CN111275176B - Distributed computing method and distributed computing system - Google Patents
Distributed computing method and distributed computing system Download PDFInfo
- Publication number
- CN111275176B CN111275176B CN202010123789.5A CN202010123789A CN111275176B CN 111275176 B CN111275176 B CN 111275176B CN 202010123789 A CN202010123789 A CN 202010123789A CN 111275176 B CN111275176 B CN 111275176B
- Authority
- CN
- China
- Prior art keywords
- target
- graph
- embedding vector
- node
- graph embedding
- Prior art date
- Legal status (The legal status 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 status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5083—Techniques for rebalancing the load in a distributed system
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present specification provides a distributed computing method and system for running a deep learning model of an analytic graph structure. At least one computing device in the distributed computing system iteratively performs a graph embedding vector update operation for the graph structure, comprising: obtaining a target graph embedding vector of a target node in the graph structure; receiving, from at least one associated computing device in the distributed computing system, a neighboring target graph embedding vector; generating an iterated target graph embedding vector based on the target graph embedding vector and the adjacent target graph embedding vector; and transmitting the iterated target graph embedding vector to the associated computing device.
Description
Technical Field
One or more embodiments in this specification relate to the field of distributed computing, and in particular, to a distributed computing method and a distributed computing system for running a deep learning model of an analytics graph structure.
Background
Traditionally, if an image is to be classified, one way is to manually extract, for example, texture, color, or some higher level feature. These features are then input into a classifier (e.g., random forest, etc.), and a label indicating the class of these features is output. Another way is by deep learning. The image is input into a trained neural network (e.g., convolutional neural network Convolutional Neural Network "CNN" and recurrent neural network Recurrent Neural Network "RNN") and then labels indicating the categories are output directly. Compared with a manual extraction mode, the neural network feature extraction and classification are performed in one step, and the manual feature extraction or manual rule is avoided. Automatically extracting features from raw data is an end-to-end (end-to-end) learning. Deep learning enables more efficient learning of features and patterns of the image than conventional methods.
Convolutional neural networks, however, are primarily powerful in processing data such as images and audio. This is because the subject of its study is data in a Euclidean whole ring (Euclidean domains), called Euclidean data. The most prominent feature of euclidean data is a regular spatial structure, such as an image being a regular two-dimensional square grid, such as speech being a regular one-dimensional sequence. Because these data structures can be represented by a one-dimensional, two-dimensional matrix, convolutional neural networks are very efficient to process.
However, many data in real life do not have a regular spatial structure. Such as abstracted maps of recommendation systems, electronic transactions, computational geometry, 3D grids, social networks, telecommunication networks, biological networks or brain connection groups, molecular structures, etc. Each node of the graph structure is different in connection, three nodes are connected, two nodes are connected, some nodes are connected more, and the graph structure is an irregular data structure and is called non-Euclidean data (Non Euclidean data). Non-euclidean data may be conveniently represented in a Graph form, referred to as Graph data (Graph data). The graph data is widely present in various businesses of a company. Whereas expansion of traffic requires mining information in the graph data to promote traffic development. However, because the graph data structure is irregular, the number of adjacent nodes may be different for each node, and a convolution operation cannot be performed by using a convolution kernel of the same size. That is, the conventional convolutional neural network cannot effectively analyze data in various businesses of a company.
There are many attempts to apply deep learning models to feature extraction and analysis of map data. Such as a graph neural network (Graph Neural Network, "GNN") and graph annotation network (Graph Attention Network), and so forth. When analyzing graph data using these deep learning models, the graph data is conventionally loaded onto a single computing device, such as a single electronic computer (stand-alone), through graph queries. And then, the single machine runs the deep learning model to analyze some indexes of the graph data.
However, this method of analyzing the map data is quite inefficient. Because the indexes of a plurality of adjacent nodes around the node to be analyzed need to be calculated simultaneously when index analysis is performed, the following problems are unavoidable when a single machine operates: on the one hand, these neighboring nodes occupy too much memory. For some graph data up to billions of nodes and billions of edges, each single machine cannot store all the graph data. On the other hand, because the plurality of neighboring nodes around the node to be analyzed also have oversized subgraphs at the same time, the storage cannot be performed on the same single machine, and other subgraph-based schemes cannot be realized. Therefore, the traditional scheme has to sample and clip the graph data, and the model effect is reduced by reducing the input data, so that the memory is saved. On the other hand, the same graph data repeatedly appear among different nodes in the node calculation process, and each node needs to be independently analyzed in the single-machine calculation process, so that a large amount of redundant repeated calculation is inevitably generated, and the analysis efficiency is reduced.
Therefore, there is a need for a method and system that reduces or eliminates sample clipping while not producing a significant amount of redundant computation.
Disclosure of Invention
The technical solution of one or more embodiments in the present disclosure is to solve the technical problem that when a stand-alone running deep learning model is used to analyze graph data, the graph data has to be sample-cut and a large amount of redundant computation is generated in the computation process due to excessively large data volume.
To this end, one or more embodiments in the present specification provide a deep learning model that runs analysis graph data through a distributed computing method and distributed system. The method performs distributed computation on the deep learning model of the analysis chart data through homogeneous single machine operation. For those large scale graph data up to billions of nodes, trillions of edges, one or more embodiments presented herein transform the deep learning model into a distributed graph computation model by model understanding and model transforming the deep learning model. The large-scale graph data is then distributed computed by means of the distributed system running the graph computation model.
The solution proposed by one or more embodiments in the present specification includes splitting the deep learning model into an initialization module, an iteration module and a final evaluation module, so that the computation for each node in the graph data in the whole deep learning model is separately and parallelly processed by a single machine in the distributed system, and after the processing is completed, the computation structure of each node is merged and input into a neural network for final computation evaluation. Thus, each single machine in the distributed system requires little memory to process the large-scale map data. And because distributed computing breaks up data, one or more embodiments presented in this specification can process the large-scale graph data without losing any data.
