Product demand estimation for vending machines using video surveillance data: A group-lasso method

Lost sales information has significant impacts on the estimation of product demand and substitution. However, the difficulties to recognize such information in real applications make it rarely used in research. In this paper, we come up with a novel group-lasso based product demand model using the lost sales information, which is extracted from the video surveillance data provided by the cooperate retailer. A group-lasso method is used to characterize the substitution behaviours among each pair of SKUs. Then, an alternating minimization algorithm whose efficiency and convergence have been proved is designed to solve the model. To evaluate the model, we provide comparative experiments between the proposed method, time series forecasting and a naive method by applying these models to a real data set. The experiment results show that the proposed model obtains 8% higher accuracy at the total sales level and forecasts more accurately at the SKU sales level as well, which demonstrate its superiority. Moreover, we define a pair of welfare functions to measure the social impacts from both the retailer’s and customer’s end.