Source-linked AI summary
Geo-Spotting: Mining Online Location-based Services for Optimal Retail Store Placement
Dmytro Karamshuk, Anastasios Noulas, Salvatore Scellato, Vincenzo Nicosia, Cecilia Mascolo
TL;DR
The paper asks how to identify promising retail locations when traditional data are costly. Using New York Foursquare data, it evaluates geographic and user-mobility features for predicting store popularity and finds that combining features in supervised models improves ranking performance across three chains, subject to chain and urban-setting limitations.
Problem
Retail store placement is important, but traditional approaches rely on relevant demographic, revenue, and human-flow data that are usually expensive to acquire.
Method
The paper mines New York Foursquare data to evaluate geographic and user-mobility features for ranking candidate areas by predicted retail quality.
Results
Combining multiple features in supervised learning significantly improves performance, with the optimal retail spot ranked in the top 5% in more than half of the cases.
Takeaways & Limitations
Geographic context and user movements together provide useful signals for assessing retail-area quality across the three analyzed chains.
Takeaways & Limitations
The evaluation is limited to Foursquare data from New York, and nearby same-chain facilities can have similar features despite substantially different popularity.
Abstract
from arXiv · showhide
The problem of identifying the optimal location for a new retail store has been the focus of past research, especially in the field of land economy, due to its importance in the success of a business. Traditional approaches to the problem have factored in demographics, revenue and aggregated human flow statistics from nearby or remote areas. However, the acquisition of relevant data is usually expensive. With the growth of location-based social networks, fine grained data describing user mobility and popularity of places has recently become attainable. In this paper we study the predictive power of various machine learning features on the popularity of retail stores in the city through the use of a dataset collected from Foursquare in New York. The features we mine are based on two general signals: geographic, where features are formulated according to the types and density of nearby places, and user mobility, which includes transitions between venues or the incoming flow of mobile users from distant areas. Our evaluation suggests that the best performing features are common across the three different commercial chains considered in the analysis, although variations may exist too, as explained by heterogeneities in the way retail facilities attract users. We also show that performance improves significantly when combining multiple features in supervised learning algorithms, suggesting that the retail success of a business may depend on multiple factors.
1. INTRODUCTION
The paper frames retail store placement as predicting which candidate urban areas will attract many Foursquare check-ins. It mines geographic and mobility features and finds that combined supervised models improve ranking performance, while location-based data may support broader urban and commercial applications.
- Problem: Retail placement is framed as ranking candidate city areas by their prospect of attracting many Foursquare check-ins.The paper treats predicted check-in popularity as a measure of retail quality.
- Features: Geographic features encode venue types, venue density, and competitor presence, while mobility features model nearby transitions and attraction from remote locations.These features capture both static spatial structure and user movement around a target store.
- Evaluation: Competition and spatial-structure features perform best among individual ranking features, while distant-user attraction and venue-type transition quality lead the mobility features.Feature performance can vary across retail chains because customer-attraction patterns differ.
- Evaluation: Combining features in supervised learning keeps the optimal retail spot in the top 5% of predictions in more than half of the cases.The paper reports a clear improvement over individual-feature evaluation.
- Implications: The approach is presented as potentially useful for urban mining applications involving business and citizen services, house prices, development indices, and location-based marketing.The paper positions location-based services as a source of commercially relevant spatio-temporal information.
- Data and scope: The analysis examines three retail chains in New York using Foursquare venue and check-in data.The dataset includes 37,442 venues, 46,855 users, and 620,932 check-ins collected over five months.
Analysis of Retail Store Popularity
Retail venue popularity is highly heterogeneous across New York locations and chains. The analysis links store placement and popularity to nearby venue categories, check-in behavior, and chain-specific differences.
- Venue popularity: Venue popularity follows a power-law-like distribution, with check-in frequencies spanning many orders of magnitude.The CCDF compares all venues with the three analyzed retail chains.
- Chain differences: Starbucks has 186 venues and an average of 1129.97 check-ins per place, nearly five times the average of the other chains.Dunkin’ Donuts and McDonald’s have 104 and 66 stores, respectively.
- Chain differences: Similar check-in patterns across chains should not be assumed because coffee shops and fast-food restaurants may attract different kinds of visits.The paper notes that longer coffee-shop stays may increase check-in likelihood, whereas fast-food visits may be more opportunistic.
- Spatial interactions: The study compares observed interactions between venue categories with expectations under uniform random placement across the city.This is the basis of Jensen’s inter-category attraction coefficients.
