Cognitive heuristic adaptive graph representation method

Abstract

The invention discloses a cognitive heuristic adaptive graph representation method, which is characterized by comprising the following steps of: on a network micro-scale level, taking non-Markov walk as a basis for converting node probability into dynamic association graph data, and realizing node long-range dependency association independent of links; in a network mesoscale representation level, network embedding based on unlinked sub-graph matching is applied to a graph-based pre-training model representation structure; in a network macroscopic modeling level, dynamic knowledge and situation data are fused for dynamic high-order mode distribution, an adaptive learning mechanism is constructed, and an interpretable graph neural network representation technology is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of graph representation, and more particularly, to a cognitively inspired adaptive graph representation method. Background Art

[0002] Graph representation technology mainly models complex data structures in the form of graphs to facilitate more efficient data processing, analysis, and mining. Graphs are mathematical structures composed of nodes and edges, where nodes represent entities and edges represent relationships between entities. Graph representation technology abstracts a large number of complex relationships in the real world into graph structures, allowing computers to understand and process this data in a more natural way.

[0003] However, traditional graph representation methods mostly rely on fixed structures and preset rules, which makes it difficult to flexibly adapt to rapidly changing environments and dynamic data, and they still have deficiencies in terms of interpretability in line with human cognitive laws. There are mainly two problems: 1) Many graph representation technologies often rely on complex algorithms and deep learning methods, which are often difficult to match with human intuitive understanding when processing and extracting potential features in graph structures. This makes the vectorized representation results generated by them difficult for users to understand, which in turn limits their effectiveness and reliability in practical applications.

[0004] 2) The relationship between nodes and edges in the graph and the high-dimensional data model behind it make it difficult for users to track and understand the reasoning process of the model when making cognitive decisions, which may cause cognitive load in the interpretation process. When it comes to complex decision-making situations, users may not be able to intuitively capture the key influencing factors from the graph and find it difficult to quickly and accurately obtain the required information.

[0005] Therefore, a method that conforms to the laws of human cognition is needed to optimize graph representation technology to better adapt to human cognitive characteristics and improve its interpretability. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a cognitively inspired adaptive graph representation method, which solves the problem that the relational structure of network data will produce statistical dependence between samples during the learning process of graph neural network representation, and is more universal and of practical value.

[0007] To achieve the above object, the present invention provides the following technical solution: a cognitively inspired adaptive graph representation method, comprising the following steps: At the micro-scale level of the network, non-Markov walks are used as the basis for transforming node probabilities into dynamic association graph data, realizing long-range node dependency associations that do not rely on links; At the network meso-scale representation level, network embedding based on unlinked subgraph matching is applied to the representation structure of graph-based pre-trained models; At the network macro-modeling level, facing the dynamic high-order pattern distribution, dynamic knowledge and situational data are integrated to build an adaptive learning mechanism and realize interpretable graph neural network representation technology.

[0008] The present invention is further configured as follows: the non-Markov random walk comprises the following steps: Develop cognitive flows on graphs to model the reasoning process and predict interpretable final results, using the concept of attention distribution to represent the probability distribution of attention on graph nodes; Train the discriminant model to predict the destination node by inputting the source node. Given the source node of the walk, the problem of predicting the destination node is further transformed into predicting the output attention distribution on the target node given the input attention distribution. We use the message passing algorithm in graph networks to derive a learnable transfer matrix T that evolves over time steps, and design an implicit direction function d whose dynamics are controlled by latent factors such as time, position, and history to drive the attention flow. The dynamics of the attention flow is given by Drive, where and Represents two consecutive attention distributions.

[0009] The present invention is further configured to include: using graph topic modeling technology to map and solve dynamic high-order pattern modeling.

