{"id":113548,"date":"2026-08-25T00:43:21","date_gmt":"2026-08-25T00:43:21","guid":{"rendered":"https:\/\/youzum.net\/google-research-introduces-me-pois-a-mobility-informed-framework-that-adds-how-a-place-is-used-to-text-based-poi-embeddings\/"},"modified":"2026-08-25T00:43:21","modified_gmt":"2026-08-25T00:43:21","slug":"google-research-introduces-me-pois-a-mobility-informed-framework-that-adds-how-a-place-is-used-to-text-based-poi-embeddings","status":"publish","type":"post","link":"https:\/\/youzum.net\/zh\/google-research-introduces-me-pois-a-mobility-informed-framework-that-adds-how-a-place-is-used-to-text-based-poi-embeddings\/","title":{"rendered":"Google Research Introduces ME-POIs: A Mobility-Informed Framework that Adds \u201cHow a Place Is Used\u201d to Text-Based POI Embeddings"},"content":{"rendered":"

A team from Google Research<\/a> and USC has released Mobility-Embedded POIs (ME-POIs)<\/a>, a framework that folds aggregate human movement into text-based place embeddings. The premise is that language models describe what a place is<\/em>, but not how it is used<\/em>. Two coffee shops can share a category, an address block, and a text vector, while one runs commuter turnover and the other holds customers for ninety minutes. ME-POIs encodes each visit as a contextualized vector, then uses contrastive learning to align those visits with one learnable prototype per POI. Across five map-enrichment tasks on Los Angeles and Houston mobility data, adding ME-POIs to strong text encoders improved 34 of 35 model-task pairings in Los Angeles, with relative gains up to 81.9% F1 on visit intent and a 24.7% MAE reduction on busyness. Notably, a variant trained on mobility alone beat Gemini embeddings<\/a> on price-level classification.<\/p>\n

Is it deployable?<\/strong><\/h2>\n

Partially, it is a framework you rebuild, not a checkpoint you download.<\/strong> As of publication, Google Research<\/a> has released the paper<\/a> but no public code or weights. The compute bar is low: the model is ~53.7M parameters and was pretrained on a single NVIDIA Tesla V100 16GB. The real gate is data \u2014 you need licensed foot-traffic or first-party visit logs plus POI polygons.<\/p>\n

How the framework works<\/strong><\/h2>\n

Each visit is a triple: coordinates, arrival time and departure time. Three factorized encoders handle them: Space2Vec<\/a> for multi-scale location, and two Time2Vec<\/a> encoders for arrival and departure separately, so start time and dwell duration stay distinguishable. The concatenated vectors get sinusoidal positional encoding and pass through a 4-layer, 8-head Transformer (d_h = 512) to produce contextualized visit embeddings.<\/p>\n

The core objective is contrastive. Every POI owns a learnable prototype, and an InfoNCE<\/a> loss pulls each visit embedding toward its own POI\u2019s prototype while pushing away the other POIs in the minibatch. The prototype becomes a functional centroid that averages out individual user schedules.<\/p>\n

Sparsity is the hard part. Only 9.07% of Los Angeles POIs and 7.04% of Houston POIs cleared the anchor threshold (100 and 50 total visits respectively). For the long tail, ME-POIs computes normalized Gaussian kernels at three bandwidths \u2014 0.3 km, 1.0 km, 3.0 km \u2014 and transfers anchor visit histograms to sparse POIs, then adds a KL term forcing the sparse embedding to predict that prior. A second KL term supervises anchors against their own empirical distributions. A fourth loss maximizes cosine similarity with projected text embeddings, whose prompts follow the GeoLLM<\/a> recipe: coordinates, category, address, and the ten nearest POIs with distance and direction.<\/p>\n

What the numbers say<\/strong><\/h2>\n

Evaluation covers two anonymized mobility datasets \u2014 Los Angeles (39,557 POIs, 6.9M visits, full-year 2019) and Houston (28,419 POIs, 715,604 visits, 20 days in March 2020) \u2014 across five map-enrichment tasks with frozen-embedding probing. Labels come from SafeGraph for opening hours and closures, and Google Maps for visit intent, busyness, and price level.<\/p>\n

Adding ME-POIs improved 34 of 35 model-task pairings in Los Angeles. Peak relative gains: 16.2% F1 on weekly opening hours (OpenAI-large), 81.9% F1 on visit intent (Gemini<\/a>), 6.5% F1 on permanent closure (E5), and a 24.7% MAE reduction on busyness (Gemini). In Houston, price-level F1 rose 75.1% for GTR-T5. The single regression was Gemini on permanent closure, down 0.4%.<\/p>\n

The more interesting result is the mobility-only variant. Trained with no text alignment at all, it reaches 0.600 accuracy on Los Angeles price level against Gemini\u2019s 0.559 \u2014 collective behavior outperforming the words used to label the place. It also beats every trajectory-based baseline on every task.<\/p>\n

Explainer: the mechanism, step by step<\/strong><\/h2>\n
\n<\/div>\n

\n

Key Takeaways<\/strong><\/h2>\n