Abstract:Recent studies have highlighted the effectiveness of tensor decomposition methods in the Temporal Knowledge Graphs Embedding (TKGE) task. However, we found that inherent heterogeneity among factor tensors in tensor decomposition significantly hinders the tensor fusion process and further limits the performance of link prediction. To overcome this limitation, we introduce a novel method that maps factor tensors onto a unified smooth Lie group manifold to make the distribution of factor tensors approximating homogeneous in tensor decomposition. We provide the theoretical proof of our motivation that homogeneous tensors are more effective than heterogeneous tensors in tensor fusion and approximating the target for tensor decomposition based TKGE methods. The proposed method can be directly integrated into existing tensor decomposition based TKGE methods without introducing extra parameters. Extensive experiments demonstrate the effectiveness of our method in mitigating the heterogeneity and in enhancing the tensor decomposition based TKGE models.

What problem does this paper attempt to address?

### Problems the Paper Aims to Solve This paper aims to address the performance limitations in tensor decomposition-based temporal knowledge graph embedding (TKGE) methods caused by the heterogeneity of factor tensors. Specifically, the paper points out that during the tensor decomposition process, the distribution differences among different factor tensors are significant. This heterogeneity hinders the tensor fusion process, thereby limiting the accuracy of link prediction. To overcome this limitation, the authors propose a novel method that maps factor tensors onto a unified smooth Lie group manifold, making the distribution of factor tensors more homogeneous. This approach can improve the performance of tensor decomposition-based TKGE models. ### Main Contributions 1. **First study on the negative impact of factor tensor heterogeneity on tensor decomposition-based TKGE models**: The authors demonstrate through theoretical analysis and experimental validation that homogeneous factor tensors are more effective in approximating the target tensor than heterogeneous factor tensors. 2. **Provide theoretical proof**: The authors provide theoretical proof showing that homogeneous factor tensors are more effective in tensor decomposition than heterogeneous factor tensors. 3. **Propose a method that does not require additional parameters**: This method can be directly integrated into existing tensor decomposition-based TKGE models without introducing additional parameters. Extensive experimental results show that this method significantly alleviates heterogeneity and improves model performance. ### Method Overview 1. **Lie Group Manifold Mapping**: The authors map factor tensors onto a Lie group manifold, utilizing the smoothness and uniformity of Lie groups to alleviate the heterogeneity of factor tensors. 2. **Logarithm Mapping Operation**: By introducing a logarithm mapping operation, rotation matrices are converted into skew-symmetric matrices, allowing optimization in Euclidean space. 3. **N3 Regularization**: By minimizing the difference between the original factor tensors and the mapped factor tensors, the method drives the factor tensors to become more homogeneous in Euclidean space. ### Experimental Results 1. **Performance Improvement**: Experimental results on the ICEWS14 and ICEWS05-15 datasets show that applying this method leads to significant performance improvements across various metrics for multiple tensor decomposition-based TKGE models. 2. **Quantitative Analysis**: By calculating the distance differences between factor tensors, the effectiveness of the method in alleviating heterogeneity is validated. 3. **Visualization Analysis**: Using t-SNE visualization techniques, the method's effectiveness in making factor embeddings more homogeneous is further demonstrated. ### Conclusion By introducing Lie group manifold mapping and logarithm mapping operations, this paper successfully alleviates the issue of factor tensor heterogeneity in tensor decomposition-based TKGE models, significantly enhancing the model's link prediction performance. This method is not only theoretically validated but also performs excellently in practical applications.

Mitigating Heterogeneity among Factor Tensors via Lie Group Manifolds for Tensor Decomposition Based Temporal Knowledge Graph Embedding

Tucker decomposition-based Temporal Knowledge Graph Completion

Tensor Decompositions for Temporal Knowledge Graph Completion with Time Perspective

Knowledge graph fact prediction via knowledge-enriched tensor factorization

Multi-Factor Spatio-Temporal Prediction based on Graph Decomposition Learning

The Generalized Tensor Decomposition with Heterogeneous Tensor Product for Third-Order Tensors

Time-Aware Tensor Decomposition for Missing Entry Prediction

Knowledge Graph Embedding with Diversity of Structures

Distributed Flexible Nonlinear Tensor Factorization

Temporal Knowledge Graph Reasoning with Low-rank and Model-agnostic Representations

Topology-aware Tensor Decomposition for Meta-graph Learning

Adaptive Regularizing Tucker Decomposition for Knowledge Graph Completion

Enabling Tensor Decomposition for Time-Series Classification via A Simple Pseudo-Laplacian Contrast

HyperNTF: A Hypergraph Regularized Nonnegative Tensor Factorization for Dimensionality Reduction

Latent Space Model for Higher-order Networks and Generalized Tensor Decomposition

A Novel Tensor Decomposition Networks for Recommender Systems via Adaptive Moment Estimation Method

Low-tubal-rank tensor factorization on constant curvature Riemann manifold with mixture of Gaussians

Decomposition of Big Tensors With Low Multilinear Rank

Temporal Knowledge Graph Embedding Via Sparse Transfer Matrix

Multi-View Learning a Decomposable Affinity Matrix via Tensor Self-Representation on Grassmann Manifold