In particular, in one aspect, the present specification provides a distributed computing method for running a deep learning model of an analytic graph structure. The method comprises the following steps: performing, by at least one computing device in a distributed computing system, a graph embedding vector update operation of the graph structure, wherein the graph embedding vector update operation comprises: obtaining a target graph embedded vector of a target node in the graph structure, wherein the graph structure comprises a plurality of nodes, each node is connected with at least one adjacent node through an edge, and the target node is one node in the plurality of nodes; receiving, from at least one associated computing device in the distributed computing system, a neighboring target graph embedding vector, the neighboring target graph embedding vector being a target graph embedding vector of at least one neighboring node of the target node, the associated computing device being a computing device in the distributed computing system that performs an operation on the target neighboring node; generating an iterated target graph embedding vector by running an operation module in the deep learning model based on the target graph embedding vector and an adjacent target graph embedding vector, wherein the deep learning model comprises a target index evaluation model based on the graph structure, and the operation module is a part of the deep learning model for determining a target graph embedding vector value of the target node; and transmitting the iterated target graph embedding vector to the associated computing device.
According to some embodiments in the present specification, the operation module includes at least: at least one layer in the graph neural network model; at least one layer in the graph rolling neural network model; and/or annotating at least one layer in the network model.
According to some embodiments in the present specification, the target graph embedding vector comprises a floating point number vector of a preset dimension.
According to some embodiments in the present description, the obtaining the target graph embedding vector for the target node in the graph structure includes, by at least one computing device in the distributed computing system: acquiring an initial node map embedding vector of the target node; acquiring an initial edge map embedding vector of an edge corresponding to the target node; generating an initial target graph embedding vector of the target node by running the initialization module based on the initial node graph embedding vector and the initial edge graph embedding vector; and taking the initial target graph embedding vector as the target graph embedding vector.
According to some embodiments in the present specification, the obtaining the initial graph embedding vector for the target node includes: acquiring characteristic data of the target node, and generating the initial graph embedding vector of the target node based on the characteristic data of the target node; and the obtaining the initial edge map embedding vector of the edge corresponding to the target node comprises the following steps: and acquiring edge characteristic data corresponding to the target node, and generating the initial edge map embedding vector based on the edge characteristic data.
According to some embodiments in the present specification, the method further comprises, by the at least one computing device in the distributed computing system: and executing the graph embedding vector updating operation for preset times, and generating a final graph embedding vector of the target node.
According to some embodiments in the present description, the deep learning model further comprises an evaluation module configured to translate the target graph embedding vector into a score of the target index for a corresponding node in the graph structure. The method further comprises the steps of: operating the evaluation module based on the final graph embedding vector to generate a target index score corresponding to the final graph embedding vector; and outputting the target index score.
According to some embodiments in the present description, the evaluation module comprises a fully connected layer of at least one neural network.
In another aspect, one or more embodiments in the present specification provide a distributed computing system for evaluating target metrics of a deep learning model by distributed computing. The system comprises: at least one storage medium stores at least one set of instructions for executing a deep learning model of the analytic graph structure; a plurality of computing devices, each of the plurality of computing devices being communicatively coupled to at least one other computing device and in communication with the at least one storage medium when operating, wherein at least one of the plurality of computing devices operates the at least one set of instructions and performs the method described above according to the at least one set of instructions when the distributed computing system is operating.
According to some embodiments in the present description, the distributed system includes at least one of a GAS and a Pregel system.
Additional functions of the methods and systems disclosed in one or more embodiments of the present specification will be set forth in part in the description which follows. The following numbers and examples will be apparent to those of ordinary skill in the art from the description. The inventive aspects of the present application may be fully explained by the practice or use of the methods, devices, and combinations described in the following detailed examples.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments of the present description, the drawings that are needed in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments in the present description, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 illustrates a distributed computing system provided in accordance with some embodiments in the present description;
FIG. 2 is a schematic illustration of a computing device that may be applied in the distributed computing system provided in accordance with some embodiments of the present description;
FIG. 3 illustrates a social model and relationships between various entities in the social model in some embodiments of the present description;
FIG. 4 shows a graph structure after social relationship abstraction in the social model; and
FIG. 5 illustrates a deep learning model for distributed computing provided in accordance with some embodiments in the present description.
Detailed Description
The following description provides specific applications scenarios and requirements for one or more embodiments in the present specification to enable one skilled in the art to make and use the content of one or more embodiments in the present specification. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the disclosure. Thus, the present description is not limited to the embodiments shown, but is to be accorded the widest scope consistent with the claims.
The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting. For example, as used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. The terms "comprises," "comprising," "includes," and/or "including," when used in this specification, are taken to specify the presence of stated integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. As used in this specification, the term "a on B" means that a is directly adjacent (above or below) B, or that a is indirectly adjacent (i.e., a and B are separated by some material); the term "A is within B" means that A is entirely within B, or that part A is within B.
These and other features of the present specification, as well as the operation and function of the related elements of structure, as well as the combination of parts and economies of manufacture, may be significantly improved upon in view of the following description. All of which form a part of this specification, reference is made to the accompanying drawings. It is to be expressly understood, however, that the drawings are for the purpose of illustration and description only and are not intended as a definition of the limits of the description. It should also be understood that the drawings are not drawn to scale.
One or more embodiments in the present specification provide a distributed system configured to execute a deep learning model that runs analysis graph data through a distributed computing method.
Fig. 1 illustrates a distributed computing system 100 provided in accordance with an embodiment in the present specification. The distributed computing system 100 may include a set of computing devices C1, C2, … … Ck, where k is an integer greater than 1. The plurality of computing devices may communicate data and information with each other. For example, the plurality of computing devices may be connected together via a network. These computing devices C1, C2, … … Ck share a common work objective, each of which computing devices C1, C2, … … Ck completes one or more sub-objectives of the work objective and passes the results of the execution of the sub-objectives to other computing devices requiring the results of this sub-objective. Each computing device C1, C2, … … Ck may include, in addition to one or more processors, one or more storage media (e.g., memory) to store the results of the operations and other content that may need to be stored when processing the sub-objects. Of course, each computing device C1, C2, … … Ck may also share one or more storage media M (such as memory). The storage medium M may be connected to each of the computing devices C1, C2, … … Ck by a wired or wireless means.
In some embodiments, the distributed computing system 100 may be a GAS system, a PreGel system, or the like, which is not limited herein. It should be noted that "parallel computing" and "distributed computing" are conceptual and have no obvious distinction in this specification. The same system may be referred to as "parallel" and "distributed".
Fig. 2 is a schematic diagram of a computing device 200 provided in accordance with some embodiments of the present description. The computing device 200 may be implemented on the distributed computing system 100. For example, the computing device 200 may be a single computing device on the distributed computing system 100.
The computing device 200 may be a general purpose computer or a special purpose computer. For example, the computing device 200 may be a server, a personal computer, a portable computer (e.g., a notebook computer, a tablet computer, etc.), or an electronic device having other computing capabilities.
For example, the computing device 200 may include a COM port 250, which COM port 250 may be connected to or from a network to which it is connected to facilitate data communications. Computing device 200 may also include a processor 220, such as a Central Processing Unit (CPU), in the form of one or more processors, for executing program instructions. The computing device 200 may also include an internal communication bus 210 and various forms of program storage media and data storage media, such as a magnetic disk 270 (non-transitory memory), and a read-only memory (ROM) 230 or Random Access Memory (RAM) 240, etc., for storing various data files to be processed and/or transferred. The storage medium may be a storage medium local to the computing device 200 or may be a storage medium shared by the computing device 200 (such as the storage medium M shown in fig. 1). The computing device 200 may also include program instructions stored in the ROM 230, the ram 240, and/or other types of non-transitory storage media to be executed by the processor 220. The computing device 200 may also include an I/O component 260 to support data communications with other computing devices in the distributed computing system 100. The computing device 200 may also receive programming and data via network communications.
For illustrative purposes only, only one processor 220 is depicted in the computing device 200. However, those of ordinary skill in the art will appreciate that the computing device 200 in this specification may also include multiple processors. Thus, methods/steps/operations described in this specification as being performed by one processor may also be performed by multiple processors together or separately. For example, if in the present description, the processor of the computing device 200 may perform step a and step B simultaneously. It will be appreciated that step a and step B may also be performed jointly by two different processors. For example, the first processor performs step a, the second processor performs step B, or the first processor and the second processor together perform steps a and B.
A plurality of the computing devices 200 may constitute the distributed computing system 100. The distributed computing system 100 may be used to process graph data.
The graph is a data type aimed at implementing the concepts of undirected graph and directed graph in the field of graph theory in mathematics. The graph structure describes a graph formed from a finite set of nodes (and possibly variable) and edges connecting the finite nodes. In the undirected graph, the edges represent a connection relationship between a pair of nodes that does not contain order information, that is, the connection relationship is nondirectional. In the directed graph, the edges represent a connection relationship between a pair of nodes that includes order information, that is, the connection relationship is directional. The edges are also referred to as arrows in the directed graph. The edges may be assigned values to represent certain labels or attributes, such as cost, capacity, length, relationships, and so forth.
For example, FIG. 3 illustrates a social model and relationships 300 between various entities in the social model. The entities in the social model include user n1, user n2, website n3, website n5, merchant n4, and merchant n6. For example, web site n3 may be a contentThe web site is pushed, on which some articles, news and advertisements are posted. Web site n5 may be a web sales platform (e.g., jindong TM Naughty medicine TM Etc.). User n1 is active on both website n3 and website n 5: he signs up with the web site n3, posts to write articles and get rewards on the web site n3, so that he has business to and from the web site n 3; and meanwhile, the user n1 purchases and consumes on line on the website n 5. User n2 is not online shopping, he simply browses web site n3 and has some business with user n 1. Merchant n4 and merchant n6 each have a respective online store on web site n5 to sell the product. Merchant n4 simultaneously advertises on web site n 3. According to the social relationship 300, the relationship between the entities in the social model is as follows:
merchant n6 has no direct contact with web site n 3; the relationship between merchant n4 and web site n3 is advertising e4; the relationship between the user n1 and the website n3 is that the e2 is posted on the website; and the relationship between the user n2 and the web site n3 is to browse the above content e3. While the relationship between user n2 and user n1 is business e1. The relationship between merchant n6 and web site n5 is that online store e9 is registered thereon. The relationship between merchant n4 and web site n5 is that online store e8 is registered thereon; the relationship between the user n1 and the website n5 is that the user purchases e7 online, and the user completes commodity transaction e5 with merchant n4 and completes commodity transaction e9 with merchant n6 respectively; user n2 is not purchased online at ordinary times and therefore is not in direct contact with web site n 5. Meanwhile, the website n4 and the website n5 have no business transaction and no competition relationship, and are not directly contacted.
For web site n5, the entities that are associated in social relationship 300 shown in FIG. 3 are user n1, merchant n4, and merchant n6. The data for web site n5 itself may include address, asset load information, annual sales, employee count, traffic, and so forth. These data are characteristic data of the website n5 as a node. The association e7 of the website n5 with the user n1, i.e., the "online purchase" action, may include registration information, transaction information related to the commodity transaction e5 and the commodity transaction e6, etc. (such as the commodity of the transaction, the receiving address, and payment information, etc.). The user n1 can purchase goods through the website n5, and score the goods after the transaction is finished. These data may be characteristic data of the association e 7. The data about a node may be characteristic data of the node. The data about a certain side may be characteristic data of the certain side. For web site n5, the node characteristic data and the edge characteristic data may be used to do a number of things, such as evaluating the buying habit of user n1, evaluating/predicting the effect of a business push on user n1, or evaluating the recommendation index of merchant n4, merchant n6 on the web site page, etc.
The social relationship 300 in fig. 3 may be abstracted into a graph structure. Fig. 4 shows the graph structure 400 after the social relationship 300 is abstracted. It should be noted that the diagram structure 400 is merely one example of an invention disclosed in one or more embodiments in the present specification. Those of ordinary skill in the art will appreciate that the graph structure 400 may be an abstract representation of other data structures and thus may be a completely different structural modality than that represented in fig. 4.
The graph structure 400 may include a plurality of nodes n1, n2, n3 … … ni. Where i is an integer greater than 1, representing the number of the plurality of nodes. For example, in graph structure 400, the number of nodes i=6. Each node represents a unit of interest in the system depicted in the graph structure 400, namely user n1, user n2, web site n3, web site n5, merchant n4, and merchant n6 in fig. 3.
The plurality of nodes n1, n2, n3 … … ni are connected to each other by a plurality of edges e1, e2, e3 … … ej. Where j is an integer greater than 1, and represents the number of the plurality of sides. For example, in graph structure 400, the number of edges j=9. Each edge describes the relationship between nodes connected at both ends thereof. The relationship is sometimes unordered (no orientation), sometimes unilateral (single orientation arrow), sometimes bilateral (double arrow). The plurality of edges e1, e2, e3 … … ej correspond to relationships between user n1, user n2, web site n3, web site n5, merchant n4, and merchant n6, respectively. The relationship has been described in describing fig. 3, and will not be described in detail here.
