TLDR: A new AI model, the Relation-Aware LNN–Transformer, predicts a road user’s next intersection by representing trajectories on a city’s road-intersection graph. It uses direction-aware POI aggregation and a hybrid neural network to capture both short-term temporal dynamics and long-range spatial dependencies. The model significantly outperforms existing methods in accuracy and maintains high resilience to GPS and POI data noise, making it highly suitable for real-world navigation and urban planning applications.
Predicting where a person will go next is crucial for many applications, from helping you navigate to planning city traffic. Traditionally, many prediction systems assume a “closed world,” meaning they only consider a fixed list of popular places (Points of Interest or POIs). However, real-world human movement is often exploratory, leading to new or less-visited locations, and is heavily influenced by the structure of road networks.
Researchers at New York University Shanghai have introduced a new approach to tackle these limitations. Their paper, titled “Relation-Aware LNN–Transformer for Intersection-Centric Next-Step Prediction,” proposes a framework that focuses on road intersections rather than just POIs. This allows the model to predict next steps beyond a predefined set of locations, better reflecting how people move through urban environments.
A New Way to Understand Urban Space
The core idea is to represent road-user trajectories on the city’s road-intersection graph. This means the model sees the city as a network of intersections and road segments, rather than just a collection of popular spots. To give the model a rich understanding of the environment, the researchers developed a “sector-wise directional POI aggregation.” Imagine dividing the area around each intersection into pie-like slices. Within each slice, the model gathers information about POIs, including their distance, direction (bearing), density, and simply whether they are present. This helps the model understand the directional context of the urban space, making it easier to distinguish between paths that might look similar topologically but differ in direction.
Combining Temporal Dynamics and Spatial Relationships
For processing sequences of movements, the team integrated a “Relation-Aware LNN–Transformer.” This is a hybrid model that combines two powerful components: a Continuous-time Forgetting Cell (CfC-LNN) and a bearing-biased self-attention module. The CfC-LNN is excellent at capturing fine-grained changes over time, which is important for understanding the immediate dynamics of a trajectory. The bearing-biased self-attention module, on the other hand, helps the model understand long-range spatial dependencies by adding a directional bias to how it “pays attention” to different parts of the trajectory. This compact model, with only 2.34 million parameters, is designed to be efficient yet powerful.
Impressive Performance and Robustness
The model was tested on city-scale road-user trajectories from the GeoLife dataset, which includes GPS data from various modes of transport like walking, cycling, and driving. It was compared against six state-of-the-art baseline models. The results were significant: the new model outperformed baselines by up to 17 percentage points in Accuracy@1 (predicting the exact next step correctly) and 10 percentage points in Mean Reciprocal Rank (MRR), which measures how high the correct answer ranks in the prediction list. These improvements were consistent across different path lengths, from short (1 km) to extended (7 km and full trajectories).
Beyond just accuracy, the model also demonstrated remarkable resilience to noise. Even with GPS perturbations of up to 50 meters, its Accuracy@1 dropped by only 2.4 percentage points. Similarly, with 25% noise in POI features, it maintained high performance. This robustness is crucial for real-world applications where data can often be imperfect. The stability comes partly from the continuous-time nature of the CfC-LNN and the topology-aware attention mechanism.
Also Read:
- A New Model for Predicting Urban Network Dynamics with Enhanced Efficiency
- Traffic-R1: A New AI Model for Smarter, More Human-Like Traffic Control
Why This Matters
This new framework offers a significant step forward in human mobility modeling. By moving beyond fixed POI sets and explicitly incorporating the topological and directional constraints of road networks, it provides a more realistic and accurate prediction of next steps. Its efficiency and robustness make it suitable for practical applications in personalized navigation, traffic forecasting, and urban planning. For more technical details, you can read the full research paper available here.
