Globally Connected: Mutual Influence Data in E-Commerce
This article is part of the Academic Alibaba series and is taken from the paper entitled “Globally Optimized Mutual Influence Aware Ranking in E-Commerce Search” by Tao Zhuang, Wenwu Ou, and Zhirong Wang, accepted by IJCAI 2018. The full paper can be read here.
When a customer on a large e-commerce platform issues a search query, they are inundated with thousands of relevant options. At first glance, it may seem that these search results are simply a ranked list of individual products. However, the results could more accurately be viewed as a web, each item’s attributes influencing the others’ place on the list and guiding the customer towards a purchase. While mutual influences between documents have been studied in the context of web search diversity, none of these studies have considered the unique needs of e-commerce platforms. Not until now, that is.
By modeling mutual influences and focusing on the ranking task, Alibaba’s tech team has proposed a global optimization framework to maximize gross merchandise volume (GMV) for ranking search results on Taobao, Alibaba’s e-commerce platform. Through this new optimization framework, researchers aim to guide customers to the products they want to buy and, in turn, maximize profits for online sellers.
Performance Far Beyond the Baseline
Researchers first developed a recursive neural network (RNN) model to consider mutual influences related to ranking orders. The team also created an attention mechanism to work in tandem with the RNN model to capture long-distance influences between products.
From there, Alibaba’s tech team took to the front lines. Researchers performed an online A/B test on Taobao Search, one of the largest e-commerce search engines in the world. This test compared a DNN model, an RNN model, and an RNN model with an attention mechanism, all of which included an extension that took into account an item’s global features. Over the course of a month, researchers analyzed the percentage of GMV increase over a baseline DNN that did not include global mutual influence data.
Alibaba’s miRNN and miRNN with attention model brought a substantial 5% in GMV over the baseline, verifying the value of Alibaba’s RNN method and confirming the importance of mutual influence on a global scale.
The Question of Latency
While miRNN with attention achieved the greatest GMV increase over baseline, implementing the attention model may not be feasible on a large scale. The computational cost of the RNN with attention model increases 400% over the baseline when rerank sizes reach fifty, presenting major drawbacks to e-commerce platforms seeking to maximize their GMV.
That said, the results achieved by the miDNN and miRNN models achieve excellent GMV results with only a minimal latency overhead, verifying the importance of mutual influence data in e-commerce and opening the door for a new area of research. Moving forward, researchers are focusing on more-efficient attention mechanisms to increase sellers’ GMV and customers’ satisfaction with fewer computations.
The full paper can be read here.