Analysis of the Co-purchase Network of Products to Predict Amazon Sales-Rank

Abstract

Amazon sales-rank gives a relative estimate of a product item’s popularity among other items in the same category. An early prediction of the Amazon sales-rank of a product would imply an early guess of its sales-popularity relative to the other products on Amazon, which is one of the largest e-commerce hub across the globe. Traditional methods suggest use of product review related features, e.g., volume of reviews, text content of the reviews etc. for the purpose of prediction. In contrast, we propose in this paper for the first time a network-assisted approach to construct suitable features for prediction. In particular, we build a co-purchase network treating the individual products as nodes, with edges in between if two products are bought with one another. The way a product is positioned in this network (e.g., its centrality, clustering coefficient etc.) turns out to be a strong indicator of its sales-rank. This network-assisted approach has two distinct advantages over the traditional baseline method based on review analysis – (i) it works even if the product has no reviews (relevant especially in the early stages of the product launch) and (ii) it is notably more discriminative in classifying a popular (i.e., low sales-rank) product from an unpopular (i.e., high sales-rank) one. Based on this observation, we build a supervised model to early classify a popular product from an unpopular one. We report our results on two different product categories (CDs and cell phones) and obtain remarkably better classification accuracy compared to the baseline scheme. When the top 100 (700) products based on sales-rank are labelled as popular and the bottom 100 (700) are labelled as unpopular, the classification accuracy of our method is 89.85% (82.1%) for CDs and 84.11% (84.8%) for cell phones compared to 46.37% (68.75%) and 83.17% (71.95%) respectively from the baseline method.