TL;DR
This paper proposes a novel approach to model and predict customer in-store shopping behavior using Dynamic Bayesian Networks (DBNs). By leveraging RFID technology to track customer movements, the DBN model infers purchase interests and predicts future actions, paving the way for intelligent, personalized in-store marketing and recommendations.
What Problem Does This Paper Address?
Traditional retail environments struggle to understand customer behavior at a granular level. While RFID technology offers unprecedented opportunities for tracking in-store movements, effectively translating this raw data into actionable insights for personalized marketing remains a challenge. The paper addresses the need for a robust model that can infer a customer's product purchase intent and predict their next moves within the store, enabling real-time, location-aware marketing interventions.
Methods
The core of the proposed solution is a two-level Dynamic Bayesian Network (DBN) model. The upper level models a customer's purchase goals, representing what they intend to buy during a "trip segment" (a sequence of actions to purchase one product). The lower level models the transitions between store locations, capturing the customer's travel choices. Arcs in the DBN represent dependencies: purchase goals influence travel choices, and travel choices are influenced by previous locations and current goals. The model uses conditional probability distributions learned from observed shopping trip data to infer a customer's current purchase goal and predict their next location. This inference process is updated in real-time as new movement data becomes available.
Datasets
The study utilizes a real-world dataset comprising 843 RFID-collected shopping trips from the PathTracker system, developed by Sorensen Associates. This dataset includes information on 1074 distinct products, categorized into 38 product categories. The analysis focuses on customer purchase behavior at the product category level.
Key Insights and Ideas
- Bridging Movement and Intent: The DBN model effectively bridges the gap between raw spatial movement data (from RFID) and the underlying cognitive processes of a shopper (their purchase goals and decisions).
- Real-time Inference: The DBN's ability to update beliefs in real-time is crucial for enabling timely, location-aware marketing interventions.
- Predictive Power: The model demonstrates the capability to predict not only what a customer is likely to purchase but also where they are likely to go next within the store.
- Personalized Marketing Potential: The ultimate goal is to create an intelligent shopping environment where store operators can deliver targeted marketing efforts and personalized product recommendations "on the spot."
Key Results
The paper reports a preliminary evaluation of the DBN-based approach. The model demonstrated promising results in predicting both customer purchase goals and next locations. For a successful inference example, the predicted probabilities for the actual purchase goal rose significantly after the first few steps, reaching over 80% by one-third of the trip segment. Similarly, predicted next locations closely matched actual locations.
Quantitative evaluation using a 10-fold cross-validation procedure showed that the DBN method outperformed a widely-applied Bigram benchmark method in terms of average hit rates for both purchase and location predictions. The DBN achieved higher prediction accuracies across various percentages of the trip segment.
Limitations and Future Work
The authors acknowledge that this is a preliminary evaluation. Future work includes:
- Developing more sophisticated DBN network structures to incorporate additional information from the dataset, such as shopper velocity, deliberation time, and price information.
- Adding a network layer to differentiate various shopping activities (e.g., impulse vs. planned shopping).
- Extending the model to handle multiple conjunctive shopping goals within a single trip segment, as the current study assumes a single purchase goal per segment.
- Experimenting with additional benchmark methods and conducting further comparative evaluation studies.
Summary updated