Multi-Source Learning for Sales Prediction

Abstract

With the convenience and popularity of Internet, the sales on e-commerce platforms have grown significantly. This exponential growth generates massive chunks of data, which provides the opportunity to utilize the historical data for predicting the customers' behaviors and can thus offer better services. One of the major tasks is to correctly estimate the coming sales of products since the e-retailers can reserve the products in a smart way. However, even with massive data, it is still challenging to estimate the sale amount of each product due to 1) the complicated language structures, 2) difficulties in integrating the features, and 3) missing values. To address these issues, we propose a framework for predicting the sales, which contains two phases: 1) feature extraction from product attributes and reviews, and 2) tensor decomposition for multi-source learning. The experimental results show that our framework outperforms baselines by 73%.