A Normalized Graph Model Construction Method

Abstract

The present invention discloses a method for constructing a normalized graph model, comprising the following steps: S1. Data exploration: using query statements and visualization tools to analyze business data and determine the potential relationship between business data; S2. Node and edge relationship construction: determining the node and edge types of a graph based on the result of the data exploration, thereby constructing a graph schema; S3. Graph data construction: loading the data of corresponding nodes and edges based on the graph schema, and then encoding and combining the attributes of nodes and edges to form a graph required by the graph model; S4. Model selection: selecting the corresponding paradigm for modeling according to the type of business problem; S5. Model parameter setting: setting model parameters, including the number of hidden layers and loss function; S6. Model tuning: adjusting model parameters to optimize the model. The present invention allows users to focus on business data without having to worry about specific modeling details.

Description

Technical Field

[0001] The present invention relates to the field of graphic data processing technology, and in particular to a method for constructing a normalized graphic model. Background Art

[0002] A graph is a structure used to represent entities and their relationships, which can be written as G = (V, E). A graph consists of two sets, one is the set of nodes V, and the other is the set of edges E. At the same time, nodes and edges also have their own attributes.

[0003] Since graphs have strong expressive power, GNN (graph neural network) has received extensive attention in the field of machine learning. The main purpose of the GNN architecture is to learn an embedding that contains information about its neighborhood and solve some node and edge prediction problems based on this embedding.

[0004] From the perspective of the current development of artificial intelligence, with the continuous improvement of computing power and the continuous upgrading of storage methods, computational intelligence with fast computing and memory storage capabilities has been initially realized. Based on the development of computational intelligence and perceptual intelligence, artificial intelligence is extending to cognitive intelligence such as analysis, thinking, understanding, and judgment, and real intelligent solutions have already emerged.

[0005] In addition, the data analysis industry will also usher in changes. The proportion of Chinese companies using graph databases continues to rise. The Gartner report points out that by 2025, graph technology will account for 80% of data analysis innovation. As a traditional relational database optimization direction, GraphAI can provide deeper insights and better business results after the industry data volume accumulates to a certain level.

[0006] At present, graph deep learning methods can be roughly divided into two categories: semi-supervised methods and unsupervised methods. Specifically, semi-supervised methods include graph neural networks (GNNs) and graph convolutional networks (GCNs), such as Figure 1 The figure shows a schematic diagram of the existing semi-supervised learning of node-level classification. Unsupervised methods mainly include graph autoencoders (GAE). From the perspective of model users, it is necessary not only to consider the design of the graph structure and the processing of node labels and features, but also to select the model and continuously adjust the various parameters of the model. This places high demands on users, making it difficult to combine business data with model data, making GraphAI technology difficult to implement in many scenarios. Summary of the invention

[0007] In response to the current problem that GraphAI is difficult to implement in many scenarios, the present invention proposes a normalized graph model construction method that allows users to focus on business data without having to worry about specific modeling details.

[0008] The technical solution adopted by the present invention is as follows:

[0009] A method for constructing a normalized graph model includes the following steps:

[0010] S1. Data exploration: Use query statements and visualization tools to analyze business data and determine the potential relationships between business data;

[0011] S2. Node and edge relationship construction: Determine the node and edge types of the graph based on the results of the data exploration, thereby constructing a graph schema;

[0012] S3. Graph data construction: Based on the graph schema, the data of corresponding nodes and edges are loaded from the distributed file system, and then the attributes of the nodes and edges are encoded and combined to form the graph required by the graph model;

[0013] S4. Model selection: Select the corresponding paradigm for modeling according to the type of business problem;

[0014] S5. Model parameter setting: set model parameters, including the number of hidden layers and loss function;

[0015] S6. Model tuning: Adjust model parameters to make the model optimal.

[0016] Furthermore, the paradigm includes an attribute conduction paradigm, and the modeling method of the attribute conduction paradigm includes the following steps:

[0017] S401. Aggregation: After receiving the information of neighboring nodes, the node aggregates it with its own information according to different weights to form a more comprehensive information expression;

[0018] S402. Non-linear transformation: performing non-linear transformation on the expression obtained by the aggregation as the expression of the node at this layer;

[0019] S403. Propagation: Propagate the node expression obtained after the nonlinear transformation to neighboring nodes, and each node executes steps S401 to S403 in a loop.

