PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Knowledge Graph Reasoning on Graph Types: Static, Dynamic, and Multimodal

Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, Fuchun Sun

arXiv:2212.05767v7cs.AIcs.CLcs.IR

TL;DR

KGR addresses incomplete knowledge graphs by inferring new facts, but existing surveys did not comprehensively cover static, temporal, and multi-modal reasoning. This paper surveys these models using a bi-level taxonomy, summarizes performances and datasets, and identifies challenges and opportunities. It reports that extrapolation remains early at around 30% of temporal KGR models, while multi-modal KGR accounts for only 18% of models.

  • Problem

    Existing surveys largely omit recent temporal and multi-modal KGR progress, leaving no comprehensive review and open-source repository across graph types.

  • Method

    The paper systematically reviews KGR using graph types as the top-level taxonomy and fourteen techniques plus four reasoning scenarios as base-level categories.

  • Results

    Around 30% of temporal KGR models address extrapolation, while 18% of KGR models address multi-modal scenarios.

  • Takeaways & Limitations

    The survey provides a cross-graph reference covering models, performances, datasets, challenges, opportunities, and an open-source collection for the KGR community.

  • Takeaways & Limitations

    Embedding-based KGR models often have strong expressive ability but suffer from explainability, while rule- and path-based models are more explainable but computationally demanding and less expressive.

Abstract

from arXiv · show

Knowledge graph reasoning (KGR), aiming to deduce new facts from existing facts based on mined logic rules underlying knowledge graphs (KGs), has become a fast-growing research direction. It has been proven to significantly benefit the usage of KGs in many AI applications, such as question answering, recommendation systems, and etc. According to the graph types, existing KGR models can be roughly divided into three categories, i.e., static models, temporal models, and multi-modal models. Early works in this domain mainly focus on static KGR, and recent works try to leverage the temporal and multi-modal information, which are more practical and closer to real-world. However, no survey papers and open-source repositories comprehensively summarize and discuss models in this important direction. To fill the gap, we conduct a first survey for knowledge graph reasoning tracing from static to temporal and then to multi-modal KGs. Concretely, the models are reviewed based on bi-level taxonomy, i.e., top-level (graph types) and base-level (techniques and scenarios). Besides, the performances, as well as datasets, are summarized and presented. Moreover, we point out the challenges and potential opportunities to enlighten the readers. The corresponding open-source repository is shared on GitHub https://github.com/LIANGKE23/Awesome-Knowledge-Graph-Reasoning.

1 INTRODUCTION

The introduction frames KGR as a way to infer missing KG facts and surveys models across static, temporal, and multi-modal graph types. It organizes this literature by techniques and reasoning scenarios, alongside performance, datasets, challenges, and open resources.

  • KGR infers missing facts from existing KG facts by deriving underlying logic rules, supporting applications such as question answering and recommendation systems.
  • Static KGs contain uni-modal facts, whereas temporal and multi-modal KGs add time or other information to better represent real-world scenarios.
  • The survey compares prior work by graph type at the top level and by techniques and reasoning scenarios at the base level.
  • The survey addresses coverage gaps in earlier reviews, which largely emphasized static KGR or focused on construction and applications rather than multi-modal reasoning.
  • It reviews static, temporal, and multi-modal KGR models, summarizes performances and datasets, and analyzes model strengths, weaknesses, and suitable scopes.
  • The authors provide an open-source collection of 180 state-of-the-art KGR models and 67 typical datasets, and identify challenges and potential opportunities.

2 PRELIMINARY

The paper defines static, temporal, and multi-modal knowledge graphs and formulates reasoning tasks across transductive, inductive, interpolation, and extrapolation settings. It organizes reviewed models with a bi-level taxonomy based on graph types, techniques, and scenarios.

