Three papers from the newly formed Artificial Intelligence Institute were recently accepted at AAAI-2024, premier conference on Artificial Intelligence. Two papers are about predicting missing links in knowledge graphs that go beyond triple-shaped facts, and the other one is about the challenge of recognizing human activities from sequences of body poses.
Here is a short description of the three papers:
HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces
Temporal knowledge graphs are a collection of temporal facts (s, p, o, τ), linking a subject s and an object o through a relation label p at a timepoint τ. These graphs are known to be incomplete, and temporal knowledge graph embeddings are a prominent approach for predicting missing temporal facts. To this end, existing temporal knowledge graph embeddings map temporal facts to vectors that lie in a single geometric space, e.g., the Euclidean space. Pan et al. suggest mapping such facts to vectors that lie in a space that results from combining several geometric spaces. Moreover, they propose an attention mechanism that captures that some facts change more often than other facts.
J. Pan, M. Nayyeri, Y. Li, S. Staab. HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces). In: 38th AAAI Conference on Artificial Intelligence (AAAI-24), Vancouver, CA, February 20-27, 2024.
NestE: Modeling Nested Relational Structures for Knowledge Graph Reasoning.
Nested knowledge graphs comprise facts represented by quoted triples, where the subjects and objects themselves are triples. For example, that Barack Obama’s presidentship was succeeded by the one of Donald Trump may be represented in a nested fact of the following form:
((BarackObama, holds_position, President), succeed_by, (DonaldTrump, holds_position, President)).
Such nested facts allow for expressing intricate semantics like temporal constraints and logical patterns. By considering those patterns, Xiong et al. propose a novel method, NestE, that can perform both link prediction, and triple prediction.
For example, it can predict who holds the position of President succeeding Barack Obama ((BarackObama, holds_position, President), succeed_by, (?head, holds_position, President)), or the entire triple representing the successor of Barack Obama ((BarackObama, holds_position, President), succeed_by, (?triple)).
Xiong, M. Nayyeri, L. Luo, Z. Wang, S. Pan, S. Staab. NestE: Modeling Nested Relational Structures for Knowledge Graph Reasoning. In: 38th AAAI Conference on Artificial Intelligence (AAAI-24), Vancouver, CA, February 20-27, 2024.
Navigating Open Set Scenarios for Skeleton-based Action Recognition.
Our third paper tackles human activity recognition from body pose sequences, exploring it under open set conditions, where new types of human actions can occur at any time. Peng et al. reveal that conventional open set recognition methods do not work well with sparse spatiotemporal body pose data and introduce CrossMax – a new model for detecting such out-of-distribution sequences. CrossMax leverages the cross-modal alignment of skeleton joints, bones, and velocities via a novel cross-modality mean max discrepancy suppression mechanism to align latent spaces of these modalities, clearly improving the recognition quality.
K. Peng, Y. Cheng, J. Zheng, R. Liu, D. Schneider, J. Zhang, K. Yang, M. S. Sarfraz, R. Stiefelhagen, A. Roitberg. Navigating Open Set Scenarios for Skeleton-based Action Recognition. In: 38th AAAI Conference on Artificial Intelligence (AAAI-24), Vancouver, CA, February 20-27, 2024.