • the importance of each neighbor node can be determined by the feature expression vector of the current node and neighbor nodes.
  • the convolution operator for determining the feature expression of the node v of the l+1th layer can be:
  • H l+1 (v) is the feature expression vector of node v in the l+1 layer of the graph convolutional neural network
  • N(v) is the neighbor node of node v
  • d v , du are normalization factors , For example, the degree of the corresponding node, that is, the number of connected edges connected to the corresponding node, or the number of first-order neighbor nodes
  • H l (v) is the feature expression vector of node v in the first layer of the graph convolutional neural network
  • H l (u) is the feature expression vector of the node u in the first layer of the graph convolutional neural network
  • W l is the model parameter of the first layer of the corresponding node graph convolutional neural network.
  • W l can be a model parameter in the form of a matrix, which can be called a weight matrix.
  • the formula can also consider the feature aggregation of higher-order neighbor nodes of the current node, which is represented by an ellipsis here. The principle is similar to the feature aggregation of first-order neighbor nodes, and will not be repeated here. Among them, different neighbor nodes have different normalization factors and different feature expression vectors, so the product multiplied by the weight matrix is also different, so they have different neighbor weights.
  • the feature expression vector can be a vector composed of the values corresponding to each entity feature. If each entity feature corresponds to a vector, the feature expression vector can be a vector mosaic corresponding to each entity feature. The resulting vector.
  • the initial feature expression vector of each node may be predetermined.
  • the model parameters (such as weight matrix) are adjusted according to the training samples.
  • the corresponding graph neural network used to process the relational network can perform feature aggregation on the nodes corresponding to the current entity to obtain the corresponding representation vector.
  • neighbor nodes within a predetermined order (such as order 2) of the node corresponding to the current entity can be used as feature aggregation nodes, or neighbors within a predetermined order can be sampled, and the sampled Neighbor nodes do feature aggregation.
  • the way of feature aggregation may be, for example, addition, average, maximum value, weighted sum, etc., which are not limited here.
  • each layer of graph neural network can also be many.
  • each group of model parameters is a weight matrix
  • a layer of graph neural network can correspond to multiple weight matrices.
  • the model parameters can be determined through parameter adjustment during the training process.
  • the current node is node ⁇
  • the neighbor weight of neighbor node j can be:
  • matrix V (for example, called the first auxiliary matrix) and W 1 (for example, called the first weight matrix) are the model parameters determined during the training process of the graph neural network
  • b 1 is the constant parameter determined during the training process of the graph neural network
  • X u and X j are the current feature expression vectors corresponding to node ⁇ and node j, respectively
  • X j ] represents the splicing vector of the two vectors.
  • the activation functions softmax and tanh can also be replaced by other activation functions (such as Relu, etc.), which are not limited here.
  • the corresponding neighbor weight can be determined for each neighbor node.
  • the neighbor weight for the corresponding neighbor node is also different. It is worth noting that in the processing of the relational network by the graph neural network, the current node can also be regarded as its own neighbor node, for example, it is called a zero-order neighbor node.
  • the feature aggregation of each neighbor node can be performed by methods such as weighted sum.
  • N u denotes the set of neighbors of the current node
  • the current through the neighbor node layer neural network of FIG polymerization results are:
  • an aggregation result of the current layer can be obtained, for example, the aggregation result of node j (also referred to as a representation vector) is h j .
  • the current feature expression vector of each node is determined by the node feature of the corresponding node.
  • the above neighbor aggregation results can be further integrated with the feature expression vector of the current node to obtain the aggregation result of the current node in the current layer of the graph neural network.
  • FIG. 3 In Figure 3, assuming that the graph neural network is a multi-layer network, node 1, node 2, node 3...
  • the current feature expression vector of the corresponding node is the feature aggregation result of the i-1th layer (that is, the characterization vector output by the i-1th layer), as shown in Figure 3
  • a characterization vector (such as the first characterization vector) corresponding to the current node can be obtained.
  • each feature may also have a feature importance (feature weight).
  • the feature weight may be preset or obtained through training. For example, in a relationship network describing the transfer relationship between users, the feature weight of the initial transfer frequency is greater than the feature weight of the transfer amount. For example, when determining the characterization vector, it is specific to a certain node, such as the first node corresponding to the first neighbor weight, and its corresponding first feature has the first feature weight, and the first node corresponds to the first feature
  • the first feature aggregation coefficient may be the product of the first feature weight and the first neighbor weight.
  • the feature expression (such as a value or a vector) corresponding to the first feature can be multiplied by the first feature aggregation coefficient, and the resulting product can be used as the corresponding weighting item, and each neighbor node can be placed on the first feature.
  • the weighted items of is added to obtain the feature value of the current node on the first feature after neighbor feature aggregation.
  • the first representation vector is determined.
  • the graph neural network can be trained to obtain general parameters related to the importance of features in the process of processing the relational network.
  • the feature weight vector composed of the feature weights corresponding to each feature can be determined in the following way:
  • W 2 (for example, called the second weight matrix) and W 3 (for example, called the second auxiliary matrix) are the weight matrices of the i-th layer in the graph neural network, and b 2 and b 3 are constant parameters.
  • W 2 , W 3 , b 2 , and b 3 can be used as general parameters. Represents the splicing of two vectors.
  • the excitation function Relu can also be replaced by other suitable excitation functions, which will not be repeated here.
  • Each element in the feature weight vector ⁇ corresponds to the feature weight of each feature.
  • the feature aggregation result of the current node u in the current layer can be obtained.
  • the way to determine the final aggregation result according to the feature weight can be expressed as:
  • means multiplying the corresponding elements of two matrices (such as Hadamard product).
  • the k-th element in ⁇ is the same as The kth element in as the aggregation result The kth element in.
  • the result of the vector (A, B, C) ⁇ (a, b, c) is (Aa, Bb, Cc).
  • the node contribution degree and feature contribution degree can be considered at the same time, and a more accurate feature aggregation result of neighbor nodes can be obtained.
  • the feature aggregation model is a graph neural network
  • the aggregation result obtained in the last layer is the business representation vector corresponding to the current node and the current relational network.
  • each relationship network that characterizes each connection relationship type, it is possible to perform feature aggregation on neighbor nodes of the corresponding node for the current entity to obtain each service characterization vector of the current entity under each connection relationship type. For example, in the first relational network, the first service characterization vector is obtained.
  • each importance coefficient corresponding to each service characterization vector is determined. It can be understood that, for specific services, entity characteristics under different connection relationships have different importance. For example, in a user risk prediction service, a relationship network whose connection relationship type is an interaction relationship between users is more important, while a relationship network whose connection relationship type is a terminal application public network is less important.
  • the importance coefficient of the relationship network can be preset based on experience.
  • the importance coefficient of the relationship network describing the interaction relationship between users is 0.5
  • the importance of the relationship network of the terminal application public network is 0.1.
  • the importance coefficient of the relational network can be used as the model parameter of the graph neural network and determined by training with sample data.
  • the importance coefficient can describe the preference of each meta-path (connection relationship) in the current business process.
  • the attention value can be used to reflect this preference.
  • the attention value of the current entity on one of the relational networks can be determined in the following way:
  • Z ⁇ is the attention vector under the relational network ⁇ (it can be determined by training with sample data)
  • P is the set of attention vectors corresponding to all relational networks.
  • the respective attention vectors corresponding to each connection relationship type obtained by pre-training it is possible to determine the respective attention values of the current entity corresponding to each connection relationship type, and then place the current entity in The ratio of the first attention value corresponding to the first connection relationship type to the sum of the attention values of each connection relationship type is determined as the first importance coefficient corresponding to the current entity.
  • the first attention value corresponding to the first relationship network is an exponential function whose independent variable is the following value: the transpose vector of the first attention vector corresponding to the first connection relationship type, and The product of the splicing vectors of each characterization vector.
  • the first importance coefficient is the ratio of the first attention value to the sum of each attention value corresponding to each relationship network.
  • Z ⁇ can be a model parameter, which can be determined by adjusting the sample data during the graph neural network training process. It can be specific to determine the splicing vector of the representation vectors of the current entity in step 202, according to With each Z ⁇ , the importance coefficient of the current entity under different relational networks can be determined.
  • the comprehensive evaluation result is a business result used to evaluate the current entity in a specific business.
  • the comprehensive evaluation result can be the prediction score for the current entity
  • the comprehensive evaluation result can be the accuracy of target recognition
  • the comprehensive evaluation result can be the expected score. Push the information and the user's degree of interest.
  • the comprehensive evaluation result can be the score on each category and so on.
  • the importance coefficients for the current entity in each relational network can be used as weights, and each service characterization vector obtained in step 203 can be weighted and summed, and the obtained sum can be used as a comprehensive evaluation result of the current entity, or The obtained sum value is further processed to obtain a comprehensive evaluation result of the current entity.
  • the service characterization vector corresponding to the relational network with the largest importance coefficient for the current entity, or the result obtained by further processing the service characterization vector may be used as the comprehensive evaluation result of the current entity.
  • the further processing here may be, for example, scoring on specific businesses (such as the risk of repayment on a financial platform, etc.).
  • each relationship network may have multiple importance coefficients for the current entity, corresponding to each classification category. That is, the importance coefficient of each relational network for the current entity can include the importance coefficients of each classification category. Then, through the fully connected layer, each service characterization vector is used as the input of the fully connected layer, and the corresponding importance coefficient is used as the corresponding weight. The current entity is scored on each candidate category, and each scoring result is obtained to perform category prediction.
  • Fig. 4 shows a specific example of a schematic diagram of business processing for user risk judgment based on a heterogeneous graph.
  • the heterogeneous graph includes a relationship network describing different connection relationships (metapaths) of N users.
  • the current business requirement is to predict the risk of user n in the financial lending field (such as the probability of default).
  • the pre-trained graph neural network is used to analyze the heterogeneous graphs.
  • Each relationship network is processed to obtain a comprehensive vector representation for user n, that is, each service representation vector.
  • each comprehensive vector characterization is used as the input of each neuron of the fully connected neural network, and each importance coefficient is used as the weight of the corresponding neuron, and each service characterization vector is merged to obtain a comprehensive evaluation result of user n (such as risk score).
  • a comprehensive evaluation result of user n such as risk score
  • the risk of user n in the financial lending field can be output. If the risk score is higher than the risk threshold, the result of high-risk users can be output. According to the result, subsequent services can be performed, such as limiting the loan amount of the user n, prohibiting the user n from performing the loan business, and so on.
  • each relationship network is processed separately to obtain each business characterization vector, and there is no need to synthesize each relationship network, which can avoid cumbersome Further, it can automatically determine the importance coefficient (weight) of the current entity in each relational network under the current business, and realize the information fusion under each relational network, thereby making the evaluation result of the current entity more precise.
  • Fig. 5 shows a schematic block diagram of a service processing apparatus based on a heterogeneous graph according to an embodiment.
  • the heterogeneous graph is used to describe the association relationship between multiple entities under multiple predetermined connection relationship types, where the multiple connection relationship types are independent of each other, and the multiple connection relationship types include the first connection relationship type.
  • each entity corresponds to each node one-to-one, and the two entities corresponding to the two nodes connected by the connection edge have a connection relationship of the first connection relationship type, and each node Corresponding to the entity characteristics of the corresponding entity under the first connection relationship type.
  • the service processing apparatus 500 based on the heterogeneous graph includes: a node determining unit 51 configured to determine that the current entity targeted by the current service corresponds in the first relational network
  • the feature aggregation unit 52 is configured to process the first relationship network through a predetermined feature aggregation model to obtain the first service characterization vector for the current node in the first connection relationship type; the importance determination unit 53, according to the current
  • the entity characteristics corresponding to each connection relationship type of the entity respectively determine the first importance coefficient corresponding to the first business characterization vector;
  • the fusion unit 54 is configured to merge the current entity based on at least the first importance coefficient and the first business characterization vector For each service characterization vector under each relationship network corresponding to multiple connection relationship types, a comprehensive evaluation result of the current entity is obtained, so as to use the comprehensive evaluation result to perform subsequent business processing on the current entity.
  • the neighbor nodes of the current node include the first node, the first node corresponds to the first neighbor weight, the entity feature corresponding to the first node includes the first feature, and the first feature corresponds to the first feature.
  • a feature weight, the feature aggregation unit 52 is further configured to: determine the product of the first feature weight and the first neighbor weight as the first feature aggregation coefficient of the first node on the first feature; The product of the feature expression vector and the first feature aggregation coefficient determines the element value corresponding to the first feature vector in the first service feature vector.
  • the feature The aggregation unit 52 may also be configured to use the i-th layer of the first graph neural network to process the first relational network in the following manner: splicing the current feature expression vector of the current node and the current feature expression vector of the second node to obtain the first splicing Vector; Based on the product of the first weight matrix and the first splicing vector, determine the neighbor weight of the second node in the i-th layer.
  • the first weight matrix is the model parameter of the first graph neural network in the i-th layer.
  • the current feature expression vector of the current node and the current feature expression vector of the second node are respectively determined by the entity features of the current node and the second node in the first relationship network ;
  • the current feature expression vector of the current node and the current feature expression vector of the second node are the representations of the current node and the second node after being processed by the i-1th layer graph neural network, respectively vector.
  • the feature aggregation unit 52 is further configured to fuse the aforementioned neighbor aggregation vector with the current feature expression vector of the current node in the following manner to obtain the representation vector of the current node after the i-th layer graph neural network is processed:
  • the neighbor aggregation vector and the current feature expression vector of the current node are spliced to obtain the second splicing vector; based on the product of the second weight matrix and the second splicing vector, the characteristic weight vector in the i-th layer graph neural network is determined, and the second weight
  • the matrix is the model parameter of the first graph neural network in the i-th layer, which is determined when the first graph neural network is trained; the neighbor aggregation vector is corrected according to the feature weight vector to obtain the representation of the current node after the i-th layer graph neural network is processed
  • Vector when the i-th layer graph neural network is the last layer of the first graph neural network, the characterization vector obtained after correction is the first service characterization vector.
  • the feature aggregation unit 52 may be further configured to take the product of the k-th element in the feature weight vector and the k-th element in the neighbor aggregation vector as the current node through the i-th layer graph neural network processing The k-th element of the subsequent characterization vector, so as to correct the neighbor aggregation item according to the feature weight vector.
  • the importance determination unit 53 is further configured to: determine the attention values of the current entity corresponding to the respective connection relationship types according to the respective attention vectors corresponding to the respective connection relationship types obtained by pre-training; The ratio of the first attention value corresponding to the first connection relationship type to the sum of the attention values of each connection relationship type is determined as the first importance coefficient.
  • the first attention value is an exponential function whose independent variable is the following value: the transposition vector of the first attention vector corresponding to the first connection relationship type and the splicing vector of each service characterization vector product.
  • the fusion unit 54 is further configured to: use each importance coefficient as the weight of the corresponding characterization vector to determine the weighted sum of each characterization vector, where the first importance coefficient is the weight of the first characterization vector; And as the result of a comprehensive evaluation of the current entity.
  • the comprehensive evaluation result includes one of the following: the prediction score in the prediction service, and the score in each category in the classification service.
  • the above-mentioned multiple entities include a first entity, and under each connection relationship type, each node corresponding to the first entity is associated with a user identifier of the first entity under each connection relationship type.
  • the node corresponding to the same entity is represented by the same node identifier, or the relationship network corresponding to each connection relationship type is recorded in a table, and the node identifier corresponding to the same entity Correspondence.
  • the apparatus 500 shown in FIG. 5 is an apparatus embodiment corresponding to the method embodiment shown in FIG. 2, and the corresponding description in the method embodiment shown in FIG. 2 is also applicable to the apparatus 500. Go into details again.
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method described in conjunction with FIG. 2.
  • a computing device including a memory and a processor, the memory is stored with executable code, and when the processor executes the executable code, it implements the method described in conjunction with FIG. 2 method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种基于异构图进行业务处理的方法和装置,可以利用不同结构的关系网络构成的异构图直接进行业务处理。利用多个不同连接关系类型的关系网络,可以更加全面的刻画实体的特征,另一方面,针对各个关系网络分别处理得到节点的各个业务表征向量,无需对各个关系网络进行综合,可以避免繁琐的手工特征抽取,进一步地,可以自动确定在当前业务下,当前实体在每个关系网络中的重要度系数,实现在各个关系网络下的信息融合,从而使得对当前实体的评估结果更加准确。

