NAICS-Aware Graph Neural Networks for Large-Scale POI Co-visitation Prediction: A Multi-Modal Dataset and Methodology CCS: Computing methodologies Neural networks CCS: Information systems Spatial-temporal systems CCS: Information systems Location based services CCS: Information systems Recommender s

Yazeed Alrubyli, Omar Alomeir, Abrar Wafa, Diána Hidvégi, Hend Alrasheed, Mohsen Bahrami · 2025

Read NAICS-Aware Graph Neural Networks for Large-Scale POI Co-visitation Prediction: A Multi-Modal Dataset and Methodology CCS: Computing methodologies Neural networks CCS: Information systems Spatial-temporal systems CCS: Information systems Location based services CCS: Information systems Recommender s by Yazeed Alrubyli, Omar Alomeir, Abrar Wafa, Diána Hidvégi, Hend Alrasheed, Mohsen Bahrami (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-19697.txt.

Introduction

Alrubyli and five colleagues predict which places people visit in pairs, using business type codes in twenty twenty-five. Knowing that a coffee shop visit leads to a gym visit helps city and retail planning. Past models leaned on distance alone, which misses why unlike businesses draw unlike crowds. Their network learns a code for each industry type as it trains.

The data held near ninety-five million paired visits across most United States states. Their model more than doubled the fit of the best past methods. Rank quality of the top guesses rose by about a third. Splitting the huge task by state made it run at scale. Business meaning, they show, is a strong signal.

Catalog record

CREATORYazeed Alrubyli, Omar Alomeir, Abrar Wafa, Diána Hidvégi, Hend Alrasheed, Mohsen Bahrami
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-19697

