TL;DR
This paper introduces a novel method, the Potential Customer Estimator (PCE), to predict the number of potential customers for any arbitrary location and time, leveraging existing retail stores as "sensors." By modeling spatial and temporal correlations using features from location-based social networks, PCE offers a powerful tool for businesses to make informed decisions about store placement and marketing strategies.
What Problem Does This Paper Address?
Understanding where and when potential customers are likely to be is crucial for businesses, especially in urban environments. This knowledge informs critical decisions like selecting optimal locations for new retail stores and planning effective advertising campaigns. However, accurately estimating this potential customer volume for any location (even those without existing stores) and at any given time is a significant challenge. Traditional methods often focus on existing popular locations or historical data without effectively bridging the gap to predict for new, unestablished areas or specific future time points, especially considering the dynamic nature of urban mobility and temporal factors like seasons and events.
Methods
The core of the proposed solution is the Potential Customer Estimator (PCE), which operates in three main stages:
- Correlation Graph (CG) Construction: A multi-layer graph is built to represent spatial and temporal correlations. This graph connects existing stores (labeled nodes) with query locations (unlabeled nodes) and also links query locations across different time periods. The graph incorporates three categories of features:
- Geographical Features (GF): Density of venues, neighbor entropy (venue category diversity), competitiveness (proportion of same-category venues), and attractiveness (inter-venue category interactions).
- Mobility Features (TMF/CMF): Temporal and cumulative measures of area popularity, transition density (movement within a neighborhood), incoming flow (movement into a neighborhood), and transition attractiveness (preference for moving between venue categories).
- Social Features (SF): Cohesiveness (network structure of users in a neighborhood), social groups (community detection), network centrality (user importance), and geo-social metrics (user influence in an area).
- Feature Weighting: The importance of each feature is determined by evaluating its predictability of actual store check-in numbers using a "Precision@X%" metric. Features that better predict top-performing stores receive higher weights.
- Customer Inference Algorithm (CIA): An iterative algorithm that estimates the potential customer numbers for unlabeled query locations. It starts with an initial estimate based on neighboring labeled stores and then iteratively refines these estimates by averaging the potential customer numbers of neighboring labeled and unlabeled nodes, weighted by the correlations learned in the CG. This process continues until the estimates converge.
Datasets
The study utilizes two primary location-based social network datasets:
- Foursquare: Data collected for New York City, containing 47,581 geo-tagged venues and 4,337,663 check-ins over a ten-month period (December 2010 to September 2011). This dataset was accessed indirectly through public Twitter check-ins and Foursquare venue data.
- Gowalla: A dataset collected by Cho et al. [3], comprising 196,591 users, 950,327 social connections, and 6,442,890 check-ins from February 2009 to October 2010.
The experiments focus on three popular retail chains: Starbucks (SB), McDonald's (MC), and Dunkin' Donuts (DD), using their store locations and historical check-in data.
Key Insights and Ideas
- Stores as "Sensors": The core idea of treating existing retail stores as sensors to infer customer potential in unserved areas is a clever and practical approach. It leverages existing infrastructure and data to solve a prediction problem.
- Spatial-Temporal Correlation is Key: The paper emphasizes that both where people are and when they are there are critical. The multi-layer correlation graph effectively captures these intertwined dependencies.
- Feature Importance Matters: Not all features contribute equally. The method's ability to dynamically weight features based on their predictive power for actual store performance is a significant strength, allowing the model to focus on what truly drives customer presence.
- Temporal Dynamics are Crucial: The performance difference between weekly and monthly predictions highlights the importance of temporal granularity. Monthly data, with its larger accumulated popularity, proved easier to estimate, suggesting that longer-term trends are more stable predictors.
Key Results
- Superior Performance: The proposed PCE model significantly outperforms seven competitive methods across various scenarios (retail chains, weekly/monthly predictions) in terms of Hit Rate and Normalized Discounted Cumulative Gain (NDCG). For instance, PCE achieved NDCG scores around 0.7-0.88, while other methods struggled to exceed 0.7.
- Effective "Hot Zone" Detection: PCE demonstrated strong capabilities in identifying locations with high potential customer numbers, achieving precision scores of 0.8 for the top 20% of locations. This suggests practical utility for identifying promising new store sites.
- Feature Effectiveness: Temporal Mobility Features (TMF) generally showed the highest importance, outperforming Geographical, Cumulative Mobility, and Social Features when used alone. However, combining all feature sets yielded the best overall results, indicating a synergistic effect.
- Radius Sensitivity: The optimal neighborhood radius for feature extraction was found to be around 200-300 meters, aligning with urban planning recommendations, suggesting that a well-defined local context is crucial.
Limitations and Future Work
- Data Granularity: The study notes that monthly predictions are generally more accurate than weekly ones, implying that finer temporal resolutions can be more challenging due to higher volatility.
- Feature Dependency: While TMF showed strong individual performance, the paper suggests that combining GF and SF with TMF and CMF provides the most comprehensive estimation. The relative importance of these features might vary by context.
- Future Directions: The authors aim to expand beyond location-based services by incorporating more diverse urban information, such as traffic status, weather patterns, and other near-by activities, to further enrich the modeling of potential customers.
Summary updated