Time sequence perception knowledge graph reasoning method based on attention network

Abstract

The invention discloses a time sequence perception knowledge graph reasoning method based on an attention network, and belongs to the technical field of artificial intelligence knowledge reasoning, and the method comprises the steps: inputting a historical event time sequence perception knowledge graph and an attribute time sequence perception knowledge graph; establishing a relation between a target entity and a high-order neighbor entity in the historical event time sequence perception knowledge graph by adopting a graph convolutional network; an attention mechanism is introduced, and influences of different relation types on the target entity are distinguished; adopting a layer of graph convolutional network without self-circulation to obtain embedding of each entity about attributes in the attribute time sequence perception knowledge graph; a time gate mechanism is introduced, and historical information embedded in the entity time sequence is screened; updating the relation embedding matrix under different timestamps by adopting a gating circulation unit; and entity prediction and relation prediction are realized. By the adoption of the method, reasoning precision and efficiency are improved, the processing capacity of the complex time sequence knowledge graph is improved, and accurate prediction of future events and the relation of the future events is achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence knowledge reasoning technology, and in particular to a time-series-aware knowledge graph reasoning method based on an attention network. Background Art

[0002] Knowledge graphs are a series of various graphs that show the process and structural relationship of knowledge development. At present, the traditional method of using time-series knowledge graphs to reason about future events mainly relies on historical data modeling and rule reasoning, but these methods have many shortcomings. First, traditional reasoning models are usually based on static knowledge graphs and cannot effectively capture the dynamic changes of events. Secondly, the existing models lack consideration of the differences between different types of relationships and fail to accurately reflect the impact of different relationships on the reasoning results. In addition, the existing technology has limited ability to handle concurrent events, and it is difficult to model the interdependence between multiple events occurring at the same time point, resulting in low reasoning accuracy and insufficient generalization ability. Summary of the invention

[0003] The purpose of this invention is to provide a time-series-aware knowledge graph reasoning method based on attention network, to improve the reasoning accuracy and efficiency, to improve the processing ability of complex time-series knowledge graphs, and to achieve accurate prediction of future events and their relationships.

[0004] To achieve the above objectives, the present invention provides a time-series-aware knowledge graph reasoning method based on an attention network, the steps comprising:

[0005] S1. Input historical event time-series-aware knowledge graph and attribute time-series-aware knowledge graph;

[0006] S2. Use graph convolutional networks to perform multi-layer propagation on each entity in the historical event time-series-aware knowledge graph, and establish connections between target entities and high-order neighbor entities by aggregating neighbor entities layer by layer to enrich the embedding representation of entities.

[0007] S3, introduce the attention mechanism to distinguish the impact of different relationship types on the target entity;

[0008] S4. Use a graph convolutional network without self-loop to obtain the embedding of each entity in the attribute temporal-aware knowledge graph to enrich the embedding representation of the entity.

[0009] S5, introduce the time gate mechanism to filter the historical information embedded in the entity time series in step S2 and step S4 through adaptive gating;

[0010] S6. Use gated recurrent units to process relations at different timestamps, capture dynamic changes in time series, and adaptively update the relation embedding matrix through the memory and forgetting mechanism of historical information;

[0011] S7. Based on the entity and relationship embedding representations obtained after processing in step S5 and step S6, entity prediction and relationship prediction are implemented.

[0012] Preferably, the knowledge graph in step S1 uses quadruple as the basic unit, including subject, relationship, object and time, the historical event time series-aware knowledge graph represents entities and relationships, and the attribute time series-aware knowledge graph contains static characteristics of entities, including categories and attribute values.

[0013] Preferably, in step S2, a connection is established between the target entity and the high-order neighbor entity, and the formula is:

[0014]

[0015] In the formula, represents the embedding of the head entity at layer l at time T, represents the embedding of the tail entity at layer l at time T, e r,T represents the embedding of the relation at layer l at time T, represents the translation of the head entity h through the relation r to the tail entity, h a represents the head entity in the attribute temporal-aware knowledge graph, t a represents the head entity and tail entity in the attribute temporal-aware knowledge graph, ε a A collection of entities representing an attributed time-aware knowledge graph.

[0016] Preferably, step S3 includes: using To measure the impact of different relationship types on translation, we use dot product attention to learn the weights of different relationships:

[0017]

[0018] In the formula, Represents the influence of different relationship types when the head entity h is translated to the tail entity through the relationship r.