- Spatial interactions: Starbucks and McDonald’s are frequently located near train stations, while corporate offices, museums, hostels, and gas stations also appear as chain-specific attractors.The reported inter-category coefficients compare observed placement frequencies with random expectations.
Exploiting Mobility Data for Retail Analysis
Foursquare check-in transitions reveal how far users travel to the three retail chains and which venue categories disproportionately supply customers. These mobility signals support ranking candidate areas for retail placement.
- Mobility patterns: 50% of incoming transitions originate within 200–300 meters, while 80–90% occur within 1 kilometer of the three retail chains.The CDF measures distances traveled from other venues toward the chains.
- Mobility patterns: Hostels, train stations, and financial centers have the highest transition ratios toward Starbucks, reaching 17.44, 4.79, and 4.32 times random-transition levels.These categories also show high attractiveness coefficients toward Starbucks.
- Retail placement: The placement task ranks candidate geographic areas by predicted retail quality to identify the area expected to attract the largest number of visits.The paper frames this ranking problem using features mined from Foursquare data.
Problem Formulation
The paper represents each candidate retail area by geographic coordinates and a radius, then ranks areas using feature scores. It assumes observed Foursquare check-ins proxy relative place popularity.
- Candidate areas: Each candidate area l is encoded by latitude, longitude, and radius r, with r selected as 200 meters after testing different values.The selected radius produced the highest prediction performance across independent experiments.
- Candidate areas: For every candidate area l, each feature produces a score χ̂_l, and the highest-scoring area is selected for store placement.The formulation ranks geographic areas according to predicted retail quality.
- Assumption: The task assumes that empirically observed Foursquare check-ins can serve as a proxy for a place’s relative popularity.This assumption links observed user activity to the store-placement objective.
Prediction Features
The paper mines geographic and user-mobility features from Foursquare to assess the retail quality of candidate areas. These features capture surrounding venues, competition, area popularity, and movements between nearby or external places.
- The extracted signals comprise place-geographic features describing venue types and spatial interactions, and user-mobility features describing movements and transitions between places.
- Geographic features measure surrounding venue density, heterogeneity, and competitiveness within a radius around each candidate location.Density counts nearby places, entropy captures variation in venue types, and competitiveness measures the share of nearby venues matching the target venue type.
- Competition is treated cautiously because nearby competitors may either attract an ecosystem and customer crowd or divide existing customers.
- The Jensen-based quality feature weights nearby venue types by inter-category attractiveness coefficients derived from observed versus random spatial patterns.Intra-category coefficients additionally assess whether venues of the same type form spatial clusters.
- Mobility features measure neighboring check-ins, transitions occurring within the candidate area, incoming external flows, and transition quality from nearby venues.Transition quality weights potential local customer sources using transition probabilities between venue categories and check-in counts.
- The feature set is intended to identify which static and dynamic information sources best predict retail-chain popularity.The paper compares geographic and mobility signals to understand their value for assessing retail quality.
4. FEATURE EVALUATION
The evaluation assesses how well individual mined features predict highly ranked retail locations for three commercial chains. Figure 5 reports Accuracy@X% scores for Starbucks, Dunkin’ Donuts, and McDonalds.
- Individual prediction features are evaluated for their ability to forecast high rankings among the most popular geographic spots.
- Figure 5 compares Accuracy@X% scores for individual feature predictions across Starbucks, Dunkin’ Donuts, and McDonalds.
Methodology and Metrics
The methodology ranks candidate areas using feature scores and compares predicted rankings with actual check-in-based popularity rankings. Evaluation uses NDCG@k and Accuracy@X% under repeated geographic cross-validation.
- For each retailer, each feature scores candidate areas, which are then ordered into a predicted ranking for a new store.The candidate set is denoted L, and the resulting ranked list is R.
- The evaluation compares feature-based predicted rankings with ground-truth rankings based on the actual number of store check-ins.
- NDCG@k measures how highly the top-k actually popular locations appear in the predicted ranking, using relevance scores based on actual ranking positions.Scores are normalized by the ideal ranking and lie between 0 and 1.
- The random-ranker expected value of NDCG@k serves as the comparison baseline.
- Accuracy@X% measures the fraction of predictions whose optimal location falls within the top X% of the actual popularity ranking.Using percentages rather than absolute top-K sizes enables comparison across chains.
- Geographic cross-validation repeatedly samples 33% of a brand’s associated areas as candidate locations and uses the remaining areas for training.Each experiment is repeated 1000 times.