Description

基于异构图进行业务处理的方法及装置 技术领域
本说明书一个或多个实施例涉及计算机技术领域,尤其涉及通过计算机基于异构图进行业务处理的方法和装置。
背景技术
在关系描述方面,图结构(关系网络)通常具有较强的数据描述能力。对于真实世界中存在关联关系的各种实体和概念,以及他们之间的关系,可以通过图结构来描述。图结构的初衷是为了提高搜索引擎的能力,改善用户的搜索质量以及搜索体验。随着人工智能的技术发展和应用,图结构可以广泛应用于智能搜索、智能问答、个性化推荐、内容分发等领域。特别是,可以将图结构与机器学习模型相结合,增强模型的预测能力。例如,在客服机器人问答系统中,可以通过引入与产品服务相关的图结构,帮助意图识别模型对用户使用自然语言提出的问题进行语义分析和意图识别,之后根据识别到的意图查询到答案推送给用户。
在互联网数据日益庞大的情况下,平台用户的交互场景、交互关系错综复杂。例如,通讯录上的存和被存关系、终端应用关联信息、网络链路重叠信息,等等,这就造成了数据多源的特性。对于这样的网络数据,综合建立一个综合的图结构(如综合性的知识图谱)的过程较复杂。因此,希望能有更有效的方案,将多个网络之间的信息综合利用,提高基于图结构的业务处理能力。
发明内容
本说明书一个或多个实施例描述了一种基于异构图进行业务处理的方法及装置,通过不同结构的多个关系网络综合表征节点,得到更有效的处理结果。
根据第一方面,提供了一种基于异构图进行业务处理的方法,所述异构图用于描述多个实体之间在预定的多个连接关系类型下的关联关系,其中,所述多个连接关系类型相互独立,所述多个连接关系类型包括第一连接关系类型,描述所述多个实体之间在所述第一连接关系类型下的关联关系的关系网络为第一关系网络,在所述第一关系网络中,各个实体分别与各个节点一一对应,通过连接边连接的两两节点对应的两两实体之间具有所述第一连接关系类型的关联关系,各个节点分别对应有相应实体在所述第一连 接关系类型下的实体特征;所述方法包括:确定当前业务所针对的当前实体在所述第一关系网络中对应的当前节点;通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量;根据所述当前实体在各个连接关系类型下分别对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数;至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果,以利用所述综合评估结果针对所述当前实体进行后续业务处理。
根据一个实施例,在所述第一关系网络中,所述当前节点的邻居节点包括第一节点,所述第一节点对应第一邻居权重,所述第一节点对应的实体特征包括第一特征,所述第一特征对应第一特征权重,所述通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量包括:将所述第一特征权重与所述第一邻居权重的乘积确定为所述第一节点在所述第一特征上的第一特征聚合系数;基于所述第一节点在所述第一特征上的特征表达向量与所述第一特征聚合系数的乘积,确定所述第一业务表征向量中与所述第一特征向量相对应的元素值。
根据一个实施例,在所述第一关系网络中,所述当前节点的邻居节点包括第二节点,所述第二节点对应第二邻居权重,所述预先确定的特征聚合模型为第一图神经网络;所述第一图神经网络的第i层通过以下方式处理所述第一关系网络:将所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量拼接,得到第一拼接向量;基于第一权重矩阵与所述第一拼接向量的乘积,确定所述第二节点在第i层的邻居权重,所述第一权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;根据所述第二节点在第i层的邻居权重与所述第二节点的当前特征表达向量,确定所述当前节点的第i邻居聚合向量;将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量。
根据一个实施例,在i为1的情况下,所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量,分别由所述当前节点和所述第二节点在所述第一关系网络中的实体特征确定;在i为大于1的自然数的情况下,所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量,分别为所述当前节点和所述第二节点经过第i-1层图神经网络处理后的表征向量。
根据一个实施例,所述将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量包括:将所述邻居聚 合向量与所述当前节点的当前特征表达向量进行拼接,得到第二拼接向量;基于第二权重矩阵与所述第二拼接向量的乘积,确定第i层图神经网络中的特征权重向量,所述第二权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;根据所述特征权重向量对所述邻居聚合向量进行修正,得到所述当前节点经过第i层图神经网络处理后的表征向量,在第i层图神经网络为所述第一图神经网络的最后一层时,修正后得到的表征向量为所述第一业务表征向量。
根据一个实施例,所述根据所述特征权重向量对所述邻居聚合项进行修正包括,将所述特征权重向量中第k个元素与所述邻居聚合向量中第k个元素的乘积,作为所述当前节点经过第i层图神经网络处理后的表征向量的第k个元素。
根据一个实施例,所述根据所述当前实体在各个连接关系类型下对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数包括:根据预先训练得到的各个连接关系类型分别对应的各个注意力向量,确定所述当前实体分别对应于各个连接关系类型的各个注意力值;将当前实体在所述第一连接关系类型下对应的第一注意力值与各个连接关系类型的注意力值之和的比值,确定为所述第一重要度系数。
根据一个实施例,所述第一注意力值为,自变量为以下值的指数函数:第一连接关系类型对应的第一注意力向量的转置向量,与各个业务表征向量的拼接向量的乘积。
根据一个实施例,所述至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果包括:将各个重要度系数作为相应表征向量的权重,确定各个业务表征向量的加权和,其中,所述第一重要度系数为所述第一业务表征向量的权重;将所述加权和作为对所述当前实体的综合评估结果。
根据一个实施例,所述综合评估结果包括以下中的一种:预测业务中的预测分数、分类业务中在各个类别上的评分。
根据一个实施例,所述多个实体包括第一实体,所述第一实体在各个连接关系类型下分别对应的各个节点通过所述第一实体的至少一个用户标识相关联。
根据第二方面,提供了一种基于异构图进行业务处理的装置,所述异构图用于描述多个实体之间分别在预定的多个连接关系类型下的关联关系,其中,所述多个连接关系类型相互独立,所述多个连接关系类型包括第一连接关系类型,描述所述多个实体之间在所述第一连接关系类型下的关联关系的关系网络为第一关系网络,在所述第一关系 网络中,各个实体分别与各个节点一一对应,通过连接边连接的两两节点对应的两两实体之间具有所述第一连接关系类型的关联关系,各个节点分别对应有相应实体在所述第一连接关系类型下的实体特征;所述装置包括:
节点确定单元,配置为确定当前业务所针对的当前实体在所述第一关系网络中对应的当前节点;
特征聚合单元,配置为通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量;
重要度确定单元,根据所述当前实体在各个连接关系类型下分别对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数;
融合单元,配置为至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果,以利用所述综合评估结果针对所述当前实体进行后续业务处理。
根据第三方面,提供了一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行第一方面的方法。
根据第四方面,提供了一种计算设备,包括存储器和处理器,其特征在于,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现第一方面的方法。
通过本说明书实施例提供的基于异构图进行业务处理的方法及装置,可以利用不同结构的多个关系网络构成的异构图直接进行业务处理。具体地,对多个用户之间的异构图,先针对不同的连接关系类型下的多个关系网络分别进行处理,得到当前实体在各个关系网络中各自的业务表征向量,然后,根据当前业务中,各个关系网络相对于当前实体分别对应的重要度系数,对这些业务表征向量进行融合,从而得到一个综合评估结果,以利用综合评估结果针对当前实体进行后续业务处理。由于利用了多个不同连接关系类型的关系网络,可以更加全面的刻画实体的特征,另一方面,先针对各个关系网络分别处理得到业务表征向量,无需对各个关系网络进行综合,可以避免繁琐的手工特征合并和/或抽取,另外,可以自动确定当前业务下,每个关系网络中的重要度系数(权重),实现在各个关系网络的信息融合,从而,可以使得对当前实体的评估结果更加准确。
附图说明
为了更清楚地说明本申请实施例的技术方案,下面将对实施例描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1示出本说明书披露的一个实施例的实施场景示意图;
图2示出根据一个实施例的基于异构图进行业务处理的方法流程图;
图3示出一个具体例子的通过预先确定的特征聚合模型处理第一关系网络进行特征聚合的示意图;
图4示出在一个具体例子的基于异构图进行用户风险性预测的业务处理的示意图;
图5示出根据一个实施例的基于异构图进行业务处理的装置的示意性框图。
具体实施方式
下面结合附图,对本说明书提供的方案进行描述。
为了便于说明,结合图1示出的本说明书实施例的一个具体适用场景进行描述。图1示出了一个基于描述用户之间的关系的异构图,对用户进行金融风险性预测的场景示意图。
