A time-series knowledge graph error detection method fusing representation learning and logical rules

Abstract

A kind of time sequence knowledge graph error detection method based on double-path architecture, for the problem of confusing fact error and obsolete fact and other heterogeneous noise in automatically constructed time sequence knowledge graph, a kind of error detection framework combining representation learning and logical rules is proposed.The execution steps of the method mainly include: (1) construct neural representation learning path, from global historical context and local recent state conversion, model semantic evolution with time perception;(2) construct symbolic reasoning path, by mining time sequence logical rules and performing explicit logical consistency check, to provide explainable reasoning evidence;(3) introduce time perception relationship gate mechanism and adaptive weighted fusion strategy, integrate neural embedding rationality score and logical confidence score, generate unified validity evaluation for each time sequence quadruple.The invention effectively combines the generalization ability of neural time sequence modeling and the explainability of time sequence logic verification, providing a robust and highly explainable error detection scheme for engineering knowledge graph maintenance in dynamic environment.

Description

Technical Field

[0001] This invention belongs to the field of computer and artificial intelligence technology, specifically relating to an error detection method based on temporal knowledge graphs, and more particularly to a dual-path interpretable temporal knowledge graph error detection method that integrates representation learning and symbolic logic reasoning, used to achieve accurate identification and data cleaning of factual errors and outdated facts in temporal knowledge graphs. Background Technology

[0002] Knowledge graphs explicitly encode facts of the real world in the form of triples[1],[2],[3], providing a foundation for large-scale knowledge representation and reasoning, and have become the core infrastructure of intelligent question answering, recommendation systems and engineering information systems[4],[5]. With the rapid expansion of real-world knowledge, most temporal knowledge graphs rely on automated pipeline construction[6], which expands triples into quadruples

[11] by introducing a timestamp dimension to record the dynamic evolution of facts. However, this automated construction process inevitably introduces complex noise[7], and the introduction of temporal features makes error detection tasks more challenging than in static environments. Accurate automatic error detection has become a key prerequisite for building reliable dynamic knowledge-driven systems[8].

[0003] In the context of temporal knowledge graphs

[11] , erroneous facts exhibit significant heterogeneity and concealment. They include not only factual errors that violate hard logical constraints, but also a large number of outdated facts that were true at historical points in time but have become invalid as the entity's state evolves. Most existing temporal knowledge graph analysis methods are designed for link prediction tasks [9] and

[10] , and have obvious limitations when dealing with error detection tasks. On the one hand, models based on representation learning

[14] and

[15] have overgeneralization problems when capturing temporal evolution, tending to assign high confidence scores to semantically reasonable but temporally invalid quadruples, resulting in low accuracy in identifying anomalies, and their implicit vector representation cannot provide interpretable basis for decision-making.

[0004] On the other hand, logic rule-based models

[12] ,

[13] , and

[17] face serious challenges of rule sparsity when dealing with temporal evolution facts. Although symbolic logic has good interpretability, due to the discrete nature of logic rules, it is difficult to cover the complex long-tail dependencies in the temporal graph, and it is also unable to effectively model the dynamic features of the continuous change of entity states over time, resulting in a severe limitation on the recall rate during the detection process. In addition, existing temporal reasoning mechanisms often struggle to balance the relationship between the global historical background and local recent changes, and cannot accurately capture time-sensitive relational semantics.

[0005] Therefore, in the dynamic environment of temporal knowledge graphs, existing technologies struggle to simultaneously address both potential temporal semantic evolution and explicit temporal logical constraints. Faced with complex scenarios interwoven with heterogeneous noise, there is a lack of interpretable temporal knowledge graph error detection methods that can accurately distinguish between factual errors and outdated facts while overcoming rule sparsity and providing logical support. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes an error detection method for temporal knowledge graphs that integrates representation learning and logical rules. This method comprises a rule module and an embedding module. Specifically, the rule module aims to mine temporal logical rules through symbolic reasoning and check explicit logical consistency, providing interpretable evidence for error detection; the embedding module models the evolution of temporal semantics using a dual-path architecture, capturing time-aware features by utilizing global historical context and local recent transitions; finally, a weighted linear fusion strategy is employed to integrate the scores from the rule module and the embedding module, generating the final validity assessment result. The detailed construction steps of this invention are as follows:

[0007] 1) Extract time-series inference paths from historical data through time-filtered weighted sampling;

[0008] 2) Abstract the path into temporal logic rules and calculate the rule confidence to filter high-quality rule sets;

[0009] 3) Calculate the logical confidence score of the rule prediction based on the historical path instances matched by the query quadruple;

[0010] 4) Detect whether the query quadruple violates the logical mutual exclusion constraint, and use it as negative evidence to correct the logical score;

[0011] 5) Aggregate historical snapshot information using global path branches to learn the long-term attribute representation of entities;

[0012] 6) Utilize local path branches to model recent dynamic evolution and update the time-aware representation of relationships;

[0013] 7) Integrate global and local features using an adaptive gating fusion mechanism, and calculate the neural plausibility score through a decoder;

[0014] 8) Integrate the rule confidence score and neural rationality score using a weighted linear fusion strategy to obtain the final evaluation result;

[0015] Steps 1) to 4) are the rule module, steps 5) to 7) are the embedding module, and step 8) is the process of merging the rule module and the embedding module.

[0016] The main contributions of the temporal knowledge graph error detection method of this invention are as follows:

[0017] 1. To achieve a comprehensive evaluation of temporal facts, this invention designs and proposes a novel dual-path cognitive framework that evaluates query quadruples through two parallel components. The embedding path, based on deep representation learning, is used to capture long-term and short-term temporal dynamic features, while the symbolic rule-based path is used to verify hard logical constraints.