[0010] The present invention is further configured as follows: the dynamic high-order mode modeling includes the following steps: Modeling nodes using dynamic probabilistic topic techniques; The sampled random walk sequence corresponds to the word, and the set of walks starting from each node corresponds to the document; And given the graph , the length is A possible set of non-Markov walk sequences , the expected number of high-order pattern topics , learning parameters: Node-topic distribution matrix , where the line Corresponds to Represents a node Belong to The distribution of the probability of a structural theme; Walk-topic distribution matrix , where the line is The distribution on express Belong to The probability of a structural theme.

[0011] The present invention is further configured to include an adaptive graph pre-training model.

[0012] The present invention is further configured as follows: the adaptive graph pre-training model includes: Dynamic knowledge graphs are used to conveniently integrate effective knowledge into adaptive graph representation learning to guide the model's intrinsic interpretability learning mechanism and design knowledge selection functions. , automatically select context-matching factual relations and interacting knowledge for node sequences from the knowledge graph; The sequence of unlinked subgraph nodes and their semantic association information from the non-Markov walk in the input graph is expressed as a latent space vector consisting of original features, absolute roles, relative positions and path-based relative distance vectors; Incorporating different knowledge context information in a dynamic way, reformulating knowledge and network feature entities at multiple levels into a unified feature vector; A multi-view graph neural network is designed, and node-level and module-level attention mechanisms are introduced to adaptively and dynamically screen high-order pattern themes to guide node aggregation, unifying node features with dynamic high-order pattern theme features.

[0013] The present invention is further configured to include a data representation level, a computable level, and a graph representation model level.

[0014] The present invention is further configured as follows: the data representation layer is used to construct a long-range correlation dependency representation method for graph data nodes, and to generate data samples with different dimensional links through a non-Markov walk sequence set.

[0015] The present invention is further configured as follows: the computable level is used to study dynamic high-order mode fusion calculations.

[0016] The present invention is further configured as follows: the graph representation model level is used to study the adaptive graph pre-training model.

[0017] In summary, the present invention has the following beneficial effects: the present invention predicts potential dynamic random walks from information flows driven by implicit cognitive flows, models unchained subgraphs with long-range correlations of nodes, and explores a basic general framework for feasible dynamic graph structure data representation learning.

[0018] The present invention brings new explorable paradigms and different perspectives to the general field of graph representation learning technology with human cognitive inspiration characteristics. The specific technical solution breaks through the traditional graph representation method, in which node neighborhoods are uniformly aggregated to describe one-hop relationships and over-rely on the graph link structure message passing mechanism. It studies the problem of adaptive representation learning under the framework of human cognitive inspiration and solves the problem that the relational structure of network data will generate statistical dependence between samples during the graph neural network representation learning process. It is more universal and practical than previous graph representation learning technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of micro, meso and macro; Figure 2 This is a schematic diagram of the data layer, computing layer, and model layer. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Reference Figure 1 As shown, to achieve the above purpose, the present invention provides the following technical solutions: a cognitively inspired adaptive graph representation method, at the network micro-scale level, uses non-Markov walks as the basis for node probability transformation of dynamic association graph data to achieve long-range node dependency association that does not rely on links; At the network meso-scale representation level, the network embedding based on unlinked subgraph matching is applied to the graph-based pre-trained model representation structure as the basic framework and core academic idea for solving the problem of the effectiveness and generalizability of graph neural network representation learning; At the network macro-modeling level, facing the dynamic high-order pattern distribution, dynamic knowledge and situational data are naturally integrated to build an adaptive learning mechanism and realize interpretable graph neural network representation technology.

[0022] The core technical solution revolves around three parts: non-Markov random walk, dynamic high-order pattern modeling, and adaptive graph representation model.

[0023] The approach adopted for the non-Markov random walk part breaks through the constraint assumption that the node random walk obeys the Markov characteristics in the traditional graph representation learning technology, and understands the unique problem-solving process of graph representation technology from the perspective of the long-range correlation dependency of nodes mapped in the cognitive process of human network representation learning.

[0024] 1) We develop a new attention mechanism on graphs, called cognitive flow, to model the reasoning process and predict interpretable final results. We use the concept of attention distribution to represent the probability distribution of attention on graph nodes.