For a target node, the nodes that are connected by edges are their neighbors. Neighboring nodes, such as node n5 in fig. 2, are n1, n4, and n6. Correspondingly, in FIG. 3, the neighboring nodes of web site n5 are user n1, merchant n4, and merchant n6.
From the above example, it is evident that analysing the graph structure requires the use of two basic characteristics of the graph structure: first, each node has its own characteristic data. For example, if a wind control rule for evaluating fraud risk is set up for user n1 with respect to graph structure 400, it is checked whether node characteristic data such as the registration address, IP address, and shipping address of user n1 match. If the feature data do not match, the system determines that the user is at some risk of fraud. This is an application to graph node characteristic information. Second, each node in the graph structure also has structural information. For example, if there are a large number of transaction nodes connected to an IP node for a certain period of time, that is, there are a large number of edges extending from the IP node, the wind control system executing the wind control rule determines that there is a risk of fraud for this IP address. This is an application to graph node structure information. In addition, in determining the risk of fraud for user n1, the risk of fraud for other nodes connected to user n1 is also considered as an evaluation parameter. Accordingly, for the wind control system, the risk of fraud with respect to each node in the graph structure is an indicator to be evaluated by the wind control system. For example, the risk of fraud in the IP address of the user n1 is an indicator that the wind control system needs to evaluate the user n 1.
To analyze a particular target metric for each node in graph structure 400 (such as the fraud risk for each node, the abnormal transaction metrics for each node, the consumption habits of user n1, or the recommended level of website n5 to the merchant on which the store is open, etc., referred to as target metrics), the present description provides, by one or more embodiments, a method to apply a distributed deep learning model to feature extraction and analysis of graph data in graph structure 400. It will be appreciated that different distributed deep learning models may be selected depending on the target metrics. However, for clarity of explanation of the solution in one or more embodiments in this specification, the graph structure 400 will be analyzed below using a graph roll-up neural network (Graph Convolution Network "GCN") of the graph neural networks (Graph Neural Network, "GNN"). Of course, those of ordinary skill in the art will appreciate that other deep learning models, such as a graph-meaning network (Graph Attention Network "GAT") or other types of graph neural networks, may be employed in whole or in part in the analysis of the graph structure 400 without departing from the core spirit of the present description.
Fig. 5 illustrates a deep learning model 500 suitable for distributed computing provided in accordance with some embodiments in the present description. The deep learning model 500 may be executed by the distributed computing system 100 to evaluate target metrics of the graph structure 400. Since the graph structure 400 is executing in the distributed computing system 100, the distributed computing system 100 may allocate at least one computing device for each node in the graph structure 400. That is, each node in the graph structure 400 corresponds to at least one computing device in the distributed computing system 100.
The deep learning model 500 may be used to evaluate target metrics of the graph structure 400. For example, in some embodiments, the graph structure 400 may be a topology of the relationships 300 between various entities in the social model, and the target index may be a transaction risk index for each node. Because the nature of the different nodes may be the same or different, the evaluation algorithms for the risk indices of transactions may be the same or different. For example, n1 is the user and n6 is the merchant. The assessment algorithm for the transaction risk index for node n1 may be different from the assessment algorithm for the transaction risk index for node n6. In some embodiments, the transaction risk index for a node in graph structure 400 may be obtained by calculating some weighted average (corresponding to the convolution operation in the GCN) of the transaction risk indices for all neighboring nodes (including the node itself) of the node within m hops in graph structure 400, and then evaluating the weighted average itself. Where m is a natural number greater than 0. The m hops refer to m-layer proximity relations. For example, in the graph structure 400, since the nodes n1, n4, and n6 are all directly connected to the node n5, the neighboring node of one hop of the node n5, that is, m=1, is simply called neighboring node, and is n1, n4, and n6. Similarly, n 1's neighboring nodes have n2, n3, n6, and n4 in addition to n 5. Of these nodes, n2 and n3 are both separated from node n5 by node n1 and are therefore two-hop neighbors of n5, i.e., m=2. Thus, all other nodes in graph structure 400 are nodes within two hops of node n 5.
The deep learning network may comprise an iterative process followed by a neural network. The process propagates the node states until an equilibrium or preset state is reached. Accordingly, the deep learning model 500 may include an initialization module 510 and an iteration module 530. The neural network will then generate an output for each node based on its state. Accordingly, the deep learning model 500 may also include an evaluation module 550. Further, the deep learning network may be applied to a single node in the graph structure 400, or may be applied to all nodes in the graph structure 400.
The initialization module 510 may be one or more sets of instructions stored in the non-transitory storage medium of the distributed computing system 100. When the distributed computing system 100 is running, at least one of the plurality of computing devices communicates with the non-transitory storage medium, reads the one or more sets of instruction sets, and executes the initialization module 510 according to the one or more sets of instruction sets. The target node may be a node in the graph structure 400 corresponding to the at least one computing device. For example, if the target node is node n5, the at least one computing device is a computing device in the distributed computing system 100 responsible for operating on node n 5. When the at least one computing device executes the initialization module 510, the at least one computing device performs initialization operations comprising:
S511, obtaining an initial target graph embedding vector of the target node in the graph structure.
The target graph embedding vector ("meshing") may be a particular graph embedding vector of the target node that is required in evaluating the target metrics. For example, in the example where the target indicator is the transaction risk index, the target graph embedding vector may be a graph embedding vector that needs to be calculated with respect to a target node in an algorithm employed to evaluate the transaction risk index of the nodes in the graph structure 400.
When the at least one computing device performs the initialization operation, an initial value of the target graph embedding vector, referred to as an initial target graph embedding vector, is obtained. In particular, to obtain an initial target graph embedding vector for the target node, the at least one computing device may:
and obtaining an initial node map embedding vector of the target node. For example, the at least one computing device may first acquire/receive feature data of the target node and then generate the initial graph embedding vector of the target node by running a first built-in algorithm of the initialization module 510 based on the feature data of the target node. The different target metrics may correspond to different first built-in algorithms. Taking the website n5 as a target node for example, when evaluating the risk index of the transaction of the website n5, the characteristic data may include the IP address of the website n5, the address of the transaction, the expenditure, the income, the liability information, and the like. Accordingly, the first built-in algorithm may be part of the algorithm for evaluating risk index of transactions, such as preliminary parameters for risk of transactions to web site n5 based on data such as IP address of web site n5, address of transactions, spending, income, liability information, etc. The initial map embedding vector may be the preliminary parameters obtained based on the above-described feature data. The preliminary parameters may reflect the transaction condition of website n 5.