[0018] 2. A complementary weighted fusion strategy is implemented and introduced into the model to effectively integrate implicit semantic signals with explicit logical evidence. This strategy enables the model to effectively detect and distinguish different types of anomalies in temporal knowledge graphs, including factual errors and outdated facts, based on evolutionary features.

[0019] 3. Comprehensive experiments were conducted on five publicly available time-series knowledge graph benchmark datasets. The results show that the proposed method consistently outperforms existing state-of-the-art benchmark methods in both precision and recall for error detection, fully demonstrating the robustness of the designed dual-path architecture. Attached Figure Description

[0020] Figure 1 This is an overall framework diagram of the method proposed in this invention.

[0021] Figures 2(a), 2(b), 2(c), 2(d), 2(e), and 2(f) show the ablation experimental results of the rule module and the embedding module on the ICEWS14 and ICEWS18 datasets, respectively.

[0022] Figures 3(a), 3(b), 3(c), 3(d), 3(e), and 3(f) show the ablation experimental results with different noise levels added to the ICEWS14 and ICEWS18 datasets, respectively. Detailed Implementation

[0023] The following section further explains this scheme with reference to the accompanying drawings and specific implementation methods. The first part summarizes the entire framework method; the second part details the rule module and the embedding module, and explains the training process in detail; the third part covers the experimental details, including the benchmark dataset used, the initialization of experimental parameters, the presentation of results, ablation experiments, and the visualization and analysis of results; the fourth part provides a summary of the findings and contributions of this work.

[0024] 1. Overview of the Scheme

[0025] Quadruples are facts (knowledge) in a temporal knowledge graph. Each quadruple (s, r, o, t) represents the existence of a relation r between the head entity s and the tail entity o at timestamp t, where E is the set of all entities, R is the set of relations, and T is the set of timestamps. In this invention, to address the heterogeneous noise present in temporal knowledge graphs, a novel error detection framework called Dual-view Understanding and Evolution over Time (DUET) is proposed. This framework aims to evaluate the validity score of each temporal quadruple, thereby accurately distinguishing between factual errors and outdated facts.

[0026] Figure 1 The overall framework of the proposed method is shown, including a rule module and an embedding module. The specific steps are as follows: (1) The rule module uses a time-filtered weighted sampling algorithm to mine symbolic logic rules with temporal constraints from the graph based on the time influence of historical facts, and calculates their rule confidence; (2) The embedding module introduces a time-aware relation gating mechanism, dynamically adjusts the semantic representation of relations under different time spans by calculating the time interval vector, so as to capture the time sensitivity of relations; (3) For the quadruple to be tested, the rule module searches for path instances that satisfy the main body of the logical rules in the historical subgraph, and calculates the symbolic logic confidence score by combining rule support and logical mutual exclusion conflict detection; (4) The embedding module aggregates the static structural information of historical snapshots through global branches and models the recent evolution state by combining local branches. After integrating the dual-branch features by using an adaptive gating fusion mechanism, the decoder calculates the neural rationality score; (5) The logical confidence score generated by the rule module and the neural rationality score generated by the embedding module are integrated into the final unified validity assessment score by a weighted linear fusion strategy.

[0027] The rule module of this invention focuses on providing traceable evidence for error detection using explicit logical constraints, enhancing the interpretability of the results; while the embedding module focuses on capturing implicit temporal semantic evolution patterns from global history and local evolution. The fusion reasoning process of this invention effectively combines the rigor of symbolic reasoning with the generalization ability of representation learning, ensuring robustness to heterogeneous noise detection in dynamic environments.

[0028] 2. System Model

[0029] 2.1 Rules Module

[0030] The rules module consists of two components: rule mining and rule application. Rule mining extracts logical rules with temporal constraints from historical data using a time-filter-based weighted sampling method and calculates their confidence scores. Rule application retrieves path instances that satisfy the rules from the graph and, combined with logical mutual exclusion constraint detection, generates interpretable logical confidence scores for the quadruples to be tested.

[0031] 2.1.1 Rule Mining

[0032] To maintain computational efficiency while ensuring logical consistency on large-scale temporal knowledge graphs, a rule mining strategy based on temporal filtering and weighted random walks is adopted. This sampling mechanism ensures efficient discovery of complex temporal dependencies by prioritizing path traversal based on time relevance. The process effectively constructs high-confidence logical paths by identifying the historical preconditions of specific queries, thus serving as an effective criterion for anomaly detection.

[0033] The sampling process begins with selecting target quadruplets (s, r, o, t) from the training set. q To find the reasoning path, the algorithm searches for historical interactions pointing to entity s. Let a specific historical quadruple be represented as (e i r i ,s,t i ), where e i It is the precursor entity, r i It is a historical relationship, and t i <t q Timestamp. To quantify the temporal impact of this interaction, an unnormalized relevance weight w(t) is defined using an exponentially decaying function. i The specific calculation formula is as follows:

[0034] w(t i )=exp(-λ(t q -t i )), (1)

[0035] Where λ represents the decay factor. The second step involves calculating the selection of the specific historical edge (e) by normalizing the correlation weights. i r i ,s,t i The transition probability of (). The specific calculation formula is as follows:

[0036]

[0037] in, This represents the set of all historical edges pointing to entity s. The third step is to abstract the sampled path instances into general logical rules. In this abstraction stage, concrete entities are replaced with variables, while strictly preserving the relation sequence and temporal order. The generated temporal Horn rule is formally expressed as:

[0038]

[0039] Where, r i Represents the relationship of the i-th jump, a i and b i Represents entity variable, t i This represents the corresponding timestamp, and is constrained by t1≤…≤t k <t q Finally, to evaluate the predictive quality of the mining patterns, a confidence score is calculated for each candidate rule ρ. The formula for this metric is as follows:

[0040]

[0041] in, Let |·| represent a temporal knowledge graph, and let |·| represent the cardinality of the set.

[0042] 2.1.2 Rule Application

[0043] To efficiently verify the validity of query quadruples using the mined logic, a sparse matrix representation strategy is employed to optimize the rule application process. This method transforms the reasoning process into efficient matrix operations, aiming to reduce computational complexity while strictly preserving logical constraints. Specifically, for each relation r in the knowledge graph, a sparse matrix M is constructed. r Used to store all historical fact instances:

[0044]

[0045] Where N represents the total number of historical facts involving relation r. Given a mined temporal Horn rule, the application process is performed by iteratively filtering and concatenating these matrices to find relational paths leading to query subject s and object o. For the first atom, an initial candidate set... as follows:

[0046]

[0047] in, The matrix represents the historical facts of relation r1. The index symbol [:, k] refers to the k-th column, where columns 0, 1, and 2 correspond to the subject, object, and timestamp, respectively.

[0048] For subsequent step i, the candidate path set is recursively updated. This involves updating the current path set. The matrix is ​​concatenated with the next relation M. The valid path is expanded only if the tail entity of the previous step matches the head entity of the current step and satisfies the time order constraint. The recursive calculation formula is as follows:

[0049]

[0050] After completing the path retrieval process, the system uses the noisy-or

[13] strategy to synthesize these signals into a unified support score S. supp (q):

[0051]

[0052] in, Let c(ρ) represent the set of rules mined, and c(ρ) represent the confidence level of rule ρ. Furthermore, explicit verification of logical contradictions is performed based on mutual exclusion relationships. Calculate the conflict score S conf (q):

[0053]

[0054] Where φ represents the confidence level of the mutual exclusion constraint. Finally, to generate a unified logical confidence score S... rule (q) integrates the support signal and the conflict signal as follows:

[0055] S rule (q)=σ(S supp (q)-β·S conf (q)), (10)

[0056] Where β is the penalty coefficient and σ(·) is the sigmoid function. rule (q) is the output of the rule pattern.

[0057] 2.2 Embedded Module

[0058] The embedding module consists of three collaborative components: a global historical path, a local evolutionary path, and a multi-path fusion mechanism. The global historical path captures the global structural representation of entities by aggregating long-term historical snapshot information; the local evolutionary path focuses on the recent state transitions of entities to capture their dynamic evolutionary features; and the multi-path fusion mechanism uses adaptive gating to dynamically integrate dual-path features, and the decoder calculates a neural rationality score that reflects the reasonableness of the facts in the continuous semantic space.

[0059] 2.2.1 Global History Path

[0060] To capture persistent structural patterns, the global history path constructs a global representation by aggregating historical snapshots. Let... Indicates the query time tq The union of all previous historical edges. An L-layer relational graph convolutional network

[18] is used to encode these static topological attributes. Entity e i The update rule at level l+1 is formally expressed as:

[0061]

[0062] in, Let represent the initial static embedding, σ represent the nonlinear activation function, and In the global graph The set of neighbor indices of entity i under relation r. and These are the learnable relation-specific transformation matrix and the self-loop transformation matrix for layer l, respectively, and c i,i Used as a normalization constant. Final output As a time-invariant structural basis, it captures the long-term stable characteristics of entities in knowledge graphs.

[0063] 2.2.2 Local Evolution Path

[0064] Complementing the global path, the local path focuses on modeling sequence evolution using a hierarchical cyclic mechanism. At each time step τ, a snapshot G is processed. τ This invention uses a relational graph convolutional network to extract local structural features. The invention employs a relational graph convolutional network to generate structural features x at a specific time point. i，τ Specifically, embedding static entities into E static As input features, entity e i The aggregation operation is calculated as follows:

[0065]

[0066] in, Represents the neighbor index set, c i，r It is a normalization constant. In order to model the continuous state changes of entities, this invention uses a gated recurrent unit

[19] to manage the sequence evolution. First, the update gate z is calculated. i，τ and reset door n i，τ To regulate information flow:

[0067]

[0068] Among them, W {z，n} U {z，n} and b {z，n} These are learnable parameters, and σ represents the Sigmoid activation function. Next, the reset gate n is used... i，τ Calculate candidate entity embeddings that reflect the historical context filtered by the reset mechanism.

[0069]

[0070] Among them, W e U c and b e These are learnable parameters, and tanh represents the hyperbolic tangent activation function. The final evolved entity embedding... Update using linear interpolation:

[0071]

[0072] Among them, z i，τ This represents the update gate. This mechanism enables the model to adaptively determine how much historical information to retain and how much new temporal information to accept.

[0073] Meanwhile, to address the semantic drift problem of relations, this invention introduces a time-gated relation evolution mechanism. First, based on the current context representation... Calculate the dynamic gating vector g τ :

[0074]

[0075] Among them, W gate and b gate Let represent the learnable weight matrix and bias vector, respectively, and σ(·) denote the Sigmoid activation function. The final time-aware relation embedding r τ Synthesis through dynamic balancing of temporal context and static semantic information:

[0076]

[0077] in, Representing the temporal context, E r Let r represent the initial static embedding of relation r, where 1 represents an all-one vector and ⊙ represents element-wise multiplication. This gating mechanism allows the model to selectively preserve static semantics while incorporating dynamic evolution.

