Ratneswaran, Shiveswarran (Purdue University), Ukkusuri, Satish (Purdue University)

Motif-Density Aware Multi-Head Attention Knowledge Graph Embedding ((MA)2KGE) for Spatio-Temporal Representation Learning of Urban Ride-Sharing Networks

Scheduled for presentation during the Invited Session "S20c-Foundation Model-Enabled Scene Understanding, Reasoning, and Decision-Making for Autonomous Driving and ITS" (TH-LA-T20), Thursday, November 20, 2025, 16:40−17:00, Surfers Paradise 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

With the growth of urban ride-sharing services, they often face challenges such as prolonged waiting times, increased vehicle idle times due to unaware of future spatio-temporal demand requests. In parallel, recent development of LLM powered ride-hailing assistants, requires a strong knowledge of urban ride-share networks to answer the natural language queries. The two above mentioned problems demand an efficient knowledge representation of urban ride-sharing networks from multi-sourced data (Trip demands and patterns, POIs, willingness to ride-share). To this end, Knowledge graphs (KGs) are potential solution and their KG Embeddings (KGE) in a dense, low dimensional space while maintain the structures and semantics such that retrieved easily and enables reasoning and interpretation. We present an urban ride-share KG construction method. In addition, the rideshare network KGs are primarily having dense packed motifs (chains, star-hub, hierarchy). Learning such structures and semantics with much attention, paves the way for an efficient KGE. Hence, we propose a Motif-density Aware Multi-Head Attention KGE (MA2KGE) for the spatio-temporal representation learning of urban ride-share networks. Our proposed method achieved the State-of-the-art performance with 0.42 Mean Reciprocal Rank (MRR)