A time sequence knowledge graph completion method based on fuzzy geometric representation and time perception mechanism

Abstract

The application discloses a time sequence knowledge graph completion method based on fuzzy geometric representation and time perception mechanism, which aims to predict the four-tuple of missing entities or relations by using known four-tuples. The method firstly constructs a time sequence knowledge graph dataset containing time information, and carries out vectorization embedding representation on entities and relations; subsequently, a time coding module is designed to encode the time information to obtain a time embedding representation vector; on this basis, the entity, the relation and the time information are mapped into a fuzzy geometric representation space, and the entity semantic space is modeled by a box structure; then, a scoring function is constructed by calculating the intersection volume between the reasoning entity box and the candidate entity box; finally, the prediction of missing facts in the time sequence knowledge graph is realized through model training. The application can effectively improve the entity relation representation capability in the time sequence knowledge graph, and improve the accuracy of missing fact prediction and the stability of model training.

Description

Technical Field

[0001] This invention relates to the field of temporal knowledge graph completion technology, specifically a temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism. Background Technology

[0002] Temporal knowledge graphs introduce temporal information into traditional knowledge graphs to describe the dynamic characteristics of relationships between entities over time. Facts are typically represented as quadruples (head entity, relation, tail entity, time) containing temporal information. Due to incomplete data collection and frequent information updates in the real world, temporal knowledge graphs often contain a large number of missing facts. Therefore, temporal knowledge graph completion techniques are needed to predict missing entities, relations, or temporal information, thereby improving the knowledge graph structure. Temporal knowledge graph completion has significant application value in recommendation systems, intelligent question answering, and event prediction.

[0003] Existing temporal knowledge graph completion methods typically use embedding models to represent entities, relations, and temporal information as vectors, and then perform reasoning within the embedding space. However, existing methods often neglect the complex relationship between relations and temporal information during the modeling process, lack sufficient representation capabilities for temporal information, and fail to fully utilize the changing patterns of facts over time, thus affecting the model's predictive performance for missing facts in the temporal knowledge graph. Therefore, it is necessary to propose a novel temporal knowledge graph completion method to improve the modeling ability for temporal relations and enhance the accuracy of knowledge graph completion. Summary of the Invention

[0004] To address the aforementioned problems of existing temporal knowledge graph completion methods, this invention provides a temporal knowledge graph completion method based on fuzzy geometric box representation and a time-aware mechanism. This method constructs a time-aware encoding mechanism and a fuzzy geometric representation model to uniformly map entity, relation, and temporal information into a high-dimensional vector representation space, thereby improving the entity and relation representation capabilities in the temporal knowledge graph and enabling effective reasoning and prediction of missing facts.

[0005] The technical solution of this invention is: a method for completing a temporal knowledge graph based on fuzzy geometric representation and time-aware mechanism, comprising the following steps:

[0006] 1. First, the entities in the time-series knowledge graph are initialized to obtain entity embedding representation vectors; relations are defined as the position parameters of entity boxes in the vector space, which includes two specific processing steps: entity representation and relation constraint process.

[0007] (1): In the process of entity representation, use and Two vectors represent the basis vector and offset vector of the entity, respectively. By applying normalization constraints, the model's stability and geometric interpretability are enhanced, thereby limiting the numerical range and suppressing gradient explosion. Then, a bump mechanism is used to construct the initial embedding representation vectors of the inference entity and candidate entities under the influence of the current facts. The initial representation of the inference entity is jointly determined by its own basis vector and the offset vector of the candidate entity, and the initial representation of the candidate entity is jointly determined by its basis vector and the offset vector of the inference entity.

[0008] (2): During the definition of a relation, the model learns two parameters for each relation: a base vector and a width vector. The base vector represents the center position of the entity in the vector space; the width vector describes the expansion range of the box in each dimension.

[0009] Secondly, in order to enhance the dynamic semantics of entities at different times, a time-aware mechanism is proposed. The discrete timestamps of individual tuples in the temporal knowledge graph are mapped into a continuous time vector through Fourier basis functions, and then fused with the entity embedding representation vector to obtain a time-aware entity embedding representation vector, so that the entity representation changes dynamically with time.

[0010] Then, a fuzzy box representation mechanism is introduced to construct geometric regions for modeling, making the interaction of entities, relations, and time in the same space more natural and continuous. The time-aware entity embedding representation vectors are combined with relation parameters to construct inference entity boxes and prediction entity boxes:

[0011]

[0012] in Let the center position of the solid box be the vector representation space. and represent the upper and lower bounds of the entity box in the vector representation space, respectively. This box characterizes the potential region of candidate entity boxes that can be derived from the inference entity box under the action of relation r.

