Joint Prediction of Multiple Waiting Times with Recurrent Neural Nets

点击翻译按钮获取中文摘要

Demand forecasting has always been important for retailers as it drives important decisions regarding businessstrategy and supply chain management. Traditional approaches model the purchase arrival rates using classicalstatistical models, which were developed without easy access to scalable computing power and both wide(many types of data) and long (for many individuals) purchase data. Benefits brought forth by the huge influxof data during the recent years ought not to be limited to having better estimates of aggregate statistics to applyclassical statistical models, but should also include a more fine-grained modelling based on the detailedindividual-level purchase information.There have been many models proposed in classical statistics to describe the behaviour of event arrivals, asthey're useful in many areas, such as monitoring equipment remaining lifetimes and disaster forecasting. Thesemodels have been meticulously described and their properties have been rigorously proven. However, thesesame models also require a large number of assumptions, which may not necessarily hold in reality. One suchassumption is that the dependence between events occurring further apart in time also tends to be smaller. We'dlike to explore a data-driven model that aims to relax these assumptions. The recent explosion in accessibilityto computational power has re-ignited interest in the Artificial Neural Net, which has shown success in manydifferent areas of application. A modification to this is the Recurrent Neural Net. With the addition of aninternal state variable, this neural-network set-up can be applied to sequential data as well. It had also beenshown to be Turing complete, meaning that it can emulate any computer program given a large enough set ofparameters. Applications of the RNN in natural language processing have been rather successful,demonstrating that this model is much better than classical statistical models in capturing complex sequentialdependence, since the classical models require short-range dependence assumptions to work.