  • Knowledge graph definitions: Static KGs contain entity, relation, and fact sets, with facts represented as entity-relation-entity triplets.Static KGs are distinguished from other KG types by the absence of temporal or multi-modal extensions.
  • Knowledge graph definitions: Temporal KGs are sequences of timestamped static KGs, where each fact is represented by a head entity, relation, tail entity, and timestamp.A snapshot at timestamp t contains the entities, relations, and facts occurring at t.
  • Knowledge graph definitions: Multi-modal KGs contain facts with multiple modalities, represented either as entities in N-MMKGs or as attributes in A-MMKGs.The two representation modes differ in whether additional modal data becomes graph entities or new attributes.
  • Task formulation: KGR infers missing facts, with static and temporal tasks using scoring-based fact prediction and multi-modal tasks additionally requiring modality-fusion modules.Temporal reasoning predicts quadruples at specific timestamps, while multi-modal reasoning incorporates extra information during inference.
  • Reasoning scenarios: Transductive reasoning uses entities and relations seen during training, whereas inductive reasoning permits unseen entities or relations in queried facts.Interpolation predicts facts at timestamps within the observed range, while extrapolation predicts facts at future timestamps beyond that range.
  • Taxonomy design: The bi-level taxonomy first classifies models by three graph types, then by fourteen techniques and four reasoning scenarios.Technique categories differ across static, temporal, and multi-modal KGR, including embedding-, path-, rule-, RNN-, and Transformer-based families.

3 STATIC KGR MODEL

The static KGR review covers embedding-based, path-based, and rule-based models, together with transductive and inductive reasoning. It contrasts their expressive power, explainability, complexity, and suitability for unseen entities.

  • Overview: The survey systematically reviews 90 static KGR models organized by techniques and reasoning scenarios.Static models are categorized into embedding-based, path-based, and rule-based families.
  • Embedding-based models: Embedding-based models learn entity and relation representations from fact triplets and rank candidate facts using scoring functions.Their main subtypes are translational, tensor decompositional, and neural network models.
  • Embedding-based models: Tensor decompositional models represent KGs as three-way tensors decomposed into low-dimensional entity and relation vectors.RESCAL models pairwise latent-factor interactions with matrices, while later models modify the decomposition or interaction structure.
  • Path-based and rule-based models: Path-based models mine logical knowledge from paths between queried entities, while rule-based models use rules that map sets of facts to inferred facts.Path-based methods include random-walk, path-ranking, and neural multi-hop approaches; rule-based methods may mine and embed logical rules.
  • Reasoning scenarios: 56 transductive and 34 inductive models are reviewed, with 56.25% of inductive models using GNNs and 37.5% using rule-based techniques.Path-based models lack demonstrated inductive ability, whereas rule-based models can generalize entity-agnostic rules; GNNs support both scenarios.
  • Observation and discussion: Embedding-based models generally offer stronger expressive ability but less explainability, while path-based and rule-based models are more explainable but face complexity or expressiveness limits.The review identifies recent emphasis on GNN-based and inductive reasoning models because of scalability and expressive-ability concerns.

4 TEMPORAL KGR MODEL

The temporal KGR review organizes models by how they encode time, distinguishing RNN-based from RNN-agnostic approaches and interpolation from extrapolation. It finds broader scenario compatibility for RNN-based and time-operation models than for time-vector models.

  • Overview: The survey systematically reviews 58 temporal KGR models according to temporal-integration techniques and reasoning scenarios.Temporal models are divided into RNN-based and RNN-agnostic families.
  • RNN-based models: RNN-based models use recurrent networks to model temporal information, including basic RNN, LSTM, and GRU-enhanced variants.Examples combine recurrent networks with graph convolutions, temporal constraints, path reasoning, or timestamp-wise message passing.
  • RNN-agnostic models: RNN-agnostic models incorporate time without recurrent frameworks through time-vector-guided or time-operation-guided mechanisms.Time-vector methods fuse temporal embeddings with fact representations, whereas time-operation methods use mechanisms such as hyper-planes or time-related rewards.
  • Reasoning scenarios: 34 interpolation and 24 extrapolation models are reviewed, with extrapolation models including 42.57% time-operation and 47.62% RNN-based approaches.RNN-based and time-operation models show similar interpolation-to-extrapolation ratios and compatibility with both scenarios.
  • Observation and discussion: RNN-based models generally model temporal information better and are more readily adopted for extrapolation than other approaches.The review identifies extrapolation as an early-stage area, comprising only around 30% of temporal KGR models.

5 MULTI-MODAL KGR MODEL

The survey organizes multi-modal KGR models by their mechanisms for incorporating textual, visual, and other modal information. It reviews transformer-based and transformer-agnostic approaches, datasets, performance comparisons, and current trends.

