Google Research Introduces ME-POIs Framework
Google Research unveils ME-POIs, a 53.7M parameter framework incorporating real-world mobility patterns into text-based POI embeddings.
Stock photo for illustration only, not from the actual event
- Google Research releases ME-POIs, combining mobility behavior with spatial POI embeddings.
- The model contains 53.7 million parameters and trains on a single NVIDIA Tesla V100 16GB GPU.
- Utilizes a 4-layer, 8-head Transformer architecture to generate contextualized visit vectors.
- Evaluations across Los Angeles and Houston show performance improvements on map-enrichment tasks.
Google Research has introduced a novel framework called ME-POIs, designed to overcome the limitations of traditional Points of Interest (POIs) that rely strictly on text data such as names, addresses, and categories while ignoring actual human usage patterns. Rather than providing a downloadable checkpoint, the researchers built the system as a framework to be retrained from scratch.
In terms of compute requirements, the model operates on a modest scale with approximately 53.7 million parameters, pre-trained on a single NVIDIA Tesla V100 16GB hardware unit. However, the primary barrier lies in data acquisition, which requires licensed foot-traffic logs, first-party visit records, and precise POI polygons.
Stock photo for illustration only, not from the actual event
The system processes each visit as a triple consisting of coordinates, arrival time, and departure time. Three dedicated factorized encoders handle these inputs: Space2Vec for multi-scale spatial coordinates, and two Time2Vec encoders for arrival and departure times respectively, preserving the distinction of dwell duration. These vectors are then concatenated and processed through a 4-layer, 8-head Transformer architecture (d_h = 512).
The core objective relies on contrastive learning. Every POI features a learnable prototype, and an InfoNCE loss pulls each visit embedding toward its corresponding POI prototype while repelling other POIs in the minibatch. This turns the prototype into a functional centroid that averages out individual user schedules.
Integrating mobility data into geographic embeddings marks an essential evolution in spatial computing. While static map attributes describe what a place is called, human visit patterns reveal how a space functions in practice, distinguishing locations that share identical category labels but serve entirely different roles in daily urban life.
Data sparsity presents a significant hurdle, as only 9.07% of Los Angeles POIs and 7.04% of Houston POIs cleared the anchor threshold of 100 and 50 total visits respectively. For the long-tail locations, ME-POIs computes normalized Gaussian kernels across three bandwidths to transfer anchor visit histograms, supplemented by KL divergence losses and text embeddings aligned with the GeoLLM recipe.
Evaluations across anonymized mobility datasets for Los Angeles (39,557 POIs, 6.9M visits, 2019) and Houston (28,419 POIs, 715,604 visits, March 2020) demonstrated that ME-POIs improved 34 out of 35 task pairings in Los Angeles, including a notable 81.9% F1 boost on visit intent when paired with Gemini.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment