Approximate Temporal Conditioned Autoencoder and Regressor for Sales Prediction

Abstract

Lost sales can be caused by out-of-stock products, bad customer-care and over-pricing and is a major source of business failure in retail market. Predicting demand for products and customers' purchasing behaviour can help the businesses avoid these type of failures. Such predictions are usualy done using matrix factorization techniques. These are not suited to handle temporal variations in data which is often the case in retail world. Beyond factorization, deep learning methods including autoencoders and regressor are proposed without considering temporal information. Extending them with deep learning components to handle temporal data is not straightforward as these type of models requires a lot of data with variations in items and customers as well as in time. Variations in item and customer is easily achieved but variation in time is usually limited. We therefore propose a method which approximates a recurrent neural network to handle a few samples in temporal dimension. To this end we start with an existing algorithm DCAR [1] and enhance it to handle a few samples of temporal data resulting a predictor that beats the performance of classical matrix factorization and deep factorization methods as well as DCAR. The method simultaneously trains multiple autoencoders and regressors with an unfolded recurrent neural network.