Multi-target CNN-LSTM regressor for predicting urban distribution of short-term food delivery demand

Abstract

The food delivery market has increased rapidly in the last few years, becoming a well-established reality in the business world and a common feature of urban life. Food delivery platforms provide the end-to-end services that connect restaurants with consumers, including the delivery service to those people ordering food through an online portal. A key component of these platforms is logistics, specifically the logistics of drivers. Ideally, the number of drivers operating in an urban area should be just the right number to serve the demand in that area. Since the demand is extremely dynamic in space and time, the spatial–temporal distribution of drivers remains a challenging problem, partially solved by means of variable incentives in different city areas at different times. In this context, a precise demand prediction would avoid a local lack of drivers in some areas, and an inefficient concentration of drivers in some other areas. For this reason, we propose a deep neural network-based methodology to forecast short-term food delivery demand distribution over urban areas. The study, carried out on a real-world dataset from a food delivery company, focuses on hourly demands and frequent prediction updates. The sequential modeling approach, designed to catch rapid changes and sudden variations beyond the general demand trend, is based on a multi-target CNN-LSTM regressor trained on location-specific time series. The methodology uses a single model for all service areas simultaneously, and a single one-step volume inference for every area at each time update. The results disclose a better performance over baselines (historical estimates for the same time-area) and more traditional statistical approaches (moving averages and univariate time-series forecasting), demonstrating a promising implementation potential within an online delivery platform framework.