[0013] Finally, the intersection volume of the inferred entity box and the candidate entity box in the vector representation space is calculated using a scoring function based on the intersection volume. This determines the degree of matching between the inferred entity and the candidate entity under the constraints of the current relationship and time. From the perspective of dimensional modeling, the upper and lower bounds of the inferred box and the predicted box in each dimension are calculated.

[0014]

[0015] The beneficial effects of this invention are as follows: This invention proposes a temporal knowledge graph completion method based on fuzzy geometric representation and a time-aware mechanism. By introducing a time-aware mechanism to effectively model temporal information, the model can better capture the dynamic characteristics of entity relationships in the knowledge graph as they change over time. Simultaneously, fuzzy geometric representation is used to spatially model entities and relationships, improving the flexibility and expressive power of knowledge representation. Based on this, a scoring function is constructed and combined with the model training and prediction process to achieve effective inference of missing facts in incomplete temporal knowledge graphs. Compared with existing methods, this invention can make fuller use of temporal information, improve model training stability and prediction accuracy, and enhance overall computational efficiency while ensuring the completion effect. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a temporal knowledge graph completion method based on fuzzy geometric representation and time awareness mechanism according to the present invention.

[0017] Figure 2 This is a structural diagram of a time encoder according to a specific embodiment of the present invention.

[0018] Figure 3 This is a flowchart of the timing knowledge graph completion operation in a specific embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0020] Reference Figure 1 , Figure 2 and Figure 3 This invention provides a temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism, the method comprising the following steps:

[0021] Step 1: Entity representation and relation modeling.

[0022] First, entities in the temporal knowledge graph are embedded and represented in a vector space to obtain the corresponding entity vector representation. Simultaneously, relationships are modeled as positional constraint parameters for entity boxes in the vector space. This process mainly includes two parts: entity representation and relationship parameter learning.

[0023] During the entity representation phase, two vector parameters are constructed for each entity, namely the basis vector. and offset vector This is used to characterize the basic semantic location and offset features of entities in the embedding space. By applying normalization constraints to the vectors, the formula is... This keeps the embedding vectors within a reasonable numerical range, thereby improving the stability of the model training process and enhancing the interpretability of the geometric representation. Simultaneously, to ensure that the entity representation reflects the structural information of the current facts, a bump mechanism is introduced to initialize and update the entity representation, where the inference entity... The representation of a variable is determined by its own basis vectors and the offset vectors of candidate entities, while the candidate entities... The representation of an entity is determined by its own basis vector and the offset vector of the inference entity, thereby realizing interactive modeling between entities.

[0024] (3)

[0025] During the relation modeling phase, two key parameters are learned for each relation: the center vector. and width vector The center vector describes the core position of the relation in the embedding space, while the width vector characterizes the extent of the corresponding box in each dimension. In this way, relational constraint regions with spatial extent can be formed in the vector space, allowing entities to form geometrically structured representations under the influence of relations.

[0026] Step 2: Constructing the time-aware mechanism.

[0027] To describe the changing characteristics of entity relationships under different time conditions, a time-aware mechanism is introduced into the model. Specifically, discrete timestamps in the temporal knowledge graph are mapped to continuous time vectors using Fourier basis functions. Then, the time vector is fused with the entity embedding vector to obtain the entity representation containing time information, as shown in the formula. In this way, entities can be represented with different semantics at different points in time, thus better depicting the dynamic characteristics of facts changing over time in temporal knowledge graphs.

[0028] Step 3: Construction of fuzzy geometric representation.

[0029] After obtaining the time-aware entity representation, a fuzzy box representation method is introduced to construct a geometric region model in the embedding space, enabling a more flexible description of semantic relationships between entities. Specifically, the time-enhanced entity vectors are combined with relation parameters to construct inference entity boxes and candidate entity boxes, represented as follows:

[0030]

[0031] in This indicates the center position of the physical box within the embedded space. and These represent the upper and lower bounds of the box in each dimension. This geometric structure represents the relationship... Under this effect, the potential regions where candidate entities may be derived from the reasoning entity are located can be defined. This box structure allows for a more natural description of the inclusion, overlap, and spatial constraints between entities.

[0032] Step 4: Constructing the scoring function.