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

NAICS-Aware Graph Neural Networks for Large-Scale POI Co-visitation Prediction: A Multi-Modal Dataset and Methodology CCS: Computing methodologies Neural networks CCS: Information systems Spatial-temporal systems CCS: Information systems Location based services CCS: Information systems Recommender systems CCS: Applied computing Forecasting Yazeed Alrubyli, Omar Alomeir, Abrar Wafa, Diána Hidvégi, Hend Alrasheed, Mohsen Bahrami Abstract . Understanding where people go after visiting one business is crucial for urban planning, retail analytics, and location-based services. However, predicting these co-visitation patterns across millions of venues remains challenging due to extreme data sparsity and the complex interplay between spatial proximity and business relationships. Traditional approaches using only geographic distance fail to capture why coffee shops attract different customer flows than fine dining restaurants, even when co-located. We introduce NAICS-aware GraphSAGE, a novel graph neural network that integrates business taxonomy knowledge through learnable embeddings to predict population-scale co-visitation patterns. Our key insight is that business semantics—captured through detailed industry codes—provide crucial signals that pure spatial models cannot explain. The approach scales to massive datasets (4.2 billion potential venue pairs) through efficient state-wise decomposition while combining spatial, temporal, and socioeconomic features in an end-to-end framework. Evaluated on our POI-Graph dataset comprising 94.9 million co-visitation records across 92,486 brands and 48 US states, our method achieves significant improvements over state-of-the-art baselines: R 2 R^{2} increases from 0.243 to 0.625 (157% improvement), with strong gains in ranking quality (32% improvement in NDCG@10). 1. Introduction The rapid adoption of location-enabled mobile devices has created unprecedented opportunities to analyze population-scale movement between Points-of-Interest (POIs). Such data underpins applications from urban resilience and pandemic response ( Liu et al. 2020 ) to retail analytics ( Fang et al. 2022 ) and next-POI recommendation ( Yang et al. 2022 ) . In many scenarios, the critical primitive is not an isolated visit, but a co-visitation event : two venues visited by the same user within a short temporal window. Accurate co-visitation forecasts can directly benefit demand-aware marketing, multi-stop itinerary planning, and infrastructure investment analysis. Predicting co-visitation patterns across the United States presents formidable challenges: with over 92,000 brands and 276 business categories, the potential co-visitation space contains 4.2 billion brand pairs. Empirical analysis reveals extreme sparsity—over 99.9% of brand pairs exhibit zero co-visits monthly, while observed pairs span five orders of magnitude in intensity (1 to 28,000+ monthly interactions). This creates a "needle in haystack" problem where traditional collaborative filtering approaches fail. Consider predicting co-visits between a Starbucks and nearby restaurants in Manhattan. Traditional approaches using only geographic proximity miss crucial business semantics: coffee shops exhibit high complementarity with fast-casual dining (due to meal timing) but low complementarity with fine dining (different customer occasions). Capturing these nuanced relationships requires understanding both spatial proximity and semantic complementarity—precisely what business taxonomy integration provides. Accurate co-visitation prediction enables transformative applications across domains. In retail analytics, it helps optimize store placement by identifying complementary business clusters, with major chains reporting 15–25% gains in site selection accuracy ( Barbosa et al. 2023 ) . In urban planning, it informs mixed-use developments that increase foot traffic and local business revenue by 30–40% ( Luca et al. 2023 ) . For economic development, it helps identify ecosystem vulnerabilities and guide strategic investment, especially during economic disruptions. In location-based services, co-visitation models improve itinerary recommendations and raise user engagement by 20–35%. Specifically, co-visitation prediction between points of interest (POIs) addresses several computational challenges in spatial analytics and location modeling. First, it enables efficient pre-filtering for sales prediction models by identifying areas with high co-visitation connectivity, reducing the search space from millions of potential locations to those with demonstrated customer flow relationships. Second, co-visitation patterns provide location recommendation signals that incorporate semantic POI relationships beyond spatial proximity, enabling more accurate site selection based on complementary venue types and market potential. Third, predicted co-visitation flows can serve as features embedded into both sales and cannibalization prediction models, capturing indications for customer redistribution effects after opening at a new location, as well as attractiveness of the area. Despite its utility, co-visit forecasting remains technically challenging. Co-visit matrices are extremely sparse and subject to temporal drift due to seasonality and economic shifts. Co-visitation patterns are shaped by complex interactions among geography, business function, and socio-demographics. Finally, modern datasets contain millions of venues and billions of candidate edges, making classical methods computationally infeasible ( Wu et al. 2020 ) . Graph Neural Networks (GNNs) provide a principled framework to combine rich node/edge attributes with topological structure. Recent GNN-based POI models have evolved rapidly ( Yan et al. 2023 ; Wang et al. 2023b ; Xu et al. 2024 ) , but they largely focus on individual user behavior through node-level prediction tasks rather than population-level co-visitation forecasting through edge-regression. Moreover, they seldom incorporate domain-specific taxonomies like NAICS that are critical for interpretability ( Zhang et al. 2021a ; Dong et al. 2023 ) . In this work, we cast co-visit prediction as a spatio-temporal edge-regression problem and present the first end-to-end GNN that jointly embeds NAICS codes, temporal signals, and spatial relations while scaling to hundreds of millions of POI pairs. We focus on pre-pandemic data (Jan 2018–Mar 2020) to establish robust baselines on stable behavioral patterns. Unlike existing approaches targeting individual user behavior, our method predicts population-level mobility patterns through learnable business semantics. Our contributions are twofold: (1) Methodological: We formulate POI co-visitation as an edge regression task and develop a NAICS-aware GraphSAGE framework that incorporates learnable business taxonomy embeddings for population-level co-visitation prediction at scale. (2) Empirical: We achieve strong predictive performance (Test R^{2} of 0.625, representing a 157% improvement over the best baseline) through extensive experiments, demonstrating significant improvements over traditional spatial interaction models and existing graph-based approaches. 2. Related Work Graph Neural Networks have emerged as powerful tools for spatial-temporal prediction due to their ability to model complex relational dependencies ( Wu et al. 2020 ; Bronstein et al. 2023 ) . Early spatial GNN approaches like STGCN ( Zhao et al. 2019 ) and GraphWaveNet ( Wu et al. 2019 ) demonstrated effectiveness for traffic flow prediction through spatial-temporal convolutions. Recent advances include ASTGCN’s adaptive adjacency learning ( Guo et al. 2019 ) , STSGCN’s multi-component fusion ( Song et al. 2020 ) , and GMAN’s spatial-temporal attention mechanisms ( Zheng et al. 2020 ) . The field has seen rapid evolution toward more sophisticated architectures. STGODE ( Liu et al. 2023 ) introduces ordinary differential equations for continuous spatial-temporal modeling, while STGFormer ( Chen et al. 2023 ) combines transformers with graph convolutions for improved traffic forecasting. Recent work on spatio-temporal graph learning ( Jin et al. 2023 ) and trajectory-based prediction ( Shao et al. 2022 ) demonstrates the growing sophistication of GNN approaches for mobility modeling. For POI applications specifically, several architectures have shown promise: STHGCN ( Yan et al. 2023 ) employs hypergraphs for sequential POI recommendation, GETNext ( Yang et al. 2022 ) augments transformers with trajectory flow maps, and recent work explores multi-modal integration for location-based services ( Xu et al. 2024 ; Wu et al. 2024 ) . Existing POI-focused GNNs primarily target individual user behavior through node-level prediction tasks (next-POI recommendation, user preference modeling). In contrast, our work addresses population-level co-visitation forecasting through edge-regression,

Read the complete text