Introduction
In the retail industry, the success of a new convenience store is a high-stakes investment often referred to as a gamble due to limited data. While modern Geographic Information Systems (GIS) provide extensive spatial data regarding roads, schools, and residential areas, a significant hurdle remains: commercial confidentiality. A retailer typically only knows the turnover (revenue) of their own stores (labeled data) and lacks access to competitor sales figures (unlabeled data). This paper introduces LapRLSC, a semi-supervised regression model designed to bridge this gap by leveraging both a retailer's own data and unlabeled competitor locations through a novel "profitability constraint."
Methods
The authors propose a two-part solution involving sophisticated feature engineering and a novel optimization model:
- Spatial Feature Extraction: Rather than simple radius-based counts, the model utilizes a Weighted Walk Distance Feature. Using GIS, the authors identify the shortest walkable paths to points of interest (POIs) like schools or bus stops, weighted by "human flow" (e.g., the number of residents or students). This captures nuances like physical barriers (e.g., highways) that simple distance metrics ignore.
- LapRLSC (Laplacian Regularized Least Squares with Constraints):
- Manifold Regularization: The model assumes that stores in similar geographic environments should exhibit similar turnovers.
- The Profitability Constraint: The most innovative aspect is the introduction of a global performance constraint. Assuming that most existing competitor stores are rational and profitable, the model requires their average turnover to stay above a certain threshold ($h$). This anchors the "unlabeled" competitor stores to a realistic performance floor, preventing wild predictions in data-sparse areas.
Datasets
The study utilizes a real-world dataset from a major city in China, comprising 272 convenience stores. Of these, 46 are labeled (belonging to the primary retailer) and 226 are unlabeled (competitors). The GIS data includes a wide variety of elements categorized into roads, geo-environments, buildings, and population centers.
Key Results
The authors compared LapRLSC against standard supervised methods and basic semi-supervised models (LapRLS). The results demonstrated that including weighted walk distance features consistently lowered the Mean Absolute Error (MAE).
Crucially, the LapRLSC model, which incorporates the average turnover constraint, significantly outperformed standard LapRLS. As shown in the performance analysis, the semi-supervised approach remains stable and accurate even with a very small training set of only 5 to 10 labeled stores, making it highly effective for retailers expanding into new territories where they have little historical data.
Conclusion and Future Work
The "profitability prior" introduced in this research effectively treats competitor success as a constraint rather than a missing variable, providing a robust framework for retail site evaluation. However, limitations remain, such as the labor-intensive manual pre-processing required for walk-distance calculations and the model's sensitivity to the parameter $h$ (the estimated competitor turnover threshold). Future work may explore automated road-network data processing and the application of non-linear kernels to further enhance predictive power.
Summary updated