如图1所示,异构图可以异构图用于描述多个实体之间分别在预定的多个连接关系类型下的关联关系。图1示出的连接关系类型例如是:共用终端应用(APP)的连接关系类型、金融平台转账的连接关系类型、通讯录联系人保存的连接关系类型,等等。每种连接关系类型都可以构成一个独立的关系网络。如图1示出的共用APP连接网络、转账关系连接网络、通讯录连接网络,等等。
在图1的各种关系网络中,实体可以是用户。对于某个关系网络来说,各个节点和各个用户可以一一对应。在不同的关系网络中,同一用户可以通过相应连接关系下的用户标识(实体标识)相互关联。各个关系网络可以相互独立存在(各个连接关系类型相互独立)。在可选的实现中,这些关系网络也可以进行融合,得到一个综合的关系网络。这里说的融合,可以理解为节点合并,但连接关系仍然呈现多元化(各个连接关系类型相互独立)。在三元组描述的图数据中,对这些关系网络进行融合可以是将各个关 系网络中对应同一用户的节点用同一个节点标识(或实体标识)表示。由于各个连接关系类型下的关系网络始终相互独立,这多个关系网络可以称为异构图。
在图1示出的应用场景下,待处理的业务可以是预测用户A的金融风险性(例如偿还借贷款项的风险性等)。根据本说明书的技术构思,在确定待预测风险性的用户(如可以是获取其用户标识)后,在异构图的各个连接关系类型分别对应的关系网络中,都可以确定相应的节点,通过对各个关系网络中的节点数据分割处理后,再进行融合,得到针对该用户的风险评估结果,即风险分数。该风险分数可以进一步用于该用户的风险预测业务,例如风险分数超过第一阈值,判定该用户为高风险用户,禁止其在当前金融平台进行借贷业务。
可以理解的是,在各个关系网络中,也可能出现不一致的节点情况,例如金融平台转账的连接关系类型中,没有用户A对应的节点(未参与过任何平台转账行为),而其他连接关系类型下,都包含用户A对应的节点。这样,在金融平台转账的连接关系类型对应的关系网络中,针对用户A的处理结果可能为空,或者零值,此时,其他关系网络可能对应较高权重,亦即,更多地依赖其他关系进行当前业务处理。如此,通过异构图的全面性,避免单一连接关系类型导致的无法对新用户进行业务评估等情况。
下面详细描述本说明书的技术构思。
图2示出一个实施例的基于异构图进行业务处理的流程。该方法的执行主体可以是任何具有计算、处理能力的系统、设备、装置、平台或服务器。该方法适用于通过各种连接关系描述实体之间关系的异构图。具体地,一种连接关系可以作为一个维度,在每个维度可以建立相应的元路径,对应一种连接关系类型。例如,对于用户在金融领域的风险性评估,可以通过以下几种元路径对其连接关系进行描述:(a)user-(save)-user:用户通讯录路径,如A的通讯录包含B,则构成一条元路径A-save-B;(b)user-(saved)-user:用户被存储路径,如A被B的通讯录存储,则构成一条元路径A-saved-B;(c)user-(use)-app-(used)-user:终端应用共用路径,如用户A和用户B都使用了终端应用C,则构成一条元路径A-use-APP C-used-B;(d)user-(connect)-Wi-Fi-(connected)-user:网络共用路径,如用户A和用户B都通过无线网络WiFi D接入互联网,则构成一条元路径A-connect-Wi-Fi D-connected-B;(e)user-(friend)-user:交互路径,如用户A和用户B之间具有交互关系,则构成一条元路径A-friend-B;等等。
其中,(e)中的交互关系可以是相互聊天、具有转账、发红包等交互行为产生的 联系。
可以看出,在以上路径中,各条路径分别描述用户之间单一、独立的连接关系。这种关系的获取较简单,如用户通讯录路径及用户被存储路径,可以通过获取各个用户的通讯录确定,终端应用共用路径可以通过检测用户终端安装的应用,或者各个终端应用的用户群体确定,网络共用路径可以通过用户向服务端进行信息交互时的接入网络的IP地址等确定,交互路径可以通过服务端接收到的用户信息请求、记载的交互记录等信息确定。
在一些情况下,元路径数量较大,例如终端应用共用路径,任意使用同一终端应用的两两用户之间,都可以建立连接关系,当终端应用的用户群较大时,数据量剧增。因此,根据一个实施例,还可以对元路径进行采样,例如对终端应用C,可以通过预定方式选择与用户A的相关联的用户建立元路径,其他用户不考虑与用户A的关联性。这里的预定方式例如是随机选择预定数量(如5个)的用户,或者选择与用户A在地理位置上关联较大的预定数量(如5个)的用户,等等。
每种元路径上,用户还可以对应有相应的用户特征。如:用户通讯录路径中,用户特征可以包括用户通讯录人数等特征;用户被存储路径可以包括用户被存储的次数、存储(标记)关系类型等特征;终端应用共用路径可以包括用户所使用的终端应用数量、所共用的终端应用的使用人数等特征;网络共用路径中,可以包括用户连接网络的频率、两个用户之间共用网络次数、用户连接网络的变更频率等特征;交互路径中,可以对应有用户之间的交互频次、用户的交互用户数量等特征。
以上的各种元路径,分别对应各种连接关系类型。这些用户在各种连接关系类型下的关联关系,共同构成异构图。值得说明的是,在各个路径中,同一个用户可以具有相同的用户标识,例如通过终端设备唯一身份标识、用户在当前平台中的注册用户名(user ID),等等。这样,虽然异构图包含多个元路径描述的用户之间的关系,但由于通过一致的用户标识进行描述,仍然可以明确用户在各种元路径中的对应关系。当同一用户通过不同的标识描述时,还可以通过表格等记录同一用户在不同元路径中的对应关系。
在各个元路径描述的连接关系构成异构图时,可以将各种连接关系的元路径综合到一起,也可以分别存储,在此不作限定。
上述示例中,以用户为实体对异构图进行了示例性描述,但实践中,异构图中还 可以是其他实体,如文档、关键词、网络页面等等,相应地,元路径也可以是相应的各种合理元路径,描述相应实体的连接关系,进一步地,元路径中对应的实体特征也可以是其他特征,在此不再赘述。
如图2所示,该基于异构图进行业务处理方法可以包括以下步骤:步骤201,确定当前业务所针对的当前实体在各个关系网络中分别对应的当前节点;步骤202,通过预先确定的各个特征聚合模型分别处理各个关系网络,得到当前节点分别对应于各个关系网络的各个业务表征向量;步骤203,根据当前实体在各个连接关系类型下分别对应的实体特征,确定各个业务表征向量分别对应的各个重要度系数;步骤204,基于各个重要度系数,融合当前实体在多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对当前实体的综合评估结果,以利用综合评估结果针对当前实体进行后续业务处理。
首先,在步骤201中,确定当前业务所针对的当前实体在各个关系网络中对应的当前节点。可以理解,各个关系网络中,都可以具有与当前实体对应的节点。这些节点可以对应有当前实体的实体标识,或者通过表格与当前实体建立有对应关系。当确定了待进行业务处理的当前实体后,可以在各个关系网络中确定相应节点。当多个关系网络综合在一起形成的异构图为合并节点的关系网络(如前述的在三元组表示方式中统一节点标识)时,当前实体在各个连接关系类型下对应的节点可以仅有一个。
以各种连接关系类型中的任一连接关系类型(称为第一连接关系类型)为例,将该任一连接关系类型称为第一连接关系类型,相应的关系网络称为第一关系网络,可以将第一关系网络中与当前实体对应的节点称为第一关系网络中的当前节点。
接着,在步骤202中,分别通过预先训练的各个图神经网络处理相应关系网络,得到当前节点在各个连接关系类型下分别对应的各个业务表征向量。在异构图的各个连接关系分别对应的各个关系网络中,当前实体不仅对应有相应节点,还对应有在相应连接关系下的实体特征。计算机处理过程中,可以将这些实体特征通过符号进行表示,如转账频次对应的数值、页面间的跳转频次对应的数值等。当一个关系网络中,实体特征有多个时,还可以通过实体向量表示这多个特征。
可以理解,在关系网络中,各个节点可以分别对应有相应的表征向量。这种表征向量可能每一维都是对确定含义的特征的表达,也可能是每一维都没有确定含义的一种向量表达。这种表征向量也可以称为节点的特征表达向量。每个节点都可以有初始的特征表达向量(或者称为初始的表征向量)。在初始的特征表达向量每一维都是对确定含 义的特征的表达的情况下,可以根据实体特征直接确定初始的特征表达向量。例如,某个维度对应的实体特征为用户之间的转账频次,可以用一个与相应实际转账频次正相关的数作为初始的特征表达向量中该维度的值。在初始的特征表达向量每一维都没有确定含义的一种向量表达的情况下,例如词汇的语义表示等,可以根据训练样本训练图神经网络确定各个节点的初始的特征表达向量,以及图神经网络的其他模型参数,在此不再赘述。
为了区分节点的特征表达向量,本说明书实施例将经过特征聚合模型处理关系网络得到的节点上的表征向量,称为业务表征向量。顾名思义,业务表征向量,可以是在具体业务中,用于体现业务特点的表征向量。例如,在依赖单个关系网络的分类业务中,可以根据某个节点的业务表征向量通过激励函数等映射到该节点在各个分类类别上的概率。
在对关系网络的处理过程中,通常采用邻居节点特征聚合的方式,将当前节点的前一层特征表达向量与邻居节点的特征表达向量进行聚合,得到当前层的特征表达向量作为当前层的输出。在本说明书实施例中,这种特征聚合的方式通过特征聚合模型描述。特征聚合模型可以是预先设定聚合方式(例如特征加权方式等)的模型,也可以是图神经网络模型。
一个节点和其周围的邻居节点通常具有不同的关联程度。关联程度不同,对当前节点的影响也不同。例如,高阶节点对当前节点的影响小于低阶节点、转账频次较高的邻居节点对当前节点的影响小于转账频次较低的邻居节点。因此,根据一个可能的设计,各个邻居节点可以对应有邻居重要度(邻居权重),用于描述各个邻居节点相对于当前节点的重要度,
根据一个实施方式,特征聚合模型用于对多阶邻居节点进行特征聚合,各阶邻居节点分别对应的相应的邻居阶数重要度系数。例如,假设当前节点的权重为a 0,每个一阶邻居节点的邻居权重均为a 1,二阶邻居节点的邻居权重均为a 2……其中,a 0>a 1>a 2…在图神经网络的训练过程中,这些邻居权重可以作为特征聚合模型的参数,根据样本实体对应的样本特征和预先标注的样本业务结果,调整确定。可选地,在各阶邻居节点中,还可以针对各个邻居节点确定不同的邻居权重,例如一阶邻居节点中,邻居权重与邻居节点和当前节点之间的相互转账频次正相关。