[0025] 2) Considering only available source and destination nodes, a discriminant model is trained to predict the destination node by inputting the source node. Given the wandering source node, the problem of predicting the destination node is further transformed into predicting the output attention distribution on the target node given the input attention distribution.

[0026] 3) Use the message passing algorithm in the graph network to derive a learnable transfer matrix that evolves over time steps , the implicit direction function whose design dynamics are controlled by latent factors such as time, position and history , to drive the attention flow. The dynamics of the attention flow is given by Drive, where and Represents two consecutive attention distributions.

[0027] 4) Attention is transferred from the source node to the destination node, which means simulating the implicit cognitive flow in the human cognitive process through the graph. Nodes are extracted from the distribution driven by the function d to construct a random walk sequence This essentially means a strong inductive bias towards graphs, relations (sequences of consecutive nodes) and non-relations (latent directional functions that depend on time, position and history), enabling non-Markov random walks.

[0028] In the specific scheme, cognitive flow is regarded as a graph-level computation directly operated in probability space rather than discrete sample space. The research scheme considers both the inherent association relationship in the graph structure and the dynamic association characteristics based on spatiotemporal semantic context, and realizes the prediction of potential cognitive flow based on the multidimensional data association of abstract representation, laying the foundation for cognitive-inspired graph representation learning technology and its interpretability that do not rely on Markov characteristics.

[0029] Secondly, in the research plan on dynamic high-order pattern modeling, the influence characteristics of modular dynamic high-order pattern distribution on micro-scale cognitive association structure are seamlessly transferred to the topic modeling technology in natural language processing, and the dynamic high-order pattern modeling is solved by mapping with graph topic modeling technology.

[0030] 1) Documents and topics in topic models in natural language processing are defined by the distribution of topics and words. This probabilistic nature directly corresponds to the distribution of structural patterns required to describe complex high-order neighborhoods in the network. Therefore, using dynamic probabilistic topic technology to model nodes can more accurately capture the differences in local high-order pattern distributions and better help graph neural networks capture node features.

[0031] 2) Considering that the graph is related and the documents are independent samples, appropriate adjustments are made to make the structural topics technically reasonable. This is similar to topic modeling in natural language processing, where the sequence of sampled random walks corresponds to words and the set of walks starting from each node corresponds to documents.

[0032] 3) Given a graph , the length is A possible set of non-Markov walk sequences , the expected number of high-order pattern topics , learning parameters: •Node-topic distribution matrix , where the line Corresponds to Represents a node Belong to The probability distribution of a structural topic.

[0033] • Walk-topic distribution matrix , where the line is The distribution on express Belong to The probability of a structural theme.

[0034] 4) Based on the above steps, nodes are given a probabilistic description of their local high-order structural patterns. The learned walk co-occurrence matrix not only indicates the co-occurrence of walks, but also indicates the potential high-order pattern themes, that is, the high-order pattern themes are composed of the distribution of high-order patterns that indicate node characteristics, realizing dynamic high-order pattern modeling.

[0035] This technical solution essentially studies the problem from the perspective of how cognitive representation structures interact with the problem-solving process. The research solution considers both the important information of the system structure provided by the individual connections between nodes and the combination and overlap of node links to form a high-order structure representing the environment. It provides a basis for the adaptive adjustment ability of the graph representation model under different data distributions and the adaptive migration of models involved in calculations in different tasks.

[0036] Third, in the adaptive graph pre-training model, the transformation and fusion between human internal cognitive association network modules, that is, switching from a space with obvious solutions to another space with solutions, is the key psychological mechanism for the emergence of creative insights. Based on human cognitive psychological mechanisms, the research program explores the projection process guided by human knowledge that is naturally reflected in the graph representation process from the perspective of dynamic fusion of data and knowledge.