  • Model Taxonomy: The survey reviews 32 multi-modal KGR models and divides them into transformer-based and transformer-agnostic categories.The division is based on how models fuse extra multi-modal information.
  • Model Taxonomy: Directly applying static KGR models to multi-modal scenarios generally produces sub-optimal performance because they lack fusion modules for additional modalities.Relevant information can include text and images.
  • Transformer-based Models: Transformer-based models use pretrained transformer frameworks to encode multi-modal features and support unified processing across modalities.Examples include VBKGC and Knowledge-CLIP, which use pretrained transformers or CLIP-based representations.
  • Transformer-agnostic Models: Transformer-agnostic models extend unimodal KGR models with separate mechanisms for encoding and fusing extra modal information.Examples include language-based entity-description encoders, attention mechanisms for images, and multi-modal graph attention.
  • Observation and Discussion: 18% of reviewed KGR models address multi-modal scenarios, indicating that multi-modal KGR remains at an early research stage.The survey identifies substantial room for further exploration.

6 DATASETS

The survey compiles benchmark datasets across static, temporal, and multi-modal KGR. It records dataset descriptions and statistics, while collecting the datasets in an accompanying GitHub repository.

  • Dataset Overview: The survey summarizes typical KGR datasets for static, temporal, and multi-modal knowledge graphs.It covers 38 static transductive datasets, 15 static inductive datasets, 18 temporal datasets, and 11 multi-modal datasets.
  • Static Datasets: The static collection distinguishes transductive and inductive benchmarks and presents their statistics in separate tables.Table 10 covers static transductive reasoning, while Table 11 covers static inductive reasoning.
  • Static Datasets: Static datasets include biomedical, commonsense, lexical, geographic, family, and general-purpose knowledge graphs.Examples include Hetionet, ATOMIC, WordNet, Countries, FAMILY, Wikidata, and FreeBASE.
  • Temporal Datasets: Temporal datasets add timestamp information and include event, movie, political-event, Wikidata, and YAGO-derived benchmarks.Examples include GDELT, IMDB, ICEWS, Wikidata variants, and YAGO variants.
  • Multi-modal Datasets: Multi-modal datasets combine KG triples with textual descriptions, images, visual entities, or numeric literals.Examples include FB-IMG-TXT, IMGpedia, MKG, MMKG, Richpedia, and WN9-IMG-TXT.

7 CHALLENGE AND OPPORTUNITY

The survey identifies challenges spanning out-of-distribution reasoning, scalability, multi-relational facts, multi-modal fusion, explainability, applications, and KG–LLM integration. It presents these areas as opportunities for further KGR research.

  • 7.1 Out-of-distribution Reasoning: Out-of-distribution reasoning must address continuously emerging entities and relations that are under-explored in original knowledge graphs.Inductive models target unseen entities, while few-shot and text-based methods target unseen relations; performance can drop when language models are not finely trained.
  • Scalable Reasoning: Industrial-scale KGs require more efficient reasoning, motivating progressive propagation, Bellman–Ford aggregation, greedy search, and graph-clustering approaches.NBF-net replaces DFS-based aggregation with Bellman–Ford, while A∗Star Net further optimizes aggregation greedily.
  • Multi-relational Reasoning: Multi-relational facts are structurally and semantically more complex than uni-relational or bi-relational facts, yet models often omit some relations by simplifying them.This leaves multi-relational reasoning comparatively under-addressed.
  • 7.4 Multi-modal Reasoning: Existing multi-modal KGR models often concatenate modality embeddings for scoring, leaving adaptive and more fine-grained fusion as an open direction.An adaptive fusion mode could weigh the importance of different modalities.
  • 7.5 Explainable Reasoning: Embedding-based neural KGR models offer expressive ability but weaker explainability, whereas rule-based and path-based models are more explainable but computationally expensive and less expressive.Hybrid approaches such as ARGCN seek a trade-off, but most remain rough.
  • Applications and LLMs: KGR applications in domains such as medicine, finance, and plagiarism detection require further study beyond the many existing theoretical methods.The survey also identifies cooperation between KGR and LLMs as a future research hotspot.

8 CONCLUSION

The survey provides a bi-level review of KGR across static, temporal, and multi-modal graphs, together with datasets, challenges, and opportunities. Its repository collects 180 state-of-the-art models and 67 typical datasets.

  • Conclusion: The survey classifies KGR models by graph types at the top level and by techniques and reasoning scenarios at the base level.It covers three graph types, fourteen techniques, and four reasoning scenarios.
  • Conclusion: The accompanying GitHub repository collects papers and codes for 180 state-of-the-art KGR models and 67 typical datasets.The collection is intended to support the research community.
Loading 2212.05767v7…