And obtaining an initial edge map embedding vector of an edge corresponding to the target node. For example, the at least one computing device may obtain/receive edge feature data corresponding to the target node, and generate the initial edge map embedding vector by running a second built-in algorithm of the initialization module based on the edge feature data. For example, in the example of the transaction risk index of the nodes in the evaluation graph structure 400 described above, the edge feature data may be the total number of website users and the daily activity of website users (corresponding to e 7), the number of website advertisements within a predetermined period of time, the click-through rate of each advertisement (corresponding to e8, e 9), and so on. The second built-in algorithm may be part of the evaluation transaction risk index algorithm, such as may be some performance algorithm. The initial edge map embedding vector may be a result of performing the performance algorithm based on the feature data. The results may reflect the market performance of website n 5.
Then, the aggregation of the initial node map embedding vector and the initial edge map embedding vector is performed, namely, the target map embedding vector of the target node is generated by running the initialization module based on the initial node map embedding vector and the initial edge map embedding vector. For example, in the example of the transaction risk index of the node in the evaluation graph structure 400, the at least one computing device may execute the third built-in algorithm in the initialization module after obtaining the transaction condition and the market performance condition of the website n5, and synthesize the transaction condition and the market performance condition to generate a target graph embedded vector value of the target node, where the target graph embedded vector value is a parameter value related to the transaction risk index and represented by the current integrated condition of the website n 5.
In some embodiments, the first built-in algorithm, the second built-in algorithm, and the third built-in algorithm may be the same algorithm or different algorithms. In some embodiments, the target graph embedding vector comprises a floating point number vector of a preset dimension. For example, the target graph embedding vector may be a 32-bit floating point number vector of a preset dimension. The initial target graph embedding vector has the same form as the target graph embedding vector.
It will be appreciated that since the graph structure 400 is executing in the distributed computing system 100, the distributed computing system 100 may calculate an initial target graph embedding vector in parallel for each node in the graph structure 400 during execution of the initialization module 510.
The initial target-graph-embedded vector will be used to input the iteration module 530 to generate a new target-graph-embedded vector, and the target-graph-embedded vector is updated continually by repeated iterations.
The iteration module 530 may be one or more sets of instructions stored in the non-transitory storage medium of the distributed computing system 100. When the distributed computing system 100 is running, at least one of the plurality of computing devices communicates with the non-transitory storage medium, reads the one or more sets of instruction sets, and executes the iteration module 530 in accordance with the one or more sets of instruction sets. When the at least one computing device executes the iteration module 530, the at least one computing device iterates over the target graph embedding vector by repeatedly performing a graph embedding vector update operation. The graph embedding vector update operation includes:
S531: and obtaining the target graph embedded vector of the target node in the graph structure.
As previously described, the graph structure 400 includes a plurality of nodes, each node being connected by an edge to at least one neighboring node, the target node being the node of the plurality of nodes for which the at least one computing device is responsible for computing. The target graph embedding vector is the initial target graph embedding vector of the target node when the iteration module 530 is first run.
S533: a neighboring target graph embedding vector is received from at least one associated computing device in the distributed computing system.
As previously described, the adjacent target graph embedding vector may be a target graph embedding vector of at least one adjacent node of the target node. For example, in the graph structure 400 in fig. 4, when the target node is n5, the adjacent target graph embedding vectors thereof may include a target graph embedding vector of the node n1, a target graph embedding vector of the node n4, and a target graph embedding vector of the node n6, or target graph embedding vectors corresponding to one or more nodes of n1, n4, and n 6. Accordingly, the associated computing device is a computing device in the distributed computing system 100 that performs operations on the target neighboring node.
S535: and generating an iterated target graph embedding vector by running an operation module in the deep learning model based on the target graph embedding vector and the adjacent target graph embedding vector.
As previously described, the deep learning model is a target index evaluation model based on the graph structure 400; the operation module is used for determining a target graph embedded vector value part of the target node in the deep learning model. In some embodiments, the operation module in the deep learning model may include a graph neural network or a portion of the neural network, such as one layer, multiple layers, or any combination of graph neural network GNN, graph rolling network GCN, graph annotation network GAT. For example, the operation module in the deep learning model may include one or more layers of operations in the GCN.
For example, in the example of evaluating the transaction risk index of a node in graph structure 400 described above, the operation module may include a graph aggregation calculation ("pooling") operation on the target graph embedded vector and the neighboring target graph embedded vector. The graph aggregation calculation may be a maximum value graph aggregation calculation (max graph aggregation calculation) or an average value graph aggregation calculation (average graph aggregation calculation).
For another example, when the deep learning model is a graph roll-up neural network GCN, the operation module in the deep learning model may include an L-layer operation in the GCN, where L is a natural number. The value of L may be calculated by a specific formula or may be obtained empirically, and is not limited in this specification. The output of the GCN can be expressed as:
Z＝H ()
wherein Z can also be the output of the graph level; h (l) ∈R N×D Is an input to the layer-1 network (initial input is H (0) =x); n is the number of nodes in the graph data 400 that need to be considered on the one computing device; the graph embedding vector for each node is represented using a feature vector H in the D dimension. A is an adjacency matrix (adjacency) of graph structure 400matrix)。To add a self-connected adjacency matrix A so that each node still maintains its own characteristics in the next level of operation; />A degree matrix (degree matrix) is used for the +.>Normalization is performed on->W (l) ∈R D×D Is a trained parameter matrix. Sigma is the corresponding activation function, e.g. ReLU (·) or max (0, ·).
The at least one computing device may generate a new iterated target graph embedding vector using the target graph embedding vector and the adjacent target graph embedding vector, via the computation of the computing module.
S537: and transmitting the iterated target graph embedding vector to the associated computing equipment. In this step, the at least one computing device propagates the target graph embedding vector for the target node to associated computing devices within a hop.