[0019] Preferably, in step S4, a layer of graph convolutional network without self-loop is used to obtain the embedding of attributes of each entity in the attribute time-series-aware knowledge graph, and the formula is:

[0020]

[0021] In the formula, They represent the embedding of the head entity and the tail entity at time T, respectively. a represents the head entity in the attribute temporal-aware knowledge graph, t a represents the head entity and tail entity in the attribute temporal-aware knowledge graph, ε a A collection of entities representing an attributed time-aware knowledge graph.

[0022] Preferably, in step S5, a time gate mechanism is introduced to filter the historical information embedded in the entity time series in steps S2 and S4 through adaptive gating, including:

[0023] First, consider the historical events carried by the knowledge graph series at different times, and use the time gate to optimize the entity embedding. The formula is:

[0024]

[0025] U T =σ(WH T-1 + b);

[0026] In the formula, H T represents the entity embedding matrix at time T, represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T-1; represents the entity embedding matrix obtained from the attribute time series-aware knowledge graph at time T, U T represents the time gate factor;

[0027] Then consider the impact of changes in its attributes over time on the reasoning effect, the formula is:

[0028]

[0029] U T =σ(WH T-1 + b);

[0030] In the formula, represents the entity embedding matrix obtained by modeling the attribute time-series-aware knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix at time T-1.

[0031] Preferably, in step S6, a gated recurrent unit is used to process the relationships at different timestamps, including an average pooling operation and a GRU component update, and the formula is:

[0032]

[0033] In the formula, represents the embedding of relation r in relation matrix R at time T, Represents the concatenation operation of the vector. If no corresponding event occurs at time T, then By all Composition, RT Represents the relation embedding matrix.

[0034] Preferably, the entity prediction in step S7 represents the entities that may participate in an event at a predicted future time point, and the relationship prediction is used to infer the possible relationships between entities;

[0035] The probability vector for all entities is:

[0036] P(h / t|t / h,r,H T ,R T )=σ(H T ConvTransE(e r,T / e h,T ,e r,T ));

[0037] All relationship probability vectors are:

[0038] P(r|h,t,H T ,R T )=σ(R T ConvTransE(e h,T ,e t,T ));

[0039] In the formula, σ represents the activation function, e h,T and e t,T Respectively represent H T The embedding of h and t in e r,T Represents R T In the embedding of r, H represents the entity embedding matrix, and R represents the relationship embedding matrix. The entity and relationship prediction tasks are regarded as multi-label learning. and Respectively represent the label vectors of the two tasks at time T. For events and relationships that actually occurred, and The value of is 1, otherwise it is 0;

[0040] The entity prediction loss function is:

[0041]

[0042] The relationship prediction loss function is

[0043]

[0044] The total objective loss function is:

[0045] L=L e +L r ;

[0046] Where, L erepresents the entity prediction loss function, L r represents the relationship prediction loss function, L represents the total loss function, and are respectively expressed in vector and The i-th element in P i (h / t|t / h,r,H T ,R T ) and P i (r|h,t,H T ,R T ) represent the probabilities of entity i and relation i respectively, and V represents the total number of entities.

[0047] Therefore, the present invention adopts the above-mentioned time-series-aware knowledge graph reasoning method based on attention network, which has the following beneficial effects:

[0048] (1) The introduction of graph convolutional networks and attention mechanisms fully utilizes high-order neighbor entities and multiple relationship type information, significantly improving the accuracy and efficiency of reasoning;

[0049] (2) The introduction of the time gate mechanism can effectively cope with changes in time series, prevent the convergence problem of entities embedded in different timestamps, and improve the stability of time series reasoning. At the same time, by introducing the time gate mechanism, the problem of difficulty in modeling the interdependence between multiple events occurring at the same time point is solved;

[0050] (3) By using graph convolutional networks without self-loops and gated recurrent unit components, the expressive power of entity and relation embeddings is enhanced, and the generalization ability of the model is improved;

[0051] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a method block diagram of an embodiment of the present invention;

[0053] Figure 2 It is an overall framework diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0055] Example

[0056] Reference Figure 1-Figure 2 The present invention provides a time-series-aware knowledge graph reasoning method based on an attention network, the steps comprising:

[0057] S1. Input historical event time-series-aware knowledge graph and attribute time-series-aware knowledge graph.

