Massive-STEPS: Massive Semantic Trajectories for Understanding POI Check-ins -- Dataset and Benchmarks

TL;DR

Massive-STEPS tackles the reliance on aging, city-skewed POI datasets by introducing a large-scale, semantically enriched benchmark spanning $12$ cities and two periods ($2012$-$2013$ and $2017$-$2018$). Built atop Semantic Trails and aligned with Foursquare OS Places, it supports longitudinal and cross-city POI recommendation research with reproducible preprocessing and open-source benchmarks. The authors evaluate supervised and zero-shot approaches, revealing that urban features like category entropy significantly influence model accuracy and that transformer-based GNNs excel in many settings, while LLM-based zero-shot methods can rival supervised baselines in several cities. By releasing both the dataset and benchmarking pipeline under Apache 2.0, Massive-STEPS enables open, equitable, and scalable research in human mobility and POI recommendation across diverse urban contexts.

Abstract

Understanding human mobility through Point-of-Interest (POI) recommendation is increasingly important for applications such as urban planning, personalized services, and generative agent simulation. However, progress in this field is hindered by two key challenges: the over-reliance on older datasets from 2012-2013 and the lack of reproducible, city-level check-in datasets that reflect diverse global regions. To address these gaps, we present Massive-STEPS (Massive Semantic Trajectories for Understanding POI Check-ins), a large-scale, publicly available benchmark dataset built upon the Semantic Trails dataset and enriched with semantic POI metadata. Massive-STEPS spans 12 geographically and culturally diverse cities and features more recent (2017-2018) and longer-duration (24 months) check-in data than prior datasets. We benchmarked a wide range of POI recommendation models on Massive-STEPS using both supervised and zero-shot approaches, and evaluated their performance across multiple urban contexts. By releasing Massive-STEPS, we aim to facilitate reproducible and equitable research in human mobility and POI recommendation. The dataset and benchmarking code are available at: https://github.com/cruiseresearchgroup/Massive-STEPS

Figures (6)

• Figure 1: Distribution of POI recommendation studies modeled on specific cities, filtered from Table IV in zhang2025survey. We identified and counted studies that explicitly mentioned city names, revealing the skewness of existing research, which is saturated around New York and Tokyo. In addition, we include the distribution of studies by LBSN platform, showing that Foursquare is by far the most commonly used source of check-in data. The list of identified studies is shown in Table \ref{['tab:lit-rev-studies']}.
• Figure 2: World map highlighting the cities included in the Massive-STEPS dataset.
• Figure 3: Top 10 most frequent POI categories in each city, highlighting local cultural and urban preferences.
• Figure 4: Distribution of trail lengths, showing a long-tailed pattern with most trajectories consisting of a few check-ins.
• Figure 5: Distribution of user activity based on the number of trajectories per user, indicating a cold-start-heavy dataset.