[0078] 2.2.3 Multi-path fusion mechanism

[0079] To effectively integrate information from both global historical paths and local evolutionary paths, this invention employs an adaptive gating fusion mechanism. Specifically, it includes the following steps:

[0080] First, the learnable gating vector α is computed based on the concatenated context. i :

[0081]

[0082] Among them, W fusion and bfusion Let represent the learnable transformation matrix and bias vector, respectively; [·; ·] denote the concatenation operation; and σ(·) be the sigmoid activation function. Next, the final entity embedding is generated through linear interpolation.

[0083]

[0084] Here, ⊙ represents element-wise multiplication. This mechanism allows the model to dynamically weigh the importance of long-term global history against short-term local evolution.

[0085] Finally, the ConvTransE decoder is used to evaluate the query quadruple q = (s, r, o, t) q The validity of the embedding likelihood score S is calculated. cmb (q) is as follows:

[0086]

[0087] in, and This represents the subject embedding and relation embedding reshaped into a two-dimensional (2D) format. During the model training phase, to train the aforementioned framework for error detection, this invention uses a joint loss function that includes interval ranking loss. Contrastive negative samples are constructed by simulating factual errors and outdated facts. The corresponding optimization objective is defined as:

[0088]

[0089] Subsequently, multi-path comparison loss was employed. This aims to enforce semantic consistency between global historical paths and local evolutionary paths. The objective is to minimize contrastive loss by maximizing the similarity of the same entity's representation across different perspectives while increasing its distance from other entities.

[0090]

[0091] Where ε represents the entity set, and sim(·,·) represents the cosine similarity function used to measure time-invariant global embeddings. and time-sensitive local embedding The alignment degree between them; ψ is a temperature hyperparameter controlling the smoothness of the probability distribution. The final optimization objective is constructed as a direct arithmetic sum of the two components mentioned above to facilitate the unweighted joint training process:

[0092]

[0093] in, It is the combined loss function used for backpropagation, representing the sum of error detection capability and representation consistency from different perspectives.

[0094] 2.3 Score Fusion Strategy

[0095] The final stage of the framework involves integrating the outputs of the rule pattern and the embedding pattern to produce the query quadruple q = (s, r, o, t). q The overall confidence score is calculated. The specific fusion steps are as follows:

[0096] First, the rule pattern provides a uniform logical confidence score S. τulc (q), this score aggregates sparse signals from supporting paths and logical mutual exclusion constraints. Simultaneously, the embedding pattern is generated by the decoder into a probability distribution P(o|s, r, t) on the candidate entities. q The specific probability associated with the target tail entity o will be used as the embedding rationality score S. emb (q) is used to capture the underlying statistical regularities of the data.

[0097] Finally, by employing a weighted linear fusion strategy, the statistical inference of embedded patterns is combined with the logical derivation of rule patterns to calculate the overall confidence level:

[0098] C(q)=μ·S ernb (q)+(1-μ)·S rule (q), (25)

[0099] Where μ∈[0,1] is a hyperparameter used to balance the contributions of statistical patterns and logical constraints, and C(q) represents the final rationality of the quadruple. Through this joint scoring mechanism, the final evaluation of the validity of the quadruple simultaneously considers the latent semantic features and explicit logical consistency. The overall execution flow of the DUET framework proposed in this invention can be implemented through the following algorithmic steps:

[0100] Algorithm 1: A Temporal Knowledge Graph Training and Error Detection Method Based on the DUET Framework

[0101]

[0102]

[0103] 3 Experimental Design

[0104] 3.1 Dataset

[0105] In the experiments of the DUET temporal knowledge graph error detection method, five widely used public benchmark datasets were introduced for evaluation. These datasets include ICEWS14, ICEWS18, ICEWS05-15

[20] , GDELT

[21] , and YAGO

[22] . The ICEWS series datasets contain international political events with fine-grained daily timestamps; among them, ICEWS14 and ICEWS18 occurred in 2014 and 2018, respectively, and ICEWS05-15 covers events from 2005 to 2015. The GDELT dataset is constructed based on global news reports and provides broader but noisier event coverage. The YAGO dataset is derived from Wikipedia and WordNet and focuses on factual assertions with annual timestamps, showing different temporal characteristics compared to event-centric datasets. Table 1 provides statistical information for these datasets.

[0106] Table 1: Datasets Used

[0107]

[0108] 3.2 Baseline Model

[0109] Since there is a lack of dedicated methods for anomaly detection in temporal knowledge graphs, this invention follows the common practice of reusing link prediction models as anomaly detectors. Specifically, the proposed method is evaluated against a variety of state-of-the-art inference models, which are tuned to fit the anomaly detection setting by treating their plausibility scores as inverse indicators of normality. The selected benchmark models include the following:

[0110] TTransE

[14] extends the static TransE model by introducing a time projection matrix to capture the directionality and ordering of temporal facts. The negative value of its scoring function is used as anomaly score.

[0111] DE-TransE, DE-DistMult, and DE-SimplE

[15] belong to the diachronic embedding framework in which each entity representation is parameterized as a time-dependent function. These models are adapted for anomaly detection tasks by interpreting their plausibility scores as a measure of normality.

[0112] RE-NET

[16] : Models the temporal evolution of entities using a recurrent event network. The predicted probability of the triple observed at time t is used as the normality score.

[0113] TLogic

[13] : Derives temporal logic rules from historical facts. If a triple violates a high-confidence temporal rule or cannot be explained by the learned rules, the triple is considered anomalous.