[0058] The knowledge graph uses quadruple as the basic unit, including subject, relation, object and time. The historical event time-aware knowledge graph represents entities and relations, and the attribute time-aware knowledge graph contains the static characteristics of entities, including categories and attribute values. By combining these two different information sources, the model can more comprehensively understand and model entities and the dynamic relationships between them. The data sources of the two knowledge graphs are both from authoritative public datasets in related fields, and they are time-based temporal knowledge graphs. These temporal knowledge graphs include the Integrated Crisis Early Warning System (ICEWS18, ICEWS14), and the Global Database of Events, Language and Tone (GDELT). These datasets are widely used in time-series event prediction and dynamic relationship modeling, covering a large number of events and corresponding time information worldwide.

[0059] S2. Use graph convolutional networks to perform multi-layer propagation on each entity in the historical event time-series perception knowledge graph. By aggregating neighbor entities layer by layer, the target entity’s field of view is gradually expanded, and the connection between the target entity and the high-order neighbor entities is established to enrich the entity’s embedded representation. By establishing the connection between the target entity and the high-order neighbor entities, not only the local representation ability of the entity is enhanced, but also the model is helped to capture the potential structural relationships and complex dependencies in the knowledge graph.

[0060] In this embodiment, the relationship between the target entity and the high-order neighbor entity is established, and the formula is:

[0061]

[0062] In the formula, represents the embedding of the head entity at layer l at time T, represents the embedding of the tail entity at layer l at time T, e r,T represents the embedding of the relation at layer l at time T, represents the translation of the head entity h to the tail entity through the relationship r, α represents the influence of different relationship types on the translation of the head entity h to the tail entity through the relationship r, represents the weight of the first layer in the neural network, and ε represents the entity set in the knowledge graph.

[0063] S3. Introduce the attention mechanism to distinguish the impact of different types of relationships on the target entity. The attention mechanism can dynamically adjust the weight of each relationship on the target entity to ensure that the influence of key relationships is highlighted and unimportant relationships are weakened during the information dissemination process. This step ensures that the model better captures the complex and diverse relationships between entities to improve the accuracy of reasoning.

[0064] In this embodiment, using To measure the impact of different relationship types on translation, we use dot product attention to learn the weights of different relationships:

[0065]

[0066] In the formula, Represents the influence of different relationship types when the head entity h is translated to the tail entity through the relationship r.

[0067] S4. A graph convolutional network without self-loop is used to obtain the embedding of each entity in the attribute-time-aware knowledge graph to enrich the embedding representation of the entity. By obtaining the attribute information of each entity, such as category, state, etc., a more semantically layered embedding is generated to enrich the embedding representation of the entity, avoiding excessive information propagation introduced by the self-loop mechanism, and ensuring that the model can effectively capture attribute features without losing details.

[0068] In this embodiment, a layer of graph convolutional network without self-loop is used to obtain the embedding of attributes of each entity in the attribute time-series-aware knowledge graph, and the formula is:

[0069]

[0070] In the formula, They represent the embedding of the head entity and the tail entity at time T, respectively. a represents the head entity in the attribute temporal-aware knowledge graph, t a represents the head entity and tail entity in the attribute temporal-aware knowledge graph, ε a A collection of entities representing an attributed time-aware knowledge graph.

[0071] S5. Introduce the TimeGate mechanism to filter the historical information embedded in the entity time series in steps S2 and S4 through adaptive gating, so as to avoid the problem that the time-aware attribute time-series-aware knowledge graph of historical events tends to be stable in a short period of time, resulting in the entity embedding being too similar at different timestamps. At the same time, this mechanism can also effectively alleviate the gradient vanishing problem caused by too many layers of graph neural networks and improve the stability of the model.

[0072] In this embodiment, a time gate mechanism is introduced to filter the historical information embedded in the entity time series in step S2 and step S4 through adaptive gating, including:

[0073] First, consider the historical events carried by the knowledge graph series at different times, and use the time gate to optimize the entity embedding. The formula is:

[0074]

[0075] U T =σ(WH T-1 + b);

[0076] In the formula, H T represents the entity embedding matrix at time T, represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T-1; represents the entity embedding matrix obtained from the attribute time series-aware knowledge graph at time T, U T Represents the time gate factor. H T Depend on H T-1 and After the time gate, the decision is made by U T control.

[0077] Then consider the impact of changes in its attributes over time on the reasoning effect, the formula is:

[0078]

[0079] U T =σ(WH T-1 + b);

[0080] In the formula, represents the entity embedding matrix obtained by modeling the attribute time-series-aware knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix at time T-1. Depend on and H T-1 After the time gate, the decision is made by U T control.

[0081] S6. Use gated recurrent units to process relations at different timestamps to capture dynamic changes in time series, and adaptively update the relation embedding matrix through the memory and forgetting mechanism of historical information to ensure that the representation of the relation at each timestamp can reflect the actual time dependency and relation difference.

[0082] In this embodiment, a gated recurrent unit is used to process the relationships at different timestamps, including average pooling operations and GRU component updates, and the formula is:

[0083]

[0084] In the formula, represents the embedding of relation r in relation matrix R at time T, Represents the concatenation operation of the vector. If no corresponding event occurs at time T, then By all Composition, R T Represents the relation embedding matrix.

[0085] S7. Based on the entity and relationship embedding representations obtained after processing in step S5 and step S6, entity prediction and relationship prediction are implemented.

[0086] Entity prediction refers to the entities that may be involved in an event at a predicted future time point, and relationship prediction is used to infer the possible relationships between entities;

[0087] The probability vector for all entities is:

[0088] P(h / t|t / h,r,H T ,R T )=σ(H T ConvTransE(e r,T / e h,T ,e r,T ));

[0089] All relationship probability vectors are:

[0090] P(r|h,t,H T ,R T )=σ(R T ConvTransE(e h,T ,e t,T ));

[0091] In the formula, σ represents the activation function, eh,T and e t,T Respectively represent H T The embedding of h and t in e r,T Represents R T In the embedding of r, H represents the entity embedding matrix, and R represents the relationship embedding matrix. The entity and relationship prediction tasks are regarded as multi-label learning. and Respectively represent the label vectors of the two tasks at time T. For events and relationships that actually occurred, and The value of is 1, otherwise it is 0;

[0092] The entity prediction loss function is:

[0093]

[0094] The relationship prediction loss function is

[0095]

[0096] The total objective loss function is:

[0097] L=L e +L r ;

[0098] Where, L e represents the entity prediction loss function, L r represents the relationship prediction loss function, L represents the total loss function, and are respectively expressed in vector and The i-th element in P i (h / t|t / h,r,H T ,R T ) and P i (r|h,t,H T ,R T ) represent the probabilities of entity i and relation i respectively, and V represents the total number of entities.

[0099] In order to verify the effectiveness of the method of the present invention, the R-GCN model and the RE-GCN model are selected as comparison models. The three datasets of ICEWS18, ICEWS18 and GDELT are tested and compared using the mean reciprocal (MRR) indicator. The entity prediction results and the relationship prediction results are shown in Table 1 and Table 2, respectively.

[0100] Table 1 Entity prediction results

[0101]

[0102] Table 2 Relationship prediction results

[0103]

[0104] It can be seen from the experimental results that the average inverse of the indicators of the method of the present invention are higher than those of the R-GCN model and the RE-GCN model. It can be further concluded that differentially capturing the structural information and high-order information of the knowledge graph through the relational attention network, dynamically updating attributes and recursively capturing historical events are helpful to improve the prediction effect.

[0105] Therefore, the present invention uses the above-mentioned time-aware knowledge graph reasoning method based on attention network, which can differentially capture the structure and high-order connections of the knowledge graph through the attention mechanism, recursively propagate embeddings from the entity's neighbors to optimize the entity, and dynamically model the historical knowledge graph series through time gates and gated recurrent units to capture all historical events and update relationships. In addition, the model enriches the entity representation by dynamically updating entity attributes, thereby realizing simultaneous reasoning of entities and their relationships.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A time-series-aware knowledge graph reasoning method based on attention network, characterized in that the steps include: S1. Input historical event time-series-aware knowledge graph and attribute time-series-aware knowledge graph; S2. Use graph convolutional networks to perform multi-layer propagation on each entity in the historical event time-series-aware knowledge graph, and establish connections between target entities and high-order neighbor entities by aggregating neighbor entities layer by layer to enrich the embedding representation of entities. S3, introduce the attention mechanism to distinguish the impact of different relationship types on the target entity; S4. Use a graph convolutional network without self-loop to obtain the embedding of each entity in the attribute temporal-aware knowledge graph to enrich the embedding representation of the entity. S5, introduce the time gate mechanism to filter the historical information embedded in the entity time series in step S2 and step S4 through adaptive gating; S6. Use gated recurrent units to process relations at different timestamps, capture dynamic changes in time series, and adaptively update the relation embedding matrix through the memory and forgetting mechanism of historical information; S7. Based on the entity and relationship embedding representations obtained after processing in step S5 and step S6, entity prediction and relationship prediction are implemented.