S539: and assigning the iterated target graph embedding vector to the target graph embedding vector, and repeatedly executing the graph embedding vector updating operation, namely steps S531-S539, until the target graph embedding vector iterates for a preset number of times.
Next, the at least one computing device may perform a preset number of the graph embedding vector update operations to generate a final graph embedding vector for the target node. It will be appreciated that since the graph structure 400 is implemented in the distributed computing system 100, the distributed computing system 100 executes the iteration module 530 in parallel for each node in the graph structure 400, iterating through the respective target graph embedding vectors during execution of the iteration module 530. Thus, each iteration is performed at step S533 in the graph embedding vector update operation, i.e., when a neighboring target graph embedding vector is received from at least one associated computing device in the distributed computing system 100, the neighboring target graph embedding vector is also iterated. Since the target-graph embedding vector updating operation needs to generate the iterated target-graph embedding vector based on the target-graph embedding vector and the adjacent target-graph embedding vector in step S535 every iteration, the content of the target-graph embedding vector is added with the content of the target-graph embedding vector corresponding to the new one-hop node every time a new iteration is performed. For example, the initial target graph embedding vector for target node n5 includes only information about the target node itself to connect with the edges of the target node. After the first round of target graph embedding vector iteration, step S535 introduces neighboring target graph embedding vectors corresponding to the one-hop neighboring nodes n1, n4, n6 of the target node n5 to generate an iterated target graph embedding vector. Meanwhile, the target graph embedding vector corresponding to the node n1 also introduces the target graph embedding vector corresponding to the adjacent nodes n2, n3, n5 and n6 of the node n1 in the wave iteration. And nodes n2, n3 are two-hop neighbor nodes to target node n 5. Naturally, when the second round of target graph embedding vectors iterate, the target graph embedding vector of the target node n5 includes the content of the target graph embedding vector corresponding to the two-hop adjacent nodes n2 and n 3.
The number of iterations may be based on the size of the graph structure 400 or may be a preset number. For example, in the example of evaluating the transaction risk indices of nodes in the graph structure 400, since the evaluation of transaction risk involves only a certain weighted average of the transaction risk indices of all neighboring nodes within two hops of the node, the graph embedding vector update operation need only be iterated twice.
After a preset number of iterations of the graph embedding vector update operation, the at least one computing device generates a final graph embedding vector for the target node. In some embodiments, the distributed computing system 100 may generate a final graph embedding vector for all nodes in the graph structure 400 after a predetermined number of graph embedding vector update operations.
Next, the at least one computing device may input the final graph embedding vector for the node (including the final graph embedding vector for the target node) to the evaluation module 550 for scoring of the final target index.
The evaluation module 550 may be one or more sets of instructions stored in the non-transitory storage medium of the distributed computing system 100. When the distributed computing system 100 is running, at least one of the plurality of computing devices communicates with the non-transitory storage medium, reads the one or more sets of instruction sets, and executes the evaluation module 550 according to the one or more sets of instruction sets. The evaluation module 550 is configured to translate the target graph embedding vector into a score of the target indicator for the corresponding node in the graph structure 400. When the at least one computing device executes the evaluation module 550, the at least one computing device performs a target metric scoring operation. The target index scoring operation includes:
S551: and operating the evaluation module based on the final graph embedding vector to generate a target index score corresponding to the final graph embedding vector. In some embodiments, the evaluation module may be a neural network. The neural network will generate an output for each input node based on its state. For example, the neural network may be a fully connected layer; the neural network may also be part of the structure of various suitable neural networks or part of neural networks, such as one or more layers of RNN, CNN or other types of neural networks. The description is not intended to be limiting.
S553: and outputting the target index score.
In summary, the method provided in the present specification performs distributed computation on the deep learning model of the analysis chart data through homogeneous stand-alone operation. By performing model understanding and model transformation on the deep learning model, one or more embodiments in the specification split the deep learning model into an initialization module, an iteration module and a final evaluation module, so that the calculation of each node in the graph data in the whole deep learning model is separately and parallelly processed by a single machine in the distributed system, and the calculation structure of each node is merged and input into a neural network for final calculation evaluation after the processing is completed. Thus, each single machine in the distributed system requires little memory to process the large-scale map data. And because distributed computing breaks up data, one or more embodiments presented in this specification can process the large-scale graph data without losing any data.
When the method disclosed by one or more embodiments in the specification is used for predicting the target index of the large-scale graph data, the following beneficial effects are achieved.
1. The prediction result is accurate. In the process of calculating the large-scale graph neural network by a single machine, because the memory occupied by the multi-degree neighbors is too much, sampling and cutting have to be performed, and the model effect is reduced by reducing input data, so that the memory is saved. By distributing the graph data in the graph computing system, each computing node only calculates a small number of nodes, so that the occupied memory is not increased due to multi-degree neighbors, and the multi-degree neighbors only calculate the increase of iteration rounds and have no change on memory consumption. Therefore, the prediction process does not need to sample, does not lose the graph information and is not approximate, and therefore the most accurate model prediction service effect can be obtained.
2. The calculation performance is improved. In the process of calculating the large-scale graph neural network by the single machine, graph data are repeated among different single machines, and each single machine needs to be repeatedly calculated, so that a large number of redundant repeated calculations are generated. While one or more embodiments in the present description disclose a graph-in-a-graph computing system that distributes output to each node, each node performs new computations in a propagation graph embedding direction. Each node performs only one calculation. Therefore, one or more embodiments in the present specification disclose a solution that consumes less resources and operates faster. Very low CPU, MEM, IO. For the PPI data 3-hop neighbor scheme, compared with the deep learning model scheme, the resource is reduced by 93%, the resource consumption does not increase exponentially with the number of neighbor hops but increases linearly, and the network transmission does not increase exponentially with the number of hops but is constant. CPU, MEM, IO saves 25%35%67% respectively in the full-image prediction process of the PD model of the micro-lending enterprise. And in the GraphFlat+UDF prediction, the GraphFlat clipping only keeps 13% of the average length, and the rabbit shedding is the accurate prediction.
The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. Furthermore, the processes depicted in the accompanying drawings do not necessarily have to be in the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. Thus, after reading this detailed disclosure, those skilled in the art will appreciate that the foregoing detailed disclosure may be presented by way of example only and may not be limiting. Although not explicitly described herein, one of ordinary skill in the art will appreciate that one or more of the embodiments in the present specification is intended to embrace a variety of reasonable alterations, improvements and modifications to the embodiments. Such alterations, improvements, and modifications are intended to be proposed by this specification, and are intended to be within the spirit and scope of the exemplary embodiments of this specification.