Individual Feature Performance
Individual geographic and mobility features predict popular retail locations, but their strongest signals vary by chain. Competition, spatial structure, and customer-flow patterns each contribute to ranking candidate areas.
- NDCG@10 improves over a random baseline for all features across the three chains, with some features performing considerably better than others.
- Geographic features lead differently by chain: Jensen Quality scores 0.81 for Dunkin’ Donuts, while Competitiveness scores 0.70 for Starbucks and 0.78 for McDonald’s.
- Competitiveness suggests that lacking nearby competitors can help a retail facility attract customers, while additional competitors may reduce its customer share.
- Among mobility features, Transition Quality scores 0.66 for Starbucks and 0.81 for Dunkin’ Donuts, whereas Incoming Flow scores 0.81 for McDonald’s.
- McDonald’s attracts users from farther away more often, with beyond-200-meter attraction probability near 0.65 versus 0.50 for Starbucks and 0.45 for Dunkin’ Donuts.
- For Dunkin’ Donuts, Jensen Quality and Transition Quality both achieve NDCG=0.81, indicating spatial structure alone is sufficiently correlated with user mobility there.
- Accuracy@X% rankings broadly resemble NDCG results, but Competitiveness rises faster and exceeds 70% at Accuracy@15%; Starbucks performs lower, potentially because its stores are denser and geographically overlapping.
5. SUPERVISED LEARNING APPROACH
The supervised approach combines geographic and mobility features to rank candidate locations using regression and pairwise learning-to-rank models. Adding mobility information substantially improves prediction across chains, with Support Vector Regression achieving the best overall performance.
- The approach tests whether composing geographic and mobility signals predicts candidate-area popularity better than individual features.
- The supervised models include Support Vector Regression, M5 decision trees, regularized Linear Regression, and the pairwise RankNet algorithm.
- Regression models learn scores for candidate areas and then rank them, while RankNet learns ordering relations between pairs of venues from their features.
- Adding mobility features considerably improves NDCG@10 across all chains; Starbucks reaches 0.77, a 10% improvement over its best single-feature prediction.
- Supervised learning reaches best NDCG@10 values of 0.83 for Dunkin’ Donuts and 0.84 for McDonald’s, versus single-feature values of 0.81 and 0.79.
- At Top-20%, accuracy exceeds 80% across all chains and reaches 93% for Dunkin’ Donuts, outperforming best individual-feature predictions.
- Across prediction techniques, Support Vector Regression achieves the optimal performance, while geographic-plus-mobility combinations outperform single-feature predictions.
6. IMPLICATIONS
Foursquare’s rich spatio-temporal data can support retail-area assessment by representing venue semantics, competition, and human mobility. The paper also identifies noise and city-specific heterogeneity as important boundaries.
- Location-based services provide new data layers that can support commercial analysis and applications in urban mining.
- Geo-tagged venue semantics enable representations of urban commercial value that include nearby competitive facilities and venue types.
- Different retail chains attract customer crowds through different spatial and mobility patterns, despite some common attraction patterns.
- Noise can impede prediction, especially for Starbucks, where nearby same-chain stores may have similar features but different popularity.
- Local architecture and planning can alter customer flows even when venues have similar coordinates, and the authors plan to extend evaluation to more cities and chains.
7. RELATED WORK
The paper builds on spatial retail-location research and urban-mining studies by adding human-mobility traces to geographic retail analysis. It positions Foursquare as a source of fine-grained urban and commercial signals.
- Earlier retail-placement research used spatial interaction models and retail-quality indices to study store positioning and urban commercial structure.
- Jensen’s approach modeled attraction and repulsion among 55 retail-store types and assessed retail positioning in Lyon.
- Other studies linked street centrality with retail-store density in Bologna and Barcelona to examine urban structure and land use.
- This paper extends geographic retail analysis by incorporating features mined from human mobility traces, combining geographic and mobility signals to assess potential store locations.
- Urban-mining work has extracted knowledge from cellular mobility and public-transport movement data for applications such as event recommendation and spending analysis.
8. CONCLUSION
The paper uses Foursquare mobility data to study retail-store popularity in New York and evaluates geographic and user-movement features for placement assessment.
- Foursquare mobility data were collected to analyze the check-in popularity of three retail store chains in New York.
- Features model spatial, semantic, and user-movement information in the surrounding area.
- User attractors and same-type retail stores are the strongest place-semantic indicators of popularity.
- Combining geographic and mobility features in supervised learning better explains place popularity than individual features alone.
- The study plans comparisons across cities and additional place and retail-chain types.