[0114] LogCL

[23] : learns temporal representations by contrastive learning that perceives local and global history. It is adapted to anomaly detection tasks by using the contrastive loss of a given event as its anomaly score.

[0115] This series of benchmark models covers geometric translation, diachronic embedding, recurrent dynamics, symbolic rule learning, and contrastive representation learning, thus providing a comprehensive benchmark for anomaly detection in temporal knowledge graphs.

[0116] 3.3 Model Evaluation

[0117] To comprehensively evaluate the performance of the DUET framework and benchmark models in error detection of temporal knowledge graphs, this invention uses three standard evaluation metrics. Specifically, precision, recall, and F1 score are employed because they are widely used in knowledge graph quality evaluation. This multi-layered evaluation approach provides a balanced assessment of the model's error detection capabilities across various datasets.

[0118] Precision: Evaluates the accuracy of positive predictions by quantifying the proportion of truly incorrect errors among those identified. High precision indicates that the model has successfully minimized false positives that incorrectly label correct quadruples as anomalies. This feature is particularly important in applications requiring high-confidence detection.

[0119] Recall: Evaluates the model's ability to identify all actual errors present in the dataset. It effectively measures erroneous quadruplets that the framework might have failed to detect. Achieving high recall ensures comprehensive coverage, which is crucial for maintaining the overall reliability of temporal knowledge graphs.

[0120] The F1 score, as the harmonic mean of precision and recall, provides a single, balanced performance metric. It is particularly effective for evaluating detection results in scenarios where the distribution of correct and incorrect facts spans different time periods and relation types, and where data imbalance exists.

[0121] 3.4 Results

[0122] Tables 2 and 3 show the detection performance of the present invention and the benchmark model on five datasets. Table 2 includes the datasets ICEWS14, ICEWS18, and ICEWS05-15, while Table 3 includes the datasets GDELT and YAGO. The data in the tables present the performance metrics parameters under different Top-k thresholds, and all values ​​are presented as percentages. The bolded values ​​represent the best results under the corresponding evaluation conditions.

[0123] Table 2: Experimental results on the ICEWS14 and ICEWS18 datasets

[0124]

[0125]

[0126] Table 3: Experimental results on the ICEWS0515 and WIKI datasets

[0127]

[0128]

[0129] Tables 2 and 3 present the comparative evaluation results of the proposed DUET framework and representative benchmark methods on five benchmark datasets. Overall, the proposed model demonstrates a consistent and significant advantage on most evaluation metrics, particularly at the most stringent thresholds. As shown in Table 2, on the ICEWS14 dataset, the precision at the Top-1% threshold reaches 83.33%, significantly outperforming the second-best performing model, LogCL (65.72%). Similarly, on the fact-centered YAGO dataset shown in Table 3, the proposed model achieves a precision of 90.05%, outperforming purely embedding-based benchmark models (such as DE-TransE with 87.45% precision) and rule-based benchmark models (such as TLogic with 69.40% precision). These results demonstrate that by combining rule-based symbolic reasoning with embedding-based neural modeling, the proposed model effectively filters logically illogical errors, thereby significantly improving detection accuracy. Furthermore, the proposed model exhibits excellent robustness in balancing precision and recall, achieving the highest F1 score in all experimental settings.

[0130] On the large and noisy GDELT dataset, the limitations of purely rule-based methods are evident; TLogic achieves only 9.19% recall at the 1% threshold due to the sparsity of mineable rules. In contrast, the model of this invention effectively complements it with embedded signals, achieving a recall of 17.22% and an F1 score of 28.42%, surpassing the second-ranked LogCL (with an F1 score of 26.81%). This finding highlights the framework's ability to generalize beyond the coverage of explicit rules while maintaining logical consistency. Furthermore, on the longer-spanning ICEWS05-15 dataset, this invention achieves best results across all metrics at the 1% threshold, with a precision of 80.18%, a recall of 16.87%, and an F1 score of 27.87%. Compared to the top-performing benchmark model LogCL (with an F1 score of 26.03%), this result demonstrates the superior ability of this invention to capture complex temporal dependencies over long time periods.

[0131] Despite maintaining strong overall performance, the performance gap on certain metrics narrowed due to the competitive nature of the embedding-based benchmark models in dense event environments. Notably, on the ICEWS18 dataset, while the model of this invention significantly outperformed LogCL (85.67%) at a precision of 1% threshold (91.32%), their recall at the same threshold was highly similar, with the model of this invention at 18.56% and LogCL at 18.57%. This phenomenon indicates that for certain dense event distributions, the embedding components of the benchmark models are very effective in retrieving candidate objects, even with lower precision. A similar trend was observed on the GDELT dataset, where the precision of this invention at a 1% threshold was 81.36%, slightly higher than LogCL's 80.69%. This closeness is attributed to the benchmark model's use of contrastive learning to handle high noise levels. Nevertheless, when the threshold was relaxed to 5%, the model of this invention still maintained a significant advantage, consistently observing a higher F1 score compared to all competitors. This fully validates that the dual-path mechanism of this invention can effectively identify obvious anomalies as well as more subtle inconsistencies.

[0132] 3.5 Ablation Experiment

[0133] To further illustrate the beneficial effects of this invention, the internal mechanisms and robustness of the proposed framework are analyzed in depth. Specifically, this invention first verifies the complementarity between the embedding pattern and the rule pattern, then examines the model's sensitivity to the proportion of outliers, and finally evaluates the model's comprehensive ability to effectively distinguish between factual errors and outdated facts.

[0134] 3.5.1 Pattern Complementarity

[0135] To verify the necessity of the proposed dual-path architecture, this invention investigates the complementarity between data-driven embedding patterns and symbolic rule patterns. For example... Figures 2(a) to 2(f) As shown, the performance comparisons involve the full DUET framework and two ablation variants on the ICEWS14 and ICEWS18 datasets. In these visualizations, red bars represent the full DUET framework, blue bars represent the performance of the variant that only utilizes regular patterns for error detection, and yellow bars correspond to the variant that only relies on embedding patterns. The comparison results show that removing any component leads to a decrease in overall detection quality, although this impact varies across different evaluation metrics.

[0136] The variants of the embedding pattern (represented by the yellow bars) exhibit a significant drop in precision. This performance degradation suggests that while the embeddings effectively capture the underlying data distribution, the lack of strict logical constraints leads to a higher false positive rate. For example, as shown in Figure 2(a), on the ICEWS14 dataset, at the 1% threshold, precision drops from 83.3% of the full model to 69.8% when the rule module is removed. Similarly, as shown in Figure 2(d), on the ICEWS18 dataset, precision drops from 91.3% to 87.6%. This observation suggests that individual embedding modules are prone to verifying semantically plausible but historically or logically incorrect facts as correct.

[0137] Conversely, relying solely on rule-based variants (represented by the blue bars) results in a significant drop in recall and F1 score primarily due to the inherent sparsity of symbolic rules. As observed in Figures 2(e) and 2(f) at the 1% threshold for the ICEWS18 dataset, recall (as shown in Figure 2(e)) plummets from 18.6% for the full model to 7.4% when relying solely on rules, causing the F1 score (as shown in Figure 2(f)) to drop from 30.9% to 13.2%. This structural limitation confirms that rule mining cannot cover all complex temporal dependencies, and the lack of embedding-based generalization prevents the model from evaluating quadruplets that do not trigger specific rules. By effectively combining these advantages, the full DUET framework achieves optimal performance, where the embedding path ensures broad coverage to maintain high recall, while the rule path provides rigorous logical validation to ensure high precision.

[0138] 3.5.2 Sensitivity to Abnormal Proportions

[0139] To evaluate the robustness of the proposed framework under different levels of data corruption, sensitivity analyses were performed on the ICEWS14 and ICEWS18 datasets. Figures 3(a) to 3(f)As shown in the visualizations, this invention constructed three different test scenarios by injecting synthesis errors at ratios of 5%, 10%, and 15%, respectively. In these graphs, the blue line represents the model's performance at a 5% noise level, while the orange and green lines correspond to noise levels of 10% and 15%, respectively. This comparison highlights how anomaly density affects the model's detection capability at different Top-k thresholds.

[0140] First, as shown in Figures 3(a) and 3(d), a counterintuitive trade-off pattern can be observed: as the error rate increases, the model's accuracy at a strict threshold shows a significant improvement. This performance trend is attributed to the unique nature of synthetic errors. Unlike subtle fact distortions, randomly injected anomalies often blatantly violate global time rules. Therefore, in noisy scenarios, the dataset is saturated with these easily distinguishable anomalies, allowing the model to prioritize obvious errors and fill in high-confidence anomaly predictions in the top-ranked detection results.

[0141] Secondly, a key observation regarding the F1 score trend is that, as shown in Figures 3(c) and 3(f), as the threshold k increases, the model's performance at 5% noise level is eventually surpassed by noise levels of 10% and 15%. This crossover is driven by the rapid depletion of the "finite pool" of identifiable errors in low-noise environments, forcing the model to subsequently incorporate false positives, thus strictly limiting the potential growth of the F1 score. In contrast, high-noise scenarios provide a much larger pool of distinguishable anomalies, allowing the framework to maintain stable precision over a wider inspection window. This result strongly demonstrates that the proposed model can balance precision and recall extremely effectively when a sufficient number of identifiable conflicts exist in the historical context.

[0142] 3.5.3 Error Type Analysis

[0143] Table 4: Differences in the performance of factual errors and outdated facts at different thresholds

[0144]

[0145] Table 4 evaluates the detection performance of the DUET framework for two error categories: factual errors and outdated facts. As the experimental results show, significant performance differences were observed between these two anomaly types across all top-k thresholds on the ICEWS14 and ICEWS18 datasets. Specifically, on the ICEWS18 dataset, when k = 1%, the precision for factual errors reached 95.64%, while the precision for outdated facts was only 36.23%. This performance gap is primarily attributed to the inherent nature of the anomalies: factual errors typically involve triples that violate static semantic patterns or high-confidence logical rules, which are more easily identified by the multi-path embedding and rule aggregation modules of this invention. Conversely, outdated facts present a significant challenge because they still possess semantic plausibility within the global historical context; therefore, their detection requires the model to accurately track temporal state transitions at specific timestamps and the lifecycle changes of entities.

[0146] Compared to the LogCL benchmark model, the DUET framework demonstrates consistent improvements in both error categories. For factual errors, DUET achieves a higher F1 score than LogCL (e.g., on the ICEWS18 dataset, the F1 score is 62.01% when k = 5%, compared to 52.92% for LogCL), indicating that the invention more effectively integrates symbolic and statistical signals. Furthermore, in the more challenging task of detecting outdated facts, DUET consistently maintains a stable lead over the benchmark model. For example, on the ICEWS14 dataset, the F1 score improves from 13.04% to 16.36% when k = 1%. This enhancement is primarily attributed to the time-aware relational gating mechanism in this invention, which effectively prevents the framework from accepting outdated information by monitoring temporal consistency. These results confirm that, through explicit logical consistency and implicit temporal regularity in joint inference, the proposed framework achieves robust error detection even when dealing with subtle temporal variations that the benchmark model cannot capture.

[0147] 4. Conclusion

[0148] This invention proposes a method and system for error detection in temporal knowledge graphs that integrates time-aware embedding representation and symbolic temporal logic reasoning. By employing an innovative dual-path architecture, this invention effectively utilizes the generalization ability of embedding patterns to accurately capture implicit temporal evolution, while leveraging the interpretability of rule patterns to rigorously verify explicit logical consistency. Furthermore, this invention introduces a time-aware relation gating mechanism and a complementary weighted fusion strategy, effectively overcoming the technical challenge of distinguishing between "factual errors" and "outdated facts" in existing technologies. Extensive dataset validation demonstrates that this invention significantly outperforms existing single embedding or single rule models in terms of anomaly detection accuracy and noise robustness. This invention fully leverages the synergistic advantages of neural embedding and symbolic logic, providing a robust and highly interpretable solution for the construction, quality assessment, and automated maintenance of temporal knowledge graphs in dynamic environments.

[0149] References:

[0150] [1] A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, O. Yakhnenko, Translating embeddings for modeling multi-relational data, Advances in neural information processing systems 26 (2013).

[0151] [2] Z.Wang, J.Zhang, J.Feng, Z.Chen, Knowledge graph embedding by translating on hyperplanes, in: Proceedings of the AAAI conference on artificialintelligence, Vol.28, 2014.

[0152] [3] Q.Wang, Z.Mao, B.Wang, LGuo, Knowledge graph embedding: A survey of approaches and applications, IEEE transactions on knowledge and data engineering 29(12)(2017)2724-2743.

[0153] [4]H.Cheng,K.Wang,X.Tan,A link prediction method for chinese financial event knowledge graph based on graph attention networks andconvolutional neural networks.Engineering Applications of ArtificialIntelligence 138(2024)109361.

[0154] [5]Y.Liu,Z.Liu,Z.Yu,Q.Fu,W.Tang,W.Yao,Z.Sun,Knowledge augmentedexpert finding framework via knowledge graph embedding for community questionanswering,Engineering Applications ofArtificial Intelligence 142(2025)109891.

[0155] [6]X.Dong,E.Gabrilovich,G.Heitz,W.Horn,N.Lao,K.Murphy,T.Strohmann,S.Sun,W.Zhang,Knowledge vault:A web-scale ap proach to probabilisticknowledge fusion,in:Proceedings ofthe 20th ACM SIGKDD internationalconference on Knowledge discovery and data mining,2014,pp.601-610.

[0156] [7]S.Heindorf, M.Potthast,B.Stein,G.Engels,Vandalism detection inwikidata,in:Proceedings of the 25th ACM International on Conference onInformation and Knowledge Management,2016,pp.327-336.

[0157] [8]H.Paulheim,Knowledge graph refinement:A survey of approaches andevaluation methods,Semantic web 8(3)(2016)489-508.

[0158] [9]R.Xie,Z.Liu,J.Jia,H.Luan,M.Sun,Representation learning ofknowledge graphs with entity descriptions,in:Proceedings of the AAAIconference on artificial intelligence,Vol.30,2016.

[0159]

[10] X.Chen,M.Chen,W.Shi,Y.Sun,C.Zaniolo,Embedding uncertain knowledgegraphs,in:Proceedings of the AAAI conferencc on artificial intelligence,Vol.33,2019,pp.3363-3370.

[0160]

[11] T.Jiang,T.Liu,T.Ge,L.Sha,B.Chang,S.Li,Z.Sui,Towards time awareknowledge graph completion,in:Proceedings ofCOLING 2016,the 26thInternational Conference on Computational Linguistics,2016,pp.1715-1724.

[0161]

[12] L.A.Galárraga,C.Teflioudi,K.Hose,F.Suchanek,Amie:association rulemining under incomplete evidence in ontological knowledge bases,in:Proceedings of the 22nd international conference on World Wide Web,2013,pp.413--422.

[0162]

[13] Y.Liu,Y.Ma,M.Hildebrandt,M.Joblin,V.Tresp,Tlogic:Temporal logicalrules for explainable link forecasting on temporal knowledge graphs,in:Proceedings of the AAAI conference on artificial intelligence,Vol.36,2022,pp.4120--4127.

[0163]

[14] J.Leblay,M.W.Chekol,Deriving validity time in knowledge graph,in:Companion proceedings ofthe the web conference 2018,2018,pp.1771-1776.

[0164]

[15] R.Goel,S.M.Kazemi,M.Brubaker,P.Poupart,Diachronic embedding fortemporal knowledge graph completion,in:Proceedings ofthe AAAI conference onartificial intelligence,Vol.34,2020,pp.3988-3995.

[0165]

[16] W.Jin,M.Qu,X.Jin,X.Ren,Recurrent event network:Autoregressivestructure inferenceover temporal knowledge graphs,in:Proceedings ofthe2020conference on empirical methods in natural language processing,2020,pp.6669-6683.

[0166]

[17] Z.Li,X.Jin,W.Li,S.Guan,J.Guo,H.Shen,Y.Wang,X.Cheng,Temporalknowledge graph reasoning based on evolutional representation leaming,in:Proceedings of the 44th international ACM SIGIR conference on research anddevelopmennt in information retrieval,2021,pp.408-417.

[0167]

[18] M.Schlichtkrull,T.N.Kipf,P. Bloem,R.Van Den Berg,I.Titov,M.Welling,Modeling relational data with graph convolutional networks,in:European semantic web conference,Springer,2018,pp.593-607.

[0168]

[19] K.Cho,B.Van C.Gulcehre,D.Bahdanau,F.Bougares,H.Schwenk,Y.Bengio,Leaming phrase representations using rnn encoder-decoder forstatistical machine translation,in:Proceedings of the 2014Conference onEmpirical Methods in Natural Language Processing,2014,pp.1724-1734.

[0169]

[20] M.D.Ward,A.Beger,J.Cutler,M.Dickenson,C.Dorff,B.Radford,Comparinggdelt and icews event data,Analysis 21(1)(2013)267-297.

[0170]

[21] K.Leetaru,P.A.Schrodt,Gdelt:Global data on events,location,andtone,1979-2012,in:ISA annual convention,Vol.2,Citeseer,2013,pp.1-49.

[0171]

[22] F.M.Suchanek,G.Kasneci,G.Weikum,Yago:a core of semanticknowledge,in:Proceedings of the 16th international conference on World WideWeb,2007,pp.697-706.

[0172]

[23] W.Chen,H.Wan,Y.Wu,S.Zhao,J.Cheng,Y.Li,Y.Lin,Local-global history-aware contrastive learning for temporal knowledge graph reason ing,in:2024IEEE 40th International Conference on Data Engineering,IEEE,2024,pp.733-746。

Claims

1. A temporal knowledge graph error detection method integrating representation learning and logical rules. First, representation learning is used to capture the implicit temporal evolution patterns of the temporal knowledge graph. Then, logical rules are used to verify explicit logical consistency, alleviating the problem of easily confusing factual errors with outdated facts and the difficulty in accurately distinguishing them in current temporal knowledge graph error detection tasks. The method includes the following steps: S1. A time-series inference path is formed by time-filtered weighted sampling on historical data. The path is then abstracted into time-series logical rules to calculate the rule confidence. Given a query quadruple, a sparse matrix is ​​introduced to match historical fact instances and combined with logical mutual exclusion constraints to obtain a unified logical confidence score for rule prediction. S2 constructs a global historical path and a local evolution path, respectively learning the long-term global static features and recent local dynamic features of entities. At the same time, a time-gating mechanism is introduced to update the relation representation. After fusing dual-view features, given a query quadruple, the neural rationality score is calculated through the decoder. S3 integrates the unified logical confidence score and the neural rationality score to form the final comprehensive evaluation result of error detection; Step S1 specifically includes: S11, given the target query quadruple q = (s, r, o, t) q ), find the historical interaction quadruple (e) pointing to subject s i r i ,s,t i ), where the timestamp satisfies t i <t q Unnormalized correlation weights are defined using an exponential decay function: w(t i )=exp(-λ(t q -t i )), where λ denotes an attenuation factor. Normalizing the weighted score gives the probability P((e i , r i , s, t i ) | (s, t q )) of selecting this particular historical edge, which is computed as: where, denotes the set of all historical edges pointing to the subject s. After sampling the path instances, they are abstracted to general temporal Horn rules p and the confidence conf(p) of each candidate rule is computed as follows: in, Represents a time-series knowledge graph. S12, construct a sparse matrix M for each relation r in the knowledge graph r For storing all historical fact instances to iteratively find the path. Adopt the noisy-or strategy to synthesize the retrieved path signals into a unified support score S supp (q): wherein, denotes the set of mined rules, c(p) denotes the confidence of rule p. Subsequently, the set Explicitly verify logical contradictions, compute conflict score S conf (q): where φ(r, r') represents the confidence of mutual exclusion constraint. Finally, the support signal and the conflict signal are integrated to generate the unified logical confidence score S rule (q): S rule (q) = σ(S supp (q) - β · S conf (q)), Where β is the penalty coefficient and σ(·) is the sigmoid function. Step S2 specifically includes: S21, aggregates all historical snapshot information prior to the query time through the global historical path, and uses a layered relational graph convolutional network to encode static topological attributes, entity e. i The representation at level l+1 is updated as follows: in, This represents the set of neighbor indices of an entity in the global graph under relation r. and It is a learnable transformation matrix, c i，r The normalization constant is used to obtain the global structural features in the final output. S22, at each time step τ, generates a structural feature x for a specific time through a local evolution path. i，τ The sequential evolution of entities is simulated using gated cyclic units, and the update gate z is calculated respectively. i，τ and reset door n i.τ Then calculate the candidate entity embedding. The evolved local structural features were obtained through linear interpolation. Simultaneously, a time-gated relation evolution mechanism is introduced, based on the current context representation. Calculate the dynamic gating vector Time-aware relationship embedding is synthesized by dynamically balancing temporal context and static semantic information. S23, compute adaptive gating vector based on context of stitching Generate final entity embeddings by linear interpolation Finally, evaluate the validity of the given query quadruple using the ConvTransE decoder, computing the neural plausibility score S emb (q): Step S3 specifically includes: S31, which integrates the unified logical confidence score and the neural rationality score, specifically: C(q) = μ - S emb (q) + (1 - μ) - S rule (q) Among them, S rule (q) is the unified logic confidence score extracted in step S1, S emb (q) is the neural plausibility score extracted in step S2, and the hyperparameter μ∈[0,1] is used to balance the contribution weights of statistical patterns and logical constraints. C(q) represents the overall confidence level of the quadruple. When C(q) is less than a preset threshold, the quadruple is judged to be abnormal and marked as an error or outdated fact.

2. Step S1 according to claim 1 is called the rule mode module, step S2 is called the embedding mode module, and step S3 is the final anomaly detection fusion judgment result of the present invention.

Images