[0020] Furthermore, the paradigm includes a vectorized matching paradigm, and the modeling method of the vectorized matching paradigm includes: an aggregation process, in which each node aggregates neighbor information in a process similar to the attribute conduction paradigm, but the model learning goal is no longer to fit the label of the sample, but to minimize the expression of related things and maximize the expression of unrelated things; after the training is completed, the model will obtain the calculation method of the expression of things in the corresponding context, and the similarity of the expression of things reflects the correlation of things.

[0021] Furthermore, if no negative correlation is predefined, some pairs of things will be randomly generated as irrelevant relationships.

[0022] Furthermore, the K most relevant targets of a thing are retrieved through the hash inverted index.

[0023] Furthermore, the paradigm includes a causal reasoning paradigm, and the modeling method of the causal reasoning paradigm includes: obtaining the conditional probability of each node by performing maximum likelihood estimation on the values ​​of each event in the training sample, and using breadth-first search to find the graph structure that best fits the causal relationship of each event, and after obtaining the graph structure and the conditional probability of each node, a probabilistic graph model is formed.

[0024] Furthermore, when prediction is performed according to the probabilistic graph model, the value probability of the node to be queried can be calculated after inputting certain node observation values.

[0025] Furthermore, when the hypothesis is calculated according to the probabilistic graphical model, based on the causal structure learned by the model and by blocking confounding factors, the conditional probability of the query node when the intervention node takes the intervention value is calculated, that is, the probability of the query node taking the intervention value.

[0026] The beneficial effects of the present invention are:

[0027] (1) The present invention opens up the complex process from business data to graph data, so that users do not need to pay attention to how the underlying graph data is implemented and stored. They only need to focus on their own business data, discover the correlation between each type of business data, and construct a graph schema based on these correlations. This greatly reduces the difficulty of use for users.

[0028] (2) The present invention can encapsulate the details of the graph model to achieve automated modeling. The model type and model parameters are encapsulated, allowing users to start from the actual problem and select the corresponding paradigm to model according to the problem type, thereby ignoring the underlying details of the model and lowering the modeling threshold.

[0029] (3) The present invention simplifies the modeling process and lowers the modeling threshold, allowing users to quickly experiment and obtain feedback.

[0030] (4) The present invention allows users to enjoy the intelligent capabilities brought by Graph AI by directly using or combining computing paradigms without having to think about complex model construction ideas and data processing ideas. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the existing semi-supervised learning of node-level classification.

[0032] Figure 2 This is one of the flow charts of the method for constructing a normalized graph model according to an embodiment of the present invention.

[0033] Figure 3This is the second flow chart of the method for constructing a normalized graph model according to an embodiment of the present invention.

[0034] Figure 4 It is a paradigm example diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, the specific implementation methods of the present invention are now described. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0036] like Figure 2 As shown, this embodiment provides a method for constructing a normalized graph model, including a composition process and a modeling process, wherein the composition process includes the following steps:

[0037] S1. Data exploration: Use query statements and visualization tools to analyze business data and determine the potential relationships between business data;

[0038] S2. Node and edge relationship construction: Determine the node and edge types of the graph based on the results of the data exploration, thereby constructing a graph schema;

[0039] S3. Graph data construction: Based on the graph schema, the corresponding node and edge data are loaded from the distributed file system, and then the attributes of the nodes and edges are encoded and combined to form the graph required by the graph model.

[0040] The modeling process includes the following steps:

[0041] S4. Model selection: Select the corresponding paradigm for modeling according to the type of business problem (such as classification problem, recommendation problem, etc.). The specific implementation process is to split the original graph nodes into training sets and test sets in proportion for training and verification.

[0042] S5. Model parameter setting: Set model parameters, such as the number of hidden layers and loss function.

[0043] S6. Model tuning: Adjust model parameters through neural network controllers such as AutoML to make the model optimal.

[0044] Preferably, if Figure 3 As shown, the normalized graph model construction method implemented in this embodiment also includes an ETL process, which mainly performs data collection, data cleaning and data processing.

[0045] Preferably, this implementation provides three paradigms, including attribute conduction paradigm, vectorized matching paradigm and causal reasoning paradigm, such as Figure 4 Shown is an example of the use of the paradigm.

[0046] Specifically, the modeling method of the attribute conduction paradigm includes the following steps:

[0047] S401. Aggregation: After receiving the information of neighboring nodes, the node aggregates it with its own information according to different weights to form a more comprehensive information expression;

[0048] S402. Non-linear transformation: performing non-linear transformation on the expression obtained by the aggregation as the expression of the node at this layer;

[0049] S403. Propagation: Propagate the node expression obtained after the nonlinear transformation to neighboring nodes, and each node executes steps S401 to S403 in a loop.

[0050] Specifically, the modeling method of the vectorized matching paradigm includes: the aggregation process, the process of aggregating neighbor information at each node is similar to the attribute conduction paradigm, but the model learning goal is no longer to fit the label of the sample, but to minimize the expression of related things and maximize the expression of irrelevant things. If there is no pre-defined negative correlation, some pairs of things will be randomly generated as irrelevant relationships. After the training is completed, the model will obtain the calculation method of the expression of things in the corresponding context, and the similarity of the expression of things reflects the relevance of things. Based on this expression, the hash inverted index is used to help users quickly retrieve the K most relevant targets of a certain thing.

[0051] Specifically, the modeling method of the causal reasoning paradigm includes: obtaining the conditional probability of each node by performing maximum likelihood estimation on the values ​​of each event in the training sample, and using breadth-first search to find the graph structure that best fits the causal relationship of each event. After obtaining the graph structure and the conditional probability of each node, a probabilistic graph model is formed. When making predictions, the accurate value probability of the node to be queried can be calculated by inputting certain node observations. When calculating hypotheses, based on the causal structure that the model has learned, by blocking confounding factors, the conditional probability of the query node when the intervention node takes the intervention value is calculated, that is, the value probability of the query node after the intervention.

[0052] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for constructing a normalized graph model, characterized in that: The following steps are involved: S1. Data exploration: Use query statements and visualization tools to analyze business data and identify potential relationships between business data; S2. Node and edge relationship construction: Determine the node and edge types of the graph based on the results of the data exploration, thereby constructing a graph schema; S3. Graph data construction: Based on the graph schema, the data of corresponding nodes and edges are loaded from the distributed file system, and then the attributes of the nodes and edges are encoded and combined to form the graph required by the graph model; S4. Model selection: Select the corresponding paradigm for modeling according to the type of business problem; S5. Model parameter setting: set model parameters, including the number of hidden layers and loss function; S6. Model tuning: adjust model parameters to make the model optimal; The paradigm includes an attribute conduction paradigm, and the modeling method of the attribute conduction paradigm includes the following steps: S401. Aggregation: After receiving the information from neighboring nodes, the node aggregates the information with its own information according to different weights to form a more comprehensive information expression; S402. Non-linear transformation: performing non-linear transformation on the expression obtained by the aggregation as the expression of the node at this layer; S403. Propagation: Propagate the node expression obtained after the nonlinear transformation to neighboring nodes, and each node executes steps S401 to S403 in a loop; The paradigm includes a vectorized matching paradigm, and the modeling method of the vectorized matching paradigm includes: an aggregation process, in which each node aggregates neighbor information similarly to the attribute conduction paradigm, but the model learning target is no longer to fit the label of the sample, but to minimize the expression of related things and maximize the expression of unrelated things; after the training is completed, the model will obtain the calculation method of the expression of things in the corresponding context, and the similarity of the expression of things reflects the relevance of things; The paradigm includes a causal reasoning paradigm, and the modeling method of the causal reasoning paradigm includes: obtaining the conditional probability of each node by performing maximum likelihood estimation on the values ​​of each event in the training sample, and using breadth-first search to find the graph structure that best fits the causal relationship of each event, and after obtaining the graph structure and the conditional probability of each node, a probabilistic graph model is formed.

2. The method for constructing a normalized graph model according to claim 1, characterized in that: If no negative correlation is predefined, some pairs of things will be randomly generated as irrelevant relationships.

3. The method for constructing a normalized graph model according to claim 1, characterized in that: Retrieve the K most relevant targets of a thing through a hash inverted index.

4. The method for constructing a normalized graph model according to claim 1, characterized in that: When making predictions based on the probabilistic graph model, the value probability of the node to be queried can be calculated after inputting certain node observation values.

5. The method for constructing a normalized graph model according to claim 1, characterized in that: When the hypothesis is calculated according to the probabilistic graphical model, based on the causal structure learned by the model and by blocking confounding factors, the conditional probability of the query node when the intervention node takes the intervention value is calculated, that is, the probability of the query node taking the value after the intervention.

Images