Furthermore, certain terms in one or more embodiments in the specification have been used to describe embodiments of the specification. For example, "one embodiment," "an embodiment," and/or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the present description. Thus, it is emphasized and should be appreciated that two or more references to "an embodiment" or "one embodiment" or "an alternative embodiment" in various portions of this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined as suitable in one or more embodiments of the invention.
It should be appreciated that in the foregoing description of embodiments of the present specification, one or more embodiments of the specification combine various features into a single embodiment, figure, or description thereof for the purpose of simplifying the specification. However, it is not necessary to say that a combination of these features, and it is entirely possible for a person skilled in the art to extract some of them as separate embodiments when reading one or more embodiments in this description. That is, embodiments in this specification may also be understood as an integration of multiple secondary embodiments. While each secondary embodiment is satisfied by less than all of the features of a single foregoing disclosed embodiment.
Each patent, patent application, publication of patent application, and other materials, such as articles, books, specifications, publications, documents, articles, etc., cited herein are hereby incorporated by reference. The entire contents for all purposes, except for any prosecution file history associated therewith, may be any identical prosecution file history inconsistent or conflicting with this file, or any identical prosecution file history which may have a limiting influence on the broadest scope of the claims. Now or later in association with this document. For example, if there is any inconsistency or conflict between the description, definition, and/or use of terms associated with any of the incorporated materials, the terms in the present document shall prevail.
Finally, it is to be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments in this specification. Other modified embodiments are also within the scope of this specification. Accordingly, the embodiments disclosed herein are by way of example only and not limitation. Those skilled in the art can adopt alternative arrangements to achieve the same technical result in accordance with the embodiments in the specification. Therefore, the embodiments of the present specification are not limited to the embodiments precisely described in the application.
Claims (13)
1. A distributed computing method for running a deep learning model for analyzing graph structures, comprising:
performing, by at least one computing device in the distributed computing system, a graph embedding vector update operation for the graph structure, wherein the graph structure includes a plurality of nodes, each node being connected by an edge to at least one neighboring node,
wherein the graph embedding vector update operation includes, for each target node in the graph structure:
obtaining a target graph embedding vector value of a target node in the graph structure, wherein the target graph embedding vector value is related to an initial edge graph embedding vector of an edge corresponding to the target node,
receiving a neighbor target graph embedding vector from at least one associated computing device in the distributed computing system, the neighbor target graph embedding vector being a target graph embedding vector of at least one neighbor node of the target node, the associated computing device being a computing device in the distributed computing system performing an operation on the target neighbor node,
Generating an iterated target graph embedding vector by running an operation module in the deep learning model based on the target graph embedding vector and an adjacent target graph embedding vector, and
repeating the updating operation until the target graph embedding vector is iterated for a preset number of times, wherein the deep learning model comprises a target index evaluation model based on the graph structure, and the operation module is used for determining a part of the deep learning model for determining the target graph embedding vector value of the target node; and
and transmitting the iterated target graph embedding vector to the associated computing equipment.
2. The method as recited in claim 1, wherein the operation module comprises at least:
at least one layer in the graph neural network model;
at least one layer in the graph rolling neural network model; and/or
The drawing is intended for at least one layer in the network model.
3. The method of claim 1, wherein the target graph embedding vector comprises a floating point number vector of a preset dimension.
4. The method of claim 1, wherein the obtaining the target graph embedding vector for the target node in the graph structure comprises, by at least one computing device in the distributed computing system:
Acquiring an initial node map embedding vector of the target node;
acquiring the initial edge map embedding vector of the edge corresponding to the target node;
generating an initial target graph embedding vector of the target node by running the initialization module based on the initial node graph embedding vector and the initial edge graph embedding vector; and
and taking the initial target graph embedding vector as the target graph embedding vector.
5. The method as recited in claim 4, wherein:
the obtaining the initial graph embedding vector of the target node comprises the following steps: acquiring characteristic data of the target node, and generating the initial graph embedding vector of the target node based on the characteristic data of the target node; and
the obtaining the initial edge map embedding vector of the edge corresponding to the target node comprises the following steps: and acquiring edge characteristic data corresponding to the target node, and generating the initial edge map embedding vector based on the edge characteristic data.
6. The method of claim 1, further comprising, by the at least one computing device in the distributed computing system:
and executing the graph embedding vector updating operation for preset times, and generating a final graph embedding vector of the target node.
7. The method of claim 6, wherein the deep learning model further comprises an evaluation module configured to translate the target graph embedding vector into a score for the target indicator for a corresponding node in the graph structure; and
the method further comprises the steps of:
operating the evaluation module based on the final graph embedding vector to generate a target index score corresponding to the final graph embedding vector; and
and outputting the target index score.
8. The method of claim 1, wherein the evaluation module comprises a fully connected layer of at least one neural network.
9. The method of claim 1, wherein the target metrics comprise at least one of a transaction risk index, a fraud risk, an abnormal transaction index for each node, a consumption habit of a user, or a recommendation level of a merchant associated with a node in the graph structure.
10. The method of claim 1, wherein the target metrics include a transaction risk index for each node in the graph structure,
the generating, by running an operation module in the deep learning model, the iterated target graph embedding vector based on the target graph embedding vector and the adjacent target graph embedding vector includes: and calculating a weighted average value of transaction risk indexes of all the target nodes including all the neighboring nodes of the target node within m hops of the graph structure through convolution operation in the graph convolution neural network, and then evaluating the weighted average value, wherein m is a natural number larger than 0.
11. The method of claim 5, wherein the target metrics include a transaction risk index for each node in the graph structure,
the nodes of the graph structure include websites,
the edge characteristic data includes at least one of a total number of the website users and a daily activity of the website users or a number of website advertisements within a predetermined period of time and a click rate of each advertisement,
the initial edge map embedding vector includes results of performance algorithms based on the feature data, which may reflect market performance of the website.
12. A distributed computing system for evaluating target metrics of a deep learning model by distributed computing, comprising:
at least one storage medium stores at least one set of instructions for executing a deep learning model of the analytic graph structure;
a plurality of computing devices, each of the plurality of computing devices being communicatively coupled to at least one other computing device and in communication with the at least one storage medium at a time of operation,
wherein at least one computing device of the plurality of computing devices, when the distributed computing system is running, runs the at least one set of instructions and performs the method of any of claims 1-11 in accordance with the at least one set of instructions.
13. The distributed computing system of claim 12, wherein the distributed system comprises at least one of a GAS and Pregel system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010123789.5A CN111275176B (en) | 2020-02-27 | 2020-02-27 | Distributed computing method and distributed computing system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010123789.5A CN111275176B (en) | 2020-02-27 | 2020-02-27 | Distributed computing method and distributed computing system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111275176A CN111275176A (en) | 2020-06-12 |
| CN111275176B true CN111275176B (en) | 2023-09-26 |
Family
ID=70999235
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010123789.5A Active CN111275176B (en) | 2020-02-27 | 2020-02-27 | Distributed computing method and distributed computing system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111275176B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112486641A (en) * | 2020-11-18 | 2021-03-12 | 鹏城实验室 | Task scheduling method based on graph neural network |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108804576A (en) * | 2018-05-22 | 2018-11-13 | 华中科技大学 | A kind of domain name hierarchical structure detection method based on link analysis |
| CN109194707A (en) * | 2018-07-24 | 2019-01-11 | 阿里巴巴集团控股有限公司 | The method and device of distribution figure insertion |
| WO2019114412A1 (en) * | 2017-12-15 | 2019-06-20 | 阿里巴巴集团控股有限公司 | Graphical structure model-based method for credit risk control, and device and equipment |
| CN109918454A (en) * | 2019-02-22 | 2019-06-21 | 阿里巴巴集团控股有限公司 | The method and device of node insertion is carried out to relational network figure |
| CN110008348A (en) * | 2019-01-29 | 2019-07-12 | 阿里巴巴集团控股有限公司 | The method and apparatus for carrying out network insertion in conjunction with node and side |
| CN110263227A (en) * | 2019-05-15 | 2019-09-20 | 阿里巴巴集团控股有限公司 | Clique based on figure neural network finds method and system |
| CN110490274A (en) * | 2019-10-17 | 2019-11-22 | 支付宝(杭州)信息技术有限公司 | Assess the method and device of alternative events |
| CN110782044A (en) * | 2019-10-29 | 2020-02-11 | 支付宝(杭州)信息技术有限公司 | Method and device for multi-party joint training of neural network of graph |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10572510B2 (en) * | 2015-12-21 | 2020-02-25 | Sap Se | Distributed database transaction protocol |
-
2020
- 2020-02-27 CN CN202010123789.5A patent/CN111275176B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019114412A1 (en) * | 2017-12-15 | 2019-06-20 | 阿里巴巴集团控股有限公司 | Graphical structure model-based method for credit risk control, and device and equipment |
| CN108804576A (en) * | 2018-05-22 | 2018-11-13 | 华中科技大学 | A kind of domain name hierarchical structure detection method based on link analysis |
| CN109194707A (en) * | 2018-07-24 | 2019-01-11 | 阿里巴巴集团控股有限公司 | The method and device of distribution figure insertion |
| CN110008348A (en) * | 2019-01-29 | 2019-07-12 | 阿里巴巴集团控股有限公司 | The method and apparatus for carrying out network insertion in conjunction with node and side |
| CN109918454A (en) * | 2019-02-22 | 2019-06-21 | 阿里巴巴集团控股有限公司 | The method and device of node insertion is carried out to relational network figure |
| CN110263227A (en) * | 2019-05-15 | 2019-09-20 | 阿里巴巴集团控股有限公司 | Clique based on figure neural network finds method and system |
| CN110490274A (en) * | 2019-10-17 | 2019-11-22 | 支付宝(杭州)信息技术有限公司 | Assess the method and device of alternative events |
| CN110782044A (en) * | 2019-10-29 | 2020-02-11 | 支付宝(杭州)信息技术有限公司 | Method and device for multi-party joint training of neural network of graph |
Non-Patent Citations (2)
| Title |
|---|
| Bryan Perozzi.et al."DeepWalk: Online Learning of Social Representations".《arXiv: 1403.6652v2》.2014,全文. * |
| 陈思等."基于非均匀邻居节点采样的聚合式图嵌入方法".《浙江大学学报（龚学版）》.2019,第53卷(第11期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111275176A (en) | 2020-06-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11809993B2 (en) | Systems and methods for determining graph similarity | |
| US11227190B1 (en) | Graph neural network training methods and systems | |
| CN111291139A (en) | Attention mechanism-based knowledge graph long-tail relation completion method | |
| CN111985622A (en) | Graph neural network training method and system | |
| CN107786943A (en) | A kind of tenant group method and computing device | |
| CN114372573B (en) | User portrait information recognition method and device, computer equipment and storage medium | |
| Li et al. | A utility-based link prediction method in social networks | |
| US10922627B2 (en) | Determining a course of action based on aggregated data | |
| US20220138502A1 (en) | Graph neural network training methods and systems | |
| CN112214499B (en) | Graph data processing method and device, computer equipment and storage medium | |
| CN115618098B (en) | Cold-chain logistics recommendation method and device based on knowledge enhancement and cavity convolution | |
| CN113255798A (en) | Classification model training method, device, equipment and medium | |
| US20230133683A1 (en) | Feature interaction via edge search | |
| US20230267317A1 (en) | Sign-aware recommendation apparatus and method using graph neural network | |
| Wu et al. | Estimating fund-raising performance for start-up projects from a market graph perspective | |
| CN117391313A (en) | Intelligent decision method, system, equipment and medium based on AI | |
| Lux et al. | Interpolation of sparse high-dimensional data | |
| CN111275176B (en) | Distributed computing method and distributed computing system | |
| Yang et al. | A robust method for inferring network structures | |
| Martineau et al. | Graph matching as a graph convolution operator for graph neural networks | |
| Khare et al. | Bayesian inference for Gaussian graphical models beyond decomposable graphs | |
| CN115345291A (en) | Graph processing method and related device | |
| WO2022166125A1 (en) | Recommendation system with adaptive weighted baysian personalized ranking loss | |
| Yin et al. | A hybrid method for fast detection of spatial disease clusters in irregular shapes | |
| CN111428741B (en) | Network community discovery method and device, electronic equipment and readable storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |