Temporal Inductive Path Neural Network for Temporal Knowledge Graph Reasoning

TL;DR

This paper tackles extrapolated reasoning on Temporal Knowledge Graphs in an inductive setting where new entities continually emerge. It introduces Temporal Inductive Path Neural Network (TiPNN), which builds a unified History Temporal Graph to capture historical connectivity and temporal patterns, and learns query-aware temporal paths through an ω-layer path aggregation framework. The model combines query-conditioned static relation representations with time-encoded edge features, optimizes with a joint LKG loss plus a relation-orthogonality regularizer, and demonstrates state-of-the-art performance on ICEWS18, GDELT, WIKI, and YAGO3, along with strong inductive generalization and reasoning evidence. The results indicate TiPNN's practical impact for scalable, interpretable temporal reasoning in real-world dynamic knowledge graphs.

Abstract

Temporal Knowledge Graph (TKG) is an extension of traditional Knowledge Graph (KG) that incorporates the dimension of time. Reasoning on TKGs is a crucial task that aims to predict future facts based on historical occurrences. The key challenge lies in uncovering structural dependencies within historical subgraphs and temporal patterns. Most existing approaches model TKGs relying on entity modeling, as nodes in the graph play a crucial role in knowledge representation. However, the real-world scenario often involves an extensive number of entities, with new entities emerging over time. This makes it challenging for entity-dependent methods to cope with extensive volumes of entities, and effectively handling newly emerging entities also becomes a significant challenge. Therefore, we propose Temporal Inductive Path Neural Network (TiPNN), which models historical information in an entity-independent perspective. Specifically, TiPNN adopts a unified graph, namely history temporal graph, to comprehensively capture and encapsulate information from history. Subsequently, we utilize the defined query-aware temporal paths on a history temporal graph to model historical path information related to queries for reasoning. Extensive experiments illustrate that the proposed model not only attains significant performance enhancements but also handles inductive settings, while additionally facilitating the provision of reasoning evidence through history temporal graphs.

Figures (9)

• Figure 1: An example of inference based on temporal paths modeling. The red triangular nodes and red circular nodes represent the subject entity and candidate object entity of the query at timestamp $T+1$, respectively.
• Figure 2: Example of constructing the history temporal graph $\hat{G}_{t-2:t}$ with $m=3$. For illustrative purposes, we use an undirected graph to demonstrate the construction method and omit the relationship types of edges in the historical subgraph sequence and the constructed history temporal graph. The time labels attached to the edges in the history temporal graph represent the timestamps of the corresponding edges in the historical subgraph sequence.
• Figure 3: An illustrative diagram of the proposed TiPNN model for query-aware temporal path processing. For a given query $(s,r,?)$ at future timestamp, TiPNN engages in temporal path processing within the constructed history temporal graph $\hat{G}$ to perform prediction for the future timestamp. The temporal path feature $\textbf{H}_{\mathcal{V}}$ is iteratively learned and updated by the query-aware $\omega$-layers aggregation neural network. In each layer, the temporal edges in $\hat{G}$, enriched with temporal information, are individually modeled for basic static representation and temporal representation through $\bm{\Psi}_r$ and $\bm{\Upsilon}$, respectively. Subsequently, temporal edges are merged using $\textsc{Tmsg}(\cdot)$, followed by aggregation through $\textsc{Agg}(\cdot)$ to obtain the feature of the current layer's temporal path.
• Figure 4: Ablation Results on Temporal Encoding.
• Figure 5: Ablation Results on Temporal Edge Merging Method.

Theorems & Definitions (1)

• Definition 1