在一个可选的实现方式中,特征聚合模型为图卷积神经网络(以下也称为图神经网络),各个邻居节点对应的不同的邻居重要度系数。各个邻居节点的重要度可以通过 当前节点与邻居节点的特征表达向量确定。
作为一个具体示例,通过图卷积神经网络处理图数据时,确定第l+1层的节点v的特征表达的卷积算子可以为:
Figure PCTCN2021074248-appb-000001
其中:H l+1(v)是节点v在图卷积神经网络的第l+1层的特征表达向量;N(v)是节点v的邻居节点;d v、d u是归一化因子,比如是相应节点的度,即,与相应节点连接的连接边数量,或者一阶邻居节点的数量;H l(v)是节点v在图卷积神经网络的第l层的特征表达向量;H l(u)是节点u在图卷积神经网络的第l层的特征表达向量;W l是相应节点图卷积神经网络第l层的模型参数。邻居节点有多个时,W l可以是矩阵形式的模型参数,可以称为权重矩阵。公式还可以考虑当前节点的更高阶邻居节点的特征聚合,在此用省略号表示,其原理与一阶邻居节点的特征聚合类似,在此不再赘述。其中,不同的邻居节点的归一化因子不同,特征表达向量不同,从而与权重矩阵相乘的积也不同,因此具有不同的邻居权重。
另外,如果每个实体特征对应的一个值,特征表达向量可以是各个实体特征对应的值构成的向量,如果每个实体特征对应的是一个向量,特征表达向量可以是各个实体特征对应的向量拼接得到的向量。在本说明书的实施架构下,各个节点初始的特征表达向量可以是预先确定的。在图神经网络训练过程中,根据训练样本调整模型参数(如权重矩阵)。在每个关系网络中,通过用于处理该关系网络的相应图神经网络可以对当前实体对应的节点进行特征聚合,得到相应的表征向量。其中,在特征聚合过程中,可以将当前实体对应的节点的预定阶数(如2阶)内的邻居节点作为特征聚合的节点,也可以对预定阶数内的邻居进行采样,将采样得到的邻居节点做特征聚合。特征聚合的方式例如可以是:加和、求平均、取最大值、求加权和,等等,在此不作限定。
值得说明的是,上述卷积算子只是图卷积神经网络中特征聚合的一个具体示例,实践中,可以采用多种方式进行特征聚合,每一层图神经网络对应的模型参数也可以有多组,例如每组模型参数为一个权重矩阵,一层图神经网络可以对应多个权重矩阵。对于训练好的图神经网络而言,模型参数可以是经过训练过程中的参数调整确定下来的。
根据另一个具体示例,例如,在某一种连接关系类型对应的关系网络(如称为第一关系网络)中,当前节点为节点μ,邻居节点j的邻居权重可以为:
α(μ,j)=softmax j(V·tanh(W 1[X u||X j])+b 1)
其中,矩阵V(例如称为第一辅助矩阵)和W 1(例如称为第一权重矩阵)是图神经网络训练过程中确定的模型参数,b 1是图神经网络训练过程中确定的常数参数,X u、X j分别是节点μ、节点j对应的当前的特征表达向量,[X u||X j]表示两个向量的拼接向量。可以理解的是,激活函数softmax、tanh也可以用其他激活函数(如Relu等)代替,在此不作限定。
如此,可以针对各个邻居节点分别确定相应的邻居权重。在各个邻居节点的当前特征向量表达各不相同的情况下,针对相应邻居节点的邻居权重也各不相同。值得说明的是,在图神经网络对关系网络的处理过程中,也可以将当前节点看作自身的邻居节点,例如称为零阶邻居节点。
根据邻居权重对各个邻居节点进行特征聚合,可以采用诸如求加权和等方式进行。例如,通过N u表示当前节点的邻居节点集合,当前节点经过一层图神经网络的邻居聚合结果为:
Figure PCTCN2021074248-appb-000002
可以理解,对于每个节点,经过一层图神经网络之后,都可以得到一个当前层的聚合结果,如节点j的聚合结果(也可以称为表征向量)为h j。在第一层图神经网络聚合时,各个节点的当前特征表达向量由相应节点的节点特征确定。
在一个实施例中,可以将以上邻居聚合结果进一步当前节点的特征表达向量综合,得到当前节点在图神经网络的当前层的聚合结果。为了更清楚说明针对当前节点聚合得到当前层的表征向量的过程,可以参考图3所示。图3中,假设图神经网络为多层网络,节点1、节点2、节点3…为节点u的邻居节点,将它们在第i-1层(i≥2)对应的特征聚合结果分别记为
Figure PCTCN2021074248-appb-000003
当前节点u在第i-1层对应的特征聚合结果记为
Figure PCTCN2021074248-appb-000004
则在第i层,相应节点的当前特征表达向量为第i-1层的特征聚合结果(即第i-1层输出的表征向量),即图3中
Figure PCTCN2021074248-appb-000005
将节点u的各个邻居节点进行聚合,得到邻居聚合结果
Figure PCTCN2021074248-appb-000006
然后,将
Figure PCTCN2021074248-appb-000007
Figure PCTCN2021074248-appb-000008
综合,可以得到节点u在第i层的特征表达向量
Figure PCTCN2021074248-appb-000009
从而,在单个关系网络(如第一关系网络)中,经过预先训练的图神经网络的层层迭代处理,可以得到当前节点对应的一个表征向量(如第一表征向量)。
这里,将
Figure PCTCN2021074248-appb-000010
Figure PCTCN2021074248-appb-000011
综合的过程例如可以是求和、求平均或加权求和等。然而,在特征表达向量中,每个特征对节点的表达向量的贡献度也可能不同,因此,在进一步可选的实现方式中,各个特征还可以具有特征重要度(特征权重)。
根据一个实施方式,特征权重可以是预先设定或者训练得到的。例如在描述用户之间的转账关系的关系网络中,初始的转账频次的特征权重大于转账金额的特征权重。举例而言,在确定表征向量时,具体到某个节点,例如对应第一邻居权重的第一节点,其对应的第一特征具有第一特征权重,该第一节点在第一特征上对应的第一特征聚合系数可以为第一特征权重与第一邻居权重的乘积。在进行特征聚合时,可以将第一特征对应的特征表达(如一个值或一个向量),与第一特征聚合系数相乘,得到的乘积作为相应加权项,将各个邻居节点在第一特征上的加权项加和,得到当前节点经过邻居特征聚合后在第一特征上的特征值。经过对第一关系网络的预定次数(在上述处理模型为图神经网络时,与图神经网络层数一致)迭代,从而确定出第一表征向量。
但是,当特征聚合模型为图神经网络时,由于在图神经网络处理过程中,隐层的特征并不能准确确定其含义,因此不能通过认为定义特征权重。因此,根据另一个实施方式,可以通过训练图神经网络,得到处理关系网络过程中特征重要度相关的通用参数。
作为一个具体示例,在某一层图神经网络中,可以通过以下方式确定各个特征分别对应的特征权重构成的特征权重向量:
Figure PCTCN2021074248-appb-000012
其中,W 2(例如称为第二权重矩阵)、W 3(例如称为第二辅助矩阵)均为图神经网络中第i层的权重矩阵,b 2、b 3均为常数参数,这些模型参数均可以在图神经网络训练过程中根据损失函数进行调整确定。在神经网络的某一层,W 2、W 3、b 2、b 3可以作为通用参数。
Figure PCTCN2021074248-appb-000013
表示两个向量的拼接。激励函数Relu也可以通过其他合适的激励函数代替,在此不再赘述。
特征权重向量β中的各个元素分别对应各个特征的特征权重。将相应特征权重与邻居聚合结果中的相应元素一一对应相乘,可以得到当前节点u在当前层的特征聚合结果。参考图3,根据特征权重确定最终的聚合结果的方式可以表示为:
Figure PCTCN2021074248-appb-000014
其中,⊙表示将两个矩阵的对应元素相乘(如哈达玛积)。对于向量而言,β中的第k个元素与
Figure PCTCN2021074248-appb-000015
中的第k个元素作为聚合结果
Figure PCTCN2021074248-appb-000016
中的第k个元素。例如,向量(A,B, C)⊙(a,b,c)的结果为(Aa,Bb,Cc)。
如此,可以同时考虑节点贡献度和特征贡献度,得到更准确的邻居节点的特征聚合结果。当特征聚合模型为图神经网络时,最后一层得到的聚合结果就是当前节点与当前关系网络对应的业务表征向量。
在表征各个连接关系类型的各个关系网络,可以分别针对当前实体对相应节点的邻居节点进行特征聚合,得到当前实体分别在各个连接关系类型下的各个业务表征向量。如在第一关系网络中,得到第一业务表征向量。
另一方面,在步骤203中,根据当前实体在各个连接关系类型下分别对应的实体特征,确定各个业务表征向量分别对应的各个重要度系数。可以理解,对于具体业务而言,不同连接关系下的实体特征具有不同的重要性。例如,用户风险性预测业务中,连接关系类型为用户之间的交互关系的关系网络比较重要,而连接关系类型为终端应用公用网络的关系网络的重要度较小。
在一个实施例中,关系网络的重要度系数可以根据经验预先设定。例如,描述用户之间的交互关系的关系网络的重要度系数为0.5,终端应用公用网络的关系网络的重要度为0.1。
在另一个实施例中,关系网络的重要度系数可以作为图神经网络的模型参数,利用样本数据训练确定。重要度系数可以描绘当前业务处理过程中对各个元路径(连接关系)的偏好。例如可以通过注意力值来体现这种偏好。
作为示例,当前实体在其中一个关系网络上的注意力值可以通过以下方式确定:
Figure PCTCN2021074248-appb-000017
其中,Z ρ是关系网络ρ下的注意力向量(可通过样本数据训练确定),
Figure PCTCN2021074248-appb-000018
是所有关系网络的元路径下的业务表征向量的拼接向量(通过步骤202得到的各个业务表征向量的拼接得到的向量),P是所有的关系网络对应的注意力向量集合。也就是说,在第一关系网络下,可以根据预先训练得到的各个连接关系类型分别对应的各个注意力向量,确定当前实体分别对应于各个连接关系类型的各个注意力值,然后将当前实体在第一连接关系类型下对应的第一注意力值与各个连接关系类型的注意力值之和的比值,确定为当前实体对应的第一重要度系数。