2. According to claim 1, a time-series-aware knowledge graph reasoning method based on an attention network is characterized in that: In step S1, the knowledge graph uses quadruple as the basic unit, including subject, relationship, object and time. The historical event time-series-aware knowledge graph represents entities and relationships, and the attribute time-series-aware knowledge graph contains the static characteristics of entities, including categories and attribute values.

3. According to claim 2, a time-series-aware knowledge graph reasoning method based on attention network is characterized in that: In step S2, the connection between the target entity and the high-order neighbor entity is established, and the formula is: In the formula, represents the embedding of the head entity at layer l at time T, represents the embedding of the tail entity at layer l at time T, e r,T represents the embedding of the relation at layer l at time T, represents the translation of the head entity h to the tail entity through the relationship r, α represents the influence of different relationship types on the translation of the head entity h to the tail entity through the relationship r, represents the weight of the first layer in the neural network, and ε represents the entity set in the knowledge graph.

4. According to claim 3, a time-series-aware knowledge graph reasoning method based on an attention network is characterized in that: Step S3 includes: using To measure the impact of different relationship types on translation, we use dot product attention to learn the weights of different relationships: In the formula, Represents the influence of different relationship types when the head entity h is translated to the tail entity through the relationship r.

5. According to claim 4, a time-series-aware knowledge graph reasoning method based on attention network is characterized in that: In step S4, a graph convolutional network without self-loop is used to obtain the embedding of attributes of each entity in the attribute time-series-aware knowledge graph. The formula is: In the formula, They represent the embedding of the head entity and the tail entity at time T, respectively. a represents the head entity in the attribute temporal-aware knowledge graph, t a represents the head entity and tail entity in the attribute temporal-aware knowledge graph, ε a A collection of entities representing an attributed time-aware knowledge graph.

6. According to claim 5, a time-series-aware knowledge graph reasoning method based on attention network is characterized in that: In step S5, a time gate mechanism is introduced to filter the historical information embedded in the entity time series in steps S2 and S4 through adaptive gating, including: First, consider the historical events carried by the knowledge graph series at different times, and use the time gate to optimize the entity embedding. The formula is: U T =σ(WH T-1 +b); In the formula, H T represents the entity embedding matrix at time T, represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix obtained by modeling the historical event time series perception knowledge graph through the graph neural network at time T-1; represents the entity embedding matrix obtained from the attribute time series-aware knowledge graph at time T, U T represents the time gate factor; Then consider the impact of changes in its attributes over time on the reasoning effect, the formula is: U T =σ(WH T-1 +b); In the formula, represents the entity embedding matrix obtained by modeling the attribute time-aware knowledge graph through the graph neural network at time T, H T-1 Represents the entity embedding matrix at time T-1.

7. According to claim 6, a time-series-aware knowledge graph reasoning method based on attention network is characterized in that: In step S6, a gated recurrent unit is used to process the relationships at different timestamps, including average pooling operations and GRU component updates. The formula is: In the formula, represents the embedding of relation r in relation matrix R at time T, Represents the concatenation operation of the vector. If no corresponding event occurs at time T, then By all Composition, R T Represents the relation embedding matrix.

8. According to claim 7, a time-series-aware knowledge graph reasoning method based on attention network is characterized in that: In step S7, entity prediction represents entities that may be involved in an event at a predicted future time point, and relationship prediction is used to infer possible relationships between entities; The probability vector for all entities is: P(h / t|t / h,r,H T ,R T )=σ(H T ConvTransE(e r,T / e h,T ,e r,T )); All relationship probability vectors are: P(r|h,t,H T ,R T )=σ(R T ConvTransE(e h,T ,e t,T )); In the formula, σ represents the activation function, e h,T and e t,T Respectively represent H T The embedding of h and t in e r,T Represents R T In the embedding of r, H represents the entity embedding matrix, and R represents the relationship embedding matrix. The entity and relationship prediction tasks are regarded as multi-label learning. and Respectively represent the label vectors of the two tasks at time T. For events and relationships that actually occurred, and The value of is 1, otherwise it is 0; The entity prediction loss function is: The relationship prediction loss function is The total objective loss function is: L=L e +L r ; Where, L e represents the entity prediction loss function, L r represents the relationship prediction loss function, L represents the total loss function, and are respectively expressed in vector and The i-th element in P i (h / t|t / h,r,H T ,R T ) and P i (r|h,t,H T ,R T ) represent the probabilities of entity i and relation i respectively, and V represents the total number of entities.

Images