Personalized Product Assortment with Real-time 3D Perception and Bayesian Payoff Estimation

TL;DR

EdgeRec3D tackles the real-world shelf-space allocation problem for physical retailers by fusing real-time 3D perception with uncertainty-aware payoff estimation. The system combines CountNet3D-based on-edge sales signals, a two-stage SpAGMM+K-means clustering to model demographic-driven preference heterogeneity, a Robust Bayesian Payoff (RBP) model with Penalized Expected Payoff per Facing (PEPF) to rank actions under noise, a candidate-generation module, and a heuristic search to manage combinatorial search. Offline evaluation and two field experiments demonstrate substantial sales gains, with +35.03% and +27.78% DID in the two experiments and a +9.4% deployment gain in a 28-week observational study, highlighting the value of edge perception and uncertainty-aware optimization in retail. The work provides a scalable, real-time framework for store-level assortment decisions and offers practical impact for brick-and-mortar merchants through improved alignment with shopper preferences.

Abstract

Product assortment selection is a critical challenge facing physical retailers. Effectively aligning inventory with the preferences of shoppers can increase sales and decrease out-of-stocks. However, in real-world settings the problem is challenging due to the combinatorial explosion of product assortment possibilities. Consumer preferences are typically heterogeneous across space and time, making inventory-preference alignment challenging. Additionally, existing strategies rely on syndicated data, which tends to be aggregated, low resolution, and suffer from high latency. To solve these challenges, we introduce a real-time recommendation system, which we call EdgeRec3D. Our system utilizes recent advances in 3D computer vision for perception and automatic, fine grained sales estimation. These perceptual components run on the edge of the network and facilitate real-time reward signals. Additionally, we develop a Bayesian payoff model to account for noisy estimates from 3D LIDAR data. We rely on spatial clustering to allow the system to adapt to heterogeneous consumer preferences, and a graph-based candidate generation algorithm to address the combinatorial search problem. We test our system in real-world stores across two, 6-8 week A/B tests with beverage products and demonstrate a 35% and 27% increase in sales respectively. Finally, we monitor the deployed system for a period of 28 weeks with an observational study and show a 9.4% increase in sales.

Figures (10)

• Figure 1: The product assortment problem can be described as choosing the set of products, and corresponding quantities, that will maximize expected payoff, subject to the total number of discrete locations (slots) within the product display. Our system maximizes expected payoff, while accounting for uncertainty.
• Figure 2: (a) Logical design of EdgeRec3D. The user performs successive 3D scans of a display to obtain product count predictions. These predictions are differenced over time following Equation \ref{['eq:sales']} to estimate sales. We train our two-stage spatial clustering pipeline to group demographically similar stores. Within each cluster, we estimate a Robust Bayesian Payoff (RBP) model to obtain a ranking statistic (PEPF) for each product. We generate candidates for each display, rank each candidate, and produce recommendation sets with heuristic search. (b) Physical design of EdgeRec3D. 3D perception is implemented at the edge of the network to facilitate real-time feedback. Data are aggregated across nodes for value estimation and candidate generation. When recommendations are served, we observe the state of the display in real-time and apply heuristic search to improve it.
• Figure 3: Deployment phase observational study. We split stores into high a low compliance groups as in the experimental phase. We see that the two groups have roughly parallel trends in the pre-deployment period. We deploy EdgeRec3D on April 16, 2023. In the post-deployment period, the two groups are still subject to the same seasonal trends, but the high compliance group diverges from the low compliance group. See Table \ref{['tab:deployment']} for full results.
• Figure 4: Representation of the Robust Bayesian Payoff (RBP) model using plate notation. $i$ indexes product, $k$ indexes cluster, and $l$ indexes store. Circles indicate random variables and squares indicate hyperparameters. The $\beta$ parameters describe how expected payoff changes as allocated product quantity increases and have hierarchical structure. This hierarchy facilitates store- and cluster-level preference estimation. The variance parameters are not expected to vary across stores or clusters and are estimated universally for each product.
• Figure 5: Example predictions from CountNet3D on real-world data. (Best viewed in color with zoom in). For each scene we show the image, the PointBeam proposals, the ground truth, and the predicted counts. We also display the scene-level MAPE (after rounding). Each beam is given a randomly generated color to highlight the beam regions.

Theorems & Definitions (2)

• Definition 1: Product Assortment Problem
• Definition 2: Demographic Spatial Clustering Problem