在一个实施例中,针对当前实体,第一关系网络对应的第一注意力值为,自变量为以下值的指数函数:第一连接关系类型对应的第一注意力向量的转置向量,与各个表征向量的拼接向量的乘积。第一重要度系数为,第一注意力值与各个关系网络分别对应的各个注意力值之和的比值。
可以理解,Z ρ可以是模型参数,可以在图神经网络训练过程中通过样本数据调整确定,
Figure PCTCN2021074248-appb-000019
可以是具体到确定当前实体在步骤202中的各个表征向量的拼接向量,根据
Figure PCTCN2021074248-appb-000020
和各个Z ρ,可以确定当前实体在不同关系网络下的重要度系数。
然后,在步骤204中,基于各个重要度系数,融合当前实体在多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对当前实体的综合评估结果。可以理解,根据各个业务表征向量的重要度系数,可以针对当前实体确定出一个综合评估结果。其中,综合评估结果是用于在具体业务上对当前实体进行评估的业务结果。例如在预测业务中,该综合评估结果可以是针对当前实体的预测分数,在目标识别业务中,该综合评估结果可以是目标识别的准确度,在信息推送业务中,该综合评估结果可以是待推送信息和用户的兴趣程度,在分类业务中,该综合评估结果可以是在各个类别上的评分等等。
根据一个实施方式,可以将各个关系网络中针对当前实体的重要度系数作为权重,对步骤203中得到的各个业务表征向量加权求和,得到的和值作为对当前实体的综合评估结果,或者对得到的和值进一步处理得到对当前实体的综合评估结果。
根据另一个实施方式,可以将针对当前实体的重要度系数最大的关系网络对应的业务表征向量,或者对该业务表征向量进一步处理得到的结果作为对当前实体的综合评估结果。
其中,这里的进一步处理例如可以是在具体业务(如金融平台还款的风险度等)上进行评分。
根据有一个实施方式,每个关系网络针对当前实体的重要度系数可以有多个,分别对应各个分类类别。即每个关系网络针对当前实体的重要度系数可以包括在各个分类类别上的重要度系数。则还可以通过全连接层,将各个业务表征向量作为全连接层的输入,相应重要度系数作为相应的权重,对当前实体在各个候选类别上打分,得到各个打分结果,从而进行类别预测。
为了更明确本说明书实施例的应用场景,图4示出一个具体例子的基于异构图进 行用户风险性判断的业务处理的示意图。如图4所示,给出了在该具体例子中,异构图包括针对N个用户的不同连接关系(元路径)进行描述的关系网络。当前业务需求为预测用户n在金融借贷领域的风险性(如违约概率)的情况下,根据用户n在各个关系网络中对应的当前节点,通过预先训练的图神经网络分别对异构图中的各个关系网络进行处理,分别得到针对用户n的综合向量表征,即各个业务表征向量。然后,根据各个综合向量表征确定各个关系网络相对于用户n的重要度系数。将各个业务表征向量作为全连接神经网络的各个神经元的输入,各个重要度系数作为相应神经元的权重,对各个业务表征向量进行融合,得到对用户n的综合评估结果(如风险分数)。根据该综合评估结果,可以输出用户n在金融借贷领域的风险性,如风险分数高于风险阈值,输出高风险用户的结果。根据该结果,可以进行后续业务,如限制该用户n的借贷金额、禁止用户n进行借贷业务,等等。
通过上述的基于异构图进行业务处理的方法,可以对多个用户之间的异构图中,先针对不同的连接关系构成的各个关系网络分别进行处理,得到当前实体分别在各个关系网络中的业务表征向量,然后,根据当前业务中,当前实体的各个关系网络分别对应的重要度系数,对这些业务表征向量进行融合,从而得到一个综合评估结果,从而利用综合评估结果针对当前实体进行后续业务处理。由于利用了多个不同连接关系类型的关系网络,可以更加全面的刻画实体的特征,另一方面,先针对各个关系网络分别处理得到各个业务表征向量,无需对各个关系网络进行综合,可以避免繁琐的手工特征抽取,进一步地,可以自动确定当前业务下,当前实体在每个关系网络中的重要度系数(权重),实现在各个关系网络下的信息融合,从而使得对当前实体的评估结果更加准确。
根据另一方面的实施例,还提供一种基于异构图进行业务处理的装置。图5示出根据一个实施例的基于异构图进行业务处理装置的示意性框图。其中,异构图用于描述多个实体之间分别在预定的多个连接关系类型下的关联关系,其中,多个连接关系类型相互独立,多个连接关系类型包括第一连接关系类型,在第一连接关系类型构成的第一关系网络中,各个实体分别与各个节点一一对应,通过连接边连接的两两节点对应的两两实体之间具有第一连接关系类型的连接关系,各个节点分别对应有相应实体在第一连接关系类型下的实体特征。
以针对第一关系网络的处理为例,如图5所示,基于异构图进行业务处理装置500包括:节点确定单元51,配置为确定当前业务所针对的当前实体在第一关系网络中对应的当前节点;特征聚合单元52,配置为通过预先确定的特征聚合模型处理第一关系网络, 得到针对当前节点在第一连接关系类型下的第一业务表征向量;重要度确定单元53,根据当前实体在各个连接关系类型下分别对应的实体特征,确定第一业务表征向量对应的第一重要度系数;融合单元54,配置为至少基于第一重要度系数和第一业务表征向量,融合当前实体在多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对当前实体的综合评估结果,以利用综合评估结果针对当前实体进行后续业务处理。
根据一方面的实施方式,在第一关系网络中,当前节点的邻居节点包括第一节点,第一节点对应第一邻居权重,第一节点对应的实体特征包括第一特征,第一特征对应第一特征权重,特征聚合单元52进一步配置为:将第一特征权重与第一邻居权重的乘积确定为第一节点在第一特征上的第一特征聚合系数;基于第一节点在第一特征上的特征表达向量与第一特征聚合系数的乘积,确定第一业务表征向量中与第一特征向量相对应的元素值。
根据另一方面的实施方式,假设在第一关系网络中,当前节点的邻居节点包括第二节点,第二节点对应第二邻居权重,预先确定的特征聚合模型为第一图神经网络,则特征聚合单元52还配置可以为,利用第一图神经网络的第i层通过以下方式处理第一关系网络:将当前节点的当前特征表达向量和第二节点的当前特征表达向量拼接,得到第一拼接向量;基于第一权重矩阵与第一拼接向量的乘积,确定第二节点在第i层的邻居权重,第一权重矩阵是第一图神经网络在第i层的模型参数,在训练第一图神经网络时确定;根据第二节点在第i层的邻居权重与第二节点的当前特征表达向量,确定当前节点的第i邻居聚合向量;将上述邻居聚合向量与当前节点的当前特征表达向量融合,得到当前节点经过第i层图神经网络处理后的表征向量。
在进一步的实施例中,在i为1的情况下,当前节点的当前特征表达向量和第二节点的当前特征表达向量,分别由当前节点和第二节点在第一关系网络中的实体特征确定;在i为大于1的自然数的情况下,当前节点的当前特征表达向量和第二节点的当前特征表达向量,分别为当前节点和第二节点经过第i-1层图神经网络处理后的表征向量。
根据另一个进一步的实施例,特征聚合单元52进一步配置为,通过以下方式将上述邻居聚合向量与当前节点的当前特征表达向量融合,得到当前节点经过第i层图神经网络处理后的表征向量:将邻居聚合向量与当前节点的当前特征表达向量进行拼接,得到第二拼接向量;基于第二权重矩阵与第二拼接向量的乘积,确定第i层图神经网络中的特征权重向量,第二权重矩阵是第一图神经网络在第i层的模型参数,在训练第一图神经网络时确定;根据特征权重向量对邻居聚合向量进行修正,得到当前节点经过第i 层图神经网络处理后的表征向量,在第i层图神经网络为第一图神经网络的最后一层时,修正后得到的表征向量为第一业务表征向量。
在一个可选的实施例中,特征聚合单元52进一步可以配置为,将特征权重向量中第k个元素与邻居聚合向量中第k个元素的乘积,作为当前节点经过第i层图神经网络处理后的表征向量的第k个元素,从而根据特征权重向量对邻居聚合项进行修正。
根据一个实施例,重要度确定单元53进一步配置为:根据预先训练得到的各个连接关系类型分别对应的各个注意力向量,确定当前实体分别对应于各个连接关系类型的各个注意力值;将当前实体在第一连接关系类型下对应的第一注意力值与各个连接关系类型的注意力值之和的比值,确定为第一重要度系数。
在一个进一步的实施例中,第一注意力值为,自变量为以下值的指数函数:第一连接关系类型对应的第一注意力向量的转置向量,与各个业务表征向量的拼接向量的乘积。
根据一个可能的设计,融合单元54还配置为:将各个重要度系数作为相应表征向量的权重,确定各个表征向量的加权和,其中,第一重要度系数为第一表征向量的权重;将加权和作为对当前实体的综合评估结果。
综合评估结果包括以下中的一种:预测业务中的预测分数、分类业务中在各个类别上的评分。
在一个实施例中,上述多个实体包括第一实体,在各个连接关系类型下,第一实体分别对应的各个节点通过第一实体在各个连接关系类型下的用户标识相关联。例如,在各个连接关系类型对应的关系网络中,同一个实体对应的额节点通过同一个节点标识表示,或者通过表格记录各个连接关系类型对应的关系网络中,对应到同一个实体的节点标识的对应关系。
值得说明的是,图5所示的装置500是与图2示出的方法实施例相对应的装置实施例,图2示出的方法实施例中的相应描述同样适用于装置500,在此不再赘述。
根据另一方面的实施例,还提供一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行结合图2所描述的方法。
根据再一方面的实施例,还提供一种计算设备,包括存储器和处理器,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现结合图2所述的方法。
本领域技术人员应该可以意识到,在上述一个或多个示例中,本说明书实施例所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时,可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式,对本说明书的技术构思的目的、技术方案和有益效果进行了进一步详细说明,所应理解的是,以上所述仅为本说明书的技术构思的具体实施方式而已,并不用于限定本说明书的技术构思的保护范围,凡在本说明书实施例的技术方案的基础之上,所做的任何修改、等同替换、改进等,均应包括在本说明书的技术构思的保护范围之内。

Claims (24)

  1. 一种基于异构图进行业务处理的方法,所述异构图用于描述多个实体之间在预定的多个连接关系类型下的关联关系,其中,所述多个连接关系类型相互独立,所述多个连接关系类型包括第一连接关系类型,描述所述多个实体之间在所述第一连接关系类型下的关联关系的关系网络为第一关系网络,在所述第一关系网络中,各个实体分别与各个节点一一对应,通过连接边连接的两两节点对应的两两实体之间具有所述第一连接关系类型的关联关系,各个节点分别对应有相应实体在所述第一连接关系类型下的实体特征;所述方法包括:
    确定当前业务所针对的当前实体在所述第一关系网络中对应的当前节点;
    通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量;
    根据所述当前实体在各个连接关系类型下分别对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数;
    至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果,以利用所述综合评估结果针对所述当前实体进行后续业务处理。
  2. 根据权利要求1所述的方法,其中,在所述第一关系网络中,所述当前节点的邻居节点包括第一节点,所述第一节点对应第一邻居权重,所述第一节点对应的实体特征包括第一特征,所述第一特征对应第一特征权重,所述通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量包括:
    将所述第一特征权重与所述第一邻居权重的乘积确定为所述第一节点在所述第一特征上的第一特征聚合系数;
    基于所述第一节点在所述第一特征上的特征表达向量与所述第一特征聚合系数的乘积,确定所述第一业务表征向量中与所述第一特征向量相对应的元素值。
  3. 根据权利要求1所述的方法,其中,在所述第一关系网络中,所述当前节点的邻居节点包括第二节点,所述第二节点对应第二邻居权重,所述预先确定的特征聚合模型为第一图神经网络;所述第一图神经网络的第i层通过以下方式处理所述第一关系网络:
    将所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量拼接,得到第一拼接向量;
    基于第一权重矩阵与所述第一拼接向量的乘积,确定所述第二节点在第i层的邻居权重,所述第一权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;
    根据所述第二节点在第i层的邻居权重与所述第二节点的当前特征表达向量,确定所述当前节点的第i邻居聚合向量;
    将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量。
  4. 根据权利要求3所述的方法,其中:
    在i为1的情况下,所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量,分别由所述当前节点和所述第二节点在所述第一关系网络中的实体特征确定;
    在i为大于1的自然数的情况下,所述当前节点的当前特征表达向量和所述第二节 点的当前特征表达向量,分别为所述当前节点和所述第二节点经过第i-1层图神经网络处理后的表征向量。
  5. 根据权利要求3所述的方法,其中,将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量包括:
    将所述邻居聚合向量与所述当前节点的当前特征表达向量进行拼接,得到第二拼接向量;
    基于第二权重矩阵与所述第二拼接向量的乘积,确定第i层图神经网络中的特征权重向量,所述第二权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;
    根据所述特征权重向量对所述邻居聚合向量进行修正,得到所述当前节点经过第i层图神经网络处理后的表征向量,在第i层图神经网络为所述第一图神经网络的最后一层时,修正后得到的表征向量为所述第一业务表征向量。
  6. 根据权利要求5所述的方法,其中,根据所述特征权重向量对所述邻居聚合项进行修正包括,将所述特征权重向量中第k个元素与所述邻居聚合向量中第k个元素的乘积,作为所述当前节点经过第i层图神经网络处理后的表征向量的第k个元素。
  7. 根据权利要求1所述的方法,其中,根据所述当前实体在各个连接关系类型下对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数包括:
    根据预先训练得到的各个连接关系类型分别对应的各个注意力向量,确定所述当前实体分别对应于各个连接关系类型的各个注意力值;
    将当前实体在所述第一连接关系类型下对应的第一注意力值与各个连接关系类型的注意力值之和的比值,确定为所述第一重要度系数。
  8. 根据权利要求7所述的方法,其中,所述第一注意力值为,自变量为以下值的指数函数:第一连接关系类型对应的第一注意力向量的转置向量,与各个业务表征向量的拼接向量的乘积。
  9. 根据权利要求1所述的方法,其中,至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果包括:
    将各个重要度系数作为相应表征向量的权重,确定各个业务表征向量的加权和,其中,所述第一重要度系数为所述第一业务表征向量的权重;
    将所述加权和作为对所述当前实体的综合评估结果。
  10. 根据权利要求1所述的方法,其中,所述综合评估结果包括以下中的一种:预测业务中的预测分数、分类业务中在各个类别上的评分。
  11. 根据权利要求1所述的方法,其中,所述多个实体包括第一实体,所述第一实体在各个连接关系类型下分别对应的各个节点通过所述第一实体的至少一个用户标识相关联。
  12. 一种基于异构图进行业务处理的装置,所述异构图用于描述多个实体之间分别在预定的多个连接关系类型下的关联关系,其中,所述多个连接关系类型相互独立,所述多个连接关系类型包括第一连接关系类型,描述所述多个实体之间在所述第一连接关系类型下的关联关系的关系网络为第一关系网络,在所述第一关系网络中,各个实体分别与各个节点一一对应,通过连接边连接的两两节点对应的两两实体之间具有所述第一连接关系类型的关联关系,各个节点分别对应有相应实体在所述第一连接关系类型下的实体特征;所述装置包括:
    节点确定单元,配置为确定当前业务所针对的当前实体在所述第一关系网络中对应的当前节点;
    特征聚合单元,配置为通过预先确定的特征聚合模型处理第一关系网络,得到针对所述当前节点在所述第一连接关系类型下的第一业务表征向量;
    重要度确定单元,根据所述当前实体在各个连接关系类型下分别对应的实体特征,确定所述第一业务表征向量对应的第一重要度系数;
    融合单元,配置为至少基于所述第一重要度系数和所述第一业务表征向量,融合所述当前实体在所述多个连接关系类型分别对应的各个关系网络下的各个业务表征向量,得到对所述当前实体的综合评估结果,以利用所述综合评估结果针对所述当前实体进行后续业务处理。
  13. 根据权利要求12所述的装置,其中,在所述第一关系网络中,所述当前节点的邻居节点包括第一节点,所述第一节点对应第一邻居权重,所述第一节点对应的实体特征包括第一特征,所述第一特征对应第一特征权重,所述特征聚合单元进一步配置为:
    将所述第一特征权重与所述第一邻居权重的乘积确定为所述第一节点在所述第一特征上的第一特征聚合系数;
    基于所述第一节点在所述第一特征上的特征表达向量与所述第一特征聚合系数的乘积,确定所述第一业务表征向量中与所述第一特征向量相对应的元素值。
  14. 根据权利要求12所述的装置,其中,在所述第一关系网络中,所述当前节点的邻居节点包括第二节点,所述第二节点对应第二邻居权重,所述预先确定的特征聚合模型为第一图神经网络;所述特征聚合单元还配置为,利用所述第一图神经网络的第i层通过以下方式处理所述第一关系网络:
    将所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量拼接,得到第一拼接向量;
    基于第一权重矩阵与所述第一拼接向量的乘积,确定所述第二节点在第i层的邻居权重,所述第一权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;
    根据所述第二节点在第i层的邻居权重与所述第二节点的当前特征表达向量,确定所述当前节点的第i邻居聚合向量;
    将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量。
  15. 根据权利要求14所述的装置,其中:
    在i为1的情况下,所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量,分别由所述当前节点和所述第二节点在所述第一关系网络中的实体特征确定;
    在i为大于1的自然数的情况下,所述当前节点的当前特征表达向量和所述第二节点的当前特征表达向量,分别为所述当前节点和所述第二节点经过第i-1层图神经网络处理后的表征向量。
  16. 根据权利要求14所述的装置,其中,所述特征聚合单元进一步配置为通过以下方式将所述邻居聚合向量与所述当前节点的当前特征表达向量融合,得到所述当前节点经过第i层图神经网络处理后的表征向量:
    将所述邻居聚合向量与所述当前节点的当前特征表达向量进行拼接,得到第二拼接向量;
    基于第二权重矩阵与所述第二拼接向量的乘积,确定第i层图神经网络中的特征权 重向量,所述第二权重矩阵是所述第一图神经网络在第i层的模型参数,在训练所述第一图神经网络时确定;
    根据所述特征权重向量对所述邻居聚合向量进行修正,得到所述当前节点经过第i层图神经网络处理后的表征向量,在第i层图神经网络为所述第一图神经网络的最后一层时,修正后得到的表征向量为所述第一业务表征向量。
  17. 根据权利要求16所述的装置,其中,所述特征聚合单元进一步配置为,将所述特征权重向量中第k个元素与所述邻居聚合向量中第k个元素的乘积,作为所述当前节点经过第i层图神经网络处理后的表征向量的第k个元素,从而根据所述特征权重向量对所述邻居聚合项进行修正。
  18. 根据权利要求12所述的装置,其中,所述重要度确定单元进一步配置为:
    根据预先训练得到的各个连接关系类型分别对应的各个注意力向量,确定所述当前实体分别对应于各个连接关系类型的各个注意力值;
    将当前实体在所述第一连接关系类型下对应的第一注意力值与各个连接关系类型的注意力值之和的比值,确定为所述第一重要度系数。
  19. 根据权利要求18所述的装置,其中,所述第一注意力值为,自变量为以下值的指数函数:第一连接关系类型对应的第一注意力向量的转置向量,与各个业务表征向量的拼接向量的乘积。
  20. 根据权利要求12所述的装置,其中,所述融合单元还配置为:
    将各个重要度系数作为相应业务表征向量的权重,确定各个表征向量的加权和,其中,所述第一重要度系数为所述第一业务表征向量的权重;
    将所述加权和作为对所述当前实体的综合评估结果。
  21. 根据权利要求12所述的装置,其中,所述综合评估结果包括以下中的一种:预测业务中的预测分数、分类业务中在各个类别上的评分。
  22. 根据权利要求12所述的装置,其中,所述多个实体包括第一实体,所述第一实体在各个连接关系类型下分别对应的各个节点通过所述第一实体的至少一个用户标识相关联。
  23. 一种计算机可读存储介质,其上存储有计算机程序,当所述计算机程序在计算机中执行时,令计算机执行权利要求1-11中任一项所述的方法。
  24. 一种计算设备,包括存储器和处理器,其特征在于,所述存储器中存储有可执行代码,所述处理器执行所述可执行代码时,实现权利要求1-11中任一项所述的方法。
PCT/CN2021/074248 2020-03-10 2021-01-28 基于异构图进行业务处理的方法及装置 WO2021179834A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010162991.9A CN111309983B (zh) 2020-03-10 2020-03-10 基于异构图进行业务处理的方法及装置
CN202010162991.9 2020-03-10

