Time sequence knowledge graph alignment method for information incomplete scene

Abstract

The invention discloses a time sequence knowledge graph alignment method for an information incomplete scene. The method comprises the following steps: constructing a time sequence knowledge graph; modeling local and global structures of the time sequence knowledge graph to generate entity representation; extracting multi-granularity time information by using a multi-granularity time encoder; generating a self-adaptive similarity matrix and a pseudo-aligned entity pair; performing scale adaptive entity projection on the pseudo-aligned entity pair and the similarity matrix; performing multi-scale hypergraph retrieval on the similarity matrix, the pseudo entity pair and the projection hypergraph; performing multi-scale interaction enhancement fusion on the multi-scale hypergraph; and finally generating a fusion result through multi-round iterative fusion and a multi-scale hypergraph retrieval process, namely predicting the congestion area in the peak period. According to the method, in-scale interaction and cross-scale conflict detection are combined, so that the dynamic relationship can be modeled explicitly and the cross-time-scale inconsistency can be solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation planning technology, and in particular relates to a method for aligning time-series knowledge graphs for scenarios with incomplete information. Background Technology

[0002] As real-world facts continue to evolve, static knowledge graphs (KGs) become less suitable for time-sensitive tasks such as question answering and knowledge reasoning. Therefore, temporal knowledge graphs (TKGs) are attracting increasing attention. A fact in a temporal knowledge graph is typically represented as a quadruple. ,in It is a time range, including the start timestamp. and end timestamp .if If it is not None, this fact is called a valid quadruple or a temporal fact. In the field of intelligent transportation planning, historical traffic conditions, weather, and event (such as accidents) temporal data can be used to optimize real-time route planning. For example, by using urban traffic temporal maps, congestion areas during peak hours can be predicted, or by monitoring the temporal changes in sensor data of bridges and pipelines, the risk of equipment aging or failure can be predicted.

[0003] To increase coverage and eliminate outdated facts in temporal knowledge graphs (TKGs), a key step is temporal knowledge graph alignment (TKGA), which aims to match TKGs from different sources and merge relevant facts by using aligned entities as anchors. The main focus of TKGA is entity alignment (TEA), because the number of entities (typically over 10k) is much greater than the number of relations (typically around a few hundred).

[0004] Several methods have been proposed to address TKGA. Some methods aim to inject temporal information into entity representations, while others directly use temporal information for alignment. These methods have been evaluated and compared on the dicews and yw datasets. However, these benchmark datasets differ significantly from real-world scenarios and are overly simplified, allowing current methods to achieve over 90% Hits1 scores on these datasets. In practical applications, TKGA is far more complex, and current techniques and datasets remain limited to simplified temporal knowledge graph quadruple distributions and temporal relationships. Specifically, dicews is constructed by splitting the temporal knowledge graph icews into two partially overlapping subsets, which significantly reduces the difficulty of TKGA because the knowledge graphs to be aligned come from the same source and share semantics. yw originates from two real-world knowledge graphs, YAGO and Wikidata; however, it is constructed by including quadruples containing the most common temporal relationships. Therefore, if the knowledge graphs to be aligned are essentially knowledge graphs about football players and their teams, this is domain-specific and fails to reflect the realities of TKGA in the real world.

[0005] Real-world temporal knowledge graphs are far more complex than existing simplified baseline graphs, necessitating a shift in research towards TKGA-Wild (i.e., temporal knowledge graph alignment in scenarios with incomplete information). This shift highlights several key challenges: Entanglement of Time Elements at Multiple Scales: Real-world temporal knowledge graphs exhibit entanglement of time elements across two key dimensions: 1) Coexistence of Time Elements at Multiple Granularities: Existing datasets simplify temporal knowledge graphs by discarding non-annual time information to maintain a uniform representation of time intervals, which is impractical in real-world scenarios. In fact, the time information of the same entity may be presented at different granularities (year / month / day). For example, an entity may simultaneously exhibit macroscopic time cycles (patterns spanning several years) and microscopic event records (by month / day). This coexistence of time elements at multiple granularities poses a significant challenge to current TKGA methods, as they typically rely on only a single time granularity (e.g., by year), thus failing to capture complex time patterns across multiple granularities; 2) Topological Differences in Time Intervals: Furthermore, real-world temporal knowledge graphs exhibit extensive topological relationships between time intervals of aligned entities from different sources. A term from a non-TKG source may have completely non-overlapping time intervals. These relationships include disjoint (e.g., an entity spanning 2019–2020, while its aligned entity spans 1998–2000), overlapping (e.g., 2019–2024 vs. 2018–2023), containing (e.g., 2023 vs. 2020–2024), and identical (e.g., 1989–2000 vs. 1989–2000). This diversity of interval topological relationships greatly increases the complexity of temporal relational reasoning, making it difficult for current methods to accurately model and align entities.

[0006] Cross-source temporal structure imbalance: In the TKGA-Wild scenario, another major challenge is the structural asymmetry in two key aspects: 1) The difficulty of learning alignment patterns caused by incomplete multi-source temporal structures. Existing datasets mainly assume that most entities are associated with complete temporal relationships (time-annotated facts including head and tail timestamps). This simplifies real-world scenarios where time annotations from knowledge graphs from different sources are often incomplete (i.e., missing timestamps). Information about different neighbors of aligned entities from different knowledge graphs is also missing time records. This multi-source temporal incompleteness further exacerbates the asymmetry in the alignment dimension, thus hindering the ability of current models to effectively capture temporal associations and achieve effective alignment; 2) Imbalance in temporal event density. Furthermore, some existing simplified datasets often assume that the number of timestamp facts for time-aligned entity associations is roughly similar. However, in real-world scenarios, there is a severe asymmetry in the number of temporal facts between aligned entities. A term may have dozens of timestamp records, while its corresponding entity may only have two. The imbalance in temporal event density poses a significant challenge to existing methods, which fail to effectively model the incompleteness of multi-source temporal structures and the dramatic changes in temporal density. Therefore, under these complex alignment conditions, time-related reasoning becomes unreliable. Summary of the Invention