[0037] 1) Dynamic knowledge graphs are used to conveniently integrate effective knowledge into adaptive graph representation learning to guide the model's intrinsic interpretability learning mechanism. Design knowledge selection function , from the knowledge graph for the node sequence Automatically select knowledge of factual relations and interactions that match the context.

[0038] 2) The sequence of unlinked subgraph nodes of non-Markov walks in the input graph and their semantic association information are expressed as a latent space vector (Embedding) composed of original features, absolute roles, relative positions and path-based relative distance vectors, reflecting the ecological status of each node in the entire graph data association network.

[0039] 3) Incorporating different knowledge context information in a dynamic manner, the knowledge and network feature entities at multiple levels are reformulated as a unified feature vector, which essentially reflects the time series transient nature of the real-world knowledge edge.

[0040] 4) Design a multi-view graph neural network, introduce node-level and module-level attention mechanisms, adaptively and dynamically screen high-order pattern themes to guide node aggregation, unify node features with dynamic high-order pattern theme features, and realize an adaptive graph neural network pre-training model.

[0041] Based on the embedding vector of the input unlinked subgraph sampling representation, this technical solution guides the adaptive conversion between different high-order modules in the context space based on dynamic knowledge selection, and represents the dynamic overall structure of the problem-solving context space in a unified framework. It integrates data and knowledge, explores the cognitive motivation and context factors that generate links in semantic context associations, and characterizes the intrinsic learning mechanism of adaptive graph neural network representation under the condition of context changes, solving the problem of lack of interpretability and versatility in existing graph representation learning methods.

[0042] In summary, the present invention predicts potential dynamic random walks from the information flow driven by implicit cognitive flow, models unchained subgraphs with long-range correlations among nodes, and explores a basic general framework for feasible dynamic graph structure data representation learning.

[0043] The present invention brings new explorable paradigms and different perspectives to the general field of graph representation learning technology with human cognitive inspiration characteristics. The specific technical solution breaks through the traditional graph representation method, in which node neighborhoods are uniformly aggregated to describe one-hop relationships and over-rely on the graph link structure message passing mechanism. It studies the problem of adaptive representation learning under the framework of human cognitive inspiration and solves the problem that the relational structure of network data will generate statistical dependence between samples during the graph neural network representation learning process. It is more universal and practical than previous graph representation learning technologies.

[0044] Reference Figure 2 As shown, it is expressed at the data representation level, computable level and graph representation model level respectively: First, at the data representation level of graph representation, we model the long-range correlation and dependency representation method of graph data nodes, based on the characteristics that almost all node sequences in the association sequence of the human brain representation network involve long-range correlation and dependency. Through the set of non-Markov walk sequences, we study how node data can generate large samples of links of different dimensions based on a certain association metric function, naturally model the dynamic long-range dependency dimensional representation, and lay the foundation for cognitively inspired adaptive graph representation learning technology.

[0045] Secondly, at the computable level, dynamic high-order pattern fusion computing is studied. In view of the influence of modular dynamic high-order pattern distribution on micro-scale cognitive association structure in the process of human brain learning network representation, how to capture dynamic high-order pattern distribution is solved, and the general graph representation learning method that models the interaction between the multi-module dynamic representation structure of graph data and the task solving process is possible. At the same time, it provides a research basis for the adaptive migration of models participating in calculations in different tasks based on graph-based pre-training methods.

[0046] Finally, at the level of graph representation models, we study adaptive graph pre-training models. Aiming at the characteristics of adaptively capturing long-term and short-term structural changes in cognitive association structures in human cognition, we solve the problem of how to model the adaptive transformation between different modules in the context space in graph neural networks, represent the dynamic overall structure of the problem-solving context space in a unified framework, deeply capture the intrinsic learning mechanism of graph neural network representation, and provide a guarantee for the model to break through the bottleneck of unexplainable traditional graph neural network representation.

[0047] In summary, based on the fusion of knowledge and data, the intrinsic learning mechanism of cognitively inspired adaptive graph neural network representation is characterized in a unified graph representation technology framework to solve the problems of interpretability and versatility.