Publications (1)

Publication Number Publication Date
WO2021179834A1 true WO2021179834A1 (zh) 2021-09-16

Family

ID=71158851

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/074248 WO2021179834A1 (zh) 2020-03-10 2021-01-28 基于异构图进行业务处理的方法及装置

Country Status (2)

Country Link
CN (1) CN111309983B (zh)
WO (1) WO2021179834A1 (zh)

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570058A (zh) * 2021-09-22 2021-10-29 航天宏康智能科技(北京)有限公司 推荐方法和装置
CN113987200A (zh) * 2021-10-19 2022-01-28 云南大学 神经网络结合知识图谱的推荐方法、系统、终端、介质
CN114065048A (zh) * 2021-11-25 2022-02-18 西安电子科技大学 基于多异构图图神经网络的物品推荐方法
CN114780867A (zh) * 2022-05-10 2022-07-22 杭州网易云音乐科技有限公司 推荐方法、介质、装置和计算设备
CN115314398A (zh) * 2022-09-29 2022-11-08 南昌航空大学 一种评估异质信息网络关键节点的方法
CN115359654A (zh) * 2022-08-02 2022-11-18 支付宝(杭州)信息技术有限公司 流量预测系统的更新方法及装置
CN116912919A (zh) * 2023-09-12 2023-10-20 深圳须弥云图空间科技有限公司 一种图像识别模型的训练方法及装置
WO2023207790A1 (zh) * 2022-04-28 2023-11-02 华为技术有限公司 一种分类模型的训练方法及装置

