Learning from the Past: Fast NAS for price predictions

Abstract

In e-commerce sites, one of the most important tasks is pricing. The price has to be high enough to be profitable, while low enough to attract customers for competitors. One of the solutions is to use deep learning to predict a suitable price, with a suitable neural architecture. However, network architecture search has always been a time-consuming task, and yet essential for a good performance. The process depends on the experience of the person who designs it and the knowledge learned cannot be applied to similar tasks. This paper proposes a framework to utilize the architectures from similar datasets for the same tasks. Datasets are collected from 10 e-commerce websites with different types and targeted customers. Measurement is conducted on the data to obtain the correlation between different similarity measurements and the performance of a model to the datasets. It is proven that the model is 80% better than the baseline, and 20% faster than a full search.