[0033] After constructing the entity boxes and relation regions, candidate entities are evaluated using a scoring function. Specifically, the volume of the intersection region between the inferred entity box and the candidate entity box in the embedding space is calculated and used as a measure of the matching degree. From a dimension-by-dimensional analysis perspective, the upper and lower bounds of the two boxes in each dimension are calculated, and the overall intersection region is obtained accordingly. A larger intersection volume indicates that the candidate entity is a better match to the current relation and time conditions, and is more likely to be the true tail entity.

[0034] Step 5: Model training and prediction.

[0035] After the scoring function is constructed, the model is trained using known facts from the temporal knowledge graph. Positive and negative sample data are constructed, and the corresponding scores are calculated using the scoring function. The model parameters are then optimized using a loss function to ensure that true facts receive higher scores and erroneous facts receive lower scores. After model training is complete, the trained entity representations, relation parameters, and temporal representations can be used to infer and predict unknown facts. By ranking candidate entities by score, the entity with the highest score is selected as the prediction result, thereby completing the missing information in the temporal knowledge graph.

[0036] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism, characterized in that: This method maps entities, relationships, and temporal information from a temporal knowledge graph to a continuous vector space during model training. It then models the semantic space of entities by constructing a fuzzy geometric box representation model, thereby enabling the prediction of missing facts in the knowledge graph. By introducing a temporal encoding mechanism to characterize temporal information and describing the semantic relationships between entities through geometric box interactions, the model is able to perform knowledge reasoning under time constraints. The specific operating steps include: Step 1: Preprocess the time-series knowledge graph data, represent the facts in the knowledge graph as quadruple structures with time information, and initialize the entity set, relation set, and time set with vectors to construct entity embedding representation, relation embedding representation, and time embedding representation. Step 2: Vectorize the timestamps using a time encoding function to map the time information into a continuous vector space, and combine it with the entity embedding representation vector to construct a time-aware temporal embedding representation vector to capture the evolutionary features of entities in the knowledge graph over time. Step 3: Construct a fuzzy geometric box structure based on entity embedding representation. Define the high-dimensional hyperrectangular region of the entity in the semantic space through the center vector and offset vector. Perform spatial transformation on the entity box through relation transformation function to generate inference entity box and candidate entity box. Step 4: Construct a scoring function by calculating the volume of the intersection region between the inferred entity box and the candidate entity box, and optimize the model parameters using the loss function to obtain a stable entity semantic space representation. Step 5: After the model training is completed, the missing facts in the knowledge graph are predicted. The candidate entities are ranked by calculating the matching degree between the reasoning entity box and the candidate entity box, so as to obtain the final knowledge graph completion result.

2. The temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism according to claim 1, characterized in that: The process of preprocessing time-series knowledge graph data and constructing a four-tuple structure represents facts in the knowledge graph as four-tuples containing a head entity, relation, tail entity, and timestamp. And through the entity set Relation sets and time sets Perform vector initialization, in Generate the corresponding embedding representation vector in the dimensional vector space.

3. The temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism according to claim 1, characterized in that: The time encoder described above is used to vectorize time information. It constructs a periodic time encoding function using sine and cosine functions to map the timestamp τ into a time embedding vector. By combining periodic functions of different frequencies, the model can simultaneously capture long-term time trends and short-term time fluctuations.

4. The temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism according to claim 1, characterized in that: The fuzzy geometric box structure is used to represent the distribution range of entities in the semantic space. Each entity is represented by a box structure, which is determined by the center vector and the offset vector. The box structure forms a high-dimensional hyperrectangular region in the embedding space, thereby representing the range of the entity's semantic space. The fuzzy boundary mechanism improves the model's ability to express uncertain semantic relationships.

5. The temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism according to claim 1, characterized in that: The scoring function measures the degree of matching between candidate entities by calculating the volume of the intersection region between the inferred entity box and the candidate entity box. The larger the intersection volume, the higher the semantic matching degree between the inferred entity and the candidate entity.

6. The temporal knowledge graph completion method based on fuzzy geometric representation and time-aware mechanism according to claim 1, characterized in that: During the model training process, the model parameters are optimized by constructing a loss function, training samples are generated through a negative sampling strategy, and a loss function is constructed to make the score of the real fact higher than the score of the negative sample. In this way, the entity embedding, relation embedding, and temporal embedding parameters are continuously updated through the gradient descent algorithm until the model training converges.

Images