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111309983B (zh) * 2020-03-10 2021-09-21 支付宝(杭州)信息技术有限公司 基于异构图进行业务处理的方法及装置
CN111581450B (zh) * 2020-06-24 2023-07-14 支付宝(杭州)信息技术有限公司 确定用户的业务属性的方法及装置
CN111832924B (zh) * 2020-06-30 2021-06-15 北方工业大学 基于图神经网络的社区燃气系统动态风险评估方法及装置
CN112037038B (zh) * 2020-09-02 2024-05-28 中国银行股份有限公司 银行信贷风险预测方法及装置
CN112036418A (zh) * 2020-09-04 2020-12-04 京东数字科技控股股份有限公司 用于提取用户特征的方法和装置
CN111931069B (zh) * 2020-09-25 2021-01-22 浙江口碑网络技术有限公司 用户兴趣确定方法、装置及计算机设备
CN112541575B (zh) * 2020-12-06 2023-03-10 支付宝(杭州)信息技术有限公司 图神经网络的训练方法及装置
CN114648345B (zh) * 2020-12-17 2024-08-02 支付宝(杭州)信息技术有限公司 训练表征模型及确定实体表征向量的方法及装置
CN112766500B (zh) * 2021-02-07 2022-05-17 支付宝(杭州)信息技术有限公司 图神经网络的训练方法及装置
CN113298116B (zh) * 2021-04-26 2024-04-02 上海淇玥信息技术有限公司 基于注意力权重的图嵌入特征提取方法、装置及电子设备
CN113554100B (zh) * 2021-07-28 2023-04-07 湖南科技大学 异构图注意力网络增强的Web服务分类方法
CN113626650A (zh) * 2021-08-04 2021-11-09 支付宝(杭州)信息技术有限公司 业务处理方法、装置和电子设备
CN113724073A (zh) * 2021-09-09 2021-11-30 支付宝(杭州)信息技术有限公司 一种风险识别和控制方法及装置
CN114022058A (zh) * 2022-01-06 2022-02-08 成都晓多科技有限公司 基于时序知识图谱的中小企业失信风险预测方法
CN114470790A (zh) * 2022-02-09 2022-05-13 腾讯科技(深圳)有限公司 虚拟资源处理方法、装置、设备、计算机程序及存储介质
CN114707488B (zh) * 2022-02-25 2024-02-09 马上消费金融股份有限公司 数据处理方法、装置、计算机设备及存储介质
CN115798722B (zh) * 2023-02-02 2023-05-26 神州医疗科技股份有限公司 基于知识图谱的免疫药物人群高低危筛选方法及系统
CN117591755B (zh) * 2024-01-18 2024-06-14 杭州海康威视数字技术股份有限公司 动态异构网络潜在异常关系预测方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110046698A (zh) * 2019-04-28 2019-07-23 北京邮电大学 异质图神经网络生成方法、装置、电子设备及存储介质
US20190325075A1 (en) * 2018-04-18 2019-10-24 Oracle International Corporation Efficient, in-memory, relational representation for heterogeneous graphs
CN110717047A (zh) * 2019-10-22 2020-01-21 湖南科技大学 一种基于图卷积神经网络的Web服务分类方法
CN111309983A (zh) * 2020-03-10 2020-06-19 支付宝(杭州)信息技术有限公司 基于异构图进行业务处理的方法及装置

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140025670A1 (en) * 2012-07-19 2014-01-23 Berrin, Llc Location based recommendations
US9946798B2 (en) * 2015-06-18 2018-04-17 International Business Machines Corporation Identification of target audience for content delivery in social networks by quantifying semantic relations and crowdsourcing
CN106156483B (zh) * 2016-01-18 2018-10-02 李雪 一种基于文献数据的风险评价方法、装置及服务器
CN106980659A (zh) * 2017-03-20 2017-07-25 华中科技大学鄂州工业技术研究院 一种基于异构图模型的社交活动推荐方法
CN108399268B (zh) * 2018-03-29 2022-04-29 浙江大学 一种基于博弈论的增量式异构图聚类方法
CN109299373B (zh) * 2018-10-20 2021-10-29 上海交通大学 基于图卷积技术的推荐系统
CN110263780B (zh) * 2018-10-30 2022-09-02 腾讯科技(深圳)有限公司 实现异构图、分子空间结构性质识别的方法、装置和设备
CN109615167A (zh) * 2018-11-06 2019-04-12 阿里巴巴集团控股有限公司 确定疑似批量风险交易事件的方法、装置和电子设备
CN110097472B (zh) * 2019-05-08 2022-12-27 腾讯科技(深圳)有限公司 一种社团属性识别方法及相关设备

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190325075A1 (en) * 2018-04-18 2019-10-24 Oracle International Corporation Efficient, in-memory, relational representation for heterogeneous graphs
CN110046698A (zh) * 2019-04-28 2019-07-23 北京邮电大学 异质图神经网络生成方法、装置、电子设备及存储介质
CN110717047A (zh) * 2019-10-22 2020-01-21 湖南科技大学 一种基于图卷积神经网络的Web服务分类方法
CN111309983A (zh) * 2020-03-10 2020-06-19 支付宝(杭州)信息技术有限公司 基于异构图进行业务处理的方法及装置

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570058A (zh) * 2021-09-22 2021-10-29 航天宏康智能科技(北京)有限公司 推荐方法和装置
CN113987200A (zh) * 2021-10-19 2022-01-28 云南大学 神经网络结合知识图谱的推荐方法、系统、终端、介质
CN113987200B (zh) * 2021-10-19 2024-03-15 云南大学 神经网络结合知识图谱的推荐方法、系统、终端、介质
CN114065048A (zh) * 2021-11-25 2022-02-18 西安电子科技大学 基于多异构图图神经网络的物品推荐方法
CN114065048B (zh) * 2021-11-25 2024-02-09 西安电子科技大学 基于多异构图图神经网络的物品推荐方法
WO2023207790A1 (zh) * 2022-04-28 2023-11-02 华为技术有限公司 一种分类模型的训练方法及装置
CN114780867A (zh) * 2022-05-10 2022-07-22 杭州网易云音乐科技有限公司 推荐方法、介质、装置和计算设备
CN114780867B (zh) * 2022-05-10 2023-11-03 杭州网易云音乐科技有限公司 推荐方法、介质、装置和计算设备
CN115359654B (zh) * 2022-08-02 2023-09-08 支付宝(杭州)信息技术有限公司 流量预测系统的更新方法及装置
CN115359654A (zh) * 2022-08-02 2022-11-18 支付宝(杭州)信息技术有限公司 流量预测系统的更新方法及装置
CN115314398B (zh) * 2022-09-29 2022-12-23 南昌航空大学 一种评估异质信息网络关键节点的方法
CN115314398A (zh) * 2022-09-29 2022-11-08 南昌航空大学 一种评估异质信息网络关键节点的方法
CN116912919A (zh) * 2023-09-12 2023-10-20 深圳须弥云图空间科技有限公司 一种图像识别模型的训练方法及装置
CN116912919B (zh) * 2023-09-12 2024-03-15 深圳须弥云图空间科技有限公司 一种图像识别模型的训练方法及装置

Also Published As

Publication number Publication date
CN111309983A (zh) 2020-06-19
CN111309983B (zh) 2021-09-21

Similar Documents

Publication Publication Date Title
WO2021179834A1 (zh) 基于异构图进行业务处理的方法及装置
CN111061856B (zh) 一种基于知识感知的新闻推荐方法
TWI689871B (zh) 梯度提升決策樹(gbdt)模型的特徵解釋方法和裝置
WO2020135535A1 (zh) 一种推荐模型训练方法及相关装置
US11227217B1 (en) Entity transaction attribute determination method and apparatus
WO2019205325A1 (zh) 确定用户风险等级的方法、终端设备及计算机可读存储介质
CN111581450B (zh) 确定用户的业务属性的方法及装置
CN109766454A (zh) 一种投资者分类方法、装置、设备及介质
CN112231592B (zh) 基于图的网络社团发现方法、装置、设备以及存储介质
CN114491263B (zh) 推荐模型训练方法及装置、推荐方法及装置
Feng et al. Computational social indicators: a case study of chinese university ranking
CN111797320A (zh) 数据处理方法、装置、设备及存储介质
CN113609345B (zh) 目标对象关联方法和装置、计算设备以及存储介质
US20200394448A1 (en) Methods for more effectively moderating one or more images and devices thereof
CN111695084A (zh) 模型生成方法、信用评分生成方法、装置、设备及存储介质
CN114330499A (zh) 分类模型的训练方法、装置、设备、存储介质及程序产品
CN113792110A (zh) 一种基于社交物联网的设备信任值评估方法
CN115203550A (zh) 一种增强邻居关系的社交推荐方法及系统
CN111309823A (zh) 用于知识图谱的数据预处理方法及装置
CN112016850A (zh) 业务评估方法以及装置
CN115511606A (zh) 对象识别方法、装置、设备及存储介质
CN111079930A (zh) 数据集质量参数的确定方法、装置及电子设备
US12052174B1 (en) Reducing network traffic by filtering network requests based on network request-related information systems and methods
CN113468403A (zh) 基于大数据挖掘的用户信息预测方法及云计算ai服务系统
KR20200130767A (ko) 인공신경망을 이용한 암호화폐 거래소 상장여부 평가방법

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21768931

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21768931

Country of ref document: EP

Kind code of ref document: A1