[0048] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A cognitively inspired adaptive graph representation method, characterized in that: The following steps are included: At the micro-scale level of the network, non-Markov walks are used as the basis for transforming node probabilities into dynamic association graph data, realizing long-range node dependency associations that do not rely on links; At the network meso-scale representation level, network embedding based on unlinked subgraph matching is applied to the representation structure of graph-based pre-trained models; At the network macro-modeling level, facing the dynamic high-order pattern distribution, dynamic knowledge and situational data are integrated to build an adaptive learning mechanism and realize interpretable graph neural network representation technology.

2. The cognitively inspired adaptive graph representation method according to claim 1, characterized in that: The node probability conversion dynamic association graph data includes the following steps: Develop cognitive flows on graphs to model the reasoning process and predict interpretable final results, using the concept of attention distribution to represent the probability distribution of attention on graph nodes; Train the discriminant model to predict the destination node by inputting the source node. Given the source node of the walk, the problem of predicting the destination node is further transformed into predicting the output attention distribution on the target node given the input attention distribution. We use the message passing algorithm in graph networks to derive a learnable transfer matrix T that evolves over time steps, and design an implicit direction function d whose dynamics are controlled by time, position, and history latent factors to drive the attention flow. The dynamics of the attention flow is given by Drive, where and Represents two consecutive attention distributions.

3. A cognitively inspired adaptive graph representation method according to claim 2, characterized in that: It also includes the use of graph topic modeling technology to map and solve dynamic high-order pattern modeling.

4. The cognitively inspired adaptive graph representation method according to claim 3, characterized in that: The dynamic high-order mode modeling includes the following steps: Modeling nodes using dynamic probabilistic topic techniques; The sampled random walk sequence corresponds to the word, and the set of walks starting from each node corresponds to the document; And given the graph , the length is A possible set of non-Markov walk sequences , the expected number of high-order pattern topics , learning parameters: Node-topic distribution matrix , where the line Corresponds to Represents a node Belong to The distribution of the probability of a structural theme; Walk-topic distribution matrix , where the line is The distribution on express Belong to The probability of a structural theme.

5. The cognitively inspired adaptive graph representation method according to claim 1, characterized in that: Also included are adaptive graph pre-trained models.

6. A cognitively inspired adaptive graph representation method according to claim 5, characterized in that: The adaptive graph pre-training model includes: Dynamic knowledge graphs are used to conveniently integrate effective knowledge into adaptive graph representation learning to guide the model's intrinsic interpretability learning mechanism and design knowledge selection functions. , automatically select context-matching factual relations and interacting knowledge for node sequences from the knowledge graph; The sequence of unlinked subgraph nodes and their semantic association information from the non-Markov walk in the input graph is expressed as a latent space vector consisting of original features, absolute roles, relative positions and path-based relative distance vectors; Incorporating different knowledge context information in a dynamic way, reformulating knowledge and network feature entities at multiple levels into a unified feature vector; A multi-view graph neural network is designed, and node-level and module-level attention mechanisms are introduced to adaptively and dynamically screen high-order pattern themes to guide node aggregation, unifying node features with dynamic high-order pattern theme features.

7. A cognitively inspired adaptive graph representation method according to any one of claims 1 to 6, characterized in that: It also includes the data representation level, the computable level, and the graph representation model level.

8. A cognitively inspired adaptive graph representation method according to claim 7, characterized in that: The data representation layer is used to construct a long-range correlation dependency representation method for graph data nodes, and to generate data samples with different dimensional links through a non-Markov walk sequence set.

9. The cognitively inspired adaptive graph representation method according to claim 7, characterized in that: The computable level is used to study dynamic high-order mode fusion calculations.

10. The cognitively inspired adaptive graph representation method according to claim 7, characterized in that: The graph representation model level is used to study the adaptive graph pre-training model.

Images