[0007] To address the issue that current TKGA datasets are overly simplistic and unrealistic, this application proposes a benchmark challenge for Realistic Temporal Knowledge Graph Alignment (TKGA-Wild). Specifically, it creates an initial seed by employing prior mappings between YAGO and Wikidata entities, avoiding resource-intensive manual annotation. Then, these two knowledge graphs are selected, and temporal and non-temporal facts are added to construct a subset of the knowledge graphs to be aligned. These new datasets feature multi-scale temporal element entanglement (i.e., multi-granularity temporal coexistence and temporal interval topological differences), more realistic cross-source temporal structure imbalance (i.e., incomplete multi-source temporal structure and imbalanced temporal event density), and novel challenging alignment scenarios.

[0008] Furthermore, to address key challenges in TKGA-Wild, this application proposes a hypergraph-driven deconstructible retrieval enhancement framework. Technically, the TKGA process is reformulated as a multi-scale hypergraph retrieval enhancement generation method. First, it encodes multi-granularity temporal, structural, and semantic features, and enhances multi-granularity temporal understanding by generating similarity matrices and pseudo-alignment pairs through alignment-based inference. Then, based on the similarity matrix and pseudo-alignment pairs, a scale-adaptive entity projection module decouples and aligns entities across different temporal and relational scales, constructing a projected hypergraph that captures complex temporal topological differences and balances the distribution of temporal event density. This hypergraph, along with similarity cues, is processed by a multi-scale hypergraph retrieval module to construct rich high-order representations, i.e., a multi-scale hypergraph. Finally, a multi-scale interaction enhancement fusion module integrates information from the multi-scale hypergraph through a scale-weaving collaborative mechanism (i.e., intra-scale interactions and conflict detection) to infer the final entity pairs and effectively address the inconsistencies caused by multi-source temporal incompleteness and complex and unbalanced temporal event density distribution. Through iterative optimization, this application generates high-quality entity pairs in the TKGA-Wild scenario.

[0009] To achieve the above objectives, the temporal knowledge graph alignment method disclosed in this application for scenarios with incomplete information includes the following steps: S1: Acquire historical traffic conditions, weather, and event time-series data, and integrate them to construct a time-series knowledge graph; S2: Use a meta-structure encoder to generate entity representations by modeling the local and global structures of the temporal knowledge graph; use a multi-granularity temporal encoder to extract multi-granularity temporal information; adaptively fuse the entity representations and multi-granularity temporal information to generate an adaptive similarity matrix, and generate pseudo-aligned entity pairs accordingly. S3: Perform scale-adaptive entity projection on the pseudo-aligned entity pairs and similarity matrix, specifically including: the pseudo-aligned entities are adaptively aligned according to time and relational features at different scales, thereby generating more informative scale-adaptive alignment projections, and these projections are fused to construct a projection hypergraph. S4: Perform multi-scale hypergraph retrieval on the similarity matrix, pseudo-entity pairs, and projected hypergraph, specifically including: storing information in a projection memory and retrieving relevant entities based on the memory to generate a multi-scale hypergraph; S5: Perform multi-scale interaction enhancement fusion on the multi-scale hypergraph, specifically including: performing inference through a scale weaving collaboration mechanism to fuse temporal information between different scales and obtain a fused graph. The scale weaving collaboration mechanism includes intra-scale interaction and inter-scale conflict detection. The graph feedback updates the metastructure encoder and the multi-granularity temporal encoder, and steps S2-S5 are repeated. S6: Through multiple rounds of iterative fusion and multi-scale hypergraph retrieval processes, the final fusion result is generated, which predicts the congested areas during peak hours.

[0010] Preferably, the metastructure encoder combines a partial random walk with a skip-gram-based embedding framework. Specifically, the random walk process preserves relational information by generating edge-annotated paths, with the path format being... , where r l-1 and e l These are the (l-1)th relation and the l-th entity, respectively, and the jump probability is affected by the hyperparameter. Controls are implemented to strike a balance between breadth-first and depth-first search behaviors; From the entity Jump to entity The probability is defined as: ; in, Representing entities A set of one-hop neighbors, excluding its preceding entity. e i+1 It is the next entity; each jump also retrieves the triple. Corresponding relationships The generated path is treated as a sentence, with entities and relations treated as tokens. Then, a Skip-gram model is employed, and a linear transformation matrix is ​​introduced. To learn the structural embedding representation of entities and relations dw is the structure embedding, and N is the total number of entities.

[0011] Preferably, the multi-granularity time encoder employs the Time2Vec method to encode temporal features at multiple scales; Time is decomposed into three granularities: year, month, and day; for each entity, its time occurrence is marked based on the timestamps of relevant facts in the temporal knowledge graph. For each entity Each particle size Each generates a binary time vector. , where |T g | represents the total number of time granularities. For the i-th binary time vector generated at granularity g, if the entity In particle size The next If someone participates in a certain fact at a certain point in time, then... Otherwise, it is 0; For each activated time point in the time vector Its Time2Vec code is calculated using the following formula: ; in, and These are learnable parameters, where i is the time node and kim is the number of dimensions. Ultimately, at a certain granularity Below, entity The temporal embedding representation is aggregated from the Time2Vec representations of all its activation time points, and is transformed by a linear transformation: ; in Representing entities In particle size The set of time points in the time vector that have a value of 1. Granularity The trainable projection matrix is ​​as follows; Therefore, the encoder ultimately generates a set of multi-granularity temporal embeddings for each entity: These are yearly time embedding, monthly time embedding, and daily time embedding, respectively.

[0012] Preferably, the adaptive fusion is as follows: ; Among them, name embedding It is obtained through the unsupervised BERT double transformation method, where N is the total number of entities; During the training phase of adaptive fusion, the boundary ranking loss function is used as the optimization objective for temporal knowledge graph fusion inference; during the testing phase, cross-domain similarity local scaling is used as a similarity metric between entity embeddings to calculate the adaptive similarity matrix between entities in the source and target graphs; based on the adaptive similarity matrix... For each source entity, the target entity with the highest similarity score is retrieved and used as a pseudo-fusion entity pair.

[0013] Preferably, the scale-adaptive entity projection pairs the adaptive similarity matrix. Each temporal entity in the graph undergoes multi-dimensional temporal and relational projections, and a projection hypergraph is constructed. The time projection employs a timestamp masking projection operator for the first k similar target entities corresponding to each source entity in the adaptive similarity matrix. Remove timestamps and their corresponding facts from the target entity that do not appear in the source entity's time set; The relation projection uses a relation-type masking projection operator for the first k similar target entities corresponding to each source entity in the adaptive similarity matrix. Remove relationship types and related facts from the target entity that did not appear in the source entity; Specifically, given a source entity and a target entity Its time projection and relational projection processes are formally represented as follows: ; in, and These represent the set of timestamps associated with an entity and the set of relation types, respectively; Mask represents the covering. and Information that is not available; through time projection and relational projection, for each target entity In its similar source entities During the matching process, time projection and relational projection results are generated. Each maintains consistency in its time and relationship, ultimately forming a size of The set of projections, where Indicates the number of source entities; Based on the above projection set, a dynamic projection hypergraph is constructed. ; where the set of nodes in the hypergraph Includes the original target entity and all its corresponding projections; hyperedge set Dynamically generated based on the top k most similar target entities obtained by each source entity in the fusion similarity matrix.

[0014] Preferably, the multi-scale hypergraph retrieval constructs a projection memory library, dynamically selects the top k most relevant entity projections, integrates them to construct a multi-scale hypergraph structure, and finally fuses them to generate a multi-scale hypergraph structure, so as to more efficiently perform temporal representation in low-quality data scenarios. The projection memory library It includes the target entity and its related projection information; the multi-scale hypergraph retrieval retrieves the k most relevant projection information for each source entity from the memory; Combining the projected hypergraph and the retrieved front Projection By fusing with abstract projection, a multi-scale hypergraph is constructed; the multi-scale hypergraph consists of three layers: The first layer captures the most detailed parts and consists of a projection hypergraph generated by scale-adaptive entity projection. The projection hypergraph originates from the metastructure encoder, the multi-granularity temporal encoder, and the similarity matrix generated by adaptive ensemble. Combined with scale-adaptive solid projection; The second layer consists of relevant projections retrieved from the projection memory; in this layer, each supernode represents a retrieved relevant projection, each superedge represents a query, and each superedge corresponds to a relevant projection retrieved under a specific source entity. The third layer performs abstract fusion and simplification, merging and compressing the information from the second layer. Specifically, it merges the supernodes in the second layer based on whether they correspond to the same target entity, ultimately forming a supernode representing the target entity. Each superedge is consistent with the second layer, representing a query, and each superedge corresponds to the relevant target entity fused under the source entity. In form, multi-scale hypergraph The structure is as follows: ; in, Let V1, V2, and V3 represent the sets of supernodes in the first, second, and third layers, respectively. Let E1, E2, and E3 represent the set of hyperedges connecting the relevant projections and the target entity in the three layers, respectively.

[0015] Preferably, the multi-scale interaction enhancement fusion includes: a scale weaving collaboration mechanism and multi-scale fusion inference; wherein, the scale weaving collaboration mechanism integrates same-scale interaction and cross-scale conflict detection capabilities to achieve scale adaptive supplementation and pruning of time-incomplete or unbalanced data, adjust the density of time events, and solve the problem of information inconsistency between different scales; The intra-scale interactions include: for a certain scale, for similar entity pairs in the hypergraph, the temporal fact supplementation operation uses a large language model to analyze the context of the entity pair and supplement the missing core temporal information; the temporal fact pruning operation removes irrelevant temporal information. The multi-scale fusion inference includes: transforming the cross-scale temporal alignment problem into a selection task to handle two types of multi-scale temporal information inputs: 1) outputs from intra-scale interactive components, i.e., the updated multi-scale hypergraph; 2) outputs from conflict detection, i.e., conflict subgraphs. For the first type of input, the module combines similar temporal fusion subgraphs at the same source entity and the same scale, transforming the problem into a selection task for each source entity. The large language model selects relevant temporal fusion subgraphs from the corresponding entities to obtain the fusion result at that scale. E s and E t These are the source entity set and the target entity set, respectively. For the second type of input, it is transformed into a cross-scale selection task, whereby the large language model selects the most relevant target entity for each source entity from the conflict fusion subgraph across all scales, generating the fusion result after conflict resolution. .

[0016] Preferably, for conflicts between scales, a conflict detection mechanism is used to perform secondary feedback based on the multi-scale fusion inference results. Specifically, the fusion subgraph at each scale is extracted from the fusion inference results. 'l' is the scale number, which checks whether there are multiple target entities corresponding to the same source entity; if so, it is a conflict. The fused subgraphs containing the same source entity are grouped into a set and fed back to the fusion inference step for further processing; Iterative update: The final set of fused subgraphs obtained through multi-scale fusion inference ; The fused subgraph is fed back to the multi-granularity temporal encoder and the meta-structure encoder, thereby triggering a new round of iterations; after I n After rounds of iteration, the final fusion result is generated.

[0017] The beneficial effects of this application are as follows: We propose a real-world temporal knowledge graph alignment approach, which addresses the entanglement of multi-scale temporal elements and more realistic cross-source temporal structural imbalances. To better reflect these challenges, we also establish a more realistic benchmark dataset, WildBETA.

[0018] A novel and efficient method, HyDRA, is proposed, which extends the concept of multi-granularity encoding. This application is the first to reformulate the TKGA process as a multi-scale hypergraph retrieval enhancement generation method, and combines key modules, including scale-adaptive entity projection, multi-scale hypergraph retrieval, and multi-scale interactive enhancement fusion, which together address the main challenges of TKGA-Wild.

[0019] A scale-weaving coordination mechanism was designed for HyDRA to explicitly address the challenges of multi-source temporal incompleteness and inconsistencies arising from complex and uneven distributions of temporal event densities. These mechanisms effectively combine intra-scale interactions with cross-scale conflict detection, facilitating dynamic reasoning capabilities in complex temporal alignments. Attached Figure Description

[0020] Figure 1 The alignment method flow of this application; Figure 2 Overall architecture diagram of this application; Figure 3 Schematic diagram of a multi-scale hypergraph; Figure 4 Explanation of the symbols used in this application. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0022] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0023] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0024] The technical solutions provided in this application involve technologies such as machine learning and natural language processing in artificial intelligence, which are specifically described and illustrated through the following embodiments.

[0025] Before introducing the embodiments of this application, some terms involved in this application will be explained.

[0026] 1. Temporal Knowledge Graph Alignment: Most existing methods extend traditional static knowledge graph alignment methods by introducing temporal encoding, thereby improving alignment performance. In particular, TEAGNN and TREA were among the earliest methods to explore the TKGA task. Unlike temporal GNN models that decompose the temporal graph into multiple snapshots, these models treat timestamps as attributes of connections between entities.

[0027] Unlike existing methods, this application formulates the TKGA-Wild task as a multi-scale hypergraph retrieval enhancement generation paradigm. This application focuses on modeling many neglected key issues, including multi-granularity temporal information, topological differences in time intervals, complex alignment patterns caused by multi-source temporal incompleteness, and unbalanced distribution of temporal event density.

[0028] 2. Retrieval-Enhanced Generation: Retrieval-enhanced generation (RAG) has attracted widespread attention in various fields in recent years because it can improve task performance by capturing information and enhancing model performance in knowledge-intensive tasks. For example, Self-RAG improves the generation quality of large language models (LLMs) by retrieving external information on demand and combining self-reflection. RAKG improves knowledge graph construction by organizing training data in a retrieval-enhanced manner, allowing the retrieval of relevant text and graphical content. RAGTrans stores training data in multimodal memory for knowledge retrieval and aggregates neighborhood information through a multimodal hypergraph transformer to enhance user content generation.

[0029] 3. Temporal Knowledge Graph Alignment Datasets: Currently, the most widely used datasets, dicews and yw, are extracted from icews, yago, and wikidata. Existing temporal knowledge graph alignment datasets, either derived from a single knowledge graph or with simplified distributions of quadruples and time elements, may not serve as fair benchmarks for evaluating TKGA performance, especially in complex scenarios with incomplete information.

[0030] refer to Figure 1 and Figure 2This application first utilizes two key encoding modules for feature extraction: a metastructure encoder module generates entity representations by modeling the local and global structures of temporal knowledge graphs (TKGs); while a multi-granularity temporal encoder module focuses on modeling multi-granularity temporal information. The features learned by the two modules are then fed into an adaptive fusion module, which generates an adaptive similarity matrix and produces pseudo-aligned entity pairs accordingly.

[0031] Next, the pseudo-entity pairs and the adaptively fused similarity matrix are input into the scale-adaptive entity projection module. In this module, entities are adaptively aligned at different scales based on temporal and relational features, thereby generating more informative scale-adaptive aligned projections. These projections are further fused to construct a projection hypergraph.

[0032] Subsequently, the adaptively fused similarity matrix, pseudo-entity pairs, and projected hypergraph are fed into the multi-scale hypergraph retrieval module. This module generates a multi-scale hypergraph with high-order information representation capabilities by storing information in a projection memory and retrieving relevant entities based on that memory. These multi-scale hypergraphs are then input into the multi-scale interaction enhancement fusion module, where inference is performed through a scale weaving collaborative mechanism (including intra-scale interactions and inter-scale conflict detection) to fuse temporal information across different scales and derive the final fused graph. This result is then fed back to update the two aforementioned key encoding modules.

[0033] Through a multi-round iterative fusion and multi-scale hypergraph retrieval process, this application can effectively overcome the challenge of low data quality, effectively characterize and fuse data, and ultimately generate high-quality fusion results.

[0034] To effectively support subsequent multi-scale retrieval and fusion processes, the system must provide rich, fine-grained, and semantically consistent entity representations. Therefore, in one embodiment, this application first introduces a set of multi-granularity information encoders to model the structure and temporal context of entities from complementary perspectives. These encoders lay the foundation for addressing the core challenge of incomplete information scenarios—namely, multi-granularity temporal coexistence—capturing both static topological patterns and dynamic temporal behavior. The generated multi-view embeddings are then integrated into an adaptive fusion module to generate an initial alignment signal, providing the basis for initiating the entire fusion framework.

[0035] Metastructure Encoder: First, a metastructure encoder is introduced to effectively capture local and global structural patterns in TKGs. This encoder combines biased random walks with a skip-gram-based embedding framework and does not rely on supervision signals. Specifically, the random walk process preserves relational information by generating edge-annotated paths, with the path format as follows: The jump probability is affected by the hyperparameter. Control is needed to strike a balance between breadth-first and depth-first search behavior. From entities Jump to entity The probability is defined as: ; in, Representing entities A set of one-hop neighbors, excluding its preceding entity. Each redirect also requires searching for the triple. Corresponding relationships The generated path is treated as a sentence, with entities and relations treated as tokens. Next, a Skip-gram model is employed, and a linear transformation matrix is ​​introduced. To learn the structural embedding representation of entities and relations This process can effectively encode key structural information in incomplete temporal knowledge graphs.

[0036] Multi-granularity temporal encoder: The multi-granularity temporal encoder aims to capture the temporal patterns of entities with disordered granular information at different time granularities (i.e., year, month, day). By modeling coarse-grained and fine-grained temporal relationships, it enhances the fusion capability of temporal knowledge graphs over a larger time range. The Time2Vec method is employed to encode temporal features at multiple scales.

[0037] For each entity, its temporal occurrence is labeled based on the timestamps of related facts in the temporal knowledge graph. Here, time is decomposed into three granularities: year, month, and day. The year-level representation aims to capture the overall temporal trend and similarity between entities, while the month-level and day-level representations focus on more refined time alignment features.

[0038] For each entity Each particle size Each of these will generate a binary time vector. If the entity In particle size The next If someone participates in a certain fact at a certain point in time, then... Otherwise, it is 0.

[0039] For each activated time point in the time vector Its Time2Vec code can be calculated using the following formula: ; in and These are learnable parameters. This design enables the model to capture both linear and periodic time patterns.

[0040] Ultimately, at a certain granularity Below, entity The temporal embedding representation is aggregated from the Time2Vec representations of all its activation time points, and is transformed by a linear transformation: ; in Representing entities In particle size The set of time points in the time vector that have a value of 1. Granularity The trainable projection matrix below.

[0041] Therefore, the encoder ultimately generates a set of multi-granularity temporal embeddings for each entity: This set will be further used in subsequent fusion of incomplete temporal knowledge graphs.

[0042] Adaptive Fusion: Finally, by adaptively fusing different types of embeddings, a high-quality entity embedding representation is generated, as shown below: ; Among them, name embedding It is obtained through an unsupervised BERT dual transform method, which can efficiently encode entity names into the same dimensional space as other embeddings. This design reflects HyDRA's intention to further introduce multimodal semantic information to improve embedding quality.

[0043] In the training phase, the Margin Ranking Loss function is used as the optimization objective for temporal knowledge graph fusion inference. In the testing phase, Cross-domain Similarity Local Scaling (CSLS) is used as the similarity metric between entity embeddings to calculate an adaptive similarity matrix between entities in the source and target graphs. Based on this similarity matrix... For each source entity, the target entity with the highest similarity score can be retrieved and treated as a pseudo-fused entity pair.

[0044] Although the fusion similarity matrix obtained through adaptive fusion initially reveals the similarity relationships between temporal entities at different time granularities in low-quality data scenarios, this simple multi-granularity temporal information embedding cannot fully reflect the true correlation between entities. This is mainly due to the complex scale time intervals and incomplete information of temporal events commonly found in low-quality data scenarios. To further address these challenges, in one embodiment, this application proposes a scale-adaptive entity projection module: this module, through adaptive similarity matrix... By performing deep decoupling, each temporal entity is projected with fine-grained scale awareness along multiple time and relational dimensions, and a projection hypergraph is constructed to achieve adaptive representation of entities at different spatiotemporal scales.

[0045] Adaptive Temporal & Relational Projection: The core of this module lies in constructing a fine-grained temporal alignment mapping mechanism. For the top-k similar target entities corresponding to each source entity in the adaptive similarity matrix, this application designs a dual-channel projection operator: 1) In the time dimension, a timestamp masking projection operator is used. 1) Remove timestamps and their corresponding facts from the target entity that do not appear in the source entity's time set; 2) At the relational dimension, use the relation type masking projection operator. This involves removing relation types and related facts from the target entity that do not appear in the source entity. Specifically, given a source entity... and a target entity Its projection process can be formally represented as: ; in and These represent the set of timestamps and the set of relation types associated with the entity, respectively. Through the aforementioned dual-channel selective projection, for each target entity... In its similar source entities During the matching process, two projection results will be generated. Each maintains consistency in its time and relationship, ultimately forming a size of The set of projections, where This indicates the number of source entities. This decoupled projection mechanism effectively solves the problems of semantic and temporal drift and missing data in temporal knowledge graphs under low-quality data scenarios, while preserving the distinguishable features of entities across different dimensions.

[0046] Projective Hypergraph Construction: Based on the above set of projections, construct a dynamic projective hypergraph. The set of nodes in a hypergraph. Includes the original target entity and all its corresponding projections; hyperedge set Then, based on the similarity obtained by each source entity in the fusion similarity matrix... The most similar target entity is dynamically generated.

[0047] This design endows the projective hypergraph with the following two key properties: 1) By utilizing multi-dimensional projection, the hypergraph can capture multi-scale temporal interaction patterns, enabling the model to distinguish and align asynchronous or uneven time intervals, thereby explicitly representing complex temporal interval differences; 2) Combined with projection after relation masking, the model can adaptively capture the true semantic correlations between different relation types in low-quality data scenarios. Notably, the hypergraph dynamically updates with changes in the adaptive fusion matrix, thus achieving the gradual modeling of complex dependencies between temporal entities. This mechanism effectively overcomes the representational rigidity problem faced by traditional static graph models when dealing with incomplete temporal knowledge graphs in open, time-varying environments, providing a more flexible representation space for knowledge graph fusion.

[0048] This application aims to retrieve projected entities from a target temporal knowledge graph that are genuinely relevant to the source entities and aligned with their time scales. This eliminates irrelevant or misleading temporal information interference, thereby contributing to the achievement of the goal of temporal knowledge graph fusion tasks in low-quality data scenarios. In one embodiment, this application re-represents the temporal knowledge graph fusion task in low-quality data scenarios as a retrieval enhancement generation process and proposes a multi-scale hypergraph retrieval module. This module dynamically selects projected entities by constructing a projection memory. The most relevant entity projections are selected and integrated into a multi-scale hypergraph structure. Finally, the multi-scale hypergraph structure is fused to generate a more efficient temporal representation of low-quality data scenarios.

[0049] Projection memory and retrieval mechanism: First, construct a projection memory. This includes the target entity and its associated projection information. The projection information covers multiple aspects of the target entity's characteristics, such as different scales of time and relational information, aiming to reflect the multi-dimensional characteristics of the target entity and thus achieve more granular relevance assessment. To efficiently utilize this resource, a relevant entity retrieval module is designed to retrieve the k most relevant projection information for each source entity from the memory. This design effectively alleviates the challenges posed by low-quality data by differentially processing the projection information, thereby improving retrieval performance. Specifically, the retrieval operation enables HyDRA to accurately extract the target entity projection information related to the source entity from the memory.

[0050] Multi-scale hypergraph structure: Subsequently, in order to effectively integrate information at different scales, this is achieved by combining projected hypergraphs and retrieved front-end hypergraphs. Projection By fusing with abstract projection, a multi-scale hypergraph was constructed. For example... Figure 3 As shown, this structure captures higher-dimensional and more refined temporal information representations. Specifically, the constructed multi-scale hypergraph consists of three layers: - Level 1 ( The most detailed parts are captured, consisting of a projected hypergraph generated by the scale-adaptive entity projection module. This hypergraph mainly originates from the metastructure encoder and the multi-granularity temporal encoder, as well as the similarity matrix generated by adaptive ensemble. It combines scale-adaptive solid projection. It can initially capture effective temporal information under conditions of multi-granularity temporal coexistence, topological differences in time intervals, and unbalanced temporal event density.

[0051] - Level 2 ( This layer consists of relevant projections retrieved from the projection memory. In this layer, each hypernode represents a retrieved relevant projection, and each hyperedge represents a query (i.e., a source entity). Each hyperedge corresponds to a relevant projection retrieved for a specific source entity. By adopting the RAG perspective, multi-scale information is reorganized, thereby capturing more relevant temporal information.

[0052] - Level 3 ( This process performs abstract fusion and simplification, merging and compressing information from Layer 2 into a more abstract form. Specifically, supernodes in Layer 2 are merged based on whether they correspond to the same target entity, ultimately forming a single supernode representing the target entity. Each hyperedge remains consistent with Layer 2, representing a query (i.e., the source entity), and each hyperedge corresponds to a relevant target entity merged under the source entity. This further compresses temporal information and focuses on time alignment at a higher level.

[0053] This hierarchical design allows multi-scale hypergraphs to simultaneously preserve low-level temporal details (derived from multi-granularity and broad scale-adaptive projections, i.e., Layer 1) and abstract high-level alignments (derived from relevant projection retrieval and abstract fusion, i.e., Layers 2 and 3), thus more expressively and adaptively displaying the relationships between source and target entities. Formally, multi-scale hypergraphs... The structure is as follows: ; in, This represents the set of supernodes in a three-layer structure. This represents the set of hyperedges connecting the relevant projections and the target entity in the three layers. This multi-scale structure can dynamically adjust the temporal dependencies between entities and ensure the flexibility of the alignment process.

[0054] To facilitate interaction and information utilization of temporally aligned patterns across different scales, a direct approach is to mix entity feature information from different scales. However, this application argues that intra-scale interaction and cross-scale interaction represent different dimensions of temporal feature pattern interaction. The former mainly describes detailed interactions at similar time scales, while the latter focuses on changes and conflicts in entity features at a macroscopic level. Therefore, in one embodiment, this application proposes a multi-scale interaction enhancement fusion module to replace the simple mixing approach. This module mainly comprises two components: a scale weaving coordination mechanism and multi-scale fusion inference. The scale weaving coordination mechanism integrates intra-scale interaction and cross-scale conflict detection capabilities, enabling scale-adaptive supplementation and pruning of incomplete or unbalanced temporal data, which helps adjust the density of temporal events and resolve information inconsistencies between different scales. Based on this, the multi-scale fusion inference module further adaptively aggregates the multi-scale representation enhanced by interaction, promoting a more robust and consistent understanding of temporal features.

[0055] Intra-scale interaction: Due to the low data quality in open, time-varying environments, traditional fusion models are prone to information bottlenecks. Recent research shows that group interaction mechanisms can provide richer insights into incomplete temporal information. To achieve the alignment of similar temporal entities at the same scale, this application introduces temporal fact supplementation and temporal fact pruning operations into the intra-scale interaction mechanism. Specifically, for a certain scale $l$, for similar entity pairs in the hypergraph, the temporal fact supplementation operation allows LLM analysis of the context of the entity pair and supplements missing core temporal information (including time points, relationships, or facts); while the temporal fact pruning operation is responsible for removing irrelevant temporal information.

[0056] Multi-scale fusion reasoning: To reduce the search space and alleviate the "illusion phenomenon" in fusion reasoning, this application proposes a multi-scale fusion reasoning component. This component transforms the cross-scale temporal alignment problem into a selection task, which can handle two types of multi-scale temporal information inputs: 1) the output from intra-scale interaction components, i.e., the updated multi-scale hypergraph; 2) the output from conflict detection, i.e., conflict subgraphs. For the first type of input, the module combines similar temporal fusion subgraphs at the same source entity and the same scale, transforming the problem into a selection task for each source entity. The LLM selects relevant temporal fusion subgraphs from the corresponding entities to obtain the fusion result at that scale, denoted as . For the second type of input, the module transforms it into a cross-scale selection task. Specifically, LLM needs to select the most relevant target entity for each source entity from the conflict fusion subgraphs across all scales, generating the fusion result after conflict resolution. This fusion method combines a scale weaving coordination mechanism (i.e., intra-scale interaction and inter-scale conflict detection) to achieve consistent time-based reasoning while maintaining computational efficiency.

[0057] Conflict Detection: Relying solely on same-scale interaction mechanisms may not effectively avoid inter-scale conflicts, thus affecting the efficiency of fusion inference. A direct solution is to model and detect conflicts for entity interactions at all scales, but this introduces a large amount of redundant information and increases computational complexity. In fact, many subtle conflicts can be resolved automatically by the multi-scale fusion inference module. Therefore, this application proposes a conflict detection mechanism that uses secondary feedback based on the multi-scale fusion inference results to resolve conflicts arising from macro-micro scale variations. Specifically, it extracts the fusion subgraph for each scale from the fusion inference results. The system checks if there are multiple target entities corresponding to the same source entity. If so, they are considered to be in conflict. These fusion subgraphs containing the same source entity are grouped into a small set and fed back to the fusion inference module for further processing.

[0058] Iterative update: The final set of fused subgraphs obtained through multi-scale fusion inference ; This will be used as a seed fusion subgraph and fed back to the multi-granularity temporal encoder and metastructure encoder, thereby triggering a new round of iterations. After I... n After several iterations, the adaptive integration module is finally able to generate high-quality fusion results.

[0059] The symbols used in this application are explained in reference to [reference]. Figure 4 .

[0060] The beneficial effects of this application are as follows: Based on multi-scale temporal element entanglement (i.e., multi-granularity temporal coexistence and topological differences in time intervals) and cross-source temporal structure imbalance (i.e., incomplete multi-source temporal structure and imbalance of temporal event density). A novel TKGA-Wild framework is proposed to efficiently address the challenges of TKGA-Wild from the perspective of enhanced generation of multi-scale hypergraph retrieval.

[0061] To further mitigate the multi-source temporal incompleteness and temporal event density imbalance in TKGA-Wild, a scale weaving coordination mechanism was designed. Specifically, these mechanisms combine intra-scale interactions and cross-scale conflict detection, enabling explicit modeling of dynamic relationships and resolving inconsistencies across time scales.

[0062] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0063] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component (e.g., is functionally equivalent to it), even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although specific features of this disclosure have been disclosed with respect to only one of several implementations, such features may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0064] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0065] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A temporal knowledge graph alignment method for scenarios with incomplete information, characterized in that, Includes the following steps: S1: Acquire historical traffic conditions, weather, and event time-series data, and integrate them to construct a time-series knowledge graph; S2: Use a meta-structure encoder to generate entity representations by modeling the local and global structures of the temporal knowledge graph; use a multi-granularity temporal encoder to extract multi-granularity temporal information; adaptively fuse the entity representations and multi-granularity temporal information to generate an adaptive similarity matrix, and generate pseudo-aligned entity pairs accordingly. S3: Perform scale-adaptive entity projection on the pseudo-aligned entity pairs and similarity matrix, specifically including: the pseudo-aligned entities are adaptively aligned according to time and relational features at different scales, thereby generating more informative scale-adaptive alignment projections, and these projections are fused to construct a projection hypergraph. S4: Perform multi-scale hypergraph retrieval on the similarity matrix, pseudo-entity pairs, and projected hypergraph, specifically including: storing information in a projection memory and retrieving relevant entities based on the memory to generate a multi-scale hypergraph; S5: Perform multi-scale interaction enhancement fusion on the multi-scale hypergraph, specifically including: performing inference through a scale weaving collaboration mechanism to fuse temporal information between different scales and obtain a fused graph. The scale weaving collaboration mechanism includes intra-scale interaction and inter-scale conflict detection. The graph feedback updates the metastructure encoder and the multi-granularity temporal encoder, and steps S2-S5 are repeated. S6: Through multiple rounds of iterative fusion and multi-scale hypergraph retrieval processes, the final fusion result is generated, which is the predicted congestion area during peak hours.

2. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 1, characterized in that, The metastructure encoder combines a partial random walk with a skip-gram-based embedding framework. Specifically, the random walk process preserves relational information by generating edge-annotated paths, with the path format being... , where r l-1 and e l These are the (l-1)th relation and the l-th entity, respectively, and the jump probability is affected by the hyperparameter. Controls are implemented to strike a balance between breadth-first and depth-first search behaviors; From the entity Jump to entity The probability is defined as: ； in, Representing entities A set of one-hop neighbors, excluding its preceding entity. e i+1 It is the next entity. Representing entities and entity e i+1 The distance between them; each jump also retrieves the triple. Corresponding relationships The generated path is treated as a sentence, with entities and relations treated as tokens. Then, a Skip-gram model is employed, and a linear transformation matrix is ​​introduced. To learn the structural embedding representation of entities and relations dw is the structure embedding, N is the total number of entities, and n is the entity label.

3. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 2, characterized in that, The multi-granularity temporal encoder employs the Time2Vec method to encode temporal features at multiple scales; Time is broken down into three granularities: year, month, and day; For each entity, its timing is marked based on the timestamps of relevant facts in the temporal knowledge graph; For each entity Each particle size Each generates a binary time vector. , where |T g | represents the total number of time granularities. For the i-th binary time vector generated at granularity g, if the entity In particle size The next If someone participates in a certain fact at a certain point in time, then... Otherwise, it is 0; For each activated time point in the time vector Its Time2Vec code is calculated using the following formula: ； in, and These are learnable parameters, where i is the time node and kim is the number of dimensions. Ultimately, at a certain granularity Below, entity The temporal embedding representation is aggregated from the Time2Vec representations of all its activation time points, and is transformed by a linear transformation: ； in Representing entities In particle size The set of time points in the lower time vector that have a value of 1. Granularity The trainable projection matrix is ​​as follows; Therefore, the encoder ultimately generates a set of multi-granularity temporal embeddings for each entity: These are yearly time embedding, monthly time embedding, and daily time embedding, respectively.

4. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 3, characterized in that, The adaptive fusion is described in detail below: ； Among them, name embedding It is obtained through the unsupervised BERT double transformation method, where N is the total number of entities; During the training phase of adaptive fusion, the boundary ranking loss function is used as the optimization objective for temporal knowledge graph fusion inference; during the testing phase, cross-domain similarity local scaling is used as a similarity metric between entity embeddings to calculate the adaptive similarity matrix between entities in the source and target graphs; based on the adaptive similarity matrix... For each source entity, the target entity with the highest similarity score is retrieved and used as a pseudo-fusion entity pair.

5. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 4, characterized in that, The scale-adaptive entity projection pairs with the adaptive similarity matrix. Each temporal entity in the graph undergoes multi-dimensional temporal and relational projections, and a projection hypergraph is constructed. The time projection employs a timestamp masking projection operator for the first k similar target entities corresponding to each source entity in the adaptive similarity matrix. Remove timestamps and their corresponding facts from the target entity that do not appear in the source entity's time set; The relation projection uses a relation-type masking projection operator for the first k similar target entities corresponding to each source entity in the adaptive similarity matrix. Remove relationship types and related facts from the target entity that did not appear in the source entity; Specifically, given a source entity and a target entity Its time projection and relational projection processes are formally represented as follows: ； in, and These represent the set of timestamps associated with an entity and the set of relation types, respectively; Mask represents the covering. and Information that is not available; through time projection and relational projection, for each target entity In its similar source entities During the matching process, time projection and relational projection results are generated. Each maintains consistency in its time and relationship, ultimately forming a size of The set of projections, where Indicates the number of source entities; Based on the above projection set, a dynamic projection hypergraph is constructed. ; where the set of nodes in the hypergraph Includes the original target entity and all its corresponding projections; hyperedge set The projection is dynamically generated based on the top k most similar target entities obtained by each source entity in the fusion similarity matrix, with the subscript P indicating projection.

6. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 5, characterized in that, The multi-scale hypergraph retrieval constructs a projection memory library, dynamically selects the top k most relevant entity projections, integrates them to construct a multi-scale hypergraph structure, and finally fuses them to generate a multi-scale hypergraph structure, so as to more efficiently perform temporal representation in low-quality data scenarios. The projection memory library It includes the target entity and its related projection information; the multi-scale hypergraph retrieval retrieves the k most relevant projection information for each source entity from the memory; Combining the projected hypermap and the retrieved front Projection By fusing with abstract projection, a multi-scale hypergraph is constructed; the multi-scale hypergraph consists of three layers: The first layer captures the most detailed parts and consists of a projection hypergraph generated by scale-adaptive entity projection. The projection hypergraph originates from the metastructure encoder, the multi-granularity temporal encoder, and the similarity matrix generated by adaptive ensemble. Combined with scale-adaptive solid projection; The second layer consists of relevant projections retrieved from the projection memory; in this layer, each supernode represents a retrieved relevant projection, each superedge represents a query, and each superedge corresponds to a relevant projection retrieved under a specific source entity. The third layer performs abstract fusion and simplification, merging and compressing the information from the second layer. Specifically, it merges the supernodes in the second layer based on whether they correspond to the same target entity, ultimately forming a supernode representing the target entity. Each superedge is consistent with the second layer, representing a query, and each superedge corresponds to the relevant target entity fused under the source entity. In form, multi-scale hypergraph The structure is as follows: ； in, Let V1, V2, and V3 represent the sets of supernodes in the first, second, and third layers, respectively. E1, E2, and E3 represent the hyperedge sets connecting the relevant projections and the target entity in the three layers, respectively. The subscript m indicates multi-scale.

7. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 6, characterized in that, Multi-scale interaction enhancement fusion includes: scale weaving collaboration mechanism and multi-scale fusion inference; among them, scale weaving collaboration mechanism integrates same-scale interaction and cross-scale conflict detection capabilities to achieve scale adaptive supplementation and pruning of time-incomplete or unbalanced data, adjust the density of time events and solve the problem of information inconsistency between different scales. The intra-scale interactions include: for a certain scale, for similar entity pairs in the hypergraph, the temporal fact supplementation operation uses a large language model to analyze the context of the entity pair and supplement the missing core temporal information; and the temporal fact pruning operation removes irrelevant temporal information. The multi-scale fusion inference includes: transforming the cross-scale temporal alignment problem into a selection task to handle two types of multi-scale temporal information inputs: 1) outputs from intra-scale interactive components, i.e., the updated multi-scale hypergraph; 2) outputs from conflict detection, i.e., conflict subgraphs. For the first type of input, the module combines similar temporal fusion subgraphs at the same source entity and the same scale, transforming the problem into a selection task for each source entity. The large language model selects relevant temporal fusion subgraphs from the corresponding entities to obtain the fusion result at that scale. E s and E t These are the source entity set and the target entity set, respectively, where l is the scale number. For the second type of input, it is transformed into a cross-scale selection task, whereby the large language model selects the most relevant target entity for each source entity from the conflict fusion subgraph across all scales, generating the fusion result after conflict resolution. .

8. The temporal knowledge graph alignment method for scenarios with incomplete information as described in claim 7, characterized in that, For conflicts between scales, a conflict detection mechanism is used to perform secondary feedback based on the multi-scale fusion inference results. Specifically, the fusion subgraph at each scale is extracted from the fusion inference results. 'l' is the scale number, which checks whether there are multiple target entities corresponding to the same source entity; if so, it is a conflict. The fused subgraphs containing the same source entity are grouped into a set and fed back to the fusion inference step for further processing; Iterative update: The final set of fused subgraphs obtained through multi-scale fusion inference ； The fused subgraph is fed back to the multi-granularity temporal encoder and the metastructure encoder, thereby triggering a new round of iteration; After I n After rounds of iteration, the final fusion result is generated.

Images